]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
bf0f6f24 IM |
2 | /* |
3 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | |
4 | * | |
5 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
6 | * | |
7 | * Interactivity improvements by Mike Galbraith | |
8 | * (C) 2007 Mike Galbraith <efault@gmx.de> | |
9 | * | |
10 | * Various enhancements by Dmitry Adamushko. | |
11 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | |
12 | * | |
13 | * Group scheduling enhancements by Srivatsa Vaddagiri | |
14 | * Copyright IBM Corporation, 2007 | |
15 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | |
16 | * | |
17 | * Scaled math optimizations by Thomas Gleixner | |
18 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | |
21805085 PZ |
19 | * |
20 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
90eec103 | 21 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra |
bf0f6f24 | 22 | */ |
c4ad6fcb IM |
23 | #include <linux/energy_model.h> |
24 | #include <linux/mmap_lock.h> | |
25 | #include <linux/hugetlb_inline.h> | |
26 | #include <linux/jiffies.h> | |
27 | #include <linux/mm_api.h> | |
28 | #include <linux/highmem.h> | |
29 | #include <linux/spinlock_api.h> | |
30 | #include <linux/cpumask_api.h> | |
31 | #include <linux/lockdep_api.h> | |
32 | #include <linux/softirq.h> | |
33 | #include <linux/refcount_api.h> | |
34 | #include <linux/topology.h> | |
35 | #include <linux/sched/clock.h> | |
36 | #include <linux/sched/cond_resched.h> | |
37 | #include <linux/sched/cputime.h> | |
38 | #include <linux/sched/isolation.h> | |
d664e399 | 39 | #include <linux/sched/nohz.h> |
c4ad6fcb IM |
40 | |
41 | #include <linux/cpuidle.h> | |
42 | #include <linux/interrupt.h> | |
43 | #include <linux/mempolicy.h> | |
44 | #include <linux/mutex_api.h> | |
45 | #include <linux/profile.h> | |
46 | #include <linux/psi.h> | |
47 | #include <linux/ratelimit.h> | |
1930a6e7 | 48 | #include <linux/task_work.h> |
c4ad6fcb IM |
49 | |
50 | #include <asm/switch_to.h> | |
51 | ||
52 | #include <linux/sched/cond_resched.h> | |
53 | ||
325ea10c | 54 | #include "sched.h" |
b9e9c6ca IM |
55 | #include "stats.h" |
56 | #include "autogroup.h" | |
029632fb | 57 | |
bf0f6f24 | 58 | /* |
21805085 | 59 | * Targeted preemption latency for CPU-bound tasks: |
bf0f6f24 | 60 | * |
21805085 | 61 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
62 | * 'timeslice length' - timeslices in CFS are of variable length |
63 | * and have no persistent notion like in traditional, time-slice | |
64 | * based scheduling concepts. | |
bf0f6f24 | 65 | * |
d274a4ce IM |
66 | * (to see the precise effective timeslice length of your workload, |
67 | * run vmstat and monitor the context-switches (cs) field) | |
2b4d5b25 IM |
68 | * |
69 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 70 | */ |
2b4d5b25 | 71 | unsigned int sysctl_sched_latency = 6000000ULL; |
ed8885a1 | 72 | static unsigned int normalized_sysctl_sched_latency = 6000000ULL; |
2bd8e6d4 | 73 | |
1983a922 CE |
74 | /* |
75 | * The initial- and re-scaling of tunables is configurable | |
1983a922 CE |
76 | * |
77 | * Options are: | |
2b4d5b25 IM |
78 | * |
79 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
80 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
81 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
82 | * | |
83 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
1983a922 | 84 | */ |
8a99b683 | 85 | unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG; |
1983a922 | 86 | |
2bd8e6d4 | 87 | /* |
b2be5e96 | 88 | * Minimal preemption granularity for CPU-bound tasks: |
2b4d5b25 | 89 | * |
864616ee | 90 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 91 | */ |
ed8885a1 MS |
92 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
93 | static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 | 94 | |
51ce83ed JD |
95 | /* |
96 | * Minimal preemption granularity for CPU-bound SCHED_IDLE tasks. | |
97 | * Applies only when SCHED_IDLE tasks compete with normal tasks. | |
98 | * | |
99 | * (default: 0.75 msec) | |
100 | */ | |
101 | unsigned int sysctl_sched_idle_min_granularity = 750000ULL; | |
102 | ||
21805085 | 103 | /* |
2b4d5b25 | 104 | * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity |
b2be5e96 | 105 | */ |
0bf377bb | 106 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
107 | |
108 | /* | |
2bba22c5 | 109 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 110 | * parent will (try to) run first. |
21805085 | 111 | */ |
2bba22c5 | 112 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 113 | |
bf0f6f24 IM |
114 | /* |
115 | * SCHED_OTHER wake-up granularity. | |
bf0f6f24 IM |
116 | * |
117 | * This option delays the preemption effects of decoupled workloads | |
118 | * and reduces their over-scheduling. Synchronous workloads will still | |
119 | * have immediate wakeup/sleep latencies. | |
2b4d5b25 IM |
120 | * |
121 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 122 | */ |
ed8885a1 MS |
123 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
124 | static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; | |
bf0f6f24 | 125 | |
2b4d5b25 | 126 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
da84d961 | 127 | |
05289b90 TG |
128 | int sched_thermal_decay_shift; |
129 | static int __init setup_sched_thermal_decay_shift(char *str) | |
130 | { | |
131 | int _shift = 0; | |
132 | ||
133 | if (kstrtoint(str, 0, &_shift)) | |
134 | pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n"); | |
135 | ||
136 | sched_thermal_decay_shift = clamp(_shift, 0, 10); | |
137 | return 1; | |
138 | } | |
139 | __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift); | |
140 | ||
afe06efd TC |
141 | #ifdef CONFIG_SMP |
142 | /* | |
97fb7a0a | 143 | * For asym packing, by default the lower numbered CPU has higher priority. |
afe06efd TC |
144 | */ |
145 | int __weak arch_asym_cpu_priority(int cpu) | |
146 | { | |
147 | return -cpu; | |
148 | } | |
6d101ba6 OJ |
149 | |
150 | /* | |
60e17f5c | 151 | * The margin used when comparing utilization with CPU capacity. |
6d101ba6 OJ |
152 | * |
153 | * (default: ~20%) | |
154 | */ | |
60e17f5c VK |
155 | #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024) |
156 | ||
4aed8aa4 VS |
157 | /* |
158 | * The margin used when comparing CPU capacities. | |
159 | * is 'cap1' noticeably greater than 'cap2' | |
160 | * | |
161 | * (default: ~5%) | |
162 | */ | |
163 | #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078) | |
afe06efd TC |
164 | #endif |
165 | ||
ec12cb7f PT |
166 | #ifdef CONFIG_CFS_BANDWIDTH |
167 | /* | |
168 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
169 | * each time a cfs_rq requests quota. | |
170 | * | |
171 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
172 | * to consumption or the quota being specified to be smaller than the slice) | |
173 | * we will always only issue the remaining available time. | |
174 | * | |
2b4d5b25 IM |
175 | * (default: 5 msec, units: microseconds) |
176 | */ | |
d4ae80ff ZN |
177 | static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; |
178 | #endif | |
179 | ||
180 | #ifdef CONFIG_SYSCTL | |
181 | static struct ctl_table sched_fair_sysctls[] = { | |
182 | { | |
183 | .procname = "sched_child_runs_first", | |
184 | .data = &sysctl_sched_child_runs_first, | |
185 | .maxlen = sizeof(unsigned int), | |
186 | .mode = 0644, | |
187 | .proc_handler = proc_dointvec, | |
188 | }, | |
189 | #ifdef CONFIG_CFS_BANDWIDTH | |
190 | { | |
191 | .procname = "sched_cfs_bandwidth_slice_us", | |
192 | .data = &sysctl_sched_cfs_bandwidth_slice, | |
193 | .maxlen = sizeof(unsigned int), | |
194 | .mode = 0644, | |
195 | .proc_handler = proc_dointvec_minmax, | |
196 | .extra1 = SYSCTL_ONE, | |
197 | }, | |
198 | #endif | |
199 | {} | |
200 | }; | |
201 | ||
202 | static int __init sched_fair_sysctl_init(void) | |
203 | { | |
204 | register_sysctl_init("kernel", sched_fair_sysctls); | |
205 | return 0; | |
206 | } | |
207 | late_initcall(sched_fair_sysctl_init); | |
ec12cb7f PT |
208 | #endif |
209 | ||
8527632d PG |
210 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
211 | { | |
212 | lw->weight += inc; | |
213 | lw->inv_weight = 0; | |
214 | } | |
215 | ||
216 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
217 | { | |
218 | lw->weight -= dec; | |
219 | lw->inv_weight = 0; | |
220 | } | |
221 | ||
222 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
223 | { | |
224 | lw->weight = w; | |
225 | lw->inv_weight = 0; | |
226 | } | |
227 | ||
029632fb PZ |
228 | /* |
229 | * Increase the granularity value when there are more CPUs, | |
230 | * because with more CPUs the 'effective latency' as visible | |
231 | * to users decreases. But the relationship is not linear, | |
232 | * so pick a second-best guess by going with the log2 of the | |
233 | * number of CPUs. | |
234 | * | |
235 | * This idea comes from the SD scheduler of Con Kolivas: | |
236 | */ | |
58ac93e4 | 237 | static unsigned int get_update_sysctl_factor(void) |
029632fb | 238 | { |
58ac93e4 | 239 | unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); |
029632fb PZ |
240 | unsigned int factor; |
241 | ||
242 | switch (sysctl_sched_tunable_scaling) { | |
243 | case SCHED_TUNABLESCALING_NONE: | |
244 | factor = 1; | |
245 | break; | |
246 | case SCHED_TUNABLESCALING_LINEAR: | |
247 | factor = cpus; | |
248 | break; | |
249 | case SCHED_TUNABLESCALING_LOG: | |
250 | default: | |
251 | factor = 1 + ilog2(cpus); | |
252 | break; | |
253 | } | |
254 | ||
255 | return factor; | |
256 | } | |
257 | ||
258 | static void update_sysctl(void) | |
259 | { | |
260 | unsigned int factor = get_update_sysctl_factor(); | |
261 | ||
262 | #define SET_SYSCTL(name) \ | |
263 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
264 | SET_SYSCTL(sched_min_granularity); | |
265 | SET_SYSCTL(sched_latency); | |
266 | SET_SYSCTL(sched_wakeup_granularity); | |
267 | #undef SET_SYSCTL | |
268 | } | |
269 | ||
f38f12d1 | 270 | void __init sched_init_granularity(void) |
029632fb PZ |
271 | { |
272 | update_sysctl(); | |
273 | } | |
274 | ||
9dbdb155 | 275 | #define WMULT_CONST (~0U) |
029632fb PZ |
276 | #define WMULT_SHIFT 32 |
277 | ||
9dbdb155 PZ |
278 | static void __update_inv_weight(struct load_weight *lw) |
279 | { | |
280 | unsigned long w; | |
281 | ||
282 | if (likely(lw->inv_weight)) | |
283 | return; | |
284 | ||
285 | w = scale_load_down(lw->weight); | |
286 | ||
287 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
288 | lw->inv_weight = 1; | |
289 | else if (unlikely(!w)) | |
290 | lw->inv_weight = WMULT_CONST; | |
291 | else | |
292 | lw->inv_weight = WMULT_CONST / w; | |
293 | } | |
029632fb PZ |
294 | |
295 | /* | |
9dbdb155 PZ |
296 | * delta_exec * weight / lw.weight |
297 | * OR | |
298 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
299 | * | |
1c3de5e1 | 300 | * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case |
9dbdb155 PZ |
301 | * we're guaranteed shift stays positive because inv_weight is guaranteed to |
302 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
303 | * | |
304 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
305 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 306 | */ |
9dbdb155 | 307 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 308 | { |
9dbdb155 | 309 | u64 fact = scale_load_down(weight); |
1e17fb8e | 310 | u32 fact_hi = (u32)(fact >> 32); |
9dbdb155 | 311 | int shift = WMULT_SHIFT; |
1e17fb8e | 312 | int fs; |
029632fb | 313 | |
9dbdb155 | 314 | __update_inv_weight(lw); |
029632fb | 315 | |
1e17fb8e CC |
316 | if (unlikely(fact_hi)) { |
317 | fs = fls(fact_hi); | |
318 | shift -= fs; | |
319 | fact >>= fs; | |
029632fb PZ |
320 | } |
321 | ||
2eeb01a2 | 322 | fact = mul_u32_u32(fact, lw->inv_weight); |
029632fb | 323 | |
1e17fb8e CC |
324 | fact_hi = (u32)(fact >> 32); |
325 | if (fact_hi) { | |
326 | fs = fls(fact_hi); | |
327 | shift -= fs; | |
328 | fact >>= fs; | |
9dbdb155 | 329 | } |
029632fb | 330 | |
9dbdb155 | 331 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
332 | } |
333 | ||
334 | ||
335 | const struct sched_class fair_sched_class; | |
a4c2f00f | 336 | |
bf0f6f24 IM |
337 | /************************************************************** |
338 | * CFS operations on generic schedulable entities: | |
339 | */ | |
340 | ||
62160e3f | 341 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8f48894f | 342 | |
b758149c PZ |
343 | /* Walk up scheduling entities hierarchy */ |
344 | #define for_each_sched_entity(se) \ | |
345 | for (; se; se = se->parent) | |
346 | ||
f6783319 | 347 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 348 | { |
5d299eab PZ |
349 | struct rq *rq = rq_of(cfs_rq); |
350 | int cpu = cpu_of(rq); | |
351 | ||
352 | if (cfs_rq->on_list) | |
f6783319 | 353 | return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; |
5d299eab PZ |
354 | |
355 | cfs_rq->on_list = 1; | |
356 | ||
357 | /* | |
358 | * Ensure we either appear before our parent (if already | |
359 | * enqueued) or force our parent to appear after us when it is | |
360 | * enqueued. The fact that we always enqueue bottom-up | |
361 | * reduces this to two cases and a special case for the root | |
362 | * cfs_rq. Furthermore, it also means that we will always reset | |
363 | * tmp_alone_branch either when the branch is connected | |
364 | * to a tree or when we reach the top of the tree | |
365 | */ | |
366 | if (cfs_rq->tg->parent && | |
367 | cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { | |
67e86250 | 368 | /* |
5d299eab PZ |
369 | * If parent is already on the list, we add the child |
370 | * just before. Thanks to circular linked property of | |
371 | * the list, this means to put the child at the tail | |
372 | * of the list that starts by parent. | |
67e86250 | 373 | */ |
5d299eab PZ |
374 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, |
375 | &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); | |
376 | /* | |
377 | * The branch is now connected to its tree so we can | |
378 | * reset tmp_alone_branch to the beginning of the | |
379 | * list. | |
380 | */ | |
381 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 382 | return true; |
5d299eab | 383 | } |
3d4b47b4 | 384 | |
5d299eab PZ |
385 | if (!cfs_rq->tg->parent) { |
386 | /* | |
387 | * cfs rq without parent should be put | |
388 | * at the tail of the list. | |
389 | */ | |
390 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
391 | &rq->leaf_cfs_rq_list); | |
392 | /* | |
393 | * We have reach the top of a tree so we can reset | |
394 | * tmp_alone_branch to the beginning of the list. | |
395 | */ | |
396 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
f6783319 | 397 | return true; |
3d4b47b4 | 398 | } |
5d299eab PZ |
399 | |
400 | /* | |
401 | * The parent has not already been added so we want to | |
402 | * make sure that it will be put after us. | |
403 | * tmp_alone_branch points to the begin of the branch | |
404 | * where we will add parent. | |
405 | */ | |
406 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); | |
407 | /* | |
408 | * update tmp_alone_branch to points to the new begin | |
409 | * of the branch | |
410 | */ | |
411 | rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; | |
f6783319 | 412 | return false; |
3d4b47b4 PZ |
413 | } |
414 | ||
415 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
416 | { | |
417 | if (cfs_rq->on_list) { | |
31bc6aea VG |
418 | struct rq *rq = rq_of(cfs_rq); |
419 | ||
420 | /* | |
421 | * With cfs_rq being unthrottled/throttled during an enqueue, | |
422 | * it can happen the tmp_alone_branch points the a leaf that | |
423 | * we finally want to del. In this case, tmp_alone_branch moves | |
424 | * to the prev element but it will point to rq->leaf_cfs_rq_list | |
425 | * at the end of the enqueue. | |
426 | */ | |
427 | if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) | |
428 | rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; | |
429 | ||
3d4b47b4 PZ |
430 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); |
431 | cfs_rq->on_list = 0; | |
432 | } | |
433 | } | |
434 | ||
5d299eab PZ |
435 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
436 | { | |
437 | SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); | |
438 | } | |
439 | ||
039ae8bc VG |
440 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
441 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ | |
442 | list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \ | |
443 | leaf_cfs_rq_list) | |
b758149c PZ |
444 | |
445 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 446 | static inline struct cfs_rq * |
b758149c PZ |
447 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
448 | { | |
449 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 450 | return se->cfs_rq; |
b758149c | 451 | |
fed14d45 | 452 | return NULL; |
b758149c PZ |
453 | } |
454 | ||
455 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
456 | { | |
457 | return se->parent; | |
458 | } | |
459 | ||
464b7527 PZ |
460 | static void |
461 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
462 | { | |
463 | int se_depth, pse_depth; | |
464 | ||
465 | /* | |
466 | * preemption test can be made between sibling entities who are in the | |
467 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
468 | * both tasks until we find their ancestors who are siblings of common | |
469 | * parent. | |
470 | */ | |
471 | ||
472 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
473 | se_depth = (*se)->depth; |
474 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
475 | |
476 | while (se_depth > pse_depth) { | |
477 | se_depth--; | |
478 | *se = parent_entity(*se); | |
479 | } | |
480 | ||
481 | while (pse_depth > se_depth) { | |
482 | pse_depth--; | |
483 | *pse = parent_entity(*pse); | |
484 | } | |
485 | ||
486 | while (!is_same_group(*se, *pse)) { | |
487 | *se = parent_entity(*se); | |
488 | *pse = parent_entity(*pse); | |
489 | } | |
490 | } | |
491 | ||
30400039 JD |
492 | static int tg_is_idle(struct task_group *tg) |
493 | { | |
494 | return tg->idle > 0; | |
495 | } | |
496 | ||
497 | static int cfs_rq_is_idle(struct cfs_rq *cfs_rq) | |
498 | { | |
499 | return cfs_rq->idle > 0; | |
500 | } | |
501 | ||
502 | static int se_is_idle(struct sched_entity *se) | |
503 | { | |
504 | if (entity_is_task(se)) | |
505 | return task_has_idle_policy(task_of(se)); | |
506 | return cfs_rq_is_idle(group_cfs_rq(se)); | |
507 | } | |
508 | ||
8f48894f PZ |
509 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
510 | ||
b758149c PZ |
511 | #define for_each_sched_entity(se) \ |
512 | for (; se; se = NULL) | |
bf0f6f24 | 513 | |
f6783319 | 514 | static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
3d4b47b4 | 515 | { |
f6783319 | 516 | return true; |
3d4b47b4 PZ |
517 | } |
518 | ||
519 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
520 | { | |
521 | } | |
522 | ||
5d299eab PZ |
523 | static inline void assert_list_leaf_cfs_rq(struct rq *rq) |
524 | { | |
525 | } | |
526 | ||
039ae8bc VG |
527 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ |
528 | for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos) | |
b758149c | 529 | |
b758149c PZ |
530 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
531 | { | |
532 | return NULL; | |
533 | } | |
534 | ||
464b7527 PZ |
535 | static inline void |
536 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
537 | { | |
538 | } | |
539 | ||
366e7ad6 | 540 | static inline int tg_is_idle(struct task_group *tg) |
30400039 JD |
541 | { |
542 | return 0; | |
543 | } | |
544 | ||
545 | static int cfs_rq_is_idle(struct cfs_rq *cfs_rq) | |
546 | { | |
547 | return 0; | |
548 | } | |
549 | ||
550 | static int se_is_idle(struct sched_entity *se) | |
551 | { | |
552 | return 0; | |
553 | } | |
554 | ||
b758149c PZ |
555 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
556 | ||
6c16a6dc | 557 | static __always_inline |
9dbdb155 | 558 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
559 | |
560 | /************************************************************** | |
561 | * Scheduling class tree data structure manipulation methods: | |
562 | */ | |
563 | ||
1bf08230 | 564 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 565 | { |
1bf08230 | 566 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 567 | if (delta > 0) |
1bf08230 | 568 | max_vruntime = vruntime; |
02e0431a | 569 | |
1bf08230 | 570 | return max_vruntime; |
02e0431a PZ |
571 | } |
572 | ||
0702e3eb | 573 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
574 | { |
575 | s64 delta = (s64)(vruntime - min_vruntime); | |
576 | if (delta < 0) | |
577 | min_vruntime = vruntime; | |
578 | ||
579 | return min_vruntime; | |
580 | } | |
581 | ||
bf9be9a1 | 582 | static inline bool entity_before(struct sched_entity *a, |
54fdc581 FC |
583 | struct sched_entity *b) |
584 | { | |
585 | return (s64)(a->vruntime - b->vruntime) < 0; | |
586 | } | |
587 | ||
bf9be9a1 PZ |
588 | #define __node_2_se(node) \ |
589 | rb_entry((node), struct sched_entity, run_node) | |
590 | ||
1af5f730 PZ |
591 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
592 | { | |
b60205c7 | 593 | struct sched_entity *curr = cfs_rq->curr; |
bfb06889 | 594 | struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); |
b60205c7 | 595 | |
1af5f730 PZ |
596 | u64 vruntime = cfs_rq->min_vruntime; |
597 | ||
b60205c7 PZ |
598 | if (curr) { |
599 | if (curr->on_rq) | |
600 | vruntime = curr->vruntime; | |
601 | else | |
602 | curr = NULL; | |
603 | } | |
1af5f730 | 604 | |
bfb06889 | 605 | if (leftmost) { /* non-empty tree */ |
bf9be9a1 | 606 | struct sched_entity *se = __node_2_se(leftmost); |
1af5f730 | 607 | |
b60205c7 | 608 | if (!curr) |
1af5f730 PZ |
609 | vruntime = se->vruntime; |
610 | else | |
611 | vruntime = min_vruntime(vruntime, se->vruntime); | |
612 | } | |
613 | ||
1bf08230 | 614 | /* ensure we never gain time by being placed backwards. */ |
d05b4305 VD |
615 | u64_u32_store(cfs_rq->min_vruntime, |
616 | max_vruntime(cfs_rq->min_vruntime, vruntime)); | |
1af5f730 PZ |
617 | } |
618 | ||
bf9be9a1 PZ |
619 | static inline bool __entity_less(struct rb_node *a, const struct rb_node *b) |
620 | { | |
621 | return entity_before(__node_2_se(a), __node_2_se(b)); | |
622 | } | |
623 | ||
bf0f6f24 IM |
624 | /* |
625 | * Enqueue an entity into the rb-tree: | |
626 | */ | |
0702e3eb | 627 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 628 | { |
bf9be9a1 | 629 | rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less); |
bf0f6f24 IM |
630 | } |
631 | ||
0702e3eb | 632 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 633 | { |
bfb06889 | 634 | rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
635 | } |
636 | ||
029632fb | 637 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 638 | { |
bfb06889 | 639 | struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); |
f4b6755f PZ |
640 | |
641 | if (!left) | |
642 | return NULL; | |
643 | ||
bf9be9a1 | 644 | return __node_2_se(left); |
bf0f6f24 IM |
645 | } |
646 | ||
ac53db59 RR |
647 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
648 | { | |
649 | struct rb_node *next = rb_next(&se->run_node); | |
650 | ||
651 | if (!next) | |
652 | return NULL; | |
653 | ||
bf9be9a1 | 654 | return __node_2_se(next); |
ac53db59 RR |
655 | } |
656 | ||
657 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 658 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 659 | { |
bfb06889 | 660 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); |
aeb73b04 | 661 | |
70eee74b BS |
662 | if (!last) |
663 | return NULL; | |
7eee3e67 | 664 | |
bf9be9a1 | 665 | return __node_2_se(last); |
aeb73b04 PZ |
666 | } |
667 | ||
bf0f6f24 IM |
668 | /************************************************************** |
669 | * Scheduling class statistics methods: | |
670 | */ | |
671 | ||
8a99b683 | 672 | int sched_update_scaling(void) |
b2be5e96 | 673 | { |
58ac93e4 | 674 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 | 675 | |
b2be5e96 PZ |
676 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, |
677 | sysctl_sched_min_granularity); | |
678 | ||
acb4a848 CE |
679 | #define WRT_SYSCTL(name) \ |
680 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
681 | WRT_SYSCTL(sched_min_granularity); | |
682 | WRT_SYSCTL(sched_latency); | |
683 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
684 | #undef WRT_SYSCTL |
685 | ||
b2be5e96 PZ |
686 | return 0; |
687 | } | |
688 | #endif | |
647e7cac | 689 | |
a7be37ac | 690 | /* |
f9c0b095 | 691 | * delta /= w |
a7be37ac | 692 | */ |
9dbdb155 | 693 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 694 | { |
f9c0b095 | 695 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 696 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
697 | |
698 | return delta; | |
699 | } | |
700 | ||
647e7cac IM |
701 | /* |
702 | * The idea is to set a period in which each task runs once. | |
703 | * | |
532b1858 | 704 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
705 | * this period because otherwise the slices get too small. |
706 | * | |
707 | * p = (nr <= nl) ? l : l*nr/nl | |
708 | */ | |
4d78e7b6 PZ |
709 | static u64 __sched_period(unsigned long nr_running) |
710 | { | |
8e2b0bf3 BF |
711 | if (unlikely(nr_running > sched_nr_latency)) |
712 | return nr_running * sysctl_sched_min_granularity; | |
713 | else | |
714 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
715 | } |
716 | ||
51ce83ed JD |
717 | static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq); |
718 | ||
647e7cac IM |
719 | /* |
720 | * We calculate the wall-time slice from the period by taking a part | |
721 | * proportional to the weight. | |
722 | * | |
f9c0b095 | 723 | * s = p*P[w/rw] |
647e7cac | 724 | */ |
6d0f0ebd | 725 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 726 | { |
0c2de3f0 | 727 | unsigned int nr_running = cfs_rq->nr_running; |
51ce83ed JD |
728 | struct sched_entity *init_se = se; |
729 | unsigned int min_gran; | |
0c2de3f0 PZ |
730 | u64 slice; |
731 | ||
732 | if (sched_feat(ALT_PERIOD)) | |
733 | nr_running = rq_of(cfs_rq)->cfs.h_nr_running; | |
734 | ||
735 | slice = __sched_period(nr_running + !se->on_rq); | |
f9c0b095 | 736 | |
0a582440 | 737 | for_each_sched_entity(se) { |
6272d68c | 738 | struct load_weight *load; |
3104bf03 | 739 | struct load_weight lw; |
51ce83ed | 740 | struct cfs_rq *qcfs_rq; |
6272d68c | 741 | |
51ce83ed JD |
742 | qcfs_rq = cfs_rq_of(se); |
743 | load = &qcfs_rq->load; | |
f9c0b095 | 744 | |
0a582440 | 745 | if (unlikely(!se->on_rq)) { |
51ce83ed | 746 | lw = qcfs_rq->load; |
0a582440 MG |
747 | |
748 | update_load_add(&lw, se->load.weight); | |
749 | load = &lw; | |
750 | } | |
9dbdb155 | 751 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 | 752 | } |
0c2de3f0 | 753 | |
51ce83ed JD |
754 | if (sched_feat(BASE_SLICE)) { |
755 | if (se_is_idle(init_se) && !sched_idle_cfs_rq(cfs_rq)) | |
756 | min_gran = sysctl_sched_idle_min_granularity; | |
757 | else | |
758 | min_gran = sysctl_sched_min_granularity; | |
759 | ||
760 | slice = max_t(u64, slice, min_gran); | |
761 | } | |
0c2de3f0 | 762 | |
0a582440 | 763 | return slice; |
bf0f6f24 IM |
764 | } |
765 | ||
647e7cac | 766 | /* |
660cc00f | 767 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 768 | * |
f9c0b095 | 769 | * vs = s/w |
647e7cac | 770 | */ |
f9c0b095 | 771 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 772 | { |
f9c0b095 | 773 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
774 | } |
775 | ||
c0796298 | 776 | #include "pelt.h" |
23127296 | 777 | #ifdef CONFIG_SMP |
283e2ed3 | 778 | |
772bd008 | 779 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee | 780 | static unsigned long task_h_load(struct task_struct *p); |
3b1baa64 | 781 | static unsigned long capacity_of(int cpu); |
fb13c7ee | 782 | |
540247fb YD |
783 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
784 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 785 | { |
540247fb | 786 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 787 | |
f207934f PZ |
788 | memset(sa, 0, sizeof(*sa)); |
789 | ||
b5a9b340 | 790 | /* |
dfcb245e | 791 | * Tasks are initialized with full load to be seen as heavy tasks until |
b5a9b340 | 792 | * they get a chance to stabilize to their real load level. |
dfcb245e | 793 | * Group entities are initialized with zero load to reflect the fact that |
b5a9b340 VG |
794 | * nothing has been attached to the task group yet. |
795 | */ | |
796 | if (entity_is_task(se)) | |
0dacee1b | 797 | sa->load_avg = scale_load_down(se->load.weight); |
f207934f | 798 | |
9d89c257 | 799 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 800 | } |
7ea241af | 801 | |
2b8c41da YD |
802 | /* |
803 | * With new tasks being created, their initial util_avgs are extrapolated | |
804 | * based on the cfs_rq's current util_avg: | |
805 | * | |
806 | * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight | |
807 | * | |
808 | * However, in many cases, the above util_avg does not give a desired | |
809 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
810 | * as when the series is a harmonic series. | |
811 | * | |
812 | * To solve this problem, we also cap the util_avg of successive tasks to | |
813 | * only 1/2 of the left utilization budget: | |
814 | * | |
8fe5c5a9 | 815 | * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n |
2b8c41da | 816 | * |
8fe5c5a9 | 817 | * where n denotes the nth task and cpu_scale the CPU capacity. |
2b8c41da | 818 | * |
8fe5c5a9 QP |
819 | * For example, for a CPU with 1024 of capacity, a simplest series from |
820 | * the beginning would be like: | |
2b8c41da YD |
821 | * |
822 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
823 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
824 | * | |
825 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
826 | * if util_avg > util_avg_cap. | |
827 | */ | |
d0fe0b9c | 828 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da | 829 | { |
d0fe0b9c | 830 | struct sched_entity *se = &p->se; |
2b8c41da YD |
831 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
832 | struct sched_avg *sa = &se->avg; | |
8ec59c0f | 833 | long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))); |
8fe5c5a9 | 834 | long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; |
2b8c41da YD |
835 | |
836 | if (cap > 0) { | |
837 | if (cfs_rq->avg.util_avg != 0) { | |
838 | sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; | |
839 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); | |
840 | ||
841 | if (sa->util_avg > cap) | |
842 | sa->util_avg = cap; | |
843 | } else { | |
844 | sa->util_avg = cap; | |
845 | } | |
2b8c41da | 846 | } |
7dc603c9 | 847 | |
e21cf434 | 848 | sa->runnable_avg = sa->util_avg; |
9f683953 | 849 | |
d0fe0b9c DE |
850 | if (p->sched_class != &fair_sched_class) { |
851 | /* | |
852 | * For !fair tasks do: | |
853 | * | |
854 | update_cfs_rq_load_avg(now, cfs_rq); | |
a4f9a0e5 | 855 | attach_entity_load_avg(cfs_rq, se); |
d0fe0b9c DE |
856 | switched_from_fair(rq, p); |
857 | * | |
858 | * such that the next switched_to_fair() has the | |
859 | * expected state. | |
860 | */ | |
861 | se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); | |
862 | return; | |
7dc603c9 | 863 | } |
2b8c41da YD |
864 | } |
865 | ||
7dc603c9 | 866 | #else /* !CONFIG_SMP */ |
540247fb | 867 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
868 | { |
869 | } | |
d0fe0b9c | 870 | void post_init_entity_util_avg(struct task_struct *p) |
2b8c41da YD |
871 | { |
872 | } | |
fe749158 | 873 | static void update_tg_load_avg(struct cfs_rq *cfs_rq) |
3d30544f PZ |
874 | { |
875 | } | |
7dc603c9 | 876 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 877 | |
bf0f6f24 | 878 | /* |
9dbdb155 | 879 | * Update the current task's runtime statistics. |
bf0f6f24 | 880 | */ |
b7cc0896 | 881 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 882 | { |
429d43bc | 883 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 884 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 885 | u64 delta_exec; |
bf0f6f24 IM |
886 | |
887 | if (unlikely(!curr)) | |
888 | return; | |
889 | ||
9dbdb155 PZ |
890 | delta_exec = now - curr->exec_start; |
891 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 892 | return; |
bf0f6f24 | 893 | |
8ebc91d9 | 894 | curr->exec_start = now; |
d842de87 | 895 | |
ceeadb83 YS |
896 | if (schedstat_enabled()) { |
897 | struct sched_statistics *stats; | |
898 | ||
899 | stats = __schedstats_from_se(curr); | |
900 | __schedstat_set(stats->exec_max, | |
901 | max(delta_exec, stats->exec_max)); | |
902 | } | |
9dbdb155 PZ |
903 | |
904 | curr->sum_exec_runtime += delta_exec; | |
ae92882e | 905 | schedstat_add(cfs_rq->exec_clock, delta_exec); |
9dbdb155 PZ |
906 | |
907 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
908 | update_min_vruntime(cfs_rq); | |
909 | ||
d842de87 SV |
910 | if (entity_is_task(curr)) { |
911 | struct task_struct *curtask = task_of(curr); | |
912 | ||
f977bb49 | 913 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d2cc5ed6 | 914 | cgroup_account_cputime(curtask, delta_exec); |
f06febc9 | 915 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 916 | } |
ec12cb7f PT |
917 | |
918 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
919 | } |
920 | ||
6e998916 SG |
921 | static void update_curr_fair(struct rq *rq) |
922 | { | |
923 | update_curr(cfs_rq_of(&rq->curr->se)); | |
924 | } | |
925 | ||
bf0f6f24 | 926 | static inline void |
60f2415e | 927 | update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 928 | { |
ceeadb83 | 929 | struct sched_statistics *stats; |
60f2415e | 930 | struct task_struct *p = NULL; |
4fa8d299 JP |
931 | |
932 | if (!schedstat_enabled()) | |
933 | return; | |
934 | ||
ceeadb83 YS |
935 | stats = __schedstats_from_se(se); |
936 | ||
60f2415e YS |
937 | if (entity_is_task(se)) |
938 | p = task_of(se); | |
3ea94de1 | 939 | |
60f2415e | 940 | __update_stats_wait_start(rq_of(cfs_rq), p, stats); |
bf0f6f24 IM |
941 | } |
942 | ||
4fa8d299 | 943 | static inline void |
60f2415e | 944 | update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3ea94de1 | 945 | { |
ceeadb83 YS |
946 | struct sched_statistics *stats; |
947 | struct task_struct *p = NULL; | |
cb251765 | 948 | |
4fa8d299 JP |
949 | if (!schedstat_enabled()) |
950 | return; | |
951 | ||
ceeadb83 YS |
952 | stats = __schedstats_from_se(se); |
953 | ||
b9c88f75 | 954 | /* |
955 | * When the sched_schedstat changes from 0 to 1, some sched se | |
956 | * maybe already in the runqueue, the se->statistics.wait_start | |
957 | * will be 0.So it will let the delta wrong. We need to avoid this | |
958 | * scenario. | |
959 | */ | |
ceeadb83 | 960 | if (unlikely(!schedstat_val(stats->wait_start))) |
b9c88f75 | 961 | return; |
962 | ||
60f2415e | 963 | if (entity_is_task(se)) |
3ea94de1 | 964 | p = task_of(se); |
3ea94de1 | 965 | |
60f2415e | 966 | __update_stats_wait_end(rq_of(cfs_rq), p, stats); |
3ea94de1 | 967 | } |
3ea94de1 | 968 | |
4fa8d299 | 969 | static inline void |
60f2415e | 970 | update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se) |
1a3d027c | 971 | { |
ceeadb83 | 972 | struct sched_statistics *stats; |
1a3d027c | 973 | struct task_struct *tsk = NULL; |
4fa8d299 JP |
974 | |
975 | if (!schedstat_enabled()) | |
976 | return; | |
977 | ||
ceeadb83 YS |
978 | stats = __schedstats_from_se(se); |
979 | ||
1a3d027c JP |
980 | if (entity_is_task(se)) |
981 | tsk = task_of(se); | |
982 | ||
60f2415e | 983 | __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats); |
3ea94de1 | 984 | } |
3ea94de1 | 985 | |
bf0f6f24 IM |
986 | /* |
987 | * Task is being enqueued - update stats: | |
988 | */ | |
cb251765 | 989 | static inline void |
60f2415e | 990 | update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 991 | { |
4fa8d299 JP |
992 | if (!schedstat_enabled()) |
993 | return; | |
994 | ||
bf0f6f24 IM |
995 | /* |
996 | * Are we enqueueing a waiting task? (for current tasks | |
997 | * a dequeue/enqueue event is a NOP) | |
998 | */ | |
429d43bc | 999 | if (se != cfs_rq->curr) |
60f2415e | 1000 | update_stats_wait_start_fair(cfs_rq, se); |
1a3d027c JP |
1001 | |
1002 | if (flags & ENQUEUE_WAKEUP) | |
60f2415e | 1003 | update_stats_enqueue_sleeper_fair(cfs_rq, se); |
bf0f6f24 IM |
1004 | } |
1005 | ||
bf0f6f24 | 1006 | static inline void |
60f2415e | 1007 | update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1008 | { |
4fa8d299 JP |
1009 | |
1010 | if (!schedstat_enabled()) | |
1011 | return; | |
1012 | ||
bf0f6f24 IM |
1013 | /* |
1014 | * Mark the end of the wait period if dequeueing a | |
1015 | * waiting task: | |
1016 | */ | |
429d43bc | 1017 | if (se != cfs_rq->curr) |
60f2415e | 1018 | update_stats_wait_end_fair(cfs_rq, se); |
cb251765 | 1019 | |
4fa8d299 JP |
1020 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
1021 | struct task_struct *tsk = task_of(se); | |
2f064a59 | 1022 | unsigned int state; |
cb251765 | 1023 | |
2f064a59 PZ |
1024 | /* XXX racy against TTWU */ |
1025 | state = READ_ONCE(tsk->__state); | |
1026 | if (state & TASK_INTERRUPTIBLE) | |
ceeadb83 | 1027 | __schedstat_set(tsk->stats.sleep_start, |
4fa8d299 | 1028 | rq_clock(rq_of(cfs_rq))); |
2f064a59 | 1029 | if (state & TASK_UNINTERRUPTIBLE) |
ceeadb83 | 1030 | __schedstat_set(tsk->stats.block_start, |
4fa8d299 | 1031 | rq_clock(rq_of(cfs_rq))); |
cb251765 | 1032 | } |
cb251765 MG |
1033 | } |
1034 | ||
bf0f6f24 IM |
1035 | /* |
1036 | * We are picking a new current task - update its stats: | |
1037 | */ | |
1038 | static inline void | |
79303e9e | 1039 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
1040 | { |
1041 | /* | |
1042 | * We are starting a new run period: | |
1043 | */ | |
78becc27 | 1044 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1045 | } |
1046 | ||
bf0f6f24 IM |
1047 | /************************************************** |
1048 | * Scheduling class queueing methods: | |
1049 | */ | |
1050 | ||
cb29a5c1 MG |
1051 | #ifdef CONFIG_NUMA |
1052 | #define NUMA_IMBALANCE_MIN 2 | |
1053 | ||
1054 | static inline long | |
1055 | adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr) | |
1056 | { | |
1057 | /* | |
1058 | * Allow a NUMA imbalance if busy CPUs is less than the maximum | |
1059 | * threshold. Above this threshold, individual tasks may be contending | |
1060 | * for both memory bandwidth and any shared HT resources. This is an | |
1061 | * approximation as the number of running tasks may not be related to | |
1062 | * the number of busy CPUs due to sched_setaffinity. | |
1063 | */ | |
1064 | if (dst_running > imb_numa_nr) | |
1065 | return imbalance; | |
1066 | ||
1067 | /* | |
1068 | * Allow a small imbalance based on a simple pair of communicating | |
1069 | * tasks that remain local when the destination is lightly loaded. | |
1070 | */ | |
1071 | if (imbalance <= NUMA_IMBALANCE_MIN) | |
1072 | return 0; | |
1073 | ||
1074 | return imbalance; | |
1075 | } | |
1076 | #endif /* CONFIG_NUMA */ | |
1077 | ||
cbee9f88 PZ |
1078 | #ifdef CONFIG_NUMA_BALANCING |
1079 | /* | |
598f0ec0 MG |
1080 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1081 | * calculated based on the tasks virtual memory size and | |
1082 | * numa_balancing_scan_size. | |
cbee9f88 | 1083 | */ |
598f0ec0 MG |
1084 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1085 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1086 | |
1087 | /* Portion of address space to scan in MB */ | |
1088 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1089 | |
4b96a29b PZ |
1090 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1091 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1092 | ||
b5dd77c8 | 1093 | struct numa_group { |
c45a7795 | 1094 | refcount_t refcount; |
b5dd77c8 RR |
1095 | |
1096 | spinlock_t lock; /* nr_tasks, tasks */ | |
1097 | int nr_tasks; | |
1098 | pid_t gid; | |
1099 | int active_nodes; | |
1100 | ||
1101 | struct rcu_head rcu; | |
1102 | unsigned long total_faults; | |
1103 | unsigned long max_faults_cpu; | |
1104 | /* | |
5b763a14 BR |
1105 | * faults[] array is split into two regions: faults_mem and faults_cpu. |
1106 | * | |
b5dd77c8 RR |
1107 | * Faults_cpu is used to decide whether memory should move |
1108 | * towards the CPU. As a consequence, these stats are weighted | |
1109 | * more by CPU use than by memory faults. | |
1110 | */ | |
04f5c362 | 1111 | unsigned long faults[]; |
b5dd77c8 RR |
1112 | }; |
1113 | ||
cb361d8c JH |
1114 | /* |
1115 | * For functions that can be called in multiple contexts that permit reading | |
1116 | * ->numa_group (see struct task_struct for locking rules). | |
1117 | */ | |
1118 | static struct numa_group *deref_task_numa_group(struct task_struct *p) | |
1119 | { | |
1120 | return rcu_dereference_check(p->numa_group, p == current || | |
9ef7e7e3 | 1121 | (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu))); |
cb361d8c JH |
1122 | } |
1123 | ||
1124 | static struct numa_group *deref_curr_numa_group(struct task_struct *p) | |
1125 | { | |
1126 | return rcu_dereference_protected(p->numa_group, p == current); | |
1127 | } | |
1128 | ||
b5dd77c8 RR |
1129 | static inline unsigned long group_faults_priv(struct numa_group *ng); |
1130 | static inline unsigned long group_faults_shared(struct numa_group *ng); | |
1131 | ||
598f0ec0 MG |
1132 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1133 | { | |
1134 | unsigned long rss = 0; | |
1135 | unsigned long nr_scan_pages; | |
1136 | ||
1137 | /* | |
1138 | * Calculations based on RSS as non-present and empty pages are skipped | |
1139 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1140 | * on resident pages | |
1141 | */ | |
1142 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1143 | rss = get_mm_rss(p->mm); | |
1144 | if (!rss) | |
1145 | rss = nr_scan_pages; | |
1146 | ||
1147 | rss = round_up(rss, nr_scan_pages); | |
1148 | return rss / nr_scan_pages; | |
1149 | } | |
1150 | ||
3b03706f | 1151 | /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ |
598f0ec0 MG |
1152 | #define MAX_SCAN_WINDOW 2560 |
1153 | ||
1154 | static unsigned int task_scan_min(struct task_struct *p) | |
1155 | { | |
316c1608 | 1156 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1157 | unsigned int scan, floor; |
1158 | unsigned int windows = 1; | |
1159 | ||
64192658 KT |
1160 | if (scan_size < MAX_SCAN_WINDOW) |
1161 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1162 | floor = 1000 / windows; |
1163 | ||
1164 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1165 | return max_t(unsigned int, floor, scan); | |
1166 | } | |
1167 | ||
b5dd77c8 RR |
1168 | static unsigned int task_scan_start(struct task_struct *p) |
1169 | { | |
1170 | unsigned long smin = task_scan_min(p); | |
1171 | unsigned long period = smin; | |
cb361d8c | 1172 | struct numa_group *ng; |
b5dd77c8 RR |
1173 | |
1174 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1175 | rcu_read_lock(); |
1176 | ng = rcu_dereference(p->numa_group); | |
1177 | if (ng) { | |
b5dd77c8 RR |
1178 | unsigned long shared = group_faults_shared(ng); |
1179 | unsigned long private = group_faults_priv(ng); | |
1180 | ||
c45a7795 | 1181 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1182 | period *= shared + 1; |
1183 | period /= private + shared + 1; | |
1184 | } | |
cb361d8c | 1185 | rcu_read_unlock(); |
b5dd77c8 RR |
1186 | |
1187 | return max(smin, period); | |
1188 | } | |
1189 | ||
598f0ec0 MG |
1190 | static unsigned int task_scan_max(struct task_struct *p) |
1191 | { | |
b5dd77c8 RR |
1192 | unsigned long smin = task_scan_min(p); |
1193 | unsigned long smax; | |
cb361d8c | 1194 | struct numa_group *ng; |
598f0ec0 MG |
1195 | |
1196 | /* Watch for min being lower than max due to floor calculations */ | |
1197 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
b5dd77c8 RR |
1198 | |
1199 | /* Scale the maximum scan period with the amount of shared memory. */ | |
cb361d8c JH |
1200 | ng = deref_curr_numa_group(p); |
1201 | if (ng) { | |
b5dd77c8 RR |
1202 | unsigned long shared = group_faults_shared(ng); |
1203 | unsigned long private = group_faults_priv(ng); | |
1204 | unsigned long period = smax; | |
1205 | ||
c45a7795 | 1206 | period *= refcount_read(&ng->refcount); |
b5dd77c8 RR |
1207 | period *= shared + 1; |
1208 | period /= private + shared + 1; | |
1209 | ||
1210 | smax = max(smax, period); | |
1211 | } | |
1212 | ||
598f0ec0 MG |
1213 | return max(smin, smax); |
1214 | } | |
1215 | ||
0ec8aa00 PZ |
1216 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1217 | { | |
98fa15f3 | 1218 | rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1219 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); |
1220 | } | |
1221 | ||
1222 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1223 | { | |
98fa15f3 | 1224 | rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE); |
0ec8aa00 PZ |
1225 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); |
1226 | } | |
1227 | ||
be1e4e76 RR |
1228 | /* Shared or private faults. */ |
1229 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1230 | ||
1231 | /* Memory and CPU locality */ | |
1232 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1233 | ||
1234 | /* Averaged statistics, and temporary buffers. */ | |
1235 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1236 | ||
e29cf08b MG |
1237 | pid_t task_numa_group_id(struct task_struct *p) |
1238 | { | |
cb361d8c JH |
1239 | struct numa_group *ng; |
1240 | pid_t gid = 0; | |
1241 | ||
1242 | rcu_read_lock(); | |
1243 | ng = rcu_dereference(p->numa_group); | |
1244 | if (ng) | |
1245 | gid = ng->gid; | |
1246 | rcu_read_unlock(); | |
1247 | ||
1248 | return gid; | |
e29cf08b MG |
1249 | } |
1250 | ||
44dba3d5 | 1251 | /* |
97fb7a0a | 1252 | * The averaged statistics, shared & private, memory & CPU, |
44dba3d5 IM |
1253 | * occupy the first half of the array. The second half of the |
1254 | * array is for current counters, which are averaged into the | |
1255 | * first set by task_numa_placement. | |
1256 | */ | |
1257 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1258 | { |
44dba3d5 | 1259 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1260 | } |
1261 | ||
1262 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1263 | { | |
44dba3d5 | 1264 | if (!p->numa_faults) |
ac8e895b MG |
1265 | return 0; |
1266 | ||
44dba3d5 IM |
1267 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1268 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1269 | } |
1270 | ||
83e1d2cd MG |
1271 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1272 | { | |
cb361d8c JH |
1273 | struct numa_group *ng = deref_task_numa_group(p); |
1274 | ||
1275 | if (!ng) | |
83e1d2cd MG |
1276 | return 0; |
1277 | ||
cb361d8c JH |
1278 | return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1279 | ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1280 | } |
1281 | ||
20e07dea RR |
1282 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1283 | { | |
5b763a14 BR |
1284 | return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] + |
1285 | group->faults[task_faults_idx(NUMA_CPU, nid, 1)]; | |
20e07dea RR |
1286 | } |
1287 | ||
b5dd77c8 RR |
1288 | static inline unsigned long group_faults_priv(struct numa_group *ng) |
1289 | { | |
1290 | unsigned long faults = 0; | |
1291 | int node; | |
1292 | ||
1293 | for_each_online_node(node) { | |
1294 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
1295 | } | |
1296 | ||
1297 | return faults; | |
1298 | } | |
1299 | ||
1300 | static inline unsigned long group_faults_shared(struct numa_group *ng) | |
1301 | { | |
1302 | unsigned long faults = 0; | |
1303 | int node; | |
1304 | ||
1305 | for_each_online_node(node) { | |
1306 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
1307 | } | |
1308 | ||
1309 | return faults; | |
1310 | } | |
1311 | ||
4142c3eb RR |
1312 | /* |
1313 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1314 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1315 | * between these nodes are slowed down, to allow things to settle down. | |
1316 | */ | |
1317 | #define ACTIVE_NODE_FRACTION 3 | |
1318 | ||
1319 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1320 | { | |
1321 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1322 | } | |
1323 | ||
6c6b1193 RR |
1324 | /* Handle placement on systems where not all nodes are directly connected. */ |
1325 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
0fb3978b | 1326 | int lim_dist, bool task) |
6c6b1193 RR |
1327 | { |
1328 | unsigned long score = 0; | |
0fb3978b | 1329 | int node, max_dist; |
6c6b1193 RR |
1330 | |
1331 | /* | |
1332 | * All nodes are directly connected, and the same distance | |
1333 | * from each other. No need for fancy placement algorithms. | |
1334 | */ | |
1335 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1336 | return 0; | |
1337 | ||
0fb3978b HY |
1338 | /* sched_max_numa_distance may be changed in parallel. */ |
1339 | max_dist = READ_ONCE(sched_max_numa_distance); | |
6c6b1193 RR |
1340 | /* |
1341 | * This code is called for each node, introducing N^2 complexity, | |
1342 | * which should be ok given the number of nodes rarely exceeds 8. | |
1343 | */ | |
1344 | for_each_online_node(node) { | |
1345 | unsigned long faults; | |
1346 | int dist = node_distance(nid, node); | |
1347 | ||
1348 | /* | |
1349 | * The furthest away nodes in the system are not interesting | |
1350 | * for placement; nid was already counted. | |
1351 | */ | |
0fb3978b | 1352 | if (dist >= max_dist || node == nid) |
6c6b1193 RR |
1353 | continue; |
1354 | ||
1355 | /* | |
1356 | * On systems with a backplane NUMA topology, compare groups | |
1357 | * of nodes, and move tasks towards the group with the most | |
1358 | * memory accesses. When comparing two nodes at distance | |
1359 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1360 | * of each group. Skip other nodes. | |
1361 | */ | |
0fb3978b | 1362 | if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist) |
6c6b1193 RR |
1363 | continue; |
1364 | ||
1365 | /* Add up the faults from nearby nodes. */ | |
1366 | if (task) | |
1367 | faults = task_faults(p, node); | |
1368 | else | |
1369 | faults = group_faults(p, node); | |
1370 | ||
1371 | /* | |
1372 | * On systems with a glueless mesh NUMA topology, there are | |
1373 | * no fixed "groups of nodes". Instead, nodes that are not | |
1374 | * directly connected bounce traffic through intermediate | |
1375 | * nodes; a numa_group can occupy any set of nodes. | |
1376 | * The further away a node is, the less the faults count. | |
1377 | * This seems to result in good task placement. | |
1378 | */ | |
1379 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
0fb3978b HY |
1380 | faults *= (max_dist - dist); |
1381 | faults /= (max_dist - LOCAL_DISTANCE); | |
6c6b1193 RR |
1382 | } |
1383 | ||
1384 | score += faults; | |
1385 | } | |
1386 | ||
1387 | return score; | |
1388 | } | |
1389 | ||
83e1d2cd MG |
1390 | /* |
1391 | * These return the fraction of accesses done by a particular task, or | |
1392 | * task group, on a particular numa node. The group weight is given a | |
1393 | * larger multiplier, in order to group tasks together that are almost | |
1394 | * evenly spread out between numa nodes. | |
1395 | */ | |
7bd95320 RR |
1396 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1397 | int dist) | |
83e1d2cd | 1398 | { |
7bd95320 | 1399 | unsigned long faults, total_faults; |
83e1d2cd | 1400 | |
44dba3d5 | 1401 | if (!p->numa_faults) |
83e1d2cd MG |
1402 | return 0; |
1403 | ||
1404 | total_faults = p->total_numa_faults; | |
1405 | ||
1406 | if (!total_faults) | |
1407 | return 0; | |
1408 | ||
7bd95320 | 1409 | faults = task_faults(p, nid); |
6c6b1193 RR |
1410 | faults += score_nearby_nodes(p, nid, dist, true); |
1411 | ||
7bd95320 | 1412 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1413 | } |
1414 | ||
7bd95320 RR |
1415 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1416 | int dist) | |
83e1d2cd | 1417 | { |
cb361d8c | 1418 | struct numa_group *ng = deref_task_numa_group(p); |
7bd95320 RR |
1419 | unsigned long faults, total_faults; |
1420 | ||
cb361d8c | 1421 | if (!ng) |
7bd95320 RR |
1422 | return 0; |
1423 | ||
cb361d8c | 1424 | total_faults = ng->total_faults; |
7bd95320 RR |
1425 | |
1426 | if (!total_faults) | |
83e1d2cd MG |
1427 | return 0; |
1428 | ||
7bd95320 | 1429 | faults = group_faults(p, nid); |
6c6b1193 RR |
1430 | faults += score_nearby_nodes(p, nid, dist, false); |
1431 | ||
7bd95320 | 1432 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1433 | } |
1434 | ||
10f39042 RR |
1435 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1436 | int src_nid, int dst_cpu) | |
1437 | { | |
cb361d8c | 1438 | struct numa_group *ng = deref_curr_numa_group(p); |
10f39042 RR |
1439 | int dst_nid = cpu_to_node(dst_cpu); |
1440 | int last_cpupid, this_cpupid; | |
1441 | ||
1442 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
37355bdc MG |
1443 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); |
1444 | ||
1445 | /* | |
1446 | * Allow first faults or private faults to migrate immediately early in | |
1447 | * the lifetime of a task. The magic number 4 is based on waiting for | |
1448 | * two full passes of the "multi-stage node selection" test that is | |
1449 | * executed below. | |
1450 | */ | |
98fa15f3 | 1451 | if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && |
37355bdc MG |
1452 | (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid))) |
1453 | return true; | |
10f39042 RR |
1454 | |
1455 | /* | |
1456 | * Multi-stage node selection is used in conjunction with a periodic | |
1457 | * migration fault to build a temporal task<->page relation. By using | |
1458 | * a two-stage filter we remove short/unlikely relations. | |
1459 | * | |
1460 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1461 | * a task's usage of a particular page (n_p) per total usage of this | |
1462 | * page (n_t) (in a given time-span) to a probability. | |
1463 | * | |
1464 | * Our periodic faults will sample this probability and getting the | |
1465 | * same result twice in a row, given these samples are fully | |
1466 | * independent, is then given by P(n)^2, provided our sample period | |
1467 | * is sufficiently short compared to the usage pattern. | |
1468 | * | |
1469 | * This quadric squishes small probabilities, making it less likely we | |
1470 | * act on an unlikely task<->page relation. | |
1471 | */ | |
10f39042 RR |
1472 | if (!cpupid_pid_unset(last_cpupid) && |
1473 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1474 | return false; | |
1475 | ||
1476 | /* Always allow migrate on private faults */ | |
1477 | if (cpupid_match_pid(p, last_cpupid)) | |
1478 | return true; | |
1479 | ||
1480 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1481 | if (!ng) | |
1482 | return true; | |
1483 | ||
1484 | /* | |
4142c3eb RR |
1485 | * Destination node is much more heavily used than the source |
1486 | * node? Allow migration. | |
10f39042 | 1487 | */ |
4142c3eb RR |
1488 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1489 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1490 | return true; |
1491 | ||
1492 | /* | |
4142c3eb RR |
1493 | * Distribute memory according to CPU & memory use on each node, |
1494 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1495 | * | |
1496 | * faults_cpu(dst) 3 faults_cpu(src) | |
1497 | * --------------- * - > --------------- | |
1498 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1499 | */ |
4142c3eb RR |
1500 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1501 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1502 | } |
1503 | ||
6499b1b2 VG |
1504 | /* |
1505 | * 'numa_type' describes the node at the moment of load balancing. | |
1506 | */ | |
1507 | enum numa_type { | |
1508 | /* The node has spare capacity that can be used to run more tasks. */ | |
1509 | node_has_spare = 0, | |
1510 | /* | |
1511 | * The node is fully used and the tasks don't compete for more CPU | |
1512 | * cycles. Nevertheless, some tasks might wait before running. | |
1513 | */ | |
1514 | node_fully_busy, | |
1515 | /* | |
1516 | * The node is overloaded and can't provide expected CPU cycles to all | |
1517 | * tasks. | |
1518 | */ | |
1519 | node_overloaded | |
1520 | }; | |
58d081b5 | 1521 | |
fb13c7ee | 1522 | /* Cached statistics for all CPUs within a node */ |
58d081b5 MG |
1523 | struct numa_stats { |
1524 | unsigned long load; | |
8e0e0eda | 1525 | unsigned long runnable; |
6499b1b2 | 1526 | unsigned long util; |
fb13c7ee | 1527 | /* Total compute capacity of CPUs on a node */ |
5ef20ca1 | 1528 | unsigned long compute_capacity; |
6499b1b2 VG |
1529 | unsigned int nr_running; |
1530 | unsigned int weight; | |
1531 | enum numa_type node_type; | |
ff7db0bf | 1532 | int idle_cpu; |
58d081b5 | 1533 | }; |
e6628d5b | 1534 | |
ff7db0bf MG |
1535 | static inline bool is_core_idle(int cpu) |
1536 | { | |
1537 | #ifdef CONFIG_SCHED_SMT | |
1538 | int sibling; | |
1539 | ||
1540 | for_each_cpu(sibling, cpu_smt_mask(cpu)) { | |
1541 | if (cpu == sibling) | |
1542 | continue; | |
1543 | ||
1c6829cf | 1544 | if (!idle_cpu(sibling)) |
ff7db0bf MG |
1545 | return false; |
1546 | } | |
1547 | #endif | |
1548 | ||
1549 | return true; | |
1550 | } | |
1551 | ||
58d081b5 MG |
1552 | struct task_numa_env { |
1553 | struct task_struct *p; | |
e6628d5b | 1554 | |
58d081b5 MG |
1555 | int src_cpu, src_nid; |
1556 | int dst_cpu, dst_nid; | |
e496132e | 1557 | int imb_numa_nr; |
e6628d5b | 1558 | |
58d081b5 | 1559 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1560 | |
40ea2b42 | 1561 | int imbalance_pct; |
7bd95320 | 1562 | int dist; |
fb13c7ee MG |
1563 | |
1564 | struct task_struct *best_task; | |
1565 | long best_imp; | |
58d081b5 MG |
1566 | int best_cpu; |
1567 | }; | |
1568 | ||
6499b1b2 | 1569 | static unsigned long cpu_load(struct rq *rq); |
8e0e0eda | 1570 | static unsigned long cpu_runnable(struct rq *rq); |
6499b1b2 VG |
1571 | |
1572 | static inline enum | |
1573 | numa_type numa_classify(unsigned int imbalance_pct, | |
1574 | struct numa_stats *ns) | |
1575 | { | |
1576 | if ((ns->nr_running > ns->weight) && | |
8e0e0eda VG |
1577 | (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) || |
1578 | ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100)))) | |
6499b1b2 VG |
1579 | return node_overloaded; |
1580 | ||
1581 | if ((ns->nr_running < ns->weight) || | |
8e0e0eda VG |
1582 | (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) && |
1583 | ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100)))) | |
6499b1b2 VG |
1584 | return node_has_spare; |
1585 | ||
1586 | return node_fully_busy; | |
1587 | } | |
1588 | ||
76c389ab VS |
1589 | #ifdef CONFIG_SCHED_SMT |
1590 | /* Forward declarations of select_idle_sibling helpers */ | |
1591 | static inline bool test_idle_cores(int cpu, bool def); | |
ff7db0bf MG |
1592 | static inline int numa_idle_core(int idle_core, int cpu) |
1593 | { | |
ff7db0bf MG |
1594 | if (!static_branch_likely(&sched_smt_present) || |
1595 | idle_core >= 0 || !test_idle_cores(cpu, false)) | |
1596 | return idle_core; | |
1597 | ||
1598 | /* | |
1599 | * Prefer cores instead of packing HT siblings | |
1600 | * and triggering future load balancing. | |
1601 | */ | |
1602 | if (is_core_idle(cpu)) | |
1603 | idle_core = cpu; | |
ff7db0bf MG |
1604 | |
1605 | return idle_core; | |
1606 | } | |
76c389ab VS |
1607 | #else |
1608 | static inline int numa_idle_core(int idle_core, int cpu) | |
1609 | { | |
1610 | return idle_core; | |
1611 | } | |
1612 | #endif | |
ff7db0bf | 1613 | |
6499b1b2 | 1614 | /* |
ff7db0bf MG |
1615 | * Gather all necessary information to make NUMA balancing placement |
1616 | * decisions that are compatible with standard load balancer. This | |
1617 | * borrows code and logic from update_sg_lb_stats but sharing a | |
1618 | * common implementation is impractical. | |
6499b1b2 VG |
1619 | */ |
1620 | static void update_numa_stats(struct task_numa_env *env, | |
ff7db0bf MG |
1621 | struct numa_stats *ns, int nid, |
1622 | bool find_idle) | |
6499b1b2 | 1623 | { |
ff7db0bf | 1624 | int cpu, idle_core = -1; |
6499b1b2 VG |
1625 | |
1626 | memset(ns, 0, sizeof(*ns)); | |
ff7db0bf MG |
1627 | ns->idle_cpu = -1; |
1628 | ||
0621df31 | 1629 | rcu_read_lock(); |
6499b1b2 VG |
1630 | for_each_cpu(cpu, cpumask_of_node(nid)) { |
1631 | struct rq *rq = cpu_rq(cpu); | |
1632 | ||
1633 | ns->load += cpu_load(rq); | |
8e0e0eda | 1634 | ns->runnable += cpu_runnable(rq); |
82762d2a | 1635 | ns->util += cpu_util_cfs(cpu); |
6499b1b2 VG |
1636 | ns->nr_running += rq->cfs.h_nr_running; |
1637 | ns->compute_capacity += capacity_of(cpu); | |
ff7db0bf MG |
1638 | |
1639 | if (find_idle && !rq->nr_running && idle_cpu(cpu)) { | |
1640 | if (READ_ONCE(rq->numa_migrate_on) || | |
1641 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) | |
1642 | continue; | |
1643 | ||
1644 | if (ns->idle_cpu == -1) | |
1645 | ns->idle_cpu = cpu; | |
1646 | ||
1647 | idle_core = numa_idle_core(idle_core, cpu); | |
1648 | } | |
6499b1b2 | 1649 | } |
0621df31 | 1650 | rcu_read_unlock(); |
6499b1b2 VG |
1651 | |
1652 | ns->weight = cpumask_weight(cpumask_of_node(nid)); | |
1653 | ||
1654 | ns->node_type = numa_classify(env->imbalance_pct, ns); | |
ff7db0bf MG |
1655 | |
1656 | if (idle_core >= 0) | |
1657 | ns->idle_cpu = idle_core; | |
6499b1b2 VG |
1658 | } |
1659 | ||
fb13c7ee MG |
1660 | static void task_numa_assign(struct task_numa_env *env, |
1661 | struct task_struct *p, long imp) | |
1662 | { | |
a4739eca SD |
1663 | struct rq *rq = cpu_rq(env->dst_cpu); |
1664 | ||
5fb52dd9 MG |
1665 | /* Check if run-queue part of active NUMA balance. */ |
1666 | if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) { | |
1667 | int cpu; | |
1668 | int start = env->dst_cpu; | |
1669 | ||
1670 | /* Find alternative idle CPU. */ | |
1671 | for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) { | |
1672 | if (cpu == env->best_cpu || !idle_cpu(cpu) || | |
1673 | !cpumask_test_cpu(cpu, env->p->cpus_ptr)) { | |
1674 | continue; | |
1675 | } | |
1676 | ||
1677 | env->dst_cpu = cpu; | |
1678 | rq = cpu_rq(env->dst_cpu); | |
1679 | if (!xchg(&rq->numa_migrate_on, 1)) | |
1680 | goto assign; | |
1681 | } | |
1682 | ||
1683 | /* Failed to find an alternative idle CPU */ | |
a4739eca | 1684 | return; |
5fb52dd9 | 1685 | } |
a4739eca | 1686 | |
5fb52dd9 | 1687 | assign: |
a4739eca SD |
1688 | /* |
1689 | * Clear previous best_cpu/rq numa-migrate flag, since task now | |
1690 | * found a better CPU to move/swap. | |
1691 | */ | |
5fb52dd9 | 1692 | if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) { |
a4739eca SD |
1693 | rq = cpu_rq(env->best_cpu); |
1694 | WRITE_ONCE(rq->numa_migrate_on, 0); | |
1695 | } | |
1696 | ||
fb13c7ee MG |
1697 | if (env->best_task) |
1698 | put_task_struct(env->best_task); | |
bac78573 ON |
1699 | if (p) |
1700 | get_task_struct(p); | |
fb13c7ee MG |
1701 | |
1702 | env->best_task = p; | |
1703 | env->best_imp = imp; | |
1704 | env->best_cpu = env->dst_cpu; | |
1705 | } | |
1706 | ||
28a21745 | 1707 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1708 | struct task_numa_env *env) |
1709 | { | |
e4991b24 RR |
1710 | long imb, old_imb; |
1711 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1712 | long src_capacity, dst_capacity; |
1713 | ||
1714 | /* | |
1715 | * The load is corrected for the CPU capacity available on each node. | |
1716 | * | |
1717 | * src_load dst_load | |
1718 | * ------------ vs --------- | |
1719 | * src_capacity dst_capacity | |
1720 | */ | |
1721 | src_capacity = env->src_stats.compute_capacity; | |
1722 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 | 1723 | |
5f95ba7a | 1724 | imb = abs(dst_load * src_capacity - src_load * dst_capacity); |
e63da036 | 1725 | |
28a21745 | 1726 | orig_src_load = env->src_stats.load; |
e4991b24 | 1727 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1728 | |
5f95ba7a | 1729 | old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); |
e4991b24 RR |
1730 | |
1731 | /* Would this change make things worse? */ | |
1732 | return (imb > old_imb); | |
e63da036 RR |
1733 | } |
1734 | ||
6fd98e77 SD |
1735 | /* |
1736 | * Maximum NUMA importance can be 1998 (2*999); | |
1737 | * SMALLIMP @ 30 would be close to 1998/64. | |
1738 | * Used to deter task migration. | |
1739 | */ | |
1740 | #define SMALLIMP 30 | |
1741 | ||
fb13c7ee MG |
1742 | /* |
1743 | * This checks if the overall compute and NUMA accesses of the system would | |
1744 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1745 | * into account that it might be best if task running on the dst_cpu should | |
1746 | * be exchanged with the source task | |
1747 | */ | |
a0f03b61 | 1748 | static bool task_numa_compare(struct task_numa_env *env, |
305c1fac | 1749 | long taskimp, long groupimp, bool maymove) |
fb13c7ee | 1750 | { |
cb361d8c | 1751 | struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); |
fb13c7ee | 1752 | struct rq *dst_rq = cpu_rq(env->dst_cpu); |
cb361d8c | 1753 | long imp = p_ng ? groupimp : taskimp; |
fb13c7ee | 1754 | struct task_struct *cur; |
28a21745 | 1755 | long src_load, dst_load; |
7bd95320 | 1756 | int dist = env->dist; |
cb361d8c JH |
1757 | long moveimp = imp; |
1758 | long load; | |
a0f03b61 | 1759 | bool stopsearch = false; |
fb13c7ee | 1760 | |
a4739eca | 1761 | if (READ_ONCE(dst_rq->numa_migrate_on)) |
a0f03b61 | 1762 | return false; |
a4739eca | 1763 | |
fb13c7ee | 1764 | rcu_read_lock(); |
154abafc | 1765 | cur = rcu_dereference(dst_rq->curr); |
bac78573 | 1766 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) |
fb13c7ee MG |
1767 | cur = NULL; |
1768 | ||
7af68335 PZ |
1769 | /* |
1770 | * Because we have preemption enabled we can get migrated around and | |
1771 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1772 | */ | |
a0f03b61 MG |
1773 | if (cur == env->p) { |
1774 | stopsearch = true; | |
7af68335 | 1775 | goto unlock; |
a0f03b61 | 1776 | } |
7af68335 | 1777 | |
305c1fac | 1778 | if (!cur) { |
6fd98e77 | 1779 | if (maymove && moveimp >= env->best_imp) |
305c1fac SD |
1780 | goto assign; |
1781 | else | |
1782 | goto unlock; | |
1783 | } | |
1784 | ||
88cca72c MG |
1785 | /* Skip this swap candidate if cannot move to the source cpu. */ |
1786 | if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) | |
1787 | goto unlock; | |
1788 | ||
1789 | /* | |
1790 | * Skip this swap candidate if it is not moving to its preferred | |
1791 | * node and the best task is. | |
1792 | */ | |
1793 | if (env->best_task && | |
1794 | env->best_task->numa_preferred_nid == env->src_nid && | |
1795 | cur->numa_preferred_nid != env->src_nid) { | |
1796 | goto unlock; | |
1797 | } | |
1798 | ||
fb13c7ee MG |
1799 | /* |
1800 | * "imp" is the fault differential for the source task between the | |
1801 | * source and destination node. Calculate the total differential for | |
1802 | * the source task and potential destination task. The more negative | |
305c1fac | 1803 | * the value is, the more remote accesses that would be expected to |
fb13c7ee | 1804 | * be incurred if the tasks were swapped. |
88cca72c | 1805 | * |
305c1fac SD |
1806 | * If dst and source tasks are in the same NUMA group, or not |
1807 | * in any group then look only at task weights. | |
1808 | */ | |
cb361d8c JH |
1809 | cur_ng = rcu_dereference(cur->numa_group); |
1810 | if (cur_ng == p_ng) { | |
13ede331 MG |
1811 | /* |
1812 | * Do not swap within a group or between tasks that have | |
1813 | * no group if there is spare capacity. Swapping does | |
1814 | * not address the load imbalance and helps one task at | |
1815 | * the cost of punishing another. | |
1816 | */ | |
1817 | if (env->dst_stats.node_type == node_has_spare) | |
1818 | goto unlock; | |
1819 | ||
305c1fac SD |
1820 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1821 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1822 | /* |
305c1fac SD |
1823 | * Add some hysteresis to prevent swapping the |
1824 | * tasks within a group over tiny differences. | |
887c290e | 1825 | */ |
cb361d8c | 1826 | if (cur_ng) |
305c1fac SD |
1827 | imp -= imp / 16; |
1828 | } else { | |
1829 | /* | |
1830 | * Compare the group weights. If a task is all by itself | |
1831 | * (not part of a group), use the task weight instead. | |
1832 | */ | |
cb361d8c | 1833 | if (cur_ng && p_ng) |
305c1fac SD |
1834 | imp += group_weight(cur, env->src_nid, dist) - |
1835 | group_weight(cur, env->dst_nid, dist); | |
1836 | else | |
1837 | imp += task_weight(cur, env->src_nid, dist) - | |
1838 | task_weight(cur, env->dst_nid, dist); | |
fb13c7ee MG |
1839 | } |
1840 | ||
88cca72c MG |
1841 | /* Discourage picking a task already on its preferred node */ |
1842 | if (cur->numa_preferred_nid == env->dst_nid) | |
1843 | imp -= imp / 16; | |
1844 | ||
1845 | /* | |
1846 | * Encourage picking a task that moves to its preferred node. | |
1847 | * This potentially makes imp larger than it's maximum of | |
1848 | * 1998 (see SMALLIMP and task_weight for why) but in this | |
1849 | * case, it does not matter. | |
1850 | */ | |
1851 | if (cur->numa_preferred_nid == env->src_nid) | |
1852 | imp += imp / 8; | |
1853 | ||
305c1fac | 1854 | if (maymove && moveimp > imp && moveimp > env->best_imp) { |
6fd98e77 | 1855 | imp = moveimp; |
305c1fac | 1856 | cur = NULL; |
fb13c7ee | 1857 | goto assign; |
305c1fac | 1858 | } |
fb13c7ee | 1859 | |
88cca72c MG |
1860 | /* |
1861 | * Prefer swapping with a task moving to its preferred node over a | |
1862 | * task that is not. | |
1863 | */ | |
1864 | if (env->best_task && cur->numa_preferred_nid == env->src_nid && | |
1865 | env->best_task->numa_preferred_nid != env->src_nid) { | |
1866 | goto assign; | |
1867 | } | |
1868 | ||
6fd98e77 SD |
1869 | /* |
1870 | * If the NUMA importance is less than SMALLIMP, | |
1871 | * task migration might only result in ping pong | |
1872 | * of tasks and also hurt performance due to cache | |
1873 | * misses. | |
1874 | */ | |
1875 | if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) | |
1876 | goto unlock; | |
1877 | ||
fb13c7ee MG |
1878 | /* |
1879 | * In the overloaded case, try and keep the load balanced. | |
1880 | */ | |
305c1fac SD |
1881 | load = task_h_load(env->p) - task_h_load(cur); |
1882 | if (!load) | |
1883 | goto assign; | |
1884 | ||
e720fff6 PZ |
1885 | dst_load = env->dst_stats.load + load; |
1886 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1887 | |
28a21745 | 1888 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1889 | goto unlock; |
1890 | ||
305c1fac | 1891 | assign: |
ff7db0bf | 1892 | /* Evaluate an idle CPU for a task numa move. */ |
10e2f1ac | 1893 | if (!cur) { |
ff7db0bf MG |
1894 | int cpu = env->dst_stats.idle_cpu; |
1895 | ||
1896 | /* Nothing cached so current CPU went idle since the search. */ | |
1897 | if (cpu < 0) | |
1898 | cpu = env->dst_cpu; | |
1899 | ||
10e2f1ac | 1900 | /* |
ff7db0bf MG |
1901 | * If the CPU is no longer truly idle and the previous best CPU |
1902 | * is, keep using it. | |
10e2f1ac | 1903 | */ |
ff7db0bf MG |
1904 | if (!idle_cpu(cpu) && env->best_cpu >= 0 && |
1905 | idle_cpu(env->best_cpu)) { | |
1906 | cpu = env->best_cpu; | |
1907 | } | |
1908 | ||
ff7db0bf | 1909 | env->dst_cpu = cpu; |
10e2f1ac | 1910 | } |
ba7e5a27 | 1911 | |
fb13c7ee | 1912 | task_numa_assign(env, cur, imp); |
a0f03b61 MG |
1913 | |
1914 | /* | |
1915 | * If a move to idle is allowed because there is capacity or load | |
1916 | * balance improves then stop the search. While a better swap | |
1917 | * candidate may exist, a search is not free. | |
1918 | */ | |
1919 | if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu)) | |
1920 | stopsearch = true; | |
1921 | ||
1922 | /* | |
1923 | * If a swap candidate must be identified and the current best task | |
1924 | * moves its preferred node then stop the search. | |
1925 | */ | |
1926 | if (!maymove && env->best_task && | |
1927 | env->best_task->numa_preferred_nid == env->src_nid) { | |
1928 | stopsearch = true; | |
1929 | } | |
fb13c7ee MG |
1930 | unlock: |
1931 | rcu_read_unlock(); | |
a0f03b61 MG |
1932 | |
1933 | return stopsearch; | |
fb13c7ee MG |
1934 | } |
1935 | ||
887c290e RR |
1936 | static void task_numa_find_cpu(struct task_numa_env *env, |
1937 | long taskimp, long groupimp) | |
2c8a50aa | 1938 | { |
305c1fac | 1939 | bool maymove = false; |
2c8a50aa MG |
1940 | int cpu; |
1941 | ||
305c1fac | 1942 | /* |
fb86f5b2 MG |
1943 | * If dst node has spare capacity, then check if there is an |
1944 | * imbalance that would be overruled by the load balancer. | |
305c1fac | 1945 | */ |
fb86f5b2 MG |
1946 | if (env->dst_stats.node_type == node_has_spare) { |
1947 | unsigned int imbalance; | |
1948 | int src_running, dst_running; | |
1949 | ||
1950 | /* | |
1951 | * Would movement cause an imbalance? Note that if src has | |
1952 | * more running tasks that the imbalance is ignored as the | |
1953 | * move improves the imbalance from the perspective of the | |
1954 | * CPU load balancer. | |
1955 | * */ | |
1956 | src_running = env->src_stats.nr_running - 1; | |
1957 | dst_running = env->dst_stats.nr_running + 1; | |
1958 | imbalance = max(0, dst_running - src_running); | |
7d2b5dd0 | 1959 | imbalance = adjust_numa_imbalance(imbalance, dst_running, |
e496132e | 1960 | env->imb_numa_nr); |
fb86f5b2 MG |
1961 | |
1962 | /* Use idle CPU if there is no imbalance */ | |
ff7db0bf | 1963 | if (!imbalance) { |
fb86f5b2 | 1964 | maymove = true; |
ff7db0bf MG |
1965 | if (env->dst_stats.idle_cpu >= 0) { |
1966 | env->dst_cpu = env->dst_stats.idle_cpu; | |
1967 | task_numa_assign(env, NULL, 0); | |
1968 | return; | |
1969 | } | |
1970 | } | |
fb86f5b2 MG |
1971 | } else { |
1972 | long src_load, dst_load, load; | |
1973 | /* | |
1974 | * If the improvement from just moving env->p direction is better | |
1975 | * than swapping tasks around, check if a move is possible. | |
1976 | */ | |
1977 | load = task_h_load(env->p); | |
1978 | dst_load = env->dst_stats.load + load; | |
1979 | src_load = env->src_stats.load - load; | |
1980 | maymove = !load_too_imbalanced(src_load, dst_load, env); | |
1981 | } | |
305c1fac | 1982 | |
2c8a50aa MG |
1983 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { |
1984 | /* Skip this CPU if the source task cannot migrate */ | |
3bd37062 | 1985 | if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) |
2c8a50aa MG |
1986 | continue; |
1987 | ||
1988 | env->dst_cpu = cpu; | |
a0f03b61 MG |
1989 | if (task_numa_compare(env, taskimp, groupimp, maymove)) |
1990 | break; | |
2c8a50aa MG |
1991 | } |
1992 | } | |
1993 | ||
58d081b5 MG |
1994 | static int task_numa_migrate(struct task_struct *p) |
1995 | { | |
58d081b5 MG |
1996 | struct task_numa_env env = { |
1997 | .p = p, | |
fb13c7ee | 1998 | |
58d081b5 | 1999 | .src_cpu = task_cpu(p), |
b32e86b4 | 2000 | .src_nid = task_node(p), |
fb13c7ee MG |
2001 | |
2002 | .imbalance_pct = 112, | |
2003 | ||
2004 | .best_task = NULL, | |
2005 | .best_imp = 0, | |
4142c3eb | 2006 | .best_cpu = -1, |
58d081b5 | 2007 | }; |
cb361d8c | 2008 | unsigned long taskweight, groupweight; |
58d081b5 | 2009 | struct sched_domain *sd; |
cb361d8c JH |
2010 | long taskimp, groupimp; |
2011 | struct numa_group *ng; | |
a4739eca | 2012 | struct rq *best_rq; |
7bd95320 | 2013 | int nid, ret, dist; |
e6628d5b | 2014 | |
58d081b5 | 2015 | /* |
fb13c7ee MG |
2016 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
2017 | * imbalance and would be the first to start moving tasks about. | |
2018 | * | |
2019 | * And we want to avoid any moving of tasks about, as that would create | |
2020 | * random movement of tasks -- counter the numa conditions we're trying | |
2021 | * to satisfy here. | |
58d081b5 MG |
2022 | */ |
2023 | rcu_read_lock(); | |
fb13c7ee | 2024 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
e496132e | 2025 | if (sd) { |
46a73e8a | 2026 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; |
e496132e MG |
2027 | env.imb_numa_nr = sd->imb_numa_nr; |
2028 | } | |
e6628d5b MG |
2029 | rcu_read_unlock(); |
2030 | ||
46a73e8a RR |
2031 | /* |
2032 | * Cpusets can break the scheduler domain tree into smaller | |
2033 | * balance domains, some of which do not cross NUMA boundaries. | |
2034 | * Tasks that are "trapped" in such domains cannot be migrated | |
2035 | * elsewhere, so there is no point in (re)trying. | |
2036 | */ | |
2037 | if (unlikely(!sd)) { | |
8cd45eee | 2038 | sched_setnuma(p, task_node(p)); |
46a73e8a RR |
2039 | return -EINVAL; |
2040 | } | |
2041 | ||
2c8a50aa | 2042 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
2043 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
2044 | taskweight = task_weight(p, env.src_nid, dist); | |
2045 | groupweight = group_weight(p, env.src_nid, dist); | |
ff7db0bf | 2046 | update_numa_stats(&env, &env.src_stats, env.src_nid, false); |
7bd95320 RR |
2047 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; |
2048 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
ff7db0bf | 2049 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
58d081b5 | 2050 | |
a43455a1 | 2051 | /* Try to find a spot on the preferred nid. */ |
2d4056fa | 2052 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 | 2053 | |
9de05d48 RR |
2054 | /* |
2055 | * Look at other nodes in these cases: | |
2056 | * - there is no space available on the preferred_nid | |
2057 | * - the task is part of a numa_group that is interleaved across | |
2058 | * multiple NUMA nodes; in order to better consolidate the group, | |
2059 | * we need to check other locations. | |
2060 | */ | |
cb361d8c JH |
2061 | ng = deref_curr_numa_group(p); |
2062 | if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { | |
5c7b1aaf | 2063 | for_each_node_state(nid, N_CPU) { |
2c8a50aa MG |
2064 | if (nid == env.src_nid || nid == p->numa_preferred_nid) |
2065 | continue; | |
58d081b5 | 2066 | |
7bd95320 | 2067 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
2068 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
2069 | dist != env.dist) { | |
2070 | taskweight = task_weight(p, env.src_nid, dist); | |
2071 | groupweight = group_weight(p, env.src_nid, dist); | |
2072 | } | |
7bd95320 | 2073 | |
83e1d2cd | 2074 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
2075 | taskimp = task_weight(p, nid, dist) - taskweight; |
2076 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 2077 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
2078 | continue; |
2079 | ||
7bd95320 | 2080 | env.dist = dist; |
2c8a50aa | 2081 | env.dst_nid = nid; |
ff7db0bf | 2082 | update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); |
2d4056fa | 2083 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
2084 | } |
2085 | } | |
2086 | ||
68d1b02a RR |
2087 | /* |
2088 | * If the task is part of a workload that spans multiple NUMA nodes, | |
2089 | * and is migrating into one of the workload's active nodes, remember | |
2090 | * this node as the task's preferred numa node, so the workload can | |
2091 | * settle down. | |
2092 | * A task that migrated to a second choice node will be better off | |
2093 | * trying for a better one later. Do not set the preferred node here. | |
2094 | */ | |
cb361d8c | 2095 | if (ng) { |
db015dae RR |
2096 | if (env.best_cpu == -1) |
2097 | nid = env.src_nid; | |
2098 | else | |
8cd45eee | 2099 | nid = cpu_to_node(env.best_cpu); |
db015dae | 2100 | |
8cd45eee SD |
2101 | if (nid != p->numa_preferred_nid) |
2102 | sched_setnuma(p, nid); | |
db015dae RR |
2103 | } |
2104 | ||
2105 | /* No better CPU than the current one was found. */ | |
f22aef4a | 2106 | if (env.best_cpu == -1) { |
b2b2042b | 2107 | trace_sched_stick_numa(p, env.src_cpu, NULL, -1); |
db015dae | 2108 | return -EAGAIN; |
f22aef4a | 2109 | } |
0ec8aa00 | 2110 | |
a4739eca | 2111 | best_rq = cpu_rq(env.best_cpu); |
fb13c7ee | 2112 | if (env.best_task == NULL) { |
286549dc | 2113 | ret = migrate_task_to(p, env.best_cpu); |
a4739eca | 2114 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
286549dc | 2115 | if (ret != 0) |
b2b2042b | 2116 | trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu); |
fb13c7ee MG |
2117 | return ret; |
2118 | } | |
2119 | ||
0ad4e3df | 2120 | ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu); |
a4739eca | 2121 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
0ad4e3df | 2122 | |
286549dc | 2123 | if (ret != 0) |
b2b2042b | 2124 | trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu); |
fb13c7ee MG |
2125 | put_task_struct(env.best_task); |
2126 | return ret; | |
e6628d5b MG |
2127 | } |
2128 | ||
6b9a7460 MG |
2129 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
2130 | static void numa_migrate_preferred(struct task_struct *p) | |
2131 | { | |
5085e2a3 RR |
2132 | unsigned long interval = HZ; |
2133 | ||
2739d3ee | 2134 | /* This task has no NUMA fault statistics yet */ |
98fa15f3 | 2135 | if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) |
6b9a7460 MG |
2136 | return; |
2137 | ||
2739d3ee | 2138 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 | 2139 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
789ba280 | 2140 | p->numa_migrate_retry = jiffies + interval; |
2739d3ee RR |
2141 | |
2142 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 2143 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
2144 | return; |
2145 | ||
2146 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 2147 | task_numa_migrate(p); |
6b9a7460 MG |
2148 | } |
2149 | ||
20e07dea | 2150 | /* |
7d380f24 | 2151 | * Find out how many nodes the workload is actively running on. Do this by |
20e07dea RR |
2152 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
2153 | * be different from the set of nodes where the workload's memory is currently | |
2154 | * located. | |
20e07dea | 2155 | */ |
4142c3eb | 2156 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
2157 | { |
2158 | unsigned long faults, max_faults = 0; | |
4142c3eb | 2159 | int nid, active_nodes = 0; |
20e07dea | 2160 | |
5c7b1aaf | 2161 | for_each_node_state(nid, N_CPU) { |
20e07dea RR |
2162 | faults = group_faults_cpu(numa_group, nid); |
2163 | if (faults > max_faults) | |
2164 | max_faults = faults; | |
2165 | } | |
2166 | ||
5c7b1aaf | 2167 | for_each_node_state(nid, N_CPU) { |
20e07dea | 2168 | faults = group_faults_cpu(numa_group, nid); |
4142c3eb RR |
2169 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
2170 | active_nodes++; | |
20e07dea | 2171 | } |
4142c3eb RR |
2172 | |
2173 | numa_group->max_faults_cpu = max_faults; | |
2174 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
2175 | } |
2176 | ||
04bb2f94 RR |
2177 | /* |
2178 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
2179 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
2180 | * period will be for the next scan window. If local/(local+remote) ratio is |
2181 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
2182 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
2183 | */ |
2184 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 2185 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
2186 | |
2187 | /* | |
2188 | * Increase the scan period (slow down scanning) if the majority of | |
2189 | * our memory is already on our local node, or if the majority of | |
2190 | * the page accesses are shared with other processes. | |
2191 | * Otherwise, decrease the scan period. | |
2192 | */ | |
2193 | static void update_task_scan_period(struct task_struct *p, | |
2194 | unsigned long shared, unsigned long private) | |
2195 | { | |
2196 | unsigned int period_slot; | |
37ec97de | 2197 | int lr_ratio, ps_ratio; |
04bb2f94 RR |
2198 | int diff; |
2199 | ||
2200 | unsigned long remote = p->numa_faults_locality[0]; | |
2201 | unsigned long local = p->numa_faults_locality[1]; | |
2202 | ||
2203 | /* | |
2204 | * If there were no record hinting faults then either the task is | |
7d380f24 | 2205 | * completely idle or all activity is in areas that are not of interest |
074c2381 MG |
2206 | * to automatic numa balancing. Related to that, if there were failed |
2207 | * migration then it implies we are migrating too quickly or the local | |
2208 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 2209 | */ |
074c2381 | 2210 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
2211 | p->numa_scan_period = min(p->numa_scan_period_max, |
2212 | p->numa_scan_period << 1); | |
2213 | ||
2214 | p->mm->numa_next_scan = jiffies + | |
2215 | msecs_to_jiffies(p->numa_scan_period); | |
2216 | ||
2217 | return; | |
2218 | } | |
2219 | ||
2220 | /* | |
2221 | * Prepare to scale scan period relative to the current period. | |
2222 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
2223 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
2224 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
2225 | */ | |
2226 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
37ec97de RR |
2227 | lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); |
2228 | ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); | |
2229 | ||
2230 | if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2231 | /* | |
2232 | * Most memory accesses are local. There is no need to | |
2233 | * do fast NUMA scanning, since memory is already local. | |
2234 | */ | |
2235 | int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; | |
2236 | if (!slot) | |
2237 | slot = 1; | |
2238 | diff = slot * period_slot; | |
2239 | } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { | |
2240 | /* | |
2241 | * Most memory accesses are shared with other tasks. | |
2242 | * There is no point in continuing fast NUMA scanning, | |
2243 | * since other tasks may just move the memory elsewhere. | |
2244 | */ | |
2245 | int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; | |
04bb2f94 RR |
2246 | if (!slot) |
2247 | slot = 1; | |
2248 | diff = slot * period_slot; | |
2249 | } else { | |
04bb2f94 | 2250 | /* |
37ec97de RR |
2251 | * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, |
2252 | * yet they are not on the local NUMA node. Speed up | |
2253 | * NUMA scanning to get the memory moved over. | |
04bb2f94 | 2254 | */ |
37ec97de RR |
2255 | int ratio = max(lr_ratio, ps_ratio); |
2256 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
04bb2f94 RR |
2257 | } |
2258 | ||
2259 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
2260 | task_scan_min(p), task_scan_max(p)); | |
2261 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
2262 | } | |
2263 | ||
7e2703e6 RR |
2264 | /* |
2265 | * Get the fraction of time the task has been running since the last | |
2266 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
2267 | * decays those on a 32ms period, which is orders of magnitude off | |
2268 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
2269 | * stats only if the task is so new there are no NUMA statistics yet. | |
2270 | */ | |
2271 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
2272 | { | |
2273 | u64 runtime, delta, now; | |
2274 | /* Use the start of this time slice to avoid calculations. */ | |
2275 | now = p->se.exec_start; | |
2276 | runtime = p->se.sum_exec_runtime; | |
2277 | ||
2278 | if (p->last_task_numa_placement) { | |
2279 | delta = runtime - p->last_sum_exec_runtime; | |
2280 | *period = now - p->last_task_numa_placement; | |
a860fa7b XX |
2281 | |
2282 | /* Avoid time going backwards, prevent potential divide error: */ | |
2283 | if (unlikely((s64)*period < 0)) | |
2284 | *period = 0; | |
7e2703e6 | 2285 | } else { |
c7b50216 | 2286 | delta = p->se.avg.load_sum; |
9d89c257 | 2287 | *period = LOAD_AVG_MAX; |
7e2703e6 RR |
2288 | } |
2289 | ||
2290 | p->last_sum_exec_runtime = runtime; | |
2291 | p->last_task_numa_placement = now; | |
2292 | ||
2293 | return delta; | |
2294 | } | |
2295 | ||
54009416 RR |
2296 | /* |
2297 | * Determine the preferred nid for a task in a numa_group. This needs to | |
2298 | * be done in a way that produces consistent results with group_weight, | |
2299 | * otherwise workloads might not converge. | |
2300 | */ | |
2301 | static int preferred_group_nid(struct task_struct *p, int nid) | |
2302 | { | |
2303 | nodemask_t nodes; | |
2304 | int dist; | |
2305 | ||
2306 | /* Direct connections between all NUMA nodes. */ | |
2307 | if (sched_numa_topology_type == NUMA_DIRECT) | |
2308 | return nid; | |
2309 | ||
2310 | /* | |
2311 | * On a system with glueless mesh NUMA topology, group_weight | |
2312 | * scores nodes according to the number of NUMA hinting faults on | |
2313 | * both the node itself, and on nearby nodes. | |
2314 | */ | |
2315 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
2316 | unsigned long score, max_score = 0; | |
2317 | int node, max_node = nid; | |
2318 | ||
2319 | dist = sched_max_numa_distance; | |
2320 | ||
5c7b1aaf | 2321 | for_each_node_state(node, N_CPU) { |
54009416 RR |
2322 | score = group_weight(p, node, dist); |
2323 | if (score > max_score) { | |
2324 | max_score = score; | |
2325 | max_node = node; | |
2326 | } | |
2327 | } | |
2328 | return max_node; | |
2329 | } | |
2330 | ||
2331 | /* | |
2332 | * Finding the preferred nid in a system with NUMA backplane | |
2333 | * interconnect topology is more involved. The goal is to locate | |
2334 | * tasks from numa_groups near each other in the system, and | |
2335 | * untangle workloads from different sides of the system. This requires | |
2336 | * searching down the hierarchy of node groups, recursively searching | |
2337 | * inside the highest scoring group of nodes. The nodemask tricks | |
2338 | * keep the complexity of the search down. | |
2339 | */ | |
5c7b1aaf | 2340 | nodes = node_states[N_CPU]; |
54009416 RR |
2341 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { |
2342 | unsigned long max_faults = 0; | |
81907478 | 2343 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
2344 | int a, b; |
2345 | ||
2346 | /* Are there nodes at this distance from each other? */ | |
2347 | if (!find_numa_distance(dist)) | |
2348 | continue; | |
2349 | ||
2350 | for_each_node_mask(a, nodes) { | |
2351 | unsigned long faults = 0; | |
2352 | nodemask_t this_group; | |
2353 | nodes_clear(this_group); | |
2354 | ||
2355 | /* Sum group's NUMA faults; includes a==b case. */ | |
2356 | for_each_node_mask(b, nodes) { | |
2357 | if (node_distance(a, b) < dist) { | |
2358 | faults += group_faults(p, b); | |
2359 | node_set(b, this_group); | |
2360 | node_clear(b, nodes); | |
2361 | } | |
2362 | } | |
2363 | ||
2364 | /* Remember the top group. */ | |
2365 | if (faults > max_faults) { | |
2366 | max_faults = faults; | |
2367 | max_group = this_group; | |
2368 | /* | |
2369 | * subtle: at the smallest distance there is | |
2370 | * just one node left in each "group", the | |
2371 | * winner is the preferred nid. | |
2372 | */ | |
2373 | nid = a; | |
2374 | } | |
2375 | } | |
2376 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2377 | if (!max_faults) |
2378 | break; | |
54009416 RR |
2379 | nodes = max_group; |
2380 | } | |
2381 | return nid; | |
2382 | } | |
2383 | ||
cbee9f88 PZ |
2384 | static void task_numa_placement(struct task_struct *p) |
2385 | { | |
98fa15f3 | 2386 | int seq, nid, max_nid = NUMA_NO_NODE; |
f03bb676 | 2387 | unsigned long max_faults = 0; |
04bb2f94 | 2388 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2389 | unsigned long total_faults; |
2390 | u64 runtime, period; | |
7dbd13ed | 2391 | spinlock_t *group_lock = NULL; |
cb361d8c | 2392 | struct numa_group *ng; |
cbee9f88 | 2393 | |
7e5a2c17 JL |
2394 | /* |
2395 | * The p->mm->numa_scan_seq field gets updated without | |
2396 | * exclusive access. Use READ_ONCE() here to ensure | |
2397 | * that the field is read in a single access: | |
2398 | */ | |
316c1608 | 2399 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2400 | if (p->numa_scan_seq == seq) |
2401 | return; | |
2402 | p->numa_scan_seq = seq; | |
598f0ec0 | 2403 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2404 | |
7e2703e6 RR |
2405 | total_faults = p->numa_faults_locality[0] + |
2406 | p->numa_faults_locality[1]; | |
2407 | runtime = numa_get_avg_runtime(p, &period); | |
2408 | ||
7dbd13ed | 2409 | /* If the task is part of a group prevent parallel updates to group stats */ |
cb361d8c JH |
2410 | ng = deref_curr_numa_group(p); |
2411 | if (ng) { | |
2412 | group_lock = &ng->lock; | |
60e69eed | 2413 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2414 | } |
2415 | ||
688b7585 MG |
2416 | /* Find the node with the highest number of faults */ |
2417 | for_each_online_node(nid) { | |
44dba3d5 IM |
2418 | /* Keep track of the offsets in numa_faults array */ |
2419 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2420 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2421 | int priv; |
745d6147 | 2422 | |
be1e4e76 | 2423 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2424 | long diff, f_diff, f_weight; |
8c8a743c | 2425 | |
44dba3d5 IM |
2426 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2427 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2428 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2429 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2430 | |
ac8e895b | 2431 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2432 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2433 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2434 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2435 | |
7e2703e6 RR |
2436 | /* |
2437 | * Normalize the faults_from, so all tasks in a group | |
2438 | * count according to CPU use, instead of by the raw | |
2439 | * number of faults. Tasks with little runtime have | |
2440 | * little over-all impact on throughput, and thus their | |
2441 | * faults are less important. | |
2442 | */ | |
2443 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2444 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2445 | (total_faults + 1); |
44dba3d5 IM |
2446 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2447 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2448 | |
44dba3d5 IM |
2449 | p->numa_faults[mem_idx] += diff; |
2450 | p->numa_faults[cpu_idx] += f_diff; | |
2451 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2452 | p->total_numa_faults += diff; |
cb361d8c | 2453 | if (ng) { |
44dba3d5 IM |
2454 | /* |
2455 | * safe because we can only change our own group | |
2456 | * | |
2457 | * mem_idx represents the offset for a given | |
2458 | * nid and priv in a specific region because it | |
2459 | * is at the beginning of the numa_faults array. | |
2460 | */ | |
cb361d8c | 2461 | ng->faults[mem_idx] += diff; |
5b763a14 | 2462 | ng->faults[cpu_idx] += f_diff; |
cb361d8c JH |
2463 | ng->total_faults += diff; |
2464 | group_faults += ng->faults[mem_idx]; | |
8c8a743c | 2465 | } |
ac8e895b MG |
2466 | } |
2467 | ||
cb361d8c | 2468 | if (!ng) { |
f03bb676 SD |
2469 | if (faults > max_faults) { |
2470 | max_faults = faults; | |
2471 | max_nid = nid; | |
2472 | } | |
2473 | } else if (group_faults > max_faults) { | |
2474 | max_faults = group_faults; | |
688b7585 MG |
2475 | max_nid = nid; |
2476 | } | |
83e1d2cd MG |
2477 | } |
2478 | ||
5c7b1aaf | 2479 | /* Cannot migrate task to CPU-less node */ |
ab31c7fd | 2480 | if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) { |
5c7b1aaf HY |
2481 | int near_nid = max_nid; |
2482 | int distance, near_distance = INT_MAX; | |
2483 | ||
2484 | for_each_node_state(nid, N_CPU) { | |
2485 | distance = node_distance(max_nid, nid); | |
2486 | if (distance < near_distance) { | |
2487 | near_nid = nid; | |
2488 | near_distance = distance; | |
2489 | } | |
2490 | } | |
2491 | max_nid = near_nid; | |
2492 | } | |
2493 | ||
cb361d8c JH |
2494 | if (ng) { |
2495 | numa_group_count_active_nodes(ng); | |
60e69eed | 2496 | spin_unlock_irq(group_lock); |
f03bb676 | 2497 | max_nid = preferred_group_nid(p, max_nid); |
688b7585 MG |
2498 | } |
2499 | ||
bb97fc31 RR |
2500 | if (max_faults) { |
2501 | /* Set the new preferred node */ | |
2502 | if (max_nid != p->numa_preferred_nid) | |
2503 | sched_setnuma(p, max_nid); | |
3a7053b3 | 2504 | } |
30619c89 SD |
2505 | |
2506 | update_task_scan_period(p, fault_types[0], fault_types[1]); | |
cbee9f88 PZ |
2507 | } |
2508 | ||
8c8a743c PZ |
2509 | static inline int get_numa_group(struct numa_group *grp) |
2510 | { | |
c45a7795 | 2511 | return refcount_inc_not_zero(&grp->refcount); |
8c8a743c PZ |
2512 | } |
2513 | ||
2514 | static inline void put_numa_group(struct numa_group *grp) | |
2515 | { | |
c45a7795 | 2516 | if (refcount_dec_and_test(&grp->refcount)) |
8c8a743c PZ |
2517 | kfree_rcu(grp, rcu); |
2518 | } | |
2519 | ||
3e6a9418 MG |
2520 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2521 | int *priv) | |
8c8a743c PZ |
2522 | { |
2523 | struct numa_group *grp, *my_grp; | |
2524 | struct task_struct *tsk; | |
2525 | bool join = false; | |
2526 | int cpu = cpupid_to_cpu(cpupid); | |
2527 | int i; | |
2528 | ||
cb361d8c | 2529 | if (unlikely(!deref_curr_numa_group(p))) { |
8c8a743c | 2530 | unsigned int size = sizeof(struct numa_group) + |
7a2341fc BR |
2531 | NR_NUMA_HINT_FAULT_STATS * |
2532 | nr_node_ids * sizeof(unsigned long); | |
8c8a743c PZ |
2533 | |
2534 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2535 | if (!grp) | |
2536 | return; | |
2537 | ||
c45a7795 | 2538 | refcount_set(&grp->refcount, 1); |
4142c3eb RR |
2539 | grp->active_nodes = 1; |
2540 | grp->max_faults_cpu = 0; | |
8c8a743c | 2541 | spin_lock_init(&grp->lock); |
e29cf08b | 2542 | grp->gid = p->pid; |
8c8a743c | 2543 | |
be1e4e76 | 2544 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2545 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2546 | |
989348b5 | 2547 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2548 | |
8c8a743c PZ |
2549 | grp->nr_tasks++; |
2550 | rcu_assign_pointer(p->numa_group, grp); | |
2551 | } | |
2552 | ||
2553 | rcu_read_lock(); | |
316c1608 | 2554 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2555 | |
2556 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2557 | goto no_join; |
8c8a743c PZ |
2558 | |
2559 | grp = rcu_dereference(tsk->numa_group); | |
2560 | if (!grp) | |
3354781a | 2561 | goto no_join; |
8c8a743c | 2562 | |
cb361d8c | 2563 | my_grp = deref_curr_numa_group(p); |
8c8a743c | 2564 | if (grp == my_grp) |
3354781a | 2565 | goto no_join; |
8c8a743c PZ |
2566 | |
2567 | /* | |
2568 | * Only join the other group if its bigger; if we're the bigger group, | |
2569 | * the other task will join us. | |
2570 | */ | |
2571 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2572 | goto no_join; |
8c8a743c PZ |
2573 | |
2574 | /* | |
2575 | * Tie-break on the grp address. | |
2576 | */ | |
2577 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2578 | goto no_join; |
8c8a743c | 2579 | |
dabe1d99 RR |
2580 | /* Always join threads in the same process. */ |
2581 | if (tsk->mm == current->mm) | |
2582 | join = true; | |
2583 | ||
2584 | /* Simple filter to avoid false positives due to PID collisions */ | |
2585 | if (flags & TNF_SHARED) | |
2586 | join = true; | |
8c8a743c | 2587 | |
3e6a9418 MG |
2588 | /* Update priv based on whether false sharing was detected */ |
2589 | *priv = !join; | |
2590 | ||
dabe1d99 | 2591 | if (join && !get_numa_group(grp)) |
3354781a | 2592 | goto no_join; |
8c8a743c | 2593 | |
8c8a743c PZ |
2594 | rcu_read_unlock(); |
2595 | ||
2596 | if (!join) | |
2597 | return; | |
2598 | ||
60e69eed MG |
2599 | BUG_ON(irqs_disabled()); |
2600 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 2601 | |
be1e4e76 | 2602 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2603 | my_grp->faults[i] -= p->numa_faults[i]; |
2604 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2605 | } |
989348b5 MG |
2606 | my_grp->total_faults -= p->total_numa_faults; |
2607 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2608 | |
8c8a743c PZ |
2609 | my_grp->nr_tasks--; |
2610 | grp->nr_tasks++; | |
2611 | ||
2612 | spin_unlock(&my_grp->lock); | |
60e69eed | 2613 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2614 | |
2615 | rcu_assign_pointer(p->numa_group, grp); | |
2616 | ||
2617 | put_numa_group(my_grp); | |
3354781a PZ |
2618 | return; |
2619 | ||
2620 | no_join: | |
2621 | rcu_read_unlock(); | |
2622 | return; | |
8c8a743c PZ |
2623 | } |
2624 | ||
16d51a59 | 2625 | /* |
3b03706f | 2626 | * Get rid of NUMA statistics associated with a task (either current or dead). |
16d51a59 JH |
2627 | * If @final is set, the task is dead and has reached refcount zero, so we can |
2628 | * safely free all relevant data structures. Otherwise, there might be | |
2629 | * concurrent reads from places like load balancing and procfs, and we should | |
2630 | * reset the data back to default state without freeing ->numa_faults. | |
2631 | */ | |
2632 | void task_numa_free(struct task_struct *p, bool final) | |
8c8a743c | 2633 | { |
cb361d8c JH |
2634 | /* safe: p either is current or is being freed by current */ |
2635 | struct numa_group *grp = rcu_dereference_raw(p->numa_group); | |
16d51a59 | 2636 | unsigned long *numa_faults = p->numa_faults; |
e9dd685c SR |
2637 | unsigned long flags; |
2638 | int i; | |
8c8a743c | 2639 | |
16d51a59 JH |
2640 | if (!numa_faults) |
2641 | return; | |
2642 | ||
8c8a743c | 2643 | if (grp) { |
e9dd685c | 2644 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2645 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2646 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2647 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2648 | |
8c8a743c | 2649 | grp->nr_tasks--; |
e9dd685c | 2650 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2651 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2652 | put_numa_group(grp); |
2653 | } | |
2654 | ||
16d51a59 JH |
2655 | if (final) { |
2656 | p->numa_faults = NULL; | |
2657 | kfree(numa_faults); | |
2658 | } else { | |
2659 | p->total_numa_faults = 0; | |
2660 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) | |
2661 | numa_faults[i] = 0; | |
2662 | } | |
8c8a743c PZ |
2663 | } |
2664 | ||
cbee9f88 PZ |
2665 | /* |
2666 | * Got a PROT_NONE fault for a page on @node. | |
2667 | */ | |
58b46da3 | 2668 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2669 | { |
2670 | struct task_struct *p = current; | |
6688cc05 | 2671 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2672 | int cpu_node = task_node(current); |
792568ec | 2673 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 2674 | struct numa_group *ng; |
ac8e895b | 2675 | int priv; |
cbee9f88 | 2676 | |
2a595721 | 2677 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2678 | return; |
2679 | ||
9ff1d9ff MG |
2680 | /* for example, ksmd faulting in a user's mm */ |
2681 | if (!p->mm) | |
2682 | return; | |
2683 | ||
f809ca9a | 2684 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2685 | if (unlikely(!p->numa_faults)) { |
2686 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2687 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2688 | |
44dba3d5 IM |
2689 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2690 | if (!p->numa_faults) | |
f809ca9a | 2691 | return; |
745d6147 | 2692 | |
83e1d2cd | 2693 | p->total_numa_faults = 0; |
04bb2f94 | 2694 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2695 | } |
cbee9f88 | 2696 | |
8c8a743c PZ |
2697 | /* |
2698 | * First accesses are treated as private, otherwise consider accesses | |
2699 | * to be private if the accessing pid has not changed | |
2700 | */ | |
2701 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2702 | priv = 1; | |
2703 | } else { | |
2704 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2705 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2706 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2707 | } |
2708 | ||
792568ec RR |
2709 | /* |
2710 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2711 | * occurs wholly within the set of nodes that the workload is | |
2712 | * actively using should be counted as local. This allows the | |
2713 | * scan rate to slow down when a workload has settled down. | |
2714 | */ | |
cb361d8c | 2715 | ng = deref_curr_numa_group(p); |
4142c3eb RR |
2716 | if (!priv && !local && ng && ng->active_nodes > 1 && |
2717 | numa_is_active_node(cpu_node, ng) && | |
2718 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
2719 | local = 1; |
2720 | ||
2739d3ee | 2721 | /* |
e1ff516a YW |
2722 | * Retry to migrate task to preferred node periodically, in case it |
2723 | * previously failed, or the scheduler moved us. | |
2739d3ee | 2724 | */ |
b6a60cf3 SD |
2725 | if (time_after(jiffies, p->numa_migrate_retry)) { |
2726 | task_numa_placement(p); | |
6b9a7460 | 2727 | numa_migrate_preferred(p); |
b6a60cf3 | 2728 | } |
6b9a7460 | 2729 | |
b32e86b4 IM |
2730 | if (migrated) |
2731 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2732 | if (flags & TNF_MIGRATE_FAIL) |
2733 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2734 | |
44dba3d5 IM |
2735 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2736 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2737 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2738 | } |
2739 | ||
6e5fb223 PZ |
2740 | static void reset_ptenuma_scan(struct task_struct *p) |
2741 | { | |
7e5a2c17 JL |
2742 | /* |
2743 | * We only did a read acquisition of the mmap sem, so | |
2744 | * p->mm->numa_scan_seq is written to without exclusive access | |
2745 | * and the update is not guaranteed to be atomic. That's not | |
2746 | * much of an issue though, since this is just used for | |
2747 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2748 | * expensive, to avoid any form of compiler optimizations: | |
2749 | */ | |
316c1608 | 2750 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2751 | p->mm->numa_scan_offset = 0; |
2752 | } | |
2753 | ||
cbee9f88 PZ |
2754 | /* |
2755 | * The expensive part of numa migration is done from task_work context. | |
2756 | * Triggered from task_tick_numa(). | |
2757 | */ | |
9434f9f5 | 2758 | static void task_numa_work(struct callback_head *work) |
cbee9f88 PZ |
2759 | { |
2760 | unsigned long migrate, next_scan, now = jiffies; | |
2761 | struct task_struct *p = current; | |
2762 | struct mm_struct *mm = p->mm; | |
51170840 | 2763 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 2764 | struct vm_area_struct *vma; |
9f40604c | 2765 | unsigned long start, end; |
598f0ec0 | 2766 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2767 | long pages, virtpages; |
cbee9f88 | 2768 | |
9148a3a1 | 2769 | SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); |
cbee9f88 | 2770 | |
b34920d4 | 2771 | work->next = work; |
cbee9f88 PZ |
2772 | /* |
2773 | * Who cares about NUMA placement when they're dying. | |
2774 | * | |
2775 | * NOTE: make sure not to dereference p->mm before this check, | |
2776 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2777 | * without p->mm even though we still had it when we enqueued this | |
2778 | * work. | |
2779 | */ | |
2780 | if (p->flags & PF_EXITING) | |
2781 | return; | |
2782 | ||
930aa174 | 2783 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2784 | mm->numa_next_scan = now + |
2785 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2786 | } |
2787 | ||
cbee9f88 PZ |
2788 | /* |
2789 | * Enforce maximal scan/migration frequency.. | |
2790 | */ | |
2791 | migrate = mm->numa_next_scan; | |
2792 | if (time_before(now, migrate)) | |
2793 | return; | |
2794 | ||
598f0ec0 MG |
2795 | if (p->numa_scan_period == 0) { |
2796 | p->numa_scan_period_max = task_scan_max(p); | |
b5dd77c8 | 2797 | p->numa_scan_period = task_scan_start(p); |
598f0ec0 | 2798 | } |
cbee9f88 | 2799 | |
fb003b80 | 2800 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2801 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2802 | return; | |
2803 | ||
19a78d11 PZ |
2804 | /* |
2805 | * Delay this task enough that another task of this mm will likely win | |
2806 | * the next time around. | |
2807 | */ | |
2808 | p->node_stamp += 2 * TICK_NSEC; | |
2809 | ||
9f40604c MG |
2810 | start = mm->numa_scan_offset; |
2811 | pages = sysctl_numa_balancing_scan_size; | |
2812 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2813 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2814 | if (!pages) |
2815 | return; | |
cbee9f88 | 2816 | |
4620f8c1 | 2817 | |
d8ed45c5 | 2818 | if (!mmap_read_trylock(mm)) |
8655d549 | 2819 | return; |
9f40604c | 2820 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2821 | if (!vma) { |
2822 | reset_ptenuma_scan(p); | |
9f40604c | 2823 | start = 0; |
6e5fb223 PZ |
2824 | vma = mm->mmap; |
2825 | } | |
9f40604c | 2826 | for (; vma; vma = vma->vm_next) { |
6b79c57b | 2827 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 2828 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 2829 | continue; |
6b79c57b | 2830 | } |
6e5fb223 | 2831 | |
4591ce4f MG |
2832 | /* |
2833 | * Shared library pages mapped by multiple processes are not | |
2834 | * migrated as it is expected they are cache replicated. Avoid | |
2835 | * hinting faults in read-only file-backed mappings or the vdso | |
2836 | * as migrating the pages will be of marginal benefit. | |
2837 | */ | |
2838 | if (!vma->vm_mm || | |
2839 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2840 | continue; | |
2841 | ||
3c67f474 MG |
2842 | /* |
2843 | * Skip inaccessible VMAs to avoid any confusion between | |
2844 | * PROT_NONE and NUMA hinting ptes | |
2845 | */ | |
3122e80e | 2846 | if (!vma_is_accessible(vma)) |
3c67f474 | 2847 | continue; |
4591ce4f | 2848 | |
9f40604c MG |
2849 | do { |
2850 | start = max(start, vma->vm_start); | |
2851 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2852 | end = min(end, vma->vm_end); | |
4620f8c1 | 2853 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
2854 | |
2855 | /* | |
4620f8c1 RR |
2856 | * Try to scan sysctl_numa_balancing_size worth of |
2857 | * hpages that have at least one present PTE that | |
2858 | * is not already pte-numa. If the VMA contains | |
2859 | * areas that are unused or already full of prot_numa | |
2860 | * PTEs, scan up to virtpages, to skip through those | |
2861 | * areas faster. | |
598f0ec0 MG |
2862 | */ |
2863 | if (nr_pte_updates) | |
2864 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 2865 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 2866 | |
9f40604c | 2867 | start = end; |
4620f8c1 | 2868 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 2869 | goto out; |
3cf1962c RR |
2870 | |
2871 | cond_resched(); | |
9f40604c | 2872 | } while (end != vma->vm_end); |
cbee9f88 | 2873 | } |
6e5fb223 | 2874 | |
9f40604c | 2875 | out: |
6e5fb223 | 2876 | /* |
c69307d5 PZ |
2877 | * It is possible to reach the end of the VMA list but the last few |
2878 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2879 | * would find the !migratable VMA on the next scan but not reset the | |
2880 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2881 | */ |
2882 | if (vma) | |
9f40604c | 2883 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2884 | else |
2885 | reset_ptenuma_scan(p); | |
d8ed45c5 | 2886 | mmap_read_unlock(mm); |
51170840 RR |
2887 | |
2888 | /* | |
2889 | * Make sure tasks use at least 32x as much time to run other code | |
2890 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
2891 | * Usually update_task_scan_period slows down scanning enough; on an | |
2892 | * overloaded system we need to limit overhead on a per task basis. | |
2893 | */ | |
2894 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
2895 | u64 diff = p->se.sum_exec_runtime - runtime; | |
2896 | p->node_stamp += 32 * diff; | |
2897 | } | |
cbee9f88 PZ |
2898 | } |
2899 | ||
d35927a1 VS |
2900 | void init_numa_balancing(unsigned long clone_flags, struct task_struct *p) |
2901 | { | |
2902 | int mm_users = 0; | |
2903 | struct mm_struct *mm = p->mm; | |
2904 | ||
2905 | if (mm) { | |
2906 | mm_users = atomic_read(&mm->mm_users); | |
2907 | if (mm_users == 1) { | |
2908 | mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
2909 | mm->numa_scan_seq = 0; | |
2910 | } | |
2911 | } | |
2912 | p->node_stamp = 0; | |
2913 | p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; | |
2914 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; | |
70ce3ea9 | 2915 | p->numa_migrate_retry = 0; |
b34920d4 | 2916 | /* Protect against double add, see task_tick_numa and task_numa_work */ |
d35927a1 VS |
2917 | p->numa_work.next = &p->numa_work; |
2918 | p->numa_faults = NULL; | |
12bf8a7e HW |
2919 | p->numa_pages_migrated = 0; |
2920 | p->total_numa_faults = 0; | |
d35927a1 VS |
2921 | RCU_INIT_POINTER(p->numa_group, NULL); |
2922 | p->last_task_numa_placement = 0; | |
2923 | p->last_sum_exec_runtime = 0; | |
2924 | ||
b34920d4 VS |
2925 | init_task_work(&p->numa_work, task_numa_work); |
2926 | ||
d35927a1 VS |
2927 | /* New address space, reset the preferred nid */ |
2928 | if (!(clone_flags & CLONE_VM)) { | |
2929 | p->numa_preferred_nid = NUMA_NO_NODE; | |
2930 | return; | |
2931 | } | |
2932 | ||
2933 | /* | |
2934 | * New thread, keep existing numa_preferred_nid which should be copied | |
2935 | * already by arch_dup_task_struct but stagger when scans start. | |
2936 | */ | |
2937 | if (mm) { | |
2938 | unsigned int delay; | |
2939 | ||
2940 | delay = min_t(unsigned int, task_scan_max(current), | |
2941 | current->numa_scan_period * mm_users * NSEC_PER_MSEC); | |
2942 | delay += 2 * TICK_NSEC; | |
2943 | p->node_stamp = delay; | |
2944 | } | |
2945 | } | |
2946 | ||
cbee9f88 PZ |
2947 | /* |
2948 | * Drive the periodic memory faults.. | |
2949 | */ | |
b1546edc | 2950 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) |
cbee9f88 PZ |
2951 | { |
2952 | struct callback_head *work = &curr->numa_work; | |
2953 | u64 period, now; | |
2954 | ||
2955 | /* | |
2956 | * We don't care about NUMA placement if we don't have memory. | |
2957 | */ | |
b3f9916d | 2958 | if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work) |
cbee9f88 PZ |
2959 | return; |
2960 | ||
2961 | /* | |
2962 | * Using runtime rather than walltime has the dual advantage that | |
2963 | * we (mostly) drive the selection from busy threads and that the | |
2964 | * task needs to have done some actual work before we bother with | |
2965 | * NUMA placement. | |
2966 | */ | |
2967 | now = curr->se.sum_exec_runtime; | |
2968 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2969 | ||
25b3e5a3 | 2970 | if (now > curr->node_stamp + period) { |
4b96a29b | 2971 | if (!curr->node_stamp) |
b5dd77c8 | 2972 | curr->numa_scan_period = task_scan_start(curr); |
19a78d11 | 2973 | curr->node_stamp += period; |
cbee9f88 | 2974 | |
b34920d4 | 2975 | if (!time_before(jiffies, curr->mm->numa_next_scan)) |
91989c70 | 2976 | task_work_add(curr, work, TWA_RESUME); |
cbee9f88 PZ |
2977 | } |
2978 | } | |
3fed382b | 2979 | |
3f9672ba SD |
2980 | static void update_scan_period(struct task_struct *p, int new_cpu) |
2981 | { | |
2982 | int src_nid = cpu_to_node(task_cpu(p)); | |
2983 | int dst_nid = cpu_to_node(new_cpu); | |
2984 | ||
05cbdf4f MG |
2985 | if (!static_branch_likely(&sched_numa_balancing)) |
2986 | return; | |
2987 | ||
3f9672ba SD |
2988 | if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) |
2989 | return; | |
2990 | ||
05cbdf4f MG |
2991 | if (src_nid == dst_nid) |
2992 | return; | |
2993 | ||
2994 | /* | |
2995 | * Allow resets if faults have been trapped before one scan | |
2996 | * has completed. This is most likely due to a new task that | |
2997 | * is pulled cross-node due to wakeups or load balancing. | |
2998 | */ | |
2999 | if (p->numa_scan_seq) { | |
3000 | /* | |
3001 | * Avoid scan adjustments if moving to the preferred | |
3002 | * node or if the task was not previously running on | |
3003 | * the preferred node. | |
3004 | */ | |
3005 | if (dst_nid == p->numa_preferred_nid || | |
98fa15f3 AK |
3006 | (p->numa_preferred_nid != NUMA_NO_NODE && |
3007 | src_nid != p->numa_preferred_nid)) | |
05cbdf4f MG |
3008 | return; |
3009 | } | |
3010 | ||
3011 | p->numa_scan_period = task_scan_start(p); | |
3f9672ba SD |
3012 | } |
3013 | ||
cbee9f88 PZ |
3014 | #else |
3015 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
3016 | { | |
3017 | } | |
0ec8aa00 PZ |
3018 | |
3019 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
3020 | { | |
3021 | } | |
3022 | ||
3023 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
3024 | { | |
3025 | } | |
3fed382b | 3026 | |
3f9672ba SD |
3027 | static inline void update_scan_period(struct task_struct *p, int new_cpu) |
3028 | { | |
3029 | } | |
3030 | ||
cbee9f88 PZ |
3031 | #endif /* CONFIG_NUMA_BALANCING */ |
3032 | ||
30cfdcfc DA |
3033 | static void |
3034 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3035 | { | |
3036 | update_load_add(&cfs_rq->load, se->load.weight); | |
367456c7 | 3037 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
3038 | if (entity_is_task(se)) { |
3039 | struct rq *rq = rq_of(cfs_rq); | |
3040 | ||
3041 | account_numa_enqueue(rq, task_of(se)); | |
3042 | list_add(&se->group_node, &rq->cfs_tasks); | |
3043 | } | |
367456c7 | 3044 | #endif |
30cfdcfc | 3045 | cfs_rq->nr_running++; |
a480adde JD |
3046 | if (se_is_idle(se)) |
3047 | cfs_rq->idle_nr_running++; | |
30cfdcfc DA |
3048 | } |
3049 | ||
3050 | static void | |
3051 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3052 | { | |
3053 | update_load_sub(&cfs_rq->load, se->load.weight); | |
bfdb198c | 3054 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
3055 | if (entity_is_task(se)) { |
3056 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 3057 | list_del_init(&se->group_node); |
0ec8aa00 | 3058 | } |
bfdb198c | 3059 | #endif |
30cfdcfc | 3060 | cfs_rq->nr_running--; |
a480adde JD |
3061 | if (se_is_idle(se)) |
3062 | cfs_rq->idle_nr_running--; | |
30cfdcfc DA |
3063 | } |
3064 | ||
8d5b9025 PZ |
3065 | /* |
3066 | * Signed add and clamp on underflow. | |
3067 | * | |
3068 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3069 | * memory. This allows lockless observations without ever seeing the negative | |
3070 | * values. | |
3071 | */ | |
3072 | #define add_positive(_ptr, _val) do { \ | |
3073 | typeof(_ptr) ptr = (_ptr); \ | |
3074 | typeof(_val) val = (_val); \ | |
3075 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3076 | \ | |
3077 | res = var + val; \ | |
3078 | \ | |
3079 | if (val < 0 && res > var) \ | |
3080 | res = 0; \ | |
3081 | \ | |
3082 | WRITE_ONCE(*ptr, res); \ | |
3083 | } while (0) | |
3084 | ||
3085 | /* | |
3086 | * Unsigned subtract and clamp on underflow. | |
3087 | * | |
3088 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3089 | * memory. This allows lockless observations without ever seeing the negative | |
3090 | * values. | |
3091 | */ | |
3092 | #define sub_positive(_ptr, _val) do { \ | |
3093 | typeof(_ptr) ptr = (_ptr); \ | |
3094 | typeof(*ptr) val = (_val); \ | |
3095 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3096 | res = var - val; \ | |
3097 | if (res > var) \ | |
3098 | res = 0; \ | |
3099 | WRITE_ONCE(*ptr, res); \ | |
3100 | } while (0) | |
3101 | ||
b5c0ce7b PB |
3102 | /* |
3103 | * Remove and clamp on negative, from a local variable. | |
3104 | * | |
3105 | * A variant of sub_positive(), which does not use explicit load-store | |
3106 | * and is thus optimized for local variable updates. | |
3107 | */ | |
3108 | #define lsub_positive(_ptr, _val) do { \ | |
3109 | typeof(_ptr) ptr = (_ptr); \ | |
3110 | *ptr -= min_t(typeof(*ptr), *ptr, _val); \ | |
3111 | } while (0) | |
3112 | ||
8d5b9025 | 3113 | #ifdef CONFIG_SMP |
8d5b9025 PZ |
3114 | static inline void |
3115 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3116 | { | |
3117 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
3118 | cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; | |
3119 | } | |
3120 | ||
3121 | static inline void | |
3122 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3123 | { | |
3124 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); | |
2d02fa8c VG |
3125 | sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); |
3126 | /* See update_cfs_rq_load_avg() */ | |
3127 | cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum, | |
3128 | cfs_rq->avg.load_avg * PELT_MIN_DIVIDER); | |
8d5b9025 PZ |
3129 | } |
3130 | #else | |
3131 | static inline void | |
8d5b9025 PZ |
3132 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } |
3133 | static inline void | |
3134 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
3135 | #endif | |
3136 | ||
9059393e | 3137 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
0dacee1b | 3138 | unsigned long weight) |
9059393e VG |
3139 | { |
3140 | if (se->on_rq) { | |
3141 | /* commit outstanding execution time */ | |
3142 | if (cfs_rq->curr == se) | |
3143 | update_curr(cfs_rq); | |
1724b95b | 3144 | update_load_sub(&cfs_rq->load, se->load.weight); |
9059393e VG |
3145 | } |
3146 | dequeue_load_avg(cfs_rq, se); | |
3147 | ||
3148 | update_load_set(&se->load, weight); | |
3149 | ||
3150 | #ifdef CONFIG_SMP | |
1ea6c46a | 3151 | do { |
87e867b4 | 3152 | u32 divider = get_pelt_divider(&se->avg); |
1ea6c46a PZ |
3153 | |
3154 | se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); | |
1ea6c46a | 3155 | } while (0); |
9059393e VG |
3156 | #endif |
3157 | ||
3158 | enqueue_load_avg(cfs_rq, se); | |
0dacee1b | 3159 | if (se->on_rq) |
1724b95b | 3160 | update_load_add(&cfs_rq->load, se->load.weight); |
0dacee1b | 3161 | |
9059393e VG |
3162 | } |
3163 | ||
3164 | void reweight_task(struct task_struct *p, int prio) | |
3165 | { | |
3166 | struct sched_entity *se = &p->se; | |
3167 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3168 | struct load_weight *load = &se->load; | |
3169 | unsigned long weight = scale_load(sched_prio_to_weight[prio]); | |
3170 | ||
0dacee1b | 3171 | reweight_entity(cfs_rq, se, weight); |
9059393e VG |
3172 | load->inv_weight = sched_prio_to_wmult[prio]; |
3173 | } | |
3174 | ||
51bf903b CZ |
3175 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
3176 | ||
3ff6dcac | 3177 | #ifdef CONFIG_FAIR_GROUP_SCHED |
387f77cc | 3178 | #ifdef CONFIG_SMP |
cef27403 PZ |
3179 | /* |
3180 | * All this does is approximate the hierarchical proportion which includes that | |
3181 | * global sum we all love to hate. | |
3182 | * | |
3183 | * That is, the weight of a group entity, is the proportional share of the | |
3184 | * group weight based on the group runqueue weights. That is: | |
3185 | * | |
3186 | * tg->weight * grq->load.weight | |
3187 | * ge->load.weight = ----------------------------- (1) | |
08f7c2f4 | 3188 | * \Sum grq->load.weight |
cef27403 PZ |
3189 | * |
3190 | * Now, because computing that sum is prohibitively expensive to compute (been | |
3191 | * there, done that) we approximate it with this average stuff. The average | |
3192 | * moves slower and therefore the approximation is cheaper and more stable. | |
3193 | * | |
3194 | * So instead of the above, we substitute: | |
3195 | * | |
3196 | * grq->load.weight -> grq->avg.load_avg (2) | |
3197 | * | |
3198 | * which yields the following: | |
3199 | * | |
3200 | * tg->weight * grq->avg.load_avg | |
3201 | * ge->load.weight = ------------------------------ (3) | |
08f7c2f4 | 3202 | * tg->load_avg |
cef27403 PZ |
3203 | * |
3204 | * Where: tg->load_avg ~= \Sum grq->avg.load_avg | |
3205 | * | |
3206 | * That is shares_avg, and it is right (given the approximation (2)). | |
3207 | * | |
3208 | * The problem with it is that because the average is slow -- it was designed | |
3209 | * to be exactly that of course -- this leads to transients in boundary | |
3210 | * conditions. In specific, the case where the group was idle and we start the | |
3211 | * one task. It takes time for our CPU's grq->avg.load_avg to build up, | |
3212 | * yielding bad latency etc.. | |
3213 | * | |
3214 | * Now, in that special case (1) reduces to: | |
3215 | * | |
3216 | * tg->weight * grq->load.weight | |
17de4ee0 | 3217 | * ge->load.weight = ----------------------------- = tg->weight (4) |
08f7c2f4 | 3218 | * grp->load.weight |
cef27403 PZ |
3219 | * |
3220 | * That is, the sum collapses because all other CPUs are idle; the UP scenario. | |
3221 | * | |
3222 | * So what we do is modify our approximation (3) to approach (4) in the (near) | |
3223 | * UP case, like: | |
3224 | * | |
3225 | * ge->load.weight = | |
3226 | * | |
3227 | * tg->weight * grq->load.weight | |
3228 | * --------------------------------------------------- (5) | |
3229 | * tg->load_avg - grq->avg.load_avg + grq->load.weight | |
3230 | * | |
17de4ee0 PZ |
3231 | * But because grq->load.weight can drop to 0, resulting in a divide by zero, |
3232 | * we need to use grq->avg.load_avg as its lower bound, which then gives: | |
3233 | * | |
3234 | * | |
3235 | * tg->weight * grq->load.weight | |
3236 | * ge->load.weight = ----------------------------- (6) | |
08f7c2f4 | 3237 | * tg_load_avg' |
17de4ee0 PZ |
3238 | * |
3239 | * Where: | |
3240 | * | |
3241 | * tg_load_avg' = tg->load_avg - grq->avg.load_avg + | |
3242 | * max(grq->load.weight, grq->avg.load_avg) | |
cef27403 PZ |
3243 | * |
3244 | * And that is shares_weight and is icky. In the (near) UP case it approaches | |
3245 | * (4) while in the normal case it approaches (3). It consistently | |
3246 | * overestimates the ge->load.weight and therefore: | |
3247 | * | |
3248 | * \Sum ge->load.weight >= tg->weight | |
3249 | * | |
3250 | * hence icky! | |
3251 | */ | |
2c8e4dce | 3252 | static long calc_group_shares(struct cfs_rq *cfs_rq) |
cf5f0acf | 3253 | { |
7c80cfc9 PZ |
3254 | long tg_weight, tg_shares, load, shares; |
3255 | struct task_group *tg = cfs_rq->tg; | |
3256 | ||
3257 | tg_shares = READ_ONCE(tg->shares); | |
cf5f0acf | 3258 | |
3d4b60d3 | 3259 | load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); |
cf5f0acf | 3260 | |
ea1dc6fc | 3261 | tg_weight = atomic_long_read(&tg->load_avg); |
3ff6dcac | 3262 | |
ea1dc6fc PZ |
3263 | /* Ensure tg_weight >= load */ |
3264 | tg_weight -= cfs_rq->tg_load_avg_contrib; | |
3265 | tg_weight += load; | |
3ff6dcac | 3266 | |
7c80cfc9 | 3267 | shares = (tg_shares * load); |
cf5f0acf PZ |
3268 | if (tg_weight) |
3269 | shares /= tg_weight; | |
3ff6dcac | 3270 | |
b8fd8423 DE |
3271 | /* |
3272 | * MIN_SHARES has to be unscaled here to support per-CPU partitioning | |
3273 | * of a group with small tg->shares value. It is a floor value which is | |
3274 | * assigned as a minimum load.weight to the sched_entity representing | |
3275 | * the group on a CPU. | |
3276 | * | |
3277 | * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 | |
3278 | * on an 8-core system with 8 tasks each runnable on one CPU shares has | |
3279 | * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In | |
3280 | * case no task is runnable on a CPU MIN_SHARES=2 should be returned | |
3281 | * instead of 0. | |
3282 | */ | |
7c80cfc9 | 3283 | return clamp_t(long, shares, MIN_SHARES, tg_shares); |
3ff6dcac | 3284 | } |
387f77cc | 3285 | #endif /* CONFIG_SMP */ |
ea1dc6fc | 3286 | |
1ea6c46a PZ |
3287 | /* |
3288 | * Recomputes the group entity based on the current state of its group | |
3289 | * runqueue. | |
3290 | */ | |
3291 | static void update_cfs_group(struct sched_entity *se) | |
2069dd75 | 3292 | { |
1ea6c46a | 3293 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); |
0dacee1b | 3294 | long shares; |
2069dd75 | 3295 | |
1ea6c46a | 3296 | if (!gcfs_rq) |
89ee048f VG |
3297 | return; |
3298 | ||
1ea6c46a | 3299 | if (throttled_hierarchy(gcfs_rq)) |
2069dd75 | 3300 | return; |
89ee048f | 3301 | |
3ff6dcac | 3302 | #ifndef CONFIG_SMP |
0dacee1b | 3303 | shares = READ_ONCE(gcfs_rq->tg->shares); |
7c80cfc9 PZ |
3304 | |
3305 | if (likely(se->load.weight == shares)) | |
3ff6dcac | 3306 | return; |
7c80cfc9 | 3307 | #else |
2c8e4dce | 3308 | shares = calc_group_shares(gcfs_rq); |
3ff6dcac | 3309 | #endif |
2069dd75 | 3310 | |
0dacee1b | 3311 | reweight_entity(cfs_rq_of(se), se, shares); |
2069dd75 | 3312 | } |
89ee048f | 3313 | |
2069dd75 | 3314 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
1ea6c46a | 3315 | static inline void update_cfs_group(struct sched_entity *se) |
2069dd75 PZ |
3316 | { |
3317 | } | |
3318 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
3319 | ||
ea14b57e | 3320 | static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags) |
a030d738 | 3321 | { |
43964409 LT |
3322 | struct rq *rq = rq_of(cfs_rq); |
3323 | ||
a4f9a0e5 | 3324 | if (&rq->cfs == cfs_rq) { |
a030d738 VK |
3325 | /* |
3326 | * There are a few boundary cases this might miss but it should | |
3327 | * get called often enough that that should (hopefully) not be | |
9783be2c | 3328 | * a real problem. |
a030d738 VK |
3329 | * |
3330 | * It will not get called when we go idle, because the idle | |
3331 | * thread is a different class (!fair), nor will the utilization | |
3332 | * number include things like RT tasks. | |
3333 | * | |
3334 | * As is, the util number is not freq-invariant (we'd have to | |
3335 | * implement arch_scale_freq_capacity() for that). | |
3336 | * | |
82762d2a | 3337 | * See cpu_util_cfs(). |
a030d738 | 3338 | */ |
ea14b57e | 3339 | cpufreq_update_util(rq, flags); |
a030d738 VK |
3340 | } |
3341 | } | |
3342 | ||
141965c7 | 3343 | #ifdef CONFIG_SMP |
e2f3e35f VD |
3344 | static inline bool load_avg_is_decayed(struct sched_avg *sa) |
3345 | { | |
3346 | if (sa->load_sum) | |
3347 | return false; | |
3348 | ||
3349 | if (sa->util_sum) | |
3350 | return false; | |
3351 | ||
3352 | if (sa->runnable_sum) | |
3353 | return false; | |
3354 | ||
3355 | /* | |
3356 | * _avg must be null when _sum are null because _avg = _sum / divider | |
3357 | * Make sure that rounding and/or propagation of PELT values never | |
3358 | * break this. | |
3359 | */ | |
3360 | SCHED_WARN_ON(sa->load_avg || | |
3361 | sa->util_avg || | |
3362 | sa->runnable_avg); | |
3363 | ||
3364 | return true; | |
3365 | } | |
3366 | ||
d05b4305 VD |
3367 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3368 | { | |
3369 | return u64_u32_load_copy(cfs_rq->avg.last_update_time, | |
3370 | cfs_rq->last_update_time_copy); | |
3371 | } | |
c566e8e9 | 3372 | #ifdef CONFIG_FAIR_GROUP_SCHED |
fdaba61e RR |
3373 | /* |
3374 | * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list | |
3375 | * immediately before a parent cfs_rq, and cfs_rqs are removed from the list | |
3376 | * bottom-up, we only have to test whether the cfs_rq before us on the list | |
3377 | * is our child. | |
3378 | * If cfs_rq is not on the list, test whether a child needs its to be added to | |
3379 | * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details). | |
3380 | */ | |
3381 | static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq) | |
3382 | { | |
3383 | struct cfs_rq *prev_cfs_rq; | |
3384 | struct list_head *prev; | |
3385 | ||
3386 | if (cfs_rq->on_list) { | |
3387 | prev = cfs_rq->leaf_cfs_rq_list.prev; | |
3388 | } else { | |
3389 | struct rq *rq = rq_of(cfs_rq); | |
3390 | ||
3391 | prev = rq->tmp_alone_branch; | |
3392 | } | |
3393 | ||
3394 | prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list); | |
3395 | ||
3396 | return (prev_cfs_rq->tg->parent == cfs_rq->tg); | |
3397 | } | |
a7b359fc OU |
3398 | |
3399 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) | |
3400 | { | |
3401 | if (cfs_rq->load.weight) | |
3402 | return false; | |
3403 | ||
e2f3e35f | 3404 | if (!load_avg_is_decayed(&cfs_rq->avg)) |
a7b359fc OU |
3405 | return false; |
3406 | ||
fdaba61e RR |
3407 | if (child_cfs_rq_on_list(cfs_rq)) |
3408 | return false; | |
3409 | ||
a7b359fc OU |
3410 | return true; |
3411 | } | |
3412 | ||
7c3edd2c PZ |
3413 | /** |
3414 | * update_tg_load_avg - update the tg's load avg | |
3415 | * @cfs_rq: the cfs_rq whose avg changed | |
7c3edd2c PZ |
3416 | * |
3417 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
3418 | * However, because tg->load_avg is a global value there are performance | |
3419 | * considerations. | |
3420 | * | |
3421 | * In order to avoid having to look at the other cfs_rq's, we use a | |
3422 | * differential update where we store the last value we propagated. This in | |
3423 | * turn allows skipping updates if the differential is 'small'. | |
3424 | * | |
815abf5a | 3425 | * Updating tg's load_avg is necessary before update_cfs_share(). |
bb17f655 | 3426 | */ |
fe749158 | 3427 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) |
bb17f655 | 3428 | { |
9d89c257 | 3429 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 3430 | |
aa0b7ae0 WL |
3431 | /* |
3432 | * No need to update load_avg for root_task_group as it is not used. | |
3433 | */ | |
3434 | if (cfs_rq->tg == &root_task_group) | |
3435 | return; | |
3436 | ||
fe749158 | 3437 | if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
9d89c257 YD |
3438 | atomic_long_add(delta, &cfs_rq->tg->load_avg); |
3439 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 3440 | } |
8165e145 | 3441 | } |
f5f9739d | 3442 | |
ad936d86 | 3443 | /* |
97fb7a0a | 3444 | * Called within set_task_rq() right before setting a task's CPU. The |
ad936d86 BP |
3445 | * caller only guarantees p->pi_lock is held; no other assumptions, |
3446 | * including the state of rq->lock, should be made. | |
3447 | */ | |
3448 | void set_task_rq_fair(struct sched_entity *se, | |
3449 | struct cfs_rq *prev, struct cfs_rq *next) | |
3450 | { | |
0ccb977f PZ |
3451 | u64 p_last_update_time; |
3452 | u64 n_last_update_time; | |
3453 | ||
ad936d86 BP |
3454 | if (!sched_feat(ATTACH_AGE_LOAD)) |
3455 | return; | |
3456 | ||
3457 | /* | |
3458 | * We are supposed to update the task to "current" time, then its up to | |
3459 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
3460 | * getting what current time is, so simply throw away the out-of-date | |
3461 | * time. This will result in the wakee task is less decayed, but giving | |
3462 | * the wakee more load sounds not bad. | |
3463 | */ | |
0ccb977f PZ |
3464 | if (!(se->avg.last_update_time && prev)) |
3465 | return; | |
ad936d86 | 3466 | |
d05b4305 VD |
3467 | p_last_update_time = cfs_rq_last_update_time(prev); |
3468 | n_last_update_time = cfs_rq_last_update_time(next); | |
ad936d86 | 3469 | |
23127296 | 3470 | __update_load_avg_blocked_se(p_last_update_time, se); |
0ccb977f | 3471 | se->avg.last_update_time = n_last_update_time; |
ad936d86 | 3472 | } |
09a43ace | 3473 | |
0e2d2aaa PZ |
3474 | /* |
3475 | * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to | |
3476 | * propagate its contribution. The key to this propagation is the invariant | |
3477 | * that for each group: | |
3478 | * | |
3479 | * ge->avg == grq->avg (1) | |
3480 | * | |
3481 | * _IFF_ we look at the pure running and runnable sums. Because they | |
3482 | * represent the very same entity, just at different points in the hierarchy. | |
3483 | * | |
9f683953 VG |
3484 | * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial |
3485 | * and simply copies the running/runnable sum over (but still wrong, because | |
3486 | * the group entity and group rq do not have their PELT windows aligned). | |
0e2d2aaa | 3487 | * |
0dacee1b | 3488 | * However, update_tg_cfs_load() is more complex. So we have: |
0e2d2aaa PZ |
3489 | * |
3490 | * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) | |
3491 | * | |
3492 | * And since, like util, the runnable part should be directly transferable, | |
3493 | * the following would _appear_ to be the straight forward approach: | |
3494 | * | |
a4c3c049 | 3495 | * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3) |
0e2d2aaa PZ |
3496 | * |
3497 | * And per (1) we have: | |
3498 | * | |
a4c3c049 | 3499 | * ge->avg.runnable_avg == grq->avg.runnable_avg |
0e2d2aaa PZ |
3500 | * |
3501 | * Which gives: | |
3502 | * | |
3503 | * ge->load.weight * grq->avg.load_avg | |
3504 | * ge->avg.load_avg = ----------------------------------- (4) | |
3505 | * grq->load.weight | |
3506 | * | |
3507 | * Except that is wrong! | |
3508 | * | |
3509 | * Because while for entities historical weight is not important and we | |
3510 | * really only care about our future and therefore can consider a pure | |
3511 | * runnable sum, runqueues can NOT do this. | |
3512 | * | |
3513 | * We specifically want runqueues to have a load_avg that includes | |
3514 | * historical weights. Those represent the blocked load, the load we expect | |
3515 | * to (shortly) return to us. This only works by keeping the weights as | |
3516 | * integral part of the sum. We therefore cannot decompose as per (3). | |
3517 | * | |
a4c3c049 VG |
3518 | * Another reason this doesn't work is that runnable isn't a 0-sum entity. |
3519 | * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the | |
3520 | * rq itself is runnable anywhere between 2/3 and 1 depending on how the | |
3521 | * runnable section of these tasks overlap (or not). If they were to perfectly | |
3522 | * align the rq as a whole would be runnable 2/3 of the time. If however we | |
3523 | * always have at least 1 runnable task, the rq as a whole is always runnable. | |
0e2d2aaa | 3524 | * |
a4c3c049 | 3525 | * So we'll have to approximate.. :/ |
0e2d2aaa | 3526 | * |
a4c3c049 | 3527 | * Given the constraint: |
0e2d2aaa | 3528 | * |
a4c3c049 | 3529 | * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX |
0e2d2aaa | 3530 | * |
a4c3c049 VG |
3531 | * We can construct a rule that adds runnable to a rq by assuming minimal |
3532 | * overlap. | |
0e2d2aaa | 3533 | * |
a4c3c049 | 3534 | * On removal, we'll assume each task is equally runnable; which yields: |
0e2d2aaa | 3535 | * |
a4c3c049 | 3536 | * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight |
0e2d2aaa | 3537 | * |
a4c3c049 | 3538 | * XXX: only do this for the part of runnable > running ? |
0e2d2aaa | 3539 | * |
0e2d2aaa | 3540 | */ |
09a43ace | 3541 | static inline void |
0e2d2aaa | 3542 | update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3543 | { |
7ceb7710 VG |
3544 | long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg; |
3545 | u32 new_sum, divider; | |
09a43ace VG |
3546 | |
3547 | /* Nothing to update */ | |
7ceb7710 | 3548 | if (!delta_avg) |
09a43ace VG |
3549 | return; |
3550 | ||
87e867b4 VG |
3551 | /* |
3552 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3553 | * See ___update_load_avg() for details. | |
3554 | */ | |
3555 | divider = get_pelt_divider(&cfs_rq->avg); | |
3556 | ||
7ceb7710 | 3557 | |
09a43ace VG |
3558 | /* Set new sched_entity's utilization */ |
3559 | se->avg.util_avg = gcfs_rq->avg.util_avg; | |
7ceb7710 VG |
3560 | new_sum = se->avg.util_avg * divider; |
3561 | delta_sum = (long)new_sum - (long)se->avg.util_sum; | |
3562 | se->avg.util_sum = new_sum; | |
09a43ace VG |
3563 | |
3564 | /* Update parent cfs_rq utilization */ | |
7ceb7710 VG |
3565 | add_positive(&cfs_rq->avg.util_avg, delta_avg); |
3566 | add_positive(&cfs_rq->avg.util_sum, delta_sum); | |
3567 | ||
3568 | /* See update_cfs_rq_load_avg() */ | |
3569 | cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum, | |
3570 | cfs_rq->avg.util_avg * PELT_MIN_DIVIDER); | |
09a43ace VG |
3571 | } |
3572 | ||
9f683953 VG |
3573 | static inline void |
3574 | update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) | |
3575 | { | |
95246d1e VG |
3576 | long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg; |
3577 | u32 new_sum, divider; | |
9f683953 VG |
3578 | |
3579 | /* Nothing to update */ | |
95246d1e | 3580 | if (!delta_avg) |
9f683953 VG |
3581 | return; |
3582 | ||
87e867b4 VG |
3583 | /* |
3584 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3585 | * See ___update_load_avg() for details. | |
3586 | */ | |
3587 | divider = get_pelt_divider(&cfs_rq->avg); | |
3588 | ||
9f683953 VG |
3589 | /* Set new sched_entity's runnable */ |
3590 | se->avg.runnable_avg = gcfs_rq->avg.runnable_avg; | |
95246d1e VG |
3591 | new_sum = se->avg.runnable_avg * divider; |
3592 | delta_sum = (long)new_sum - (long)se->avg.runnable_sum; | |
3593 | se->avg.runnable_sum = new_sum; | |
9f683953 VG |
3594 | |
3595 | /* Update parent cfs_rq runnable */ | |
95246d1e VG |
3596 | add_positive(&cfs_rq->avg.runnable_avg, delta_avg); |
3597 | add_positive(&cfs_rq->avg.runnable_sum, delta_sum); | |
3598 | /* See update_cfs_rq_load_avg() */ | |
3599 | cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum, | |
3600 | cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER); | |
9f683953 VG |
3601 | } |
3602 | ||
09a43ace | 3603 | static inline void |
0dacee1b | 3604 | update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3605 | { |
2d02fa8c | 3606 | long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; |
0dacee1b VG |
3607 | unsigned long load_avg; |
3608 | u64 load_sum = 0; | |
2d02fa8c | 3609 | s64 delta_sum; |
95d68593 | 3610 | u32 divider; |
09a43ace | 3611 | |
0e2d2aaa PZ |
3612 | if (!runnable_sum) |
3613 | return; | |
09a43ace | 3614 | |
0e2d2aaa | 3615 | gcfs_rq->prop_runnable_sum = 0; |
09a43ace | 3616 | |
95d68593 VG |
3617 | /* |
3618 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3619 | * See ___update_load_avg() for details. | |
3620 | */ | |
87e867b4 | 3621 | divider = get_pelt_divider(&cfs_rq->avg); |
95d68593 | 3622 | |
a4c3c049 VG |
3623 | if (runnable_sum >= 0) { |
3624 | /* | |
3625 | * Add runnable; clip at LOAD_AVG_MAX. Reflects that until | |
3626 | * the CPU is saturated running == runnable. | |
3627 | */ | |
3628 | runnable_sum += se->avg.load_sum; | |
95d68593 | 3629 | runnable_sum = min_t(long, runnable_sum, divider); |
a4c3c049 VG |
3630 | } else { |
3631 | /* | |
3632 | * Estimate the new unweighted runnable_sum of the gcfs_rq by | |
3633 | * assuming all tasks are equally runnable. | |
3634 | */ | |
3635 | if (scale_load_down(gcfs_rq->load.weight)) { | |
2d02fa8c | 3636 | load_sum = div_u64(gcfs_rq->avg.load_sum, |
a4c3c049 VG |
3637 | scale_load_down(gcfs_rq->load.weight)); |
3638 | } | |
3639 | ||
3640 | /* But make sure to not inflate se's runnable */ | |
3641 | runnable_sum = min(se->avg.load_sum, load_sum); | |
3642 | } | |
3643 | ||
3644 | /* | |
3645 | * runnable_sum can't be lower than running_sum | |
23127296 VG |
3646 | * Rescale running sum to be in the same range as runnable sum |
3647 | * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT] | |
3648 | * runnable_sum is in [0 : LOAD_AVG_MAX] | |
a4c3c049 | 3649 | */ |
23127296 | 3650 | running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; |
a4c3c049 VG |
3651 | runnable_sum = max(runnable_sum, running_sum); |
3652 | ||
2d02fa8c VG |
3653 | load_sum = se_weight(se) * runnable_sum; |
3654 | load_avg = div_u64(load_sum, divider); | |
83c5e9d5 | 3655 | |
2d02fa8c VG |
3656 | delta_avg = load_avg - se->avg.load_avg; |
3657 | if (!delta_avg) | |
83c5e9d5 | 3658 | return; |
09a43ace | 3659 | |
2d02fa8c | 3660 | delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum; |
7c7ad626 | 3661 | |
2d02fa8c VG |
3662 | se->avg.load_sum = runnable_sum; |
3663 | se->avg.load_avg = load_avg; | |
3664 | add_positive(&cfs_rq->avg.load_avg, delta_avg); | |
3665 | add_positive(&cfs_rq->avg.load_sum, delta_sum); | |
3666 | /* See update_cfs_rq_load_avg() */ | |
3667 | cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum, | |
3668 | cfs_rq->avg.load_avg * PELT_MIN_DIVIDER); | |
09a43ace VG |
3669 | } |
3670 | ||
0e2d2aaa | 3671 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) |
09a43ace | 3672 | { |
0e2d2aaa PZ |
3673 | cfs_rq->propagate = 1; |
3674 | cfs_rq->prop_runnable_sum += runnable_sum; | |
09a43ace VG |
3675 | } |
3676 | ||
3677 | /* Update task and its cfs_rq load average */ | |
3678 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3679 | { | |
0e2d2aaa | 3680 | struct cfs_rq *cfs_rq, *gcfs_rq; |
09a43ace VG |
3681 | |
3682 | if (entity_is_task(se)) | |
3683 | return 0; | |
3684 | ||
0e2d2aaa PZ |
3685 | gcfs_rq = group_cfs_rq(se); |
3686 | if (!gcfs_rq->propagate) | |
09a43ace VG |
3687 | return 0; |
3688 | ||
0e2d2aaa PZ |
3689 | gcfs_rq->propagate = 0; |
3690 | ||
09a43ace VG |
3691 | cfs_rq = cfs_rq_of(se); |
3692 | ||
0e2d2aaa | 3693 | add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); |
09a43ace | 3694 | |
0e2d2aaa | 3695 | update_tg_cfs_util(cfs_rq, se, gcfs_rq); |
9f683953 | 3696 | update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); |
0dacee1b | 3697 | update_tg_cfs_load(cfs_rq, se, gcfs_rq); |
09a43ace | 3698 | |
ba19f51f | 3699 | trace_pelt_cfs_tp(cfs_rq); |
8de6242c | 3700 | trace_pelt_se_tp(se); |
ba19f51f | 3701 | |
09a43ace VG |
3702 | return 1; |
3703 | } | |
3704 | ||
bc427898 VG |
3705 | /* |
3706 | * Check if we need to update the load and the utilization of a blocked | |
3707 | * group_entity: | |
3708 | */ | |
3709 | static inline bool skip_blocked_update(struct sched_entity *se) | |
3710 | { | |
3711 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3712 | ||
3713 | /* | |
3714 | * If sched_entity still have not zero load or utilization, we have to | |
3715 | * decay it: | |
3716 | */ | |
3717 | if (se->avg.load_avg || se->avg.util_avg) | |
3718 | return false; | |
3719 | ||
3720 | /* | |
3721 | * If there is a pending propagation, we have to update the load and | |
3722 | * the utilization of the sched_entity: | |
3723 | */ | |
0e2d2aaa | 3724 | if (gcfs_rq->propagate) |
bc427898 VG |
3725 | return false; |
3726 | ||
3727 | /* | |
3728 | * Otherwise, the load and the utilization of the sched_entity is | |
3729 | * already zero and there is no pending propagation, so it will be a | |
3730 | * waste of time to try to decay it: | |
3731 | */ | |
3732 | return true; | |
3733 | } | |
3734 | ||
6e83125c | 3735 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
09a43ace | 3736 | |
fe749158 | 3737 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {} |
09a43ace VG |
3738 | |
3739 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3740 | { | |
3741 | return 0; | |
3742 | } | |
3743 | ||
0e2d2aaa | 3744 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} |
09a43ace | 3745 | |
6e83125c | 3746 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 3747 | |
e2f3e35f VD |
3748 | #ifdef CONFIG_NO_HZ_COMMON |
3749 | static inline void migrate_se_pelt_lag(struct sched_entity *se) | |
3750 | { | |
3751 | u64 throttled = 0, now, lut; | |
3752 | struct cfs_rq *cfs_rq; | |
3753 | struct rq *rq; | |
3754 | bool is_idle; | |
3755 | ||
3756 | if (load_avg_is_decayed(&se->avg)) | |
3757 | return; | |
3758 | ||
3759 | cfs_rq = cfs_rq_of(se); | |
3760 | rq = rq_of(cfs_rq); | |
3761 | ||
3762 | rcu_read_lock(); | |
3763 | is_idle = is_idle_task(rcu_dereference(rq->curr)); | |
3764 | rcu_read_unlock(); | |
3765 | ||
3766 | /* | |
3767 | * The lag estimation comes with a cost we don't want to pay all the | |
3768 | * time. Hence, limiting to the case where the source CPU is idle and | |
3769 | * we know we are at the greatest risk to have an outdated clock. | |
3770 | */ | |
3771 | if (!is_idle) | |
3772 | return; | |
3773 | ||
3774 | /* | |
3775 | * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where: | |
3776 | * | |
3777 | * last_update_time (the cfs_rq's last_update_time) | |
3778 | * = cfs_rq_clock_pelt()@cfs_rq_idle | |
3779 | * = rq_clock_pelt()@cfs_rq_idle | |
3780 | * - cfs->throttled_clock_pelt_time@cfs_rq_idle | |
3781 | * | |
3782 | * cfs_idle_lag (delta between rq's update and cfs_rq's update) | |
3783 | * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle | |
3784 | * | |
3785 | * rq_idle_lag (delta between now and rq's update) | |
3786 | * = sched_clock_cpu() - rq_clock()@rq_idle | |
3787 | * | |
3788 | * We can then write: | |
3789 | * | |
3790 | * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time + | |
3791 | * sched_clock_cpu() - rq_clock()@rq_idle | |
3792 | * Where: | |
3793 | * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle | |
3794 | * rq_clock()@rq_idle is rq->clock_idle | |
3795 | * cfs->throttled_clock_pelt_time@cfs_rq_idle | |
3796 | * is cfs_rq->throttled_pelt_idle | |
3797 | */ | |
3798 | ||
3799 | #ifdef CONFIG_CFS_BANDWIDTH | |
3800 | throttled = u64_u32_load(cfs_rq->throttled_pelt_idle); | |
3801 | /* The clock has been stopped for throttling */ | |
3802 | if (throttled == U64_MAX) | |
3803 | return; | |
3804 | #endif | |
3805 | now = u64_u32_load(rq->clock_pelt_idle); | |
3806 | /* | |
3807 | * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case | |
3808 | * is observed the old clock_pelt_idle value and the new clock_idle, | |
3809 | * which lead to an underestimation. The opposite would lead to an | |
3810 | * overestimation. | |
3811 | */ | |
3812 | smp_rmb(); | |
3813 | lut = cfs_rq_last_update_time(cfs_rq); | |
3814 | ||
3815 | now -= throttled; | |
3816 | if (now < lut) | |
3817 | /* | |
3818 | * cfs_rq->avg.last_update_time is more recent than our | |
3819 | * estimation, let's use it. | |
3820 | */ | |
3821 | now = lut; | |
3822 | else | |
3823 | now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle); | |
3824 | ||
3825 | __update_load_avg_blocked_se(now, se); | |
3826 | } | |
3827 | #else | |
3828 | static void migrate_se_pelt_lag(struct sched_entity *se) {} | |
3829 | #endif | |
3830 | ||
3d30544f PZ |
3831 | /** |
3832 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
23127296 | 3833 | * @now: current time, as per cfs_rq_clock_pelt() |
3d30544f | 3834 | * @cfs_rq: cfs_rq to update |
3d30544f PZ |
3835 | * |
3836 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
d6531ab6 | 3837 | * avg. The immediate corollary is that all (fair) tasks must be attached. |
3d30544f PZ |
3838 | * |
3839 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
3840 | * | |
a315da5e | 3841 | * Return: true if the load decayed or we removed load. |
7c3edd2c PZ |
3842 | * |
3843 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
3844 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 3845 | */ |
a2c6c91f | 3846 | static inline int |
3a123bbb | 3847 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) |
2dac754e | 3848 | { |
9f683953 | 3849 | unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0; |
9d89c257 | 3850 | struct sched_avg *sa = &cfs_rq->avg; |
2a2f5d4e | 3851 | int decayed = 0; |
2dac754e | 3852 | |
2a2f5d4e PZ |
3853 | if (cfs_rq->removed.nr) { |
3854 | unsigned long r; | |
87e867b4 | 3855 | u32 divider = get_pelt_divider(&cfs_rq->avg); |
2a2f5d4e PZ |
3856 | |
3857 | raw_spin_lock(&cfs_rq->removed.lock); | |
3858 | swap(cfs_rq->removed.util_avg, removed_util); | |
3859 | swap(cfs_rq->removed.load_avg, removed_load); | |
9f683953 | 3860 | swap(cfs_rq->removed.runnable_avg, removed_runnable); |
2a2f5d4e PZ |
3861 | cfs_rq->removed.nr = 0; |
3862 | raw_spin_unlock(&cfs_rq->removed.lock); | |
3863 | ||
2a2f5d4e | 3864 | r = removed_load; |
89741892 | 3865 | sub_positive(&sa->load_avg, r); |
2d02fa8c VG |
3866 | sub_positive(&sa->load_sum, r * divider); |
3867 | /* See sa->util_sum below */ | |
3868 | sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER); | |
2dac754e | 3869 | |
2a2f5d4e | 3870 | r = removed_util; |
89741892 | 3871 | sub_positive(&sa->util_avg, r); |
98b0d890 VG |
3872 | sub_positive(&sa->util_sum, r * divider); |
3873 | /* | |
3874 | * Because of rounding, se->util_sum might ends up being +1 more than | |
3875 | * cfs->util_sum. Although this is not a problem by itself, detaching | |
3876 | * a lot of tasks with the rounding problem between 2 updates of | |
3877 | * util_avg (~1ms) can make cfs->util_sum becoming null whereas | |
3878 | * cfs_util_avg is not. | |
3879 | * Check that util_sum is still above its lower bound for the new | |
3880 | * util_avg. Given that period_contrib might have moved since the last | |
3881 | * sync, we are only sure that util_sum must be above or equal to | |
3882 | * util_avg * minimum possible divider | |
3883 | */ | |
3884 | sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER); | |
2a2f5d4e | 3885 | |
9f683953 VG |
3886 | r = removed_runnable; |
3887 | sub_positive(&sa->runnable_avg, r); | |
95246d1e VG |
3888 | sub_positive(&sa->runnable_sum, r * divider); |
3889 | /* See sa->util_sum above */ | |
3890 | sa->runnable_sum = max_t(u32, sa->runnable_sum, | |
3891 | sa->runnable_avg * PELT_MIN_DIVIDER); | |
9f683953 VG |
3892 | |
3893 | /* | |
3894 | * removed_runnable is the unweighted version of removed_load so we | |
3895 | * can use it to estimate removed_load_sum. | |
3896 | */ | |
3897 | add_tg_cfs_propagate(cfs_rq, | |
3898 | -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT); | |
2a2f5d4e PZ |
3899 | |
3900 | decayed = 1; | |
9d89c257 | 3901 | } |
36ee28e4 | 3902 | |
23127296 | 3903 | decayed |= __update_load_avg_cfs_rq(now, cfs_rq); |
d05b4305 VD |
3904 | u64_u32_store_copy(sa->last_update_time, |
3905 | cfs_rq->last_update_time_copy, | |
3906 | sa->last_update_time); | |
2a2f5d4e | 3907 | return decayed; |
21e96f88 SM |
3908 | } |
3909 | ||
3d30544f PZ |
3910 | /** |
3911 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
3912 | * @cfs_rq: cfs_rq to attach to | |
3913 | * @se: sched_entity to attach | |
3914 | * | |
3915 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3916 | * cfs_rq->avg.last_update_time being current. | |
3917 | */ | |
a4f9a0e5 | 3918 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
a05e8c51 | 3919 | { |
95d68593 VG |
3920 | /* |
3921 | * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. | |
3922 | * See ___update_load_avg() for details. | |
3923 | */ | |
87e867b4 | 3924 | u32 divider = get_pelt_divider(&cfs_rq->avg); |
f207934f PZ |
3925 | |
3926 | /* | |
3927 | * When we attach the @se to the @cfs_rq, we must align the decay | |
3928 | * window because without that, really weird and wonderful things can | |
3929 | * happen. | |
3930 | * | |
3931 | * XXX illustrate | |
3932 | */ | |
a05e8c51 | 3933 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
f207934f PZ |
3934 | se->avg.period_contrib = cfs_rq->avg.period_contrib; |
3935 | ||
3936 | /* | |
3937 | * Hell(o) Nasty stuff.. we need to recompute _sum based on the new | |
3938 | * period_contrib. This isn't strictly correct, but since we're | |
3939 | * entirely outside of the PELT hierarchy, nobody cares if we truncate | |
3940 | * _sum a little. | |
3941 | */ | |
3942 | se->avg.util_sum = se->avg.util_avg * divider; | |
3943 | ||
9f683953 VG |
3944 | se->avg.runnable_sum = se->avg.runnable_avg * divider; |
3945 | ||
40f5aa4c | 3946 | se->avg.load_sum = se->avg.load_avg * divider; |
3947 | if (se_weight(se) < se->avg.load_sum) | |
3948 | se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se)); | |
3949 | else | |
3950 | se->avg.load_sum = 1; | |
f207934f | 3951 | |
8d5b9025 | 3952 | enqueue_load_avg(cfs_rq, se); |
a05e8c51 BP |
3953 | cfs_rq->avg.util_avg += se->avg.util_avg; |
3954 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
9f683953 VG |
3955 | cfs_rq->avg.runnable_avg += se->avg.runnable_avg; |
3956 | cfs_rq->avg.runnable_sum += se->avg.runnable_sum; | |
0e2d2aaa PZ |
3957 | |
3958 | add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); | |
a2c6c91f | 3959 | |
a4f9a0e5 | 3960 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
3961 | |
3962 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
3963 | } |
3964 | ||
3d30544f PZ |
3965 | /** |
3966 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
3967 | * @cfs_rq: cfs_rq to detach from | |
3968 | * @se: sched_entity to detach | |
3969 | * | |
3970 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3971 | * cfs_rq->avg.last_update_time being current. | |
3972 | */ | |
a05e8c51 BP |
3973 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3974 | { | |
8d5b9025 | 3975 | dequeue_load_avg(cfs_rq, se); |
89741892 | 3976 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); |
7ceb7710 VG |
3977 | sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); |
3978 | /* See update_cfs_rq_load_avg() */ | |
3979 | cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum, | |
3980 | cfs_rq->avg.util_avg * PELT_MIN_DIVIDER); | |
3981 | ||
9f683953 | 3982 | sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg); |
95246d1e VG |
3983 | sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum); |
3984 | /* See update_cfs_rq_load_avg() */ | |
3985 | cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum, | |
3986 | cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER); | |
0e2d2aaa PZ |
3987 | |
3988 | add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); | |
a2c6c91f | 3989 | |
ea14b57e | 3990 | cfs_rq_util_change(cfs_rq, 0); |
ba19f51f QY |
3991 | |
3992 | trace_pelt_cfs_tp(cfs_rq); | |
a05e8c51 BP |
3993 | } |
3994 | ||
b382a531 PZ |
3995 | /* |
3996 | * Optional action to be done while updating the load average | |
3997 | */ | |
3998 | #define UPDATE_TG 0x1 | |
3999 | #define SKIP_AGE_LOAD 0x2 | |
4000 | #define DO_ATTACH 0x4 | |
e1f078f5 | 4001 | #define DO_DETACH 0x8 |
b382a531 PZ |
4002 | |
4003 | /* Update task and its cfs_rq load average */ | |
4004 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
4005 | { | |
23127296 | 4006 | u64 now = cfs_rq_clock_pelt(cfs_rq); |
b382a531 PZ |
4007 | int decayed; |
4008 | ||
4009 | /* | |
4010 | * Track task load average for carrying it to new CPU after migrated, and | |
4011 | * track group sched_entity load average for task_h_load calc in migration | |
4012 | */ | |
4013 | if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) | |
23127296 | 4014 | __update_load_avg_se(now, cfs_rq, se); |
b382a531 PZ |
4015 | |
4016 | decayed = update_cfs_rq_load_avg(now, cfs_rq); | |
4017 | decayed |= propagate_entity_load_avg(se); | |
4018 | ||
4019 | if (!se->avg.last_update_time && (flags & DO_ATTACH)) { | |
4020 | ||
ea14b57e PZ |
4021 | /* |
4022 | * DO_ATTACH means we're here from enqueue_entity(). | |
4023 | * !last_update_time means we've passed through | |
4024 | * migrate_task_rq_fair() indicating we migrated. | |
4025 | * | |
4026 | * IOW we're enqueueing a task on a new CPU. | |
4027 | */ | |
a4f9a0e5 | 4028 | attach_entity_load_avg(cfs_rq, se); |
fe749158 | 4029 | update_tg_load_avg(cfs_rq); |
b382a531 | 4030 | |
e1f078f5 CZ |
4031 | } else if (flags & DO_DETACH) { |
4032 | /* | |
4033 | * DO_DETACH means we're here from dequeue_entity() | |
4034 | * and we are migrating task out of the CPU. | |
4035 | */ | |
4036 | detach_entity_load_avg(cfs_rq, se); | |
4037 | update_tg_load_avg(cfs_rq); | |
bef69dd8 VG |
4038 | } else if (decayed) { |
4039 | cfs_rq_util_change(cfs_rq, 0); | |
4040 | ||
4041 | if (flags & UPDATE_TG) | |
fe749158 | 4042 | update_tg_load_avg(cfs_rq); |
bef69dd8 | 4043 | } |
b382a531 PZ |
4044 | } |
4045 | ||
104cb16d MR |
4046 | /* |
4047 | * Synchronize entity load avg of dequeued entity without locking | |
4048 | * the previous rq. | |
4049 | */ | |
71b47eaf | 4050 | static void sync_entity_load_avg(struct sched_entity *se) |
104cb16d MR |
4051 | { |
4052 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4053 | u64 last_update_time; | |
4054 | ||
4055 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
23127296 | 4056 | __update_load_avg_blocked_se(last_update_time, se); |
104cb16d MR |
4057 | } |
4058 | ||
0905f04e YD |
4059 | /* |
4060 | * Task first catches up with cfs_rq, and then subtract | |
4061 | * itself from the cfs_rq (task must be off the queue now). | |
4062 | */ | |
71b47eaf | 4063 | static void remove_entity_load_avg(struct sched_entity *se) |
0905f04e YD |
4064 | { |
4065 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2a2f5d4e | 4066 | unsigned long flags; |
0905f04e YD |
4067 | |
4068 | /* | |
7dc603c9 | 4069 | * tasks cannot exit without having gone through wake_up_new_task() -> |
d6531ab6 CZ |
4070 | * enqueue_task_fair() which will have added things to the cfs_rq, |
4071 | * so we can remove unconditionally. | |
0905f04e | 4072 | */ |
0905f04e | 4073 | |
104cb16d | 4074 | sync_entity_load_avg(se); |
2a2f5d4e PZ |
4075 | |
4076 | raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); | |
4077 | ++cfs_rq->removed.nr; | |
4078 | cfs_rq->removed.util_avg += se->avg.util_avg; | |
4079 | cfs_rq->removed.load_avg += se->avg.load_avg; | |
9f683953 | 4080 | cfs_rq->removed.runnable_avg += se->avg.runnable_avg; |
2a2f5d4e | 4081 | raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); |
2dac754e | 4082 | } |
642dbc39 | 4083 | |
9f683953 VG |
4084 | static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq) |
4085 | { | |
4086 | return cfs_rq->avg.runnable_avg; | |
4087 | } | |
4088 | ||
7ea241af YD |
4089 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) |
4090 | { | |
4091 | return cfs_rq->avg.load_avg; | |
4092 | } | |
4093 | ||
d91cecc1 CY |
4094 | static int newidle_balance(struct rq *this_rq, struct rq_flags *rf); |
4095 | ||
7f65ea42 PB |
4096 | static inline unsigned long task_util(struct task_struct *p) |
4097 | { | |
4098 | return READ_ONCE(p->se.avg.util_avg); | |
4099 | } | |
4100 | ||
4101 | static inline unsigned long _task_util_est(struct task_struct *p) | |
4102 | { | |
4103 | struct util_est ue = READ_ONCE(p->se.avg.util_est); | |
4104 | ||
68d7a190 | 4105 | return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED)); |
7f65ea42 PB |
4106 | } |
4107 | ||
4108 | static inline unsigned long task_util_est(struct task_struct *p) | |
4109 | { | |
4110 | return max(task_util(p), _task_util_est(p)); | |
4111 | } | |
4112 | ||
a7008c07 VS |
4113 | #ifdef CONFIG_UCLAMP_TASK |
4114 | static inline unsigned long uclamp_task_util(struct task_struct *p) | |
4115 | { | |
4116 | return clamp(task_util_est(p), | |
4117 | uclamp_eff_value(p, UCLAMP_MIN), | |
4118 | uclamp_eff_value(p, UCLAMP_MAX)); | |
4119 | } | |
4120 | #else | |
4121 | static inline unsigned long uclamp_task_util(struct task_struct *p) | |
4122 | { | |
4123 | return task_util_est(p); | |
4124 | } | |
4125 | #endif | |
4126 | ||
7f65ea42 PB |
4127 | static inline void util_est_enqueue(struct cfs_rq *cfs_rq, |
4128 | struct task_struct *p) | |
4129 | { | |
4130 | unsigned int enqueued; | |
4131 | ||
4132 | if (!sched_feat(UTIL_EST)) | |
4133 | return; | |
4134 | ||
4135 | /* Update root cfs_rq's estimated utilization */ | |
4136 | enqueued = cfs_rq->avg.util_est.enqueued; | |
92a801e5 | 4137 | enqueued += _task_util_est(p); |
7f65ea42 | 4138 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); |
4581bea8 VD |
4139 | |
4140 | trace_sched_util_est_cfs_tp(cfs_rq); | |
7f65ea42 PB |
4141 | } |
4142 | ||
8c1f560c XY |
4143 | static inline void util_est_dequeue(struct cfs_rq *cfs_rq, |
4144 | struct task_struct *p) | |
4145 | { | |
4146 | unsigned int enqueued; | |
4147 | ||
4148 | if (!sched_feat(UTIL_EST)) | |
4149 | return; | |
4150 | ||
4151 | /* Update root cfs_rq's estimated utilization */ | |
4152 | enqueued = cfs_rq->avg.util_est.enqueued; | |
4153 | enqueued -= min_t(unsigned int, enqueued, _task_util_est(p)); | |
4154 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); | |
4155 | ||
4156 | trace_sched_util_est_cfs_tp(cfs_rq); | |
4157 | } | |
4158 | ||
b89997aa VD |
4159 | #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100) |
4160 | ||
7f65ea42 PB |
4161 | /* |
4162 | * Check if a (signed) value is within a specified (unsigned) margin, | |
4163 | * based on the observation that: | |
4164 | * | |
4165 | * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1) | |
4166 | * | |
3b03706f | 4167 | * NOTE: this only works when value + margin < INT_MAX. |
7f65ea42 PB |
4168 | */ |
4169 | static inline bool within_margin(int value, int margin) | |
4170 | { | |
4171 | return ((unsigned int)(value + margin - 1) < (2 * margin - 1)); | |
4172 | } | |
4173 | ||
8c1f560c XY |
4174 | static inline void util_est_update(struct cfs_rq *cfs_rq, |
4175 | struct task_struct *p, | |
4176 | bool task_sleep) | |
7f65ea42 | 4177 | { |
b89997aa | 4178 | long last_ewma_diff, last_enqueued_diff; |
7f65ea42 PB |
4179 | struct util_est ue; |
4180 | ||
4181 | if (!sched_feat(UTIL_EST)) | |
4182 | return; | |
4183 | ||
7f65ea42 PB |
4184 | /* |
4185 | * Skip update of task's estimated utilization when the task has not | |
4186 | * yet completed an activation, e.g. being migrated. | |
4187 | */ | |
4188 | if (!task_sleep) | |
4189 | return; | |
4190 | ||
d519329f PB |
4191 | /* |
4192 | * If the PELT values haven't changed since enqueue time, | |
4193 | * skip the util_est update. | |
4194 | */ | |
4195 | ue = p->se.avg.util_est; | |
4196 | if (ue.enqueued & UTIL_AVG_UNCHANGED) | |
4197 | return; | |
4198 | ||
b89997aa VD |
4199 | last_enqueued_diff = ue.enqueued; |
4200 | ||
b8c96361 PB |
4201 | /* |
4202 | * Reset EWMA on utilization increases, the moving average is used only | |
4203 | * to smooth utilization decreases. | |
4204 | */ | |
68d7a190 | 4205 | ue.enqueued = task_util(p); |
b8c96361 PB |
4206 | if (sched_feat(UTIL_EST_FASTUP)) { |
4207 | if (ue.ewma < ue.enqueued) { | |
4208 | ue.ewma = ue.enqueued; | |
4209 | goto done; | |
4210 | } | |
4211 | } | |
4212 | ||
7f65ea42 | 4213 | /* |
b89997aa | 4214 | * Skip update of task's estimated utilization when its members are |
7f65ea42 PB |
4215 | * already ~1% close to its last activation value. |
4216 | */ | |
7f65ea42 | 4217 | last_ewma_diff = ue.enqueued - ue.ewma; |
b89997aa VD |
4218 | last_enqueued_diff -= ue.enqueued; |
4219 | if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) { | |
4220 | if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN)) | |
4221 | goto done; | |
4222 | ||
7f65ea42 | 4223 | return; |
b89997aa | 4224 | } |
7f65ea42 | 4225 | |
10a35e68 VG |
4226 | /* |
4227 | * To avoid overestimation of actual task utilization, skip updates if | |
4228 | * we cannot grant there is idle time in this CPU. | |
4229 | */ | |
8c1f560c | 4230 | if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq)))) |
10a35e68 VG |
4231 | return; |
4232 | ||
7f65ea42 PB |
4233 | /* |
4234 | * Update Task's estimated utilization | |
4235 | * | |
4236 | * When *p completes an activation we can consolidate another sample | |
4237 | * of the task size. This is done by storing the current PELT value | |
4238 | * as ue.enqueued and by using this value to update the Exponential | |
4239 | * Weighted Moving Average (EWMA): | |
4240 | * | |
4241 | * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) | |
4242 | * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) | |
4243 | * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) | |
4244 | * = w * ( last_ewma_diff ) + ewma(t-1) | |
4245 | * = w * (last_ewma_diff + ewma(t-1) / w) | |
4246 | * | |
4247 | * Where 'w' is the weight of new samples, which is configured to be | |
4248 | * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT) | |
4249 | */ | |
4250 | ue.ewma <<= UTIL_EST_WEIGHT_SHIFT; | |
4251 | ue.ewma += last_ewma_diff; | |
4252 | ue.ewma >>= UTIL_EST_WEIGHT_SHIFT; | |
b8c96361 | 4253 | done: |
68d7a190 | 4254 | ue.enqueued |= UTIL_AVG_UNCHANGED; |
7f65ea42 | 4255 | WRITE_ONCE(p->se.avg.util_est, ue); |
4581bea8 VD |
4256 | |
4257 | trace_sched_util_est_se_tp(&p->se); | |
7f65ea42 PB |
4258 | } |
4259 | ||
ef8df979 VD |
4260 | static inline int task_fits_capacity(struct task_struct *p, |
4261 | unsigned long capacity) | |
3b1baa64 | 4262 | { |
a7008c07 | 4263 | return fits_capacity(uclamp_task_util(p), capacity); |
3b1baa64 MR |
4264 | } |
4265 | ||
4266 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) | |
4267 | { | |
740cf8a7 | 4268 | if (!sched_asym_cpucap_active()) |
3b1baa64 MR |
4269 | return; |
4270 | ||
0ae78eec | 4271 | if (!p || p->nr_cpus_allowed == 1) { |
3b1baa64 MR |
4272 | rq->misfit_task_load = 0; |
4273 | return; | |
4274 | } | |
4275 | ||
4276 | if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) { | |
4277 | rq->misfit_task_load = 0; | |
4278 | return; | |
4279 | } | |
4280 | ||
01cfcde9 VG |
4281 | /* |
4282 | * Make sure that misfit_task_load will not be null even if | |
4283 | * task_h_load() returns 0. | |
4284 | */ | |
4285 | rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1); | |
3b1baa64 MR |
4286 | } |
4287 | ||
38033c37 PZ |
4288 | #else /* CONFIG_SMP */ |
4289 | ||
a7b359fc OU |
4290 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) |
4291 | { | |
4292 | return true; | |
4293 | } | |
4294 | ||
d31b1a66 VG |
4295 | #define UPDATE_TG 0x0 |
4296 | #define SKIP_AGE_LOAD 0x0 | |
b382a531 | 4297 | #define DO_ATTACH 0x0 |
e1f078f5 | 4298 | #define DO_DETACH 0x0 |
d31b1a66 | 4299 | |
88c0616e | 4300 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) |
536bd00c | 4301 | { |
ea14b57e | 4302 | cfs_rq_util_change(cfs_rq, 0); |
536bd00c RW |
4303 | } |
4304 | ||
9d89c257 | 4305 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 4306 | |
a05e8c51 | 4307 | static inline void |
a4f9a0e5 | 4308 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} |
a05e8c51 BP |
4309 | static inline void |
4310 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
4311 | ||
d91cecc1 | 4312 | static inline int newidle_balance(struct rq *rq, struct rq_flags *rf) |
6e83125c PZ |
4313 | { |
4314 | return 0; | |
4315 | } | |
4316 | ||
7f65ea42 PB |
4317 | static inline void |
4318 | util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {} | |
4319 | ||
4320 | static inline void | |
8c1f560c XY |
4321 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {} |
4322 | ||
4323 | static inline void | |
4324 | util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p, | |
4325 | bool task_sleep) {} | |
3b1baa64 | 4326 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} |
7f65ea42 | 4327 | |
38033c37 | 4328 | #endif /* CONFIG_SMP */ |
9d85f21c | 4329 | |
ddc97297 PZ |
4330 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4331 | { | |
4332 | #ifdef CONFIG_SCHED_DEBUG | |
4333 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
4334 | ||
4335 | if (d < 0) | |
4336 | d = -d; | |
4337 | ||
4338 | if (d > 3*sysctl_sched_latency) | |
ae92882e | 4339 | schedstat_inc(cfs_rq->nr_spread_over); |
ddc97297 PZ |
4340 | #endif |
4341 | } | |
4342 | ||
aeb73b04 PZ |
4343 | static void |
4344 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
4345 | { | |
1af5f730 | 4346 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 4347 | |
2cb8600e PZ |
4348 | /* |
4349 | * The 'current' period is already promised to the current tasks, | |
4350 | * however the extra weight of the new task will slow them down a | |
4351 | * little, place the new task so that it fits in the slot that | |
4352 | * stays open at the end. | |
4353 | */ | |
94dfb5e7 | 4354 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 4355 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 4356 | |
a2e7a7eb | 4357 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 4358 | if (!initial) { |
2cae3948 JD |
4359 | unsigned long thresh; |
4360 | ||
4361 | if (se_is_idle(se)) | |
4362 | thresh = sysctl_sched_min_granularity; | |
4363 | else | |
4364 | thresh = sysctl_sched_latency; | |
a7be37ac | 4365 | |
a2e7a7eb MG |
4366 | /* |
4367 | * Halve their sleep time's effect, to allow | |
4368 | * for a gentler effect of sleepers: | |
4369 | */ | |
4370 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
4371 | thresh >>= 1; | |
51e0304c | 4372 | |
a2e7a7eb | 4373 | vruntime -= thresh; |
aeb73b04 PZ |
4374 | } |
4375 | ||
b5d9d734 | 4376 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 4377 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
4378 | } |
4379 | ||
d3d9dc33 PT |
4380 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
4381 | ||
fe61468b | 4382 | static inline bool cfs_bandwidth_used(void); |
b5179ac7 PZ |
4383 | |
4384 | /* | |
4385 | * MIGRATION | |
4386 | * | |
4387 | * dequeue | |
4388 | * update_curr() | |
4389 | * update_min_vruntime() | |
4390 | * vruntime -= min_vruntime | |
4391 | * | |
4392 | * enqueue | |
4393 | * update_curr() | |
4394 | * update_min_vruntime() | |
4395 | * vruntime += min_vruntime | |
4396 | * | |
4397 | * this way the vruntime transition between RQs is done when both | |
4398 | * min_vruntime are up-to-date. | |
4399 | * | |
4400 | * WAKEUP (remote) | |
4401 | * | |
59efa0ba | 4402 | * ->migrate_task_rq_fair() (p->state == TASK_WAKING) |
b5179ac7 PZ |
4403 | * vruntime -= min_vruntime |
4404 | * | |
4405 | * enqueue | |
4406 | * update_curr() | |
4407 | * update_min_vruntime() | |
4408 | * vruntime += min_vruntime | |
4409 | * | |
4410 | * this way we don't have the most up-to-date min_vruntime on the originating | |
4411 | * CPU and an up-to-date min_vruntime on the destination CPU. | |
4412 | */ | |
4413 | ||
bf0f6f24 | 4414 | static void |
88ec22d3 | 4415 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4416 | { |
2f950354 PZ |
4417 | bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); |
4418 | bool curr = cfs_rq->curr == se; | |
4419 | ||
88ec22d3 | 4420 | /* |
2f950354 PZ |
4421 | * If we're the current task, we must renormalise before calling |
4422 | * update_curr(). | |
88ec22d3 | 4423 | */ |
2f950354 | 4424 | if (renorm && curr) |
88ec22d3 PZ |
4425 | se->vruntime += cfs_rq->min_vruntime; |
4426 | ||
2f950354 PZ |
4427 | update_curr(cfs_rq); |
4428 | ||
bf0f6f24 | 4429 | /* |
2f950354 PZ |
4430 | * Otherwise, renormalise after, such that we're placed at the current |
4431 | * moment in time, instead of some random moment in the past. Being | |
4432 | * placed in the past could significantly boost this task to the | |
4433 | * fairness detriment of existing tasks. | |
bf0f6f24 | 4434 | */ |
2f950354 PZ |
4435 | if (renorm && !curr) |
4436 | se->vruntime += cfs_rq->min_vruntime; | |
4437 | ||
89ee048f VG |
4438 | /* |
4439 | * When enqueuing a sched_entity, we must: | |
4440 | * - Update loads to have both entity and cfs_rq synced with now. | |
859f2062 CZ |
4441 | * - For group_entity, update its runnable_weight to reflect the new |
4442 | * h_nr_running of its group cfs_rq. | |
89ee048f VG |
4443 | * - For group_entity, update its weight to reflect the new share of |
4444 | * its group cfs_rq | |
4445 | * - Add its new weight to cfs_rq->load.weight | |
4446 | */ | |
b382a531 | 4447 | update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); |
9f683953 | 4448 | se_update_runnable(se); |
1ea6c46a | 4449 | update_cfs_group(se); |
17bc14b7 | 4450 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 4451 | |
1a3d027c | 4452 | if (flags & ENQUEUE_WAKEUP) |
aeb73b04 | 4453 | place_entity(cfs_rq, se, 0); |
bf0f6f24 | 4454 | |
cb251765 | 4455 | check_schedstat_required(); |
60f2415e | 4456 | update_stats_enqueue_fair(cfs_rq, se, flags); |
4fa8d299 | 4457 | check_spread(cfs_rq, se); |
2f950354 | 4458 | if (!curr) |
83b699ed | 4459 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 4460 | se->on_rq = 1; |
3d4b47b4 | 4461 | |
51bf903b | 4462 | if (cfs_rq->nr_running == 1) { |
d3d9dc33 | 4463 | check_enqueue_throttle(cfs_rq); |
51bf903b CZ |
4464 | if (!throttled_hierarchy(cfs_rq)) |
4465 | list_add_leaf_cfs_rq(cfs_rq); | |
4466 | } | |
bf0f6f24 IM |
4467 | } |
4468 | ||
2c13c919 | 4469 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 4470 | { |
2c13c919 RR |
4471 | for_each_sched_entity(se) { |
4472 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4473 | if (cfs_rq->last != se) |
2c13c919 | 4474 | break; |
f1044799 PZ |
4475 | |
4476 | cfs_rq->last = NULL; | |
2c13c919 RR |
4477 | } |
4478 | } | |
2002c695 | 4479 | |
2c13c919 RR |
4480 | static void __clear_buddies_next(struct sched_entity *se) |
4481 | { | |
4482 | for_each_sched_entity(se) { | |
4483 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4484 | if (cfs_rq->next != se) |
2c13c919 | 4485 | break; |
f1044799 PZ |
4486 | |
4487 | cfs_rq->next = NULL; | |
2c13c919 | 4488 | } |
2002c695 PZ |
4489 | } |
4490 | ||
ac53db59 RR |
4491 | static void __clear_buddies_skip(struct sched_entity *se) |
4492 | { | |
4493 | for_each_sched_entity(se) { | |
4494 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4495 | if (cfs_rq->skip != se) |
ac53db59 | 4496 | break; |
f1044799 PZ |
4497 | |
4498 | cfs_rq->skip = NULL; | |
ac53db59 RR |
4499 | } |
4500 | } | |
4501 | ||
a571bbea PZ |
4502 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4503 | { | |
2c13c919 RR |
4504 | if (cfs_rq->last == se) |
4505 | __clear_buddies_last(se); | |
4506 | ||
4507 | if (cfs_rq->next == se) | |
4508 | __clear_buddies_next(se); | |
ac53db59 RR |
4509 | |
4510 | if (cfs_rq->skip == se) | |
4511 | __clear_buddies_skip(se); | |
a571bbea PZ |
4512 | } |
4513 | ||
6c16a6dc | 4514 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 4515 | |
bf0f6f24 | 4516 | static void |
371fd7e7 | 4517 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4518 | { |
e1f078f5 CZ |
4519 | int action = UPDATE_TG; |
4520 | ||
4521 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se))) | |
4522 | action |= DO_DETACH; | |
4523 | ||
a2a2d680 DA |
4524 | /* |
4525 | * Update run-time statistics of the 'current'. | |
4526 | */ | |
4527 | update_curr(cfs_rq); | |
89ee048f VG |
4528 | |
4529 | /* | |
4530 | * When dequeuing a sched_entity, we must: | |
4531 | * - Update loads to have both entity and cfs_rq synced with now. | |
859f2062 CZ |
4532 | * - For group_entity, update its runnable_weight to reflect the new |
4533 | * h_nr_running of its group cfs_rq. | |
dfcb245e | 4534 | * - Subtract its previous weight from cfs_rq->load.weight. |
89ee048f VG |
4535 | * - For group entity, update its weight to reflect the new share |
4536 | * of its group cfs_rq. | |
4537 | */ | |
e1f078f5 | 4538 | update_load_avg(cfs_rq, se, action); |
9f683953 | 4539 | se_update_runnable(se); |
a2a2d680 | 4540 | |
60f2415e | 4541 | update_stats_dequeue_fair(cfs_rq, se, flags); |
67e9fb2a | 4542 | |
2002c695 | 4543 | clear_buddies(cfs_rq, se); |
4793241b | 4544 | |
83b699ed | 4545 | if (se != cfs_rq->curr) |
30cfdcfc | 4546 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 4547 | se->on_rq = 0; |
30cfdcfc | 4548 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
4549 | |
4550 | /* | |
b60205c7 PZ |
4551 | * Normalize after update_curr(); which will also have moved |
4552 | * min_vruntime if @se is the one holding it back. But before doing | |
4553 | * update_min_vruntime() again, which will discount @se's position and | |
4554 | * can move min_vruntime forward still more. | |
88ec22d3 | 4555 | */ |
371fd7e7 | 4556 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 4557 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 4558 | |
d8b4986d PT |
4559 | /* return excess runtime on last dequeue */ |
4560 | return_cfs_rq_runtime(cfs_rq); | |
4561 | ||
1ea6c46a | 4562 | update_cfs_group(se); |
b60205c7 PZ |
4563 | |
4564 | /* | |
4565 | * Now advance min_vruntime if @se was the entity holding it back, | |
4566 | * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be | |
4567 | * put back on, and if we advance min_vruntime, we'll be placed back | |
4568 | * further than we started -- ie. we'll be penalized. | |
4569 | */ | |
9845c49c | 4570 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE) |
b60205c7 | 4571 | update_min_vruntime(cfs_rq); |
e2f3e35f VD |
4572 | |
4573 | if (cfs_rq->nr_running == 0) | |
4574 | update_idle_cfs_rq_clock_pelt(cfs_rq); | |
bf0f6f24 IM |
4575 | } |
4576 | ||
4577 | /* | |
4578 | * Preempt the current task with a newly woken task if needed: | |
4579 | */ | |
7c92e54f | 4580 | static void |
2e09bf55 | 4581 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 4582 | { |
11697830 | 4583 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
4584 | struct sched_entity *se; |
4585 | s64 delta; | |
11697830 | 4586 | |
6d0f0ebd | 4587 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 4588 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 4589 | if (delta_exec > ideal_runtime) { |
8875125e | 4590 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
4591 | /* |
4592 | * The current task ran long enough, ensure it doesn't get | |
4593 | * re-elected due to buddy favours. | |
4594 | */ | |
4595 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
4596 | return; |
4597 | } | |
4598 | ||
4599 | /* | |
4600 | * Ensure that a task that missed wakeup preemption by a | |
4601 | * narrow margin doesn't have to wait for a full slice. | |
4602 | * This also mitigates buddy induced latencies under load. | |
4603 | */ | |
f685ceac MG |
4604 | if (delta_exec < sysctl_sched_min_granularity) |
4605 | return; | |
4606 | ||
f4cfb33e WX |
4607 | se = __pick_first_entity(cfs_rq); |
4608 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 4609 | |
f4cfb33e WX |
4610 | if (delta < 0) |
4611 | return; | |
d7d82944 | 4612 | |
f4cfb33e | 4613 | if (delta > ideal_runtime) |
8875125e | 4614 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
4615 | } |
4616 | ||
83b699ed | 4617 | static void |
8494f412 | 4618 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 4619 | { |
21f56ffe PZ |
4620 | clear_buddies(cfs_rq, se); |
4621 | ||
83b699ed SV |
4622 | /* 'current' is not kept within the tree. */ |
4623 | if (se->on_rq) { | |
4624 | /* | |
4625 | * Any task has to be enqueued before it get to execute on | |
4626 | * a CPU. So account for the time it spent waiting on the | |
4627 | * runqueue. | |
4628 | */ | |
60f2415e | 4629 | update_stats_wait_end_fair(cfs_rq, se); |
83b699ed | 4630 | __dequeue_entity(cfs_rq, se); |
88c0616e | 4631 | update_load_avg(cfs_rq, se, UPDATE_TG); |
83b699ed SV |
4632 | } |
4633 | ||
79303e9e | 4634 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 4635 | cfs_rq->curr = se; |
4fa8d299 | 4636 | |
eba1ed4b IM |
4637 | /* |
4638 | * Track our maximum slice length, if the CPU's load is at | |
4639 | * least twice that of our own weight (i.e. dont track it | |
4640 | * when there are only lesser-weight tasks around): | |
4641 | */ | |
f2bedc47 DE |
4642 | if (schedstat_enabled() && |
4643 | rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { | |
ceeadb83 YS |
4644 | struct sched_statistics *stats; |
4645 | ||
4646 | stats = __schedstats_from_se(se); | |
4647 | __schedstat_set(stats->slice_max, | |
4648 | max((u64)stats->slice_max, | |
a2dcb276 | 4649 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); |
eba1ed4b | 4650 | } |
4fa8d299 | 4651 | |
4a55b450 | 4652 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
4653 | } |
4654 | ||
3f3a4904 PZ |
4655 | static int |
4656 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
4657 | ||
ac53db59 RR |
4658 | /* |
4659 | * Pick the next process, keeping these things in mind, in this order: | |
4660 | * 1) keep things fair between processes/task groups | |
4661 | * 2) pick the "next" process, since someone really wants that to run | |
4662 | * 3) pick the "last" process, for cache locality | |
4663 | * 4) do not run the "skip" process, if something else is available | |
4664 | */ | |
678d5718 PZ |
4665 | static struct sched_entity * |
4666 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 4667 | { |
678d5718 PZ |
4668 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
4669 | struct sched_entity *se; | |
4670 | ||
4671 | /* | |
4672 | * If curr is set we have to see if its left of the leftmost entity | |
4673 | * still in the tree, provided there was anything in the tree at all. | |
4674 | */ | |
4675 | if (!left || (curr && entity_before(curr, left))) | |
4676 | left = curr; | |
4677 | ||
4678 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 4679 | |
ac53db59 RR |
4680 | /* |
4681 | * Avoid running the skip buddy, if running something else can | |
4682 | * be done without getting too unfair. | |
4683 | */ | |
21f56ffe | 4684 | if (cfs_rq->skip && cfs_rq->skip == se) { |
678d5718 PZ |
4685 | struct sched_entity *second; |
4686 | ||
4687 | if (se == curr) { | |
4688 | second = __pick_first_entity(cfs_rq); | |
4689 | } else { | |
4690 | second = __pick_next_entity(se); | |
4691 | if (!second || (curr && entity_before(curr, second))) | |
4692 | second = curr; | |
4693 | } | |
4694 | ||
ac53db59 RR |
4695 | if (second && wakeup_preempt_entity(second, left) < 1) |
4696 | se = second; | |
4697 | } | |
aa2ac252 | 4698 | |
9abb8973 PO |
4699 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) { |
4700 | /* | |
4701 | * Someone really wants this to run. If it's not unfair, run it. | |
4702 | */ | |
ac53db59 | 4703 | se = cfs_rq->next; |
9abb8973 PO |
4704 | } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) { |
4705 | /* | |
4706 | * Prefer last buddy, try to return the CPU to a preempted task. | |
4707 | */ | |
4708 | se = cfs_rq->last; | |
4709 | } | |
ac53db59 | 4710 | |
4793241b | 4711 | return se; |
aa2ac252 PZ |
4712 | } |
4713 | ||
678d5718 | 4714 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 4715 | |
ab6cde26 | 4716 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
4717 | { |
4718 | /* | |
4719 | * If still on the runqueue then deactivate_task() | |
4720 | * was not called and update_curr() has to be done: | |
4721 | */ | |
4722 | if (prev->on_rq) | |
b7cc0896 | 4723 | update_curr(cfs_rq); |
bf0f6f24 | 4724 | |
d3d9dc33 PT |
4725 | /* throttle cfs_rqs exceeding runtime */ |
4726 | check_cfs_rq_runtime(cfs_rq); | |
4727 | ||
4fa8d299 | 4728 | check_spread(cfs_rq, prev); |
cb251765 | 4729 | |
30cfdcfc | 4730 | if (prev->on_rq) { |
60f2415e | 4731 | update_stats_wait_start_fair(cfs_rq, prev); |
30cfdcfc DA |
4732 | /* Put 'current' back into the tree. */ |
4733 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 4734 | /* in !on_rq case, update occurred at dequeue */ |
88c0616e | 4735 | update_load_avg(cfs_rq, prev, 0); |
30cfdcfc | 4736 | } |
429d43bc | 4737 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
4738 | } |
4739 | ||
8f4d37ec PZ |
4740 | static void |
4741 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 4742 | { |
bf0f6f24 | 4743 | /* |
30cfdcfc | 4744 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 4745 | */ |
30cfdcfc | 4746 | update_curr(cfs_rq); |
bf0f6f24 | 4747 | |
9d85f21c PT |
4748 | /* |
4749 | * Ensure that runnable average is periodically updated. | |
4750 | */ | |
88c0616e | 4751 | update_load_avg(cfs_rq, curr, UPDATE_TG); |
1ea6c46a | 4752 | update_cfs_group(curr); |
9d85f21c | 4753 | |
8f4d37ec PZ |
4754 | #ifdef CONFIG_SCHED_HRTICK |
4755 | /* | |
4756 | * queued ticks are scheduled to match the slice, so don't bother | |
4757 | * validating it and just reschedule. | |
4758 | */ | |
983ed7a6 | 4759 | if (queued) { |
8875125e | 4760 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
4761 | return; |
4762 | } | |
8f4d37ec PZ |
4763 | /* |
4764 | * don't let the period tick interfere with the hrtick preemption | |
4765 | */ | |
4766 | if (!sched_feat(DOUBLE_TICK) && | |
4767 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
4768 | return; | |
4769 | #endif | |
4770 | ||
2c2efaed | 4771 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 4772 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
4773 | } |
4774 | ||
ab84d31e PT |
4775 | |
4776 | /************************************************** | |
4777 | * CFS bandwidth control machinery | |
4778 | */ | |
4779 | ||
4780 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb | 4781 | |
e9666d10 | 4782 | #ifdef CONFIG_JUMP_LABEL |
c5905afb | 4783 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
4784 | |
4785 | static inline bool cfs_bandwidth_used(void) | |
4786 | { | |
c5905afb | 4787 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
4788 | } |
4789 | ||
1ee14e6c | 4790 | void cfs_bandwidth_usage_inc(void) |
029632fb | 4791 | { |
ce48c146 | 4792 | static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used); |
1ee14e6c BS |
4793 | } |
4794 | ||
4795 | void cfs_bandwidth_usage_dec(void) | |
4796 | { | |
ce48c146 | 4797 | static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used); |
029632fb | 4798 | } |
e9666d10 | 4799 | #else /* CONFIG_JUMP_LABEL */ |
029632fb PZ |
4800 | static bool cfs_bandwidth_used(void) |
4801 | { | |
4802 | return true; | |
4803 | } | |
4804 | ||
1ee14e6c BS |
4805 | void cfs_bandwidth_usage_inc(void) {} |
4806 | void cfs_bandwidth_usage_dec(void) {} | |
e9666d10 | 4807 | #endif /* CONFIG_JUMP_LABEL */ |
029632fb | 4808 | |
ab84d31e PT |
4809 | /* |
4810 | * default period for cfs group bandwidth. | |
4811 | * default: 0.1s, units: nanoseconds | |
4812 | */ | |
4813 | static inline u64 default_cfs_period(void) | |
4814 | { | |
4815 | return 100000000ULL; | |
4816 | } | |
ec12cb7f PT |
4817 | |
4818 | static inline u64 sched_cfs_bandwidth_slice(void) | |
4819 | { | |
4820 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
4821 | } | |
4822 | ||
a9cf55b2 | 4823 | /* |
763a9ec0 QC |
4824 | * Replenish runtime according to assigned quota. We use sched_clock_cpu |
4825 | * directly instead of rq->clock to avoid adding additional synchronization | |
4826 | * around rq->lock. | |
a9cf55b2 PT |
4827 | * |
4828 | * requires cfs_b->lock | |
4829 | */ | |
029632fb | 4830 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 | 4831 | { |
bcb1704a HC |
4832 | s64 runtime; |
4833 | ||
f4183717 HC |
4834 | if (unlikely(cfs_b->quota == RUNTIME_INF)) |
4835 | return; | |
4836 | ||
4837 | cfs_b->runtime += cfs_b->quota; | |
bcb1704a HC |
4838 | runtime = cfs_b->runtime_snap - cfs_b->runtime; |
4839 | if (runtime > 0) { | |
4840 | cfs_b->burst_time += runtime; | |
4841 | cfs_b->nr_burst++; | |
4842 | } | |
4843 | ||
f4183717 | 4844 | cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst); |
bcb1704a | 4845 | cfs_b->runtime_snap = cfs_b->runtime; |
a9cf55b2 PT |
4846 | } |
4847 | ||
029632fb PZ |
4848 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4849 | { | |
4850 | return &tg->cfs_bandwidth; | |
4851 | } | |
4852 | ||
85dac906 | 4853 | /* returns 0 on failure to allocate runtime */ |
e98fa02c PT |
4854 | static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b, |
4855 | struct cfs_rq *cfs_rq, u64 target_runtime) | |
ec12cb7f | 4856 | { |
e98fa02c PT |
4857 | u64 min_amount, amount = 0; |
4858 | ||
4859 | lockdep_assert_held(&cfs_b->lock); | |
ec12cb7f PT |
4860 | |
4861 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
e98fa02c | 4862 | min_amount = target_runtime - cfs_rq->runtime_remaining; |
ec12cb7f | 4863 | |
ec12cb7f PT |
4864 | if (cfs_b->quota == RUNTIME_INF) |
4865 | amount = min_amount; | |
58088ad0 | 4866 | else { |
77a4d1a1 | 4867 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
4868 | |
4869 | if (cfs_b->runtime > 0) { | |
4870 | amount = min(cfs_b->runtime, min_amount); | |
4871 | cfs_b->runtime -= amount; | |
4872 | cfs_b->idle = 0; | |
4873 | } | |
ec12cb7f | 4874 | } |
ec12cb7f PT |
4875 | |
4876 | cfs_rq->runtime_remaining += amount; | |
85dac906 PT |
4877 | |
4878 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
4879 | } |
4880 | ||
e98fa02c PT |
4881 | /* returns 0 on failure to allocate runtime */ |
4882 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4883 | { | |
4884 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4885 | int ret; | |
4886 | ||
4887 | raw_spin_lock(&cfs_b->lock); | |
4888 | ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice()); | |
4889 | raw_spin_unlock(&cfs_b->lock); | |
4890 | ||
4891 | return ret; | |
4892 | } | |
4893 | ||
9dbdb155 | 4894 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
4895 | { |
4896 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 4897 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
4898 | |
4899 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
4900 | return; |
4901 | ||
5e2d2cc2 L |
4902 | if (cfs_rq->throttled) |
4903 | return; | |
85dac906 PT |
4904 | /* |
4905 | * if we're unable to extend our runtime we resched so that the active | |
4906 | * hierarchy can be throttled | |
4907 | */ | |
4908 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 4909 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
4910 | } |
4911 | ||
6c16a6dc | 4912 | static __always_inline |
9dbdb155 | 4913 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 4914 | { |
56f570e5 | 4915 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
4916 | return; |
4917 | ||
4918 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
4919 | } | |
4920 | ||
85dac906 PT |
4921 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
4922 | { | |
56f570e5 | 4923 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
4924 | } |
4925 | ||
64660c86 PT |
4926 | /* check whether cfs_rq, or any parent, is throttled */ |
4927 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4928 | { | |
56f570e5 | 4929 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
4930 | } |
4931 | ||
4932 | /* | |
4933 | * Ensure that neither of the group entities corresponding to src_cpu or | |
4934 | * dest_cpu are members of a throttled hierarchy when performing group | |
4935 | * load-balance operations. | |
4936 | */ | |
4937 | static inline int throttled_lb_pair(struct task_group *tg, | |
4938 | int src_cpu, int dest_cpu) | |
4939 | { | |
4940 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
4941 | ||
4942 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
4943 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
4944 | ||
4945 | return throttled_hierarchy(src_cfs_rq) || | |
4946 | throttled_hierarchy(dest_cfs_rq); | |
4947 | } | |
4948 | ||
64660c86 PT |
4949 | static int tg_unthrottle_up(struct task_group *tg, void *data) |
4950 | { | |
4951 | struct rq *rq = data; | |
4952 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4953 | ||
4954 | cfs_rq->throttle_count--; | |
64660c86 | 4955 | if (!cfs_rq->throttle_count) { |
64eaf507 CZ |
4956 | cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) - |
4957 | cfs_rq->throttled_clock_pelt; | |
31bc6aea | 4958 | |
a7b359fc | 4959 | /* Add cfs_rq with load or one or more already running entities to the list */ |
0a00a354 | 4960 | if (!cfs_rq_is_decayed(cfs_rq)) |
31bc6aea | 4961 | list_add_leaf_cfs_rq(cfs_rq); |
64660c86 | 4962 | } |
64660c86 PT |
4963 | |
4964 | return 0; | |
4965 | } | |
4966 | ||
4967 | static int tg_throttle_down(struct task_group *tg, void *data) | |
4968 | { | |
4969 | struct rq *rq = data; | |
4970 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4971 | ||
82958366 | 4972 | /* group is entering throttled state, stop time */ |
31bc6aea | 4973 | if (!cfs_rq->throttle_count) { |
64eaf507 | 4974 | cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq); |
31bc6aea VG |
4975 | list_del_leaf_cfs_rq(cfs_rq); |
4976 | } | |
64660c86 PT |
4977 | cfs_rq->throttle_count++; |
4978 | ||
4979 | return 0; | |
4980 | } | |
4981 | ||
e98fa02c | 4982 | static bool throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
4983 | { |
4984 | struct rq *rq = rq_of(cfs_rq); | |
4985 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4986 | struct sched_entity *se; | |
43e9f7f2 | 4987 | long task_delta, idle_task_delta, dequeue = 1; |
e98fa02c PT |
4988 | |
4989 | raw_spin_lock(&cfs_b->lock); | |
4990 | /* This will start the period timer if necessary */ | |
4991 | if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) { | |
4992 | /* | |
4993 | * We have raced with bandwidth becoming available, and if we | |
4994 | * actually throttled the timer might not unthrottle us for an | |
4995 | * entire period. We additionally needed to make sure that any | |
4996 | * subsequent check_cfs_rq_runtime calls agree not to throttle | |
4997 | * us, as we may commit to do cfs put_prev+pick_next, so we ask | |
4998 | * for 1ns of runtime rather than just check cfs_b. | |
4999 | */ | |
5000 | dequeue = 0; | |
5001 | } else { | |
5002 | list_add_tail_rcu(&cfs_rq->throttled_list, | |
5003 | &cfs_b->throttled_cfs_rq); | |
5004 | } | |
5005 | raw_spin_unlock(&cfs_b->lock); | |
5006 | ||
5007 | if (!dequeue) | |
5008 | return false; /* Throttle no longer required. */ | |
85dac906 PT |
5009 | |
5010 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
5011 | ||
f1b17280 | 5012 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
5013 | rcu_read_lock(); |
5014 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
5015 | rcu_read_unlock(); | |
85dac906 PT |
5016 | |
5017 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 5018 | idle_task_delta = cfs_rq->idle_h_nr_running; |
85dac906 PT |
5019 | for_each_sched_entity(se) { |
5020 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
5021 | /* throttled entity or throttle-on-deactivate */ | |
5022 | if (!se->on_rq) | |
b6d37a76 | 5023 | goto done; |
85dac906 | 5024 | |
b6d37a76 | 5025 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); |
6212437f | 5026 | |
30400039 JD |
5027 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
5028 | idle_task_delta = cfs_rq->h_nr_running; | |
5029 | ||
85dac906 | 5030 | qcfs_rq->h_nr_running -= task_delta; |
43e9f7f2 | 5031 | qcfs_rq->idle_h_nr_running -= idle_task_delta; |
85dac906 | 5032 | |
b6d37a76 PW |
5033 | if (qcfs_rq->load.weight) { |
5034 | /* Avoid re-evaluating load for this entity: */ | |
5035 | se = parent_entity(se); | |
5036 | break; | |
5037 | } | |
5038 | } | |
5039 | ||
5040 | for_each_sched_entity(se) { | |
5041 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
5042 | /* throttled entity or throttle-on-deactivate */ | |
5043 | if (!se->on_rq) | |
5044 | goto done; | |
5045 | ||
5046 | update_load_avg(qcfs_rq, se, 0); | |
5047 | se_update_runnable(se); | |
5048 | ||
30400039 JD |
5049 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
5050 | idle_task_delta = cfs_rq->h_nr_running; | |
5051 | ||
b6d37a76 PW |
5052 | qcfs_rq->h_nr_running -= task_delta; |
5053 | qcfs_rq->idle_h_nr_running -= idle_task_delta; | |
85dac906 PT |
5054 | } |
5055 | ||
b6d37a76 PW |
5056 | /* At this point se is NULL and we are at root level*/ |
5057 | sub_nr_running(rq, task_delta); | |
85dac906 | 5058 | |
b6d37a76 | 5059 | done: |
c06f04c7 | 5060 | /* |
e98fa02c PT |
5061 | * Note: distribution will already see us throttled via the |
5062 | * throttled-list. rq->lock protects completion. | |
c06f04c7 | 5063 | */ |
e98fa02c PT |
5064 | cfs_rq->throttled = 1; |
5065 | cfs_rq->throttled_clock = rq_clock(rq); | |
5066 | return true; | |
85dac906 PT |
5067 | } |
5068 | ||
029632fb | 5069 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
5070 | { |
5071 | struct rq *rq = rq_of(cfs_rq); | |
5072 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5073 | struct sched_entity *se; | |
43e9f7f2 | 5074 | long task_delta, idle_task_delta; |
671fd9da | 5075 | |
22b958d8 | 5076 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
5077 | |
5078 | cfs_rq->throttled = 0; | |
1a55af2e FW |
5079 | |
5080 | update_rq_clock(rq); | |
5081 | ||
671fd9da | 5082 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 5083 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
5084 | list_del_rcu(&cfs_rq->throttled_list); |
5085 | raw_spin_unlock(&cfs_b->lock); | |
5086 | ||
64660c86 PT |
5087 | /* update hierarchical throttle state */ |
5088 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
5089 | ||
2630cde2 | 5090 | if (!cfs_rq->load.weight) { |
51bf903b CZ |
5091 | if (!cfs_rq->on_list) |
5092 | return; | |
5093 | /* | |
5094 | * Nothing to run but something to decay (on_list)? | |
5095 | * Complete the branch. | |
5096 | */ | |
5097 | for_each_sched_entity(se) { | |
5098 | if (list_add_leaf_cfs_rq(cfs_rq_of(se))) | |
5099 | break; | |
5100 | } | |
5101 | goto unthrottle_throttle; | |
2630cde2 | 5102 | } |
671fd9da PT |
5103 | |
5104 | task_delta = cfs_rq->h_nr_running; | |
43e9f7f2 | 5105 | idle_task_delta = cfs_rq->idle_h_nr_running; |
671fd9da | 5106 | for_each_sched_entity(se) { |
30400039 JD |
5107 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); |
5108 | ||
671fd9da | 5109 | if (se->on_rq) |
39f23ce0 | 5110 | break; |
30400039 JD |
5111 | enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP); |
5112 | ||
5113 | if (cfs_rq_is_idle(group_cfs_rq(se))) | |
5114 | idle_task_delta = cfs_rq->h_nr_running; | |
39f23ce0 | 5115 | |
30400039 JD |
5116 | qcfs_rq->h_nr_running += task_delta; |
5117 | qcfs_rq->idle_h_nr_running += idle_task_delta; | |
39f23ce0 VG |
5118 | |
5119 | /* end evaluation on encountering a throttled cfs_rq */ | |
30400039 | 5120 | if (cfs_rq_throttled(qcfs_rq)) |
39f23ce0 VG |
5121 | goto unthrottle_throttle; |
5122 | } | |
671fd9da | 5123 | |
39f23ce0 | 5124 | for_each_sched_entity(se) { |
30400039 | 5125 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); |
39f23ce0 | 5126 | |
30400039 | 5127 | update_load_avg(qcfs_rq, se, UPDATE_TG); |
39f23ce0 | 5128 | se_update_runnable(se); |
6212437f | 5129 | |
30400039 JD |
5130 | if (cfs_rq_is_idle(group_cfs_rq(se))) |
5131 | idle_task_delta = cfs_rq->h_nr_running; | |
671fd9da | 5132 | |
30400039 JD |
5133 | qcfs_rq->h_nr_running += task_delta; |
5134 | qcfs_rq->idle_h_nr_running += idle_task_delta; | |
39f23ce0 VG |
5135 | |
5136 | /* end evaluation on encountering a throttled cfs_rq */ | |
30400039 | 5137 | if (cfs_rq_throttled(qcfs_rq)) |
39f23ce0 | 5138 | goto unthrottle_throttle; |
671fd9da PT |
5139 | } |
5140 | ||
39f23ce0 VG |
5141 | /* At this point se is NULL and we are at root level*/ |
5142 | add_nr_running(rq, task_delta); | |
671fd9da | 5143 | |
39f23ce0 | 5144 | unthrottle_throttle: |
fe61468b VG |
5145 | assert_list_leaf_cfs_rq(rq); |
5146 | ||
97fb7a0a | 5147 | /* Determine whether we need to wake up potentially idle CPU: */ |
671fd9da | 5148 | if (rq->curr == rq->idle && rq->cfs.nr_running) |
8875125e | 5149 | resched_curr(rq); |
671fd9da PT |
5150 | } |
5151 | ||
26a8b127 | 5152 | static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b) |
671fd9da PT |
5153 | { |
5154 | struct cfs_rq *cfs_rq; | |
26a8b127 | 5155 | u64 runtime, remaining = 1; |
671fd9da PT |
5156 | |
5157 | rcu_read_lock(); | |
5158 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
5159 | throttled_list) { | |
5160 | struct rq *rq = rq_of(cfs_rq); | |
8a8c69c3 | 5161 | struct rq_flags rf; |
671fd9da | 5162 | |
c0ad4aa4 | 5163 | rq_lock_irqsave(rq, &rf); |
671fd9da PT |
5164 | if (!cfs_rq_throttled(cfs_rq)) |
5165 | goto next; | |
5166 | ||
5e2d2cc2 L |
5167 | /* By the above check, this should never be true */ |
5168 | SCHED_WARN_ON(cfs_rq->runtime_remaining > 0); | |
5169 | ||
26a8b127 | 5170 | raw_spin_lock(&cfs_b->lock); |
671fd9da | 5171 | runtime = -cfs_rq->runtime_remaining + 1; |
26a8b127 HC |
5172 | if (runtime > cfs_b->runtime) |
5173 | runtime = cfs_b->runtime; | |
5174 | cfs_b->runtime -= runtime; | |
5175 | remaining = cfs_b->runtime; | |
5176 | raw_spin_unlock(&cfs_b->lock); | |
671fd9da PT |
5177 | |
5178 | cfs_rq->runtime_remaining += runtime; | |
671fd9da PT |
5179 | |
5180 | /* we check whether we're throttled above */ | |
5181 | if (cfs_rq->runtime_remaining > 0) | |
5182 | unthrottle_cfs_rq(cfs_rq); | |
5183 | ||
5184 | next: | |
c0ad4aa4 | 5185 | rq_unlock_irqrestore(rq, &rf); |
671fd9da PT |
5186 | |
5187 | if (!remaining) | |
5188 | break; | |
5189 | } | |
5190 | rcu_read_unlock(); | |
671fd9da PT |
5191 | } |
5192 | ||
58088ad0 PT |
5193 | /* |
5194 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
5195 | * cfs_rqs as appropriate. If there has been no activity within the last | |
5196 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
5197 | * used to track this state. | |
5198 | */ | |
c0ad4aa4 | 5199 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) |
58088ad0 | 5200 | { |
51f2176d | 5201 | int throttled; |
58088ad0 | 5202 | |
58088ad0 PT |
5203 | /* no need to continue the timer with no bandwidth constraint */ |
5204 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 5205 | goto out_deactivate; |
58088ad0 | 5206 | |
671fd9da | 5207 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 5208 | cfs_b->nr_periods += overrun; |
671fd9da | 5209 | |
f4183717 HC |
5210 | /* Refill extra burst quota even if cfs_b->idle */ |
5211 | __refill_cfs_bandwidth_runtime(cfs_b); | |
5212 | ||
51f2176d BS |
5213 | /* |
5214 | * idle depends on !throttled (for the case of a large deficit), and if | |
5215 | * we're going inactive then everything else can be deferred | |
5216 | */ | |
5217 | if (cfs_b->idle && !throttled) | |
5218 | goto out_deactivate; | |
a9cf55b2 | 5219 | |
671fd9da PT |
5220 | if (!throttled) { |
5221 | /* mark as potentially idle for the upcoming period */ | |
5222 | cfs_b->idle = 1; | |
51f2176d | 5223 | return 0; |
671fd9da PT |
5224 | } |
5225 | ||
e8da1b18 NR |
5226 | /* account preceding periods in which throttling occurred */ |
5227 | cfs_b->nr_throttled += overrun; | |
5228 | ||
671fd9da | 5229 | /* |
26a8b127 | 5230 | * This check is repeated as we release cfs_b->lock while we unthrottle. |
671fd9da | 5231 | */ |
ab93a4bc | 5232 | while (throttled && cfs_b->runtime > 0) { |
c0ad4aa4 | 5233 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
671fd9da | 5234 | /* we can't nest cfs_b->lock while distributing bandwidth */ |
26a8b127 | 5235 | distribute_cfs_runtime(cfs_b); |
c0ad4aa4 | 5236 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
671fd9da PT |
5237 | |
5238 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
5239 | } | |
58088ad0 | 5240 | |
671fd9da PT |
5241 | /* |
5242 | * While we are ensured activity in the period following an | |
5243 | * unthrottle, this also covers the case in which the new bandwidth is | |
5244 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
5245 | * timer to remain active while there are any throttled entities.) | |
5246 | */ | |
5247 | cfs_b->idle = 0; | |
58088ad0 | 5248 | |
51f2176d BS |
5249 | return 0; |
5250 | ||
5251 | out_deactivate: | |
51f2176d | 5252 | return 1; |
58088ad0 | 5253 | } |
d3d9dc33 | 5254 | |
d8b4986d PT |
5255 | /* a cfs_rq won't donate quota below this amount */ |
5256 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
5257 | /* minimum remaining period time to redistribute slack quota */ | |
5258 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
5259 | /* how long we wait to gather additional slack before distributing */ | |
5260 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
5261 | ||
db06e78c BS |
5262 | /* |
5263 | * Are we near the end of the current quota period? | |
5264 | * | |
5265 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 5266 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
5267 | * migrate_hrtimers, base is never cleared, so we are fine. |
5268 | */ | |
d8b4986d PT |
5269 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
5270 | { | |
5271 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
72d0ad7c | 5272 | s64 remaining; |
d8b4986d PT |
5273 | |
5274 | /* if the call-back is running a quota refresh is already occurring */ | |
5275 | if (hrtimer_callback_running(refresh_timer)) | |
5276 | return 1; | |
5277 | ||
5278 | /* is a quota refresh about to occur? */ | |
5279 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
72d0ad7c | 5280 | if (remaining < (s64)min_expire) |
d8b4986d PT |
5281 | return 1; |
5282 | ||
5283 | return 0; | |
5284 | } | |
5285 | ||
5286 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
5287 | { | |
5288 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
5289 | ||
5290 | /* if there's a quota refresh soon don't bother with slack */ | |
5291 | if (runtime_refresh_within(cfs_b, min_left)) | |
5292 | return; | |
5293 | ||
66567fcb | 5294 | /* don't push forwards an existing deferred unthrottle */ |
5295 | if (cfs_b->slack_started) | |
5296 | return; | |
5297 | cfs_b->slack_started = true; | |
5298 | ||
4cfafd30 PZ |
5299 | hrtimer_start(&cfs_b->slack_timer, |
5300 | ns_to_ktime(cfs_bandwidth_slack_period), | |
5301 | HRTIMER_MODE_REL); | |
d8b4986d PT |
5302 | } |
5303 | ||
5304 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
5305 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5306 | { | |
5307 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
5308 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
5309 | ||
5310 | if (slack_runtime <= 0) | |
5311 | return; | |
5312 | ||
5313 | raw_spin_lock(&cfs_b->lock); | |
de53fd7a | 5314 | if (cfs_b->quota != RUNTIME_INF) { |
d8b4986d PT |
5315 | cfs_b->runtime += slack_runtime; |
5316 | ||
5317 | /* we are under rq->lock, defer unthrottling using a timer */ | |
5318 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
5319 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
5320 | start_cfs_slack_bandwidth(cfs_b); | |
5321 | } | |
5322 | raw_spin_unlock(&cfs_b->lock); | |
5323 | ||
5324 | /* even if it's not valid for return we don't want to try again */ | |
5325 | cfs_rq->runtime_remaining -= slack_runtime; | |
5326 | } | |
5327 | ||
5328 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5329 | { | |
56f570e5 PT |
5330 | if (!cfs_bandwidth_used()) |
5331 | return; | |
5332 | ||
fccfdc6f | 5333 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
5334 | return; |
5335 | ||
5336 | __return_cfs_rq_runtime(cfs_rq); | |
5337 | } | |
5338 | ||
5339 | /* | |
5340 | * This is done with a timer (instead of inline with bandwidth return) since | |
5341 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
5342 | */ | |
5343 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
5344 | { | |
5345 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
c0ad4aa4 | 5346 | unsigned long flags; |
d8b4986d PT |
5347 | |
5348 | /* confirm we're still not at a refresh boundary */ | |
c0ad4aa4 | 5349 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
66567fcb | 5350 | cfs_b->slack_started = false; |
baa9be4f | 5351 | |
db06e78c | 5352 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { |
c0ad4aa4 | 5353 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d | 5354 | return; |
db06e78c | 5355 | } |
d8b4986d | 5356 | |
c06f04c7 | 5357 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 5358 | runtime = cfs_b->runtime; |
c06f04c7 | 5359 | |
c0ad4aa4 | 5360 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
5361 | |
5362 | if (!runtime) | |
5363 | return; | |
5364 | ||
26a8b127 | 5365 | distribute_cfs_runtime(cfs_b); |
d8b4986d PT |
5366 | } |
5367 | ||
d3d9dc33 PT |
5368 | /* |
5369 | * When a group wakes up we want to make sure that its quota is not already | |
5370 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
c034f48e | 5371 | * runtime as update_curr() throttling can not trigger until it's on-rq. |
d3d9dc33 PT |
5372 | */ |
5373 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
5374 | { | |
56f570e5 PT |
5375 | if (!cfs_bandwidth_used()) |
5376 | return; | |
5377 | ||
d3d9dc33 PT |
5378 | /* an active group must be handled by the update_curr()->put() path */ |
5379 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
5380 | return; | |
5381 | ||
5382 | /* ensure the group is not already throttled */ | |
5383 | if (cfs_rq_throttled(cfs_rq)) | |
5384 | return; | |
5385 | ||
5386 | /* update runtime allocation */ | |
5387 | account_cfs_rq_runtime(cfs_rq, 0); | |
5388 | if (cfs_rq->runtime_remaining <= 0) | |
5389 | throttle_cfs_rq(cfs_rq); | |
5390 | } | |
5391 | ||
55e16d30 PZ |
5392 | static void sync_throttle(struct task_group *tg, int cpu) |
5393 | { | |
5394 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
5395 | ||
5396 | if (!cfs_bandwidth_used()) | |
5397 | return; | |
5398 | ||
5399 | if (!tg->parent) | |
5400 | return; | |
5401 | ||
5402 | cfs_rq = tg->cfs_rq[cpu]; | |
5403 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
5404 | ||
5405 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
64eaf507 | 5406 | cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu)); |
55e16d30 PZ |
5407 | } |
5408 | ||
d3d9dc33 | 5409 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 5410 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 5411 | { |
56f570e5 | 5412 | if (!cfs_bandwidth_used()) |
678d5718 | 5413 | return false; |
56f570e5 | 5414 | |
d3d9dc33 | 5415 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 5416 | return false; |
d3d9dc33 PT |
5417 | |
5418 | /* | |
5419 | * it's possible for a throttled entity to be forced into a running | |
5420 | * state (e.g. set_curr_task), in this case we're finished. | |
5421 | */ | |
5422 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 5423 | return true; |
d3d9dc33 | 5424 | |
e98fa02c | 5425 | return throttle_cfs_rq(cfs_rq); |
d3d9dc33 | 5426 | } |
029632fb | 5427 | |
029632fb PZ |
5428 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
5429 | { | |
5430 | struct cfs_bandwidth *cfs_b = | |
5431 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 5432 | |
029632fb PZ |
5433 | do_sched_cfs_slack_timer(cfs_b); |
5434 | ||
5435 | return HRTIMER_NORESTART; | |
5436 | } | |
5437 | ||
2e8e1922 PA |
5438 | extern const u64 max_cfs_quota_period; |
5439 | ||
029632fb PZ |
5440 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) |
5441 | { | |
5442 | struct cfs_bandwidth *cfs_b = | |
5443 | container_of(timer, struct cfs_bandwidth, period_timer); | |
c0ad4aa4 | 5444 | unsigned long flags; |
029632fb PZ |
5445 | int overrun; |
5446 | int idle = 0; | |
2e8e1922 | 5447 | int count = 0; |
029632fb | 5448 | |
c0ad4aa4 | 5449 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
029632fb | 5450 | for (;;) { |
77a4d1a1 | 5451 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
5452 | if (!overrun) |
5453 | break; | |
5454 | ||
5a6d6a6c HC |
5455 | idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); |
5456 | ||
2e8e1922 PA |
5457 | if (++count > 3) { |
5458 | u64 new, old = ktime_to_ns(cfs_b->period); | |
5459 | ||
4929a4e6 XZ |
5460 | /* |
5461 | * Grow period by a factor of 2 to avoid losing precision. | |
5462 | * Precision loss in the quota/period ratio can cause __cfs_schedulable | |
5463 | * to fail. | |
5464 | */ | |
5465 | new = old * 2; | |
5466 | if (new < max_cfs_quota_period) { | |
5467 | cfs_b->period = ns_to_ktime(new); | |
5468 | cfs_b->quota *= 2; | |
f4183717 | 5469 | cfs_b->burst *= 2; |
4929a4e6 XZ |
5470 | |
5471 | pr_warn_ratelimited( | |
5472 | "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
5473 | smp_processor_id(), | |
5474 | div_u64(new, NSEC_PER_USEC), | |
5475 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
5476 | } else { | |
5477 | pr_warn_ratelimited( | |
5478 | "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n", | |
5479 | smp_processor_id(), | |
5480 | div_u64(old, NSEC_PER_USEC), | |
5481 | div_u64(cfs_b->quota, NSEC_PER_USEC)); | |
5482 | } | |
2e8e1922 PA |
5483 | |
5484 | /* reset count so we don't come right back in here */ | |
5485 | count = 0; | |
5486 | } | |
029632fb | 5487 | } |
4cfafd30 PZ |
5488 | if (idle) |
5489 | cfs_b->period_active = 0; | |
c0ad4aa4 | 5490 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
029632fb PZ |
5491 | |
5492 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
5493 | } | |
5494 | ||
5495 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5496 | { | |
5497 | raw_spin_lock_init(&cfs_b->lock); | |
5498 | cfs_b->runtime = 0; | |
5499 | cfs_b->quota = RUNTIME_INF; | |
5500 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
f4183717 | 5501 | cfs_b->burst = 0; |
029632fb PZ |
5502 | |
5503 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 5504 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5505 | cfs_b->period_timer.function = sched_cfs_period_timer; |
5506 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
5507 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
66567fcb | 5508 | cfs_b->slack_started = false; |
029632fb PZ |
5509 | } |
5510 | ||
5511 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
5512 | { | |
5513 | cfs_rq->runtime_enabled = 0; | |
5514 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
5515 | } | |
5516 | ||
77a4d1a1 | 5517 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 5518 | { |
4cfafd30 | 5519 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 5520 | |
f1d1be8a XP |
5521 | if (cfs_b->period_active) |
5522 | return; | |
5523 | ||
5524 | cfs_b->period_active = 1; | |
763a9ec0 | 5525 | hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); |
f1d1be8a | 5526 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
5527 | } |
5528 | ||
5529 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
5530 | { | |
7f1a169b TH |
5531 | /* init_cfs_bandwidth() was not called */ |
5532 | if (!cfs_b->throttled_cfs_rq.next) | |
5533 | return; | |
5534 | ||
029632fb PZ |
5535 | hrtimer_cancel(&cfs_b->period_timer); |
5536 | hrtimer_cancel(&cfs_b->slack_timer); | |
5537 | } | |
5538 | ||
502ce005 | 5539 | /* |
97fb7a0a | 5540 | * Both these CPU hotplug callbacks race against unregister_fair_sched_group() |
502ce005 PZ |
5541 | * |
5542 | * The race is harmless, since modifying bandwidth settings of unhooked group | |
5543 | * bits doesn't do much. | |
5544 | */ | |
5545 | ||
3b03706f | 5546 | /* cpu online callback */ |
0e59bdae KT |
5547 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
5548 | { | |
502ce005 | 5549 | struct task_group *tg; |
0e59bdae | 5550 | |
5cb9eaa3 | 5551 | lockdep_assert_rq_held(rq); |
502ce005 PZ |
5552 | |
5553 | rcu_read_lock(); | |
5554 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5555 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
5556 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
0e59bdae KT |
5557 | |
5558 | raw_spin_lock(&cfs_b->lock); | |
5559 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
5560 | raw_spin_unlock(&cfs_b->lock); | |
5561 | } | |
502ce005 | 5562 | rcu_read_unlock(); |
0e59bdae KT |
5563 | } |
5564 | ||
502ce005 | 5565 | /* cpu offline callback */ |
38dc3348 | 5566 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb | 5567 | { |
502ce005 PZ |
5568 | struct task_group *tg; |
5569 | ||
5cb9eaa3 | 5570 | lockdep_assert_rq_held(rq); |
502ce005 PZ |
5571 | |
5572 | rcu_read_lock(); | |
5573 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5574 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
029632fb | 5575 | |
029632fb PZ |
5576 | if (!cfs_rq->runtime_enabled) |
5577 | continue; | |
5578 | ||
5579 | /* | |
5580 | * clock_task is not advancing so we just need to make sure | |
5581 | * there's some valid quota amount | |
5582 | */ | |
51f2176d | 5583 | cfs_rq->runtime_remaining = 1; |
0e59bdae | 5584 | /* |
97fb7a0a | 5585 | * Offline rq is schedulable till CPU is completely disabled |
0e59bdae KT |
5586 | * in take_cpu_down(), so we prevent new cfs throttling here. |
5587 | */ | |
5588 | cfs_rq->runtime_enabled = 0; | |
5589 | ||
029632fb PZ |
5590 | if (cfs_rq_throttled(cfs_rq)) |
5591 | unthrottle_cfs_rq(cfs_rq); | |
5592 | } | |
502ce005 | 5593 | rcu_read_unlock(); |
029632fb PZ |
5594 | } |
5595 | ||
5596 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f6783319 VG |
5597 | |
5598 | static inline bool cfs_bandwidth_used(void) | |
5599 | { | |
5600 | return false; | |
5601 | } | |
5602 | ||
9dbdb155 | 5603 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 5604 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 5605 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 5606 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 5607 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
5608 | |
5609 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
5610 | { | |
5611 | return 0; | |
5612 | } | |
64660c86 PT |
5613 | |
5614 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
5615 | { | |
5616 | return 0; | |
5617 | } | |
5618 | ||
5619 | static inline int throttled_lb_pair(struct task_group *tg, | |
5620 | int src_cpu, int dest_cpu) | |
5621 | { | |
5622 | return 0; | |
5623 | } | |
029632fb PZ |
5624 | |
5625 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
5626 | ||
5627 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5628 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
5629 | #endif |
5630 | ||
029632fb PZ |
5631 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
5632 | { | |
5633 | return NULL; | |
5634 | } | |
5635 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 5636 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 5637 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
5638 | |
5639 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
5640 | ||
bf0f6f24 IM |
5641 | /************************************************** |
5642 | * CFS operations on tasks: | |
5643 | */ | |
5644 | ||
8f4d37ec PZ |
5645 | #ifdef CONFIG_SCHED_HRTICK |
5646 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5647 | { | |
8f4d37ec PZ |
5648 | struct sched_entity *se = &p->se; |
5649 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5650 | ||
9148a3a1 | 5651 | SCHED_WARN_ON(task_rq(p) != rq); |
8f4d37ec | 5652 | |
8bf46a39 | 5653 | if (rq->cfs.h_nr_running > 1) { |
8f4d37ec PZ |
5654 | u64 slice = sched_slice(cfs_rq, se); |
5655 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
5656 | s64 delta = slice - ran; | |
5657 | ||
5658 | if (delta < 0) { | |
65bcf072 | 5659 | if (task_current(rq, p)) |
8875125e | 5660 | resched_curr(rq); |
8f4d37ec PZ |
5661 | return; |
5662 | } | |
31656519 | 5663 | hrtick_start(rq, delta); |
8f4d37ec PZ |
5664 | } |
5665 | } | |
a4c2f00f PZ |
5666 | |
5667 | /* | |
5668 | * called from enqueue/dequeue and updates the hrtick when the | |
5669 | * current task is from our class and nr_running is low enough | |
5670 | * to matter. | |
5671 | */ | |
5672 | static void hrtick_update(struct rq *rq) | |
5673 | { | |
5674 | struct task_struct *curr = rq->curr; | |
5675 | ||
e0ee463c | 5676 | if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
5677 | return; |
5678 | ||
5679 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
5680 | hrtick_start_fair(rq, curr); | |
5681 | } | |
55e12e5e | 5682 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
5683 | static inline void |
5684 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5685 | { | |
5686 | } | |
a4c2f00f PZ |
5687 | |
5688 | static inline void hrtick_update(struct rq *rq) | |
5689 | { | |
5690 | } | |
8f4d37ec PZ |
5691 | #endif |
5692 | ||
2802bf3c | 5693 | #ifdef CONFIG_SMP |
2802bf3c MR |
5694 | static inline bool cpu_overutilized(int cpu) |
5695 | { | |
82762d2a | 5696 | return !fits_capacity(cpu_util_cfs(cpu), capacity_of(cpu)); |
2802bf3c MR |
5697 | } |
5698 | ||
5699 | static inline void update_overutilized_status(struct rq *rq) | |
5700 | { | |
f9f240f9 | 5701 | if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) { |
2802bf3c | 5702 | WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED); |
f9f240f9 QY |
5703 | trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED); |
5704 | } | |
2802bf3c MR |
5705 | } |
5706 | #else | |
5707 | static inline void update_overutilized_status(struct rq *rq) { } | |
5708 | #endif | |
5709 | ||
323af6de VK |
5710 | /* Runqueue only has SCHED_IDLE tasks enqueued */ |
5711 | static int sched_idle_rq(struct rq *rq) | |
5712 | { | |
5713 | return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running && | |
5714 | rq->nr_running); | |
5715 | } | |
5716 | ||
a480adde JD |
5717 | /* |
5718 | * Returns true if cfs_rq only has SCHED_IDLE entities enqueued. Note the use | |
5719 | * of idle_nr_running, which does not consider idle descendants of normal | |
5720 | * entities. | |
5721 | */ | |
5722 | static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq) | |
5723 | { | |
5724 | return cfs_rq->nr_running && | |
5725 | cfs_rq->nr_running == cfs_rq->idle_nr_running; | |
5726 | } | |
5727 | ||
afa70d94 | 5728 | #ifdef CONFIG_SMP |
323af6de VK |
5729 | static int sched_idle_cpu(int cpu) |
5730 | { | |
5731 | return sched_idle_rq(cpu_rq(cpu)); | |
5732 | } | |
afa70d94 | 5733 | #endif |
323af6de | 5734 | |
bf0f6f24 IM |
5735 | /* |
5736 | * The enqueue_task method is called before nr_running is | |
5737 | * increased. Here we update the fair scheduling stats and | |
5738 | * then put the task into the rbtree: | |
5739 | */ | |
ea87bb78 | 5740 | static void |
371fd7e7 | 5741 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5742 | { |
5743 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5744 | struct sched_entity *se = &p->se; |
43e9f7f2 | 5745 | int idle_h_nr_running = task_has_idle_policy(p); |
8e1ac429 | 5746 | int task_new = !(flags & ENQUEUE_WAKEUP); |
bf0f6f24 | 5747 | |
2539fc82 PB |
5748 | /* |
5749 | * The code below (indirectly) updates schedutil which looks at | |
5750 | * the cfs_rq utilization to select a frequency. | |
5751 | * Let's add the task's estimated utilization to the cfs_rq's | |
5752 | * estimated utilization, before we update schedutil. | |
5753 | */ | |
5754 | util_est_enqueue(&rq->cfs, p); | |
5755 | ||
8c34ab19 RW |
5756 | /* |
5757 | * If in_iowait is set, the code below may not trigger any cpufreq | |
5758 | * utilization updates, so do it here explicitly with the IOWAIT flag | |
5759 | * passed. | |
5760 | */ | |
5761 | if (p->in_iowait) | |
674e7541 | 5762 | cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); |
8c34ab19 | 5763 | |
bf0f6f24 | 5764 | for_each_sched_entity(se) { |
62fb1851 | 5765 | if (se->on_rq) |
bf0f6f24 IM |
5766 | break; |
5767 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 5768 | enqueue_entity(cfs_rq, se, flags); |
85dac906 | 5769 | |
953bfcd1 | 5770 | cfs_rq->h_nr_running++; |
43e9f7f2 | 5771 | cfs_rq->idle_h_nr_running += idle_h_nr_running; |
85dac906 | 5772 | |
30400039 JD |
5773 | if (cfs_rq_is_idle(cfs_rq)) |
5774 | idle_h_nr_running = 1; | |
5775 | ||
6d4d2246 VG |
5776 | /* end evaluation on encountering a throttled cfs_rq */ |
5777 | if (cfs_rq_throttled(cfs_rq)) | |
5778 | goto enqueue_throttle; | |
5779 | ||
88ec22d3 | 5780 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 5781 | } |
8f4d37ec | 5782 | |
2069dd75 | 5783 | for_each_sched_entity(se) { |
0f317143 | 5784 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 5785 | |
88c0616e | 5786 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 5787 | se_update_runnable(se); |
1ea6c46a | 5788 | update_cfs_group(se); |
6d4d2246 VG |
5789 | |
5790 | cfs_rq->h_nr_running++; | |
5791 | cfs_rq->idle_h_nr_running += idle_h_nr_running; | |
5ab297ba | 5792 | |
30400039 JD |
5793 | if (cfs_rq_is_idle(cfs_rq)) |
5794 | idle_h_nr_running = 1; | |
5795 | ||
5ab297ba VG |
5796 | /* end evaluation on encountering a throttled cfs_rq */ |
5797 | if (cfs_rq_throttled(cfs_rq)) | |
5798 | goto enqueue_throttle; | |
2069dd75 PZ |
5799 | } |
5800 | ||
7d148be6 VG |
5801 | /* At this point se is NULL and we are at root level*/ |
5802 | add_nr_running(rq, 1); | |
2802bf3c | 5803 | |
7d148be6 VG |
5804 | /* |
5805 | * Since new tasks are assigned an initial util_avg equal to | |
5806 | * half of the spare capacity of their CPU, tiny tasks have the | |
5807 | * ability to cross the overutilized threshold, which will | |
5808 | * result in the load balancer ruining all the task placement | |
5809 | * done by EAS. As a way to mitigate that effect, do not account | |
5810 | * for the first enqueue operation of new tasks during the | |
5811 | * overutilized flag detection. | |
5812 | * | |
5813 | * A better way of solving this problem would be to wait for | |
5814 | * the PELT signals of tasks to converge before taking them | |
5815 | * into account, but that is not straightforward to implement, | |
5816 | * and the following generally works well enough in practice. | |
5817 | */ | |
8e1ac429 | 5818 | if (!task_new) |
7d148be6 | 5819 | update_overutilized_status(rq); |
cd126afe | 5820 | |
7d148be6 | 5821 | enqueue_throttle: |
5d299eab PZ |
5822 | assert_list_leaf_cfs_rq(rq); |
5823 | ||
a4c2f00f | 5824 | hrtick_update(rq); |
bf0f6f24 IM |
5825 | } |
5826 | ||
2f36825b VP |
5827 | static void set_next_buddy(struct sched_entity *se); |
5828 | ||
bf0f6f24 IM |
5829 | /* |
5830 | * The dequeue_task method is called before nr_running is | |
5831 | * decreased. We remove the task from the rbtree and | |
5832 | * update the fair scheduling stats: | |
5833 | */ | |
371fd7e7 | 5834 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5835 | { |
5836 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5837 | struct sched_entity *se = &p->se; |
2f36825b | 5838 | int task_sleep = flags & DEQUEUE_SLEEP; |
43e9f7f2 | 5839 | int idle_h_nr_running = task_has_idle_policy(p); |
323af6de | 5840 | bool was_sched_idle = sched_idle_rq(rq); |
bf0f6f24 | 5841 | |
8c1f560c XY |
5842 | util_est_dequeue(&rq->cfs, p); |
5843 | ||
bf0f6f24 IM |
5844 | for_each_sched_entity(se) { |
5845 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 5846 | dequeue_entity(cfs_rq, se, flags); |
85dac906 | 5847 | |
953bfcd1 | 5848 | cfs_rq->h_nr_running--; |
43e9f7f2 | 5849 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; |
2069dd75 | 5850 | |
30400039 JD |
5851 | if (cfs_rq_is_idle(cfs_rq)) |
5852 | idle_h_nr_running = 1; | |
5853 | ||
6d4d2246 VG |
5854 | /* end evaluation on encountering a throttled cfs_rq */ |
5855 | if (cfs_rq_throttled(cfs_rq)) | |
5856 | goto dequeue_throttle; | |
5857 | ||
bf0f6f24 | 5858 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 5859 | if (cfs_rq->load.weight) { |
754bd598 KK |
5860 | /* Avoid re-evaluating load for this entity: */ |
5861 | se = parent_entity(se); | |
2f36825b VP |
5862 | /* |
5863 | * Bias pick_next to pick a task from this cfs_rq, as | |
5864 | * p is sleeping when it is within its sched_slice. | |
5865 | */ | |
754bd598 KK |
5866 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
5867 | set_next_buddy(se); | |
bf0f6f24 | 5868 | break; |
2f36825b | 5869 | } |
371fd7e7 | 5870 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 5871 | } |
8f4d37ec | 5872 | |
2069dd75 | 5873 | for_each_sched_entity(se) { |
0f317143 | 5874 | cfs_rq = cfs_rq_of(se); |
2069dd75 | 5875 | |
88c0616e | 5876 | update_load_avg(cfs_rq, se, UPDATE_TG); |
9f683953 | 5877 | se_update_runnable(se); |
1ea6c46a | 5878 | update_cfs_group(se); |
6d4d2246 VG |
5879 | |
5880 | cfs_rq->h_nr_running--; | |
5881 | cfs_rq->idle_h_nr_running -= idle_h_nr_running; | |
5ab297ba | 5882 | |
30400039 JD |
5883 | if (cfs_rq_is_idle(cfs_rq)) |
5884 | idle_h_nr_running = 1; | |
5885 | ||
5ab297ba VG |
5886 | /* end evaluation on encountering a throttled cfs_rq */ |
5887 | if (cfs_rq_throttled(cfs_rq)) | |
5888 | goto dequeue_throttle; | |
5889 | ||
2069dd75 PZ |
5890 | } |
5891 | ||
423d02e1 PW |
5892 | /* At this point se is NULL and we are at root level*/ |
5893 | sub_nr_running(rq, 1); | |
cd126afe | 5894 | |
323af6de VK |
5895 | /* balance early to pull high priority tasks */ |
5896 | if (unlikely(!was_sched_idle && sched_idle_rq(rq))) | |
5897 | rq->next_balance = jiffies; | |
5898 | ||
423d02e1 | 5899 | dequeue_throttle: |
8c1f560c | 5900 | util_est_update(&rq->cfs, p, task_sleep); |
a4c2f00f | 5901 | hrtick_update(rq); |
bf0f6f24 IM |
5902 | } |
5903 | ||
e7693a36 | 5904 | #ifdef CONFIG_SMP |
10e2f1ac PZ |
5905 | |
5906 | /* Working cpumask for: load_balance, load_balance_newidle. */ | |
18c31c97 BH |
5907 | static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); |
5908 | static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask); | |
10e2f1ac | 5909 | |
9fd81dd5 | 5910 | #ifdef CONFIG_NO_HZ_COMMON |
e022e0d3 PZ |
5911 | |
5912 | static struct { | |
5913 | cpumask_var_t idle_cpus_mask; | |
5914 | atomic_t nr_cpus; | |
f643ea22 | 5915 | int has_blocked; /* Idle CPUS has blocked load */ |
7fd7a9e0 | 5916 | int needs_update; /* Newly idle CPUs need their next_balance collated */ |
e022e0d3 | 5917 | unsigned long next_balance; /* in jiffy units */ |
f643ea22 | 5918 | unsigned long next_blocked; /* Next update of blocked load in jiffies */ |
e022e0d3 PZ |
5919 | } nohz ____cacheline_aligned; |
5920 | ||
9fd81dd5 | 5921 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 5922 | |
b0fb1eb4 VG |
5923 | static unsigned long cpu_load(struct rq *rq) |
5924 | { | |
5925 | return cfs_rq_load_avg(&rq->cfs); | |
5926 | } | |
5927 | ||
3318544b VG |
5928 | /* |
5929 | * cpu_load_without - compute CPU load without any contributions from *p | |
5930 | * @cpu: the CPU which load is requested | |
5931 | * @p: the task which load should be discounted | |
5932 | * | |
5933 | * The load of a CPU is defined by the load of tasks currently enqueued on that | |
5934 | * CPU as well as tasks which are currently sleeping after an execution on that | |
5935 | * CPU. | |
5936 | * | |
5937 | * This method returns the load of the specified CPU by discounting the load of | |
5938 | * the specified task, whenever the task is currently contributing to the CPU | |
5939 | * load. | |
5940 | */ | |
5941 | static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p) | |
5942 | { | |
5943 | struct cfs_rq *cfs_rq; | |
5944 | unsigned int load; | |
5945 | ||
5946 | /* Task has no contribution or is new */ | |
5947 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
5948 | return cpu_load(rq); | |
5949 | ||
5950 | cfs_rq = &rq->cfs; | |
5951 | load = READ_ONCE(cfs_rq->avg.load_avg); | |
5952 | ||
5953 | /* Discount task's util from CPU's util */ | |
5954 | lsub_positive(&load, task_h_load(p)); | |
5955 | ||
5956 | return load; | |
5957 | } | |
5958 | ||
9f683953 VG |
5959 | static unsigned long cpu_runnable(struct rq *rq) |
5960 | { | |
5961 | return cfs_rq_runnable_avg(&rq->cfs); | |
5962 | } | |
5963 | ||
070f5e86 VG |
5964 | static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p) |
5965 | { | |
5966 | struct cfs_rq *cfs_rq; | |
5967 | unsigned int runnable; | |
5968 | ||
5969 | /* Task has no contribution or is new */ | |
5970 | if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
5971 | return cpu_runnable(rq); | |
5972 | ||
5973 | cfs_rq = &rq->cfs; | |
5974 | runnable = READ_ONCE(cfs_rq->avg.runnable_avg); | |
5975 | ||
5976 | /* Discount task's runnable from CPU's runnable */ | |
5977 | lsub_positive(&runnable, p->se.avg.runnable_avg); | |
5978 | ||
5979 | return runnable; | |
5980 | } | |
5981 | ||
ced549fa | 5982 | static unsigned long capacity_of(int cpu) |
029632fb | 5983 | { |
ced549fa | 5984 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
5985 | } |
5986 | ||
c58d25f3 PZ |
5987 | static void record_wakee(struct task_struct *p) |
5988 | { | |
5989 | /* | |
5990 | * Only decay a single time; tasks that have less then 1 wakeup per | |
5991 | * jiffy will not have built up many flips. | |
5992 | */ | |
5993 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
5994 | current->wakee_flips >>= 1; | |
5995 | current->wakee_flip_decay_ts = jiffies; | |
5996 | } | |
5997 | ||
5998 | if (current->last_wakee != p) { | |
5999 | current->last_wakee = p; | |
6000 | current->wakee_flips++; | |
6001 | } | |
6002 | } | |
6003 | ||
63b0e9ed MG |
6004 | /* |
6005 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 6006 | * |
63b0e9ed | 6007 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
6008 | * at a frequency roughly N times higher than one of its wakees. |
6009 | * | |
6010 | * In order to determine whether we should let the load spread vs consolidating | |
6011 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
6012 | * partner, and a factor of lls_size higher frequency in the other. | |
6013 | * | |
6014 | * With both conditions met, we can be relatively sure that the relationship is | |
6015 | * non-monogamous, with partner count exceeding socket size. | |
6016 | * | |
6017 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
6018 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
6019 | * socket size. | |
63b0e9ed | 6020 | */ |
62470419 MW |
6021 | static int wake_wide(struct task_struct *p) |
6022 | { | |
63b0e9ed MG |
6023 | unsigned int master = current->wakee_flips; |
6024 | unsigned int slave = p->wakee_flips; | |
17c891ab | 6025 | int factor = __this_cpu_read(sd_llc_size); |
62470419 | 6026 | |
63b0e9ed MG |
6027 | if (master < slave) |
6028 | swap(master, slave); | |
6029 | if (slave < factor || master < slave * factor) | |
6030 | return 0; | |
6031 | return 1; | |
62470419 MW |
6032 | } |
6033 | ||
90001d67 | 6034 | /* |
d153b153 PZ |
6035 | * The purpose of wake_affine() is to quickly determine on which CPU we can run |
6036 | * soonest. For the purpose of speed we only consider the waking and previous | |
6037 | * CPU. | |
90001d67 | 6038 | * |
7332dec0 MG |
6039 | * wake_affine_idle() - only considers 'now', it check if the waking CPU is |
6040 | * cache-affine and is (or will be) idle. | |
f2cdd9cc PZ |
6041 | * |
6042 | * wake_affine_weight() - considers the weight to reflect the average | |
6043 | * scheduling latency of the CPUs. This seems to work | |
6044 | * for the overloaded case. | |
90001d67 | 6045 | */ |
3b76c4a3 | 6046 | static int |
89a55f56 | 6047 | wake_affine_idle(int this_cpu, int prev_cpu, int sync) |
90001d67 | 6048 | { |
7332dec0 MG |
6049 | /* |
6050 | * If this_cpu is idle, it implies the wakeup is from interrupt | |
6051 | * context. Only allow the move if cache is shared. Otherwise an | |
6052 | * interrupt intensive workload could force all tasks onto one | |
6053 | * node depending on the IO topology or IRQ affinity settings. | |
806486c3 MG |
6054 | * |
6055 | * If the prev_cpu is idle and cache affine then avoid a migration. | |
6056 | * There is no guarantee that the cache hot data from an interrupt | |
6057 | * is more important than cache hot data on the prev_cpu and from | |
6058 | * a cpufreq perspective, it's better to have higher utilisation | |
6059 | * on one CPU. | |
7332dec0 | 6060 | */ |
943d355d RJ |
6061 | if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu)) |
6062 | return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu; | |
90001d67 | 6063 | |
d153b153 | 6064 | if (sync && cpu_rq(this_cpu)->nr_running == 1) |
3b76c4a3 | 6065 | return this_cpu; |
90001d67 | 6066 | |
d8fcb81f JL |
6067 | if (available_idle_cpu(prev_cpu)) |
6068 | return prev_cpu; | |
6069 | ||
3b76c4a3 | 6070 | return nr_cpumask_bits; |
90001d67 PZ |
6071 | } |
6072 | ||
3b76c4a3 | 6073 | static int |
f2cdd9cc PZ |
6074 | wake_affine_weight(struct sched_domain *sd, struct task_struct *p, |
6075 | int this_cpu, int prev_cpu, int sync) | |
90001d67 | 6076 | { |
90001d67 PZ |
6077 | s64 this_eff_load, prev_eff_load; |
6078 | unsigned long task_load; | |
6079 | ||
11f10e54 | 6080 | this_eff_load = cpu_load(cpu_rq(this_cpu)); |
90001d67 | 6081 | |
90001d67 PZ |
6082 | if (sync) { |
6083 | unsigned long current_load = task_h_load(current); | |
6084 | ||
f2cdd9cc | 6085 | if (current_load > this_eff_load) |
3b76c4a3 | 6086 | return this_cpu; |
90001d67 | 6087 | |
f2cdd9cc | 6088 | this_eff_load -= current_load; |
90001d67 PZ |
6089 | } |
6090 | ||
90001d67 PZ |
6091 | task_load = task_h_load(p); |
6092 | ||
f2cdd9cc PZ |
6093 | this_eff_load += task_load; |
6094 | if (sched_feat(WA_BIAS)) | |
6095 | this_eff_load *= 100; | |
6096 | this_eff_load *= capacity_of(prev_cpu); | |
90001d67 | 6097 | |
11f10e54 | 6098 | prev_eff_load = cpu_load(cpu_rq(prev_cpu)); |
f2cdd9cc PZ |
6099 | prev_eff_load -= task_load; |
6100 | if (sched_feat(WA_BIAS)) | |
6101 | prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; | |
6102 | prev_eff_load *= capacity_of(this_cpu); | |
90001d67 | 6103 | |
082f764a MG |
6104 | /* |
6105 | * If sync, adjust the weight of prev_eff_load such that if | |
6106 | * prev_eff == this_eff that select_idle_sibling() will consider | |
6107 | * stacking the wakee on top of the waker if no other CPU is | |
6108 | * idle. | |
6109 | */ | |
6110 | if (sync) | |
6111 | prev_eff_load += 1; | |
6112 | ||
6113 | return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits; | |
90001d67 PZ |
6114 | } |
6115 | ||
772bd008 | 6116 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
7ebb66a1 | 6117 | int this_cpu, int prev_cpu, int sync) |
098fb9db | 6118 | { |
3b76c4a3 | 6119 | int target = nr_cpumask_bits; |
098fb9db | 6120 | |
89a55f56 | 6121 | if (sched_feat(WA_IDLE)) |
3b76c4a3 | 6122 | target = wake_affine_idle(this_cpu, prev_cpu, sync); |
90001d67 | 6123 | |
3b76c4a3 MG |
6124 | if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits) |
6125 | target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); | |
098fb9db | 6126 | |
ceeadb83 | 6127 | schedstat_inc(p->stats.nr_wakeups_affine_attempts); |
3b76c4a3 MG |
6128 | if (target == nr_cpumask_bits) |
6129 | return prev_cpu; | |
098fb9db | 6130 | |
3b76c4a3 | 6131 | schedstat_inc(sd->ttwu_move_affine); |
ceeadb83 | 6132 | schedstat_inc(p->stats.nr_wakeups_affine); |
3b76c4a3 | 6133 | return target; |
098fb9db IM |
6134 | } |
6135 | ||
aaee1203 | 6136 | static struct sched_group * |
45da2773 | 6137 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu); |
aaee1203 PZ |
6138 | |
6139 | /* | |
97fb7a0a | 6140 | * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group. |
aaee1203 PZ |
6141 | */ |
6142 | static int | |
18bd1b4b | 6143 | find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
aaee1203 PZ |
6144 | { |
6145 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
6146 | unsigned int min_exit_latency = UINT_MAX; |
6147 | u64 latest_idle_timestamp = 0; | |
6148 | int least_loaded_cpu = this_cpu; | |
17346452 | 6149 | int shallowest_idle_cpu = -1; |
aaee1203 PZ |
6150 | int i; |
6151 | ||
eaecf41f MR |
6152 | /* Check if we have any choice: */ |
6153 | if (group->group_weight == 1) | |
ae4df9d6 | 6154 | return cpumask_first(sched_group_span(group)); |
eaecf41f | 6155 | |
aaee1203 | 6156 | /* Traverse only the allowed CPUs */ |
3bd37062 | 6157 | for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { |
97886d9d AL |
6158 | struct rq *rq = cpu_rq(i); |
6159 | ||
6160 | if (!sched_core_cookie_match(rq, p)) | |
6161 | continue; | |
6162 | ||
17346452 VK |
6163 | if (sched_idle_cpu(i)) |
6164 | return i; | |
6165 | ||
943d355d | 6166 | if (available_idle_cpu(i)) { |
83a0a96a NP |
6167 | struct cpuidle_state *idle = idle_get_state(rq); |
6168 | if (idle && idle->exit_latency < min_exit_latency) { | |
6169 | /* | |
6170 | * We give priority to a CPU whose idle state | |
6171 | * has the smallest exit latency irrespective | |
6172 | * of any idle timestamp. | |
6173 | */ | |
6174 | min_exit_latency = idle->exit_latency; | |
6175 | latest_idle_timestamp = rq->idle_stamp; | |
6176 | shallowest_idle_cpu = i; | |
6177 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
6178 | rq->idle_stamp > latest_idle_timestamp) { | |
6179 | /* | |
6180 | * If equal or no active idle state, then | |
6181 | * the most recently idled CPU might have | |
6182 | * a warmer cache. | |
6183 | */ | |
6184 | latest_idle_timestamp = rq->idle_stamp; | |
6185 | shallowest_idle_cpu = i; | |
6186 | } | |
17346452 | 6187 | } else if (shallowest_idle_cpu == -1) { |
11f10e54 | 6188 | load = cpu_load(cpu_rq(i)); |
18cec7e0 | 6189 | if (load < min_load) { |
83a0a96a NP |
6190 | min_load = load; |
6191 | least_loaded_cpu = i; | |
6192 | } | |
e7693a36 GH |
6193 | } |
6194 | } | |
6195 | ||
17346452 | 6196 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 6197 | } |
e7693a36 | 6198 | |
18bd1b4b BJ |
6199 | static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p, |
6200 | int cpu, int prev_cpu, int sd_flag) | |
6201 | { | |
93f50f90 | 6202 | int new_cpu = cpu; |
18bd1b4b | 6203 | |
3bd37062 | 6204 | if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) |
6fee85cc BJ |
6205 | return prev_cpu; |
6206 | ||
c976a862 | 6207 | /* |
57abff06 | 6208 | * We need task's util for cpu_util_without, sync it up to |
c469933e | 6209 | * prev_cpu's last_update_time. |
c976a862 VK |
6210 | */ |
6211 | if (!(sd_flag & SD_BALANCE_FORK)) | |
6212 | sync_entity_load_avg(&p->se); | |
6213 | ||
18bd1b4b BJ |
6214 | while (sd) { |
6215 | struct sched_group *group; | |
6216 | struct sched_domain *tmp; | |
6217 | int weight; | |
6218 | ||
6219 | if (!(sd->flags & sd_flag)) { | |
6220 | sd = sd->child; | |
6221 | continue; | |
6222 | } | |
6223 | ||
45da2773 | 6224 | group = find_idlest_group(sd, p, cpu); |
18bd1b4b BJ |
6225 | if (!group) { |
6226 | sd = sd->child; | |
6227 | continue; | |
6228 | } | |
6229 | ||
6230 | new_cpu = find_idlest_group_cpu(group, p, cpu); | |
e90381ea | 6231 | if (new_cpu == cpu) { |
97fb7a0a | 6232 | /* Now try balancing at a lower domain level of 'cpu': */ |
18bd1b4b BJ |
6233 | sd = sd->child; |
6234 | continue; | |
6235 | } | |
6236 | ||
97fb7a0a | 6237 | /* Now try balancing at a lower domain level of 'new_cpu': */ |
18bd1b4b BJ |
6238 | cpu = new_cpu; |
6239 | weight = sd->span_weight; | |
6240 | sd = NULL; | |
6241 | for_each_domain(cpu, tmp) { | |
6242 | if (weight <= tmp->span_weight) | |
6243 | break; | |
6244 | if (tmp->flags & sd_flag) | |
6245 | sd = tmp; | |
6246 | } | |
18bd1b4b BJ |
6247 | } |
6248 | ||
6249 | return new_cpu; | |
6250 | } | |
6251 | ||
97886d9d | 6252 | static inline int __select_idle_cpu(int cpu, struct task_struct *p) |
9fe1f127 | 6253 | { |
97886d9d AL |
6254 | if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) && |
6255 | sched_cpu_cookie_match(cpu_rq(cpu), p)) | |
9fe1f127 MG |
6256 | return cpu; |
6257 | ||
6258 | return -1; | |
6259 | } | |
6260 | ||
10e2f1ac | 6261 | #ifdef CONFIG_SCHED_SMT |
ba2591a5 | 6262 | DEFINE_STATIC_KEY_FALSE(sched_smt_present); |
b284909a | 6263 | EXPORT_SYMBOL_GPL(sched_smt_present); |
10e2f1ac PZ |
6264 | |
6265 | static inline void set_idle_cores(int cpu, int val) | |
6266 | { | |
6267 | struct sched_domain_shared *sds; | |
6268 | ||
6269 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
6270 | if (sds) | |
6271 | WRITE_ONCE(sds->has_idle_cores, val); | |
6272 | } | |
6273 | ||
6274 | static inline bool test_idle_cores(int cpu, bool def) | |
6275 | { | |
6276 | struct sched_domain_shared *sds; | |
6277 | ||
c722f35b RR |
6278 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
6279 | if (sds) | |
6280 | return READ_ONCE(sds->has_idle_cores); | |
10e2f1ac PZ |
6281 | |
6282 | return def; | |
6283 | } | |
6284 | ||
6285 | /* | |
6286 | * Scans the local SMT mask to see if the entire core is idle, and records this | |
6287 | * information in sd_llc_shared->has_idle_cores. | |
6288 | * | |
6289 | * Since SMT siblings share all cache levels, inspecting this limited remote | |
6290 | * state should be fairly cheap. | |
6291 | */ | |
1b568f0a | 6292 | void __update_idle_core(struct rq *rq) |
10e2f1ac PZ |
6293 | { |
6294 | int core = cpu_of(rq); | |
6295 | int cpu; | |
6296 | ||
6297 | rcu_read_lock(); | |
6298 | if (test_idle_cores(core, true)) | |
6299 | goto unlock; | |
6300 | ||
6301 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
6302 | if (cpu == core) | |
6303 | continue; | |
6304 | ||
943d355d | 6305 | if (!available_idle_cpu(cpu)) |
10e2f1ac PZ |
6306 | goto unlock; |
6307 | } | |
6308 | ||
6309 | set_idle_cores(core, 1); | |
6310 | unlock: | |
6311 | rcu_read_unlock(); | |
6312 | } | |
6313 | ||
6314 | /* | |
6315 | * Scan the entire LLC domain for idle cores; this dynamically switches off if | |
6316 | * there are no idle cores left in the system; tracked through | |
6317 | * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. | |
6318 | */ | |
9fe1f127 | 6319 | static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) |
10e2f1ac | 6320 | { |
9fe1f127 MG |
6321 | bool idle = true; |
6322 | int cpu; | |
10e2f1ac | 6323 | |
1b568f0a | 6324 | if (!static_branch_likely(&sched_smt_present)) |
97886d9d | 6325 | return __select_idle_cpu(core, p); |
10e2f1ac | 6326 | |
9fe1f127 MG |
6327 | for_each_cpu(cpu, cpu_smt_mask(core)) { |
6328 | if (!available_idle_cpu(cpu)) { | |
6329 | idle = false; | |
6330 | if (*idle_cpu == -1) { | |
6331 | if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) { | |
6332 | *idle_cpu = cpu; | |
6333 | break; | |
6334 | } | |
6335 | continue; | |
bec2860a | 6336 | } |
9fe1f127 | 6337 | break; |
10e2f1ac | 6338 | } |
9fe1f127 MG |
6339 | if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr)) |
6340 | *idle_cpu = cpu; | |
10e2f1ac PZ |
6341 | } |
6342 | ||
9fe1f127 MG |
6343 | if (idle) |
6344 | return core; | |
10e2f1ac | 6345 | |
9fe1f127 | 6346 | cpumask_andnot(cpus, cpus, cpu_smt_mask(core)); |
10e2f1ac PZ |
6347 | return -1; |
6348 | } | |
6349 | ||
c722f35b RR |
6350 | /* |
6351 | * Scan the local SMT mask for idle CPUs. | |
6352 | */ | |
6353 | static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) | |
6354 | { | |
6355 | int cpu; | |
6356 | ||
6357 | for_each_cpu(cpu, cpu_smt_mask(target)) { | |
6358 | if (!cpumask_test_cpu(cpu, p->cpus_ptr) || | |
6359 | !cpumask_test_cpu(cpu, sched_domain_span(sd))) | |
6360 | continue; | |
6361 | if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) | |
6362 | return cpu; | |
6363 | } | |
6364 | ||
6365 | return -1; | |
6366 | } | |
6367 | ||
10e2f1ac PZ |
6368 | #else /* CONFIG_SCHED_SMT */ |
6369 | ||
9fe1f127 | 6370 | static inline void set_idle_cores(int cpu, int val) |
10e2f1ac | 6371 | { |
9fe1f127 MG |
6372 | } |
6373 | ||
6374 | static inline bool test_idle_cores(int cpu, bool def) | |
6375 | { | |
6376 | return def; | |
6377 | } | |
6378 | ||
6379 | static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) | |
6380 | { | |
97886d9d | 6381 | return __select_idle_cpu(core, p); |
10e2f1ac PZ |
6382 | } |
6383 | ||
c722f35b RR |
6384 | static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) |
6385 | { | |
6386 | return -1; | |
6387 | } | |
6388 | ||
10e2f1ac PZ |
6389 | #endif /* CONFIG_SCHED_SMT */ |
6390 | ||
6391 | /* | |
6392 | * Scan the LLC domain for idle CPUs; this is dynamically regulated by | |
6393 | * comparing the average scan cost (tracked in sd->avg_scan_cost) against the | |
6394 | * average idle time for this rq (as found in rq->avg_idle). | |
a50bde51 | 6395 | */ |
c722f35b | 6396 | static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target) |
10e2f1ac | 6397 | { |
ec4fc801 | 6398 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
9fe1f127 | 6399 | int i, cpu, idle_cpu = -1, nr = INT_MAX; |
70fb5ccf | 6400 | struct sched_domain_shared *sd_share; |
94aafc3e | 6401 | struct rq *this_rq = this_rq(); |
9fe1f127 | 6402 | int this = smp_processor_id(); |
9cfb38a7 | 6403 | struct sched_domain *this_sd; |
94aafc3e | 6404 | u64 time = 0; |
10e2f1ac | 6405 | |
9cfb38a7 WL |
6406 | this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); |
6407 | if (!this_sd) | |
6408 | return -1; | |
6409 | ||
bae4ec13 MG |
6410 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
6411 | ||
c722f35b | 6412 | if (sched_feat(SIS_PROP) && !has_idle_core) { |
e6e0dc2d | 6413 | u64 avg_cost, avg_idle, span_avg; |
94aafc3e | 6414 | unsigned long now = jiffies; |
1ad3aaf3 | 6415 | |
e6e0dc2d | 6416 | /* |
94aafc3e PZ |
6417 | * If we're busy, the assumption that the last idle period |
6418 | * predicts the future is flawed; age away the remaining | |
6419 | * predicted idle time. | |
e6e0dc2d | 6420 | */ |
94aafc3e PZ |
6421 | if (unlikely(this_rq->wake_stamp < now)) { |
6422 | while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) { | |
6423 | this_rq->wake_stamp++; | |
6424 | this_rq->wake_avg_idle >>= 1; | |
6425 | } | |
6426 | } | |
6427 | ||
6428 | avg_idle = this_rq->wake_avg_idle; | |
e6e0dc2d | 6429 | avg_cost = this_sd->avg_scan_cost + 1; |
10e2f1ac | 6430 | |
e6e0dc2d | 6431 | span_avg = sd->span_weight * avg_idle; |
1ad3aaf3 PZ |
6432 | if (span_avg > 4*avg_cost) |
6433 | nr = div_u64(span_avg, avg_cost); | |
6434 | else | |
6435 | nr = 4; | |
10e2f1ac | 6436 | |
bae4ec13 MG |
6437 | time = cpu_clock(this); |
6438 | } | |
60588bfa | 6439 | |
70fb5ccf CY |
6440 | if (sched_feat(SIS_UTIL)) { |
6441 | sd_share = rcu_dereference(per_cpu(sd_llc_shared, target)); | |
6442 | if (sd_share) { | |
6443 | /* because !--nr is the condition to stop scan */ | |
6444 | nr = READ_ONCE(sd_share->nr_idle_scan) + 1; | |
6445 | /* overloaded LLC is unlikely to have idle cpu/core */ | |
6446 | if (nr == 1) | |
6447 | return -1; | |
6448 | } | |
6449 | } | |
6450 | ||
56498cfb | 6451 | for_each_cpu_wrap(cpu, cpus, target + 1) { |
c722f35b | 6452 | if (has_idle_core) { |
9fe1f127 MG |
6453 | i = select_idle_core(p, cpu, cpus, &idle_cpu); |
6454 | if ((unsigned int)i < nr_cpumask_bits) | |
6455 | return i; | |
6456 | ||
6457 | } else { | |
6458 | if (!--nr) | |
6459 | return -1; | |
97886d9d | 6460 | idle_cpu = __select_idle_cpu(cpu, p); |
9fe1f127 MG |
6461 | if ((unsigned int)idle_cpu < nr_cpumask_bits) |
6462 | break; | |
6463 | } | |
10e2f1ac PZ |
6464 | } |
6465 | ||
c722f35b | 6466 | if (has_idle_core) |
02dbb724 | 6467 | set_idle_cores(target, false); |
9fe1f127 | 6468 | |
c722f35b | 6469 | if (sched_feat(SIS_PROP) && !has_idle_core) { |
bae4ec13 | 6470 | time = cpu_clock(this) - time; |
94aafc3e PZ |
6471 | |
6472 | /* | |
6473 | * Account for the scan cost of wakeups against the average | |
6474 | * idle time. | |
6475 | */ | |
6476 | this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time); | |
6477 | ||
bae4ec13 MG |
6478 | update_avg(&this_sd->avg_scan_cost, time); |
6479 | } | |
10e2f1ac | 6480 | |
9fe1f127 | 6481 | return idle_cpu; |
10e2f1ac PZ |
6482 | } |
6483 | ||
b7a33161 MR |
6484 | /* |
6485 | * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which | |
6486 | * the task fits. If no CPU is big enough, but there are idle ones, try to | |
6487 | * maximize capacity. | |
6488 | */ | |
6489 | static int | |
6490 | select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target) | |
6491 | { | |
b4c9c9f1 | 6492 | unsigned long task_util, best_cap = 0; |
b7a33161 MR |
6493 | int cpu, best_cpu = -1; |
6494 | struct cpumask *cpus; | |
6495 | ||
ec4fc801 | 6496 | cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
b7a33161 MR |
6497 | cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); |
6498 | ||
b4c9c9f1 VG |
6499 | task_util = uclamp_task_util(p); |
6500 | ||
b7a33161 MR |
6501 | for_each_cpu_wrap(cpu, cpus, target) { |
6502 | unsigned long cpu_cap = capacity_of(cpu); | |
6503 | ||
6504 | if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu)) | |
6505 | continue; | |
b4c9c9f1 | 6506 | if (fits_capacity(task_util, cpu_cap)) |
b7a33161 MR |
6507 | return cpu; |
6508 | ||
6509 | if (cpu_cap > best_cap) { | |
6510 | best_cap = cpu_cap; | |
6511 | best_cpu = cpu; | |
6512 | } | |
6513 | } | |
6514 | ||
6515 | return best_cpu; | |
6516 | } | |
6517 | ||
ef8df979 | 6518 | static inline bool asym_fits_capacity(unsigned long task_util, int cpu) |
b4c9c9f1 | 6519 | { |
740cf8a7 | 6520 | if (sched_asym_cpucap_active()) |
b4c9c9f1 VG |
6521 | return fits_capacity(task_util, capacity_of(cpu)); |
6522 | ||
6523 | return true; | |
6524 | } | |
6525 | ||
10e2f1ac PZ |
6526 | /* |
6527 | * Try and locate an idle core/thread in the LLC cache domain. | |
a50bde51 | 6528 | */ |
772bd008 | 6529 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 6530 | { |
c722f35b | 6531 | bool has_idle_core = false; |
99bd5e2f | 6532 | struct sched_domain *sd; |
b4c9c9f1 | 6533 | unsigned long task_util; |
32e839dd | 6534 | int i, recent_used_cpu; |
a50bde51 | 6535 | |
b7a33161 | 6536 | /* |
b4c9c9f1 VG |
6537 | * On asymmetric system, update task utilization because we will check |
6538 | * that the task fits with cpu's capacity. | |
b7a33161 | 6539 | */ |
740cf8a7 | 6540 | if (sched_asym_cpucap_active()) { |
b4c9c9f1 VG |
6541 | sync_entity_load_avg(&p->se); |
6542 | task_util = uclamp_task_util(p); | |
b7a33161 MR |
6543 | } |
6544 | ||
9099a147 | 6545 | /* |
ec4fc801 | 6546 | * per-cpu select_rq_mask usage |
9099a147 PZ |
6547 | */ |
6548 | lockdep_assert_irqs_disabled(); | |
6549 | ||
b4c9c9f1 VG |
6550 | if ((available_idle_cpu(target) || sched_idle_cpu(target)) && |
6551 | asym_fits_capacity(task_util, target)) | |
e0a79f52 | 6552 | return target; |
99bd5e2f SS |
6553 | |
6554 | /* | |
97fb7a0a | 6555 | * If the previous CPU is cache affine and idle, don't be stupid: |
99bd5e2f | 6556 | */ |
3c29e651 | 6557 | if (prev != target && cpus_share_cache(prev, target) && |
b4c9c9f1 VG |
6558 | (available_idle_cpu(prev) || sched_idle_cpu(prev)) && |
6559 | asym_fits_capacity(task_util, prev)) | |
772bd008 | 6560 | return prev; |
a50bde51 | 6561 | |
52262ee5 MG |
6562 | /* |
6563 | * Allow a per-cpu kthread to stack with the wakee if the | |
6564 | * kworker thread and the tasks previous CPUs are the same. | |
6565 | * The assumption is that the wakee queued work for the | |
6566 | * per-cpu kthread that is now complete and the wakeup is | |
6567 | * essentially a sync wakeup. An obvious example of this | |
6568 | * pattern is IO completions. | |
6569 | */ | |
6570 | if (is_per_cpu_kthread(current) && | |
8b4e74cc | 6571 | in_task() && |
52262ee5 | 6572 | prev == smp_processor_id() && |
014ba44e VD |
6573 | this_rq()->nr_running <= 1 && |
6574 | asym_fits_capacity(task_util, prev)) { | |
52262ee5 MG |
6575 | return prev; |
6576 | } | |
6577 | ||
97fb7a0a | 6578 | /* Check a recently used CPU as a potential idle candidate: */ |
32e839dd | 6579 | recent_used_cpu = p->recent_used_cpu; |
89aafd67 | 6580 | p->recent_used_cpu = prev; |
32e839dd MG |
6581 | if (recent_used_cpu != prev && |
6582 | recent_used_cpu != target && | |
6583 | cpus_share_cache(recent_used_cpu, target) && | |
3c29e651 | 6584 | (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) && |
b4c9c9f1 VG |
6585 | cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) && |
6586 | asym_fits_capacity(task_util, recent_used_cpu)) { | |
32e839dd MG |
6587 | return recent_used_cpu; |
6588 | } | |
6589 | ||
b4c9c9f1 VG |
6590 | /* |
6591 | * For asymmetric CPU capacity systems, our domain of interest is | |
6592 | * sd_asym_cpucapacity rather than sd_llc. | |
6593 | */ | |
740cf8a7 | 6594 | if (sched_asym_cpucap_active()) { |
b4c9c9f1 VG |
6595 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target)); |
6596 | /* | |
6597 | * On an asymmetric CPU capacity system where an exclusive | |
6598 | * cpuset defines a symmetric island (i.e. one unique | |
6599 | * capacity_orig value through the cpuset), the key will be set | |
6600 | * but the CPUs within that cpuset will not have a domain with | |
6601 | * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric | |
6602 | * capacity path. | |
6603 | */ | |
6604 | if (sd) { | |
6605 | i = select_idle_capacity(p, sd, target); | |
6606 | return ((unsigned)i < nr_cpumask_bits) ? i : target; | |
6607 | } | |
6608 | } | |
6609 | ||
518cd623 | 6610 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
10e2f1ac PZ |
6611 | if (!sd) |
6612 | return target; | |
772bd008 | 6613 | |
c722f35b RR |
6614 | if (sched_smt_active()) { |
6615 | has_idle_core = test_idle_cores(target, false); | |
6616 | ||
6617 | if (!has_idle_core && cpus_share_cache(prev, target)) { | |
6618 | i = select_idle_smt(p, sd, prev); | |
6619 | if ((unsigned int)i < nr_cpumask_bits) | |
6620 | return i; | |
6621 | } | |
6622 | } | |
6623 | ||
6624 | i = select_idle_cpu(p, sd, has_idle_core, target); | |
10e2f1ac PZ |
6625 | if ((unsigned)i < nr_cpumask_bits) |
6626 | return i; | |
6627 | ||
a50bde51 PZ |
6628 | return target; |
6629 | } | |
231678b7 | 6630 | |
104cb16d | 6631 | /* |
4e3c7d33 DE |
6632 | * Predicts what cpu_util(@cpu) would return if @p was removed from @cpu |
6633 | * (@dst_cpu = -1) or migrated to @dst_cpu. | |
390031e4 QP |
6634 | */ |
6635 | static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) | |
6636 | { | |
6637 | struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; | |
4e3c7d33 | 6638 | unsigned long util = READ_ONCE(cfs_rq->avg.util_avg); |
390031e4 QP |
6639 | |
6640 | /* | |
4e3c7d33 DE |
6641 | * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its |
6642 | * contribution. If @p migrates from another CPU to @cpu add its | |
6643 | * contribution. In all the other cases @cpu is not impacted by the | |
6644 | * migration so its util_avg is already correct. | |
390031e4 QP |
6645 | */ |
6646 | if (task_cpu(p) == cpu && dst_cpu != cpu) | |
736cc6b3 | 6647 | lsub_positive(&util, task_util(p)); |
390031e4 QP |
6648 | else if (task_cpu(p) != cpu && dst_cpu == cpu) |
6649 | util += task_util(p); | |
6650 | ||
6651 | if (sched_feat(UTIL_EST)) { | |
4e3c7d33 DE |
6652 | unsigned long util_est; |
6653 | ||
390031e4 QP |
6654 | util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued); |
6655 | ||
6656 | /* | |
4e3c7d33 DE |
6657 | * During wake-up @p isn't enqueued yet and doesn't contribute |
6658 | * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued. | |
6659 | * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p | |
6660 | * has been enqueued. | |
6661 | * | |
6662 | * During exec (@dst_cpu = -1) @p is enqueued and does | |
6663 | * contribute to cpu_rq(cpu)->cfs.util_est.enqueued. | |
6664 | * Remove it to "simulate" cpu_util without @p's contribution. | |
6665 | * | |
6666 | * Despite the task_on_rq_queued(@p) check there is still a | |
6667 | * small window for a possible race when an exec | |
6668 | * select_task_rq_fair() races with LB's detach_task(). | |
6669 | * | |
6670 | * detach_task() | |
6671 | * deactivate_task() | |
6672 | * p->on_rq = TASK_ON_RQ_MIGRATING; | |
6673 | * -------------------------------- A | |
6674 | * dequeue_task() \ | |
6675 | * dequeue_task_fair() + Race Time | |
6676 | * util_est_dequeue() / | |
6677 | * -------------------------------- B | |
6678 | * | |
6679 | * The additional check "current == p" is required to further | |
6680 | * reduce the race window. | |
390031e4 QP |
6681 | */ |
6682 | if (dst_cpu == cpu) | |
6683 | util_est += _task_util_est(p); | |
4e3c7d33 DE |
6684 | else if (unlikely(task_on_rq_queued(p) || current == p)) |
6685 | lsub_positive(&util_est, _task_util_est(p)); | |
390031e4 QP |
6686 | |
6687 | util = max(util, util_est); | |
6688 | } | |
6689 | ||
6690 | return min(util, capacity_orig_of(cpu)); | |
6691 | } | |
6692 | ||
4e3c7d33 DE |
6693 | /* |
6694 | * cpu_util_without: compute cpu utilization without any contributions from *p | |
6695 | * @cpu: the CPU which utilization is requested | |
6696 | * @p: the task which utilization should be discounted | |
6697 | * | |
6698 | * The utilization of a CPU is defined by the utilization of tasks currently | |
6699 | * enqueued on that CPU as well as tasks which are currently sleeping after an | |
6700 | * execution on that CPU. | |
6701 | * | |
6702 | * This method returns the utilization of the specified CPU by discounting the | |
6703 | * utilization of the specified task, whenever the task is currently | |
6704 | * contributing to the CPU utilization. | |
6705 | */ | |
6706 | static unsigned long cpu_util_without(int cpu, struct task_struct *p) | |
6707 | { | |
6708 | /* Task has no contribution or is new */ | |
6709 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
6710 | return cpu_util_cfs(cpu); | |
6711 | ||
6712 | return cpu_util_next(cpu, p, -1); | |
6713 | } | |
6714 | ||
390031e4 | 6715 | /* |
3e8c6c9a VD |
6716 | * energy_env - Utilization landscape for energy estimation. |
6717 | * @task_busy_time: Utilization contribution by the task for which we test the | |
6718 | * placement. Given by eenv_task_busy_time(). | |
6719 | * @pd_busy_time: Utilization of the whole perf domain without the task | |
6720 | * contribution. Given by eenv_pd_busy_time(). | |
6721 | * @cpu_cap: Maximum CPU capacity for the perf domain. | |
6722 | * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap). | |
390031e4 | 6723 | */ |
3e8c6c9a VD |
6724 | struct energy_env { |
6725 | unsigned long task_busy_time; | |
6726 | unsigned long pd_busy_time; | |
6727 | unsigned long cpu_cap; | |
6728 | unsigned long pd_cap; | |
6729 | }; | |
6730 | ||
6731 | /* | |
6732 | * Compute the task busy time for compute_energy(). This time cannot be | |
6733 | * injected directly into effective_cpu_util() because of the IRQ scaling. | |
6734 | * The latter only makes sense with the most recent CPUs where the task has | |
6735 | * run. | |
6736 | */ | |
6737 | static inline void eenv_task_busy_time(struct energy_env *eenv, | |
6738 | struct task_struct *p, int prev_cpu) | |
390031e4 | 6739 | { |
3e8c6c9a VD |
6740 | unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu); |
6741 | unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu)); | |
6742 | ||
6743 | if (unlikely(irq >= max_cap)) | |
6744 | busy_time = max_cap; | |
6745 | else | |
6746 | busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap); | |
6747 | ||
6748 | eenv->task_busy_time = busy_time; | |
6749 | } | |
6750 | ||
6751 | /* | |
6752 | * Compute the perf_domain (PD) busy time for compute_energy(). Based on the | |
6753 | * utilization for each @pd_cpus, it however doesn't take into account | |
6754 | * clamping since the ratio (utilization / cpu_capacity) is already enough to | |
6755 | * scale the EM reported power consumption at the (eventually clamped) | |
6756 | * cpu_capacity. | |
6757 | * | |
6758 | * The contribution of the task @p for which we want to estimate the | |
6759 | * energy cost is removed (by cpu_util_next()) and must be calculated | |
6760 | * separately (see eenv_task_busy_time). This ensures: | |
6761 | * | |
6762 | * - A stable PD utilization, no matter which CPU of that PD we want to place | |
6763 | * the task on. | |
6764 | * | |
6765 | * - A fair comparison between CPUs as the task contribution (task_util()) | |
6766 | * will always be the same no matter which CPU utilization we rely on | |
6767 | * (util_avg or util_est). | |
6768 | * | |
6769 | * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't | |
6770 | * exceed @eenv->pd_cap. | |
6771 | */ | |
6772 | static inline void eenv_pd_busy_time(struct energy_env *eenv, | |
6773 | struct cpumask *pd_cpus, | |
6774 | struct task_struct *p) | |
6775 | { | |
6776 | unsigned long busy_time = 0; | |
390031e4 QP |
6777 | int cpu; |
6778 | ||
3e8c6c9a VD |
6779 | for_each_cpu(cpu, pd_cpus) { |
6780 | unsigned long util = cpu_util_next(cpu, p, -1); | |
489f1645 | 6781 | |
3e8c6c9a VD |
6782 | busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL); |
6783 | } | |
0372e1cf | 6784 | |
3e8c6c9a VD |
6785 | eenv->pd_busy_time = min(eenv->pd_cap, busy_time); |
6786 | } | |
af24bde8 | 6787 | |
3e8c6c9a VD |
6788 | /* |
6789 | * Compute the maximum utilization for compute_energy() when the task @p | |
6790 | * is placed on the cpu @dst_cpu. | |
6791 | * | |
6792 | * Returns the maximum utilization among @eenv->cpus. This utilization can't | |
6793 | * exceed @eenv->cpu_cap. | |
6794 | */ | |
6795 | static inline unsigned long | |
6796 | eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus, | |
6797 | struct task_struct *p, int dst_cpu) | |
6798 | { | |
6799 | unsigned long max_util = 0; | |
6800 | int cpu; | |
489f1645 | 6801 | |
3e8c6c9a VD |
6802 | for_each_cpu(cpu, pd_cpus) { |
6803 | struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL; | |
6804 | unsigned long util = cpu_util_next(cpu, p, dst_cpu); | |
6805 | unsigned long cpu_util; | |
af24bde8 | 6806 | |
390031e4 | 6807 | /* |
eb92692b QP |
6808 | * Performance domain frequency: utilization clamping |
6809 | * must be considered since it affects the selection | |
6810 | * of the performance domain frequency. | |
6811 | * NOTE: in case RT tasks are running, by default the | |
6812 | * FREQUENCY_UTIL's utilization can be max OPP. | |
390031e4 | 6813 | */ |
3e8c6c9a VD |
6814 | cpu_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk); |
6815 | max_util = max(max_util, cpu_util); | |
390031e4 QP |
6816 | } |
6817 | ||
3e8c6c9a VD |
6818 | return min(max_util, eenv->cpu_cap); |
6819 | } | |
6820 | ||
6821 | /* | |
6822 | * compute_energy(): Use the Energy Model to estimate the energy that @pd would | |
6823 | * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task | |
6824 | * contribution is ignored. | |
6825 | */ | |
6826 | static inline unsigned long | |
6827 | compute_energy(struct energy_env *eenv, struct perf_domain *pd, | |
6828 | struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu) | |
6829 | { | |
6830 | unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu); | |
6831 | unsigned long busy_time = eenv->pd_busy_time; | |
6832 | ||
6833 | if (dst_cpu >= 0) | |
6834 | busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time); | |
6835 | ||
6836 | return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap); | |
390031e4 QP |
6837 | } |
6838 | ||
732cd75b QP |
6839 | /* |
6840 | * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the | |
6841 | * waking task. find_energy_efficient_cpu() looks for the CPU with maximum | |
6842 | * spare capacity in each performance domain and uses it as a potential | |
6843 | * candidate to execute the task. Then, it uses the Energy Model to figure | |
6844 | * out which of the CPU candidates is the most energy-efficient. | |
6845 | * | |
6846 | * The rationale for this heuristic is as follows. In a performance domain, | |
6847 | * all the most energy efficient CPU candidates (according to the Energy | |
6848 | * Model) are those for which we'll request a low frequency. When there are | |
6849 | * several CPUs for which the frequency request will be the same, we don't | |
6850 | * have enough data to break the tie between them, because the Energy Model | |
6851 | * only includes active power costs. With this model, if we assume that | |
6852 | * frequency requests follow utilization (e.g. using schedutil), the CPU with | |
6853 | * the maximum spare capacity in a performance domain is guaranteed to be among | |
6854 | * the best candidates of the performance domain. | |
6855 | * | |
6856 | * In practice, it could be preferable from an energy standpoint to pack | |
6857 | * small tasks on a CPU in order to let other CPUs go in deeper idle states, | |
6858 | * but that could also hurt our chances to go cluster idle, and we have no | |
6859 | * ways to tell with the current Energy Model if this is actually a good | |
6860 | * idea or not. So, find_energy_efficient_cpu() basically favors | |
6861 | * cluster-packing, and spreading inside a cluster. That should at least be | |
6862 | * a good thing for latency, and this is consistent with the idea that most | |
6863 | * of the energy savings of EAS come from the asymmetry of the system, and | |
6864 | * not so much from breaking the tie between identical CPUs. That's also the | |
6865 | * reason why EAS is enabled in the topology code only for systems where | |
6866 | * SD_ASYM_CPUCAPACITY is set. | |
6867 | * | |
6868 | * NOTE: Forkees are not accepted in the energy-aware wake-up path because | |
6869 | * they don't have any useful utilization data yet and it's not possible to | |
6870 | * forecast their impact on energy consumption. Consequently, they will be | |
6871 | * placed by find_idlest_cpu() on the least loaded CPU, which might turn out | |
6872 | * to be energy-inefficient in some use-cases. The alternative would be to | |
6873 | * bias new tasks towards specific types of CPUs first, or to try to infer | |
6874 | * their util_avg from the parent task, but those heuristics could hurt | |
6875 | * other use-cases too. So, until someone finds a better way to solve this, | |
6876 | * let's keep things simple by re-using the existing slow path. | |
6877 | */ | |
732cd75b QP |
6878 | static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) |
6879 | { | |
9b340131 | 6880 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
eb92692b | 6881 | unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; |
3e8c6c9a | 6882 | struct root_domain *rd = this_rq()->rd; |
b812fc97 | 6883 | int cpu, best_energy_cpu, target = -1; |
732cd75b | 6884 | struct sched_domain *sd; |
eb92692b | 6885 | struct perf_domain *pd; |
3e8c6c9a | 6886 | struct energy_env eenv; |
732cd75b QP |
6887 | |
6888 | rcu_read_lock(); | |
6889 | pd = rcu_dereference(rd->pd); | |
6890 | if (!pd || READ_ONCE(rd->overutilized)) | |
619e090c | 6891 | goto unlock; |
732cd75b QP |
6892 | |
6893 | /* | |
6894 | * Energy-aware wake-up happens on the lowest sched_domain starting | |
6895 | * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. | |
6896 | */ | |
6897 | sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); | |
6898 | while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
6899 | sd = sd->parent; | |
6900 | if (!sd) | |
619e090c PG |
6901 | goto unlock; |
6902 | ||
6903 | target = prev_cpu; | |
732cd75b QP |
6904 | |
6905 | sync_entity_load_avg(&p->se); | |
6906 | if (!task_util_est(p)) | |
6907 | goto unlock; | |
6908 | ||
3e8c6c9a VD |
6909 | eenv_task_busy_time(&eenv, p, prev_cpu); |
6910 | ||
732cd75b | 6911 | for (; pd; pd = pd->next) { |
3e8c6c9a VD |
6912 | unsigned long cpu_cap, cpu_thermal_cap, util; |
6913 | unsigned long cur_delta, max_spare_cap = 0; | |
8d4c97c1 | 6914 | bool compute_prev_delta = false; |
732cd75b | 6915 | int max_spare_cap_cpu = -1; |
b812fc97 | 6916 | unsigned long base_energy; |
732cd75b | 6917 | |
9b340131 DE |
6918 | cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask); |
6919 | ||
3e8c6c9a VD |
6920 | if (cpumask_empty(cpus)) |
6921 | continue; | |
6922 | ||
6923 | /* Account thermal pressure for the energy estimation */ | |
6924 | cpu = cpumask_first(cpus); | |
6925 | cpu_thermal_cap = arch_scale_cpu_capacity(cpu); | |
6926 | cpu_thermal_cap -= arch_scale_thermal_pressure(cpu); | |
6927 | ||
6928 | eenv.cpu_cap = cpu_thermal_cap; | |
6929 | eenv.pd_cap = 0; | |
6930 | ||
6931 | for_each_cpu(cpu, cpus) { | |
6932 | eenv.pd_cap += cpu_thermal_cap; | |
6933 | ||
6934 | if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) | |
6935 | continue; | |
6936 | ||
3bd37062 | 6937 | if (!cpumask_test_cpu(cpu, p->cpus_ptr)) |
732cd75b QP |
6938 | continue; |
6939 | ||
732cd75b QP |
6940 | util = cpu_util_next(cpu, p, cpu); |
6941 | cpu_cap = capacity_of(cpu); | |
1d42509e VS |
6942 | |
6943 | /* | |
6944 | * Skip CPUs that cannot satisfy the capacity request. | |
6945 | * IOW, placing the task there would make the CPU | |
6946 | * overutilized. Take uclamp into account to see how | |
6947 | * much capacity we can get out of the CPU; this is | |
a5418be9 | 6948 | * aligned with sched_cpu_util(). |
1d42509e VS |
6949 | */ |
6950 | util = uclamp_rq_util_with(cpu_rq(cpu), util, p); | |
60e17f5c | 6951 | if (!fits_capacity(util, cpu_cap)) |
732cd75b QP |
6952 | continue; |
6953 | ||
3e8c6c9a VD |
6954 | lsub_positive(&cpu_cap, util); |
6955 | ||
732cd75b | 6956 | if (cpu == prev_cpu) { |
8d4c97c1 PG |
6957 | /* Always use prev_cpu as a candidate. */ |
6958 | compute_prev_delta = true; | |
3e8c6c9a | 6959 | } else if (cpu_cap > max_spare_cap) { |
8d4c97c1 PG |
6960 | /* |
6961 | * Find the CPU with the maximum spare capacity | |
6962 | * in the performance domain. | |
6963 | */ | |
3e8c6c9a | 6964 | max_spare_cap = cpu_cap; |
732cd75b QP |
6965 | max_spare_cap_cpu = cpu; |
6966 | } | |
6967 | } | |
6968 | ||
8d4c97c1 PG |
6969 | if (max_spare_cap_cpu < 0 && !compute_prev_delta) |
6970 | continue; | |
6971 | ||
3e8c6c9a | 6972 | eenv_pd_busy_time(&eenv, cpus, p); |
8d4c97c1 | 6973 | /* Compute the 'base' energy of the pd, without @p */ |
b812fc97 | 6974 | base_energy = compute_energy(&eenv, pd, cpus, p, -1); |
8d4c97c1 PG |
6975 | |
6976 | /* Evaluate the energy impact of using prev_cpu. */ | |
6977 | if (compute_prev_delta) { | |
3e8c6c9a VD |
6978 | prev_delta = compute_energy(&eenv, pd, cpus, p, |
6979 | prev_cpu); | |
6980 | /* CPU utilization has changed */ | |
b812fc97 | 6981 | if (prev_delta < base_energy) |
619e090c | 6982 | goto unlock; |
b812fc97 | 6983 | prev_delta -= base_energy; |
8d4c97c1 PG |
6984 | best_delta = min(best_delta, prev_delta); |
6985 | } | |
6986 | ||
6987 | /* Evaluate the energy impact of using max_spare_cap_cpu. */ | |
6988 | if (max_spare_cap_cpu >= 0) { | |
3e8c6c9a VD |
6989 | cur_delta = compute_energy(&eenv, pd, cpus, p, |
6990 | max_spare_cap_cpu); | |
6991 | /* CPU utilization has changed */ | |
b812fc97 | 6992 | if (cur_delta < base_energy) |
619e090c | 6993 | goto unlock; |
b812fc97 | 6994 | cur_delta -= base_energy; |
eb92692b QP |
6995 | if (cur_delta < best_delta) { |
6996 | best_delta = cur_delta; | |
732cd75b QP |
6997 | best_energy_cpu = max_spare_cap_cpu; |
6998 | } | |
6999 | } | |
7000 | } | |
732cd75b QP |
7001 | rcu_read_unlock(); |
7002 | ||
b812fc97 | 7003 | if (best_delta < prev_delta) |
619e090c | 7004 | target = best_energy_cpu; |
732cd75b | 7005 | |
619e090c | 7006 | return target; |
732cd75b | 7007 | |
619e090c | 7008 | unlock: |
732cd75b QP |
7009 | rcu_read_unlock(); |
7010 | ||
619e090c | 7011 | return target; |
732cd75b QP |
7012 | } |
7013 | ||
aaee1203 | 7014 | /* |
de91b9cb | 7015 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
3aef1551 | 7016 | * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE, |
de91b9cb | 7017 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. |
aaee1203 | 7018 | * |
97fb7a0a IM |
7019 | * Balances load by selecting the idlest CPU in the idlest group, or under |
7020 | * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 7021 | * |
97fb7a0a | 7022 | * Returns the target CPU number. |
aaee1203 | 7023 | */ |
0017d735 | 7024 | static int |
3aef1551 | 7025 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags) |
aaee1203 | 7026 | { |
3aef1551 | 7027 | int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); |
f1d88b44 | 7028 | struct sched_domain *tmp, *sd = NULL; |
c88d5910 | 7029 | int cpu = smp_processor_id(); |
63b0e9ed | 7030 | int new_cpu = prev_cpu; |
99bd5e2f | 7031 | int want_affine = 0; |
3aef1551 VS |
7032 | /* SD_flags and WF_flags share the first nibble */ |
7033 | int sd_flag = wake_flags & 0xF; | |
c88d5910 | 7034 | |
9099a147 PZ |
7035 | /* |
7036 | * required for stable ->cpus_allowed | |
7037 | */ | |
7038 | lockdep_assert_held(&p->pi_lock); | |
dc824eb8 | 7039 | if (wake_flags & WF_TTWU) { |
c58d25f3 | 7040 | record_wakee(p); |
732cd75b | 7041 | |
f8a696f2 | 7042 | if (sched_energy_enabled()) { |
732cd75b QP |
7043 | new_cpu = find_energy_efficient_cpu(p, prev_cpu); |
7044 | if (new_cpu >= 0) | |
7045 | return new_cpu; | |
7046 | new_cpu = prev_cpu; | |
7047 | } | |
7048 | ||
00061968 | 7049 | want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr); |
c58d25f3 | 7050 | } |
aaee1203 | 7051 | |
dce840a0 | 7052 | rcu_read_lock(); |
aaee1203 | 7053 | for_each_domain(cpu, tmp) { |
fe3bcfe1 | 7054 | /* |
97fb7a0a | 7055 | * If both 'cpu' and 'prev_cpu' are part of this domain, |
99bd5e2f | 7056 | * cpu is a valid SD_WAKE_AFFINE target. |
fe3bcfe1 | 7057 | */ |
99bd5e2f SS |
7058 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
7059 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
f1d88b44 VK |
7060 | if (cpu != prev_cpu) |
7061 | new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync); | |
7062 | ||
7063 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
29cd8bae | 7064 | break; |
f03542a7 | 7065 | } |
29cd8bae | 7066 | |
2917406c BS |
7067 | /* |
7068 | * Usually only true for WF_EXEC and WF_FORK, as sched_domains | |
7069 | * usually do not have SD_BALANCE_WAKE set. That means wakeup | |
7070 | * will usually go to the fast path. | |
7071 | */ | |
f03542a7 | 7072 | if (tmp->flags & sd_flag) |
29cd8bae | 7073 | sd = tmp; |
63b0e9ed MG |
7074 | else if (!want_affine) |
7075 | break; | |
29cd8bae PZ |
7076 | } |
7077 | ||
f1d88b44 VK |
7078 | if (unlikely(sd)) { |
7079 | /* Slow path */ | |
18bd1b4b | 7080 | new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag); |
dc824eb8 | 7081 | } else if (wake_flags & WF_TTWU) { /* XXX always ? */ |
f1d88b44 | 7082 | /* Fast path */ |
f1d88b44 | 7083 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); |
e7693a36 | 7084 | } |
dce840a0 | 7085 | rcu_read_unlock(); |
e7693a36 | 7086 | |
c88d5910 | 7087 | return new_cpu; |
e7693a36 | 7088 | } |
0a74bef8 PT |
7089 | |
7090 | /* | |
97fb7a0a | 7091 | * Called immediately before a task is migrated to a new CPU; task_cpu(p) and |
0a74bef8 | 7092 | * cfs_rq_of(p) references at time of call are still valid and identify the |
97fb7a0a | 7093 | * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 7094 | */ |
3f9672ba | 7095 | static void migrate_task_rq_fair(struct task_struct *p, int new_cpu) |
0a74bef8 | 7096 | { |
e2f3e35f VD |
7097 | struct sched_entity *se = &p->se; |
7098 | ||
59efa0ba PZ |
7099 | /* |
7100 | * As blocked tasks retain absolute vruntime the migration needs to | |
7101 | * deal with this by subtracting the old and adding the new | |
7102 | * min_vruntime -- the latter is done by enqueue_entity() when placing | |
7103 | * the task on the new runqueue. | |
7104 | */ | |
2f064a59 | 7105 | if (READ_ONCE(p->__state) == TASK_WAKING) { |
59efa0ba | 7106 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
59efa0ba | 7107 | |
d05b4305 | 7108 | se->vruntime -= u64_u32_load(cfs_rq->min_vruntime); |
59efa0ba PZ |
7109 | } |
7110 | ||
e1f078f5 | 7111 | if (!task_on_rq_migrating(p)) { |
e2f3e35f VD |
7112 | remove_entity_load_avg(se); |
7113 | ||
144d8487 | 7114 | /* |
e2f3e35f VD |
7115 | * Here, the task's PELT values have been updated according to |
7116 | * the current rq's clock. But if that clock hasn't been | |
7117 | * updated in a while, a substantial idle time will be missed, | |
7118 | * leading to an inflation after wake-up on the new rq. | |
7119 | * | |
7120 | * Estimate the missing time from the cfs_rq last_update_time | |
7121 | * and update sched_avg to improve the PELT continuity after | |
7122 | * migration. | |
144d8487 | 7123 | */ |
e2f3e35f | 7124 | migrate_se_pelt_lag(se); |
144d8487 | 7125 | } |
9d89c257 YD |
7126 | |
7127 | /* Tell new CPU we are migrated */ | |
e2f3e35f | 7128 | se->avg.last_update_time = 0; |
3944a927 BS |
7129 | |
7130 | /* We have migrated, no longer consider this task hot */ | |
e2f3e35f | 7131 | se->exec_start = 0; |
3f9672ba SD |
7132 | |
7133 | update_scan_period(p, new_cpu); | |
0a74bef8 | 7134 | } |
12695578 YD |
7135 | |
7136 | static void task_dead_fair(struct task_struct *p) | |
7137 | { | |
7138 | remove_entity_load_avg(&p->se); | |
7139 | } | |
6e2df058 PZ |
7140 | |
7141 | static int | |
7142 | balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) | |
7143 | { | |
7144 | if (rq->nr_running) | |
7145 | return 1; | |
7146 | ||
7147 | return newidle_balance(rq, rf) != 0; | |
7148 | } | |
e7693a36 GH |
7149 | #endif /* CONFIG_SMP */ |
7150 | ||
a555e9d8 | 7151 | static unsigned long wakeup_gran(struct sched_entity *se) |
0bbd3336 PZ |
7152 | { |
7153 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
7154 | ||
7155 | /* | |
e52fb7c0 PZ |
7156 | * Since its curr running now, convert the gran from real-time |
7157 | * to virtual-time in his units. | |
13814d42 MG |
7158 | * |
7159 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
7160 | * they get preempted easier. That is, if 'se' < 'curr' then | |
7161 | * the resulting gran will be larger, therefore penalizing the | |
7162 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
7163 | * be smaller, again penalizing the lighter task. | |
7164 | * | |
7165 | * This is especially important for buddies when the leftmost | |
7166 | * task is higher priority than the buddy. | |
0bbd3336 | 7167 | */ |
f4ad9bd2 | 7168 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
7169 | } |
7170 | ||
464b7527 PZ |
7171 | /* |
7172 | * Should 'se' preempt 'curr'. | |
7173 | * | |
7174 | * |s1 | |
7175 | * |s2 | |
7176 | * |s3 | |
7177 | * g | |
7178 | * |<--->|c | |
7179 | * | |
7180 | * w(c, s1) = -1 | |
7181 | * w(c, s2) = 0 | |
7182 | * w(c, s3) = 1 | |
7183 | * | |
7184 | */ | |
7185 | static int | |
7186 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
7187 | { | |
7188 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
7189 | ||
7190 | if (vdiff <= 0) | |
7191 | return -1; | |
7192 | ||
a555e9d8 | 7193 | gran = wakeup_gran(se); |
464b7527 PZ |
7194 | if (vdiff > gran) |
7195 | return 1; | |
7196 | ||
7197 | return 0; | |
7198 | } | |
7199 | ||
02479099 PZ |
7200 | static void set_last_buddy(struct sched_entity *se) |
7201 | { | |
c5ae366e DA |
7202 | for_each_sched_entity(se) { |
7203 | if (SCHED_WARN_ON(!se->on_rq)) | |
7204 | return; | |
30400039 JD |
7205 | if (se_is_idle(se)) |
7206 | return; | |
69c80f3e | 7207 | cfs_rq_of(se)->last = se; |
c5ae366e | 7208 | } |
02479099 PZ |
7209 | } |
7210 | ||
7211 | static void set_next_buddy(struct sched_entity *se) | |
7212 | { | |
c5ae366e DA |
7213 | for_each_sched_entity(se) { |
7214 | if (SCHED_WARN_ON(!se->on_rq)) | |
7215 | return; | |
30400039 JD |
7216 | if (se_is_idle(se)) |
7217 | return; | |
69c80f3e | 7218 | cfs_rq_of(se)->next = se; |
c5ae366e | 7219 | } |
02479099 PZ |
7220 | } |
7221 | ||
ac53db59 RR |
7222 | static void set_skip_buddy(struct sched_entity *se) |
7223 | { | |
69c80f3e VP |
7224 | for_each_sched_entity(se) |
7225 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
7226 | } |
7227 | ||
bf0f6f24 IM |
7228 | /* |
7229 | * Preempt the current task with a newly woken task if needed: | |
7230 | */ | |
5a9b86f6 | 7231 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
7232 | { |
7233 | struct task_struct *curr = rq->curr; | |
8651a86c | 7234 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 7235 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 7236 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 7237 | int next_buddy_marked = 0; |
30400039 | 7238 | int cse_is_idle, pse_is_idle; |
bf0f6f24 | 7239 | |
4ae7d5ce IM |
7240 | if (unlikely(se == pse)) |
7241 | return; | |
7242 | ||
5238cdd3 | 7243 | /* |
163122b7 | 7244 | * This is possible from callers such as attach_tasks(), in which we |
3b03706f | 7245 | * unconditionally check_preempt_curr() after an enqueue (which may have |
5238cdd3 PT |
7246 | * lead to a throttle). This both saves work and prevents false |
7247 | * next-buddy nomination below. | |
7248 | */ | |
7249 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
7250 | return; | |
7251 | ||
2f36825b | 7252 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 7253 | set_next_buddy(pse); |
2f36825b VP |
7254 | next_buddy_marked = 1; |
7255 | } | |
57fdc26d | 7256 | |
aec0a514 BR |
7257 | /* |
7258 | * We can come here with TIF_NEED_RESCHED already set from new task | |
7259 | * wake up path. | |
5238cdd3 PT |
7260 | * |
7261 | * Note: this also catches the edge-case of curr being in a throttled | |
7262 | * group (e.g. via set_curr_task), since update_curr() (in the | |
7263 | * enqueue of curr) will have resulted in resched being set. This | |
7264 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
7265 | * below. | |
aec0a514 BR |
7266 | */ |
7267 | if (test_tsk_need_resched(curr)) | |
7268 | return; | |
7269 | ||
a2f5c9ab | 7270 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
1da1843f VK |
7271 | if (unlikely(task_has_idle_policy(curr)) && |
7272 | likely(!task_has_idle_policy(p))) | |
a2f5c9ab DH |
7273 | goto preempt; |
7274 | ||
91c234b4 | 7275 | /* |
a2f5c9ab DH |
7276 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
7277 | * is driven by the tick): | |
91c234b4 | 7278 | */ |
8ed92e51 | 7279 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 7280 | return; |
bf0f6f24 | 7281 | |
464b7527 | 7282 | find_matching_se(&se, &pse); |
002f128b | 7283 | BUG_ON(!pse); |
30400039 JD |
7284 | |
7285 | cse_is_idle = se_is_idle(se); | |
7286 | pse_is_idle = se_is_idle(pse); | |
7287 | ||
7288 | /* | |
7289 | * Preempt an idle group in favor of a non-idle group (and don't preempt | |
7290 | * in the inverse case). | |
7291 | */ | |
7292 | if (cse_is_idle && !pse_is_idle) | |
7293 | goto preempt; | |
7294 | if (cse_is_idle != pse_is_idle) | |
7295 | return; | |
7296 | ||
7297 | update_curr(cfs_rq_of(se)); | |
2f36825b VP |
7298 | if (wakeup_preempt_entity(se, pse) == 1) { |
7299 | /* | |
7300 | * Bias pick_next to pick the sched entity that is | |
7301 | * triggering this preemption. | |
7302 | */ | |
7303 | if (!next_buddy_marked) | |
7304 | set_next_buddy(pse); | |
3a7e73a2 | 7305 | goto preempt; |
2f36825b | 7306 | } |
464b7527 | 7307 | |
3a7e73a2 | 7308 | return; |
a65ac745 | 7309 | |
3a7e73a2 | 7310 | preempt: |
8875125e | 7311 | resched_curr(rq); |
3a7e73a2 PZ |
7312 | /* |
7313 | * Only set the backward buddy when the current task is still | |
7314 | * on the rq. This can happen when a wakeup gets interleaved | |
7315 | * with schedule on the ->pre_schedule() or idle_balance() | |
7316 | * point, either of which can * drop the rq lock. | |
7317 | * | |
7318 | * Also, during early boot the idle thread is in the fair class, | |
7319 | * for obvious reasons its a bad idea to schedule back to it. | |
7320 | */ | |
7321 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
7322 | return; | |
7323 | ||
7324 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
7325 | set_last_buddy(se); | |
bf0f6f24 IM |
7326 | } |
7327 | ||
21f56ffe PZ |
7328 | #ifdef CONFIG_SMP |
7329 | static struct task_struct *pick_task_fair(struct rq *rq) | |
7330 | { | |
7331 | struct sched_entity *se; | |
7332 | struct cfs_rq *cfs_rq; | |
7333 | ||
7334 | again: | |
7335 | cfs_rq = &rq->cfs; | |
7336 | if (!cfs_rq->nr_running) | |
7337 | return NULL; | |
7338 | ||
7339 | do { | |
7340 | struct sched_entity *curr = cfs_rq->curr; | |
7341 | ||
7342 | /* When we pick for a remote RQ, we'll not have done put_prev_entity() */ | |
7343 | if (curr) { | |
7344 | if (curr->on_rq) | |
7345 | update_curr(cfs_rq); | |
7346 | else | |
7347 | curr = NULL; | |
7348 | ||
7349 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | |
7350 | goto again; | |
7351 | } | |
7352 | ||
7353 | se = pick_next_entity(cfs_rq, curr); | |
7354 | cfs_rq = group_cfs_rq(se); | |
7355 | } while (cfs_rq); | |
7356 | ||
7357 | return task_of(se); | |
7358 | } | |
7359 | #endif | |
7360 | ||
5d7d6056 | 7361 | struct task_struct * |
d8ac8971 | 7362 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
bf0f6f24 IM |
7363 | { |
7364 | struct cfs_rq *cfs_rq = &rq->cfs; | |
7365 | struct sched_entity *se; | |
678d5718 | 7366 | struct task_struct *p; |
37e117c0 | 7367 | int new_tasks; |
678d5718 | 7368 | |
6e83125c | 7369 | again: |
6e2df058 | 7370 | if (!sched_fair_runnable(rq)) |
38033c37 | 7371 | goto idle; |
678d5718 | 7372 | |
9674f5ca | 7373 | #ifdef CONFIG_FAIR_GROUP_SCHED |
67692435 | 7374 | if (!prev || prev->sched_class != &fair_sched_class) |
678d5718 PZ |
7375 | goto simple; |
7376 | ||
7377 | /* | |
7378 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
7379 | * likely that a next task is from the same cgroup as the current. | |
7380 | * | |
7381 | * Therefore attempt to avoid putting and setting the entire cgroup | |
7382 | * hierarchy, only change the part that actually changes. | |
7383 | */ | |
7384 | ||
7385 | do { | |
7386 | struct sched_entity *curr = cfs_rq->curr; | |
7387 | ||
7388 | /* | |
7389 | * Since we got here without doing put_prev_entity() we also | |
7390 | * have to consider cfs_rq->curr. If it is still a runnable | |
7391 | * entity, update_curr() will update its vruntime, otherwise | |
7392 | * forget we've ever seen it. | |
7393 | */ | |
54d27365 BS |
7394 | if (curr) { |
7395 | if (curr->on_rq) | |
7396 | update_curr(cfs_rq); | |
7397 | else | |
7398 | curr = NULL; | |
678d5718 | 7399 | |
54d27365 BS |
7400 | /* |
7401 | * This call to check_cfs_rq_runtime() will do the | |
7402 | * throttle and dequeue its entity in the parent(s). | |
9674f5ca | 7403 | * Therefore the nr_running test will indeed |
54d27365 BS |
7404 | * be correct. |
7405 | */ | |
9674f5ca VK |
7406 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) { |
7407 | cfs_rq = &rq->cfs; | |
7408 | ||
7409 | if (!cfs_rq->nr_running) | |
7410 | goto idle; | |
7411 | ||
54d27365 | 7412 | goto simple; |
9674f5ca | 7413 | } |
54d27365 | 7414 | } |
678d5718 PZ |
7415 | |
7416 | se = pick_next_entity(cfs_rq, curr); | |
7417 | cfs_rq = group_cfs_rq(se); | |
7418 | } while (cfs_rq); | |
7419 | ||
7420 | p = task_of(se); | |
7421 | ||
7422 | /* | |
7423 | * Since we haven't yet done put_prev_entity and if the selected task | |
7424 | * is a different task than we started out with, try and touch the | |
7425 | * least amount of cfs_rqs. | |
7426 | */ | |
7427 | if (prev != p) { | |
7428 | struct sched_entity *pse = &prev->se; | |
7429 | ||
7430 | while (!(cfs_rq = is_same_group(se, pse))) { | |
7431 | int se_depth = se->depth; | |
7432 | int pse_depth = pse->depth; | |
7433 | ||
7434 | if (se_depth <= pse_depth) { | |
7435 | put_prev_entity(cfs_rq_of(pse), pse); | |
7436 | pse = parent_entity(pse); | |
7437 | } | |
7438 | if (se_depth >= pse_depth) { | |
7439 | set_next_entity(cfs_rq_of(se), se); | |
7440 | se = parent_entity(se); | |
7441 | } | |
7442 | } | |
7443 | ||
7444 | put_prev_entity(cfs_rq, pse); | |
7445 | set_next_entity(cfs_rq, se); | |
7446 | } | |
7447 | ||
93824900 | 7448 | goto done; |
678d5718 | 7449 | simple: |
678d5718 | 7450 | #endif |
67692435 PZ |
7451 | if (prev) |
7452 | put_prev_task(rq, prev); | |
606dba2e | 7453 | |
bf0f6f24 | 7454 | do { |
678d5718 | 7455 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 7456 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
7457 | cfs_rq = group_cfs_rq(se); |
7458 | } while (cfs_rq); | |
7459 | ||
8f4d37ec | 7460 | p = task_of(se); |
678d5718 | 7461 | |
13a453c2 | 7462 | done: __maybe_unused; |
93824900 UR |
7463 | #ifdef CONFIG_SMP |
7464 | /* | |
7465 | * Move the next running task to the front of | |
7466 | * the list, so our cfs_tasks list becomes MRU | |
7467 | * one. | |
7468 | */ | |
7469 | list_move(&p->se.group_node, &rq->cfs_tasks); | |
7470 | #endif | |
7471 | ||
e0ee463c | 7472 | if (hrtick_enabled_fair(rq)) |
b39e66ea | 7473 | hrtick_start_fair(rq, p); |
8f4d37ec | 7474 | |
3b1baa64 MR |
7475 | update_misfit_status(p, rq); |
7476 | ||
8f4d37ec | 7477 | return p; |
38033c37 PZ |
7478 | |
7479 | idle: | |
67692435 PZ |
7480 | if (!rf) |
7481 | return NULL; | |
7482 | ||
5ba553ef | 7483 | new_tasks = newidle_balance(rq, rf); |
46f69fa3 | 7484 | |
37e117c0 | 7485 | /* |
5ba553ef | 7486 | * Because newidle_balance() releases (and re-acquires) rq->lock, it is |
37e117c0 PZ |
7487 | * possible for any higher priority task to appear. In that case we |
7488 | * must re-start the pick_next_entity() loop. | |
7489 | */ | |
e4aa358b | 7490 | if (new_tasks < 0) |
37e117c0 PZ |
7491 | return RETRY_TASK; |
7492 | ||
e4aa358b | 7493 | if (new_tasks > 0) |
38033c37 | 7494 | goto again; |
38033c37 | 7495 | |
23127296 VG |
7496 | /* |
7497 | * rq is about to be idle, check if we need to update the | |
7498 | * lost_idle_time of clock_pelt | |
7499 | */ | |
7500 | update_idle_rq_clock_pelt(rq); | |
7501 | ||
38033c37 | 7502 | return NULL; |
bf0f6f24 IM |
7503 | } |
7504 | ||
98c2f700 PZ |
7505 | static struct task_struct *__pick_next_task_fair(struct rq *rq) |
7506 | { | |
7507 | return pick_next_task_fair(rq, NULL, NULL); | |
7508 | } | |
7509 | ||
bf0f6f24 IM |
7510 | /* |
7511 | * Account for a descheduled task: | |
7512 | */ | |
6e2df058 | 7513 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
7514 | { |
7515 | struct sched_entity *se = &prev->se; | |
7516 | struct cfs_rq *cfs_rq; | |
7517 | ||
7518 | for_each_sched_entity(se) { | |
7519 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 7520 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
7521 | } |
7522 | } | |
7523 | ||
ac53db59 RR |
7524 | /* |
7525 | * sched_yield() is very simple | |
7526 | * | |
7527 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
7528 | */ | |
7529 | static void yield_task_fair(struct rq *rq) | |
7530 | { | |
7531 | struct task_struct *curr = rq->curr; | |
7532 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
7533 | struct sched_entity *se = &curr->se; | |
7534 | ||
7535 | /* | |
7536 | * Are we the only task in the tree? | |
7537 | */ | |
7538 | if (unlikely(rq->nr_running == 1)) | |
7539 | return; | |
7540 | ||
7541 | clear_buddies(cfs_rq, se); | |
7542 | ||
7543 | if (curr->policy != SCHED_BATCH) { | |
7544 | update_rq_clock(rq); | |
7545 | /* | |
7546 | * Update run-time statistics of the 'current'. | |
7547 | */ | |
7548 | update_curr(cfs_rq); | |
916671c0 MG |
7549 | /* |
7550 | * Tell update_rq_clock() that we've just updated, | |
7551 | * so we don't do microscopic update in schedule() | |
7552 | * and double the fastpath cost. | |
7553 | */ | |
adcc8da8 | 7554 | rq_clock_skip_update(rq); |
ac53db59 RR |
7555 | } |
7556 | ||
7557 | set_skip_buddy(se); | |
7558 | } | |
7559 | ||
0900acf2 | 7560 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p) |
d95f4122 MG |
7561 | { |
7562 | struct sched_entity *se = &p->se; | |
7563 | ||
5238cdd3 PT |
7564 | /* throttled hierarchies are not runnable */ |
7565 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
7566 | return false; |
7567 | ||
7568 | /* Tell the scheduler that we'd really like pse to run next. */ | |
7569 | set_next_buddy(se); | |
7570 | ||
d95f4122 MG |
7571 | yield_task_fair(rq); |
7572 | ||
7573 | return true; | |
7574 | } | |
7575 | ||
681f3e68 | 7576 | #ifdef CONFIG_SMP |
bf0f6f24 | 7577 | /************************************************** |
e9c84cb8 PZ |
7578 | * Fair scheduling class load-balancing methods. |
7579 | * | |
7580 | * BASICS | |
7581 | * | |
7582 | * The purpose of load-balancing is to achieve the same basic fairness the | |
97fb7a0a | 7583 | * per-CPU scheduler provides, namely provide a proportional amount of compute |
e9c84cb8 PZ |
7584 | * time to each task. This is expressed in the following equation: |
7585 | * | |
7586 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
7587 | * | |
97fb7a0a | 7588 | * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight |
e9c84cb8 PZ |
7589 | * W_i,0 is defined as: |
7590 | * | |
7591 | * W_i,0 = \Sum_j w_i,j (2) | |
7592 | * | |
97fb7a0a | 7593 | * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight |
1c3de5e1 | 7594 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
7595 | * |
7596 | * The weight average is an exponential decay average of the instantaneous | |
7597 | * weight: | |
7598 | * | |
7599 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
7600 | * | |
97fb7a0a | 7601 | * C_i is the compute capacity of CPU i, typically it is the |
e9c84cb8 PZ |
7602 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
7603 | * can also include other factors [XXX]. | |
7604 | * | |
7605 | * To achieve this balance we define a measure of imbalance which follows | |
7606 | * directly from (1): | |
7607 | * | |
ced549fa | 7608 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
7609 | * |
7610 | * We them move tasks around to minimize the imbalance. In the continuous | |
7611 | * function space it is obvious this converges, in the discrete case we get | |
7612 | * a few fun cases generally called infeasible weight scenarios. | |
7613 | * | |
7614 | * [XXX expand on: | |
7615 | * - infeasible weights; | |
7616 | * - local vs global optima in the discrete case. ] | |
7617 | * | |
7618 | * | |
7619 | * SCHED DOMAINS | |
7620 | * | |
7621 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
97fb7a0a | 7622 | * for all i,j solution, we create a tree of CPUs that follows the hardware |
e9c84cb8 | 7623 | * topology where each level pairs two lower groups (or better). This results |
97fb7a0a | 7624 | * in O(log n) layers. Furthermore we reduce the number of CPUs going up the |
e9c84cb8 | 7625 | * tree to only the first of the previous level and we decrease the frequency |
97fb7a0a | 7626 | * of load-balance at each level inv. proportional to the number of CPUs in |
e9c84cb8 PZ |
7627 | * the groups. |
7628 | * | |
7629 | * This yields: | |
7630 | * | |
7631 | * log_2 n 1 n | |
7632 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
7633 | * i = 0 2^i 2^i | |
7634 | * `- size of each group | |
97fb7a0a | 7635 | * | | `- number of CPUs doing load-balance |
e9c84cb8 PZ |
7636 | * | `- freq |
7637 | * `- sum over all levels | |
7638 | * | |
7639 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
7640 | * this makes (5) the runtime complexity of the balancer. | |
7641 | * | |
7642 | * An important property here is that each CPU is still (indirectly) connected | |
97fb7a0a | 7643 | * to every other CPU in at most O(log n) steps: |
e9c84cb8 PZ |
7644 | * |
7645 | * The adjacency matrix of the resulting graph is given by: | |
7646 | * | |
97a7142f | 7647 | * log_2 n |
e9c84cb8 PZ |
7648 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
7649 | * k = 0 | |
7650 | * | |
7651 | * And you'll find that: | |
7652 | * | |
7653 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
7654 | * | |
97fb7a0a | 7655 | * Showing there's indeed a path between every CPU in at most O(log n) steps. |
e9c84cb8 PZ |
7656 | * The task movement gives a factor of O(m), giving a convergence complexity |
7657 | * of: | |
7658 | * | |
7659 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
7660 | * | |
7661 | * | |
7662 | * WORK CONSERVING | |
7663 | * | |
7664 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
97fb7a0a | 7665 | * balancing is more aggressive and has the newly idle CPU iterate up the domain |
e9c84cb8 PZ |
7666 | * tree itself instead of relying on other CPUs to bring it work. |
7667 | * | |
7668 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
7669 | * time. | |
7670 | * | |
7671 | * [XXX more?] | |
7672 | * | |
7673 | * | |
7674 | * CGROUPS | |
7675 | * | |
7676 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
7677 | * | |
7678 | * s_k,i | |
7679 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
7680 | * S_k | |
7681 | * | |
7682 | * Where | |
7683 | * | |
7684 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
7685 | * | |
97fb7a0a | 7686 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i. |
e9c84cb8 PZ |
7687 | * |
7688 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
7689 | * property. | |
7690 | * | |
7691 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
7692 | * rewrite all of this once again.] | |
97a7142f | 7693 | */ |
bf0f6f24 | 7694 | |
ed387b78 HS |
7695 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
7696 | ||
0ec8aa00 PZ |
7697 | enum fbq_type { regular, remote, all }; |
7698 | ||
0b0695f2 | 7699 | /* |
a9723389 VG |
7700 | * 'group_type' describes the group of CPUs at the moment of load balancing. |
7701 | * | |
0b0695f2 | 7702 | * The enum is ordered by pulling priority, with the group with lowest priority |
a9723389 VG |
7703 | * first so the group_type can simply be compared when selecting the busiest |
7704 | * group. See update_sd_pick_busiest(). | |
0b0695f2 | 7705 | */ |
3b1baa64 | 7706 | enum group_type { |
a9723389 | 7707 | /* The group has spare capacity that can be used to run more tasks. */ |
0b0695f2 | 7708 | group_has_spare = 0, |
a9723389 VG |
7709 | /* |
7710 | * The group is fully used and the tasks don't compete for more CPU | |
7711 | * cycles. Nevertheless, some tasks might wait before running. | |
7712 | */ | |
0b0695f2 | 7713 | group_fully_busy, |
a9723389 | 7714 | /* |
c82a6962 VG |
7715 | * One task doesn't fit with CPU's capacity and must be migrated to a |
7716 | * more powerful CPU. | |
a9723389 | 7717 | */ |
3b1baa64 | 7718 | group_misfit_task, |
a9723389 VG |
7719 | /* |
7720 | * SD_ASYM_PACKING only: One local CPU with higher capacity is available, | |
7721 | * and the task should be migrated to it instead of running on the | |
7722 | * current CPU. | |
7723 | */ | |
0b0695f2 | 7724 | group_asym_packing, |
a9723389 VG |
7725 | /* |
7726 | * The tasks' affinity constraints previously prevented the scheduler | |
7727 | * from balancing the load across the system. | |
7728 | */ | |
3b1baa64 | 7729 | group_imbalanced, |
a9723389 VG |
7730 | /* |
7731 | * The CPU is overloaded and can't provide expected CPU cycles to all | |
7732 | * tasks. | |
7733 | */ | |
0b0695f2 VG |
7734 | group_overloaded |
7735 | }; | |
7736 | ||
7737 | enum migration_type { | |
7738 | migrate_load = 0, | |
7739 | migrate_util, | |
7740 | migrate_task, | |
7741 | migrate_misfit | |
3b1baa64 MR |
7742 | }; |
7743 | ||
ddcdf6e7 | 7744 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 7745 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
7746 | #define LBF_DST_PINNED 0x04 |
7747 | #define LBF_SOME_PINNED 0x08 | |
23fb06d9 | 7748 | #define LBF_ACTIVE_LB 0x10 |
ddcdf6e7 PZ |
7749 | |
7750 | struct lb_env { | |
7751 | struct sched_domain *sd; | |
7752 | ||
ddcdf6e7 | 7753 | struct rq *src_rq; |
85c1e7da | 7754 | int src_cpu; |
ddcdf6e7 PZ |
7755 | |
7756 | int dst_cpu; | |
7757 | struct rq *dst_rq; | |
7758 | ||
88b8dac0 SV |
7759 | struct cpumask *dst_grpmask; |
7760 | int new_dst_cpu; | |
ddcdf6e7 | 7761 | enum cpu_idle_type idle; |
bd939f45 | 7762 | long imbalance; |
b9403130 MW |
7763 | /* The set of CPUs under consideration for load-balancing */ |
7764 | struct cpumask *cpus; | |
7765 | ||
ddcdf6e7 | 7766 | unsigned int flags; |
367456c7 PZ |
7767 | |
7768 | unsigned int loop; | |
7769 | unsigned int loop_break; | |
7770 | unsigned int loop_max; | |
0ec8aa00 PZ |
7771 | |
7772 | enum fbq_type fbq_type; | |
0b0695f2 | 7773 | enum migration_type migration_type; |
163122b7 | 7774 | struct list_head tasks; |
ddcdf6e7 PZ |
7775 | }; |
7776 | ||
029632fb PZ |
7777 | /* |
7778 | * Is this task likely cache-hot: | |
7779 | */ | |
5d5e2b1b | 7780 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
7781 | { |
7782 | s64 delta; | |
7783 | ||
5cb9eaa3 | 7784 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 7785 | |
029632fb PZ |
7786 | if (p->sched_class != &fair_sched_class) |
7787 | return 0; | |
7788 | ||
1da1843f | 7789 | if (unlikely(task_has_idle_policy(p))) |
029632fb PZ |
7790 | return 0; |
7791 | ||
ec73240b JD |
7792 | /* SMT siblings share cache */ |
7793 | if (env->sd->flags & SD_SHARE_CPUCAPACITY) | |
7794 | return 0; | |
7795 | ||
029632fb PZ |
7796 | /* |
7797 | * Buddy candidates are cache hot: | |
7798 | */ | |
5d5e2b1b | 7799 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
7800 | (&p->se == cfs_rq_of(&p->se)->next || |
7801 | &p->se == cfs_rq_of(&p->se)->last)) | |
7802 | return 1; | |
7803 | ||
7804 | if (sysctl_sched_migration_cost == -1) | |
7805 | return 1; | |
97886d9d AL |
7806 | |
7807 | /* | |
7808 | * Don't migrate task if the task's cookie does not match | |
7809 | * with the destination CPU's core cookie. | |
7810 | */ | |
7811 | if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p)) | |
7812 | return 1; | |
7813 | ||
029632fb PZ |
7814 | if (sysctl_sched_migration_cost == 0) |
7815 | return 0; | |
7816 | ||
5d5e2b1b | 7817 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
7818 | |
7819 | return delta < (s64)sysctl_sched_migration_cost; | |
7820 | } | |
7821 | ||
3a7053b3 | 7822 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 7823 | /* |
2a1ed24c SD |
7824 | * Returns 1, if task migration degrades locality |
7825 | * Returns 0, if task migration improves locality i.e migration preferred. | |
7826 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 7827 | */ |
2a1ed24c | 7828 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 7829 | { |
b1ad065e | 7830 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
f35678b6 SD |
7831 | unsigned long src_weight, dst_weight; |
7832 | int src_nid, dst_nid, dist; | |
3a7053b3 | 7833 | |
2a595721 | 7834 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
7835 | return -1; |
7836 | ||
c3b9bc5b | 7837 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 7838 | return -1; |
7a0f3083 MG |
7839 | |
7840 | src_nid = cpu_to_node(env->src_cpu); | |
7841 | dst_nid = cpu_to_node(env->dst_cpu); | |
7842 | ||
83e1d2cd | 7843 | if (src_nid == dst_nid) |
2a1ed24c | 7844 | return -1; |
7a0f3083 | 7845 | |
2a1ed24c SD |
7846 | /* Migrating away from the preferred node is always bad. */ |
7847 | if (src_nid == p->numa_preferred_nid) { | |
7848 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
7849 | return 1; | |
7850 | else | |
7851 | return -1; | |
7852 | } | |
b1ad065e | 7853 | |
c1ceac62 RR |
7854 | /* Encourage migration to the preferred node. */ |
7855 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 7856 | return 0; |
b1ad065e | 7857 | |
739294fb | 7858 | /* Leaving a core idle is often worse than degrading locality. */ |
f35678b6 | 7859 | if (env->idle == CPU_IDLE) |
739294fb RR |
7860 | return -1; |
7861 | ||
f35678b6 | 7862 | dist = node_distance(src_nid, dst_nid); |
c1ceac62 | 7863 | if (numa_group) { |
f35678b6 SD |
7864 | src_weight = group_weight(p, src_nid, dist); |
7865 | dst_weight = group_weight(p, dst_nid, dist); | |
c1ceac62 | 7866 | } else { |
f35678b6 SD |
7867 | src_weight = task_weight(p, src_nid, dist); |
7868 | dst_weight = task_weight(p, dst_nid, dist); | |
b1ad065e RR |
7869 | } |
7870 | ||
f35678b6 | 7871 | return dst_weight < src_weight; |
7a0f3083 MG |
7872 | } |
7873 | ||
3a7053b3 | 7874 | #else |
2a1ed24c | 7875 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
7876 | struct lb_env *env) |
7877 | { | |
2a1ed24c | 7878 | return -1; |
7a0f3083 | 7879 | } |
3a7053b3 MG |
7880 | #endif |
7881 | ||
1e3c88bd PZ |
7882 | /* |
7883 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
7884 | */ | |
7885 | static | |
8e45cb54 | 7886 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 7887 | { |
2a1ed24c | 7888 | int tsk_cache_hot; |
e5673f28 | 7889 | |
5cb9eaa3 | 7890 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 7891 | |
1e3c88bd PZ |
7892 | /* |
7893 | * We do not migrate tasks that are: | |
d3198084 | 7894 | * 1) throttled_lb_pair, or |
3bd37062 | 7895 | * 2) cannot be migrated to this CPU due to cpus_ptr, or |
d3198084 JK |
7896 | * 3) running (obviously), or |
7897 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 7898 | */ |
d3198084 JK |
7899 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
7900 | return 0; | |
7901 | ||
9bcb959d | 7902 | /* Disregard pcpu kthreads; they are where they need to be. */ |
3a7956e2 | 7903 | if (kthread_is_per_cpu(p)) |
9bcb959d LC |
7904 | return 0; |
7905 | ||
3bd37062 | 7906 | if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { |
e02e60c1 | 7907 | int cpu; |
88b8dac0 | 7908 | |
ceeadb83 | 7909 | schedstat_inc(p->stats.nr_failed_migrations_affine); |
88b8dac0 | 7910 | |
6263322c PZ |
7911 | env->flags |= LBF_SOME_PINNED; |
7912 | ||
88b8dac0 | 7913 | /* |
97fb7a0a | 7914 | * Remember if this task can be migrated to any other CPU in |
88b8dac0 SV |
7915 | * our sched_group. We may want to revisit it if we couldn't |
7916 | * meet load balance goals by pulling other tasks on src_cpu. | |
7917 | * | |
23fb06d9 VS |
7918 | * Avoid computing new_dst_cpu |
7919 | * - for NEWLY_IDLE | |
7920 | * - if we have already computed one in current iteration | |
7921 | * - if it's an active balance | |
88b8dac0 | 7922 | */ |
23fb06d9 VS |
7923 | if (env->idle == CPU_NEWLY_IDLE || |
7924 | env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB)) | |
88b8dac0 SV |
7925 | return 0; |
7926 | ||
97fb7a0a | 7927 | /* Prevent to re-select dst_cpu via env's CPUs: */ |
e02e60c1 | 7928 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { |
3bd37062 | 7929 | if (cpumask_test_cpu(cpu, p->cpus_ptr)) { |
6263322c | 7930 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
7931 | env->new_dst_cpu = cpu; |
7932 | break; | |
7933 | } | |
88b8dac0 | 7934 | } |
e02e60c1 | 7935 | |
1e3c88bd PZ |
7936 | return 0; |
7937 | } | |
88b8dac0 | 7938 | |
3b03706f | 7939 | /* Record that we found at least one task that could run on dst_cpu */ |
8e45cb54 | 7940 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 7941 | |
ddcdf6e7 | 7942 | if (task_running(env->src_rq, p)) { |
ceeadb83 | 7943 | schedstat_inc(p->stats.nr_failed_migrations_running); |
1e3c88bd PZ |
7944 | return 0; |
7945 | } | |
7946 | ||
7947 | /* | |
7948 | * Aggressive migration if: | |
23fb06d9 VS |
7949 | * 1) active balance |
7950 | * 2) destination numa is preferred | |
7951 | * 3) task is cache cold, or | |
7952 | * 4) too many balance attempts have failed. | |
1e3c88bd | 7953 | */ |
23fb06d9 VS |
7954 | if (env->flags & LBF_ACTIVE_LB) |
7955 | return 1; | |
7956 | ||
2a1ed24c SD |
7957 | tsk_cache_hot = migrate_degrades_locality(p, env); |
7958 | if (tsk_cache_hot == -1) | |
7959 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 7960 | |
2a1ed24c | 7961 | if (tsk_cache_hot <= 0 || |
7a96c231 | 7962 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 7963 | if (tsk_cache_hot == 1) { |
ae92882e | 7964 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
ceeadb83 | 7965 | schedstat_inc(p->stats.nr_forced_migrations); |
3a7053b3 | 7966 | } |
1e3c88bd PZ |
7967 | return 1; |
7968 | } | |
7969 | ||
ceeadb83 | 7970 | schedstat_inc(p->stats.nr_failed_migrations_hot); |
4e2dcb73 | 7971 | return 0; |
1e3c88bd PZ |
7972 | } |
7973 | ||
897c395f | 7974 | /* |
163122b7 KT |
7975 | * detach_task() -- detach the task for the migration specified in env |
7976 | */ | |
7977 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
7978 | { | |
5cb9eaa3 | 7979 | lockdep_assert_rq_held(env->src_rq); |
163122b7 | 7980 | |
5704ac0a | 7981 | deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); |
163122b7 KT |
7982 | set_task_cpu(p, env->dst_cpu); |
7983 | } | |
7984 | ||
897c395f | 7985 | /* |
e5673f28 | 7986 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 7987 | * part of active balancing operations within "domain". |
897c395f | 7988 | * |
e5673f28 | 7989 | * Returns a task if successful and NULL otherwise. |
897c395f | 7990 | */ |
e5673f28 | 7991 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f | 7992 | { |
93824900 | 7993 | struct task_struct *p; |
897c395f | 7994 | |
5cb9eaa3 | 7995 | lockdep_assert_rq_held(env->src_rq); |
e5673f28 | 7996 | |
93824900 UR |
7997 | list_for_each_entry_reverse(p, |
7998 | &env->src_rq->cfs_tasks, se.group_node) { | |
367456c7 PZ |
7999 | if (!can_migrate_task(p, env)) |
8000 | continue; | |
897c395f | 8001 | |
163122b7 | 8002 | detach_task(p, env); |
e5673f28 | 8003 | |
367456c7 | 8004 | /* |
e5673f28 | 8005 | * Right now, this is only the second place where |
163122b7 | 8006 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 8007 | * so we can safely collect stats here rather than |
163122b7 | 8008 | * inside detach_tasks(). |
367456c7 | 8009 | */ |
ae92882e | 8010 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 8011 | return p; |
897c395f | 8012 | } |
e5673f28 | 8013 | return NULL; |
897c395f PZ |
8014 | } |
8015 | ||
eb95308e PZ |
8016 | static const unsigned int sched_nr_migrate_break = 32; |
8017 | ||
5d6523eb | 8018 | /* |
0b0695f2 | 8019 | * detach_tasks() -- tries to detach up to imbalance load/util/tasks from |
163122b7 | 8020 | * busiest_rq, as part of a balancing operation within domain "sd". |
5d6523eb | 8021 | * |
163122b7 | 8022 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 8023 | */ |
163122b7 | 8024 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 8025 | { |
5d6523eb | 8026 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
0b0695f2 | 8027 | unsigned long util, load; |
5d6523eb | 8028 | struct task_struct *p; |
163122b7 KT |
8029 | int detached = 0; |
8030 | ||
5cb9eaa3 | 8031 | lockdep_assert_rq_held(env->src_rq); |
1e3c88bd | 8032 | |
acb4decc AL |
8033 | /* |
8034 | * Source run queue has been emptied by another CPU, clear | |
8035 | * LBF_ALL_PINNED flag as we will not test any task. | |
8036 | */ | |
8037 | if (env->src_rq->nr_running <= 1) { | |
8038 | env->flags &= ~LBF_ALL_PINNED; | |
8039 | return 0; | |
8040 | } | |
8041 | ||
bd939f45 | 8042 | if (env->imbalance <= 0) |
5d6523eb | 8043 | return 0; |
1e3c88bd | 8044 | |
5d6523eb | 8045 | while (!list_empty(tasks)) { |
985d3a4c YD |
8046 | /* |
8047 | * We don't want to steal all, otherwise we may be treated likewise, | |
8048 | * which could at worst lead to a livelock crash. | |
8049 | */ | |
8050 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
8051 | break; | |
8052 | ||
93824900 | 8053 | p = list_last_entry(tasks, struct task_struct, se.group_node); |
1e3c88bd | 8054 | |
367456c7 PZ |
8055 | env->loop++; |
8056 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 8057 | if (env->loop > env->loop_max) |
367456c7 | 8058 | break; |
5d6523eb PZ |
8059 | |
8060 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 8061 | if (env->loop > env->loop_break) { |
eb95308e | 8062 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 8063 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 8064 | break; |
a195f004 | 8065 | } |
1e3c88bd | 8066 | |
d3198084 | 8067 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
8068 | goto next; |
8069 | ||
0b0695f2 VG |
8070 | switch (env->migration_type) { |
8071 | case migrate_load: | |
01cfcde9 VG |
8072 | /* |
8073 | * Depending of the number of CPUs and tasks and the | |
8074 | * cgroup hierarchy, task_h_load() can return a null | |
8075 | * value. Make sure that env->imbalance decreases | |
8076 | * otherwise detach_tasks() will stop only after | |
8077 | * detaching up to loop_max tasks. | |
8078 | */ | |
8079 | load = max_t(unsigned long, task_h_load(p), 1); | |
5d6523eb | 8080 | |
0b0695f2 VG |
8081 | if (sched_feat(LB_MIN) && |
8082 | load < 16 && !env->sd->nr_balance_failed) | |
8083 | goto next; | |
367456c7 | 8084 | |
6cf82d55 VG |
8085 | /* |
8086 | * Make sure that we don't migrate too much load. | |
8087 | * Nevertheless, let relax the constraint if | |
8088 | * scheduler fails to find a good waiting task to | |
8089 | * migrate. | |
8090 | */ | |
39a2a6eb | 8091 | if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance) |
0b0695f2 VG |
8092 | goto next; |
8093 | ||
8094 | env->imbalance -= load; | |
8095 | break; | |
8096 | ||
8097 | case migrate_util: | |
8098 | util = task_util_est(p); | |
8099 | ||
8100 | if (util > env->imbalance) | |
8101 | goto next; | |
8102 | ||
8103 | env->imbalance -= util; | |
8104 | break; | |
8105 | ||
8106 | case migrate_task: | |
8107 | env->imbalance--; | |
8108 | break; | |
8109 | ||
8110 | case migrate_misfit: | |
c63be7be VG |
8111 | /* This is not a misfit task */ |
8112 | if (task_fits_capacity(p, capacity_of(env->src_cpu))) | |
0b0695f2 VG |
8113 | goto next; |
8114 | ||
8115 | env->imbalance = 0; | |
8116 | break; | |
8117 | } | |
1e3c88bd | 8118 | |
163122b7 KT |
8119 | detach_task(p, env); |
8120 | list_add(&p->se.group_node, &env->tasks); | |
8121 | ||
8122 | detached++; | |
1e3c88bd | 8123 | |
c1a280b6 | 8124 | #ifdef CONFIG_PREEMPTION |
ee00e66f PZ |
8125 | /* |
8126 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 8127 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
8128 | * the critical section. |
8129 | */ | |
5d6523eb | 8130 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 8131 | break; |
1e3c88bd PZ |
8132 | #endif |
8133 | ||
ee00e66f PZ |
8134 | /* |
8135 | * We only want to steal up to the prescribed amount of | |
0b0695f2 | 8136 | * load/util/tasks. |
ee00e66f | 8137 | */ |
bd939f45 | 8138 | if (env->imbalance <= 0) |
ee00e66f | 8139 | break; |
367456c7 PZ |
8140 | |
8141 | continue; | |
8142 | next: | |
93824900 | 8143 | list_move(&p->se.group_node, tasks); |
1e3c88bd | 8144 | } |
5d6523eb | 8145 | |
1e3c88bd | 8146 | /* |
163122b7 KT |
8147 | * Right now, this is one of only two places we collect this stat |
8148 | * so we can safely collect detach_one_task() stats here rather | |
8149 | * than inside detach_one_task(). | |
1e3c88bd | 8150 | */ |
ae92882e | 8151 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 8152 | |
163122b7 KT |
8153 | return detached; |
8154 | } | |
8155 | ||
8156 | /* | |
8157 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
8158 | */ | |
8159 | static void attach_task(struct rq *rq, struct task_struct *p) | |
8160 | { | |
5cb9eaa3 | 8161 | lockdep_assert_rq_held(rq); |
163122b7 KT |
8162 | |
8163 | BUG_ON(task_rq(p) != rq); | |
5704ac0a | 8164 | activate_task(rq, p, ENQUEUE_NOCLOCK); |
163122b7 KT |
8165 | check_preempt_curr(rq, p, 0); |
8166 | } | |
8167 | ||
8168 | /* | |
8169 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
8170 | * its new rq. | |
8171 | */ | |
8172 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
8173 | { | |
8a8c69c3 PZ |
8174 | struct rq_flags rf; |
8175 | ||
8176 | rq_lock(rq, &rf); | |
5704ac0a | 8177 | update_rq_clock(rq); |
163122b7 | 8178 | attach_task(rq, p); |
8a8c69c3 | 8179 | rq_unlock(rq, &rf); |
163122b7 KT |
8180 | } |
8181 | ||
8182 | /* | |
8183 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
8184 | * new rq. | |
8185 | */ | |
8186 | static void attach_tasks(struct lb_env *env) | |
8187 | { | |
8188 | struct list_head *tasks = &env->tasks; | |
8189 | struct task_struct *p; | |
8a8c69c3 | 8190 | struct rq_flags rf; |
163122b7 | 8191 | |
8a8c69c3 | 8192 | rq_lock(env->dst_rq, &rf); |
5704ac0a | 8193 | update_rq_clock(env->dst_rq); |
163122b7 KT |
8194 | |
8195 | while (!list_empty(tasks)) { | |
8196 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
8197 | list_del_init(&p->se.group_node); | |
1e3c88bd | 8198 | |
163122b7 KT |
8199 | attach_task(env->dst_rq, p); |
8200 | } | |
8201 | ||
8a8c69c3 | 8202 | rq_unlock(env->dst_rq, &rf); |
1e3c88bd PZ |
8203 | } |
8204 | ||
b0c79224 | 8205 | #ifdef CONFIG_NO_HZ_COMMON |
1936c53c VG |
8206 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) |
8207 | { | |
8208 | if (cfs_rq->avg.load_avg) | |
8209 | return true; | |
8210 | ||
8211 | if (cfs_rq->avg.util_avg) | |
8212 | return true; | |
8213 | ||
8214 | return false; | |
8215 | } | |
8216 | ||
91c27493 | 8217 | static inline bool others_have_blocked(struct rq *rq) |
371bf427 VG |
8218 | { |
8219 | if (READ_ONCE(rq->avg_rt.util_avg)) | |
8220 | return true; | |
8221 | ||
3727e0e1 VG |
8222 | if (READ_ONCE(rq->avg_dl.util_avg)) |
8223 | return true; | |
8224 | ||
b4eccf5f TG |
8225 | if (thermal_load_avg(rq)) |
8226 | return true; | |
8227 | ||
11d4afd4 | 8228 | #ifdef CONFIG_HAVE_SCHED_AVG_IRQ |
91c27493 VG |
8229 | if (READ_ONCE(rq->avg_irq.util_avg)) |
8230 | return true; | |
8231 | #endif | |
8232 | ||
371bf427 VG |
8233 | return false; |
8234 | } | |
8235 | ||
39b6a429 | 8236 | static inline void update_blocked_load_tick(struct rq *rq) |
b0c79224 | 8237 | { |
39b6a429 VG |
8238 | WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies); |
8239 | } | |
b0c79224 | 8240 | |
39b6a429 VG |
8241 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) |
8242 | { | |
b0c79224 VS |
8243 | if (!has_blocked) |
8244 | rq->has_blocked_load = 0; | |
8245 | } | |
8246 | #else | |
8247 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; } | |
8248 | static inline bool others_have_blocked(struct rq *rq) { return false; } | |
39b6a429 | 8249 | static inline void update_blocked_load_tick(struct rq *rq) {} |
b0c79224 VS |
8250 | static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {} |
8251 | #endif | |
8252 | ||
bef69dd8 VG |
8253 | static bool __update_blocked_others(struct rq *rq, bool *done) |
8254 | { | |
8255 | const struct sched_class *curr_class; | |
8256 | u64 now = rq_clock_pelt(rq); | |
b4eccf5f | 8257 | unsigned long thermal_pressure; |
bef69dd8 VG |
8258 | bool decayed; |
8259 | ||
8260 | /* | |
8261 | * update_load_avg() can call cpufreq_update_util(). Make sure that RT, | |
8262 | * DL and IRQ signals have been updated before updating CFS. | |
8263 | */ | |
8264 | curr_class = rq->curr->sched_class; | |
8265 | ||
b4eccf5f TG |
8266 | thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq)); |
8267 | ||
bef69dd8 VG |
8268 | decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) | |
8269 | update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) | | |
05289b90 | 8270 | update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) | |
bef69dd8 VG |
8271 | update_irq_load_avg(rq, 0); |
8272 | ||
8273 | if (others_have_blocked(rq)) | |
8274 | *done = false; | |
8275 | ||
8276 | return decayed; | |
8277 | } | |
8278 | ||
1936c53c VG |
8279 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8280 | ||
bef69dd8 | 8281 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 8282 | { |
039ae8bc | 8283 | struct cfs_rq *cfs_rq, *pos; |
bef69dd8 VG |
8284 | bool decayed = false; |
8285 | int cpu = cpu_of(rq); | |
b90f7c9d | 8286 | |
9763b67f PZ |
8287 | /* |
8288 | * Iterates the task_group tree in a bottom up fashion, see | |
8289 | * list_add_leaf_cfs_rq() for details. | |
8290 | */ | |
039ae8bc | 8291 | for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { |
bc427898 VG |
8292 | struct sched_entity *se; |
8293 | ||
bef69dd8 | 8294 | if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) { |
fe749158 | 8295 | update_tg_load_avg(cfs_rq); |
4e516076 | 8296 | |
e2f3e35f VD |
8297 | if (cfs_rq->nr_running == 0) |
8298 | update_idle_cfs_rq_clock_pelt(cfs_rq); | |
8299 | ||
bef69dd8 VG |
8300 | if (cfs_rq == &rq->cfs) |
8301 | decayed = true; | |
8302 | } | |
8303 | ||
bc427898 VG |
8304 | /* Propagate pending load changes to the parent, if any: */ |
8305 | se = cfs_rq->tg->se[cpu]; | |
8306 | if (se && !skip_blocked_update(se)) | |
02da26ad | 8307 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
a9e7f654 | 8308 | |
039ae8bc VG |
8309 | /* |
8310 | * There can be a lot of idle CPU cgroups. Don't let fully | |
8311 | * decayed cfs_rqs linger on the list. | |
8312 | */ | |
8313 | if (cfs_rq_is_decayed(cfs_rq)) | |
8314 | list_del_leaf_cfs_rq(cfs_rq); | |
8315 | ||
1936c53c VG |
8316 | /* Don't need periodic decay once load/util_avg are null */ |
8317 | if (cfs_rq_has_blocked(cfs_rq)) | |
bef69dd8 | 8318 | *done = false; |
9d89c257 | 8319 | } |
12b04875 | 8320 | |
bef69dd8 | 8321 | return decayed; |
9e3081ca PZ |
8322 | } |
8323 | ||
9763b67f | 8324 | /* |
68520796 | 8325 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
8326 | * This needs to be done in a top-down fashion because the load of a child |
8327 | * group is a fraction of its parents load. | |
8328 | */ | |
68520796 | 8329 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 8330 | { |
68520796 VD |
8331 | struct rq *rq = rq_of(cfs_rq); |
8332 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 8333 | unsigned long now = jiffies; |
68520796 | 8334 | unsigned long load; |
a35b6466 | 8335 | |
68520796 | 8336 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
8337 | return; |
8338 | ||
0e9f0245 | 8339 | WRITE_ONCE(cfs_rq->h_load_next, NULL); |
68520796 VD |
8340 | for_each_sched_entity(se) { |
8341 | cfs_rq = cfs_rq_of(se); | |
0e9f0245 | 8342 | WRITE_ONCE(cfs_rq->h_load_next, se); |
68520796 VD |
8343 | if (cfs_rq->last_h_load_update == now) |
8344 | break; | |
8345 | } | |
a35b6466 | 8346 | |
68520796 | 8347 | if (!se) { |
7ea241af | 8348 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
8349 | cfs_rq->last_h_load_update = now; |
8350 | } | |
8351 | ||
0e9f0245 | 8352 | while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { |
68520796 | 8353 | load = cfs_rq->h_load; |
7ea241af YD |
8354 | load = div64_ul(load * se->avg.load_avg, |
8355 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
8356 | cfs_rq = group_cfs_rq(se); |
8357 | cfs_rq->h_load = load; | |
8358 | cfs_rq->last_h_load_update = now; | |
8359 | } | |
9763b67f PZ |
8360 | } |
8361 | ||
367456c7 | 8362 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 8363 | { |
367456c7 | 8364 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 8365 | |
68520796 | 8366 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 8367 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 8368 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
8369 | } |
8370 | #else | |
bef69dd8 | 8371 | static bool __update_blocked_fair(struct rq *rq, bool *done) |
9e3081ca | 8372 | { |
6c1d47c0 | 8373 | struct cfs_rq *cfs_rq = &rq->cfs; |
bef69dd8 | 8374 | bool decayed; |
b90f7c9d | 8375 | |
bef69dd8 VG |
8376 | decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); |
8377 | if (cfs_rq_has_blocked(cfs_rq)) | |
8378 | *done = false; | |
b90f7c9d | 8379 | |
bef69dd8 | 8380 | return decayed; |
9e3081ca PZ |
8381 | } |
8382 | ||
367456c7 | 8383 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 8384 | { |
9d89c257 | 8385 | return p->se.avg.load_avg; |
1e3c88bd | 8386 | } |
230059de | 8387 | #endif |
1e3c88bd | 8388 | |
bef69dd8 VG |
8389 | static void update_blocked_averages(int cpu) |
8390 | { | |
8391 | bool decayed = false, done = true; | |
8392 | struct rq *rq = cpu_rq(cpu); | |
8393 | struct rq_flags rf; | |
8394 | ||
8395 | rq_lock_irqsave(rq, &rf); | |
39b6a429 | 8396 | update_blocked_load_tick(rq); |
bef69dd8 VG |
8397 | update_rq_clock(rq); |
8398 | ||
8399 | decayed |= __update_blocked_others(rq, &done); | |
8400 | decayed |= __update_blocked_fair(rq, &done); | |
8401 | ||
8402 | update_blocked_load_status(rq, !done); | |
8403 | if (decayed) | |
8404 | cpufreq_update_util(rq, 0); | |
8405 | rq_unlock_irqrestore(rq, &rf); | |
8406 | } | |
8407 | ||
1e3c88bd | 8408 | /********** Helpers for find_busiest_group ************************/ |
caeb178c | 8409 | |
1e3c88bd PZ |
8410 | /* |
8411 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
8412 | */ | |
8413 | struct sg_lb_stats { | |
8414 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
8415 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
63b2ca30 | 8416 | unsigned long group_capacity; |
070f5e86 VG |
8417 | unsigned long group_util; /* Total utilization over the CPUs of the group */ |
8418 | unsigned long group_runnable; /* Total runnable time over the CPUs of the group */ | |
5e23e474 | 8419 | unsigned int sum_nr_running; /* Nr of tasks running in the group */ |
a3498347 | 8420 | unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */ |
147c5fc2 PZ |
8421 | unsigned int idle_cpus; |
8422 | unsigned int group_weight; | |
caeb178c | 8423 | enum group_type group_type; |
490ba971 | 8424 | unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */ |
3b1baa64 | 8425 | unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */ |
0ec8aa00 PZ |
8426 | #ifdef CONFIG_NUMA_BALANCING |
8427 | unsigned int nr_numa_running; | |
8428 | unsigned int nr_preferred_running; | |
8429 | #endif | |
1e3c88bd PZ |
8430 | }; |
8431 | ||
56cf515b JK |
8432 | /* |
8433 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
8434 | * during load balancing. | |
8435 | */ | |
8436 | struct sd_lb_stats { | |
8437 | struct sched_group *busiest; /* Busiest group in this sd */ | |
8438 | struct sched_group *local; /* Local group in this sd */ | |
8439 | unsigned long total_load; /* Total load of all groups in sd */ | |
63b2ca30 | 8440 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b | 8441 | unsigned long avg_load; /* Average load across all groups in sd */ |
0b0695f2 | 8442 | unsigned int prefer_sibling; /* tasks should go to sibling first */ |
56cf515b | 8443 | |
56cf515b | 8444 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 8445 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
8446 | }; |
8447 | ||
147c5fc2 PZ |
8448 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
8449 | { | |
8450 | /* | |
8451 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
8452 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
0b0695f2 VG |
8453 | * We must however set busiest_stat::group_type and |
8454 | * busiest_stat::idle_cpus to the worst busiest group because | |
8455 | * update_sd_pick_busiest() reads these before assignment. | |
147c5fc2 PZ |
8456 | */ |
8457 | *sds = (struct sd_lb_stats){ | |
8458 | .busiest = NULL, | |
8459 | .local = NULL, | |
8460 | .total_load = 0UL, | |
63b2ca30 | 8461 | .total_capacity = 0UL, |
147c5fc2 | 8462 | .busiest_stat = { |
0b0695f2 VG |
8463 | .idle_cpus = UINT_MAX, |
8464 | .group_type = group_has_spare, | |
147c5fc2 PZ |
8465 | }, |
8466 | }; | |
8467 | } | |
8468 | ||
1ca2034e | 8469 | static unsigned long scale_rt_capacity(int cpu) |
1e3c88bd PZ |
8470 | { |
8471 | struct rq *rq = cpu_rq(cpu); | |
8ec59c0f | 8472 | unsigned long max = arch_scale_cpu_capacity(cpu); |
523e979d | 8473 | unsigned long used, free; |
523e979d | 8474 | unsigned long irq; |
b654f7de | 8475 | |
2e62c474 | 8476 | irq = cpu_util_irq(rq); |
cadefd3d | 8477 | |
523e979d VG |
8478 | if (unlikely(irq >= max)) |
8479 | return 1; | |
aa483808 | 8480 | |
467b7d01 TG |
8481 | /* |
8482 | * avg_rt.util_avg and avg_dl.util_avg track binary signals | |
8483 | * (running and not running) with weights 0 and 1024 respectively. | |
8484 | * avg_thermal.load_avg tracks thermal pressure and the weighted | |
8485 | * average uses the actual delta max capacity(load). | |
8486 | */ | |
523e979d VG |
8487 | used = READ_ONCE(rq->avg_rt.util_avg); |
8488 | used += READ_ONCE(rq->avg_dl.util_avg); | |
467b7d01 | 8489 | used += thermal_load_avg(rq); |
1e3c88bd | 8490 | |
523e979d VG |
8491 | if (unlikely(used >= max)) |
8492 | return 1; | |
1e3c88bd | 8493 | |
523e979d | 8494 | free = max - used; |
2e62c474 VG |
8495 | |
8496 | return scale_irq_capacity(free, irq, max); | |
1e3c88bd PZ |
8497 | } |
8498 | ||
ced549fa | 8499 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 8500 | { |
1ca2034e | 8501 | unsigned long capacity = scale_rt_capacity(cpu); |
1e3c88bd PZ |
8502 | struct sched_group *sdg = sd->groups; |
8503 | ||
8ec59c0f | 8504 | cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu); |
1e3c88bd | 8505 | |
ced549fa NP |
8506 | if (!capacity) |
8507 | capacity = 1; | |
1e3c88bd | 8508 | |
ced549fa | 8509 | cpu_rq(cpu)->cpu_capacity = capacity; |
51cf18c9 VD |
8510 | trace_sched_cpu_capacity_tp(cpu_rq(cpu)); |
8511 | ||
ced549fa | 8512 | sdg->sgc->capacity = capacity; |
bf475ce0 | 8513 | sdg->sgc->min_capacity = capacity; |
e3d6d0cb | 8514 | sdg->sgc->max_capacity = capacity; |
1e3c88bd PZ |
8515 | } |
8516 | ||
63b2ca30 | 8517 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
8518 | { |
8519 | struct sched_domain *child = sd->child; | |
8520 | struct sched_group *group, *sdg = sd->groups; | |
e3d6d0cb | 8521 | unsigned long capacity, min_capacity, max_capacity; |
4ec4412e VG |
8522 | unsigned long interval; |
8523 | ||
8524 | interval = msecs_to_jiffies(sd->balance_interval); | |
8525 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 8526 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
8527 | |
8528 | if (!child) { | |
ced549fa | 8529 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
8530 | return; |
8531 | } | |
8532 | ||
dc7ff76e | 8533 | capacity = 0; |
bf475ce0 | 8534 | min_capacity = ULONG_MAX; |
e3d6d0cb | 8535 | max_capacity = 0; |
1e3c88bd | 8536 | |
74a5ce20 PZ |
8537 | if (child->flags & SD_OVERLAP) { |
8538 | /* | |
8539 | * SD_OVERLAP domains cannot assume that child groups | |
8540 | * span the current group. | |
8541 | */ | |
8542 | ||
ae4df9d6 | 8543 | for_each_cpu(cpu, sched_group_span(sdg)) { |
4c58f57f | 8544 | unsigned long cpu_cap = capacity_of(cpu); |
863bffc8 | 8545 | |
4c58f57f PL |
8546 | capacity += cpu_cap; |
8547 | min_capacity = min(cpu_cap, min_capacity); | |
8548 | max_capacity = max(cpu_cap, max_capacity); | |
863bffc8 | 8549 | } |
74a5ce20 PZ |
8550 | } else { |
8551 | /* | |
8552 | * !SD_OVERLAP domains can assume that child groups | |
8553 | * span the current group. | |
97a7142f | 8554 | */ |
74a5ce20 PZ |
8555 | |
8556 | group = child->groups; | |
8557 | do { | |
bf475ce0 MR |
8558 | struct sched_group_capacity *sgc = group->sgc; |
8559 | ||
8560 | capacity += sgc->capacity; | |
8561 | min_capacity = min(sgc->min_capacity, min_capacity); | |
e3d6d0cb | 8562 | max_capacity = max(sgc->max_capacity, max_capacity); |
74a5ce20 PZ |
8563 | group = group->next; |
8564 | } while (group != child->groups); | |
8565 | } | |
1e3c88bd | 8566 | |
63b2ca30 | 8567 | sdg->sgc->capacity = capacity; |
bf475ce0 | 8568 | sdg->sgc->min_capacity = min_capacity; |
e3d6d0cb | 8569 | sdg->sgc->max_capacity = max_capacity; |
1e3c88bd PZ |
8570 | } |
8571 | ||
9d5efe05 | 8572 | /* |
ea67821b VG |
8573 | * Check whether the capacity of the rq has been noticeably reduced by side |
8574 | * activity. The imbalance_pct is used for the threshold. | |
8575 | * Return true is the capacity is reduced | |
9d5efe05 SV |
8576 | */ |
8577 | static inline int | |
ea67821b | 8578 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 8579 | { |
ea67821b VG |
8580 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
8581 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
8582 | } |
8583 | ||
a0fe2cf0 VS |
8584 | /* |
8585 | * Check whether a rq has a misfit task and if it looks like we can actually | |
8586 | * help that task: we can migrate the task to a CPU of higher capacity, or | |
8587 | * the task's current CPU is heavily pressured. | |
8588 | */ | |
8589 | static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd) | |
8590 | { | |
8591 | return rq->misfit_task_load && | |
8592 | (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity || | |
8593 | check_cpu_capacity(rq, sd)); | |
8594 | } | |
8595 | ||
30ce5dab PZ |
8596 | /* |
8597 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
3bd37062 | 8598 | * groups is inadequate due to ->cpus_ptr constraints. |
30ce5dab | 8599 | * |
97fb7a0a IM |
8600 | * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a |
8601 | * cpumask covering 1 CPU of the first group and 3 CPUs of the second group. | |
30ce5dab PZ |
8602 | * Something like: |
8603 | * | |
2b4d5b25 IM |
8604 | * { 0 1 2 3 } { 4 5 6 7 } |
8605 | * * * * * | |
30ce5dab PZ |
8606 | * |
8607 | * If we were to balance group-wise we'd place two tasks in the first group and | |
8608 | * two tasks in the second group. Clearly this is undesired as it will overload | |
97fb7a0a | 8609 | * cpu 3 and leave one of the CPUs in the second group unused. |
30ce5dab PZ |
8610 | * |
8611 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
8612 | * by noticing the lower domain failed to reach balance and had difficulty |
8613 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
8614 | * |
8615 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 8616 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 8617 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
8618 | * to create an effective group imbalance. |
8619 | * | |
8620 | * This is a somewhat tricky proposition since the next run might not find the | |
8621 | * group imbalance and decide the groups need to be balanced again. A most | |
8622 | * subtle and fragile situation. | |
8623 | */ | |
8624 | ||
6263322c | 8625 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 8626 | { |
63b2ca30 | 8627 | return group->sgc->imbalance; |
30ce5dab PZ |
8628 | } |
8629 | ||
b37d9316 | 8630 | /* |
ea67821b VG |
8631 | * group_has_capacity returns true if the group has spare capacity that could |
8632 | * be used by some tasks. | |
fb95a5a0 | 8633 | * We consider that a group has spare capacity if the number of task is |
9e91d61d DE |
8634 | * smaller than the number of CPUs or if the utilization is lower than the |
8635 | * available capacity for CFS tasks. | |
ea67821b VG |
8636 | * For the latter, we use a threshold to stabilize the state, to take into |
8637 | * account the variance of the tasks' load and to return true if the available | |
8638 | * capacity in meaningful for the load balancer. | |
8639 | * As an example, an available capacity of 1% can appear but it doesn't make | |
8640 | * any benefit for the load balance. | |
b37d9316 | 8641 | */ |
ea67821b | 8642 | static inline bool |
57abff06 | 8643 | group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
b37d9316 | 8644 | { |
5e23e474 | 8645 | if (sgs->sum_nr_running < sgs->group_weight) |
ea67821b | 8646 | return true; |
c61037e9 | 8647 | |
070f5e86 VG |
8648 | if ((sgs->group_capacity * imbalance_pct) < |
8649 | (sgs->group_runnable * 100)) | |
8650 | return false; | |
8651 | ||
ea67821b | 8652 | if ((sgs->group_capacity * 100) > |
57abff06 | 8653 | (sgs->group_util * imbalance_pct)) |
ea67821b | 8654 | return true; |
b37d9316 | 8655 | |
ea67821b VG |
8656 | return false; |
8657 | } | |
8658 | ||
8659 | /* | |
8660 | * group_is_overloaded returns true if the group has more tasks than it can | |
8661 | * handle. | |
8662 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
8663 | * with the exact right number of tasks, has no more spare capacity but is not | |
8664 | * overloaded so both group_has_capacity and group_is_overloaded return | |
8665 | * false. | |
8666 | */ | |
8667 | static inline bool | |
57abff06 | 8668 | group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs) |
ea67821b | 8669 | { |
5e23e474 | 8670 | if (sgs->sum_nr_running <= sgs->group_weight) |
ea67821b | 8671 | return false; |
b37d9316 | 8672 | |
ea67821b | 8673 | if ((sgs->group_capacity * 100) < |
57abff06 | 8674 | (sgs->group_util * imbalance_pct)) |
ea67821b | 8675 | return true; |
b37d9316 | 8676 | |
070f5e86 VG |
8677 | if ((sgs->group_capacity * imbalance_pct) < |
8678 | (sgs->group_runnable * 100)) | |
8679 | return true; | |
8680 | ||
ea67821b | 8681 | return false; |
b37d9316 PZ |
8682 | } |
8683 | ||
79a89f92 | 8684 | static inline enum |
57abff06 | 8685 | group_type group_classify(unsigned int imbalance_pct, |
0b0695f2 | 8686 | struct sched_group *group, |
79a89f92 | 8687 | struct sg_lb_stats *sgs) |
caeb178c | 8688 | { |
57abff06 | 8689 | if (group_is_overloaded(imbalance_pct, sgs)) |
caeb178c RR |
8690 | return group_overloaded; |
8691 | ||
8692 | if (sg_imbalanced(group)) | |
8693 | return group_imbalanced; | |
8694 | ||
0b0695f2 VG |
8695 | if (sgs->group_asym_packing) |
8696 | return group_asym_packing; | |
8697 | ||
3b1baa64 MR |
8698 | if (sgs->group_misfit_task_load) |
8699 | return group_misfit_task; | |
8700 | ||
57abff06 | 8701 | if (!group_has_capacity(imbalance_pct, sgs)) |
0b0695f2 VG |
8702 | return group_fully_busy; |
8703 | ||
8704 | return group_has_spare; | |
caeb178c RR |
8705 | } |
8706 | ||
4006a72b RN |
8707 | /** |
8708 | * asym_smt_can_pull_tasks - Check whether the load balancing CPU can pull tasks | |
8709 | * @dst_cpu: Destination CPU of the load balancing | |
8710 | * @sds: Load-balancing data with statistics of the local group | |
8711 | * @sgs: Load-balancing statistics of the candidate busiest group | |
8712 | * @sg: The candidate busiest group | |
8713 | * | |
8714 | * Check the state of the SMT siblings of both @sds::local and @sg and decide | |
8715 | * if @dst_cpu can pull tasks. | |
8716 | * | |
8717 | * If @dst_cpu does not have SMT siblings, it can pull tasks if two or more of | |
8718 | * the SMT siblings of @sg are busy. If only one CPU in @sg is busy, pull tasks | |
8719 | * only if @dst_cpu has higher priority. | |
8720 | * | |
8721 | * If both @dst_cpu and @sg have SMT siblings, and @sg has exactly one more | |
8722 | * busy CPU than @sds::local, let @dst_cpu pull tasks if it has higher priority. | |
8723 | * Bigger imbalances in the number of busy CPUs will be dealt with in | |
8724 | * update_sd_pick_busiest(). | |
8725 | * | |
8726 | * If @sg does not have SMT siblings, only pull tasks if all of the SMT siblings | |
8727 | * of @dst_cpu are idle and @sg has lower priority. | |
a315da5e RD |
8728 | * |
8729 | * Return: true if @dst_cpu can pull tasks, false otherwise. | |
4006a72b RN |
8730 | */ |
8731 | static bool asym_smt_can_pull_tasks(int dst_cpu, struct sd_lb_stats *sds, | |
8732 | struct sg_lb_stats *sgs, | |
8733 | struct sched_group *sg) | |
8734 | { | |
8735 | #ifdef CONFIG_SCHED_SMT | |
8736 | bool local_is_smt, sg_is_smt; | |
8737 | int sg_busy_cpus; | |
8738 | ||
8739 | local_is_smt = sds->local->flags & SD_SHARE_CPUCAPACITY; | |
8740 | sg_is_smt = sg->flags & SD_SHARE_CPUCAPACITY; | |
8741 | ||
8742 | sg_busy_cpus = sgs->group_weight - sgs->idle_cpus; | |
8743 | ||
8744 | if (!local_is_smt) { | |
8745 | /* | |
8746 | * If we are here, @dst_cpu is idle and does not have SMT | |
8747 | * siblings. Pull tasks if candidate group has two or more | |
8748 | * busy CPUs. | |
8749 | */ | |
8750 | if (sg_busy_cpus >= 2) /* implies sg_is_smt */ | |
8751 | return true; | |
8752 | ||
8753 | /* | |
8754 | * @dst_cpu does not have SMT siblings. @sg may have SMT | |
8755 | * siblings and only one is busy. In such case, @dst_cpu | |
8756 | * can help if it has higher priority and is idle (i.e., | |
8757 | * it has no running tasks). | |
8758 | */ | |
8759 | return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu); | |
8760 | } | |
8761 | ||
8762 | /* @dst_cpu has SMT siblings. */ | |
8763 | ||
8764 | if (sg_is_smt) { | |
8765 | int local_busy_cpus = sds->local->group_weight - | |
8766 | sds->local_stat.idle_cpus; | |
8767 | int busy_cpus_delta = sg_busy_cpus - local_busy_cpus; | |
8768 | ||
8769 | if (busy_cpus_delta == 1) | |
8770 | return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu); | |
8771 | ||
8772 | return false; | |
8773 | } | |
8774 | ||
8775 | /* | |
8776 | * @sg does not have SMT siblings. Ensure that @sds::local does not end | |
8777 | * up with more than one busy SMT sibling and only pull tasks if there | |
8778 | * are not busy CPUs (i.e., no CPU has running tasks). | |
8779 | */ | |
8780 | if (!sds->local_stat.sum_nr_running) | |
8781 | return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu); | |
8782 | ||
8783 | return false; | |
8784 | #else | |
8785 | /* Always return false so that callers deal with non-SMT cases. */ | |
8786 | return false; | |
8787 | #endif | |
8788 | } | |
8789 | ||
aafc917a RN |
8790 | static inline bool |
8791 | sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs, | |
8792 | struct sched_group *group) | |
8793 | { | |
4006a72b RN |
8794 | /* Only do SMT checks if either local or candidate have SMT siblings */ |
8795 | if ((sds->local->flags & SD_SHARE_CPUCAPACITY) || | |
8796 | (group->flags & SD_SHARE_CPUCAPACITY)) | |
8797 | return asym_smt_can_pull_tasks(env->dst_cpu, sds, sgs, group); | |
8798 | ||
aafc917a RN |
8799 | return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu); |
8800 | } | |
8801 | ||
c82a6962 VG |
8802 | static inline bool |
8803 | sched_reduced_capacity(struct rq *rq, struct sched_domain *sd) | |
8804 | { | |
8805 | /* | |
8806 | * When there is more than 1 task, the group_overloaded case already | |
8807 | * takes care of cpu with reduced capacity | |
8808 | */ | |
8809 | if (rq->cfs.h_nr_running != 1) | |
8810 | return false; | |
8811 | ||
8812 | return check_cpu_capacity(rq, sd); | |
8813 | } | |
8814 | ||
1e3c88bd PZ |
8815 | /** |
8816 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 8817 | * @env: The load balancing environment. |
a315da5e | 8818 | * @sds: Load-balancing data with statistics of the local group. |
1e3c88bd | 8819 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 8820 | * @sgs: variable to hold the statistics for this group. |
630246a0 | 8821 | * @sg_status: Holds flag indicating the status of the sched_group |
1e3c88bd | 8822 | */ |
bd939f45 | 8823 | static inline void update_sg_lb_stats(struct lb_env *env, |
c0d14b57 | 8824 | struct sd_lb_stats *sds, |
630246a0 QP |
8825 | struct sched_group *group, |
8826 | struct sg_lb_stats *sgs, | |
8827 | int *sg_status) | |
1e3c88bd | 8828 | { |
0b0695f2 | 8829 | int i, nr_running, local_group; |
1e3c88bd | 8830 | |
b72ff13c PZ |
8831 | memset(sgs, 0, sizeof(*sgs)); |
8832 | ||
c0d14b57 | 8833 | local_group = group == sds->local; |
0b0695f2 | 8834 | |
ae4df9d6 | 8835 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
1e3c88bd | 8836 | struct rq *rq = cpu_rq(i); |
c82a6962 | 8837 | unsigned long load = cpu_load(rq); |
1e3c88bd | 8838 | |
c82a6962 | 8839 | sgs->group_load += load; |
82762d2a | 8840 | sgs->group_util += cpu_util_cfs(i); |
070f5e86 | 8841 | sgs->group_runnable += cpu_runnable(rq); |
a3498347 | 8842 | sgs->sum_h_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 8843 | |
a426f99c | 8844 | nr_running = rq->nr_running; |
5e23e474 VG |
8845 | sgs->sum_nr_running += nr_running; |
8846 | ||
a426f99c | 8847 | if (nr_running > 1) |
630246a0 | 8848 | *sg_status |= SG_OVERLOAD; |
4486edd1 | 8849 | |
2802bf3c MR |
8850 | if (cpu_overutilized(i)) |
8851 | *sg_status |= SG_OVERUTILIZED; | |
4486edd1 | 8852 | |
0ec8aa00 PZ |
8853 | #ifdef CONFIG_NUMA_BALANCING |
8854 | sgs->nr_numa_running += rq->nr_numa_running; | |
8855 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
8856 | #endif | |
a426f99c WL |
8857 | /* |
8858 | * No need to call idle_cpu() if nr_running is not 0 | |
8859 | */ | |
0b0695f2 | 8860 | if (!nr_running && idle_cpu(i)) { |
aae6d3dd | 8861 | sgs->idle_cpus++; |
0b0695f2 VG |
8862 | /* Idle cpu can't have misfit task */ |
8863 | continue; | |
8864 | } | |
8865 | ||
8866 | if (local_group) | |
8867 | continue; | |
3b1baa64 | 8868 | |
c82a6962 VG |
8869 | if (env->sd->flags & SD_ASYM_CPUCAPACITY) { |
8870 | /* Check for a misfit task on the cpu */ | |
8871 | if (sgs->group_misfit_task_load < rq->misfit_task_load) { | |
8872 | sgs->group_misfit_task_load = rq->misfit_task_load; | |
8873 | *sg_status |= SG_OVERLOAD; | |
8874 | } | |
8875 | } else if ((env->idle != CPU_NOT_IDLE) && | |
8876 | sched_reduced_capacity(rq, env->sd)) { | |
8877 | /* Check for a task running on a CPU with reduced capacity */ | |
8878 | if (sgs->group_misfit_task_load < load) | |
8879 | sgs->group_misfit_task_load = load; | |
757ffdd7 | 8880 | } |
1e3c88bd PZ |
8881 | } |
8882 | ||
aafc917a RN |
8883 | sgs->group_capacity = group->sgc->capacity; |
8884 | ||
8885 | sgs->group_weight = group->group_weight; | |
8886 | ||
0b0695f2 | 8887 | /* Check if dst CPU is idle and preferred to this group */ |
60256435 | 8888 | if (!local_group && env->sd->flags & SD_ASYM_PACKING && |
aafc917a RN |
8889 | env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running && |
8890 | sched_asym(env, sds, sgs, group)) { | |
0b0695f2 VG |
8891 | sgs->group_asym_packing = 1; |
8892 | } | |
8893 | ||
57abff06 | 8894 | sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs); |
0b0695f2 VG |
8895 | |
8896 | /* Computing avg_load makes sense only when group is overloaded */ | |
8897 | if (sgs->group_type == group_overloaded) | |
8898 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / | |
8899 | sgs->group_capacity; | |
1e3c88bd PZ |
8900 | } |
8901 | ||
532cb4c4 MN |
8902 | /** |
8903 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 8904 | * @env: The load balancing environment. |
532cb4c4 MN |
8905 | * @sds: sched_domain statistics |
8906 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 8907 | * @sgs: sched_group statistics |
532cb4c4 MN |
8908 | * |
8909 | * Determine if @sg is a busier group than the previously selected | |
8910 | * busiest group. | |
e69f6186 YB |
8911 | * |
8912 | * Return: %true if @sg is a busier group than the previously selected | |
8913 | * busiest group. %false otherwise. | |
532cb4c4 | 8914 | */ |
bd939f45 | 8915 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
8916 | struct sd_lb_stats *sds, |
8917 | struct sched_group *sg, | |
bd939f45 | 8918 | struct sg_lb_stats *sgs) |
532cb4c4 | 8919 | { |
caeb178c | 8920 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 8921 | |
0b0695f2 VG |
8922 | /* Make sure that there is at least one task to pull */ |
8923 | if (!sgs->sum_h_nr_running) | |
8924 | return false; | |
8925 | ||
cad68e55 MR |
8926 | /* |
8927 | * Don't try to pull misfit tasks we can't help. | |
8928 | * We can use max_capacity here as reduction in capacity on some | |
8929 | * CPUs in the group should either be possible to resolve | |
8930 | * internally or be covered by avg_load imbalance (eventually). | |
8931 | */ | |
c82a6962 VG |
8932 | if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && |
8933 | (sgs->group_type == group_misfit_task) && | |
4aed8aa4 | 8934 | (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) || |
0b0695f2 | 8935 | sds->local_stat.group_type != group_has_spare)) |
cad68e55 MR |
8936 | return false; |
8937 | ||
caeb178c | 8938 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
8939 | return true; |
8940 | ||
caeb178c RR |
8941 | if (sgs->group_type < busiest->group_type) |
8942 | return false; | |
8943 | ||
9e0994c0 | 8944 | /* |
0b0695f2 VG |
8945 | * The candidate and the current busiest group are the same type of |
8946 | * group. Let check which one is the busiest according to the type. | |
9e0994c0 | 8947 | */ |
9e0994c0 | 8948 | |
0b0695f2 VG |
8949 | switch (sgs->group_type) { |
8950 | case group_overloaded: | |
8951 | /* Select the overloaded group with highest avg_load. */ | |
8952 | if (sgs->avg_load <= busiest->avg_load) | |
8953 | return false; | |
8954 | break; | |
8955 | ||
8956 | case group_imbalanced: | |
8957 | /* | |
8958 | * Select the 1st imbalanced group as we don't have any way to | |
8959 | * choose one more than another. | |
8960 | */ | |
9e0994c0 MR |
8961 | return false; |
8962 | ||
0b0695f2 VG |
8963 | case group_asym_packing: |
8964 | /* Prefer to move from lowest priority CPU's work */ | |
8965 | if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu)) | |
8966 | return false; | |
8967 | break; | |
532cb4c4 | 8968 | |
0b0695f2 VG |
8969 | case group_misfit_task: |
8970 | /* | |
8971 | * If we have more than one misfit sg go with the biggest | |
8972 | * misfit. | |
8973 | */ | |
8974 | if (sgs->group_misfit_task_load < busiest->group_misfit_task_load) | |
8975 | return false; | |
8976 | break; | |
532cb4c4 | 8977 | |
0b0695f2 VG |
8978 | case group_fully_busy: |
8979 | /* | |
8980 | * Select the fully busy group with highest avg_load. In | |
8981 | * theory, there is no need to pull task from such kind of | |
8982 | * group because tasks have all compute capacity that they need | |
8983 | * but we can still improve the overall throughput by reducing | |
8984 | * contention when accessing shared HW resources. | |
8985 | * | |
8986 | * XXX for now avg_load is not computed and always 0 so we | |
8987 | * select the 1st one. | |
8988 | */ | |
8989 | if (sgs->avg_load <= busiest->avg_load) | |
8990 | return false; | |
8991 | break; | |
8992 | ||
8993 | case group_has_spare: | |
8994 | /* | |
5f68eb19 VG |
8995 | * Select not overloaded group with lowest number of idle cpus |
8996 | * and highest number of running tasks. We could also compare | |
8997 | * the spare capacity which is more stable but it can end up | |
8998 | * that the group has less spare capacity but finally more idle | |
0b0695f2 VG |
8999 | * CPUs which means less opportunity to pull tasks. |
9000 | */ | |
5f68eb19 | 9001 | if (sgs->idle_cpus > busiest->idle_cpus) |
0b0695f2 | 9002 | return false; |
5f68eb19 VG |
9003 | else if ((sgs->idle_cpus == busiest->idle_cpus) && |
9004 | (sgs->sum_nr_running <= busiest->sum_nr_running)) | |
9005 | return false; | |
9006 | ||
0b0695f2 | 9007 | break; |
532cb4c4 MN |
9008 | } |
9009 | ||
0b0695f2 VG |
9010 | /* |
9011 | * Candidate sg has no more than one task per CPU and has higher | |
9012 | * per-CPU capacity. Migrating tasks to less capable CPUs may harm | |
9013 | * throughput. Maximize throughput, power/energy consequences are not | |
9014 | * considered. | |
9015 | */ | |
9016 | if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && | |
9017 | (sgs->group_type <= group_fully_busy) && | |
4aed8aa4 | 9018 | (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu)))) |
0b0695f2 VG |
9019 | return false; |
9020 | ||
9021 | return true; | |
532cb4c4 MN |
9022 | } |
9023 | ||
0ec8aa00 PZ |
9024 | #ifdef CONFIG_NUMA_BALANCING |
9025 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
9026 | { | |
a3498347 | 9027 | if (sgs->sum_h_nr_running > sgs->nr_numa_running) |
0ec8aa00 | 9028 | return regular; |
a3498347 | 9029 | if (sgs->sum_h_nr_running > sgs->nr_preferred_running) |
0ec8aa00 PZ |
9030 | return remote; |
9031 | return all; | |
9032 | } | |
9033 | ||
9034 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
9035 | { | |
9036 | if (rq->nr_running > rq->nr_numa_running) | |
9037 | return regular; | |
9038 | if (rq->nr_running > rq->nr_preferred_running) | |
9039 | return remote; | |
9040 | return all; | |
9041 | } | |
9042 | #else | |
9043 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
9044 | { | |
9045 | return all; | |
9046 | } | |
9047 | ||
9048 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
9049 | { | |
9050 | return regular; | |
9051 | } | |
9052 | #endif /* CONFIG_NUMA_BALANCING */ | |
9053 | ||
57abff06 VG |
9054 | |
9055 | struct sg_lb_stats; | |
9056 | ||
3318544b VG |
9057 | /* |
9058 | * task_running_on_cpu - return 1 if @p is running on @cpu. | |
9059 | */ | |
9060 | ||
9061 | static unsigned int task_running_on_cpu(int cpu, struct task_struct *p) | |
9062 | { | |
9063 | /* Task has no contribution or is new */ | |
9064 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) | |
9065 | return 0; | |
9066 | ||
9067 | if (task_on_rq_queued(p)) | |
9068 | return 1; | |
9069 | ||
9070 | return 0; | |
9071 | } | |
9072 | ||
9073 | /** | |
9074 | * idle_cpu_without - would a given CPU be idle without p ? | |
9075 | * @cpu: the processor on which idleness is tested. | |
9076 | * @p: task which should be ignored. | |
9077 | * | |
9078 | * Return: 1 if the CPU would be idle. 0 otherwise. | |
9079 | */ | |
9080 | static int idle_cpu_without(int cpu, struct task_struct *p) | |
9081 | { | |
9082 | struct rq *rq = cpu_rq(cpu); | |
9083 | ||
9084 | if (rq->curr != rq->idle && rq->curr != p) | |
9085 | return 0; | |
9086 | ||
9087 | /* | |
9088 | * rq->nr_running can't be used but an updated version without the | |
9089 | * impact of p on cpu must be used instead. The updated nr_running | |
9090 | * be computed and tested before calling idle_cpu_without(). | |
9091 | */ | |
9092 | ||
9093 | #ifdef CONFIG_SMP | |
126c2092 | 9094 | if (rq->ttwu_pending) |
3318544b VG |
9095 | return 0; |
9096 | #endif | |
9097 | ||
9098 | return 1; | |
9099 | } | |
9100 | ||
57abff06 VG |
9101 | /* |
9102 | * update_sg_wakeup_stats - Update sched_group's statistics for wakeup. | |
3318544b | 9103 | * @sd: The sched_domain level to look for idlest group. |
57abff06 VG |
9104 | * @group: sched_group whose statistics are to be updated. |
9105 | * @sgs: variable to hold the statistics for this group. | |
3318544b | 9106 | * @p: The task for which we look for the idlest group/CPU. |
57abff06 VG |
9107 | */ |
9108 | static inline void update_sg_wakeup_stats(struct sched_domain *sd, | |
9109 | struct sched_group *group, | |
9110 | struct sg_lb_stats *sgs, | |
9111 | struct task_struct *p) | |
9112 | { | |
9113 | int i, nr_running; | |
9114 | ||
9115 | memset(sgs, 0, sizeof(*sgs)); | |
9116 | ||
9117 | for_each_cpu(i, sched_group_span(group)) { | |
9118 | struct rq *rq = cpu_rq(i); | |
3318544b | 9119 | unsigned int local; |
57abff06 | 9120 | |
3318544b | 9121 | sgs->group_load += cpu_load_without(rq, p); |
57abff06 | 9122 | sgs->group_util += cpu_util_without(i, p); |
070f5e86 | 9123 | sgs->group_runnable += cpu_runnable_without(rq, p); |
3318544b VG |
9124 | local = task_running_on_cpu(i, p); |
9125 | sgs->sum_h_nr_running += rq->cfs.h_nr_running - local; | |
57abff06 | 9126 | |
3318544b | 9127 | nr_running = rq->nr_running - local; |
57abff06 VG |
9128 | sgs->sum_nr_running += nr_running; |
9129 | ||
9130 | /* | |
3318544b | 9131 | * No need to call idle_cpu_without() if nr_running is not 0 |
57abff06 | 9132 | */ |
3318544b | 9133 | if (!nr_running && idle_cpu_without(i, p)) |
57abff06 VG |
9134 | sgs->idle_cpus++; |
9135 | ||
57abff06 VG |
9136 | } |
9137 | ||
9138 | /* Check if task fits in the group */ | |
9139 | if (sd->flags & SD_ASYM_CPUCAPACITY && | |
9140 | !task_fits_capacity(p, group->sgc->max_capacity)) { | |
9141 | sgs->group_misfit_task_load = 1; | |
9142 | } | |
9143 | ||
9144 | sgs->group_capacity = group->sgc->capacity; | |
9145 | ||
289de359 VG |
9146 | sgs->group_weight = group->group_weight; |
9147 | ||
57abff06 VG |
9148 | sgs->group_type = group_classify(sd->imbalance_pct, group, sgs); |
9149 | ||
9150 | /* | |
9151 | * Computing avg_load makes sense only when group is fully busy or | |
9152 | * overloaded | |
9153 | */ | |
6c8116c9 TZ |
9154 | if (sgs->group_type == group_fully_busy || |
9155 | sgs->group_type == group_overloaded) | |
57abff06 VG |
9156 | sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / |
9157 | sgs->group_capacity; | |
9158 | } | |
9159 | ||
9160 | static bool update_pick_idlest(struct sched_group *idlest, | |
9161 | struct sg_lb_stats *idlest_sgs, | |
9162 | struct sched_group *group, | |
9163 | struct sg_lb_stats *sgs) | |
9164 | { | |
9165 | if (sgs->group_type < idlest_sgs->group_type) | |
9166 | return true; | |
9167 | ||
9168 | if (sgs->group_type > idlest_sgs->group_type) | |
9169 | return false; | |
9170 | ||
9171 | /* | |
9172 | * The candidate and the current idlest group are the same type of | |
9173 | * group. Let check which one is the idlest according to the type. | |
9174 | */ | |
9175 | ||
9176 | switch (sgs->group_type) { | |
9177 | case group_overloaded: | |
9178 | case group_fully_busy: | |
9179 | /* Select the group with lowest avg_load. */ | |
9180 | if (idlest_sgs->avg_load <= sgs->avg_load) | |
9181 | return false; | |
9182 | break; | |
9183 | ||
9184 | case group_imbalanced: | |
9185 | case group_asym_packing: | |
9186 | /* Those types are not used in the slow wakeup path */ | |
9187 | return false; | |
9188 | ||
9189 | case group_misfit_task: | |
9190 | /* Select group with the highest max capacity */ | |
9191 | if (idlest->sgc->max_capacity >= group->sgc->max_capacity) | |
9192 | return false; | |
9193 | break; | |
9194 | ||
9195 | case group_has_spare: | |
9196 | /* Select group with most idle CPUs */ | |
3edecfef | 9197 | if (idlest_sgs->idle_cpus > sgs->idle_cpus) |
57abff06 | 9198 | return false; |
3edecfef PP |
9199 | |
9200 | /* Select group with lowest group_util */ | |
9201 | if (idlest_sgs->idle_cpus == sgs->idle_cpus && | |
9202 | idlest_sgs->group_util <= sgs->group_util) | |
9203 | return false; | |
9204 | ||
57abff06 VG |
9205 | break; |
9206 | } | |
9207 | ||
9208 | return true; | |
9209 | } | |
9210 | ||
9211 | /* | |
9212 | * find_idlest_group() finds and returns the least busy CPU group within the | |
9213 | * domain. | |
9214 | * | |
9215 | * Assumes p is allowed on at least one CPU in sd. | |
9216 | */ | |
9217 | static struct sched_group * | |
45da2773 | 9218 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) |
57abff06 VG |
9219 | { |
9220 | struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups; | |
9221 | struct sg_lb_stats local_sgs, tmp_sgs; | |
9222 | struct sg_lb_stats *sgs; | |
9223 | unsigned long imbalance; | |
9224 | struct sg_lb_stats idlest_sgs = { | |
9225 | .avg_load = UINT_MAX, | |
9226 | .group_type = group_overloaded, | |
9227 | }; | |
9228 | ||
57abff06 VG |
9229 | do { |
9230 | int local_group; | |
9231 | ||
9232 | /* Skip over this group if it has no CPUs allowed */ | |
9233 | if (!cpumask_intersects(sched_group_span(group), | |
9234 | p->cpus_ptr)) | |
9235 | continue; | |
9236 | ||
97886d9d AL |
9237 | /* Skip over this group if no cookie matched */ |
9238 | if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group)) | |
9239 | continue; | |
9240 | ||
57abff06 VG |
9241 | local_group = cpumask_test_cpu(this_cpu, |
9242 | sched_group_span(group)); | |
9243 | ||
9244 | if (local_group) { | |
9245 | sgs = &local_sgs; | |
9246 | local = group; | |
9247 | } else { | |
9248 | sgs = &tmp_sgs; | |
9249 | } | |
9250 | ||
9251 | update_sg_wakeup_stats(sd, group, sgs, p); | |
9252 | ||
9253 | if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) { | |
9254 | idlest = group; | |
9255 | idlest_sgs = *sgs; | |
9256 | } | |
9257 | ||
9258 | } while (group = group->next, group != sd->groups); | |
9259 | ||
9260 | ||
9261 | /* There is no idlest group to push tasks to */ | |
9262 | if (!idlest) | |
9263 | return NULL; | |
9264 | ||
7ed735c3 VG |
9265 | /* The local group has been skipped because of CPU affinity */ |
9266 | if (!local) | |
9267 | return idlest; | |
9268 | ||
57abff06 VG |
9269 | /* |
9270 | * If the local group is idler than the selected idlest group | |
9271 | * don't try and push the task. | |
9272 | */ | |
9273 | if (local_sgs.group_type < idlest_sgs.group_type) | |
9274 | return NULL; | |
9275 | ||
9276 | /* | |
9277 | * If the local group is busier than the selected idlest group | |
9278 | * try and push the task. | |
9279 | */ | |
9280 | if (local_sgs.group_type > idlest_sgs.group_type) | |
9281 | return idlest; | |
9282 | ||
9283 | switch (local_sgs.group_type) { | |
9284 | case group_overloaded: | |
9285 | case group_fully_busy: | |
5c339005 MG |
9286 | |
9287 | /* Calculate allowed imbalance based on load */ | |
9288 | imbalance = scale_load_down(NICE_0_LOAD) * | |
9289 | (sd->imbalance_pct-100) / 100; | |
9290 | ||
57abff06 VG |
9291 | /* |
9292 | * When comparing groups across NUMA domains, it's possible for | |
9293 | * the local domain to be very lightly loaded relative to the | |
9294 | * remote domains but "imbalance" skews the comparison making | |
9295 | * remote CPUs look much more favourable. When considering | |
9296 | * cross-domain, add imbalance to the load on the remote node | |
9297 | * and consider staying local. | |
9298 | */ | |
9299 | ||
9300 | if ((sd->flags & SD_NUMA) && | |
9301 | ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load)) | |
9302 | return NULL; | |
9303 | ||
9304 | /* | |
9305 | * If the local group is less loaded than the selected | |
9306 | * idlest group don't try and push any tasks. | |
9307 | */ | |
9308 | if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance)) | |
9309 | return NULL; | |
9310 | ||
9311 | if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load) | |
9312 | return NULL; | |
9313 | break; | |
9314 | ||
9315 | case group_imbalanced: | |
9316 | case group_asym_packing: | |
9317 | /* Those type are not used in the slow wakeup path */ | |
9318 | return NULL; | |
9319 | ||
9320 | case group_misfit_task: | |
9321 | /* Select group with the highest max capacity */ | |
9322 | if (local->sgc->max_capacity >= idlest->sgc->max_capacity) | |
9323 | return NULL; | |
9324 | break; | |
9325 | ||
9326 | case group_has_spare: | |
cb29a5c1 | 9327 | #ifdef CONFIG_NUMA |
57abff06 | 9328 | if (sd->flags & SD_NUMA) { |
f5b2eeb4 | 9329 | int imb_numa_nr = sd->imb_numa_nr; |
57abff06 VG |
9330 | #ifdef CONFIG_NUMA_BALANCING |
9331 | int idlest_cpu; | |
9332 | /* | |
9333 | * If there is spare capacity at NUMA, try to select | |
9334 | * the preferred node | |
9335 | */ | |
9336 | if (cpu_to_node(this_cpu) == p->numa_preferred_nid) | |
9337 | return NULL; | |
9338 | ||
9339 | idlest_cpu = cpumask_first(sched_group_span(idlest)); | |
9340 | if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid) | |
9341 | return idlest; | |
cb29a5c1 | 9342 | #endif /* CONFIG_NUMA_BALANCING */ |
57abff06 | 9343 | /* |
2cfb7a1b MG |
9344 | * Otherwise, keep the task close to the wakeup source |
9345 | * and improve locality if the number of running tasks | |
9346 | * would remain below threshold where an imbalance is | |
f5b2eeb4 PN |
9347 | * allowed while accounting for the possibility the |
9348 | * task is pinned to a subset of CPUs. If there is a | |
9349 | * real need of migration, periodic load balance will | |
9350 | * take care of it. | |
57abff06 | 9351 | */ |
f5b2eeb4 | 9352 | if (p->nr_cpus_allowed != NR_CPUS) { |
ec4fc801 | 9353 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask); |
f5b2eeb4 PN |
9354 | |
9355 | cpumask_and(cpus, sched_group_span(local), p->cpus_ptr); | |
9356 | imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr); | |
9357 | } | |
9358 | ||
cb29a5c1 MG |
9359 | imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus); |
9360 | if (!adjust_numa_imbalance(imbalance, | |
9361 | local_sgs.sum_nr_running + 1, | |
f5b2eeb4 | 9362 | imb_numa_nr)) { |
57abff06 | 9363 | return NULL; |
cb29a5c1 | 9364 | } |
57abff06 | 9365 | } |
cb29a5c1 | 9366 | #endif /* CONFIG_NUMA */ |
57abff06 VG |
9367 | |
9368 | /* | |
9369 | * Select group with highest number of idle CPUs. We could also | |
9370 | * compare the utilization which is more stable but it can end | |
9371 | * up that the group has less spare capacity but finally more | |
9372 | * idle CPUs which means more opportunity to run task. | |
9373 | */ | |
9374 | if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus) | |
9375 | return NULL; | |
9376 | break; | |
9377 | } | |
9378 | ||
9379 | return idlest; | |
9380 | } | |
9381 | ||
70fb5ccf CY |
9382 | static void update_idle_cpu_scan(struct lb_env *env, |
9383 | unsigned long sum_util) | |
9384 | { | |
9385 | struct sched_domain_shared *sd_share; | |
9386 | int llc_weight, pct; | |
9387 | u64 x, y, tmp; | |
9388 | /* | |
9389 | * Update the number of CPUs to scan in LLC domain, which could | |
9390 | * be used as a hint in select_idle_cpu(). The update of sd_share | |
9391 | * could be expensive because it is within a shared cache line. | |
9392 | * So the write of this hint only occurs during periodic load | |
9393 | * balancing, rather than CPU_NEWLY_IDLE, because the latter | |
9394 | * can fire way more frequently than the former. | |
9395 | */ | |
9396 | if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE) | |
9397 | return; | |
9398 | ||
9399 | llc_weight = per_cpu(sd_llc_size, env->dst_cpu); | |
9400 | if (env->sd->span_weight != llc_weight) | |
9401 | return; | |
9402 | ||
9403 | sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu)); | |
9404 | if (!sd_share) | |
9405 | return; | |
9406 | ||
9407 | /* | |
9408 | * The number of CPUs to search drops as sum_util increases, when | |
9409 | * sum_util hits 85% or above, the scan stops. | |
9410 | * The reason to choose 85% as the threshold is because this is the | |
9411 | * imbalance_pct(117) when a LLC sched group is overloaded. | |
9412 | * | |
9413 | * let y = SCHED_CAPACITY_SCALE - p * x^2 [1] | |
9414 | * and y'= y / SCHED_CAPACITY_SCALE | |
9415 | * | |
9416 | * x is the ratio of sum_util compared to the CPU capacity: | |
9417 | * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE) | |
9418 | * y' is the ratio of CPUs to be scanned in the LLC domain, | |
9419 | * and the number of CPUs to scan is calculated by: | |
9420 | * | |
9421 | * nr_scan = llc_weight * y' [2] | |
9422 | * | |
9423 | * When x hits the threshold of overloaded, AKA, when | |
9424 | * x = 100 / pct, y drops to 0. According to [1], | |
9425 | * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000 | |
9426 | * | |
9427 | * Scale x by SCHED_CAPACITY_SCALE: | |
9428 | * x' = sum_util / llc_weight; [3] | |
9429 | * | |
9430 | * and finally [1] becomes: | |
9431 | * y = SCHED_CAPACITY_SCALE - | |
9432 | * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4] | |
9433 | * | |
9434 | */ | |
9435 | /* equation [3] */ | |
9436 | x = sum_util; | |
9437 | do_div(x, llc_weight); | |
9438 | ||
9439 | /* equation [4] */ | |
9440 | pct = env->sd->imbalance_pct; | |
9441 | tmp = x * x * pct * pct; | |
9442 | do_div(tmp, 10000 * SCHED_CAPACITY_SCALE); | |
9443 | tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE); | |
9444 | y = SCHED_CAPACITY_SCALE - tmp; | |
9445 | ||
9446 | /* equation [2] */ | |
9447 | y *= llc_weight; | |
9448 | do_div(y, SCHED_CAPACITY_SCALE); | |
9449 | if ((int)y != sd_share->nr_idle_scan) | |
9450 | WRITE_ONCE(sd_share->nr_idle_scan, (int)y); | |
9451 | } | |
9452 | ||
1e3c88bd | 9453 | /** |
461819ac | 9454 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 9455 | * @env: The load balancing environment. |
1e3c88bd PZ |
9456 | * @sds: variable to hold the statistics for this sched_domain. |
9457 | */ | |
0b0695f2 | 9458 | |
0ec8aa00 | 9459 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 9460 | { |
bd939f45 PZ |
9461 | struct sched_domain *child = env->sd->child; |
9462 | struct sched_group *sg = env->sd->groups; | |
05b40e05 | 9463 | struct sg_lb_stats *local = &sds->local_stat; |
56cf515b | 9464 | struct sg_lb_stats tmp_sgs; |
70fb5ccf | 9465 | unsigned long sum_util = 0; |
630246a0 | 9466 | int sg_status = 0; |
1e3c88bd | 9467 | |
1e3c88bd | 9468 | do { |
56cf515b | 9469 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
9470 | int local_group; |
9471 | ||
ae4df9d6 | 9472 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); |
56cf515b JK |
9473 | if (local_group) { |
9474 | sds->local = sg; | |
05b40e05 | 9475 | sgs = local; |
b72ff13c PZ |
9476 | |
9477 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
9478 | time_after_eq(jiffies, sg->sgc->next_update)) |
9479 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 9480 | } |
1e3c88bd | 9481 | |
c0d14b57 | 9482 | update_sg_lb_stats(env, sds, sg, sgs, &sg_status); |
1e3c88bd | 9483 | |
b72ff13c PZ |
9484 | if (local_group) |
9485 | goto next_group; | |
9486 | ||
1e3c88bd | 9487 | |
b72ff13c | 9488 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 9489 | sds->busiest = sg; |
56cf515b | 9490 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
9491 | } |
9492 | ||
b72ff13c PZ |
9493 | next_group: |
9494 | /* Now, start updating sd_lb_stats */ | |
9495 | sds->total_load += sgs->group_load; | |
63b2ca30 | 9496 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 9497 | |
70fb5ccf | 9498 | sum_util += sgs->group_util; |
532cb4c4 | 9499 | sg = sg->next; |
bd939f45 | 9500 | } while (sg != env->sd->groups); |
0ec8aa00 | 9501 | |
0b0695f2 VG |
9502 | /* Tag domain that child domain prefers tasks go to siblings first */ |
9503 | sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING; | |
9504 | ||
f643ea22 | 9505 | |
0ec8aa00 PZ |
9506 | if (env->sd->flags & SD_NUMA) |
9507 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
9508 | |
9509 | if (!env->sd->parent) { | |
2802bf3c MR |
9510 | struct root_domain *rd = env->dst_rq->rd; |
9511 | ||
4486edd1 | 9512 | /* update overload indicator if we are at root domain */ |
2802bf3c MR |
9513 | WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD); |
9514 | ||
9515 | /* Update over-utilization (tipping point, U >= 0) indicator */ | |
9516 | WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED); | |
f9f240f9 | 9517 | trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED); |
2802bf3c | 9518 | } else if (sg_status & SG_OVERUTILIZED) { |
f9f240f9 QY |
9519 | struct root_domain *rd = env->dst_rq->rd; |
9520 | ||
9521 | WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED); | |
9522 | trace_sched_overutilized_tp(rd, SG_OVERUTILIZED); | |
4486edd1 | 9523 | } |
70fb5ccf CY |
9524 | |
9525 | update_idle_cpu_scan(env, sum_util); | |
532cb4c4 MN |
9526 | } |
9527 | ||
1e3c88bd PZ |
9528 | /** |
9529 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
9530 | * groups of a given sched_domain during load balance. | |
bd939f45 | 9531 | * @env: load balance environment |
1e3c88bd | 9532 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 9533 | */ |
bd939f45 | 9534 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 9535 | { |
56cf515b JK |
9536 | struct sg_lb_stats *local, *busiest; |
9537 | ||
9538 | local = &sds->local_stat; | |
56cf515b | 9539 | busiest = &sds->busiest_stat; |
dd5feea1 | 9540 | |
0b0695f2 | 9541 | if (busiest->group_type == group_misfit_task) { |
c82a6962 VG |
9542 | if (env->sd->flags & SD_ASYM_CPUCAPACITY) { |
9543 | /* Set imbalance to allow misfit tasks to be balanced. */ | |
9544 | env->migration_type = migrate_misfit; | |
9545 | env->imbalance = 1; | |
9546 | } else { | |
9547 | /* | |
9548 | * Set load imbalance to allow moving task from cpu | |
9549 | * with reduced capacity. | |
9550 | */ | |
9551 | env->migration_type = migrate_load; | |
9552 | env->imbalance = busiest->group_misfit_task_load; | |
9553 | } | |
0b0695f2 VG |
9554 | return; |
9555 | } | |
9556 | ||
9557 | if (busiest->group_type == group_asym_packing) { | |
9558 | /* | |
9559 | * In case of asym capacity, we will try to migrate all load to | |
9560 | * the preferred CPU. | |
9561 | */ | |
9562 | env->migration_type = migrate_task; | |
9563 | env->imbalance = busiest->sum_h_nr_running; | |
9564 | return; | |
9565 | } | |
9566 | ||
9567 | if (busiest->group_type == group_imbalanced) { | |
9568 | /* | |
9569 | * In the group_imb case we cannot rely on group-wide averages | |
9570 | * to ensure CPU-load equilibrium, try to move any task to fix | |
9571 | * the imbalance. The next load balance will take care of | |
9572 | * balancing back the system. | |
9573 | */ | |
9574 | env->migration_type = migrate_task; | |
9575 | env->imbalance = 1; | |
490ba971 VG |
9576 | return; |
9577 | } | |
9578 | ||
1e3c88bd | 9579 | /* |
0b0695f2 | 9580 | * Try to use spare capacity of local group without overloading it or |
a9723389 | 9581 | * emptying busiest. |
1e3c88bd | 9582 | */ |
0b0695f2 | 9583 | if (local->group_type == group_has_spare) { |
16b0a7a1 VG |
9584 | if ((busiest->group_type > group_fully_busy) && |
9585 | !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) { | |
0b0695f2 VG |
9586 | /* |
9587 | * If busiest is overloaded, try to fill spare | |
9588 | * capacity. This might end up creating spare capacity | |
9589 | * in busiest or busiest still being overloaded but | |
9590 | * there is no simple way to directly compute the | |
9591 | * amount of load to migrate in order to balance the | |
9592 | * system. | |
9593 | */ | |
9594 | env->migration_type = migrate_util; | |
9595 | env->imbalance = max(local->group_capacity, local->group_util) - | |
9596 | local->group_util; | |
9597 | ||
9598 | /* | |
9599 | * In some cases, the group's utilization is max or even | |
9600 | * higher than capacity because of migrations but the | |
9601 | * local CPU is (newly) idle. There is at least one | |
9602 | * waiting task in this overloaded busiest group. Let's | |
9603 | * try to pull it. | |
9604 | */ | |
9605 | if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) { | |
9606 | env->migration_type = migrate_task; | |
9607 | env->imbalance = 1; | |
9608 | } | |
9609 | ||
9610 | return; | |
9611 | } | |
9612 | ||
9613 | if (busiest->group_weight == 1 || sds->prefer_sibling) { | |
5e23e474 | 9614 | unsigned int nr_diff = busiest->sum_nr_running; |
0b0695f2 VG |
9615 | /* |
9616 | * When prefer sibling, evenly spread running tasks on | |
9617 | * groups. | |
9618 | */ | |
9619 | env->migration_type = migrate_task; | |
5e23e474 | 9620 | lsub_positive(&nr_diff, local->sum_nr_running); |
cb29a5c1 | 9621 | env->imbalance = nr_diff; |
b396f523 | 9622 | } else { |
0b0695f2 | 9623 | |
b396f523 MG |
9624 | /* |
9625 | * If there is no overload, we just want to even the number of | |
9626 | * idle cpus. | |
9627 | */ | |
9628 | env->migration_type = migrate_task; | |
cb29a5c1 MG |
9629 | env->imbalance = max_t(long, 0, |
9630 | (local->idle_cpus - busiest->idle_cpus)); | |
b396f523 MG |
9631 | } |
9632 | ||
cb29a5c1 | 9633 | #ifdef CONFIG_NUMA |
b396f523 | 9634 | /* Consider allowing a small imbalance between NUMA groups */ |
7d2b5dd0 | 9635 | if (env->sd->flags & SD_NUMA) { |
fb86f5b2 | 9636 | env->imbalance = adjust_numa_imbalance(env->imbalance, |
cb29a5c1 MG |
9637 | local->sum_nr_running + 1, |
9638 | env->sd->imb_numa_nr); | |
7d2b5dd0 | 9639 | } |
cb29a5c1 MG |
9640 | #endif |
9641 | ||
9642 | /* Number of tasks to move to restore balance */ | |
9643 | env->imbalance >>= 1; | |
b396f523 | 9644 | |
fcf0553d | 9645 | return; |
1e3c88bd PZ |
9646 | } |
9647 | ||
9a5d9ba6 | 9648 | /* |
0b0695f2 VG |
9649 | * Local is fully busy but has to take more load to relieve the |
9650 | * busiest group | |
9a5d9ba6 | 9651 | */ |
0b0695f2 VG |
9652 | if (local->group_type < group_overloaded) { |
9653 | /* | |
9654 | * Local will become overloaded so the avg_load metrics are | |
9655 | * finally needed. | |
9656 | */ | |
9657 | ||
9658 | local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) / | |
9659 | local->group_capacity; | |
9660 | ||
111688ca AL |
9661 | /* |
9662 | * If the local group is more loaded than the selected | |
9663 | * busiest group don't try to pull any tasks. | |
9664 | */ | |
9665 | if (local->avg_load >= busiest->avg_load) { | |
9666 | env->imbalance = 0; | |
9667 | return; | |
9668 | } | |
06354900 | 9669 | |
9670 | sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) / | |
9671 | sds->total_capacity; | |
dd5feea1 SS |
9672 | } |
9673 | ||
9674 | /* | |
0b0695f2 VG |
9675 | * Both group are or will become overloaded and we're trying to get all |
9676 | * the CPUs to the average_load, so we don't want to push ourselves | |
9677 | * above the average load, nor do we wish to reduce the max loaded CPU | |
9678 | * below the average load. At the same time, we also don't want to | |
9679 | * reduce the group load below the group capacity. Thus we look for | |
9680 | * the minimum possible imbalance. | |
dd5feea1 | 9681 | */ |
0b0695f2 | 9682 | env->migration_type = migrate_load; |
56cf515b | 9683 | env->imbalance = min( |
0b0695f2 | 9684 | (busiest->avg_load - sds->avg_load) * busiest->group_capacity, |
63b2ca30 | 9685 | (sds->avg_load - local->avg_load) * local->group_capacity |
ca8ce3d0 | 9686 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd | 9687 | } |
fab47622 | 9688 | |
1e3c88bd PZ |
9689 | /******* find_busiest_group() helpers end here *********************/ |
9690 | ||
0b0695f2 VG |
9691 | /* |
9692 | * Decision matrix according to the local and busiest group type: | |
9693 | * | |
9694 | * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded | |
9695 | * has_spare nr_idle balanced N/A N/A balanced balanced | |
9696 | * fully_busy nr_idle nr_idle N/A N/A balanced balanced | |
a6583531 | 9697 | * misfit_task force N/A N/A N/A N/A N/A |
0b0695f2 VG |
9698 | * asym_packing force force N/A N/A force force |
9699 | * imbalanced force force N/A N/A force force | |
9700 | * overloaded force force N/A N/A force avg_load | |
9701 | * | |
9702 | * N/A : Not Applicable because already filtered while updating | |
9703 | * statistics. | |
9704 | * balanced : The system is balanced for these 2 groups. | |
9705 | * force : Calculate the imbalance as load migration is probably needed. | |
9706 | * avg_load : Only if imbalance is significant enough. | |
9707 | * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite | |
9708 | * different in groups. | |
9709 | */ | |
9710 | ||
1e3c88bd PZ |
9711 | /** |
9712 | * find_busiest_group - Returns the busiest group within the sched_domain | |
0a9b23ce | 9713 | * if there is an imbalance. |
a315da5e | 9714 | * @env: The load balancing environment. |
1e3c88bd | 9715 | * |
a3df0679 | 9716 | * Also calculates the amount of runnable load which should be moved |
1e3c88bd PZ |
9717 | * to restore balance. |
9718 | * | |
e69f6186 | 9719 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 9720 | */ |
56cf515b | 9721 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 9722 | { |
56cf515b | 9723 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
9724 | struct sd_lb_stats sds; |
9725 | ||
147c5fc2 | 9726 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
9727 | |
9728 | /* | |
b0fb1eb4 | 9729 | * Compute the various statistics relevant for load balancing at |
1e3c88bd PZ |
9730 | * this level. |
9731 | */ | |
23f0d209 | 9732 | update_sd_lb_stats(env, &sds); |
2802bf3c | 9733 | |
f8a696f2 | 9734 | if (sched_energy_enabled()) { |
2802bf3c MR |
9735 | struct root_domain *rd = env->dst_rq->rd; |
9736 | ||
9737 | if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)) | |
9738 | goto out_balanced; | |
9739 | } | |
9740 | ||
56cf515b JK |
9741 | local = &sds.local_stat; |
9742 | busiest = &sds.busiest_stat; | |
1e3c88bd | 9743 | |
cc57aa8f | 9744 | /* There is no busy sibling group to pull tasks from */ |
0b0695f2 | 9745 | if (!sds.busiest) |
1e3c88bd PZ |
9746 | goto out_balanced; |
9747 | ||
0b0695f2 VG |
9748 | /* Misfit tasks should be dealt with regardless of the avg load */ |
9749 | if (busiest->group_type == group_misfit_task) | |
9750 | goto force_balance; | |
9751 | ||
9752 | /* ASYM feature bypasses nice load balance check */ | |
9753 | if (busiest->group_type == group_asym_packing) | |
9754 | goto force_balance; | |
b0432d8f | 9755 | |
866ab43e PZ |
9756 | /* |
9757 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 9758 | * work because they assume all things are equal, which typically |
3bd37062 | 9759 | * isn't true due to cpus_ptr constraints and the like. |
866ab43e | 9760 | */ |
caeb178c | 9761 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
9762 | goto force_balance; |
9763 | ||
cc57aa8f | 9764 | /* |
9c58c79a | 9765 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
9766 | * don't try and pull any tasks. |
9767 | */ | |
0b0695f2 | 9768 | if (local->group_type > busiest->group_type) |
1e3c88bd PZ |
9769 | goto out_balanced; |
9770 | ||
cc57aa8f | 9771 | /* |
0b0695f2 VG |
9772 | * When groups are overloaded, use the avg_load to ensure fairness |
9773 | * between tasks. | |
cc57aa8f | 9774 | */ |
0b0695f2 VG |
9775 | if (local->group_type == group_overloaded) { |
9776 | /* | |
9777 | * If the local group is more loaded than the selected | |
9778 | * busiest group don't try to pull any tasks. | |
9779 | */ | |
9780 | if (local->avg_load >= busiest->avg_load) | |
9781 | goto out_balanced; | |
9782 | ||
9783 | /* XXX broken for overlapping NUMA groups */ | |
9784 | sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) / | |
9785 | sds.total_capacity; | |
1e3c88bd | 9786 | |
aae6d3dd | 9787 | /* |
0b0695f2 VG |
9788 | * Don't pull any tasks if this group is already above the |
9789 | * domain average load. | |
aae6d3dd | 9790 | */ |
0b0695f2 | 9791 | if (local->avg_load >= sds.avg_load) |
aae6d3dd | 9792 | goto out_balanced; |
0b0695f2 | 9793 | |
c186fafe | 9794 | /* |
0b0695f2 VG |
9795 | * If the busiest group is more loaded, use imbalance_pct to be |
9796 | * conservative. | |
c186fafe | 9797 | */ |
56cf515b JK |
9798 | if (100 * busiest->avg_load <= |
9799 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 9800 | goto out_balanced; |
aae6d3dd | 9801 | } |
1e3c88bd | 9802 | |
0b0695f2 VG |
9803 | /* Try to move all excess tasks to child's sibling domain */ |
9804 | if (sds.prefer_sibling && local->group_type == group_has_spare && | |
5e23e474 | 9805 | busiest->sum_nr_running > local->sum_nr_running + 1) |
0b0695f2 VG |
9806 | goto force_balance; |
9807 | ||
2ab4092f VG |
9808 | if (busiest->group_type != group_overloaded) { |
9809 | if (env->idle == CPU_NOT_IDLE) | |
9810 | /* | |
9811 | * If the busiest group is not overloaded (and as a | |
9812 | * result the local one too) but this CPU is already | |
9813 | * busy, let another idle CPU try to pull task. | |
9814 | */ | |
9815 | goto out_balanced; | |
9816 | ||
9817 | if (busiest->group_weight > 1 && | |
9818 | local->idle_cpus <= (busiest->idle_cpus + 1)) | |
9819 | /* | |
9820 | * If the busiest group is not overloaded | |
9821 | * and there is no imbalance between this and busiest | |
9822 | * group wrt idle CPUs, it is balanced. The imbalance | |
9823 | * becomes significant if the diff is greater than 1 | |
9824 | * otherwise we might end up to just move the imbalance | |
9825 | * on another group. Of course this applies only if | |
9826 | * there is more than 1 CPU per group. | |
9827 | */ | |
9828 | goto out_balanced; | |
9829 | ||
9830 | if (busiest->sum_h_nr_running == 1) | |
9831 | /* | |
9832 | * busiest doesn't have any tasks waiting to run | |
9833 | */ | |
9834 | goto out_balanced; | |
9835 | } | |
0b0695f2 | 9836 | |
fab47622 | 9837 | force_balance: |
1e3c88bd | 9838 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 9839 | calculate_imbalance(env, &sds); |
bb3485c8 | 9840 | return env->imbalance ? sds.busiest : NULL; |
1e3c88bd PZ |
9841 | |
9842 | out_balanced: | |
bd939f45 | 9843 | env->imbalance = 0; |
1e3c88bd PZ |
9844 | return NULL; |
9845 | } | |
9846 | ||
9847 | /* | |
97fb7a0a | 9848 | * find_busiest_queue - find the busiest runqueue among the CPUs in the group. |
1e3c88bd | 9849 | */ |
bd939f45 | 9850 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 9851 | struct sched_group *group) |
1e3c88bd PZ |
9852 | { |
9853 | struct rq *busiest = NULL, *rq; | |
0b0695f2 VG |
9854 | unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1; |
9855 | unsigned int busiest_nr = 0; | |
1e3c88bd PZ |
9856 | int i; |
9857 | ||
ae4df9d6 | 9858 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
0b0695f2 VG |
9859 | unsigned long capacity, load, util; |
9860 | unsigned int nr_running; | |
0ec8aa00 PZ |
9861 | enum fbq_type rt; |
9862 | ||
9863 | rq = cpu_rq(i); | |
9864 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 9865 | |
0ec8aa00 PZ |
9866 | /* |
9867 | * We classify groups/runqueues into three groups: | |
9868 | * - regular: there are !numa tasks | |
9869 | * - remote: there are numa tasks that run on the 'wrong' node | |
9870 | * - all: there is no distinction | |
9871 | * | |
9872 | * In order to avoid migrating ideally placed numa tasks, | |
9873 | * ignore those when there's better options. | |
9874 | * | |
9875 | * If we ignore the actual busiest queue to migrate another | |
9876 | * task, the next balance pass can still reduce the busiest | |
9877 | * queue by moving tasks around inside the node. | |
9878 | * | |
9879 | * If we cannot move enough load due to this classification | |
9880 | * the next pass will adjust the group classification and | |
9881 | * allow migration of more tasks. | |
9882 | * | |
9883 | * Both cases only affect the total convergence complexity. | |
9884 | */ | |
9885 | if (rt > env->fbq_type) | |
9886 | continue; | |
9887 | ||
0b0695f2 | 9888 | nr_running = rq->cfs.h_nr_running; |
fc488ffd VG |
9889 | if (!nr_running) |
9890 | continue; | |
9891 | ||
9892 | capacity = capacity_of(i); | |
9d5efe05 | 9893 | |
4ad3831a CR |
9894 | /* |
9895 | * For ASYM_CPUCAPACITY domains, don't pick a CPU that could | |
9896 | * eventually lead to active_balancing high->low capacity. | |
9897 | * Higher per-CPU capacity is considered better than balancing | |
9898 | * average load. | |
9899 | */ | |
9900 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && | |
4aed8aa4 | 9901 | !capacity_greater(capacity_of(env->dst_cpu), capacity) && |
0b0695f2 | 9902 | nr_running == 1) |
4ad3831a CR |
9903 | continue; |
9904 | ||
4006a72b RN |
9905 | /* Make sure we only pull tasks from a CPU of lower priority */ |
9906 | if ((env->sd->flags & SD_ASYM_PACKING) && | |
9907 | sched_asym_prefer(i, env->dst_cpu) && | |
9908 | nr_running == 1) | |
9909 | continue; | |
9910 | ||
0b0695f2 VG |
9911 | switch (env->migration_type) { |
9912 | case migrate_load: | |
9913 | /* | |
b0fb1eb4 VG |
9914 | * When comparing with load imbalance, use cpu_load() |
9915 | * which is not scaled with the CPU capacity. | |
0b0695f2 | 9916 | */ |
b0fb1eb4 | 9917 | load = cpu_load(rq); |
1e3c88bd | 9918 | |
0b0695f2 VG |
9919 | if (nr_running == 1 && load > env->imbalance && |
9920 | !check_cpu_capacity(rq, env->sd)) | |
9921 | break; | |
ea67821b | 9922 | |
0b0695f2 VG |
9923 | /* |
9924 | * For the load comparisons with the other CPUs, | |
b0fb1eb4 VG |
9925 | * consider the cpu_load() scaled with the CPU |
9926 | * capacity, so that the load can be moved away | |
9927 | * from the CPU that is potentially running at a | |
9928 | * lower capacity. | |
0b0695f2 VG |
9929 | * |
9930 | * Thus we're looking for max(load_i / capacity_i), | |
9931 | * crosswise multiplication to rid ourselves of the | |
9932 | * division works out to: | |
9933 | * load_i * capacity_j > load_j * capacity_i; | |
9934 | * where j is our previous maximum. | |
9935 | */ | |
9936 | if (load * busiest_capacity > busiest_load * capacity) { | |
9937 | busiest_load = load; | |
9938 | busiest_capacity = capacity; | |
9939 | busiest = rq; | |
9940 | } | |
9941 | break; | |
9942 | ||
9943 | case migrate_util: | |
82762d2a | 9944 | util = cpu_util_cfs(i); |
0b0695f2 | 9945 | |
c32b4308 VG |
9946 | /* |
9947 | * Don't try to pull utilization from a CPU with one | |
9948 | * running task. Whatever its utilization, we will fail | |
9949 | * detach the task. | |
9950 | */ | |
9951 | if (nr_running <= 1) | |
9952 | continue; | |
9953 | ||
0b0695f2 VG |
9954 | if (busiest_util < util) { |
9955 | busiest_util = util; | |
9956 | busiest = rq; | |
9957 | } | |
9958 | break; | |
9959 | ||
9960 | case migrate_task: | |
9961 | if (busiest_nr < nr_running) { | |
9962 | busiest_nr = nr_running; | |
9963 | busiest = rq; | |
9964 | } | |
9965 | break; | |
9966 | ||
9967 | case migrate_misfit: | |
9968 | /* | |
9969 | * For ASYM_CPUCAPACITY domains with misfit tasks we | |
9970 | * simply seek the "biggest" misfit task. | |
9971 | */ | |
9972 | if (rq->misfit_task_load > busiest_load) { | |
9973 | busiest_load = rq->misfit_task_load; | |
9974 | busiest = rq; | |
9975 | } | |
9976 | ||
9977 | break; | |
1e3c88bd | 9978 | |
1e3c88bd PZ |
9979 | } |
9980 | } | |
9981 | ||
9982 | return busiest; | |
9983 | } | |
9984 | ||
9985 | /* | |
9986 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
9987 | * so long as it is large enough. | |
9988 | */ | |
9989 | #define MAX_PINNED_INTERVAL 512 | |
9990 | ||
46a745d9 VG |
9991 | static inline bool |
9992 | asym_active_balance(struct lb_env *env) | |
1af3ed3d | 9993 | { |
46a745d9 VG |
9994 | /* |
9995 | * ASYM_PACKING needs to force migrate tasks from busy but | |
9996 | * lower priority CPUs in order to pack all tasks in the | |
9997 | * highest priority CPUs. | |
9998 | */ | |
9999 | return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) && | |
10000 | sched_asym_prefer(env->dst_cpu, env->src_cpu); | |
10001 | } | |
bd939f45 | 10002 | |
46a745d9 | 10003 | static inline bool |
e9b9734b VG |
10004 | imbalanced_active_balance(struct lb_env *env) |
10005 | { | |
10006 | struct sched_domain *sd = env->sd; | |
10007 | ||
10008 | /* | |
10009 | * The imbalanced case includes the case of pinned tasks preventing a fair | |
10010 | * distribution of the load on the system but also the even distribution of the | |
10011 | * threads on a system with spare capacity | |
10012 | */ | |
10013 | if ((env->migration_type == migrate_task) && | |
10014 | (sd->nr_balance_failed > sd->cache_nice_tries+2)) | |
10015 | return 1; | |
10016 | ||
10017 | return 0; | |
10018 | } | |
10019 | ||
10020 | static int need_active_balance(struct lb_env *env) | |
46a745d9 VG |
10021 | { |
10022 | struct sched_domain *sd = env->sd; | |
532cb4c4 | 10023 | |
46a745d9 VG |
10024 | if (asym_active_balance(env)) |
10025 | return 1; | |
1af3ed3d | 10026 | |
e9b9734b VG |
10027 | if (imbalanced_active_balance(env)) |
10028 | return 1; | |
10029 | ||
1aaf90a4 VG |
10030 | /* |
10031 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
10032 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
10033 | * because of other sched_class or IRQs if more capacity stays | |
10034 | * available on dst_cpu. | |
10035 | */ | |
10036 | if ((env->idle != CPU_NOT_IDLE) && | |
10037 | (env->src_rq->cfs.h_nr_running == 1)) { | |
10038 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
10039 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
10040 | return 1; | |
10041 | } | |
10042 | ||
0b0695f2 | 10043 | if (env->migration_type == migrate_misfit) |
cad68e55 MR |
10044 | return 1; |
10045 | ||
46a745d9 VG |
10046 | return 0; |
10047 | } | |
10048 | ||
969c7921 TH |
10049 | static int active_load_balance_cpu_stop(void *data); |
10050 | ||
23f0d209 JK |
10051 | static int should_we_balance(struct lb_env *env) |
10052 | { | |
10053 | struct sched_group *sg = env->sd->groups; | |
64297f2b | 10054 | int cpu; |
23f0d209 | 10055 | |
024c9d2f PZ |
10056 | /* |
10057 | * Ensure the balancing environment is consistent; can happen | |
10058 | * when the softirq triggers 'during' hotplug. | |
10059 | */ | |
10060 | if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) | |
10061 | return 0; | |
10062 | ||
23f0d209 | 10063 | /* |
97fb7a0a | 10064 | * In the newly idle case, we will allow all the CPUs |
23f0d209 | 10065 | * to do the newly idle load balance. |
792b9f65 JD |
10066 | * |
10067 | * However, we bail out if we already have tasks or a wakeup pending, | |
10068 | * to optimize wakeup latency. | |
23f0d209 | 10069 | */ |
792b9f65 JD |
10070 | if (env->idle == CPU_NEWLY_IDLE) { |
10071 | if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending) | |
10072 | return 0; | |
23f0d209 | 10073 | return 1; |
792b9f65 | 10074 | } |
23f0d209 | 10075 | |
97fb7a0a | 10076 | /* Try to find first idle CPU */ |
e5c14b1f | 10077 | for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { |
af218122 | 10078 | if (!idle_cpu(cpu)) |
23f0d209 JK |
10079 | continue; |
10080 | ||
64297f2b PW |
10081 | /* Are we the first idle CPU? */ |
10082 | return cpu == env->dst_cpu; | |
23f0d209 JK |
10083 | } |
10084 | ||
64297f2b PW |
10085 | /* Are we the first CPU of this group ? */ |
10086 | return group_balance_cpu(sg) == env->dst_cpu; | |
23f0d209 JK |
10087 | } |
10088 | ||
1e3c88bd PZ |
10089 | /* |
10090 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
10091 | * tasks if there is an imbalance. | |
10092 | */ | |
10093 | static int load_balance(int this_cpu, struct rq *this_rq, | |
10094 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 10095 | int *continue_balancing) |
1e3c88bd | 10096 | { |
88b8dac0 | 10097 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 10098 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 10099 | struct sched_group *group; |
1e3c88bd | 10100 | struct rq *busiest; |
8a8c69c3 | 10101 | struct rq_flags rf; |
4ba29684 | 10102 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 10103 | |
8e45cb54 PZ |
10104 | struct lb_env env = { |
10105 | .sd = sd, | |
ddcdf6e7 PZ |
10106 | .dst_cpu = this_cpu, |
10107 | .dst_rq = this_rq, | |
ae4df9d6 | 10108 | .dst_grpmask = sched_group_span(sd->groups), |
8e45cb54 | 10109 | .idle = idle, |
eb95308e | 10110 | .loop_break = sched_nr_migrate_break, |
b9403130 | 10111 | .cpus = cpus, |
0ec8aa00 | 10112 | .fbq_type = all, |
163122b7 | 10113 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
10114 | }; |
10115 | ||
65a4433a | 10116 | cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); |
1e3c88bd | 10117 | |
ae92882e | 10118 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
10119 | |
10120 | redo: | |
23f0d209 JK |
10121 | if (!should_we_balance(&env)) { |
10122 | *continue_balancing = 0; | |
1e3c88bd | 10123 | goto out_balanced; |
23f0d209 | 10124 | } |
1e3c88bd | 10125 | |
23f0d209 | 10126 | group = find_busiest_group(&env); |
1e3c88bd | 10127 | if (!group) { |
ae92882e | 10128 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
10129 | goto out_balanced; |
10130 | } | |
10131 | ||
b9403130 | 10132 | busiest = find_busiest_queue(&env, group); |
1e3c88bd | 10133 | if (!busiest) { |
ae92882e | 10134 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
10135 | goto out_balanced; |
10136 | } | |
10137 | ||
78feefc5 | 10138 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 10139 | |
ae92882e | 10140 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 10141 | |
1aaf90a4 VG |
10142 | env.src_cpu = busiest->cpu; |
10143 | env.src_rq = busiest; | |
10144 | ||
1e3c88bd | 10145 | ld_moved = 0; |
8a41dfcd VG |
10146 | /* Clear this flag as soon as we find a pullable task */ |
10147 | env.flags |= LBF_ALL_PINNED; | |
1e3c88bd PZ |
10148 | if (busiest->nr_running > 1) { |
10149 | /* | |
10150 | * Attempt to move tasks. If find_busiest_group has found | |
10151 | * an imbalance but busiest->nr_running <= 1, the group is | |
10152 | * still unbalanced. ld_moved simply stays zero, so it is | |
10153 | * correctly treated as an imbalance. | |
10154 | */ | |
c82513e5 | 10155 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 10156 | |
5d6523eb | 10157 | more_balance: |
8a8c69c3 | 10158 | rq_lock_irqsave(busiest, &rf); |
3bed5e21 | 10159 | update_rq_clock(busiest); |
88b8dac0 SV |
10160 | |
10161 | /* | |
10162 | * cur_ld_moved - load moved in current iteration | |
10163 | * ld_moved - cumulative load moved across iterations | |
10164 | */ | |
163122b7 | 10165 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
10166 | |
10167 | /* | |
163122b7 KT |
10168 | * We've detached some tasks from busiest_rq. Every |
10169 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
10170 | * unlock busiest->lock, and we are able to be sure | |
10171 | * that nobody can manipulate the tasks in parallel. | |
10172 | * See task_rq_lock() family for the details. | |
1e3c88bd | 10173 | */ |
163122b7 | 10174 | |
8a8c69c3 | 10175 | rq_unlock(busiest, &rf); |
163122b7 KT |
10176 | |
10177 | if (cur_ld_moved) { | |
10178 | attach_tasks(&env); | |
10179 | ld_moved += cur_ld_moved; | |
10180 | } | |
10181 | ||
8a8c69c3 | 10182 | local_irq_restore(rf.flags); |
88b8dac0 | 10183 | |
f1cd0858 JK |
10184 | if (env.flags & LBF_NEED_BREAK) { |
10185 | env.flags &= ~LBF_NEED_BREAK; | |
10186 | goto more_balance; | |
10187 | } | |
10188 | ||
88b8dac0 SV |
10189 | /* |
10190 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
10191 | * us and move them to an alternate dst_cpu in our sched_group | |
10192 | * where they can run. The upper limit on how many times we | |
97fb7a0a | 10193 | * iterate on same src_cpu is dependent on number of CPUs in our |
88b8dac0 SV |
10194 | * sched_group. |
10195 | * | |
10196 | * This changes load balance semantics a bit on who can move | |
10197 | * load to a given_cpu. In addition to the given_cpu itself | |
10198 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
10199 | * nohz-idle), we now have balance_cpu in a position to move | |
10200 | * load to given_cpu. In rare situations, this may cause | |
10201 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
10202 | * _independently_ and at _same_ time to move some load to | |
3b03706f | 10203 | * given_cpu) causing excess load to be moved to given_cpu. |
88b8dac0 SV |
10204 | * This however should not happen so much in practice and |
10205 | * moreover subsequent load balance cycles should correct the | |
10206 | * excess load moved. | |
10207 | */ | |
6263322c | 10208 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 10209 | |
97fb7a0a | 10210 | /* Prevent to re-select dst_cpu via env's CPUs */ |
c89d92ed | 10211 | __cpumask_clear_cpu(env.dst_cpu, env.cpus); |
7aff2e3a | 10212 | |
78feefc5 | 10213 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 10214 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 10215 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
10216 | env.loop = 0; |
10217 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 10218 | |
88b8dac0 SV |
10219 | /* |
10220 | * Go back to "more_balance" rather than "redo" since we | |
10221 | * need to continue with same src_cpu. | |
10222 | */ | |
10223 | goto more_balance; | |
10224 | } | |
1e3c88bd | 10225 | |
6263322c PZ |
10226 | /* |
10227 | * We failed to reach balance because of affinity. | |
10228 | */ | |
10229 | if (sd_parent) { | |
63b2ca30 | 10230 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 10231 | |
afdeee05 | 10232 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 10233 | *group_imbalance = 1; |
6263322c PZ |
10234 | } |
10235 | ||
1e3c88bd | 10236 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 10237 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
c89d92ed | 10238 | __cpumask_clear_cpu(cpu_of(busiest), cpus); |
65a4433a JH |
10239 | /* |
10240 | * Attempting to continue load balancing at the current | |
10241 | * sched_domain level only makes sense if there are | |
10242 | * active CPUs remaining as possible busiest CPUs to | |
10243 | * pull load from which are not contained within the | |
10244 | * destination group that is receiving any migrated | |
10245 | * load. | |
10246 | */ | |
10247 | if (!cpumask_subset(cpus, env.dst_grpmask)) { | |
bbf18b19 PN |
10248 | env.loop = 0; |
10249 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 10250 | goto redo; |
bbf18b19 | 10251 | } |
afdeee05 | 10252 | goto out_all_pinned; |
1e3c88bd PZ |
10253 | } |
10254 | } | |
10255 | ||
10256 | if (!ld_moved) { | |
ae92882e | 10257 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
10258 | /* |
10259 | * Increment the failure counter only on periodic balance. | |
10260 | * We do not want newidle balance, which can be very | |
10261 | * frequent, pollute the failure counter causing | |
10262 | * excessive cache_hot migrations and active balances. | |
10263 | */ | |
10264 | if (idle != CPU_NEWLY_IDLE) | |
10265 | sd->nr_balance_failed++; | |
1e3c88bd | 10266 | |
bd939f45 | 10267 | if (need_active_balance(&env)) { |
8a8c69c3 PZ |
10268 | unsigned long flags; |
10269 | ||
5cb9eaa3 | 10270 | raw_spin_rq_lock_irqsave(busiest, flags); |
1e3c88bd | 10271 | |
97fb7a0a IM |
10272 | /* |
10273 | * Don't kick the active_load_balance_cpu_stop, | |
10274 | * if the curr task on busiest CPU can't be | |
10275 | * moved to this_cpu: | |
1e3c88bd | 10276 | */ |
3bd37062 | 10277 | if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { |
5cb9eaa3 | 10278 | raw_spin_rq_unlock_irqrestore(busiest, flags); |
1e3c88bd PZ |
10279 | goto out_one_pinned; |
10280 | } | |
10281 | ||
8a41dfcd VG |
10282 | /* Record that we found at least one task that could run on this_cpu */ |
10283 | env.flags &= ~LBF_ALL_PINNED; | |
10284 | ||
969c7921 TH |
10285 | /* |
10286 | * ->active_balance synchronizes accesses to | |
10287 | * ->active_balance_work. Once set, it's cleared | |
10288 | * only after active load balance is finished. | |
10289 | */ | |
1e3c88bd PZ |
10290 | if (!busiest->active_balance) { |
10291 | busiest->active_balance = 1; | |
10292 | busiest->push_cpu = this_cpu; | |
10293 | active_balance = 1; | |
10294 | } | |
5cb9eaa3 | 10295 | raw_spin_rq_unlock_irqrestore(busiest, flags); |
969c7921 | 10296 | |
bd939f45 | 10297 | if (active_balance) { |
969c7921 TH |
10298 | stop_one_cpu_nowait(cpu_of(busiest), |
10299 | active_load_balance_cpu_stop, busiest, | |
10300 | &busiest->active_balance_work); | |
bd939f45 | 10301 | } |
1e3c88bd | 10302 | } |
e9b9734b | 10303 | } else { |
1e3c88bd | 10304 | sd->nr_balance_failed = 0; |
e9b9734b | 10305 | } |
1e3c88bd | 10306 | |
e9b9734b | 10307 | if (likely(!active_balance) || need_active_balance(&env)) { |
1e3c88bd PZ |
10308 | /* We were unbalanced, so reset the balancing interval */ |
10309 | sd->balance_interval = sd->min_interval; | |
1e3c88bd PZ |
10310 | } |
10311 | ||
1e3c88bd PZ |
10312 | goto out; |
10313 | ||
10314 | out_balanced: | |
afdeee05 VG |
10315 | /* |
10316 | * We reach balance although we may have faced some affinity | |
f6cad8df VG |
10317 | * constraints. Clear the imbalance flag only if other tasks got |
10318 | * a chance to move and fix the imbalance. | |
afdeee05 | 10319 | */ |
f6cad8df | 10320 | if (sd_parent && !(env.flags & LBF_ALL_PINNED)) { |
afdeee05 VG |
10321 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
10322 | ||
10323 | if (*group_imbalance) | |
10324 | *group_imbalance = 0; | |
10325 | } | |
10326 | ||
10327 | out_all_pinned: | |
10328 | /* | |
10329 | * We reach balance because all tasks are pinned at this level so | |
10330 | * we can't migrate them. Let the imbalance flag set so parent level | |
10331 | * can try to migrate them. | |
10332 | */ | |
ae92882e | 10333 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
10334 | |
10335 | sd->nr_balance_failed = 0; | |
10336 | ||
10337 | out_one_pinned: | |
3f130a37 VS |
10338 | ld_moved = 0; |
10339 | ||
10340 | /* | |
5ba553ef PZ |
10341 | * newidle_balance() disregards balance intervals, so we could |
10342 | * repeatedly reach this code, which would lead to balance_interval | |
3b03706f | 10343 | * skyrocketing in a short amount of time. Skip the balance_interval |
5ba553ef | 10344 | * increase logic to avoid that. |
3f130a37 VS |
10345 | */ |
10346 | if (env.idle == CPU_NEWLY_IDLE) | |
10347 | goto out; | |
10348 | ||
1e3c88bd | 10349 | /* tune up the balancing interval */ |
47b7aee1 VS |
10350 | if ((env.flags & LBF_ALL_PINNED && |
10351 | sd->balance_interval < MAX_PINNED_INTERVAL) || | |
10352 | sd->balance_interval < sd->max_interval) | |
1e3c88bd | 10353 | sd->balance_interval *= 2; |
1e3c88bd | 10354 | out: |
1e3c88bd PZ |
10355 | return ld_moved; |
10356 | } | |
10357 | ||
52a08ef1 JL |
10358 | static inline unsigned long |
10359 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
10360 | { | |
10361 | unsigned long interval = sd->balance_interval; | |
10362 | ||
10363 | if (cpu_busy) | |
10364 | interval *= sd->busy_factor; | |
10365 | ||
10366 | /* scale ms to jiffies */ | |
10367 | interval = msecs_to_jiffies(interval); | |
e4d32e4d VG |
10368 | |
10369 | /* | |
10370 | * Reduce likelihood of busy balancing at higher domains racing with | |
10371 | * balancing at lower domains by preventing their balancing periods | |
10372 | * from being multiples of each other. | |
10373 | */ | |
10374 | if (cpu_busy) | |
10375 | interval -= 1; | |
10376 | ||
52a08ef1 JL |
10377 | interval = clamp(interval, 1UL, max_load_balance_interval); |
10378 | ||
10379 | return interval; | |
10380 | } | |
10381 | ||
10382 | static inline void | |
31851a98 | 10383 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
10384 | { |
10385 | unsigned long interval, next; | |
10386 | ||
31851a98 LY |
10387 | /* used by idle balance, so cpu_busy = 0 */ |
10388 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
10389 | next = sd->last_balance + interval; |
10390 | ||
10391 | if (time_after(*next_balance, next)) | |
10392 | *next_balance = next; | |
10393 | } | |
10394 | ||
1e3c88bd | 10395 | /* |
97fb7a0a | 10396 | * active_load_balance_cpu_stop is run by the CPU stopper. It pushes |
969c7921 TH |
10397 | * running tasks off the busiest CPU onto idle CPUs. It requires at |
10398 | * least 1 task to be running on each physical CPU where possible, and | |
10399 | * avoids physical / logical imbalances. | |
1e3c88bd | 10400 | */ |
969c7921 | 10401 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 10402 | { |
969c7921 TH |
10403 | struct rq *busiest_rq = data; |
10404 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 10405 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 10406 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 10407 | struct sched_domain *sd; |
e5673f28 | 10408 | struct task_struct *p = NULL; |
8a8c69c3 | 10409 | struct rq_flags rf; |
969c7921 | 10410 | |
8a8c69c3 | 10411 | rq_lock_irq(busiest_rq, &rf); |
edd8e41d PZ |
10412 | /* |
10413 | * Between queueing the stop-work and running it is a hole in which | |
10414 | * CPUs can become inactive. We should not move tasks from or to | |
10415 | * inactive CPUs. | |
10416 | */ | |
10417 | if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) | |
10418 | goto out_unlock; | |
969c7921 | 10419 | |
97fb7a0a | 10420 | /* Make sure the requested CPU hasn't gone down in the meantime: */ |
969c7921 TH |
10421 | if (unlikely(busiest_cpu != smp_processor_id() || |
10422 | !busiest_rq->active_balance)) | |
10423 | goto out_unlock; | |
1e3c88bd PZ |
10424 | |
10425 | /* Is there any task to move? */ | |
10426 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 10427 | goto out_unlock; |
1e3c88bd PZ |
10428 | |
10429 | /* | |
10430 | * This condition is "impossible", if it occurs | |
10431 | * we need to fix it. Originally reported by | |
97fb7a0a | 10432 | * Bjorn Helgaas on a 128-CPU setup. |
1e3c88bd PZ |
10433 | */ |
10434 | BUG_ON(busiest_rq == target_rq); | |
10435 | ||
1e3c88bd | 10436 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 10437 | rcu_read_lock(); |
1e3c88bd | 10438 | for_each_domain(target_cpu, sd) { |
e669ac8a VS |
10439 | if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) |
10440 | break; | |
1e3c88bd PZ |
10441 | } |
10442 | ||
10443 | if (likely(sd)) { | |
8e45cb54 PZ |
10444 | struct lb_env env = { |
10445 | .sd = sd, | |
ddcdf6e7 PZ |
10446 | .dst_cpu = target_cpu, |
10447 | .dst_rq = target_rq, | |
10448 | .src_cpu = busiest_rq->cpu, | |
10449 | .src_rq = busiest_rq, | |
8e45cb54 | 10450 | .idle = CPU_IDLE, |
23fb06d9 | 10451 | .flags = LBF_ACTIVE_LB, |
8e45cb54 PZ |
10452 | }; |
10453 | ||
ae92882e | 10454 | schedstat_inc(sd->alb_count); |
3bed5e21 | 10455 | update_rq_clock(busiest_rq); |
1e3c88bd | 10456 | |
e5673f28 | 10457 | p = detach_one_task(&env); |
d02c0711 | 10458 | if (p) { |
ae92882e | 10459 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
10460 | /* Active balancing done, reset the failure counter. */ |
10461 | sd->nr_balance_failed = 0; | |
10462 | } else { | |
ae92882e | 10463 | schedstat_inc(sd->alb_failed); |
d02c0711 | 10464 | } |
1e3c88bd | 10465 | } |
dce840a0 | 10466 | rcu_read_unlock(); |
969c7921 TH |
10467 | out_unlock: |
10468 | busiest_rq->active_balance = 0; | |
8a8c69c3 | 10469 | rq_unlock(busiest_rq, &rf); |
e5673f28 KT |
10470 | |
10471 | if (p) | |
10472 | attach_one_task(target_rq, p); | |
10473 | ||
10474 | local_irq_enable(); | |
10475 | ||
969c7921 | 10476 | return 0; |
1e3c88bd PZ |
10477 | } |
10478 | ||
af3fe03c PZ |
10479 | static DEFINE_SPINLOCK(balancing); |
10480 | ||
10481 | /* | |
10482 | * Scale the max load_balance interval with the number of CPUs in the system. | |
10483 | * This trades load-balance latency on larger machines for less cross talk. | |
10484 | */ | |
10485 | void update_max_interval(void) | |
10486 | { | |
10487 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
10488 | } | |
10489 | ||
e60b56e4 VG |
10490 | static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost) |
10491 | { | |
10492 | if (cost > sd->max_newidle_lb_cost) { | |
10493 | /* | |
10494 | * Track max cost of a domain to make sure to not delay the | |
10495 | * next wakeup on the CPU. | |
10496 | */ | |
10497 | sd->max_newidle_lb_cost = cost; | |
10498 | sd->last_decay_max_lb_cost = jiffies; | |
10499 | } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) { | |
10500 | /* | |
10501 | * Decay the newidle max times by ~1% per second to ensure that | |
10502 | * it is not outdated and the current max cost is actually | |
10503 | * shorter. | |
10504 | */ | |
10505 | sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256; | |
10506 | sd->last_decay_max_lb_cost = jiffies; | |
10507 | ||
10508 | return true; | |
10509 | } | |
10510 | ||
10511 | return false; | |
10512 | } | |
10513 | ||
af3fe03c PZ |
10514 | /* |
10515 | * It checks each scheduling domain to see if it is due to be balanced, | |
10516 | * and initiates a balancing operation if so. | |
10517 | * | |
10518 | * Balancing parameters are set up in init_sched_domains. | |
10519 | */ | |
10520 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) | |
10521 | { | |
10522 | int continue_balancing = 1; | |
10523 | int cpu = rq->cpu; | |
323af6de | 10524 | int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
10525 | unsigned long interval; |
10526 | struct sched_domain *sd; | |
10527 | /* Earliest time when we have to do rebalance again */ | |
10528 | unsigned long next_balance = jiffies + 60*HZ; | |
10529 | int update_next_balance = 0; | |
10530 | int need_serialize, need_decay = 0; | |
10531 | u64 max_cost = 0; | |
10532 | ||
10533 | rcu_read_lock(); | |
10534 | for_each_domain(cpu, sd) { | |
10535 | /* | |
10536 | * Decay the newidle max times here because this is a regular | |
e60b56e4 | 10537 | * visit to all the domains. |
af3fe03c | 10538 | */ |
e60b56e4 | 10539 | need_decay = update_newidle_cost(sd, 0); |
af3fe03c PZ |
10540 | max_cost += sd->max_newidle_lb_cost; |
10541 | ||
af3fe03c PZ |
10542 | /* |
10543 | * Stop the load balance at this level. There is another | |
10544 | * CPU in our sched group which is doing load balancing more | |
10545 | * actively. | |
10546 | */ | |
10547 | if (!continue_balancing) { | |
10548 | if (need_decay) | |
10549 | continue; | |
10550 | break; | |
10551 | } | |
10552 | ||
323af6de | 10553 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
10554 | |
10555 | need_serialize = sd->flags & SD_SERIALIZE; | |
10556 | if (need_serialize) { | |
10557 | if (!spin_trylock(&balancing)) | |
10558 | goto out; | |
10559 | } | |
10560 | ||
10561 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
10562 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { | |
10563 | /* | |
10564 | * The LBF_DST_PINNED logic could have changed | |
10565 | * env->dst_cpu, so we can't know our idle | |
10566 | * state even if we migrated tasks. Update it. | |
10567 | */ | |
10568 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; | |
323af6de | 10569 | busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); |
af3fe03c PZ |
10570 | } |
10571 | sd->last_balance = jiffies; | |
323af6de | 10572 | interval = get_sd_balance_interval(sd, busy); |
af3fe03c PZ |
10573 | } |
10574 | if (need_serialize) | |
10575 | spin_unlock(&balancing); | |
10576 | out: | |
10577 | if (time_after(next_balance, sd->last_balance + interval)) { | |
10578 | next_balance = sd->last_balance + interval; | |
10579 | update_next_balance = 1; | |
10580 | } | |
10581 | } | |
10582 | if (need_decay) { | |
10583 | /* | |
10584 | * Ensure the rq-wide value also decays but keep it at a | |
10585 | * reasonable floor to avoid funnies with rq->avg_idle. | |
10586 | */ | |
10587 | rq->max_idle_balance_cost = | |
10588 | max((u64)sysctl_sched_migration_cost, max_cost); | |
10589 | } | |
10590 | rcu_read_unlock(); | |
10591 | ||
10592 | /* | |
10593 | * next_balance will be updated only when there is a need. | |
10594 | * When the cpu is attached to null domain for ex, it will not be | |
10595 | * updated. | |
10596 | */ | |
7a82e5f5 | 10597 | if (likely(update_next_balance)) |
af3fe03c PZ |
10598 | rq->next_balance = next_balance; |
10599 | ||
af3fe03c PZ |
10600 | } |
10601 | ||
d987fc7f MG |
10602 | static inline int on_null_domain(struct rq *rq) |
10603 | { | |
10604 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
10605 | } | |
10606 | ||
3451d024 | 10607 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
10608 | /* |
10609 | * idle load balancing details | |
83cd4fe2 VP |
10610 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
10611 | * needed, they will kick the idle load balancer, which then does idle | |
10612 | * load balancing for all the idle CPUs. | |
04d4e665 | 10613 | * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set |
9b019acb | 10614 | * anywhere yet. |
83cd4fe2 | 10615 | */ |
1e3c88bd | 10616 | |
3dd0337d | 10617 | static inline int find_new_ilb(void) |
1e3c88bd | 10618 | { |
9b019acb | 10619 | int ilb; |
031e3bd8 | 10620 | const struct cpumask *hk_mask; |
1e3c88bd | 10621 | |
04d4e665 | 10622 | hk_mask = housekeeping_cpumask(HK_TYPE_MISC); |
1e3c88bd | 10623 | |
031e3bd8 | 10624 | for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) { |
45da7a2b PZ |
10625 | |
10626 | if (ilb == smp_processor_id()) | |
10627 | continue; | |
10628 | ||
9b019acb NP |
10629 | if (idle_cpu(ilb)) |
10630 | return ilb; | |
10631 | } | |
786d6dc7 SS |
10632 | |
10633 | return nr_cpu_ids; | |
1e3c88bd | 10634 | } |
1e3c88bd | 10635 | |
83cd4fe2 | 10636 | /* |
9b019acb | 10637 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick any |
04d4e665 | 10638 | * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one). |
83cd4fe2 | 10639 | */ |
a4064fb6 | 10640 | static void kick_ilb(unsigned int flags) |
83cd4fe2 VP |
10641 | { |
10642 | int ilb_cpu; | |
10643 | ||
3ea2f097 VG |
10644 | /* |
10645 | * Increase nohz.next_balance only when if full ilb is triggered but | |
10646 | * not if we only update stats. | |
10647 | */ | |
10648 | if (flags & NOHZ_BALANCE_KICK) | |
10649 | nohz.next_balance = jiffies+1; | |
83cd4fe2 | 10650 | |
3dd0337d | 10651 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 10652 | |
0b005cf5 SS |
10653 | if (ilb_cpu >= nr_cpu_ids) |
10654 | return; | |
83cd4fe2 | 10655 | |
19a1f5ec PZ |
10656 | /* |
10657 | * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets | |
10658 | * the first flag owns it; cleared by nohz_csd_func(). | |
10659 | */ | |
a4064fb6 | 10660 | flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu)); |
b7031a02 | 10661 | if (flags & NOHZ_KICK_MASK) |
1c792db7 | 10662 | return; |
4550487a | 10663 | |
1c792db7 | 10664 | /* |
90b5363a | 10665 | * This way we generate an IPI on the target CPU which |
1c792db7 SS |
10666 | * is idle. And the softirq performing nohz idle load balance |
10667 | * will be run before returning from the IPI. | |
10668 | */ | |
90b5363a | 10669 | smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd); |
4550487a PZ |
10670 | } |
10671 | ||
10672 | /* | |
9f132742 VS |
10673 | * Current decision point for kicking the idle load balancer in the presence |
10674 | * of idle CPUs in the system. | |
4550487a PZ |
10675 | */ |
10676 | static void nohz_balancer_kick(struct rq *rq) | |
10677 | { | |
10678 | unsigned long now = jiffies; | |
10679 | struct sched_domain_shared *sds; | |
10680 | struct sched_domain *sd; | |
10681 | int nr_busy, i, cpu = rq->cpu; | |
a4064fb6 | 10682 | unsigned int flags = 0; |
4550487a PZ |
10683 | |
10684 | if (unlikely(rq->idle_balance)) | |
10685 | return; | |
10686 | ||
10687 | /* | |
10688 | * We may be recently in ticked or tickless idle mode. At the first | |
10689 | * busy tick after returning from idle, we will update the busy stats. | |
10690 | */ | |
00357f5e | 10691 | nohz_balance_exit_idle(rq); |
4550487a PZ |
10692 | |
10693 | /* | |
10694 | * None are in tickless mode and hence no need for NOHZ idle load | |
10695 | * balancing. | |
10696 | */ | |
10697 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
10698 | return; | |
10699 | ||
f643ea22 VG |
10700 | if (READ_ONCE(nohz.has_blocked) && |
10701 | time_after(now, READ_ONCE(nohz.next_blocked))) | |
a4064fb6 PZ |
10702 | flags = NOHZ_STATS_KICK; |
10703 | ||
4550487a | 10704 | if (time_before(now, nohz.next_balance)) |
a4064fb6 | 10705 | goto out; |
4550487a | 10706 | |
a0fe2cf0 | 10707 | if (rq->nr_running >= 2) { |
efd984c4 | 10708 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
4550487a PZ |
10709 | goto out; |
10710 | } | |
10711 | ||
10712 | rcu_read_lock(); | |
4550487a PZ |
10713 | |
10714 | sd = rcu_dereference(rq->sd); | |
10715 | if (sd) { | |
e25a7a94 VS |
10716 | /* |
10717 | * If there's a CFS task and the current CPU has reduced | |
10718 | * capacity; kick the ILB to see if there's a better CPU to run | |
10719 | * on. | |
10720 | */ | |
10721 | if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) { | |
efd984c4 | 10722 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
4550487a PZ |
10723 | goto unlock; |
10724 | } | |
10725 | } | |
10726 | ||
011b27bb | 10727 | sd = rcu_dereference(per_cpu(sd_asym_packing, cpu)); |
4550487a | 10728 | if (sd) { |
b9a7b883 VS |
10729 | /* |
10730 | * When ASYM_PACKING; see if there's a more preferred CPU | |
10731 | * currently idle; in which case, kick the ILB to move tasks | |
10732 | * around. | |
10733 | */ | |
7edab78d | 10734 | for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) { |
4550487a | 10735 | if (sched_asym_prefer(i, cpu)) { |
efd984c4 | 10736 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
4550487a PZ |
10737 | goto unlock; |
10738 | } | |
10739 | } | |
10740 | } | |
b9a7b883 | 10741 | |
a0fe2cf0 VS |
10742 | sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu)); |
10743 | if (sd) { | |
10744 | /* | |
10745 | * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU | |
10746 | * to run the misfit task on. | |
10747 | */ | |
10748 | if (check_misfit_status(rq, sd)) { | |
efd984c4 | 10749 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
a0fe2cf0 VS |
10750 | goto unlock; |
10751 | } | |
b9a7b883 VS |
10752 | |
10753 | /* | |
10754 | * For asymmetric systems, we do not want to nicely balance | |
10755 | * cache use, instead we want to embrace asymmetry and only | |
10756 | * ensure tasks have enough CPU capacity. | |
10757 | * | |
10758 | * Skip the LLC logic because it's not relevant in that case. | |
10759 | */ | |
10760 | goto unlock; | |
a0fe2cf0 VS |
10761 | } |
10762 | ||
b9a7b883 VS |
10763 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
10764 | if (sds) { | |
e25a7a94 | 10765 | /* |
b9a7b883 VS |
10766 | * If there is an imbalance between LLC domains (IOW we could |
10767 | * increase the overall cache use), we need some less-loaded LLC | |
10768 | * domain to pull some load. Likewise, we may need to spread | |
10769 | * load within the current LLC domain (e.g. packed SMT cores but | |
10770 | * other CPUs are idle). We can't really know from here how busy | |
10771 | * the others are - so just get a nohz balance going if it looks | |
10772 | * like this LLC domain has tasks we could move. | |
e25a7a94 | 10773 | */ |
b9a7b883 VS |
10774 | nr_busy = atomic_read(&sds->nr_busy_cpus); |
10775 | if (nr_busy > 1) { | |
efd984c4 | 10776 | flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK; |
b9a7b883 | 10777 | goto unlock; |
4550487a PZ |
10778 | } |
10779 | } | |
10780 | unlock: | |
10781 | rcu_read_unlock(); | |
10782 | out: | |
7fd7a9e0 VS |
10783 | if (READ_ONCE(nohz.needs_update)) |
10784 | flags |= NOHZ_NEXT_KICK; | |
10785 | ||
a4064fb6 PZ |
10786 | if (flags) |
10787 | kick_ilb(flags); | |
83cd4fe2 VP |
10788 | } |
10789 | ||
00357f5e | 10790 | static void set_cpu_sd_state_busy(int cpu) |
71325960 | 10791 | { |
00357f5e | 10792 | struct sched_domain *sd; |
a22e47a4 | 10793 | |
00357f5e PZ |
10794 | rcu_read_lock(); |
10795 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); | |
a22e47a4 | 10796 | |
00357f5e PZ |
10797 | if (!sd || !sd->nohz_idle) |
10798 | goto unlock; | |
10799 | sd->nohz_idle = 0; | |
10800 | ||
10801 | atomic_inc(&sd->shared->nr_busy_cpus); | |
10802 | unlock: | |
10803 | rcu_read_unlock(); | |
71325960 SS |
10804 | } |
10805 | ||
00357f5e PZ |
10806 | void nohz_balance_exit_idle(struct rq *rq) |
10807 | { | |
10808 | SCHED_WARN_ON(rq != this_rq()); | |
10809 | ||
10810 | if (likely(!rq->nohz_tick_stopped)) | |
10811 | return; | |
10812 | ||
10813 | rq->nohz_tick_stopped = 0; | |
10814 | cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); | |
10815 | atomic_dec(&nohz.nr_cpus); | |
10816 | ||
10817 | set_cpu_sd_state_busy(rq->cpu); | |
10818 | } | |
10819 | ||
10820 | static void set_cpu_sd_state_idle(int cpu) | |
69e1e811 SS |
10821 | { |
10822 | struct sched_domain *sd; | |
69e1e811 | 10823 | |
69e1e811 | 10824 | rcu_read_lock(); |
0e369d75 | 10825 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
10826 | |
10827 | if (!sd || sd->nohz_idle) | |
10828 | goto unlock; | |
10829 | sd->nohz_idle = 1; | |
10830 | ||
0e369d75 | 10831 | atomic_dec(&sd->shared->nr_busy_cpus); |
25f55d9d | 10832 | unlock: |
69e1e811 SS |
10833 | rcu_read_unlock(); |
10834 | } | |
10835 | ||
1e3c88bd | 10836 | /* |
97fb7a0a | 10837 | * This routine will record that the CPU is going idle with tick stopped. |
0b005cf5 | 10838 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 10839 | */ |
c1cc017c | 10840 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 10841 | { |
00357f5e PZ |
10842 | struct rq *rq = cpu_rq(cpu); |
10843 | ||
10844 | SCHED_WARN_ON(cpu != smp_processor_id()); | |
10845 | ||
97fb7a0a | 10846 | /* If this CPU is going down, then nothing needs to be done: */ |
71325960 SS |
10847 | if (!cpu_active(cpu)) |
10848 | return; | |
10849 | ||
387bc8b5 | 10850 | /* Spare idle load balancing on CPUs that don't want to be disturbed: */ |
04d4e665 | 10851 | if (!housekeeping_cpu(cpu, HK_TYPE_SCHED)) |
387bc8b5 FW |
10852 | return; |
10853 | ||
f643ea22 VG |
10854 | /* |
10855 | * Can be set safely without rq->lock held | |
10856 | * If a clear happens, it will have evaluated last additions because | |
10857 | * rq->lock is held during the check and the clear | |
10858 | */ | |
10859 | rq->has_blocked_load = 1; | |
10860 | ||
10861 | /* | |
10862 | * The tick is still stopped but load could have been added in the | |
10863 | * meantime. We set the nohz.has_blocked flag to trig a check of the | |
10864 | * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear | |
10865 | * of nohz.has_blocked can only happen after checking the new load | |
10866 | */ | |
00357f5e | 10867 | if (rq->nohz_tick_stopped) |
f643ea22 | 10868 | goto out; |
1e3c88bd | 10869 | |
97fb7a0a | 10870 | /* If we're a completely isolated CPU, we don't play: */ |
00357f5e | 10871 | if (on_null_domain(rq)) |
d987fc7f MG |
10872 | return; |
10873 | ||
00357f5e PZ |
10874 | rq->nohz_tick_stopped = 1; |
10875 | ||
c1cc017c AS |
10876 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
10877 | atomic_inc(&nohz.nr_cpus); | |
00357f5e | 10878 | |
f643ea22 VG |
10879 | /* |
10880 | * Ensures that if nohz_idle_balance() fails to observe our | |
10881 | * @idle_cpus_mask store, it must observe the @has_blocked | |
7fd7a9e0 | 10882 | * and @needs_update stores. |
f643ea22 VG |
10883 | */ |
10884 | smp_mb__after_atomic(); | |
10885 | ||
00357f5e | 10886 | set_cpu_sd_state_idle(cpu); |
f643ea22 | 10887 | |
7fd7a9e0 | 10888 | WRITE_ONCE(nohz.needs_update, 1); |
f643ea22 VG |
10889 | out: |
10890 | /* | |
10891 | * Each time a cpu enter idle, we assume that it has blocked load and | |
10892 | * enable the periodic update of the load of idle cpus | |
10893 | */ | |
10894 | WRITE_ONCE(nohz.has_blocked, 1); | |
1e3c88bd | 10895 | } |
1e3c88bd | 10896 | |
3f5ad914 Y |
10897 | static bool update_nohz_stats(struct rq *rq) |
10898 | { | |
10899 | unsigned int cpu = rq->cpu; | |
10900 | ||
10901 | if (!rq->has_blocked_load) | |
10902 | return false; | |
10903 | ||
10904 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) | |
10905 | return false; | |
10906 | ||
10907 | if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick))) | |
10908 | return true; | |
10909 | ||
10910 | update_blocked_averages(cpu); | |
10911 | ||
10912 | return rq->has_blocked_load; | |
10913 | } | |
10914 | ||
1e3c88bd | 10915 | /* |
31e77c93 VG |
10916 | * Internal function that runs load balance for all idle cpus. The load balance |
10917 | * can be a simple update of blocked load or a complete load balance with | |
10918 | * tasks movement depending of flags. | |
1e3c88bd | 10919 | */ |
d985ee9f | 10920 | static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags) |
83cd4fe2 | 10921 | { |
c5afb6a8 | 10922 | /* Earliest time when we have to do rebalance again */ |
a4064fb6 PZ |
10923 | unsigned long now = jiffies; |
10924 | unsigned long next_balance = now + 60*HZ; | |
f643ea22 | 10925 | bool has_blocked_load = false; |
c5afb6a8 | 10926 | int update_next_balance = 0; |
b7031a02 | 10927 | int this_cpu = this_rq->cpu; |
b7031a02 PZ |
10928 | int balance_cpu; |
10929 | struct rq *rq; | |
83cd4fe2 | 10930 | |
b7031a02 | 10931 | SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK); |
83cd4fe2 | 10932 | |
f643ea22 VG |
10933 | /* |
10934 | * We assume there will be no idle load after this update and clear | |
10935 | * the has_blocked flag. If a cpu enters idle in the mean time, it will | |
7fd7a9e0 | 10936 | * set the has_blocked flag and trigger another update of idle load. |
f643ea22 VG |
10937 | * Because a cpu that becomes idle, is added to idle_cpus_mask before |
10938 | * setting the flag, we are sure to not clear the state and not | |
10939 | * check the load of an idle cpu. | |
7fd7a9e0 VS |
10940 | * |
10941 | * Same applies to idle_cpus_mask vs needs_update. | |
f643ea22 | 10942 | */ |
efd984c4 VS |
10943 | if (flags & NOHZ_STATS_KICK) |
10944 | WRITE_ONCE(nohz.has_blocked, 0); | |
7fd7a9e0 VS |
10945 | if (flags & NOHZ_NEXT_KICK) |
10946 | WRITE_ONCE(nohz.needs_update, 0); | |
f643ea22 VG |
10947 | |
10948 | /* | |
10949 | * Ensures that if we miss the CPU, we must see the has_blocked | |
10950 | * store from nohz_balance_enter_idle(). | |
10951 | */ | |
10952 | smp_mb(); | |
10953 | ||
7a82e5f5 VG |
10954 | /* |
10955 | * Start with the next CPU after this_cpu so we will end with this_cpu and let a | |
10956 | * chance for other idle cpu to pull load. | |
10957 | */ | |
10958 | for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) { | |
10959 | if (!idle_cpu(balance_cpu)) | |
83cd4fe2 VP |
10960 | continue; |
10961 | ||
10962 | /* | |
97fb7a0a IM |
10963 | * If this CPU gets work to do, stop the load balancing |
10964 | * work being done for other CPUs. Next load | |
83cd4fe2 VP |
10965 | * balancing owner will pick it up. |
10966 | */ | |
f643ea22 | 10967 | if (need_resched()) { |
efd984c4 VS |
10968 | if (flags & NOHZ_STATS_KICK) |
10969 | has_blocked_load = true; | |
7fd7a9e0 VS |
10970 | if (flags & NOHZ_NEXT_KICK) |
10971 | WRITE_ONCE(nohz.needs_update, 1); | |
f643ea22 VG |
10972 | goto abort; |
10973 | } | |
83cd4fe2 | 10974 | |
5ed4f1d9 VG |
10975 | rq = cpu_rq(balance_cpu); |
10976 | ||
efd984c4 VS |
10977 | if (flags & NOHZ_STATS_KICK) |
10978 | has_blocked_load |= update_nohz_stats(rq); | |
f643ea22 | 10979 | |
ed61bbc6 TC |
10980 | /* |
10981 | * If time for next balance is due, | |
10982 | * do the balance. | |
10983 | */ | |
10984 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8a8c69c3 PZ |
10985 | struct rq_flags rf; |
10986 | ||
31e77c93 | 10987 | rq_lock_irqsave(rq, &rf); |
ed61bbc6 | 10988 | update_rq_clock(rq); |
31e77c93 | 10989 | rq_unlock_irqrestore(rq, &rf); |
8a8c69c3 | 10990 | |
b7031a02 PZ |
10991 | if (flags & NOHZ_BALANCE_KICK) |
10992 | rebalance_domains(rq, CPU_IDLE); | |
ed61bbc6 | 10993 | } |
83cd4fe2 | 10994 | |
c5afb6a8 VG |
10995 | if (time_after(next_balance, rq->next_balance)) { |
10996 | next_balance = rq->next_balance; | |
10997 | update_next_balance = 1; | |
10998 | } | |
83cd4fe2 | 10999 | } |
c5afb6a8 | 11000 | |
3ea2f097 VG |
11001 | /* |
11002 | * next_balance will be updated only when there is a need. | |
11003 | * When the CPU is attached to null domain for ex, it will not be | |
11004 | * updated. | |
11005 | */ | |
11006 | if (likely(update_next_balance)) | |
11007 | nohz.next_balance = next_balance; | |
11008 | ||
efd984c4 VS |
11009 | if (flags & NOHZ_STATS_KICK) |
11010 | WRITE_ONCE(nohz.next_blocked, | |
11011 | now + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
f643ea22 VG |
11012 | |
11013 | abort: | |
11014 | /* There is still blocked load, enable periodic update */ | |
11015 | if (has_blocked_load) | |
11016 | WRITE_ONCE(nohz.has_blocked, 1); | |
31e77c93 VG |
11017 | } |
11018 | ||
11019 | /* | |
11020 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the | |
11021 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
11022 | */ | |
11023 | static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) | |
11024 | { | |
19a1f5ec | 11025 | unsigned int flags = this_rq->nohz_idle_balance; |
31e77c93 | 11026 | |
19a1f5ec | 11027 | if (!flags) |
31e77c93 VG |
11028 | return false; |
11029 | ||
19a1f5ec | 11030 | this_rq->nohz_idle_balance = 0; |
31e77c93 | 11031 | |
19a1f5ec | 11032 | if (idle != CPU_IDLE) |
31e77c93 VG |
11033 | return false; |
11034 | ||
d985ee9f | 11035 | _nohz_idle_balance(this_rq, flags); |
31e77c93 | 11036 | |
b7031a02 | 11037 | return true; |
83cd4fe2 | 11038 | } |
31e77c93 | 11039 | |
c6f88654 VG |
11040 | /* |
11041 | * Check if we need to run the ILB for updating blocked load before entering | |
11042 | * idle state. | |
11043 | */ | |
11044 | void nohz_run_idle_balance(int cpu) | |
11045 | { | |
11046 | unsigned int flags; | |
11047 | ||
11048 | flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu)); | |
11049 | ||
11050 | /* | |
11051 | * Update the blocked load only if no SCHED_SOFTIRQ is about to happen | |
11052 | * (ie NOHZ_STATS_KICK set) and will do the same. | |
11053 | */ | |
11054 | if ((flags == NOHZ_NEWILB_KICK) && !need_resched()) | |
d985ee9f | 11055 | _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK); |
c6f88654 VG |
11056 | } |
11057 | ||
31e77c93 VG |
11058 | static void nohz_newidle_balance(struct rq *this_rq) |
11059 | { | |
11060 | int this_cpu = this_rq->cpu; | |
11061 | ||
11062 | /* | |
11063 | * This CPU doesn't want to be disturbed by scheduler | |
11064 | * housekeeping | |
11065 | */ | |
04d4e665 | 11066 | if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED)) |
31e77c93 VG |
11067 | return; |
11068 | ||
11069 | /* Will wake up very soon. No time for doing anything else*/ | |
11070 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
11071 | return; | |
11072 | ||
11073 | /* Don't need to update blocked load of idle CPUs*/ | |
11074 | if (!READ_ONCE(nohz.has_blocked) || | |
11075 | time_before(jiffies, READ_ONCE(nohz.next_blocked))) | |
11076 | return; | |
11077 | ||
31e77c93 | 11078 | /* |
c6f88654 VG |
11079 | * Set the need to trigger ILB in order to update blocked load |
11080 | * before entering idle state. | |
31e77c93 | 11081 | */ |
c6f88654 | 11082 | atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu)); |
31e77c93 VG |
11083 | } |
11084 | ||
dd707247 PZ |
11085 | #else /* !CONFIG_NO_HZ_COMMON */ |
11086 | static inline void nohz_balancer_kick(struct rq *rq) { } | |
11087 | ||
31e77c93 | 11088 | static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
b7031a02 PZ |
11089 | { |
11090 | return false; | |
11091 | } | |
31e77c93 VG |
11092 | |
11093 | static inline void nohz_newidle_balance(struct rq *this_rq) { } | |
dd707247 | 11094 | #endif /* CONFIG_NO_HZ_COMMON */ |
83cd4fe2 | 11095 | |
47ea5412 | 11096 | /* |
5b78f2dc | 11097 | * newidle_balance is called by schedule() if this_cpu is about to become |
47ea5412 | 11098 | * idle. Attempts to pull tasks from other CPUs. |
7277a34c PZ |
11099 | * |
11100 | * Returns: | |
11101 | * < 0 - we released the lock and there are !fair tasks present | |
11102 | * 0 - failed, no new tasks | |
11103 | * > 0 - success, new (fair) tasks present | |
47ea5412 | 11104 | */ |
d91cecc1 | 11105 | static int newidle_balance(struct rq *this_rq, struct rq_flags *rf) |
47ea5412 PZ |
11106 | { |
11107 | unsigned long next_balance = jiffies + HZ; | |
11108 | int this_cpu = this_rq->cpu; | |
9e9af819 | 11109 | u64 t0, t1, curr_cost = 0; |
47ea5412 PZ |
11110 | struct sched_domain *sd; |
11111 | int pulled_task = 0; | |
47ea5412 | 11112 | |
5ba553ef | 11113 | update_misfit_status(NULL, this_rq); |
e5e678e4 RR |
11114 | |
11115 | /* | |
11116 | * There is a task waiting to run. No need to search for one. | |
11117 | * Return 0; the task will be enqueued when switching to idle. | |
11118 | */ | |
11119 | if (this_rq->ttwu_pending) | |
11120 | return 0; | |
11121 | ||
47ea5412 PZ |
11122 | /* |
11123 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
11124 | * measure the duration of idle_balance() as idle time. | |
11125 | */ | |
11126 | this_rq->idle_stamp = rq_clock(this_rq); | |
11127 | ||
11128 | /* | |
11129 | * Do not pull tasks towards !active CPUs... | |
11130 | */ | |
11131 | if (!cpu_active(this_cpu)) | |
11132 | return 0; | |
11133 | ||
11134 | /* | |
11135 | * This is OK, because current is on_cpu, which avoids it being picked | |
11136 | * for load-balance and preemption/IRQs are still disabled avoiding | |
11137 | * further scheduler activity on it and we're being very careful to | |
11138 | * re-start the picking loop. | |
11139 | */ | |
11140 | rq_unpin_lock(this_rq, rf); | |
11141 | ||
9d783c8d VG |
11142 | rcu_read_lock(); |
11143 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
11144 | ||
c5b0a7ee | 11145 | if (!READ_ONCE(this_rq->rd->overload) || |
9d783c8d | 11146 | (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) { |
31e77c93 | 11147 | |
47ea5412 PZ |
11148 | if (sd) |
11149 | update_next_balance(sd, &next_balance); | |
11150 | rcu_read_unlock(); | |
11151 | ||
11152 | goto out; | |
11153 | } | |
9d783c8d | 11154 | rcu_read_unlock(); |
47ea5412 | 11155 | |
5cb9eaa3 | 11156 | raw_spin_rq_unlock(this_rq); |
47ea5412 | 11157 | |
9e9af819 | 11158 | t0 = sched_clock_cpu(this_cpu); |
47ea5412 | 11159 | update_blocked_averages(this_cpu); |
9e9af819 | 11160 | |
47ea5412 PZ |
11161 | rcu_read_lock(); |
11162 | for_each_domain(this_cpu, sd) { | |
11163 | int continue_balancing = 1; | |
9e9af819 | 11164 | u64 domain_cost; |
47ea5412 | 11165 | |
8ea9183d VG |
11166 | update_next_balance(sd, &next_balance); |
11167 | ||
11168 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) | |
47ea5412 | 11169 | break; |
47ea5412 PZ |
11170 | |
11171 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
47ea5412 PZ |
11172 | |
11173 | pulled_task = load_balance(this_cpu, this_rq, | |
11174 | sd, CPU_NEWLY_IDLE, | |
11175 | &continue_balancing); | |
11176 | ||
9e9af819 VG |
11177 | t1 = sched_clock_cpu(this_cpu); |
11178 | domain_cost = t1 - t0; | |
e60b56e4 | 11179 | update_newidle_cost(sd, domain_cost); |
47ea5412 PZ |
11180 | |
11181 | curr_cost += domain_cost; | |
9e9af819 | 11182 | t0 = t1; |
47ea5412 PZ |
11183 | } |
11184 | ||
47ea5412 PZ |
11185 | /* |
11186 | * Stop searching for tasks to pull if there are | |
11187 | * now runnable tasks on this rq. | |
11188 | */ | |
e5e678e4 RR |
11189 | if (pulled_task || this_rq->nr_running > 0 || |
11190 | this_rq->ttwu_pending) | |
47ea5412 PZ |
11191 | break; |
11192 | } | |
11193 | rcu_read_unlock(); | |
11194 | ||
5cb9eaa3 | 11195 | raw_spin_rq_lock(this_rq); |
47ea5412 PZ |
11196 | |
11197 | if (curr_cost > this_rq->max_idle_balance_cost) | |
11198 | this_rq->max_idle_balance_cost = curr_cost; | |
11199 | ||
11200 | /* | |
11201 | * While browsing the domains, we released the rq lock, a task could | |
11202 | * have been enqueued in the meantime. Since we're not going idle, | |
11203 | * pretend we pulled a task. | |
11204 | */ | |
11205 | if (this_rq->cfs.h_nr_running && !pulled_task) | |
11206 | pulled_task = 1; | |
11207 | ||
47ea5412 PZ |
11208 | /* Is there a task of a high priority class? */ |
11209 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) | |
11210 | pulled_task = -1; | |
11211 | ||
6553fc18 VG |
11212 | out: |
11213 | /* Move the next balance forward */ | |
11214 | if (time_after(this_rq->next_balance, next_balance)) | |
11215 | this_rq->next_balance = next_balance; | |
11216 | ||
47ea5412 PZ |
11217 | if (pulled_task) |
11218 | this_rq->idle_stamp = 0; | |
0826530d VG |
11219 | else |
11220 | nohz_newidle_balance(this_rq); | |
47ea5412 PZ |
11221 | |
11222 | rq_repin_lock(this_rq, rf); | |
11223 | ||
11224 | return pulled_task; | |
11225 | } | |
11226 | ||
83cd4fe2 VP |
11227 | /* |
11228 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
11229 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
11230 | */ | |
0766f788 | 11231 | static __latent_entropy void run_rebalance_domains(struct softirq_action *h) |
1e3c88bd | 11232 | { |
208cb16b | 11233 | struct rq *this_rq = this_rq(); |
6eb57e0d | 11234 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
11235 | CPU_IDLE : CPU_NOT_IDLE; |
11236 | ||
1e3c88bd | 11237 | /* |
97fb7a0a IM |
11238 | * If this CPU has a pending nohz_balance_kick, then do the |
11239 | * balancing on behalf of the other idle CPUs whose ticks are | |
d4573c3e | 11240 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
97fb7a0a | 11241 | * give the idle CPUs a chance to load balance. Else we may |
d4573c3e PM |
11242 | * load balance only within the local sched_domain hierarchy |
11243 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 11244 | */ |
b7031a02 PZ |
11245 | if (nohz_idle_balance(this_rq, idle)) |
11246 | return; | |
11247 | ||
11248 | /* normal load balance */ | |
11249 | update_blocked_averages(this_rq->cpu); | |
d4573c3e | 11250 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
11251 | } |
11252 | ||
1e3c88bd PZ |
11253 | /* |
11254 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 11255 | */ |
7caff66f | 11256 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 11257 | { |
e0b257c3 AMB |
11258 | /* |
11259 | * Don't need to rebalance while attached to NULL domain or | |
11260 | * runqueue CPU is not active | |
11261 | */ | |
11262 | if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq)))) | |
c726099e DL |
11263 | return; |
11264 | ||
11265 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 11266 | raise_softirq(SCHED_SOFTIRQ); |
4550487a PZ |
11267 | |
11268 | nohz_balancer_kick(rq); | |
1e3c88bd PZ |
11269 | } |
11270 | ||
0bcdcf28 CE |
11271 | static void rq_online_fair(struct rq *rq) |
11272 | { | |
11273 | update_sysctl(); | |
0e59bdae KT |
11274 | |
11275 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
11276 | } |
11277 | ||
11278 | static void rq_offline_fair(struct rq *rq) | |
11279 | { | |
11280 | update_sysctl(); | |
a4c96ae3 PB |
11281 | |
11282 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
11283 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
11284 | } |
11285 | ||
55e12e5e | 11286 | #endif /* CONFIG_SMP */ |
e1d1484f | 11287 | |
8039e96f VP |
11288 | #ifdef CONFIG_SCHED_CORE |
11289 | static inline bool | |
11290 | __entity_slice_used(struct sched_entity *se, int min_nr_tasks) | |
11291 | { | |
11292 | u64 slice = sched_slice(cfs_rq_of(se), se); | |
11293 | u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
11294 | ||
11295 | return (rtime * min_nr_tasks > slice); | |
11296 | } | |
11297 | ||
11298 | #define MIN_NR_TASKS_DURING_FORCEIDLE 2 | |
11299 | static inline void task_tick_core(struct rq *rq, struct task_struct *curr) | |
11300 | { | |
11301 | if (!sched_core_enabled(rq)) | |
11302 | return; | |
11303 | ||
11304 | /* | |
11305 | * If runqueue has only one task which used up its slice and | |
11306 | * if the sibling is forced idle, then trigger schedule to | |
11307 | * give forced idle task a chance. | |
11308 | * | |
11309 | * sched_slice() considers only this active rq and it gets the | |
11310 | * whole slice. But during force idle, we have siblings acting | |
11311 | * like a single runqueue and hence we need to consider runnable | |
cc00c198 | 11312 | * tasks on this CPU and the forced idle CPU. Ideally, we should |
8039e96f | 11313 | * go through the forced idle rq, but that would be a perf hit. |
cc00c198 | 11314 | * We can assume that the forced idle CPU has at least |
8039e96f | 11315 | * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check |
cc00c198 | 11316 | * if we need to give up the CPU. |
8039e96f | 11317 | */ |
4feee7d1 | 11318 | if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 && |
8039e96f VP |
11319 | __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE)) |
11320 | resched_curr(rq); | |
11321 | } | |
c6047c2e JFG |
11322 | |
11323 | /* | |
11324 | * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed. | |
11325 | */ | |
11326 | static void se_fi_update(struct sched_entity *se, unsigned int fi_seq, bool forceidle) | |
11327 | { | |
11328 | for_each_sched_entity(se) { | |
11329 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11330 | ||
11331 | if (forceidle) { | |
11332 | if (cfs_rq->forceidle_seq == fi_seq) | |
11333 | break; | |
11334 | cfs_rq->forceidle_seq = fi_seq; | |
11335 | } | |
11336 | ||
11337 | cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime; | |
11338 | } | |
11339 | } | |
11340 | ||
11341 | void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi) | |
11342 | { | |
11343 | struct sched_entity *se = &p->se; | |
11344 | ||
11345 | if (p->sched_class != &fair_sched_class) | |
11346 | return; | |
11347 | ||
11348 | se_fi_update(se, rq->core->core_forceidle_seq, in_fi); | |
11349 | } | |
11350 | ||
11351 | bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool in_fi) | |
11352 | { | |
11353 | struct rq *rq = task_rq(a); | |
11354 | struct sched_entity *sea = &a->se; | |
11355 | struct sched_entity *seb = &b->se; | |
11356 | struct cfs_rq *cfs_rqa; | |
11357 | struct cfs_rq *cfs_rqb; | |
11358 | s64 delta; | |
11359 | ||
11360 | SCHED_WARN_ON(task_rq(b)->core != rq->core); | |
11361 | ||
11362 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
11363 | /* | |
11364 | * Find an se in the hierarchy for tasks a and b, such that the se's | |
11365 | * are immediate siblings. | |
11366 | */ | |
11367 | while (sea->cfs_rq->tg != seb->cfs_rq->tg) { | |
11368 | int sea_depth = sea->depth; | |
11369 | int seb_depth = seb->depth; | |
11370 | ||
11371 | if (sea_depth >= seb_depth) | |
11372 | sea = parent_entity(sea); | |
11373 | if (sea_depth <= seb_depth) | |
11374 | seb = parent_entity(seb); | |
11375 | } | |
11376 | ||
11377 | se_fi_update(sea, rq->core->core_forceidle_seq, in_fi); | |
11378 | se_fi_update(seb, rq->core->core_forceidle_seq, in_fi); | |
11379 | ||
11380 | cfs_rqa = sea->cfs_rq; | |
11381 | cfs_rqb = seb->cfs_rq; | |
11382 | #else | |
11383 | cfs_rqa = &task_rq(a)->cfs; | |
11384 | cfs_rqb = &task_rq(b)->cfs; | |
11385 | #endif | |
11386 | ||
11387 | /* | |
11388 | * Find delta after normalizing se's vruntime with its cfs_rq's | |
11389 | * min_vruntime_fi, which would have been updated in prior calls | |
11390 | * to se_fi_update(). | |
11391 | */ | |
11392 | delta = (s64)(sea->vruntime - seb->vruntime) + | |
11393 | (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi); | |
11394 | ||
11395 | return delta > 0; | |
11396 | } | |
8039e96f VP |
11397 | #else |
11398 | static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {} | |
11399 | #endif | |
11400 | ||
bf0f6f24 | 11401 | /* |
d84b3131 FW |
11402 | * scheduler tick hitting a task of our scheduling class. |
11403 | * | |
11404 | * NOTE: This function can be called remotely by the tick offload that | |
11405 | * goes along full dynticks. Therefore no local assumption can be made | |
11406 | * and everything must be accessed through the @rq and @curr passed in | |
11407 | * parameters. | |
bf0f6f24 | 11408 | */ |
8f4d37ec | 11409 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
11410 | { |
11411 | struct cfs_rq *cfs_rq; | |
11412 | struct sched_entity *se = &curr->se; | |
11413 | ||
11414 | for_each_sched_entity(se) { | |
11415 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 11416 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 11417 | } |
18bf2805 | 11418 | |
b52da86e | 11419 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 11420 | task_tick_numa(rq, curr); |
3b1baa64 MR |
11421 | |
11422 | update_misfit_status(curr, rq); | |
2802bf3c | 11423 | update_overutilized_status(task_rq(curr)); |
8039e96f VP |
11424 | |
11425 | task_tick_core(rq, curr); | |
bf0f6f24 IM |
11426 | } |
11427 | ||
11428 | /* | |
cd29fe6f PZ |
11429 | * called on fork with the child task as argument from the parent's context |
11430 | * - child not yet on the tasklist | |
11431 | * - preemption disabled | |
bf0f6f24 | 11432 | */ |
cd29fe6f | 11433 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 11434 | { |
4fc420c9 DN |
11435 | struct cfs_rq *cfs_rq; |
11436 | struct sched_entity *se = &p->se, *curr; | |
cd29fe6f | 11437 | struct rq *rq = this_rq(); |
8a8c69c3 | 11438 | struct rq_flags rf; |
bf0f6f24 | 11439 | |
8a8c69c3 | 11440 | rq_lock(rq, &rf); |
861d034e PZ |
11441 | update_rq_clock(rq); |
11442 | ||
4fc420c9 DN |
11443 | cfs_rq = task_cfs_rq(current); |
11444 | curr = cfs_rq->curr; | |
e210bffd PZ |
11445 | if (curr) { |
11446 | update_curr(cfs_rq); | |
b5d9d734 | 11447 | se->vruntime = curr->vruntime; |
e210bffd | 11448 | } |
aeb73b04 | 11449 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 11450 | |
cd29fe6f | 11451 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 11452 | /* |
edcb60a3 IM |
11453 | * Upon rescheduling, sched_class::put_prev_task() will place |
11454 | * 'current' within the tree based on its new key value. | |
11455 | */ | |
4d78e7b6 | 11456 | swap(curr->vruntime, se->vruntime); |
8875125e | 11457 | resched_curr(rq); |
4d78e7b6 | 11458 | } |
bf0f6f24 | 11459 | |
88ec22d3 | 11460 | se->vruntime -= cfs_rq->min_vruntime; |
8a8c69c3 | 11461 | rq_unlock(rq, &rf); |
bf0f6f24 IM |
11462 | } |
11463 | ||
cb469845 SR |
11464 | /* |
11465 | * Priority of the task has changed. Check to see if we preempt | |
11466 | * the current task. | |
11467 | */ | |
da7a735e PZ |
11468 | static void |
11469 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 11470 | { |
da0c1e65 | 11471 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
11472 | return; |
11473 | ||
7c2e8bbd FW |
11474 | if (rq->cfs.nr_running == 1) |
11475 | return; | |
11476 | ||
cb469845 SR |
11477 | /* |
11478 | * Reschedule if we are currently running on this runqueue and | |
11479 | * our priority decreased, or if we are not currently running on | |
11480 | * this runqueue and our priority is higher than the current's | |
11481 | */ | |
65bcf072 | 11482 | if (task_current(rq, p)) { |
cb469845 | 11483 | if (p->prio > oldprio) |
8875125e | 11484 | resched_curr(rq); |
cb469845 | 11485 | } else |
15afe09b | 11486 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
11487 | } |
11488 | ||
daa59407 | 11489 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
11490 | { |
11491 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
11492 | |
11493 | /* | |
daa59407 BP |
11494 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
11495 | * the dequeue_entity(.flags=0) will already have normalized the | |
11496 | * vruntime. | |
11497 | */ | |
11498 | if (p->on_rq) | |
11499 | return true; | |
11500 | ||
11501 | /* | |
11502 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
11503 | * But there are some cases where it has already been normalized: | |
da7a735e | 11504 | * |
daa59407 BP |
11505 | * - A forked child which is waiting for being woken up by |
11506 | * wake_up_new_task(). | |
11507 | * - A task which has been woken up by try_to_wake_up() and | |
11508 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 11509 | */ |
d0cdb3ce | 11510 | if (!se->sum_exec_runtime || |
2f064a59 | 11511 | (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup)) |
daa59407 BP |
11512 | return true; |
11513 | ||
11514 | return false; | |
11515 | } | |
11516 | ||
09a43ace VG |
11517 | #ifdef CONFIG_FAIR_GROUP_SCHED |
11518 | /* | |
11519 | * Propagate the changes of the sched_entity across the tg tree to make it | |
11520 | * visible to the root | |
11521 | */ | |
11522 | static void propagate_entity_cfs_rq(struct sched_entity *se) | |
11523 | { | |
51bf903b CZ |
11524 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
11525 | ||
11526 | if (cfs_rq_throttled(cfs_rq)) | |
11527 | return; | |
09a43ace | 11528 | |
51bf903b CZ |
11529 | if (!throttled_hierarchy(cfs_rq)) |
11530 | list_add_leaf_cfs_rq(cfs_rq); | |
0258bdfa | 11531 | |
09a43ace VG |
11532 | /* Start to propagate at parent */ |
11533 | se = se->parent; | |
11534 | ||
11535 | for_each_sched_entity(se) { | |
11536 | cfs_rq = cfs_rq_of(se); | |
11537 | ||
51bf903b | 11538 | update_load_avg(cfs_rq, se, UPDATE_TG); |
09a43ace | 11539 | |
51bf903b | 11540 | if (cfs_rq_throttled(cfs_rq)) |
0258bdfa | 11541 | break; |
51bf903b CZ |
11542 | |
11543 | if (!throttled_hierarchy(cfs_rq)) | |
11544 | list_add_leaf_cfs_rq(cfs_rq); | |
09a43ace VG |
11545 | } |
11546 | } | |
11547 | #else | |
11548 | static void propagate_entity_cfs_rq(struct sched_entity *se) { } | |
11549 | #endif | |
11550 | ||
df217913 | 11551 | static void detach_entity_cfs_rq(struct sched_entity *se) |
daa59407 | 11552 | { |
daa59407 BP |
11553 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
11554 | ||
7e2edaf6 CZ |
11555 | #ifdef CONFIG_SMP |
11556 | /* | |
11557 | * In case the task sched_avg hasn't been attached: | |
11558 | * - A forked task which hasn't been woken up by wake_up_new_task(). | |
11559 | * - A task which has been woken up by try_to_wake_up() but is | |
11560 | * waiting for actually being woken up by sched_ttwu_pending(). | |
11561 | */ | |
11562 | if (!se->avg.last_update_time) | |
11563 | return; | |
11564 | #endif | |
11565 | ||
9d89c257 | 11566 | /* Catch up with the cfs_rq and remove our load when we leave */ |
88c0616e | 11567 | update_load_avg(cfs_rq, se, 0); |
a05e8c51 | 11568 | detach_entity_load_avg(cfs_rq, se); |
fe749158 | 11569 | update_tg_load_avg(cfs_rq); |
09a43ace | 11570 | propagate_entity_cfs_rq(se); |
da7a735e PZ |
11571 | } |
11572 | ||
df217913 | 11573 | static void attach_entity_cfs_rq(struct sched_entity *se) |
cb469845 | 11574 | { |
daa59407 | 11575 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a | 11576 | |
df217913 | 11577 | /* Synchronize entity with its cfs_rq */ |
88c0616e | 11578 | update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); |
a4f9a0e5 | 11579 | attach_entity_load_avg(cfs_rq, se); |
fe749158 | 11580 | update_tg_load_avg(cfs_rq); |
09a43ace | 11581 | propagate_entity_cfs_rq(se); |
df217913 VG |
11582 | } |
11583 | ||
11584 | static void detach_task_cfs_rq(struct task_struct *p) | |
11585 | { | |
11586 | struct sched_entity *se = &p->se; | |
11587 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11588 | ||
11589 | if (!vruntime_normalized(p)) { | |
11590 | /* | |
11591 | * Fix up our vruntime so that the current sleep doesn't | |
11592 | * cause 'unlimited' sleep bonus. | |
11593 | */ | |
11594 | place_entity(cfs_rq, se, 0); | |
11595 | se->vruntime -= cfs_rq->min_vruntime; | |
11596 | } | |
11597 | ||
11598 | detach_entity_cfs_rq(se); | |
11599 | } | |
11600 | ||
11601 | static void attach_task_cfs_rq(struct task_struct *p) | |
11602 | { | |
11603 | struct sched_entity *se = &p->se; | |
11604 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11605 | ||
11606 | attach_entity_cfs_rq(se); | |
daa59407 BP |
11607 | |
11608 | if (!vruntime_normalized(p)) | |
11609 | se->vruntime += cfs_rq->min_vruntime; | |
11610 | } | |
6efdb105 | 11611 | |
daa59407 BP |
11612 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
11613 | { | |
11614 | detach_task_cfs_rq(p); | |
11615 | } | |
11616 | ||
11617 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
11618 | { | |
11619 | attach_task_cfs_rq(p); | |
7855a35a | 11620 | |
daa59407 | 11621 | if (task_on_rq_queued(p)) { |
7855a35a | 11622 | /* |
daa59407 BP |
11623 | * We were most likely switched from sched_rt, so |
11624 | * kick off the schedule if running, otherwise just see | |
11625 | * if we can still preempt the current task. | |
7855a35a | 11626 | */ |
65bcf072 | 11627 | if (task_current(rq, p)) |
daa59407 BP |
11628 | resched_curr(rq); |
11629 | else | |
11630 | check_preempt_curr(rq, p, 0); | |
7855a35a | 11631 | } |
cb469845 SR |
11632 | } |
11633 | ||
83b699ed SV |
11634 | /* Account for a task changing its policy or group. |
11635 | * | |
11636 | * This routine is mostly called to set cfs_rq->curr field when a task | |
11637 | * migrates between groups/classes. | |
11638 | */ | |
a0e813f2 | 11639 | static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first) |
83b699ed | 11640 | { |
03b7fad1 PZ |
11641 | struct sched_entity *se = &p->se; |
11642 | ||
11643 | #ifdef CONFIG_SMP | |
11644 | if (task_on_rq_queued(p)) { | |
11645 | /* | |
11646 | * Move the next running task to the front of the list, so our | |
11647 | * cfs_tasks list becomes MRU one. | |
11648 | */ | |
11649 | list_move(&se->group_node, &rq->cfs_tasks); | |
11650 | } | |
11651 | #endif | |
83b699ed | 11652 | |
ec12cb7f PT |
11653 | for_each_sched_entity(se) { |
11654 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11655 | ||
11656 | set_next_entity(cfs_rq, se); | |
11657 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
11658 | account_cfs_rq_runtime(cfs_rq, 0); | |
11659 | } | |
83b699ed SV |
11660 | } |
11661 | ||
029632fb PZ |
11662 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
11663 | { | |
bfb06889 | 11664 | cfs_rq->tasks_timeline = RB_ROOT_CACHED; |
d05b4305 | 11665 | u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20))); |
141965c7 | 11666 | #ifdef CONFIG_SMP |
2a2f5d4e | 11667 | raw_spin_lock_init(&cfs_rq->removed.lock); |
9ee474f5 | 11668 | #endif |
029632fb PZ |
11669 | } |
11670 | ||
810b3817 | 11671 | #ifdef CONFIG_FAIR_GROUP_SCHED |
39c42611 | 11672 | static void task_change_group_fair(struct task_struct *p) |
810b3817 | 11673 | { |
df16b71c CZ |
11674 | /* |
11675 | * We couldn't detach or attach a forked task which | |
11676 | * hasn't been woken up by wake_up_new_task(). | |
11677 | */ | |
11678 | if (READ_ONCE(p->__state) == TASK_NEW) | |
11679 | return; | |
11680 | ||
daa59407 | 11681 | detach_task_cfs_rq(p); |
6efdb105 BP |
11682 | |
11683 | #ifdef CONFIG_SMP | |
11684 | /* Tell se's cfs_rq has been changed -- migrated */ | |
11685 | p->se.avg.last_update_time = 0; | |
11686 | #endif | |
5d6da83c | 11687 | set_task_rq(p, task_cpu(p)); |
daa59407 | 11688 | attach_task_cfs_rq(p); |
810b3817 | 11689 | } |
029632fb PZ |
11690 | |
11691 | void free_fair_sched_group(struct task_group *tg) | |
11692 | { | |
11693 | int i; | |
11694 | ||
029632fb PZ |
11695 | for_each_possible_cpu(i) { |
11696 | if (tg->cfs_rq) | |
11697 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 11698 | if (tg->se) |
029632fb PZ |
11699 | kfree(tg->se[i]); |
11700 | } | |
11701 | ||
11702 | kfree(tg->cfs_rq); | |
11703 | kfree(tg->se); | |
11704 | } | |
11705 | ||
11706 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
11707 | { | |
029632fb | 11708 | struct sched_entity *se; |
b7fa30c9 | 11709 | struct cfs_rq *cfs_rq; |
029632fb PZ |
11710 | int i; |
11711 | ||
6396bb22 | 11712 | tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); |
029632fb PZ |
11713 | if (!tg->cfs_rq) |
11714 | goto err; | |
6396bb22 | 11715 | tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); |
029632fb PZ |
11716 | if (!tg->se) |
11717 | goto err; | |
11718 | ||
11719 | tg->shares = NICE_0_LOAD; | |
11720 | ||
11721 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
11722 | ||
11723 | for_each_possible_cpu(i) { | |
11724 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
11725 | GFP_KERNEL, cpu_to_node(i)); | |
11726 | if (!cfs_rq) | |
11727 | goto err; | |
11728 | ||
ceeadb83 | 11729 | se = kzalloc_node(sizeof(struct sched_entity_stats), |
029632fb PZ |
11730 | GFP_KERNEL, cpu_to_node(i)); |
11731 | if (!se) | |
11732 | goto err_free_rq; | |
11733 | ||
11734 | init_cfs_rq(cfs_rq); | |
11735 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 11736 | init_entity_runnable_average(se); |
029632fb PZ |
11737 | } |
11738 | ||
11739 | return 1; | |
11740 | ||
11741 | err_free_rq: | |
11742 | kfree(cfs_rq); | |
11743 | err: | |
11744 | return 0; | |
11745 | } | |
11746 | ||
8663e24d PZ |
11747 | void online_fair_sched_group(struct task_group *tg) |
11748 | { | |
11749 | struct sched_entity *se; | |
a46d14ec | 11750 | struct rq_flags rf; |
8663e24d PZ |
11751 | struct rq *rq; |
11752 | int i; | |
11753 | ||
11754 | for_each_possible_cpu(i) { | |
11755 | rq = cpu_rq(i); | |
11756 | se = tg->se[i]; | |
a46d14ec | 11757 | rq_lock_irq(rq, &rf); |
4126bad6 | 11758 | update_rq_clock(rq); |
d0326691 | 11759 | attach_entity_cfs_rq(se); |
55e16d30 | 11760 | sync_throttle(tg, i); |
a46d14ec | 11761 | rq_unlock_irq(rq, &rf); |
8663e24d PZ |
11762 | } |
11763 | } | |
11764 | ||
6fe1f348 | 11765 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 11766 | { |
029632fb | 11767 | unsigned long flags; |
6fe1f348 PZ |
11768 | struct rq *rq; |
11769 | int cpu; | |
029632fb | 11770 | |
b027789e MK |
11771 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); |
11772 | ||
6fe1f348 PZ |
11773 | for_each_possible_cpu(cpu) { |
11774 | if (tg->se[cpu]) | |
11775 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 11776 | |
6fe1f348 PZ |
11777 | /* |
11778 | * Only empty task groups can be destroyed; so we can speculatively | |
11779 | * check on_list without danger of it being re-added. | |
11780 | */ | |
11781 | if (!tg->cfs_rq[cpu]->on_list) | |
11782 | continue; | |
11783 | ||
11784 | rq = cpu_rq(cpu); | |
11785 | ||
5cb9eaa3 | 11786 | raw_spin_rq_lock_irqsave(rq, flags); |
6fe1f348 | 11787 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); |
5cb9eaa3 | 11788 | raw_spin_rq_unlock_irqrestore(rq, flags); |
6fe1f348 | 11789 | } |
029632fb PZ |
11790 | } |
11791 | ||
11792 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
11793 | struct sched_entity *se, int cpu, | |
11794 | struct sched_entity *parent) | |
11795 | { | |
11796 | struct rq *rq = cpu_rq(cpu); | |
11797 | ||
11798 | cfs_rq->tg = tg; | |
11799 | cfs_rq->rq = rq; | |
029632fb PZ |
11800 | init_cfs_rq_runtime(cfs_rq); |
11801 | ||
11802 | tg->cfs_rq[cpu] = cfs_rq; | |
11803 | tg->se[cpu] = se; | |
11804 | ||
11805 | /* se could be NULL for root_task_group */ | |
11806 | if (!se) | |
11807 | return; | |
11808 | ||
fed14d45 | 11809 | if (!parent) { |
029632fb | 11810 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
11811 | se->depth = 0; |
11812 | } else { | |
029632fb | 11813 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
11814 | se->depth = parent->depth + 1; |
11815 | } | |
029632fb PZ |
11816 | |
11817 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
11818 | /* guarantee group entities always have weight */ |
11819 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
11820 | se->parent = parent; |
11821 | } | |
11822 | ||
11823 | static DEFINE_MUTEX(shares_mutex); | |
11824 | ||
30400039 | 11825 | static int __sched_group_set_shares(struct task_group *tg, unsigned long shares) |
029632fb PZ |
11826 | { |
11827 | int i; | |
029632fb | 11828 | |
30400039 JD |
11829 | lockdep_assert_held(&shares_mutex); |
11830 | ||
029632fb PZ |
11831 | /* |
11832 | * We can't change the weight of the root cgroup. | |
11833 | */ | |
11834 | if (!tg->se[0]) | |
11835 | return -EINVAL; | |
11836 | ||
11837 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
11838 | ||
029632fb | 11839 | if (tg->shares == shares) |
30400039 | 11840 | return 0; |
029632fb PZ |
11841 | |
11842 | tg->shares = shares; | |
11843 | for_each_possible_cpu(i) { | |
11844 | struct rq *rq = cpu_rq(i); | |
8a8c69c3 PZ |
11845 | struct sched_entity *se = tg->se[i]; |
11846 | struct rq_flags rf; | |
029632fb | 11847 | |
029632fb | 11848 | /* Propagate contribution to hierarchy */ |
8a8c69c3 | 11849 | rq_lock_irqsave(rq, &rf); |
71b1da46 | 11850 | update_rq_clock(rq); |
89ee048f | 11851 | for_each_sched_entity(se) { |
88c0616e | 11852 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
1ea6c46a | 11853 | update_cfs_group(se); |
89ee048f | 11854 | } |
8a8c69c3 | 11855 | rq_unlock_irqrestore(rq, &rf); |
029632fb PZ |
11856 | } |
11857 | ||
30400039 JD |
11858 | return 0; |
11859 | } | |
11860 | ||
11861 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
11862 | { | |
11863 | int ret; | |
11864 | ||
11865 | mutex_lock(&shares_mutex); | |
11866 | if (tg_is_idle(tg)) | |
11867 | ret = -EINVAL; | |
11868 | else | |
11869 | ret = __sched_group_set_shares(tg, shares); | |
11870 | mutex_unlock(&shares_mutex); | |
11871 | ||
11872 | return ret; | |
11873 | } | |
11874 | ||
11875 | int sched_group_set_idle(struct task_group *tg, long idle) | |
11876 | { | |
11877 | int i; | |
11878 | ||
11879 | if (tg == &root_task_group) | |
11880 | return -EINVAL; | |
11881 | ||
11882 | if (idle < 0 || idle > 1) | |
11883 | return -EINVAL; | |
11884 | ||
11885 | mutex_lock(&shares_mutex); | |
11886 | ||
11887 | if (tg->idle == idle) { | |
11888 | mutex_unlock(&shares_mutex); | |
11889 | return 0; | |
11890 | } | |
11891 | ||
11892 | tg->idle = idle; | |
11893 | ||
11894 | for_each_possible_cpu(i) { | |
11895 | struct rq *rq = cpu_rq(i); | |
11896 | struct sched_entity *se = tg->se[i]; | |
a480adde | 11897 | struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i]; |
30400039 JD |
11898 | bool was_idle = cfs_rq_is_idle(grp_cfs_rq); |
11899 | long idle_task_delta; | |
11900 | struct rq_flags rf; | |
11901 | ||
11902 | rq_lock_irqsave(rq, &rf); | |
11903 | ||
11904 | grp_cfs_rq->idle = idle; | |
11905 | if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq))) | |
11906 | goto next_cpu; | |
11907 | ||
a480adde JD |
11908 | if (se->on_rq) { |
11909 | parent_cfs_rq = cfs_rq_of(se); | |
11910 | if (cfs_rq_is_idle(grp_cfs_rq)) | |
11911 | parent_cfs_rq->idle_nr_running++; | |
11912 | else | |
11913 | parent_cfs_rq->idle_nr_running--; | |
11914 | } | |
11915 | ||
30400039 JD |
11916 | idle_task_delta = grp_cfs_rq->h_nr_running - |
11917 | grp_cfs_rq->idle_h_nr_running; | |
11918 | if (!cfs_rq_is_idle(grp_cfs_rq)) | |
11919 | idle_task_delta *= -1; | |
11920 | ||
11921 | for_each_sched_entity(se) { | |
11922 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
11923 | ||
11924 | if (!se->on_rq) | |
11925 | break; | |
11926 | ||
11927 | cfs_rq->idle_h_nr_running += idle_task_delta; | |
11928 | ||
11929 | /* Already accounted at parent level and above. */ | |
11930 | if (cfs_rq_is_idle(cfs_rq)) | |
11931 | break; | |
11932 | } | |
11933 | ||
11934 | next_cpu: | |
11935 | rq_unlock_irqrestore(rq, &rf); | |
11936 | } | |
11937 | ||
11938 | /* Idle groups have minimum weight. */ | |
11939 | if (tg_is_idle(tg)) | |
11940 | __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO)); | |
11941 | else | |
11942 | __sched_group_set_shares(tg, NICE_0_LOAD); | |
11943 | ||
029632fb PZ |
11944 | mutex_unlock(&shares_mutex); |
11945 | return 0; | |
11946 | } | |
30400039 | 11947 | |
029632fb PZ |
11948 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
11949 | ||
11950 | void free_fair_sched_group(struct task_group *tg) { } | |
11951 | ||
11952 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
11953 | { | |
11954 | return 1; | |
11955 | } | |
11956 | ||
8663e24d PZ |
11957 | void online_fair_sched_group(struct task_group *tg) { } |
11958 | ||
6fe1f348 | 11959 | void unregister_fair_sched_group(struct task_group *tg) { } |
029632fb PZ |
11960 | |
11961 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
11962 | ||
810b3817 | 11963 | |
6d686f45 | 11964 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
11965 | { |
11966 | struct sched_entity *se = &task->se; | |
0d721cea PW |
11967 | unsigned int rr_interval = 0; |
11968 | ||
11969 | /* | |
11970 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
11971 | * idle runqueue: | |
11972 | */ | |
0d721cea | 11973 | if (rq->cfs.load.weight) |
a59f4e07 | 11974 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
11975 | |
11976 | return rr_interval; | |
11977 | } | |
11978 | ||
bf0f6f24 IM |
11979 | /* |
11980 | * All the scheduling class methods: | |
11981 | */ | |
43c31ac0 PZ |
11982 | DEFINE_SCHED_CLASS(fair) = { |
11983 | ||
bf0f6f24 IM |
11984 | .enqueue_task = enqueue_task_fair, |
11985 | .dequeue_task = dequeue_task_fair, | |
11986 | .yield_task = yield_task_fair, | |
d95f4122 | 11987 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 11988 | |
2e09bf55 | 11989 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 | 11990 | |
98c2f700 | 11991 | .pick_next_task = __pick_next_task_fair, |
bf0f6f24 | 11992 | .put_prev_task = put_prev_task_fair, |
03b7fad1 | 11993 | .set_next_task = set_next_task_fair, |
bf0f6f24 | 11994 | |
681f3e68 | 11995 | #ifdef CONFIG_SMP |
6e2df058 | 11996 | .balance = balance_fair, |
21f56ffe | 11997 | .pick_task = pick_task_fair, |
4ce72a2c | 11998 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 11999 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 12000 | |
0bcdcf28 CE |
12001 | .rq_online = rq_online_fair, |
12002 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 12003 | |
12695578 | 12004 | .task_dead = task_dead_fair, |
c5b28038 | 12005 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 12006 | #endif |
bf0f6f24 | 12007 | |
bf0f6f24 | 12008 | .task_tick = task_tick_fair, |
cd29fe6f | 12009 | .task_fork = task_fork_fair, |
cb469845 SR |
12010 | |
12011 | .prio_changed = prio_changed_fair, | |
da7a735e | 12012 | .switched_from = switched_from_fair, |
cb469845 | 12013 | .switched_to = switched_to_fair, |
810b3817 | 12014 | |
0d721cea PW |
12015 | .get_rr_interval = get_rr_interval_fair, |
12016 | ||
6e998916 SG |
12017 | .update_curr = update_curr_fair, |
12018 | ||
810b3817 | 12019 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 12020 | .task_change_group = task_change_group_fair, |
810b3817 | 12021 | #endif |
982d9cdc PB |
12022 | |
12023 | #ifdef CONFIG_UCLAMP_TASK | |
12024 | .uclamp_enabled = 1, | |
12025 | #endif | |
bf0f6f24 IM |
12026 | }; |
12027 | ||
12028 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 12029 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 12030 | { |
039ae8bc | 12031 | struct cfs_rq *cfs_rq, *pos; |
bf0f6f24 | 12032 | |
5973e5b9 | 12033 | rcu_read_lock(); |
039ae8bc | 12034 | for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos) |
5cef9eca | 12035 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 12036 | rcu_read_unlock(); |
bf0f6f24 | 12037 | } |
397f2378 SD |
12038 | |
12039 | #ifdef CONFIG_NUMA_BALANCING | |
12040 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
12041 | { | |
12042 | int node; | |
12043 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
cb361d8c | 12044 | struct numa_group *ng; |
397f2378 | 12045 | |
cb361d8c JH |
12046 | rcu_read_lock(); |
12047 | ng = rcu_dereference(p->numa_group); | |
397f2378 SD |
12048 | for_each_online_node(node) { |
12049 | if (p->numa_faults) { | |
12050 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
12051 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
12052 | } | |
cb361d8c JH |
12053 | if (ng) { |
12054 | gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
12055 | gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
397f2378 SD |
12056 | } |
12057 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
12058 | } | |
cb361d8c | 12059 | rcu_read_unlock(); |
397f2378 SD |
12060 | } |
12061 | #endif /* CONFIG_NUMA_BALANCING */ | |
12062 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
12063 | |
12064 | __init void init_sched_fair_class(void) | |
12065 | { | |
12066 | #ifdef CONFIG_SMP | |
18c31c97 BH |
12067 | int i; |
12068 | ||
12069 | for_each_possible_cpu(i) { | |
12070 | zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i)); | |
12071 | zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i)); | |
12072 | } | |
12073 | ||
029632fb PZ |
12074 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); |
12075 | ||
3451d024 | 12076 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 12077 | nohz.next_balance = jiffies; |
f643ea22 | 12078 | nohz.next_blocked = jiffies; |
029632fb | 12079 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
12080 | #endif |
12081 | #endif /* SMP */ | |
12082 | ||
12083 | } |