]>
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 | */ |
325ea10c | 23 | #include "sched.h" |
029632fb PZ |
24 | |
25 | #include <trace/events/sched.h> | |
26 | ||
bf0f6f24 | 27 | /* |
21805085 | 28 | * Targeted preemption latency for CPU-bound tasks: |
bf0f6f24 | 29 | * |
21805085 | 30 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
31 | * 'timeslice length' - timeslices in CFS are of variable length |
32 | * and have no persistent notion like in traditional, time-slice | |
33 | * based scheduling concepts. | |
bf0f6f24 | 34 | * |
d274a4ce IM |
35 | * (to see the precise effective timeslice length of your workload, |
36 | * run vmstat and monitor the context-switches (cs) field) | |
2b4d5b25 IM |
37 | * |
38 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 39 | */ |
2b4d5b25 | 40 | unsigned int sysctl_sched_latency = 6000000ULL; |
ed8885a1 | 41 | static unsigned int normalized_sysctl_sched_latency = 6000000ULL; |
2bd8e6d4 | 42 | |
1983a922 CE |
43 | /* |
44 | * The initial- and re-scaling of tunables is configurable | |
1983a922 CE |
45 | * |
46 | * Options are: | |
2b4d5b25 IM |
47 | * |
48 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
49 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
50 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
51 | * | |
52 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
1983a922 | 53 | */ |
2b4d5b25 | 54 | enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG; |
1983a922 | 55 | |
2bd8e6d4 | 56 | /* |
b2be5e96 | 57 | * Minimal preemption granularity for CPU-bound tasks: |
2b4d5b25 | 58 | * |
864616ee | 59 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 60 | */ |
ed8885a1 MS |
61 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
62 | static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
63 | |
64 | /* | |
2b4d5b25 | 65 | * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity |
b2be5e96 | 66 | */ |
0bf377bb | 67 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
68 | |
69 | /* | |
2bba22c5 | 70 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 71 | * parent will (try to) run first. |
21805085 | 72 | */ |
2bba22c5 | 73 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 74 | |
bf0f6f24 IM |
75 | /* |
76 | * SCHED_OTHER wake-up granularity. | |
bf0f6f24 IM |
77 | * |
78 | * This option delays the preemption effects of decoupled workloads | |
79 | * and reduces their over-scheduling. Synchronous workloads will still | |
80 | * have immediate wakeup/sleep latencies. | |
2b4d5b25 IM |
81 | * |
82 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 83 | */ |
ed8885a1 MS |
84 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
85 | static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; | |
bf0f6f24 | 86 | |
2b4d5b25 | 87 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
da84d961 | 88 | |
afe06efd TC |
89 | #ifdef CONFIG_SMP |
90 | /* | |
97fb7a0a | 91 | * For asym packing, by default the lower numbered CPU has higher priority. |
afe06efd TC |
92 | */ |
93 | int __weak arch_asym_cpu_priority(int cpu) | |
94 | { | |
95 | return -cpu; | |
96 | } | |
6d101ba6 OJ |
97 | |
98 | /* | |
99 | * The margin used when comparing utilization with CPU capacity: | |
100 | * util * margin < capacity * 1024 | |
101 | * | |
102 | * (default: ~20%) | |
103 | */ | |
104 | static unsigned int capacity_margin = 1280; | |
afe06efd TC |
105 | #endif |
106 | ||
ec12cb7f PT |
107 | #ifdef CONFIG_CFS_BANDWIDTH |
108 | /* | |
109 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
110 | * each time a cfs_rq requests quota. | |
111 | * | |
112 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
113 | * to consumption or the quota being specified to be smaller than the slice) | |
114 | * we will always only issue the remaining available time. | |
115 | * | |
2b4d5b25 IM |
116 | * (default: 5 msec, units: microseconds) |
117 | */ | |
118 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
ec12cb7f PT |
119 | #endif |
120 | ||
8527632d PG |
121 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
122 | { | |
123 | lw->weight += inc; | |
124 | lw->inv_weight = 0; | |
125 | } | |
126 | ||
127 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
128 | { | |
129 | lw->weight -= dec; | |
130 | lw->inv_weight = 0; | |
131 | } | |
132 | ||
133 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
134 | { | |
135 | lw->weight = w; | |
136 | lw->inv_weight = 0; | |
137 | } | |
138 | ||
029632fb PZ |
139 | /* |
140 | * Increase the granularity value when there are more CPUs, | |
141 | * because with more CPUs the 'effective latency' as visible | |
142 | * to users decreases. But the relationship is not linear, | |
143 | * so pick a second-best guess by going with the log2 of the | |
144 | * number of CPUs. | |
145 | * | |
146 | * This idea comes from the SD scheduler of Con Kolivas: | |
147 | */ | |
58ac93e4 | 148 | static unsigned int get_update_sysctl_factor(void) |
029632fb | 149 | { |
58ac93e4 | 150 | unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); |
029632fb PZ |
151 | unsigned int factor; |
152 | ||
153 | switch (sysctl_sched_tunable_scaling) { | |
154 | case SCHED_TUNABLESCALING_NONE: | |
155 | factor = 1; | |
156 | break; | |
157 | case SCHED_TUNABLESCALING_LINEAR: | |
158 | factor = cpus; | |
159 | break; | |
160 | case SCHED_TUNABLESCALING_LOG: | |
161 | default: | |
162 | factor = 1 + ilog2(cpus); | |
163 | break; | |
164 | } | |
165 | ||
166 | return factor; | |
167 | } | |
168 | ||
169 | static void update_sysctl(void) | |
170 | { | |
171 | unsigned int factor = get_update_sysctl_factor(); | |
172 | ||
173 | #define SET_SYSCTL(name) \ | |
174 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
175 | SET_SYSCTL(sched_min_granularity); | |
176 | SET_SYSCTL(sched_latency); | |
177 | SET_SYSCTL(sched_wakeup_granularity); | |
178 | #undef SET_SYSCTL | |
179 | } | |
180 | ||
181 | void sched_init_granularity(void) | |
182 | { | |
183 | update_sysctl(); | |
184 | } | |
185 | ||
9dbdb155 | 186 | #define WMULT_CONST (~0U) |
029632fb PZ |
187 | #define WMULT_SHIFT 32 |
188 | ||
9dbdb155 PZ |
189 | static void __update_inv_weight(struct load_weight *lw) |
190 | { | |
191 | unsigned long w; | |
192 | ||
193 | if (likely(lw->inv_weight)) | |
194 | return; | |
195 | ||
196 | w = scale_load_down(lw->weight); | |
197 | ||
198 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
199 | lw->inv_weight = 1; | |
200 | else if (unlikely(!w)) | |
201 | lw->inv_weight = WMULT_CONST; | |
202 | else | |
203 | lw->inv_weight = WMULT_CONST / w; | |
204 | } | |
029632fb PZ |
205 | |
206 | /* | |
9dbdb155 PZ |
207 | * delta_exec * weight / lw.weight |
208 | * OR | |
209 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
210 | * | |
1c3de5e1 | 211 | * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case |
9dbdb155 PZ |
212 | * we're guaranteed shift stays positive because inv_weight is guaranteed to |
213 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
214 | * | |
215 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
216 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 217 | */ |
9dbdb155 | 218 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 219 | { |
9dbdb155 PZ |
220 | u64 fact = scale_load_down(weight); |
221 | int shift = WMULT_SHIFT; | |
029632fb | 222 | |
9dbdb155 | 223 | __update_inv_weight(lw); |
029632fb | 224 | |
9dbdb155 PZ |
225 | if (unlikely(fact >> 32)) { |
226 | while (fact >> 32) { | |
227 | fact >>= 1; | |
228 | shift--; | |
229 | } | |
029632fb PZ |
230 | } |
231 | ||
9dbdb155 PZ |
232 | /* hint to use a 32x32->64 mul */ |
233 | fact = (u64)(u32)fact * lw->inv_weight; | |
029632fb | 234 | |
9dbdb155 PZ |
235 | while (fact >> 32) { |
236 | fact >>= 1; | |
237 | shift--; | |
238 | } | |
029632fb | 239 | |
9dbdb155 | 240 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
241 | } |
242 | ||
243 | ||
244 | const struct sched_class fair_sched_class; | |
a4c2f00f | 245 | |
bf0f6f24 IM |
246 | /************************************************************** |
247 | * CFS operations on generic schedulable entities: | |
248 | */ | |
249 | ||
62160e3f | 250 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 251 | |
62160e3f | 252 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
253 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
254 | { | |
62160e3f | 255 | return cfs_rq->rq; |
bf0f6f24 IM |
256 | } |
257 | ||
8f48894f PZ |
258 | static inline struct task_struct *task_of(struct sched_entity *se) |
259 | { | |
9148a3a1 | 260 | SCHED_WARN_ON(!entity_is_task(se)); |
8f48894f PZ |
261 | return container_of(se, struct task_struct, se); |
262 | } | |
263 | ||
b758149c PZ |
264 | /* Walk up scheduling entities hierarchy */ |
265 | #define for_each_sched_entity(se) \ | |
266 | for (; se; se = se->parent) | |
267 | ||
268 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
269 | { | |
270 | return p->se.cfs_rq; | |
271 | } | |
272 | ||
273 | /* runqueue on which this entity is (to be) queued */ | |
274 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
275 | { | |
276 | return se->cfs_rq; | |
277 | } | |
278 | ||
279 | /* runqueue "owned" by this group */ | |
280 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
281 | { | |
282 | return grp->my_q; | |
283 | } | |
284 | ||
3d4b47b4 PZ |
285 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
286 | { | |
287 | if (!cfs_rq->on_list) { | |
9c2791f9 VG |
288 | struct rq *rq = rq_of(cfs_rq); |
289 | int cpu = cpu_of(rq); | |
67e86250 PT |
290 | /* |
291 | * Ensure we either appear before our parent (if already | |
292 | * enqueued) or force our parent to appear after us when it is | |
9c2791f9 VG |
293 | * enqueued. The fact that we always enqueue bottom-up |
294 | * reduces this to two cases and a special case for the root | |
295 | * cfs_rq. Furthermore, it also means that we will always reset | |
296 | * tmp_alone_branch either when the branch is connected | |
297 | * to a tree or when we reach the beg of the tree | |
67e86250 PT |
298 | */ |
299 | if (cfs_rq->tg->parent && | |
9c2791f9 VG |
300 | cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { |
301 | /* | |
302 | * If parent is already on the list, we add the child | |
303 | * just before. Thanks to circular linked property of | |
304 | * the list, this means to put the child at the tail | |
305 | * of the list that starts by parent. | |
306 | */ | |
307 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
308 | &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); | |
309 | /* | |
310 | * The branch is now connected to its tree so we can | |
311 | * reset tmp_alone_branch to the beginning of the | |
312 | * list. | |
313 | */ | |
314 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
315 | } else if (!cfs_rq->tg->parent) { | |
316 | /* | |
317 | * cfs rq without parent should be put | |
318 | * at the tail of the list. | |
319 | */ | |
67e86250 | 320 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, |
9c2791f9 VG |
321 | &rq->leaf_cfs_rq_list); |
322 | /* | |
323 | * We have reach the beg of a tree so we can reset | |
324 | * tmp_alone_branch to the beginning of the list. | |
325 | */ | |
326 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
327 | } else { | |
328 | /* | |
329 | * The parent has not already been added so we want to | |
330 | * make sure that it will be put after us. | |
331 | * tmp_alone_branch points to the beg of the branch | |
332 | * where we will add parent. | |
333 | */ | |
334 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
335 | rq->tmp_alone_branch); | |
336 | /* | |
337 | * update tmp_alone_branch to points to the new beg | |
338 | * of the branch | |
339 | */ | |
340 | rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; | |
67e86250 | 341 | } |
3d4b47b4 PZ |
342 | |
343 | cfs_rq->on_list = 1; | |
344 | } | |
345 | } | |
346 | ||
347 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
348 | { | |
349 | if (cfs_rq->on_list) { | |
350 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
351 | cfs_rq->on_list = 0; | |
352 | } | |
353 | } | |
354 | ||
c40f7d74 LT |
355 | /* Iterate through all leaf cfs_rq's on a runqueue: */ |
356 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
357 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
b758149c PZ |
358 | |
359 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 360 | static inline struct cfs_rq * |
b758149c PZ |
361 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
362 | { | |
363 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 364 | return se->cfs_rq; |
b758149c | 365 | |
fed14d45 | 366 | return NULL; |
b758149c PZ |
367 | } |
368 | ||
369 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
370 | { | |
371 | return se->parent; | |
372 | } | |
373 | ||
464b7527 PZ |
374 | static void |
375 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
376 | { | |
377 | int se_depth, pse_depth; | |
378 | ||
379 | /* | |
380 | * preemption test can be made between sibling entities who are in the | |
381 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
382 | * both tasks until we find their ancestors who are siblings of common | |
383 | * parent. | |
384 | */ | |
385 | ||
386 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
387 | se_depth = (*se)->depth; |
388 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
389 | |
390 | while (se_depth > pse_depth) { | |
391 | se_depth--; | |
392 | *se = parent_entity(*se); | |
393 | } | |
394 | ||
395 | while (pse_depth > se_depth) { | |
396 | pse_depth--; | |
397 | *pse = parent_entity(*pse); | |
398 | } | |
399 | ||
400 | while (!is_same_group(*se, *pse)) { | |
401 | *se = parent_entity(*se); | |
402 | *pse = parent_entity(*pse); | |
403 | } | |
404 | } | |
405 | ||
8f48894f PZ |
406 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
407 | ||
408 | static inline struct task_struct *task_of(struct sched_entity *se) | |
409 | { | |
410 | return container_of(se, struct task_struct, se); | |
411 | } | |
bf0f6f24 | 412 | |
62160e3f IM |
413 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
414 | { | |
415 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
416 | } |
417 | ||
bf0f6f24 | 418 | |
b758149c PZ |
419 | #define for_each_sched_entity(se) \ |
420 | for (; se; se = NULL) | |
bf0f6f24 | 421 | |
b758149c | 422 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 423 | { |
b758149c | 424 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
425 | } |
426 | ||
b758149c PZ |
427 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
428 | { | |
429 | struct task_struct *p = task_of(se); | |
430 | struct rq *rq = task_rq(p); | |
431 | ||
432 | return &rq->cfs; | |
433 | } | |
434 | ||
435 | /* runqueue "owned" by this group */ | |
436 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
437 | { | |
438 | return NULL; | |
439 | } | |
440 | ||
3d4b47b4 PZ |
441 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
442 | { | |
443 | } | |
444 | ||
445 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
446 | { | |
447 | } | |
448 | ||
c40f7d74 LT |
449 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
450 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
b758149c | 451 | |
b758149c PZ |
452 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
453 | { | |
454 | return NULL; | |
455 | } | |
456 | ||
464b7527 PZ |
457 | static inline void |
458 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
459 | { | |
460 | } | |
461 | ||
b758149c PZ |
462 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
463 | ||
6c16a6dc | 464 | static __always_inline |
9dbdb155 | 465 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
466 | |
467 | /************************************************************** | |
468 | * Scheduling class tree data structure manipulation methods: | |
469 | */ | |
470 | ||
1bf08230 | 471 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 472 | { |
1bf08230 | 473 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 474 | if (delta > 0) |
1bf08230 | 475 | max_vruntime = vruntime; |
02e0431a | 476 | |
1bf08230 | 477 | return max_vruntime; |
02e0431a PZ |
478 | } |
479 | ||
0702e3eb | 480 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
481 | { |
482 | s64 delta = (s64)(vruntime - min_vruntime); | |
483 | if (delta < 0) | |
484 | min_vruntime = vruntime; | |
485 | ||
486 | return min_vruntime; | |
487 | } | |
488 | ||
54fdc581 FC |
489 | static inline int entity_before(struct sched_entity *a, |
490 | struct sched_entity *b) | |
491 | { | |
492 | return (s64)(a->vruntime - b->vruntime) < 0; | |
493 | } | |
494 | ||
1af5f730 PZ |
495 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
496 | { | |
b60205c7 | 497 | struct sched_entity *curr = cfs_rq->curr; |
bfb06889 | 498 | struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); |
b60205c7 | 499 | |
1af5f730 PZ |
500 | u64 vruntime = cfs_rq->min_vruntime; |
501 | ||
b60205c7 PZ |
502 | if (curr) { |
503 | if (curr->on_rq) | |
504 | vruntime = curr->vruntime; | |
505 | else | |
506 | curr = NULL; | |
507 | } | |
1af5f730 | 508 | |
bfb06889 DB |
509 | if (leftmost) { /* non-empty tree */ |
510 | struct sched_entity *se; | |
511 | se = rb_entry(leftmost, struct sched_entity, run_node); | |
1af5f730 | 512 | |
b60205c7 | 513 | if (!curr) |
1af5f730 PZ |
514 | vruntime = se->vruntime; |
515 | else | |
516 | vruntime = min_vruntime(vruntime, se->vruntime); | |
517 | } | |
518 | ||
1bf08230 | 519 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 520 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
521 | #ifndef CONFIG_64BIT |
522 | smp_wmb(); | |
523 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
524 | #endif | |
1af5f730 PZ |
525 | } |
526 | ||
bf0f6f24 IM |
527 | /* |
528 | * Enqueue an entity into the rb-tree: | |
529 | */ | |
0702e3eb | 530 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 531 | { |
bfb06889 | 532 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node; |
bf0f6f24 IM |
533 | struct rb_node *parent = NULL; |
534 | struct sched_entity *entry; | |
bfb06889 | 535 | bool leftmost = true; |
bf0f6f24 IM |
536 | |
537 | /* | |
538 | * Find the right place in the rbtree: | |
539 | */ | |
540 | while (*link) { | |
541 | parent = *link; | |
542 | entry = rb_entry(parent, struct sched_entity, run_node); | |
543 | /* | |
544 | * We dont care about collisions. Nodes with | |
545 | * the same key stay together. | |
546 | */ | |
2bd2d6f2 | 547 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
548 | link = &parent->rb_left; |
549 | } else { | |
550 | link = &parent->rb_right; | |
bfb06889 | 551 | leftmost = false; |
bf0f6f24 IM |
552 | } |
553 | } | |
554 | ||
bf0f6f24 | 555 | rb_link_node(&se->run_node, parent, link); |
bfb06889 DB |
556 | rb_insert_color_cached(&se->run_node, |
557 | &cfs_rq->tasks_timeline, leftmost); | |
bf0f6f24 IM |
558 | } |
559 | ||
0702e3eb | 560 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 561 | { |
bfb06889 | 562 | rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
563 | } |
564 | ||
029632fb | 565 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 566 | { |
bfb06889 | 567 | struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); |
f4b6755f PZ |
568 | |
569 | if (!left) | |
570 | return NULL; | |
571 | ||
572 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
573 | } |
574 | ||
ac53db59 RR |
575 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
576 | { | |
577 | struct rb_node *next = rb_next(&se->run_node); | |
578 | ||
579 | if (!next) | |
580 | return NULL; | |
581 | ||
582 | return rb_entry(next, struct sched_entity, run_node); | |
583 | } | |
584 | ||
585 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 586 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 587 | { |
bfb06889 | 588 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); |
aeb73b04 | 589 | |
70eee74b BS |
590 | if (!last) |
591 | return NULL; | |
7eee3e67 IM |
592 | |
593 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
594 | } |
595 | ||
bf0f6f24 IM |
596 | /************************************************************** |
597 | * Scheduling class statistics methods: | |
598 | */ | |
599 | ||
acb4a848 | 600 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 601 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
602 | loff_t *ppos) |
603 | { | |
8d65af78 | 604 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
58ac93e4 | 605 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
606 | |
607 | if (ret || !write) | |
608 | return ret; | |
609 | ||
610 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
611 | sysctl_sched_min_granularity); | |
612 | ||
acb4a848 CE |
613 | #define WRT_SYSCTL(name) \ |
614 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
615 | WRT_SYSCTL(sched_min_granularity); | |
616 | WRT_SYSCTL(sched_latency); | |
617 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
618 | #undef WRT_SYSCTL |
619 | ||
b2be5e96 PZ |
620 | return 0; |
621 | } | |
622 | #endif | |
647e7cac | 623 | |
a7be37ac | 624 | /* |
f9c0b095 | 625 | * delta /= w |
a7be37ac | 626 | */ |
9dbdb155 | 627 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 628 | { |
f9c0b095 | 629 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 630 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
631 | |
632 | return delta; | |
633 | } | |
634 | ||
647e7cac IM |
635 | /* |
636 | * The idea is to set a period in which each task runs once. | |
637 | * | |
532b1858 | 638 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
639 | * this period because otherwise the slices get too small. |
640 | * | |
641 | * p = (nr <= nl) ? l : l*nr/nl | |
642 | */ | |
4d78e7b6 PZ |
643 | static u64 __sched_period(unsigned long nr_running) |
644 | { | |
8e2b0bf3 BF |
645 | if (unlikely(nr_running > sched_nr_latency)) |
646 | return nr_running * sysctl_sched_min_granularity; | |
647 | else | |
648 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
649 | } |
650 | ||
647e7cac IM |
651 | /* |
652 | * We calculate the wall-time slice from the period by taking a part | |
653 | * proportional to the weight. | |
654 | * | |
f9c0b095 | 655 | * s = p*P[w/rw] |
647e7cac | 656 | */ |
6d0f0ebd | 657 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 658 | { |
0a582440 | 659 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 660 | |
0a582440 | 661 | for_each_sched_entity(se) { |
6272d68c | 662 | struct load_weight *load; |
3104bf03 | 663 | struct load_weight lw; |
6272d68c LM |
664 | |
665 | cfs_rq = cfs_rq_of(se); | |
666 | load = &cfs_rq->load; | |
f9c0b095 | 667 | |
0a582440 | 668 | if (unlikely(!se->on_rq)) { |
3104bf03 | 669 | lw = cfs_rq->load; |
0a582440 MG |
670 | |
671 | update_load_add(&lw, se->load.weight); | |
672 | load = &lw; | |
673 | } | |
9dbdb155 | 674 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
675 | } |
676 | return slice; | |
bf0f6f24 IM |
677 | } |
678 | ||
647e7cac | 679 | /* |
660cc00f | 680 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 681 | * |
f9c0b095 | 682 | * vs = s/w |
647e7cac | 683 | */ |
f9c0b095 | 684 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 685 | { |
f9c0b095 | 686 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
687 | } |
688 | ||
a75cdaa9 | 689 | #ifdef CONFIG_SMP |
c0796298 | 690 | #include "pelt.h" |
283e2ed3 PZ |
691 | #include "sched-pelt.h" |
692 | ||
772bd008 | 693 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee | 694 | static unsigned long task_h_load(struct task_struct *p); |
3b1baa64 | 695 | static unsigned long capacity_of(int cpu); |
fb13c7ee | 696 | |
540247fb YD |
697 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
698 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 699 | { |
540247fb | 700 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 701 | |
f207934f PZ |
702 | memset(sa, 0, sizeof(*sa)); |
703 | ||
b5a9b340 | 704 | /* |
dfcb245e | 705 | * Tasks are initialized with full load to be seen as heavy tasks until |
b5a9b340 | 706 | * they get a chance to stabilize to their real load level. |
dfcb245e | 707 | * Group entities are initialized with zero load to reflect the fact that |
b5a9b340 VG |
708 | * nothing has been attached to the task group yet. |
709 | */ | |
710 | if (entity_is_task(se)) | |
1ea6c46a | 711 | sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight); |
1ea6c46a | 712 | |
f207934f PZ |
713 | se->runnable_weight = se->load.weight; |
714 | ||
9d89c257 | 715 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 716 | } |
7ea241af | 717 | |
7dc603c9 | 718 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
df217913 | 719 | static void attach_entity_cfs_rq(struct sched_entity *se); |
7dc603c9 | 720 | |
2b8c41da YD |
721 | /* |
722 | * With new tasks being created, their initial util_avgs are extrapolated | |
723 | * based on the cfs_rq's current util_avg: | |
724 | * | |
725 | * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight | |
726 | * | |
727 | * However, in many cases, the above util_avg does not give a desired | |
728 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
729 | * as when the series is a harmonic series. | |
730 | * | |
731 | * To solve this problem, we also cap the util_avg of successive tasks to | |
732 | * only 1/2 of the left utilization budget: | |
733 | * | |
8fe5c5a9 | 734 | * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n |
2b8c41da | 735 | * |
8fe5c5a9 | 736 | * where n denotes the nth task and cpu_scale the CPU capacity. |
2b8c41da | 737 | * |
8fe5c5a9 QP |
738 | * For example, for a CPU with 1024 of capacity, a simplest series from |
739 | * the beginning would be like: | |
2b8c41da YD |
740 | * |
741 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
742 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
743 | * | |
744 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
745 | * if util_avg > util_avg_cap. | |
746 | */ | |
747 | void post_init_entity_util_avg(struct sched_entity *se) | |
748 | { | |
749 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
750 | struct sched_avg *sa = &se->avg; | |
8fe5c5a9 QP |
751 | long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq))); |
752 | long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; | |
2b8c41da YD |
753 | |
754 | if (cap > 0) { | |
755 | if (cfs_rq->avg.util_avg != 0) { | |
756 | sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; | |
757 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); | |
758 | ||
759 | if (sa->util_avg > cap) | |
760 | sa->util_avg = cap; | |
761 | } else { | |
762 | sa->util_avg = cap; | |
763 | } | |
2b8c41da | 764 | } |
7dc603c9 PZ |
765 | |
766 | if (entity_is_task(se)) { | |
767 | struct task_struct *p = task_of(se); | |
768 | if (p->sched_class != &fair_sched_class) { | |
769 | /* | |
770 | * For !fair tasks do: | |
771 | * | |
3a123bbb | 772 | update_cfs_rq_load_avg(now, cfs_rq); |
ea14b57e | 773 | attach_entity_load_avg(cfs_rq, se, 0); |
7dc603c9 PZ |
774 | switched_from_fair(rq, p); |
775 | * | |
776 | * such that the next switched_to_fair() has the | |
777 | * expected state. | |
778 | */ | |
df217913 | 779 | se->avg.last_update_time = cfs_rq_clock_task(cfs_rq); |
7dc603c9 PZ |
780 | return; |
781 | } | |
782 | } | |
783 | ||
df217913 | 784 | attach_entity_cfs_rq(se); |
2b8c41da YD |
785 | } |
786 | ||
7dc603c9 | 787 | #else /* !CONFIG_SMP */ |
540247fb | 788 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
789 | { |
790 | } | |
2b8c41da YD |
791 | void post_init_entity_util_avg(struct sched_entity *se) |
792 | { | |
793 | } | |
3d30544f PZ |
794 | static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
795 | { | |
796 | } | |
7dc603c9 | 797 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 798 | |
bf0f6f24 | 799 | /* |
9dbdb155 | 800 | * Update the current task's runtime statistics. |
bf0f6f24 | 801 | */ |
b7cc0896 | 802 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 803 | { |
429d43bc | 804 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 805 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 806 | u64 delta_exec; |
bf0f6f24 IM |
807 | |
808 | if (unlikely(!curr)) | |
809 | return; | |
810 | ||
9dbdb155 PZ |
811 | delta_exec = now - curr->exec_start; |
812 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 813 | return; |
bf0f6f24 | 814 | |
8ebc91d9 | 815 | curr->exec_start = now; |
d842de87 | 816 | |
9dbdb155 PZ |
817 | schedstat_set(curr->statistics.exec_max, |
818 | max(delta_exec, curr->statistics.exec_max)); | |
819 | ||
820 | curr->sum_exec_runtime += delta_exec; | |
ae92882e | 821 | schedstat_add(cfs_rq->exec_clock, delta_exec); |
9dbdb155 PZ |
822 | |
823 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
824 | update_min_vruntime(cfs_rq); | |
825 | ||
d842de87 SV |
826 | if (entity_is_task(curr)) { |
827 | struct task_struct *curtask = task_of(curr); | |
828 | ||
f977bb49 | 829 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d2cc5ed6 | 830 | cgroup_account_cputime(curtask, delta_exec); |
f06febc9 | 831 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 832 | } |
ec12cb7f PT |
833 | |
834 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
835 | } |
836 | ||
6e998916 SG |
837 | static void update_curr_fair(struct rq *rq) |
838 | { | |
839 | update_curr(cfs_rq_of(&rq->curr->se)); | |
840 | } | |
841 | ||
bf0f6f24 | 842 | static inline void |
5870db5b | 843 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 844 | { |
4fa8d299 JP |
845 | u64 wait_start, prev_wait_start; |
846 | ||
847 | if (!schedstat_enabled()) | |
848 | return; | |
849 | ||
850 | wait_start = rq_clock(rq_of(cfs_rq)); | |
851 | prev_wait_start = schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
852 | |
853 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && | |
4fa8d299 JP |
854 | likely(wait_start > prev_wait_start)) |
855 | wait_start -= prev_wait_start; | |
3ea94de1 | 856 | |
2ed41a55 | 857 | __schedstat_set(se->statistics.wait_start, wait_start); |
bf0f6f24 IM |
858 | } |
859 | ||
4fa8d299 | 860 | static inline void |
3ea94de1 JP |
861 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
862 | { | |
863 | struct task_struct *p; | |
cb251765 MG |
864 | u64 delta; |
865 | ||
4fa8d299 JP |
866 | if (!schedstat_enabled()) |
867 | return; | |
868 | ||
869 | delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
870 | |
871 | if (entity_is_task(se)) { | |
872 | p = task_of(se); | |
873 | if (task_on_rq_migrating(p)) { | |
874 | /* | |
875 | * Preserve migrating task's wait time so wait_start | |
876 | * time stamp can be adjusted to accumulate wait time | |
877 | * prior to migration. | |
878 | */ | |
2ed41a55 | 879 | __schedstat_set(se->statistics.wait_start, delta); |
3ea94de1 JP |
880 | return; |
881 | } | |
882 | trace_sched_stat_wait(p, delta); | |
883 | } | |
884 | ||
2ed41a55 | 885 | __schedstat_set(se->statistics.wait_max, |
4fa8d299 | 886 | max(schedstat_val(se->statistics.wait_max), delta)); |
2ed41a55 PZ |
887 | __schedstat_inc(se->statistics.wait_count); |
888 | __schedstat_add(se->statistics.wait_sum, delta); | |
889 | __schedstat_set(se->statistics.wait_start, 0); | |
3ea94de1 | 890 | } |
3ea94de1 | 891 | |
4fa8d299 | 892 | static inline void |
1a3d027c JP |
893 | update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
894 | { | |
895 | struct task_struct *tsk = NULL; | |
4fa8d299 JP |
896 | u64 sleep_start, block_start; |
897 | ||
898 | if (!schedstat_enabled()) | |
899 | return; | |
900 | ||
901 | sleep_start = schedstat_val(se->statistics.sleep_start); | |
902 | block_start = schedstat_val(se->statistics.block_start); | |
1a3d027c JP |
903 | |
904 | if (entity_is_task(se)) | |
905 | tsk = task_of(se); | |
906 | ||
4fa8d299 JP |
907 | if (sleep_start) { |
908 | u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; | |
1a3d027c JP |
909 | |
910 | if ((s64)delta < 0) | |
911 | delta = 0; | |
912 | ||
4fa8d299 | 913 | if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) |
2ed41a55 | 914 | __schedstat_set(se->statistics.sleep_max, delta); |
1a3d027c | 915 | |
2ed41a55 PZ |
916 | __schedstat_set(se->statistics.sleep_start, 0); |
917 | __schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
918 | |
919 | if (tsk) { | |
920 | account_scheduler_latency(tsk, delta >> 10, 1); | |
921 | trace_sched_stat_sleep(tsk, delta); | |
922 | } | |
923 | } | |
4fa8d299 JP |
924 | if (block_start) { |
925 | u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; | |
1a3d027c JP |
926 | |
927 | if ((s64)delta < 0) | |
928 | delta = 0; | |
929 | ||
4fa8d299 | 930 | if (unlikely(delta > schedstat_val(se->statistics.block_max))) |
2ed41a55 | 931 | __schedstat_set(se->statistics.block_max, delta); |
1a3d027c | 932 | |
2ed41a55 PZ |
933 | __schedstat_set(se->statistics.block_start, 0); |
934 | __schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
935 | |
936 | if (tsk) { | |
937 | if (tsk->in_iowait) { | |
2ed41a55 PZ |
938 | __schedstat_add(se->statistics.iowait_sum, delta); |
939 | __schedstat_inc(se->statistics.iowait_count); | |
1a3d027c JP |
940 | trace_sched_stat_iowait(tsk, delta); |
941 | } | |
942 | ||
943 | trace_sched_stat_blocked(tsk, delta); | |
944 | ||
945 | /* | |
946 | * Blocking time is in units of nanosecs, so shift by | |
947 | * 20 to get a milliseconds-range estimation of the | |
948 | * amount of time that the task spent sleeping: | |
949 | */ | |
950 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
951 | profile_hits(SLEEP_PROFILING, | |
952 | (void *)get_wchan(tsk), | |
953 | delta >> 20); | |
954 | } | |
955 | account_scheduler_latency(tsk, delta >> 10, 0); | |
956 | } | |
957 | } | |
3ea94de1 | 958 | } |
3ea94de1 | 959 | |
bf0f6f24 IM |
960 | /* |
961 | * Task is being enqueued - update stats: | |
962 | */ | |
cb251765 | 963 | static inline void |
1a3d027c | 964 | update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 965 | { |
4fa8d299 JP |
966 | if (!schedstat_enabled()) |
967 | return; | |
968 | ||
bf0f6f24 IM |
969 | /* |
970 | * Are we enqueueing a waiting task? (for current tasks | |
971 | * a dequeue/enqueue event is a NOP) | |
972 | */ | |
429d43bc | 973 | if (se != cfs_rq->curr) |
5870db5b | 974 | update_stats_wait_start(cfs_rq, se); |
1a3d027c JP |
975 | |
976 | if (flags & ENQUEUE_WAKEUP) | |
977 | update_stats_enqueue_sleeper(cfs_rq, se); | |
bf0f6f24 IM |
978 | } |
979 | ||
bf0f6f24 | 980 | static inline void |
cb251765 | 981 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 982 | { |
4fa8d299 JP |
983 | |
984 | if (!schedstat_enabled()) | |
985 | return; | |
986 | ||
bf0f6f24 IM |
987 | /* |
988 | * Mark the end of the wait period if dequeueing a | |
989 | * waiting task: | |
990 | */ | |
429d43bc | 991 | if (se != cfs_rq->curr) |
9ef0a961 | 992 | update_stats_wait_end(cfs_rq, se); |
cb251765 | 993 | |
4fa8d299 JP |
994 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
995 | struct task_struct *tsk = task_of(se); | |
cb251765 | 996 | |
4fa8d299 | 997 | if (tsk->state & TASK_INTERRUPTIBLE) |
2ed41a55 | 998 | __schedstat_set(se->statistics.sleep_start, |
4fa8d299 JP |
999 | rq_clock(rq_of(cfs_rq))); |
1000 | if (tsk->state & TASK_UNINTERRUPTIBLE) | |
2ed41a55 | 1001 | __schedstat_set(se->statistics.block_start, |
4fa8d299 | 1002 | rq_clock(rq_of(cfs_rq))); |
cb251765 | 1003 | } |
cb251765 MG |
1004 | } |
1005 | ||
bf0f6f24 IM |
1006 | /* |
1007 | * We are picking a new current task - update its stats: | |
1008 | */ | |
1009 | static inline void | |
79303e9e | 1010 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
1011 | { |
1012 | /* | |
1013 | * We are starting a new run period: | |
1014 | */ | |
78becc27 | 1015 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1016 | } |
1017 | ||
bf0f6f24 IM |
1018 | /************************************************** |
1019 | * Scheduling class queueing methods: | |
1020 | */ | |
1021 | ||
cbee9f88 PZ |
1022 | #ifdef CONFIG_NUMA_BALANCING |
1023 | /* | |
598f0ec0 MG |
1024 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1025 | * calculated based on the tasks virtual memory size and | |
1026 | * numa_balancing_scan_size. | |
cbee9f88 | 1027 | */ |
598f0ec0 MG |
1028 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1029 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1030 | |
1031 | /* Portion of address space to scan in MB */ | |
1032 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1033 | |
4b96a29b PZ |
1034 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1035 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1036 | ||
b5dd77c8 RR |
1037 | struct numa_group { |
1038 | atomic_t refcount; | |
1039 | ||
1040 | spinlock_t lock; /* nr_tasks, tasks */ | |
1041 | int nr_tasks; | |
1042 | pid_t gid; | |
1043 | int active_nodes; | |
1044 | ||
1045 | struct rcu_head rcu; | |
1046 | unsigned long total_faults; | |
1047 | unsigned long max_faults_cpu; | |
1048 | /* | |
1049 | * Faults_cpu is used to decide whether memory should move | |
1050 | * towards the CPU. As a consequence, these stats are weighted | |
1051 | * more by CPU use than by memory faults. | |
1052 | */ | |
1053 | unsigned long *faults_cpu; | |
1054 | unsigned long faults[0]; | |
1055 | }; | |
1056 | ||
1057 | static inline unsigned long group_faults_priv(struct numa_group *ng); | |
1058 | static inline unsigned long group_faults_shared(struct numa_group *ng); | |
1059 | ||
598f0ec0 MG |
1060 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1061 | { | |
1062 | unsigned long rss = 0; | |
1063 | unsigned long nr_scan_pages; | |
1064 | ||
1065 | /* | |
1066 | * Calculations based on RSS as non-present and empty pages are skipped | |
1067 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1068 | * on resident pages | |
1069 | */ | |
1070 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1071 | rss = get_mm_rss(p->mm); | |
1072 | if (!rss) | |
1073 | rss = nr_scan_pages; | |
1074 | ||
1075 | rss = round_up(rss, nr_scan_pages); | |
1076 | return rss / nr_scan_pages; | |
1077 | } | |
1078 | ||
1079 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
1080 | #define MAX_SCAN_WINDOW 2560 | |
1081 | ||
1082 | static unsigned int task_scan_min(struct task_struct *p) | |
1083 | { | |
316c1608 | 1084 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1085 | unsigned int scan, floor; |
1086 | unsigned int windows = 1; | |
1087 | ||
64192658 KT |
1088 | if (scan_size < MAX_SCAN_WINDOW) |
1089 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1090 | floor = 1000 / windows; |
1091 | ||
1092 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1093 | return max_t(unsigned int, floor, scan); | |
1094 | } | |
1095 | ||
b5dd77c8 RR |
1096 | static unsigned int task_scan_start(struct task_struct *p) |
1097 | { | |
1098 | unsigned long smin = task_scan_min(p); | |
1099 | unsigned long period = smin; | |
1100 | ||
1101 | /* Scale the maximum scan period with the amount of shared memory. */ | |
1102 | if (p->numa_group) { | |
1103 | struct numa_group *ng = p->numa_group; | |
1104 | unsigned long shared = group_faults_shared(ng); | |
1105 | unsigned long private = group_faults_priv(ng); | |
1106 | ||
1107 | period *= atomic_read(&ng->refcount); | |
1108 | period *= shared + 1; | |
1109 | period /= private + shared + 1; | |
1110 | } | |
1111 | ||
1112 | return max(smin, period); | |
1113 | } | |
1114 | ||
598f0ec0 MG |
1115 | static unsigned int task_scan_max(struct task_struct *p) |
1116 | { | |
b5dd77c8 RR |
1117 | unsigned long smin = task_scan_min(p); |
1118 | unsigned long smax; | |
598f0ec0 MG |
1119 | |
1120 | /* Watch for min being lower than max due to floor calculations */ | |
1121 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
b5dd77c8 RR |
1122 | |
1123 | /* Scale the maximum scan period with the amount of shared memory. */ | |
1124 | if (p->numa_group) { | |
1125 | struct numa_group *ng = p->numa_group; | |
1126 | unsigned long shared = group_faults_shared(ng); | |
1127 | unsigned long private = group_faults_priv(ng); | |
1128 | unsigned long period = smax; | |
1129 | ||
1130 | period *= atomic_read(&ng->refcount); | |
1131 | period *= shared + 1; | |
1132 | period /= private + shared + 1; | |
1133 | ||
1134 | smax = max(smax, period); | |
1135 | } | |
1136 | ||
598f0ec0 MG |
1137 | return max(smin, smax); |
1138 | } | |
1139 | ||
13784475 MG |
1140 | void init_numa_balancing(unsigned long clone_flags, struct task_struct *p) |
1141 | { | |
1142 | int mm_users = 0; | |
1143 | struct mm_struct *mm = p->mm; | |
1144 | ||
1145 | if (mm) { | |
1146 | mm_users = atomic_read(&mm->mm_users); | |
1147 | if (mm_users == 1) { | |
1148 | mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
1149 | mm->numa_scan_seq = 0; | |
1150 | } | |
1151 | } | |
1152 | p->node_stamp = 0; | |
1153 | p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; | |
1154 | p->numa_scan_period = sysctl_numa_balancing_scan_delay; | |
1155 | p->numa_work.next = &p->numa_work; | |
1156 | p->numa_faults = NULL; | |
1157 | p->numa_group = NULL; | |
1158 | p->last_task_numa_placement = 0; | |
1159 | p->last_sum_exec_runtime = 0; | |
1160 | ||
1161 | /* New address space, reset the preferred nid */ | |
1162 | if (!(clone_flags & CLONE_VM)) { | |
1163 | p->numa_preferred_nid = -1; | |
1164 | return; | |
1165 | } | |
1166 | ||
1167 | /* | |
1168 | * New thread, keep existing numa_preferred_nid which should be copied | |
1169 | * already by arch_dup_task_struct but stagger when scans start. | |
1170 | */ | |
1171 | if (mm) { | |
1172 | unsigned int delay; | |
1173 | ||
1174 | delay = min_t(unsigned int, task_scan_max(current), | |
1175 | current->numa_scan_period * mm_users * NSEC_PER_MSEC); | |
1176 | delay += 2 * TICK_NSEC; | |
1177 | p->node_stamp = delay; | |
1178 | } | |
1179 | } | |
1180 | ||
0ec8aa00 PZ |
1181 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1182 | { | |
1183 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | |
1184 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | |
1185 | } | |
1186 | ||
1187 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1188 | { | |
1189 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | |
1190 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | |
1191 | } | |
1192 | ||
be1e4e76 RR |
1193 | /* Shared or private faults. */ |
1194 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1195 | ||
1196 | /* Memory and CPU locality */ | |
1197 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1198 | ||
1199 | /* Averaged statistics, and temporary buffers. */ | |
1200 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1201 | ||
e29cf08b MG |
1202 | pid_t task_numa_group_id(struct task_struct *p) |
1203 | { | |
1204 | return p->numa_group ? p->numa_group->gid : 0; | |
1205 | } | |
1206 | ||
44dba3d5 | 1207 | /* |
97fb7a0a | 1208 | * The averaged statistics, shared & private, memory & CPU, |
44dba3d5 IM |
1209 | * occupy the first half of the array. The second half of the |
1210 | * array is for current counters, which are averaged into the | |
1211 | * first set by task_numa_placement. | |
1212 | */ | |
1213 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1214 | { |
44dba3d5 | 1215 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1216 | } |
1217 | ||
1218 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1219 | { | |
44dba3d5 | 1220 | if (!p->numa_faults) |
ac8e895b MG |
1221 | return 0; |
1222 | ||
44dba3d5 IM |
1223 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1224 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1225 | } |
1226 | ||
83e1d2cd MG |
1227 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1228 | { | |
1229 | if (!p->numa_group) | |
1230 | return 0; | |
1231 | ||
44dba3d5 IM |
1232 | return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1233 | p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1234 | } |
1235 | ||
20e07dea RR |
1236 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1237 | { | |
44dba3d5 IM |
1238 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
1239 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | |
20e07dea RR |
1240 | } |
1241 | ||
b5dd77c8 RR |
1242 | static inline unsigned long group_faults_priv(struct numa_group *ng) |
1243 | { | |
1244 | unsigned long faults = 0; | |
1245 | int node; | |
1246 | ||
1247 | for_each_online_node(node) { | |
1248 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
1249 | } | |
1250 | ||
1251 | return faults; | |
1252 | } | |
1253 | ||
1254 | static inline unsigned long group_faults_shared(struct numa_group *ng) | |
1255 | { | |
1256 | unsigned long faults = 0; | |
1257 | int node; | |
1258 | ||
1259 | for_each_online_node(node) { | |
1260 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
1261 | } | |
1262 | ||
1263 | return faults; | |
1264 | } | |
1265 | ||
4142c3eb RR |
1266 | /* |
1267 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1268 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1269 | * between these nodes are slowed down, to allow things to settle down. | |
1270 | */ | |
1271 | #define ACTIVE_NODE_FRACTION 3 | |
1272 | ||
1273 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1274 | { | |
1275 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1276 | } | |
1277 | ||
6c6b1193 RR |
1278 | /* Handle placement on systems where not all nodes are directly connected. */ |
1279 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
1280 | int maxdist, bool task) | |
1281 | { | |
1282 | unsigned long score = 0; | |
1283 | int node; | |
1284 | ||
1285 | /* | |
1286 | * All nodes are directly connected, and the same distance | |
1287 | * from each other. No need for fancy placement algorithms. | |
1288 | */ | |
1289 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1290 | return 0; | |
1291 | ||
1292 | /* | |
1293 | * This code is called for each node, introducing N^2 complexity, | |
1294 | * which should be ok given the number of nodes rarely exceeds 8. | |
1295 | */ | |
1296 | for_each_online_node(node) { | |
1297 | unsigned long faults; | |
1298 | int dist = node_distance(nid, node); | |
1299 | ||
1300 | /* | |
1301 | * The furthest away nodes in the system are not interesting | |
1302 | * for placement; nid was already counted. | |
1303 | */ | |
1304 | if (dist == sched_max_numa_distance || node == nid) | |
1305 | continue; | |
1306 | ||
1307 | /* | |
1308 | * On systems with a backplane NUMA topology, compare groups | |
1309 | * of nodes, and move tasks towards the group with the most | |
1310 | * memory accesses. When comparing two nodes at distance | |
1311 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1312 | * of each group. Skip other nodes. | |
1313 | */ | |
1314 | if (sched_numa_topology_type == NUMA_BACKPLANE && | |
0ee7e74d | 1315 | dist >= maxdist) |
6c6b1193 RR |
1316 | continue; |
1317 | ||
1318 | /* Add up the faults from nearby nodes. */ | |
1319 | if (task) | |
1320 | faults = task_faults(p, node); | |
1321 | else | |
1322 | faults = group_faults(p, node); | |
1323 | ||
1324 | /* | |
1325 | * On systems with a glueless mesh NUMA topology, there are | |
1326 | * no fixed "groups of nodes". Instead, nodes that are not | |
1327 | * directly connected bounce traffic through intermediate | |
1328 | * nodes; a numa_group can occupy any set of nodes. | |
1329 | * The further away a node is, the less the faults count. | |
1330 | * This seems to result in good task placement. | |
1331 | */ | |
1332 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1333 | faults *= (sched_max_numa_distance - dist); | |
1334 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | |
1335 | } | |
1336 | ||
1337 | score += faults; | |
1338 | } | |
1339 | ||
1340 | return score; | |
1341 | } | |
1342 | ||
83e1d2cd MG |
1343 | /* |
1344 | * These return the fraction of accesses done by a particular task, or | |
1345 | * task group, on a particular numa node. The group weight is given a | |
1346 | * larger multiplier, in order to group tasks together that are almost | |
1347 | * evenly spread out between numa nodes. | |
1348 | */ | |
7bd95320 RR |
1349 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1350 | int dist) | |
83e1d2cd | 1351 | { |
7bd95320 | 1352 | unsigned long faults, total_faults; |
83e1d2cd | 1353 | |
44dba3d5 | 1354 | if (!p->numa_faults) |
83e1d2cd MG |
1355 | return 0; |
1356 | ||
1357 | total_faults = p->total_numa_faults; | |
1358 | ||
1359 | if (!total_faults) | |
1360 | return 0; | |
1361 | ||
7bd95320 | 1362 | faults = task_faults(p, nid); |
6c6b1193 RR |
1363 | faults += score_nearby_nodes(p, nid, dist, true); |
1364 | ||
7bd95320 | 1365 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1366 | } |
1367 | ||
7bd95320 RR |
1368 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1369 | int dist) | |
83e1d2cd | 1370 | { |
7bd95320 RR |
1371 | unsigned long faults, total_faults; |
1372 | ||
1373 | if (!p->numa_group) | |
1374 | return 0; | |
1375 | ||
1376 | total_faults = p->numa_group->total_faults; | |
1377 | ||
1378 | if (!total_faults) | |
83e1d2cd MG |
1379 | return 0; |
1380 | ||
7bd95320 | 1381 | faults = group_faults(p, nid); |
6c6b1193 RR |
1382 | faults += score_nearby_nodes(p, nid, dist, false); |
1383 | ||
7bd95320 | 1384 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1385 | } |
1386 | ||
10f39042 RR |
1387 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1388 | int src_nid, int dst_cpu) | |
1389 | { | |
1390 | struct numa_group *ng = p->numa_group; | |
1391 | int dst_nid = cpu_to_node(dst_cpu); | |
1392 | int last_cpupid, this_cpupid; | |
1393 | ||
1394 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
37355bdc MG |
1395 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); |
1396 | ||
1397 | /* | |
1398 | * Allow first faults or private faults to migrate immediately early in | |
1399 | * the lifetime of a task. The magic number 4 is based on waiting for | |
1400 | * two full passes of the "multi-stage node selection" test that is | |
1401 | * executed below. | |
1402 | */ | |
1403 | if ((p->numa_preferred_nid == -1 || p->numa_scan_seq <= 4) && | |
1404 | (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid))) | |
1405 | return true; | |
10f39042 RR |
1406 | |
1407 | /* | |
1408 | * Multi-stage node selection is used in conjunction with a periodic | |
1409 | * migration fault to build a temporal task<->page relation. By using | |
1410 | * a two-stage filter we remove short/unlikely relations. | |
1411 | * | |
1412 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1413 | * a task's usage of a particular page (n_p) per total usage of this | |
1414 | * page (n_t) (in a given time-span) to a probability. | |
1415 | * | |
1416 | * Our periodic faults will sample this probability and getting the | |
1417 | * same result twice in a row, given these samples are fully | |
1418 | * independent, is then given by P(n)^2, provided our sample period | |
1419 | * is sufficiently short compared to the usage pattern. | |
1420 | * | |
1421 | * This quadric squishes small probabilities, making it less likely we | |
1422 | * act on an unlikely task<->page relation. | |
1423 | */ | |
10f39042 RR |
1424 | if (!cpupid_pid_unset(last_cpupid) && |
1425 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1426 | return false; | |
1427 | ||
1428 | /* Always allow migrate on private faults */ | |
1429 | if (cpupid_match_pid(p, last_cpupid)) | |
1430 | return true; | |
1431 | ||
1432 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1433 | if (!ng) | |
1434 | return true; | |
1435 | ||
1436 | /* | |
4142c3eb RR |
1437 | * Destination node is much more heavily used than the source |
1438 | * node? Allow migration. | |
10f39042 | 1439 | */ |
4142c3eb RR |
1440 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1441 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1442 | return true; |
1443 | ||
1444 | /* | |
4142c3eb RR |
1445 | * Distribute memory according to CPU & memory use on each node, |
1446 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1447 | * | |
1448 | * faults_cpu(dst) 3 faults_cpu(src) | |
1449 | * --------------- * - > --------------- | |
1450 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1451 | */ |
4142c3eb RR |
1452 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1453 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1454 | } |
1455 | ||
c7132dd6 | 1456 | static unsigned long weighted_cpuload(struct rq *rq); |
58d081b5 MG |
1457 | static unsigned long source_load(int cpu, int type); |
1458 | static unsigned long target_load(int cpu, int type); | |
58d081b5 | 1459 | |
fb13c7ee | 1460 | /* Cached statistics for all CPUs within a node */ |
58d081b5 MG |
1461 | struct numa_stats { |
1462 | unsigned long load; | |
fb13c7ee MG |
1463 | |
1464 | /* Total compute capacity of CPUs on a node */ | |
5ef20ca1 | 1465 | unsigned long compute_capacity; |
58d081b5 | 1466 | }; |
e6628d5b | 1467 | |
fb13c7ee MG |
1468 | /* |
1469 | * XXX borrowed from update_sg_lb_stats | |
1470 | */ | |
1471 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
1472 | { | |
d90707eb | 1473 | int cpu; |
fb13c7ee MG |
1474 | |
1475 | memset(ns, 0, sizeof(*ns)); | |
1476 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
1477 | struct rq *rq = cpu_rq(cpu); | |
1478 | ||
c7132dd6 | 1479 | ns->load += weighted_cpuload(rq); |
ced549fa | 1480 | ns->compute_capacity += capacity_of(cpu); |
fb13c7ee MG |
1481 | } |
1482 | ||
fb13c7ee MG |
1483 | } |
1484 | ||
58d081b5 MG |
1485 | struct task_numa_env { |
1486 | struct task_struct *p; | |
e6628d5b | 1487 | |
58d081b5 MG |
1488 | int src_cpu, src_nid; |
1489 | int dst_cpu, dst_nid; | |
e6628d5b | 1490 | |
58d081b5 | 1491 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1492 | |
40ea2b42 | 1493 | int imbalance_pct; |
7bd95320 | 1494 | int dist; |
fb13c7ee MG |
1495 | |
1496 | struct task_struct *best_task; | |
1497 | long best_imp; | |
58d081b5 MG |
1498 | int best_cpu; |
1499 | }; | |
1500 | ||
fb13c7ee MG |
1501 | static void task_numa_assign(struct task_numa_env *env, |
1502 | struct task_struct *p, long imp) | |
1503 | { | |
a4739eca SD |
1504 | struct rq *rq = cpu_rq(env->dst_cpu); |
1505 | ||
1506 | /* Bail out if run-queue part of active NUMA balance. */ | |
1507 | if (xchg(&rq->numa_migrate_on, 1)) | |
1508 | return; | |
1509 | ||
1510 | /* | |
1511 | * Clear previous best_cpu/rq numa-migrate flag, since task now | |
1512 | * found a better CPU to move/swap. | |
1513 | */ | |
1514 | if (env->best_cpu != -1) { | |
1515 | rq = cpu_rq(env->best_cpu); | |
1516 | WRITE_ONCE(rq->numa_migrate_on, 0); | |
1517 | } | |
1518 | ||
fb13c7ee MG |
1519 | if (env->best_task) |
1520 | put_task_struct(env->best_task); | |
bac78573 ON |
1521 | if (p) |
1522 | get_task_struct(p); | |
fb13c7ee MG |
1523 | |
1524 | env->best_task = p; | |
1525 | env->best_imp = imp; | |
1526 | env->best_cpu = env->dst_cpu; | |
1527 | } | |
1528 | ||
28a21745 | 1529 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1530 | struct task_numa_env *env) |
1531 | { | |
e4991b24 RR |
1532 | long imb, old_imb; |
1533 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1534 | long src_capacity, dst_capacity; |
1535 | ||
1536 | /* | |
1537 | * The load is corrected for the CPU capacity available on each node. | |
1538 | * | |
1539 | * src_load dst_load | |
1540 | * ------------ vs --------- | |
1541 | * src_capacity dst_capacity | |
1542 | */ | |
1543 | src_capacity = env->src_stats.compute_capacity; | |
1544 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 | 1545 | |
5f95ba7a | 1546 | imb = abs(dst_load * src_capacity - src_load * dst_capacity); |
e63da036 | 1547 | |
28a21745 | 1548 | orig_src_load = env->src_stats.load; |
e4991b24 | 1549 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1550 | |
5f95ba7a | 1551 | old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); |
e4991b24 RR |
1552 | |
1553 | /* Would this change make things worse? */ | |
1554 | return (imb > old_imb); | |
e63da036 RR |
1555 | } |
1556 | ||
6fd98e77 SD |
1557 | /* |
1558 | * Maximum NUMA importance can be 1998 (2*999); | |
1559 | * SMALLIMP @ 30 would be close to 1998/64. | |
1560 | * Used to deter task migration. | |
1561 | */ | |
1562 | #define SMALLIMP 30 | |
1563 | ||
fb13c7ee MG |
1564 | /* |
1565 | * This checks if the overall compute and NUMA accesses of the system would | |
1566 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1567 | * into account that it might be best if task running on the dst_cpu should | |
1568 | * be exchanged with the source task | |
1569 | */ | |
887c290e | 1570 | static void task_numa_compare(struct task_numa_env *env, |
305c1fac | 1571 | long taskimp, long groupimp, bool maymove) |
fb13c7ee | 1572 | { |
fb13c7ee MG |
1573 | struct rq *dst_rq = cpu_rq(env->dst_cpu); |
1574 | struct task_struct *cur; | |
28a21745 | 1575 | long src_load, dst_load; |
fb13c7ee | 1576 | long load; |
1c5d3eb3 | 1577 | long imp = env->p->numa_group ? groupimp : taskimp; |
0132c3e1 | 1578 | long moveimp = imp; |
7bd95320 | 1579 | int dist = env->dist; |
fb13c7ee | 1580 | |
a4739eca SD |
1581 | if (READ_ONCE(dst_rq->numa_migrate_on)) |
1582 | return; | |
1583 | ||
fb13c7ee | 1584 | rcu_read_lock(); |
bac78573 ON |
1585 | cur = task_rcu_dereference(&dst_rq->curr); |
1586 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) | |
fb13c7ee MG |
1587 | cur = NULL; |
1588 | ||
7af68335 PZ |
1589 | /* |
1590 | * Because we have preemption enabled we can get migrated around and | |
1591 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1592 | */ | |
1593 | if (cur == env->p) | |
1594 | goto unlock; | |
1595 | ||
305c1fac | 1596 | if (!cur) { |
6fd98e77 | 1597 | if (maymove && moveimp >= env->best_imp) |
305c1fac SD |
1598 | goto assign; |
1599 | else | |
1600 | goto unlock; | |
1601 | } | |
1602 | ||
fb13c7ee MG |
1603 | /* |
1604 | * "imp" is the fault differential for the source task between the | |
1605 | * source and destination node. Calculate the total differential for | |
1606 | * the source task and potential destination task. The more negative | |
305c1fac | 1607 | * the value is, the more remote accesses that would be expected to |
fb13c7ee MG |
1608 | * be incurred if the tasks were swapped. |
1609 | */ | |
305c1fac SD |
1610 | /* Skip this swap candidate if cannot move to the source cpu */ |
1611 | if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed)) | |
1612 | goto unlock; | |
fb13c7ee | 1613 | |
305c1fac SD |
1614 | /* |
1615 | * If dst and source tasks are in the same NUMA group, or not | |
1616 | * in any group then look only at task weights. | |
1617 | */ | |
1618 | if (cur->numa_group == env->p->numa_group) { | |
1619 | imp = taskimp + task_weight(cur, env->src_nid, dist) - | |
1620 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1621 | /* |
305c1fac SD |
1622 | * Add some hysteresis to prevent swapping the |
1623 | * tasks within a group over tiny differences. | |
887c290e | 1624 | */ |
305c1fac SD |
1625 | if (cur->numa_group) |
1626 | imp -= imp / 16; | |
1627 | } else { | |
1628 | /* | |
1629 | * Compare the group weights. If a task is all by itself | |
1630 | * (not part of a group), use the task weight instead. | |
1631 | */ | |
1632 | if (cur->numa_group && env->p->numa_group) | |
1633 | imp += group_weight(cur, env->src_nid, dist) - | |
1634 | group_weight(cur, env->dst_nid, dist); | |
1635 | else | |
1636 | imp += task_weight(cur, env->src_nid, dist) - | |
1637 | task_weight(cur, env->dst_nid, dist); | |
fb13c7ee MG |
1638 | } |
1639 | ||
305c1fac | 1640 | if (maymove && moveimp > imp && moveimp > env->best_imp) { |
6fd98e77 | 1641 | imp = moveimp; |
305c1fac | 1642 | cur = NULL; |
fb13c7ee | 1643 | goto assign; |
305c1fac | 1644 | } |
fb13c7ee | 1645 | |
6fd98e77 SD |
1646 | /* |
1647 | * If the NUMA importance is less than SMALLIMP, | |
1648 | * task migration might only result in ping pong | |
1649 | * of tasks and also hurt performance due to cache | |
1650 | * misses. | |
1651 | */ | |
1652 | if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) | |
1653 | goto unlock; | |
1654 | ||
fb13c7ee MG |
1655 | /* |
1656 | * In the overloaded case, try and keep the load balanced. | |
1657 | */ | |
305c1fac SD |
1658 | load = task_h_load(env->p) - task_h_load(cur); |
1659 | if (!load) | |
1660 | goto assign; | |
1661 | ||
e720fff6 PZ |
1662 | dst_load = env->dst_stats.load + load; |
1663 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1664 | |
28a21745 | 1665 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1666 | goto unlock; |
1667 | ||
305c1fac | 1668 | assign: |
ba7e5a27 RR |
1669 | /* |
1670 | * One idle CPU per node is evaluated for a task numa move. | |
1671 | * Call select_idle_sibling to maybe find a better one. | |
1672 | */ | |
10e2f1ac PZ |
1673 | if (!cur) { |
1674 | /* | |
97fb7a0a | 1675 | * select_idle_siblings() uses an per-CPU cpumask that |
10e2f1ac PZ |
1676 | * can be used from IRQ context. |
1677 | */ | |
1678 | local_irq_disable(); | |
772bd008 MR |
1679 | env->dst_cpu = select_idle_sibling(env->p, env->src_cpu, |
1680 | env->dst_cpu); | |
10e2f1ac PZ |
1681 | local_irq_enable(); |
1682 | } | |
ba7e5a27 | 1683 | |
fb13c7ee MG |
1684 | task_numa_assign(env, cur, imp); |
1685 | unlock: | |
1686 | rcu_read_unlock(); | |
1687 | } | |
1688 | ||
887c290e RR |
1689 | static void task_numa_find_cpu(struct task_numa_env *env, |
1690 | long taskimp, long groupimp) | |
2c8a50aa | 1691 | { |
305c1fac SD |
1692 | long src_load, dst_load, load; |
1693 | bool maymove = false; | |
2c8a50aa MG |
1694 | int cpu; |
1695 | ||
305c1fac SD |
1696 | load = task_h_load(env->p); |
1697 | dst_load = env->dst_stats.load + load; | |
1698 | src_load = env->src_stats.load - load; | |
1699 | ||
1700 | /* | |
1701 | * If the improvement from just moving env->p direction is better | |
1702 | * than swapping tasks around, check if a move is possible. | |
1703 | */ | |
1704 | maymove = !load_too_imbalanced(src_load, dst_load, env); | |
1705 | ||
2c8a50aa MG |
1706 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { |
1707 | /* Skip this CPU if the source task cannot migrate */ | |
0c98d344 | 1708 | if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed)) |
2c8a50aa MG |
1709 | continue; |
1710 | ||
1711 | env->dst_cpu = cpu; | |
305c1fac | 1712 | task_numa_compare(env, taskimp, groupimp, maymove); |
2c8a50aa MG |
1713 | } |
1714 | } | |
1715 | ||
58d081b5 MG |
1716 | static int task_numa_migrate(struct task_struct *p) |
1717 | { | |
58d081b5 MG |
1718 | struct task_numa_env env = { |
1719 | .p = p, | |
fb13c7ee | 1720 | |
58d081b5 | 1721 | .src_cpu = task_cpu(p), |
b32e86b4 | 1722 | .src_nid = task_node(p), |
fb13c7ee MG |
1723 | |
1724 | .imbalance_pct = 112, | |
1725 | ||
1726 | .best_task = NULL, | |
1727 | .best_imp = 0, | |
4142c3eb | 1728 | .best_cpu = -1, |
58d081b5 MG |
1729 | }; |
1730 | struct sched_domain *sd; | |
a4739eca | 1731 | struct rq *best_rq; |
887c290e | 1732 | unsigned long taskweight, groupweight; |
7bd95320 | 1733 | int nid, ret, dist; |
887c290e | 1734 | long taskimp, groupimp; |
e6628d5b | 1735 | |
58d081b5 | 1736 | /* |
fb13c7ee MG |
1737 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1738 | * imbalance and would be the first to start moving tasks about. | |
1739 | * | |
1740 | * And we want to avoid any moving of tasks about, as that would create | |
1741 | * random movement of tasks -- counter the numa conditions we're trying | |
1742 | * to satisfy here. | |
58d081b5 MG |
1743 | */ |
1744 | rcu_read_lock(); | |
fb13c7ee | 1745 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1746 | if (sd) |
1747 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1748 | rcu_read_unlock(); |
1749 | ||
46a73e8a RR |
1750 | /* |
1751 | * Cpusets can break the scheduler domain tree into smaller | |
1752 | * balance domains, some of which do not cross NUMA boundaries. | |
1753 | * Tasks that are "trapped" in such domains cannot be migrated | |
1754 | * elsewhere, so there is no point in (re)trying. | |
1755 | */ | |
1756 | if (unlikely(!sd)) { | |
8cd45eee | 1757 | sched_setnuma(p, task_node(p)); |
46a73e8a RR |
1758 | return -EINVAL; |
1759 | } | |
1760 | ||
2c8a50aa | 1761 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
1762 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
1763 | taskweight = task_weight(p, env.src_nid, dist); | |
1764 | groupweight = group_weight(p, env.src_nid, dist); | |
1765 | update_numa_stats(&env.src_stats, env.src_nid); | |
1766 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; | |
1767 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
2c8a50aa | 1768 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1769 | |
a43455a1 | 1770 | /* Try to find a spot on the preferred nid. */ |
2d4056fa | 1771 | task_numa_find_cpu(&env, taskimp, groupimp); |
e1dda8a7 | 1772 | |
9de05d48 RR |
1773 | /* |
1774 | * Look at other nodes in these cases: | |
1775 | * - there is no space available on the preferred_nid | |
1776 | * - the task is part of a numa_group that is interleaved across | |
1777 | * multiple NUMA nodes; in order to better consolidate the group, | |
1778 | * we need to check other locations. | |
1779 | */ | |
4142c3eb | 1780 | if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) { |
2c8a50aa MG |
1781 | for_each_online_node(nid) { |
1782 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1783 | continue; | |
58d081b5 | 1784 | |
7bd95320 | 1785 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
1786 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
1787 | dist != env.dist) { | |
1788 | taskweight = task_weight(p, env.src_nid, dist); | |
1789 | groupweight = group_weight(p, env.src_nid, dist); | |
1790 | } | |
7bd95320 | 1791 | |
83e1d2cd | 1792 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
1793 | taskimp = task_weight(p, nid, dist) - taskweight; |
1794 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 1795 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
1796 | continue; |
1797 | ||
7bd95320 | 1798 | env.dist = dist; |
2c8a50aa MG |
1799 | env.dst_nid = nid; |
1800 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
2d4056fa | 1801 | task_numa_find_cpu(&env, taskimp, groupimp); |
58d081b5 MG |
1802 | } |
1803 | } | |
1804 | ||
68d1b02a RR |
1805 | /* |
1806 | * If the task is part of a workload that spans multiple NUMA nodes, | |
1807 | * and is migrating into one of the workload's active nodes, remember | |
1808 | * this node as the task's preferred numa node, so the workload can | |
1809 | * settle down. | |
1810 | * A task that migrated to a second choice node will be better off | |
1811 | * trying for a better one later. Do not set the preferred node here. | |
1812 | */ | |
db015dae RR |
1813 | if (p->numa_group) { |
1814 | if (env.best_cpu == -1) | |
1815 | nid = env.src_nid; | |
1816 | else | |
8cd45eee | 1817 | nid = cpu_to_node(env.best_cpu); |
db015dae | 1818 | |
8cd45eee SD |
1819 | if (nid != p->numa_preferred_nid) |
1820 | sched_setnuma(p, nid); | |
db015dae RR |
1821 | } |
1822 | ||
1823 | /* No better CPU than the current one was found. */ | |
1824 | if (env.best_cpu == -1) | |
1825 | return -EAGAIN; | |
0ec8aa00 | 1826 | |
a4739eca | 1827 | best_rq = cpu_rq(env.best_cpu); |
fb13c7ee | 1828 | if (env.best_task == NULL) { |
286549dc | 1829 | ret = migrate_task_to(p, env.best_cpu); |
a4739eca | 1830 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
286549dc MG |
1831 | if (ret != 0) |
1832 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); | |
fb13c7ee MG |
1833 | return ret; |
1834 | } | |
1835 | ||
0ad4e3df | 1836 | ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu); |
a4739eca | 1837 | WRITE_ONCE(best_rq->numa_migrate_on, 0); |
0ad4e3df | 1838 | |
286549dc MG |
1839 | if (ret != 0) |
1840 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); | |
fb13c7ee MG |
1841 | put_task_struct(env.best_task); |
1842 | return ret; | |
e6628d5b MG |
1843 | } |
1844 | ||
6b9a7460 MG |
1845 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1846 | static void numa_migrate_preferred(struct task_struct *p) | |
1847 | { | |
5085e2a3 RR |
1848 | unsigned long interval = HZ; |
1849 | ||
2739d3ee | 1850 | /* This task has no NUMA fault statistics yet */ |
44dba3d5 | 1851 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults)) |
6b9a7460 MG |
1852 | return; |
1853 | ||
2739d3ee | 1854 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 | 1855 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
789ba280 | 1856 | p->numa_migrate_retry = jiffies + interval; |
2739d3ee RR |
1857 | |
1858 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 1859 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
1860 | return; |
1861 | ||
1862 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 1863 | task_numa_migrate(p); |
6b9a7460 MG |
1864 | } |
1865 | ||
20e07dea | 1866 | /* |
4142c3eb | 1867 | * Find out how many nodes on the workload is actively running on. Do this by |
20e07dea RR |
1868 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
1869 | * be different from the set of nodes where the workload's memory is currently | |
1870 | * located. | |
20e07dea | 1871 | */ |
4142c3eb | 1872 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
1873 | { |
1874 | unsigned long faults, max_faults = 0; | |
4142c3eb | 1875 | int nid, active_nodes = 0; |
20e07dea RR |
1876 | |
1877 | for_each_online_node(nid) { | |
1878 | faults = group_faults_cpu(numa_group, nid); | |
1879 | if (faults > max_faults) | |
1880 | max_faults = faults; | |
1881 | } | |
1882 | ||
1883 | for_each_online_node(nid) { | |
1884 | faults = group_faults_cpu(numa_group, nid); | |
4142c3eb RR |
1885 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
1886 | active_nodes++; | |
20e07dea | 1887 | } |
4142c3eb RR |
1888 | |
1889 | numa_group->max_faults_cpu = max_faults; | |
1890 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
1891 | } |
1892 | ||
04bb2f94 RR |
1893 | /* |
1894 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1895 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
1896 | * period will be for the next scan window. If local/(local+remote) ratio is |
1897 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
1898 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
1899 | */ |
1900 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 1901 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
1902 | |
1903 | /* | |
1904 | * Increase the scan period (slow down scanning) if the majority of | |
1905 | * our memory is already on our local node, or if the majority of | |
1906 | * the page accesses are shared with other processes. | |
1907 | * Otherwise, decrease the scan period. | |
1908 | */ | |
1909 | static void update_task_scan_period(struct task_struct *p, | |
1910 | unsigned long shared, unsigned long private) | |
1911 | { | |
1912 | unsigned int period_slot; | |
37ec97de | 1913 | int lr_ratio, ps_ratio; |
04bb2f94 RR |
1914 | int diff; |
1915 | ||
1916 | unsigned long remote = p->numa_faults_locality[0]; | |
1917 | unsigned long local = p->numa_faults_locality[1]; | |
1918 | ||
1919 | /* | |
1920 | * If there were no record hinting faults then either the task is | |
1921 | * completely idle or all activity is areas that are not of interest | |
074c2381 MG |
1922 | * to automatic numa balancing. Related to that, if there were failed |
1923 | * migration then it implies we are migrating too quickly or the local | |
1924 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 1925 | */ |
074c2381 | 1926 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
1927 | p->numa_scan_period = min(p->numa_scan_period_max, |
1928 | p->numa_scan_period << 1); | |
1929 | ||
1930 | p->mm->numa_next_scan = jiffies + | |
1931 | msecs_to_jiffies(p->numa_scan_period); | |
1932 | ||
1933 | return; | |
1934 | } | |
1935 | ||
1936 | /* | |
1937 | * Prepare to scale scan period relative to the current period. | |
1938 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1939 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1940 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1941 | */ | |
1942 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
37ec97de RR |
1943 | lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); |
1944 | ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); | |
1945 | ||
1946 | if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { | |
1947 | /* | |
1948 | * Most memory accesses are local. There is no need to | |
1949 | * do fast NUMA scanning, since memory is already local. | |
1950 | */ | |
1951 | int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; | |
1952 | if (!slot) | |
1953 | slot = 1; | |
1954 | diff = slot * period_slot; | |
1955 | } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { | |
1956 | /* | |
1957 | * Most memory accesses are shared with other tasks. | |
1958 | * There is no point in continuing fast NUMA scanning, | |
1959 | * since other tasks may just move the memory elsewhere. | |
1960 | */ | |
1961 | int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; | |
04bb2f94 RR |
1962 | if (!slot) |
1963 | slot = 1; | |
1964 | diff = slot * period_slot; | |
1965 | } else { | |
04bb2f94 | 1966 | /* |
37ec97de RR |
1967 | * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, |
1968 | * yet they are not on the local NUMA node. Speed up | |
1969 | * NUMA scanning to get the memory moved over. | |
04bb2f94 | 1970 | */ |
37ec97de RR |
1971 | int ratio = max(lr_ratio, ps_ratio); |
1972 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
04bb2f94 RR |
1973 | } |
1974 | ||
1975 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
1976 | task_scan_min(p), task_scan_max(p)); | |
1977 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
1978 | } | |
1979 | ||
7e2703e6 RR |
1980 | /* |
1981 | * Get the fraction of time the task has been running since the last | |
1982 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
1983 | * decays those on a 32ms period, which is orders of magnitude off | |
1984 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
1985 | * stats only if the task is so new there are no NUMA statistics yet. | |
1986 | */ | |
1987 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
1988 | { | |
1989 | u64 runtime, delta, now; | |
1990 | /* Use the start of this time slice to avoid calculations. */ | |
1991 | now = p->se.exec_start; | |
1992 | runtime = p->se.sum_exec_runtime; | |
1993 | ||
1994 | if (p->last_task_numa_placement) { | |
1995 | delta = runtime - p->last_sum_exec_runtime; | |
1996 | *period = now - p->last_task_numa_placement; | |
1997 | } else { | |
c7b50216 | 1998 | delta = p->se.avg.load_sum; |
9d89c257 | 1999 | *period = LOAD_AVG_MAX; |
7e2703e6 RR |
2000 | } |
2001 | ||
2002 | p->last_sum_exec_runtime = runtime; | |
2003 | p->last_task_numa_placement = now; | |
2004 | ||
2005 | return delta; | |
2006 | } | |
2007 | ||
54009416 RR |
2008 | /* |
2009 | * Determine the preferred nid for a task in a numa_group. This needs to | |
2010 | * be done in a way that produces consistent results with group_weight, | |
2011 | * otherwise workloads might not converge. | |
2012 | */ | |
2013 | static int preferred_group_nid(struct task_struct *p, int nid) | |
2014 | { | |
2015 | nodemask_t nodes; | |
2016 | int dist; | |
2017 | ||
2018 | /* Direct connections between all NUMA nodes. */ | |
2019 | if (sched_numa_topology_type == NUMA_DIRECT) | |
2020 | return nid; | |
2021 | ||
2022 | /* | |
2023 | * On a system with glueless mesh NUMA topology, group_weight | |
2024 | * scores nodes according to the number of NUMA hinting faults on | |
2025 | * both the node itself, and on nearby nodes. | |
2026 | */ | |
2027 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
2028 | unsigned long score, max_score = 0; | |
2029 | int node, max_node = nid; | |
2030 | ||
2031 | dist = sched_max_numa_distance; | |
2032 | ||
2033 | for_each_online_node(node) { | |
2034 | score = group_weight(p, node, dist); | |
2035 | if (score > max_score) { | |
2036 | max_score = score; | |
2037 | max_node = node; | |
2038 | } | |
2039 | } | |
2040 | return max_node; | |
2041 | } | |
2042 | ||
2043 | /* | |
2044 | * Finding the preferred nid in a system with NUMA backplane | |
2045 | * interconnect topology is more involved. The goal is to locate | |
2046 | * tasks from numa_groups near each other in the system, and | |
2047 | * untangle workloads from different sides of the system. This requires | |
2048 | * searching down the hierarchy of node groups, recursively searching | |
2049 | * inside the highest scoring group of nodes. The nodemask tricks | |
2050 | * keep the complexity of the search down. | |
2051 | */ | |
2052 | nodes = node_online_map; | |
2053 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | |
2054 | unsigned long max_faults = 0; | |
81907478 | 2055 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
2056 | int a, b; |
2057 | ||
2058 | /* Are there nodes at this distance from each other? */ | |
2059 | if (!find_numa_distance(dist)) | |
2060 | continue; | |
2061 | ||
2062 | for_each_node_mask(a, nodes) { | |
2063 | unsigned long faults = 0; | |
2064 | nodemask_t this_group; | |
2065 | nodes_clear(this_group); | |
2066 | ||
2067 | /* Sum group's NUMA faults; includes a==b case. */ | |
2068 | for_each_node_mask(b, nodes) { | |
2069 | if (node_distance(a, b) < dist) { | |
2070 | faults += group_faults(p, b); | |
2071 | node_set(b, this_group); | |
2072 | node_clear(b, nodes); | |
2073 | } | |
2074 | } | |
2075 | ||
2076 | /* Remember the top group. */ | |
2077 | if (faults > max_faults) { | |
2078 | max_faults = faults; | |
2079 | max_group = this_group; | |
2080 | /* | |
2081 | * subtle: at the smallest distance there is | |
2082 | * just one node left in each "group", the | |
2083 | * winner is the preferred nid. | |
2084 | */ | |
2085 | nid = a; | |
2086 | } | |
2087 | } | |
2088 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2089 | if (!max_faults) |
2090 | break; | |
54009416 RR |
2091 | nodes = max_group; |
2092 | } | |
2093 | return nid; | |
2094 | } | |
2095 | ||
cbee9f88 PZ |
2096 | static void task_numa_placement(struct task_struct *p) |
2097 | { | |
f03bb676 SD |
2098 | int seq, nid, max_nid = -1; |
2099 | unsigned long max_faults = 0; | |
04bb2f94 | 2100 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2101 | unsigned long total_faults; |
2102 | u64 runtime, period; | |
7dbd13ed | 2103 | spinlock_t *group_lock = NULL; |
cbee9f88 | 2104 | |
7e5a2c17 JL |
2105 | /* |
2106 | * The p->mm->numa_scan_seq field gets updated without | |
2107 | * exclusive access. Use READ_ONCE() here to ensure | |
2108 | * that the field is read in a single access: | |
2109 | */ | |
316c1608 | 2110 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2111 | if (p->numa_scan_seq == seq) |
2112 | return; | |
2113 | p->numa_scan_seq = seq; | |
598f0ec0 | 2114 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2115 | |
7e2703e6 RR |
2116 | total_faults = p->numa_faults_locality[0] + |
2117 | p->numa_faults_locality[1]; | |
2118 | runtime = numa_get_avg_runtime(p, &period); | |
2119 | ||
7dbd13ed MG |
2120 | /* If the task is part of a group prevent parallel updates to group stats */ |
2121 | if (p->numa_group) { | |
2122 | group_lock = &p->numa_group->lock; | |
60e69eed | 2123 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2124 | } |
2125 | ||
688b7585 MG |
2126 | /* Find the node with the highest number of faults */ |
2127 | for_each_online_node(nid) { | |
44dba3d5 IM |
2128 | /* Keep track of the offsets in numa_faults array */ |
2129 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2130 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2131 | int priv; |
745d6147 | 2132 | |
be1e4e76 | 2133 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2134 | long diff, f_diff, f_weight; |
8c8a743c | 2135 | |
44dba3d5 IM |
2136 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2137 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2138 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2139 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2140 | |
ac8e895b | 2141 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2142 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2143 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2144 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2145 | |
7e2703e6 RR |
2146 | /* |
2147 | * Normalize the faults_from, so all tasks in a group | |
2148 | * count according to CPU use, instead of by the raw | |
2149 | * number of faults. Tasks with little runtime have | |
2150 | * little over-all impact on throughput, and thus their | |
2151 | * faults are less important. | |
2152 | */ | |
2153 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2154 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2155 | (total_faults + 1); |
44dba3d5 IM |
2156 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2157 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2158 | |
44dba3d5 IM |
2159 | p->numa_faults[mem_idx] += diff; |
2160 | p->numa_faults[cpu_idx] += f_diff; | |
2161 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2162 | p->total_numa_faults += diff; |
8c8a743c | 2163 | if (p->numa_group) { |
44dba3d5 IM |
2164 | /* |
2165 | * safe because we can only change our own group | |
2166 | * | |
2167 | * mem_idx represents the offset for a given | |
2168 | * nid and priv in a specific region because it | |
2169 | * is at the beginning of the numa_faults array. | |
2170 | */ | |
2171 | p->numa_group->faults[mem_idx] += diff; | |
2172 | p->numa_group->faults_cpu[mem_idx] += f_diff; | |
989348b5 | 2173 | p->numa_group->total_faults += diff; |
44dba3d5 | 2174 | group_faults += p->numa_group->faults[mem_idx]; |
8c8a743c | 2175 | } |
ac8e895b MG |
2176 | } |
2177 | ||
f03bb676 SD |
2178 | if (!p->numa_group) { |
2179 | if (faults > max_faults) { | |
2180 | max_faults = faults; | |
2181 | max_nid = nid; | |
2182 | } | |
2183 | } else if (group_faults > max_faults) { | |
2184 | max_faults = group_faults; | |
688b7585 MG |
2185 | max_nid = nid; |
2186 | } | |
83e1d2cd MG |
2187 | } |
2188 | ||
7dbd13ed | 2189 | if (p->numa_group) { |
4142c3eb | 2190 | numa_group_count_active_nodes(p->numa_group); |
60e69eed | 2191 | spin_unlock_irq(group_lock); |
f03bb676 | 2192 | max_nid = preferred_group_nid(p, max_nid); |
688b7585 MG |
2193 | } |
2194 | ||
bb97fc31 RR |
2195 | if (max_faults) { |
2196 | /* Set the new preferred node */ | |
2197 | if (max_nid != p->numa_preferred_nid) | |
2198 | sched_setnuma(p, max_nid); | |
3a7053b3 | 2199 | } |
30619c89 SD |
2200 | |
2201 | update_task_scan_period(p, fault_types[0], fault_types[1]); | |
cbee9f88 PZ |
2202 | } |
2203 | ||
8c8a743c PZ |
2204 | static inline int get_numa_group(struct numa_group *grp) |
2205 | { | |
2206 | return atomic_inc_not_zero(&grp->refcount); | |
2207 | } | |
2208 | ||
2209 | static inline void put_numa_group(struct numa_group *grp) | |
2210 | { | |
2211 | if (atomic_dec_and_test(&grp->refcount)) | |
2212 | kfree_rcu(grp, rcu); | |
2213 | } | |
2214 | ||
3e6a9418 MG |
2215 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2216 | int *priv) | |
8c8a743c PZ |
2217 | { |
2218 | struct numa_group *grp, *my_grp; | |
2219 | struct task_struct *tsk; | |
2220 | bool join = false; | |
2221 | int cpu = cpupid_to_cpu(cpupid); | |
2222 | int i; | |
2223 | ||
2224 | if (unlikely(!p->numa_group)) { | |
2225 | unsigned int size = sizeof(struct numa_group) + | |
50ec8a40 | 2226 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
2227 | |
2228 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2229 | if (!grp) | |
2230 | return; | |
2231 | ||
2232 | atomic_set(&grp->refcount, 1); | |
4142c3eb RR |
2233 | grp->active_nodes = 1; |
2234 | grp->max_faults_cpu = 0; | |
8c8a743c | 2235 | spin_lock_init(&grp->lock); |
e29cf08b | 2236 | grp->gid = p->pid; |
50ec8a40 | 2237 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
2238 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
2239 | nr_node_ids; | |
8c8a743c | 2240 | |
be1e4e76 | 2241 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2242 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2243 | |
989348b5 | 2244 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2245 | |
8c8a743c PZ |
2246 | grp->nr_tasks++; |
2247 | rcu_assign_pointer(p->numa_group, grp); | |
2248 | } | |
2249 | ||
2250 | rcu_read_lock(); | |
316c1608 | 2251 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2252 | |
2253 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2254 | goto no_join; |
8c8a743c PZ |
2255 | |
2256 | grp = rcu_dereference(tsk->numa_group); | |
2257 | if (!grp) | |
3354781a | 2258 | goto no_join; |
8c8a743c PZ |
2259 | |
2260 | my_grp = p->numa_group; | |
2261 | if (grp == my_grp) | |
3354781a | 2262 | goto no_join; |
8c8a743c PZ |
2263 | |
2264 | /* | |
2265 | * Only join the other group if its bigger; if we're the bigger group, | |
2266 | * the other task will join us. | |
2267 | */ | |
2268 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2269 | goto no_join; |
8c8a743c PZ |
2270 | |
2271 | /* | |
2272 | * Tie-break on the grp address. | |
2273 | */ | |
2274 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2275 | goto no_join; |
8c8a743c | 2276 | |
dabe1d99 RR |
2277 | /* Always join threads in the same process. */ |
2278 | if (tsk->mm == current->mm) | |
2279 | join = true; | |
2280 | ||
2281 | /* Simple filter to avoid false positives due to PID collisions */ | |
2282 | if (flags & TNF_SHARED) | |
2283 | join = true; | |
8c8a743c | 2284 | |
3e6a9418 MG |
2285 | /* Update priv based on whether false sharing was detected */ |
2286 | *priv = !join; | |
2287 | ||
dabe1d99 | 2288 | if (join && !get_numa_group(grp)) |
3354781a | 2289 | goto no_join; |
8c8a743c | 2290 | |
8c8a743c PZ |
2291 | rcu_read_unlock(); |
2292 | ||
2293 | if (!join) | |
2294 | return; | |
2295 | ||
60e69eed MG |
2296 | BUG_ON(irqs_disabled()); |
2297 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 2298 | |
be1e4e76 | 2299 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2300 | my_grp->faults[i] -= p->numa_faults[i]; |
2301 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2302 | } |
989348b5 MG |
2303 | my_grp->total_faults -= p->total_numa_faults; |
2304 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2305 | |
8c8a743c PZ |
2306 | my_grp->nr_tasks--; |
2307 | grp->nr_tasks++; | |
2308 | ||
2309 | spin_unlock(&my_grp->lock); | |
60e69eed | 2310 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2311 | |
2312 | rcu_assign_pointer(p->numa_group, grp); | |
2313 | ||
2314 | put_numa_group(my_grp); | |
3354781a PZ |
2315 | return; |
2316 | ||
2317 | no_join: | |
2318 | rcu_read_unlock(); | |
2319 | return; | |
8c8a743c PZ |
2320 | } |
2321 | ||
2322 | void task_numa_free(struct task_struct *p) | |
2323 | { | |
2324 | struct numa_group *grp = p->numa_group; | |
44dba3d5 | 2325 | void *numa_faults = p->numa_faults; |
e9dd685c SR |
2326 | unsigned long flags; |
2327 | int i; | |
8c8a743c PZ |
2328 | |
2329 | if (grp) { | |
e9dd685c | 2330 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2331 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2332 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2333 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2334 | |
8c8a743c | 2335 | grp->nr_tasks--; |
e9dd685c | 2336 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2337 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2338 | put_numa_group(grp); |
2339 | } | |
2340 | ||
44dba3d5 | 2341 | p->numa_faults = NULL; |
82727018 | 2342 | kfree(numa_faults); |
8c8a743c PZ |
2343 | } |
2344 | ||
cbee9f88 PZ |
2345 | /* |
2346 | * Got a PROT_NONE fault for a page on @node. | |
2347 | */ | |
58b46da3 | 2348 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2349 | { |
2350 | struct task_struct *p = current; | |
6688cc05 | 2351 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2352 | int cpu_node = task_node(current); |
792568ec | 2353 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 2354 | struct numa_group *ng; |
ac8e895b | 2355 | int priv; |
cbee9f88 | 2356 | |
2a595721 | 2357 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2358 | return; |
2359 | ||
9ff1d9ff MG |
2360 | /* for example, ksmd faulting in a user's mm */ |
2361 | if (!p->mm) | |
2362 | return; | |
2363 | ||
f809ca9a | 2364 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2365 | if (unlikely(!p->numa_faults)) { |
2366 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2367 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2368 | |
44dba3d5 IM |
2369 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2370 | if (!p->numa_faults) | |
f809ca9a | 2371 | return; |
745d6147 | 2372 | |
83e1d2cd | 2373 | p->total_numa_faults = 0; |
04bb2f94 | 2374 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2375 | } |
cbee9f88 | 2376 | |
8c8a743c PZ |
2377 | /* |
2378 | * First accesses are treated as private, otherwise consider accesses | |
2379 | * to be private if the accessing pid has not changed | |
2380 | */ | |
2381 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2382 | priv = 1; | |
2383 | } else { | |
2384 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2385 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2386 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2387 | } |
2388 | ||
792568ec RR |
2389 | /* |
2390 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2391 | * occurs wholly within the set of nodes that the workload is | |
2392 | * actively using should be counted as local. This allows the | |
2393 | * scan rate to slow down when a workload has settled down. | |
2394 | */ | |
4142c3eb RR |
2395 | ng = p->numa_group; |
2396 | if (!priv && !local && ng && ng->active_nodes > 1 && | |
2397 | numa_is_active_node(cpu_node, ng) && | |
2398 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
2399 | local = 1; |
2400 | ||
2739d3ee | 2401 | /* |
e1ff516a YW |
2402 | * Retry to migrate task to preferred node periodically, in case it |
2403 | * previously failed, or the scheduler moved us. | |
2739d3ee | 2404 | */ |
b6a60cf3 SD |
2405 | if (time_after(jiffies, p->numa_migrate_retry)) { |
2406 | task_numa_placement(p); | |
6b9a7460 | 2407 | numa_migrate_preferred(p); |
b6a60cf3 | 2408 | } |
6b9a7460 | 2409 | |
b32e86b4 IM |
2410 | if (migrated) |
2411 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2412 | if (flags & TNF_MIGRATE_FAIL) |
2413 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2414 | |
44dba3d5 IM |
2415 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2416 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2417 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2418 | } |
2419 | ||
6e5fb223 PZ |
2420 | static void reset_ptenuma_scan(struct task_struct *p) |
2421 | { | |
7e5a2c17 JL |
2422 | /* |
2423 | * We only did a read acquisition of the mmap sem, so | |
2424 | * p->mm->numa_scan_seq is written to without exclusive access | |
2425 | * and the update is not guaranteed to be atomic. That's not | |
2426 | * much of an issue though, since this is just used for | |
2427 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2428 | * expensive, to avoid any form of compiler optimizations: | |
2429 | */ | |
316c1608 | 2430 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2431 | p->mm->numa_scan_offset = 0; |
2432 | } | |
2433 | ||
cbee9f88 PZ |
2434 | /* |
2435 | * The expensive part of numa migration is done from task_work context. | |
2436 | * Triggered from task_tick_numa(). | |
2437 | */ | |
2438 | void task_numa_work(struct callback_head *work) | |
2439 | { | |
2440 | unsigned long migrate, next_scan, now = jiffies; | |
2441 | struct task_struct *p = current; | |
2442 | struct mm_struct *mm = p->mm; | |
51170840 | 2443 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 2444 | struct vm_area_struct *vma; |
9f40604c | 2445 | unsigned long start, end; |
598f0ec0 | 2446 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2447 | long pages, virtpages; |
cbee9f88 | 2448 | |
9148a3a1 | 2449 | SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); |
cbee9f88 PZ |
2450 | |
2451 | work->next = work; /* protect against double add */ | |
2452 | /* | |
2453 | * Who cares about NUMA placement when they're dying. | |
2454 | * | |
2455 | * NOTE: make sure not to dereference p->mm before this check, | |
2456 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2457 | * without p->mm even though we still had it when we enqueued this | |
2458 | * work. | |
2459 | */ | |
2460 | if (p->flags & PF_EXITING) | |
2461 | return; | |
2462 | ||
930aa174 | 2463 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2464 | mm->numa_next_scan = now + |
2465 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2466 | } |
2467 | ||
cbee9f88 PZ |
2468 | /* |
2469 | * Enforce maximal scan/migration frequency.. | |
2470 | */ | |
2471 | migrate = mm->numa_next_scan; | |
2472 | if (time_before(now, migrate)) | |
2473 | return; | |
2474 | ||
598f0ec0 MG |
2475 | if (p->numa_scan_period == 0) { |
2476 | p->numa_scan_period_max = task_scan_max(p); | |
b5dd77c8 | 2477 | p->numa_scan_period = task_scan_start(p); |
598f0ec0 | 2478 | } |
cbee9f88 | 2479 | |
fb003b80 | 2480 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2481 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2482 | return; | |
2483 | ||
19a78d11 PZ |
2484 | /* |
2485 | * Delay this task enough that another task of this mm will likely win | |
2486 | * the next time around. | |
2487 | */ | |
2488 | p->node_stamp += 2 * TICK_NSEC; | |
2489 | ||
9f40604c MG |
2490 | start = mm->numa_scan_offset; |
2491 | pages = sysctl_numa_balancing_scan_size; | |
2492 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2493 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2494 | if (!pages) |
2495 | return; | |
cbee9f88 | 2496 | |
4620f8c1 | 2497 | |
8655d549 VB |
2498 | if (!down_read_trylock(&mm->mmap_sem)) |
2499 | return; | |
9f40604c | 2500 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2501 | if (!vma) { |
2502 | reset_ptenuma_scan(p); | |
9f40604c | 2503 | start = 0; |
6e5fb223 PZ |
2504 | vma = mm->mmap; |
2505 | } | |
9f40604c | 2506 | for (; vma; vma = vma->vm_next) { |
6b79c57b | 2507 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 2508 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 2509 | continue; |
6b79c57b | 2510 | } |
6e5fb223 | 2511 | |
4591ce4f MG |
2512 | /* |
2513 | * Shared library pages mapped by multiple processes are not | |
2514 | * migrated as it is expected they are cache replicated. Avoid | |
2515 | * hinting faults in read-only file-backed mappings or the vdso | |
2516 | * as migrating the pages will be of marginal benefit. | |
2517 | */ | |
2518 | if (!vma->vm_mm || | |
2519 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2520 | continue; | |
2521 | ||
3c67f474 MG |
2522 | /* |
2523 | * Skip inaccessible VMAs to avoid any confusion between | |
2524 | * PROT_NONE and NUMA hinting ptes | |
2525 | */ | |
2526 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | |
2527 | continue; | |
4591ce4f | 2528 | |
9f40604c MG |
2529 | do { |
2530 | start = max(start, vma->vm_start); | |
2531 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2532 | end = min(end, vma->vm_end); | |
4620f8c1 | 2533 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
2534 | |
2535 | /* | |
4620f8c1 RR |
2536 | * Try to scan sysctl_numa_balancing_size worth of |
2537 | * hpages that have at least one present PTE that | |
2538 | * is not already pte-numa. If the VMA contains | |
2539 | * areas that are unused or already full of prot_numa | |
2540 | * PTEs, scan up to virtpages, to skip through those | |
2541 | * areas faster. | |
598f0ec0 MG |
2542 | */ |
2543 | if (nr_pte_updates) | |
2544 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 2545 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 2546 | |
9f40604c | 2547 | start = end; |
4620f8c1 | 2548 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 2549 | goto out; |
3cf1962c RR |
2550 | |
2551 | cond_resched(); | |
9f40604c | 2552 | } while (end != vma->vm_end); |
cbee9f88 | 2553 | } |
6e5fb223 | 2554 | |
9f40604c | 2555 | out: |
6e5fb223 | 2556 | /* |
c69307d5 PZ |
2557 | * It is possible to reach the end of the VMA list but the last few |
2558 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2559 | * would find the !migratable VMA on the next scan but not reset the | |
2560 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2561 | */ |
2562 | if (vma) | |
9f40604c | 2563 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2564 | else |
2565 | reset_ptenuma_scan(p); | |
2566 | up_read(&mm->mmap_sem); | |
51170840 RR |
2567 | |
2568 | /* | |
2569 | * Make sure tasks use at least 32x as much time to run other code | |
2570 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
2571 | * Usually update_task_scan_period slows down scanning enough; on an | |
2572 | * overloaded system we need to limit overhead on a per task basis. | |
2573 | */ | |
2574 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
2575 | u64 diff = p->se.sum_exec_runtime - runtime; | |
2576 | p->node_stamp += 32 * diff; | |
2577 | } | |
cbee9f88 PZ |
2578 | } |
2579 | ||
2580 | /* | |
2581 | * Drive the periodic memory faults.. | |
2582 | */ | |
2583 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2584 | { | |
2585 | struct callback_head *work = &curr->numa_work; | |
2586 | u64 period, now; | |
2587 | ||
2588 | /* | |
2589 | * We don't care about NUMA placement if we don't have memory. | |
2590 | */ | |
2591 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
2592 | return; | |
2593 | ||
2594 | /* | |
2595 | * Using runtime rather than walltime has the dual advantage that | |
2596 | * we (mostly) drive the selection from busy threads and that the | |
2597 | * task needs to have done some actual work before we bother with | |
2598 | * NUMA placement. | |
2599 | */ | |
2600 | now = curr->se.sum_exec_runtime; | |
2601 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2602 | ||
25b3e5a3 | 2603 | if (now > curr->node_stamp + period) { |
4b96a29b | 2604 | if (!curr->node_stamp) |
b5dd77c8 | 2605 | curr->numa_scan_period = task_scan_start(curr); |
19a78d11 | 2606 | curr->node_stamp += period; |
cbee9f88 PZ |
2607 | |
2608 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
2609 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
2610 | task_work_add(curr, work, true); | |
2611 | } | |
2612 | } | |
2613 | } | |
3fed382b | 2614 | |
3f9672ba SD |
2615 | static void update_scan_period(struct task_struct *p, int new_cpu) |
2616 | { | |
2617 | int src_nid = cpu_to_node(task_cpu(p)); | |
2618 | int dst_nid = cpu_to_node(new_cpu); | |
2619 | ||
05cbdf4f MG |
2620 | if (!static_branch_likely(&sched_numa_balancing)) |
2621 | return; | |
2622 | ||
3f9672ba SD |
2623 | if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) |
2624 | return; | |
2625 | ||
05cbdf4f MG |
2626 | if (src_nid == dst_nid) |
2627 | return; | |
2628 | ||
2629 | /* | |
2630 | * Allow resets if faults have been trapped before one scan | |
2631 | * has completed. This is most likely due to a new task that | |
2632 | * is pulled cross-node due to wakeups or load balancing. | |
2633 | */ | |
2634 | if (p->numa_scan_seq) { | |
2635 | /* | |
2636 | * Avoid scan adjustments if moving to the preferred | |
2637 | * node or if the task was not previously running on | |
2638 | * the preferred node. | |
2639 | */ | |
2640 | if (dst_nid == p->numa_preferred_nid || | |
2641 | (p->numa_preferred_nid != -1 && src_nid != p->numa_preferred_nid)) | |
2642 | return; | |
2643 | } | |
2644 | ||
2645 | p->numa_scan_period = task_scan_start(p); | |
3f9672ba SD |
2646 | } |
2647 | ||
cbee9f88 PZ |
2648 | #else |
2649 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2650 | { | |
2651 | } | |
0ec8aa00 PZ |
2652 | |
2653 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2654 | { | |
2655 | } | |
2656 | ||
2657 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2658 | { | |
2659 | } | |
3fed382b | 2660 | |
3f9672ba SD |
2661 | static inline void update_scan_period(struct task_struct *p, int new_cpu) |
2662 | { | |
2663 | } | |
2664 | ||
cbee9f88 PZ |
2665 | #endif /* CONFIG_NUMA_BALANCING */ |
2666 | ||
30cfdcfc DA |
2667 | static void |
2668 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2669 | { | |
2670 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2671 | if (!parent_entity(se)) |
029632fb | 2672 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 2673 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2674 | if (entity_is_task(se)) { |
2675 | struct rq *rq = rq_of(cfs_rq); | |
2676 | ||
2677 | account_numa_enqueue(rq, task_of(se)); | |
2678 | list_add(&se->group_node, &rq->cfs_tasks); | |
2679 | } | |
367456c7 | 2680 | #endif |
30cfdcfc | 2681 | cfs_rq->nr_running++; |
30cfdcfc DA |
2682 | } |
2683 | ||
2684 | static void | |
2685 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2686 | { | |
2687 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2688 | if (!parent_entity(se)) |
029632fb | 2689 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
bfdb198c | 2690 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2691 | if (entity_is_task(se)) { |
2692 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 2693 | list_del_init(&se->group_node); |
0ec8aa00 | 2694 | } |
bfdb198c | 2695 | #endif |
30cfdcfc | 2696 | cfs_rq->nr_running--; |
30cfdcfc DA |
2697 | } |
2698 | ||
8d5b9025 PZ |
2699 | /* |
2700 | * Signed add and clamp on underflow. | |
2701 | * | |
2702 | * Explicitly do a load-store to ensure the intermediate value never hits | |
2703 | * memory. This allows lockless observations without ever seeing the negative | |
2704 | * values. | |
2705 | */ | |
2706 | #define add_positive(_ptr, _val) do { \ | |
2707 | typeof(_ptr) ptr = (_ptr); \ | |
2708 | typeof(_val) val = (_val); \ | |
2709 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
2710 | \ | |
2711 | res = var + val; \ | |
2712 | \ | |
2713 | if (val < 0 && res > var) \ | |
2714 | res = 0; \ | |
2715 | \ | |
2716 | WRITE_ONCE(*ptr, res); \ | |
2717 | } while (0) | |
2718 | ||
2719 | /* | |
2720 | * Unsigned subtract and clamp on underflow. | |
2721 | * | |
2722 | * Explicitly do a load-store to ensure the intermediate value never hits | |
2723 | * memory. This allows lockless observations without ever seeing the negative | |
2724 | * values. | |
2725 | */ | |
2726 | #define sub_positive(_ptr, _val) do { \ | |
2727 | typeof(_ptr) ptr = (_ptr); \ | |
2728 | typeof(*ptr) val = (_val); \ | |
2729 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
2730 | res = var - val; \ | |
2731 | if (res > var) \ | |
2732 | res = 0; \ | |
2733 | WRITE_ONCE(*ptr, res); \ | |
2734 | } while (0) | |
2735 | ||
b5c0ce7b PB |
2736 | /* |
2737 | * Remove and clamp on negative, from a local variable. | |
2738 | * | |
2739 | * A variant of sub_positive(), which does not use explicit load-store | |
2740 | * and is thus optimized for local variable updates. | |
2741 | */ | |
2742 | #define lsub_positive(_ptr, _val) do { \ | |
2743 | typeof(_ptr) ptr = (_ptr); \ | |
2744 | *ptr -= min_t(typeof(*ptr), *ptr, _val); \ | |
2745 | } while (0) | |
2746 | ||
8d5b9025 | 2747 | #ifdef CONFIG_SMP |
8d5b9025 PZ |
2748 | static inline void |
2749 | enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2750 | { | |
1ea6c46a PZ |
2751 | cfs_rq->runnable_weight += se->runnable_weight; |
2752 | ||
2753 | cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg; | |
2754 | cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum; | |
8d5b9025 PZ |
2755 | } |
2756 | ||
2757 | static inline void | |
2758 | dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2759 | { | |
1ea6c46a PZ |
2760 | cfs_rq->runnable_weight -= se->runnable_weight; |
2761 | ||
2762 | sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg); | |
2763 | sub_positive(&cfs_rq->avg.runnable_load_sum, | |
2764 | se_runnable(se) * se->avg.runnable_load_sum); | |
8d5b9025 PZ |
2765 | } |
2766 | ||
2767 | static inline void | |
2768 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2769 | { | |
2770 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
2771 | cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; | |
2772 | } | |
2773 | ||
2774 | static inline void | |
2775 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2776 | { | |
2777 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); | |
2778 | sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); | |
2779 | } | |
2780 | #else | |
2781 | static inline void | |
2782 | enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2783 | static inline void | |
2784 | dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2785 | static inline void | |
2786 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2787 | static inline void | |
2788 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2789 | #endif | |
2790 | ||
9059393e | 2791 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1ea6c46a | 2792 | unsigned long weight, unsigned long runnable) |
9059393e VG |
2793 | { |
2794 | if (se->on_rq) { | |
2795 | /* commit outstanding execution time */ | |
2796 | if (cfs_rq->curr == se) | |
2797 | update_curr(cfs_rq); | |
2798 | account_entity_dequeue(cfs_rq, se); | |
2799 | dequeue_runnable_load_avg(cfs_rq, se); | |
2800 | } | |
2801 | dequeue_load_avg(cfs_rq, se); | |
2802 | ||
1ea6c46a | 2803 | se->runnable_weight = runnable; |
9059393e VG |
2804 | update_load_set(&se->load, weight); |
2805 | ||
2806 | #ifdef CONFIG_SMP | |
1ea6c46a PZ |
2807 | do { |
2808 | u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib; | |
2809 | ||
2810 | se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); | |
2811 | se->avg.runnable_load_avg = | |
2812 | div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider); | |
2813 | } while (0); | |
9059393e VG |
2814 | #endif |
2815 | ||
2816 | enqueue_load_avg(cfs_rq, se); | |
2817 | if (se->on_rq) { | |
2818 | account_entity_enqueue(cfs_rq, se); | |
2819 | enqueue_runnable_load_avg(cfs_rq, se); | |
2820 | } | |
2821 | } | |
2822 | ||
2823 | void reweight_task(struct task_struct *p, int prio) | |
2824 | { | |
2825 | struct sched_entity *se = &p->se; | |
2826 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2827 | struct load_weight *load = &se->load; | |
2828 | unsigned long weight = scale_load(sched_prio_to_weight[prio]); | |
2829 | ||
1ea6c46a | 2830 | reweight_entity(cfs_rq, se, weight, weight); |
9059393e VG |
2831 | load->inv_weight = sched_prio_to_wmult[prio]; |
2832 | } | |
2833 | ||
3ff6dcac | 2834 | #ifdef CONFIG_FAIR_GROUP_SCHED |
387f77cc | 2835 | #ifdef CONFIG_SMP |
cef27403 PZ |
2836 | /* |
2837 | * All this does is approximate the hierarchical proportion which includes that | |
2838 | * global sum we all love to hate. | |
2839 | * | |
2840 | * That is, the weight of a group entity, is the proportional share of the | |
2841 | * group weight based on the group runqueue weights. That is: | |
2842 | * | |
2843 | * tg->weight * grq->load.weight | |
2844 | * ge->load.weight = ----------------------------- (1) | |
2845 | * \Sum grq->load.weight | |
2846 | * | |
2847 | * Now, because computing that sum is prohibitively expensive to compute (been | |
2848 | * there, done that) we approximate it with this average stuff. The average | |
2849 | * moves slower and therefore the approximation is cheaper and more stable. | |
2850 | * | |
2851 | * So instead of the above, we substitute: | |
2852 | * | |
2853 | * grq->load.weight -> grq->avg.load_avg (2) | |
2854 | * | |
2855 | * which yields the following: | |
2856 | * | |
2857 | * tg->weight * grq->avg.load_avg | |
2858 | * ge->load.weight = ------------------------------ (3) | |
2859 | * tg->load_avg | |
2860 | * | |
2861 | * Where: tg->load_avg ~= \Sum grq->avg.load_avg | |
2862 | * | |
2863 | * That is shares_avg, and it is right (given the approximation (2)). | |
2864 | * | |
2865 | * The problem with it is that because the average is slow -- it was designed | |
2866 | * to be exactly that of course -- this leads to transients in boundary | |
2867 | * conditions. In specific, the case where the group was idle and we start the | |
2868 | * one task. It takes time for our CPU's grq->avg.load_avg to build up, | |
2869 | * yielding bad latency etc.. | |
2870 | * | |
2871 | * Now, in that special case (1) reduces to: | |
2872 | * | |
2873 | * tg->weight * grq->load.weight | |
17de4ee0 | 2874 | * ge->load.weight = ----------------------------- = tg->weight (4) |
cef27403 PZ |
2875 | * grp->load.weight |
2876 | * | |
2877 | * That is, the sum collapses because all other CPUs are idle; the UP scenario. | |
2878 | * | |
2879 | * So what we do is modify our approximation (3) to approach (4) in the (near) | |
2880 | * UP case, like: | |
2881 | * | |
2882 | * ge->load.weight = | |
2883 | * | |
2884 | * tg->weight * grq->load.weight | |
2885 | * --------------------------------------------------- (5) | |
2886 | * tg->load_avg - grq->avg.load_avg + grq->load.weight | |
2887 | * | |
17de4ee0 PZ |
2888 | * But because grq->load.weight can drop to 0, resulting in a divide by zero, |
2889 | * we need to use grq->avg.load_avg as its lower bound, which then gives: | |
2890 | * | |
2891 | * | |
2892 | * tg->weight * grq->load.weight | |
2893 | * ge->load.weight = ----------------------------- (6) | |
2894 | * tg_load_avg' | |
2895 | * | |
2896 | * Where: | |
2897 | * | |
2898 | * tg_load_avg' = tg->load_avg - grq->avg.load_avg + | |
2899 | * max(grq->load.weight, grq->avg.load_avg) | |
cef27403 PZ |
2900 | * |
2901 | * And that is shares_weight and is icky. In the (near) UP case it approaches | |
2902 | * (4) while in the normal case it approaches (3). It consistently | |
2903 | * overestimates the ge->load.weight and therefore: | |
2904 | * | |
2905 | * \Sum ge->load.weight >= tg->weight | |
2906 | * | |
2907 | * hence icky! | |
2908 | */ | |
2c8e4dce | 2909 | static long calc_group_shares(struct cfs_rq *cfs_rq) |
cf5f0acf | 2910 | { |
7c80cfc9 PZ |
2911 | long tg_weight, tg_shares, load, shares; |
2912 | struct task_group *tg = cfs_rq->tg; | |
2913 | ||
2914 | tg_shares = READ_ONCE(tg->shares); | |
cf5f0acf | 2915 | |
3d4b60d3 | 2916 | load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); |
cf5f0acf | 2917 | |
ea1dc6fc | 2918 | tg_weight = atomic_long_read(&tg->load_avg); |
3ff6dcac | 2919 | |
ea1dc6fc PZ |
2920 | /* Ensure tg_weight >= load */ |
2921 | tg_weight -= cfs_rq->tg_load_avg_contrib; | |
2922 | tg_weight += load; | |
3ff6dcac | 2923 | |
7c80cfc9 | 2924 | shares = (tg_shares * load); |
cf5f0acf PZ |
2925 | if (tg_weight) |
2926 | shares /= tg_weight; | |
3ff6dcac | 2927 | |
b8fd8423 DE |
2928 | /* |
2929 | * MIN_SHARES has to be unscaled here to support per-CPU partitioning | |
2930 | * of a group with small tg->shares value. It is a floor value which is | |
2931 | * assigned as a minimum load.weight to the sched_entity representing | |
2932 | * the group on a CPU. | |
2933 | * | |
2934 | * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 | |
2935 | * on an 8-core system with 8 tasks each runnable on one CPU shares has | |
2936 | * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In | |
2937 | * case no task is runnable on a CPU MIN_SHARES=2 should be returned | |
2938 | * instead of 0. | |
2939 | */ | |
7c80cfc9 | 2940 | return clamp_t(long, shares, MIN_SHARES, tg_shares); |
3ff6dcac | 2941 | } |
2c8e4dce JB |
2942 | |
2943 | /* | |
17de4ee0 PZ |
2944 | * This calculates the effective runnable weight for a group entity based on |
2945 | * the group entity weight calculated above. | |
2946 | * | |
2947 | * Because of the above approximation (2), our group entity weight is | |
2948 | * an load_avg based ratio (3). This means that it includes blocked load and | |
2949 | * does not represent the runnable weight. | |
2950 | * | |
2951 | * Approximate the group entity's runnable weight per ratio from the group | |
2952 | * runqueue: | |
2953 | * | |
2954 | * grq->avg.runnable_load_avg | |
2955 | * ge->runnable_weight = ge->load.weight * -------------------------- (7) | |
2956 | * grq->avg.load_avg | |
2957 | * | |
2958 | * However, analogous to above, since the avg numbers are slow, this leads to | |
2959 | * transients in the from-idle case. Instead we use: | |
2960 | * | |
2961 | * ge->runnable_weight = ge->load.weight * | |
2962 | * | |
2963 | * max(grq->avg.runnable_load_avg, grq->runnable_weight) | |
2964 | * ----------------------------------------------------- (8) | |
2965 | * max(grq->avg.load_avg, grq->load.weight) | |
2966 | * | |
2967 | * Where these max() serve both to use the 'instant' values to fix the slow | |
2968 | * from-idle and avoid the /0 on to-idle, similar to (6). | |
2c8e4dce JB |
2969 | */ |
2970 | static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares) | |
2971 | { | |
17de4ee0 PZ |
2972 | long runnable, load_avg; |
2973 | ||
2974 | load_avg = max(cfs_rq->avg.load_avg, | |
2975 | scale_load_down(cfs_rq->load.weight)); | |
2976 | ||
2977 | runnable = max(cfs_rq->avg.runnable_load_avg, | |
2978 | scale_load_down(cfs_rq->runnable_weight)); | |
2c8e4dce JB |
2979 | |
2980 | runnable *= shares; | |
2981 | if (load_avg) | |
2982 | runnable /= load_avg; | |
17de4ee0 | 2983 | |
2c8e4dce JB |
2984 | return clamp_t(long, runnable, MIN_SHARES, shares); |
2985 | } | |
387f77cc | 2986 | #endif /* CONFIG_SMP */ |
ea1dc6fc | 2987 | |
82958366 PT |
2988 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
2989 | ||
1ea6c46a PZ |
2990 | /* |
2991 | * Recomputes the group entity based on the current state of its group | |
2992 | * runqueue. | |
2993 | */ | |
2994 | static void update_cfs_group(struct sched_entity *se) | |
2069dd75 | 2995 | { |
1ea6c46a PZ |
2996 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); |
2997 | long shares, runnable; | |
2069dd75 | 2998 | |
1ea6c46a | 2999 | if (!gcfs_rq) |
89ee048f VG |
3000 | return; |
3001 | ||
1ea6c46a | 3002 | if (throttled_hierarchy(gcfs_rq)) |
2069dd75 | 3003 | return; |
89ee048f | 3004 | |
3ff6dcac | 3005 | #ifndef CONFIG_SMP |
1ea6c46a | 3006 | runnable = shares = READ_ONCE(gcfs_rq->tg->shares); |
7c80cfc9 PZ |
3007 | |
3008 | if (likely(se->load.weight == shares)) | |
3ff6dcac | 3009 | return; |
7c80cfc9 | 3010 | #else |
2c8e4dce JB |
3011 | shares = calc_group_shares(gcfs_rq); |
3012 | runnable = calc_group_runnable(gcfs_rq, shares); | |
3ff6dcac | 3013 | #endif |
2069dd75 | 3014 | |
1ea6c46a | 3015 | reweight_entity(cfs_rq_of(se), se, shares, runnable); |
2069dd75 | 3016 | } |
89ee048f | 3017 | |
2069dd75 | 3018 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
1ea6c46a | 3019 | static inline void update_cfs_group(struct sched_entity *se) |
2069dd75 PZ |
3020 | { |
3021 | } | |
3022 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
3023 | ||
ea14b57e | 3024 | static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags) |
a030d738 | 3025 | { |
43964409 LT |
3026 | struct rq *rq = rq_of(cfs_rq); |
3027 | ||
ea14b57e | 3028 | if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) { |
a030d738 VK |
3029 | /* |
3030 | * There are a few boundary cases this might miss but it should | |
3031 | * get called often enough that that should (hopefully) not be | |
9783be2c | 3032 | * a real problem. |
a030d738 VK |
3033 | * |
3034 | * It will not get called when we go idle, because the idle | |
3035 | * thread is a different class (!fair), nor will the utilization | |
3036 | * number include things like RT tasks. | |
3037 | * | |
3038 | * As is, the util number is not freq-invariant (we'd have to | |
3039 | * implement arch_scale_freq_capacity() for that). | |
3040 | * | |
3041 | * See cpu_util(). | |
3042 | */ | |
ea14b57e | 3043 | cpufreq_update_util(rq, flags); |
a030d738 VK |
3044 | } |
3045 | } | |
3046 | ||
141965c7 | 3047 | #ifdef CONFIG_SMP |
c566e8e9 | 3048 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7c3edd2c PZ |
3049 | /** |
3050 | * update_tg_load_avg - update the tg's load avg | |
3051 | * @cfs_rq: the cfs_rq whose avg changed | |
3052 | * @force: update regardless of how small the difference | |
3053 | * | |
3054 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
3055 | * However, because tg->load_avg is a global value there are performance | |
3056 | * considerations. | |
3057 | * | |
3058 | * In order to avoid having to look at the other cfs_rq's, we use a | |
3059 | * differential update where we store the last value we propagated. This in | |
3060 | * turn allows skipping updates if the differential is 'small'. | |
3061 | * | |
815abf5a | 3062 | * Updating tg's load_avg is necessary before update_cfs_share(). |
bb17f655 | 3063 | */ |
9d89c257 | 3064 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
bb17f655 | 3065 | { |
9d89c257 | 3066 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 3067 | |
aa0b7ae0 WL |
3068 | /* |
3069 | * No need to update load_avg for root_task_group as it is not used. | |
3070 | */ | |
3071 | if (cfs_rq->tg == &root_task_group) | |
3072 | return; | |
3073 | ||
9d89c257 YD |
3074 | if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
3075 | atomic_long_add(delta, &cfs_rq->tg->load_avg); | |
3076 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 3077 | } |
8165e145 | 3078 | } |
f5f9739d | 3079 | |
ad936d86 | 3080 | /* |
97fb7a0a | 3081 | * Called within set_task_rq() right before setting a task's CPU. The |
ad936d86 BP |
3082 | * caller only guarantees p->pi_lock is held; no other assumptions, |
3083 | * including the state of rq->lock, should be made. | |
3084 | */ | |
3085 | void set_task_rq_fair(struct sched_entity *se, | |
3086 | struct cfs_rq *prev, struct cfs_rq *next) | |
3087 | { | |
0ccb977f PZ |
3088 | u64 p_last_update_time; |
3089 | u64 n_last_update_time; | |
3090 | ||
ad936d86 BP |
3091 | if (!sched_feat(ATTACH_AGE_LOAD)) |
3092 | return; | |
3093 | ||
3094 | /* | |
3095 | * We are supposed to update the task to "current" time, then its up to | |
3096 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
3097 | * getting what current time is, so simply throw away the out-of-date | |
3098 | * time. This will result in the wakee task is less decayed, but giving | |
3099 | * the wakee more load sounds not bad. | |
3100 | */ | |
0ccb977f PZ |
3101 | if (!(se->avg.last_update_time && prev)) |
3102 | return; | |
ad936d86 BP |
3103 | |
3104 | #ifndef CONFIG_64BIT | |
0ccb977f | 3105 | { |
ad936d86 BP |
3106 | u64 p_last_update_time_copy; |
3107 | u64 n_last_update_time_copy; | |
3108 | ||
3109 | do { | |
3110 | p_last_update_time_copy = prev->load_last_update_time_copy; | |
3111 | n_last_update_time_copy = next->load_last_update_time_copy; | |
3112 | ||
3113 | smp_rmb(); | |
3114 | ||
3115 | p_last_update_time = prev->avg.last_update_time; | |
3116 | n_last_update_time = next->avg.last_update_time; | |
3117 | ||
3118 | } while (p_last_update_time != p_last_update_time_copy || | |
3119 | n_last_update_time != n_last_update_time_copy); | |
0ccb977f | 3120 | } |
ad936d86 | 3121 | #else |
0ccb977f PZ |
3122 | p_last_update_time = prev->avg.last_update_time; |
3123 | n_last_update_time = next->avg.last_update_time; | |
ad936d86 | 3124 | #endif |
0ccb977f PZ |
3125 | __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se); |
3126 | se->avg.last_update_time = n_last_update_time; | |
ad936d86 | 3127 | } |
09a43ace | 3128 | |
0e2d2aaa PZ |
3129 | |
3130 | /* | |
3131 | * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to | |
3132 | * propagate its contribution. The key to this propagation is the invariant | |
3133 | * that for each group: | |
3134 | * | |
3135 | * ge->avg == grq->avg (1) | |
3136 | * | |
3137 | * _IFF_ we look at the pure running and runnable sums. Because they | |
3138 | * represent the very same entity, just at different points in the hierarchy. | |
3139 | * | |
a4c3c049 VG |
3140 | * Per the above update_tg_cfs_util() is trivial and simply copies the running |
3141 | * sum over (but still wrong, because the group entity and group rq do not have | |
3142 | * their PELT windows aligned). | |
0e2d2aaa PZ |
3143 | * |
3144 | * However, update_tg_cfs_runnable() is more complex. So we have: | |
3145 | * | |
3146 | * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) | |
3147 | * | |
3148 | * And since, like util, the runnable part should be directly transferable, | |
3149 | * the following would _appear_ to be the straight forward approach: | |
3150 | * | |
a4c3c049 | 3151 | * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3) |
0e2d2aaa PZ |
3152 | * |
3153 | * And per (1) we have: | |
3154 | * | |
a4c3c049 | 3155 | * ge->avg.runnable_avg == grq->avg.runnable_avg |
0e2d2aaa PZ |
3156 | * |
3157 | * Which gives: | |
3158 | * | |
3159 | * ge->load.weight * grq->avg.load_avg | |
3160 | * ge->avg.load_avg = ----------------------------------- (4) | |
3161 | * grq->load.weight | |
3162 | * | |
3163 | * Except that is wrong! | |
3164 | * | |
3165 | * Because while for entities historical weight is not important and we | |
3166 | * really only care about our future and therefore can consider a pure | |
3167 | * runnable sum, runqueues can NOT do this. | |
3168 | * | |
3169 | * We specifically want runqueues to have a load_avg that includes | |
3170 | * historical weights. Those represent the blocked load, the load we expect | |
3171 | * to (shortly) return to us. This only works by keeping the weights as | |
3172 | * integral part of the sum. We therefore cannot decompose as per (3). | |
3173 | * | |
a4c3c049 VG |
3174 | * Another reason this doesn't work is that runnable isn't a 0-sum entity. |
3175 | * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the | |
3176 | * rq itself is runnable anywhere between 2/3 and 1 depending on how the | |
3177 | * runnable section of these tasks overlap (or not). If they were to perfectly | |
3178 | * align the rq as a whole would be runnable 2/3 of the time. If however we | |
3179 | * always have at least 1 runnable task, the rq as a whole is always runnable. | |
0e2d2aaa | 3180 | * |
a4c3c049 | 3181 | * So we'll have to approximate.. :/ |
0e2d2aaa | 3182 | * |
a4c3c049 | 3183 | * Given the constraint: |
0e2d2aaa | 3184 | * |
a4c3c049 | 3185 | * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX |
0e2d2aaa | 3186 | * |
a4c3c049 VG |
3187 | * We can construct a rule that adds runnable to a rq by assuming minimal |
3188 | * overlap. | |
0e2d2aaa | 3189 | * |
a4c3c049 | 3190 | * On removal, we'll assume each task is equally runnable; which yields: |
0e2d2aaa | 3191 | * |
a4c3c049 | 3192 | * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight |
0e2d2aaa | 3193 | * |
a4c3c049 | 3194 | * XXX: only do this for the part of runnable > running ? |
0e2d2aaa | 3195 | * |
0e2d2aaa PZ |
3196 | */ |
3197 | ||
09a43ace | 3198 | static inline void |
0e2d2aaa | 3199 | update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3200 | { |
09a43ace VG |
3201 | long delta = gcfs_rq->avg.util_avg - se->avg.util_avg; |
3202 | ||
3203 | /* Nothing to update */ | |
3204 | if (!delta) | |
3205 | return; | |
3206 | ||
a4c3c049 VG |
3207 | /* |
3208 | * The relation between sum and avg is: | |
3209 | * | |
3210 | * LOAD_AVG_MAX - 1024 + sa->period_contrib | |
3211 | * | |
3212 | * however, the PELT windows are not aligned between grq and gse. | |
3213 | */ | |
3214 | ||
09a43ace VG |
3215 | /* Set new sched_entity's utilization */ |
3216 | se->avg.util_avg = gcfs_rq->avg.util_avg; | |
3217 | se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX; | |
3218 | ||
3219 | /* Update parent cfs_rq utilization */ | |
3220 | add_positive(&cfs_rq->avg.util_avg, delta); | |
3221 | cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX; | |
3222 | } | |
3223 | ||
09a43ace | 3224 | static inline void |
0e2d2aaa | 3225 | update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3226 | { |
a4c3c049 VG |
3227 | long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; |
3228 | unsigned long runnable_load_avg, load_avg; | |
3229 | u64 runnable_load_sum, load_sum = 0; | |
3230 | s64 delta_sum; | |
09a43ace | 3231 | |
0e2d2aaa PZ |
3232 | if (!runnable_sum) |
3233 | return; | |
09a43ace | 3234 | |
0e2d2aaa | 3235 | gcfs_rq->prop_runnable_sum = 0; |
09a43ace | 3236 | |
a4c3c049 VG |
3237 | if (runnable_sum >= 0) { |
3238 | /* | |
3239 | * Add runnable; clip at LOAD_AVG_MAX. Reflects that until | |
3240 | * the CPU is saturated running == runnable. | |
3241 | */ | |
3242 | runnable_sum += se->avg.load_sum; | |
3243 | runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX); | |
3244 | } else { | |
3245 | /* | |
3246 | * Estimate the new unweighted runnable_sum of the gcfs_rq by | |
3247 | * assuming all tasks are equally runnable. | |
3248 | */ | |
3249 | if (scale_load_down(gcfs_rq->load.weight)) { | |
3250 | load_sum = div_s64(gcfs_rq->avg.load_sum, | |
3251 | scale_load_down(gcfs_rq->load.weight)); | |
3252 | } | |
3253 | ||
3254 | /* But make sure to not inflate se's runnable */ | |
3255 | runnable_sum = min(se->avg.load_sum, load_sum); | |
3256 | } | |
3257 | ||
3258 | /* | |
3259 | * runnable_sum can't be lower than running_sum | |
97fb7a0a | 3260 | * As running sum is scale with CPU capacity wehreas the runnable sum |
a4c3c049 VG |
3261 | * is not we rescale running_sum 1st |
3262 | */ | |
3263 | running_sum = se->avg.util_sum / | |
3264 | arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq))); | |
3265 | runnable_sum = max(runnable_sum, running_sum); | |
3266 | ||
0e2d2aaa PZ |
3267 | load_sum = (s64)se_weight(se) * runnable_sum; |
3268 | load_avg = div_s64(load_sum, LOAD_AVG_MAX); | |
09a43ace | 3269 | |
a4c3c049 VG |
3270 | delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum; |
3271 | delta_avg = load_avg - se->avg.load_avg; | |
09a43ace | 3272 | |
a4c3c049 VG |
3273 | se->avg.load_sum = runnable_sum; |
3274 | se->avg.load_avg = load_avg; | |
3275 | add_positive(&cfs_rq->avg.load_avg, delta_avg); | |
3276 | add_positive(&cfs_rq->avg.load_sum, delta_sum); | |
09a43ace | 3277 | |
1ea6c46a PZ |
3278 | runnable_load_sum = (s64)se_runnable(se) * runnable_sum; |
3279 | runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX); | |
a4c3c049 VG |
3280 | delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum; |
3281 | delta_avg = runnable_load_avg - se->avg.runnable_load_avg; | |
1ea6c46a | 3282 | |
a4c3c049 VG |
3283 | se->avg.runnable_load_sum = runnable_sum; |
3284 | se->avg.runnable_load_avg = runnable_load_avg; | |
1ea6c46a | 3285 | |
09a43ace | 3286 | if (se->on_rq) { |
a4c3c049 VG |
3287 | add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg); |
3288 | add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum); | |
09a43ace VG |
3289 | } |
3290 | } | |
3291 | ||
0e2d2aaa | 3292 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) |
09a43ace | 3293 | { |
0e2d2aaa PZ |
3294 | cfs_rq->propagate = 1; |
3295 | cfs_rq->prop_runnable_sum += runnable_sum; | |
09a43ace VG |
3296 | } |
3297 | ||
3298 | /* Update task and its cfs_rq load average */ | |
3299 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3300 | { | |
0e2d2aaa | 3301 | struct cfs_rq *cfs_rq, *gcfs_rq; |
09a43ace VG |
3302 | |
3303 | if (entity_is_task(se)) | |
3304 | return 0; | |
3305 | ||
0e2d2aaa PZ |
3306 | gcfs_rq = group_cfs_rq(se); |
3307 | if (!gcfs_rq->propagate) | |
09a43ace VG |
3308 | return 0; |
3309 | ||
0e2d2aaa PZ |
3310 | gcfs_rq->propagate = 0; |
3311 | ||
09a43ace VG |
3312 | cfs_rq = cfs_rq_of(se); |
3313 | ||
0e2d2aaa | 3314 | add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); |
09a43ace | 3315 | |
0e2d2aaa PZ |
3316 | update_tg_cfs_util(cfs_rq, se, gcfs_rq); |
3317 | update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); | |
09a43ace VG |
3318 | |
3319 | return 1; | |
3320 | } | |
3321 | ||
bc427898 VG |
3322 | /* |
3323 | * Check if we need to update the load and the utilization of a blocked | |
3324 | * group_entity: | |
3325 | */ | |
3326 | static inline bool skip_blocked_update(struct sched_entity *se) | |
3327 | { | |
3328 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3329 | ||
3330 | /* | |
3331 | * If sched_entity still have not zero load or utilization, we have to | |
3332 | * decay it: | |
3333 | */ | |
3334 | if (se->avg.load_avg || se->avg.util_avg) | |
3335 | return false; | |
3336 | ||
3337 | /* | |
3338 | * If there is a pending propagation, we have to update the load and | |
3339 | * the utilization of the sched_entity: | |
3340 | */ | |
0e2d2aaa | 3341 | if (gcfs_rq->propagate) |
bc427898 VG |
3342 | return false; |
3343 | ||
3344 | /* | |
3345 | * Otherwise, the load and the utilization of the sched_entity is | |
3346 | * already zero and there is no pending propagation, so it will be a | |
3347 | * waste of time to try to decay it: | |
3348 | */ | |
3349 | return true; | |
3350 | } | |
3351 | ||
6e83125c | 3352 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
09a43ace | 3353 | |
9d89c257 | 3354 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {} |
09a43ace VG |
3355 | |
3356 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3357 | { | |
3358 | return 0; | |
3359 | } | |
3360 | ||
0e2d2aaa | 3361 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} |
09a43ace | 3362 | |
6e83125c | 3363 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 3364 | |
3d30544f PZ |
3365 | /** |
3366 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
3367 | * @now: current time, as per cfs_rq_clock_task() | |
3368 | * @cfs_rq: cfs_rq to update | |
3d30544f PZ |
3369 | * |
3370 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
3371 | * avg. The immediate corollary is that all (fair) tasks must be attached, see | |
3372 | * post_init_entity_util_avg(). | |
3373 | * | |
3374 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
3375 | * | |
7c3edd2c PZ |
3376 | * Returns true if the load decayed or we removed load. |
3377 | * | |
3378 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
3379 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 3380 | */ |
a2c6c91f | 3381 | static inline int |
3a123bbb | 3382 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) |
2dac754e | 3383 | { |
0e2d2aaa | 3384 | unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0; |
9d89c257 | 3385 | struct sched_avg *sa = &cfs_rq->avg; |
2a2f5d4e | 3386 | int decayed = 0; |
2dac754e | 3387 | |
2a2f5d4e PZ |
3388 | if (cfs_rq->removed.nr) { |
3389 | unsigned long r; | |
9a2dd585 | 3390 | u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib; |
2a2f5d4e PZ |
3391 | |
3392 | raw_spin_lock(&cfs_rq->removed.lock); | |
3393 | swap(cfs_rq->removed.util_avg, removed_util); | |
3394 | swap(cfs_rq->removed.load_avg, removed_load); | |
0e2d2aaa | 3395 | swap(cfs_rq->removed.runnable_sum, removed_runnable_sum); |
2a2f5d4e PZ |
3396 | cfs_rq->removed.nr = 0; |
3397 | raw_spin_unlock(&cfs_rq->removed.lock); | |
3398 | ||
2a2f5d4e | 3399 | r = removed_load; |
89741892 | 3400 | sub_positive(&sa->load_avg, r); |
9a2dd585 | 3401 | sub_positive(&sa->load_sum, r * divider); |
2dac754e | 3402 | |
2a2f5d4e | 3403 | r = removed_util; |
89741892 | 3404 | sub_positive(&sa->util_avg, r); |
9a2dd585 | 3405 | sub_positive(&sa->util_sum, r * divider); |
2a2f5d4e | 3406 | |
0e2d2aaa | 3407 | add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum); |
2a2f5d4e PZ |
3408 | |
3409 | decayed = 1; | |
9d89c257 | 3410 | } |
36ee28e4 | 3411 | |
2a2f5d4e | 3412 | decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq); |
36ee28e4 | 3413 | |
9d89c257 YD |
3414 | #ifndef CONFIG_64BIT |
3415 | smp_wmb(); | |
3416 | cfs_rq->load_last_update_time_copy = sa->last_update_time; | |
3417 | #endif | |
36ee28e4 | 3418 | |
2a2f5d4e | 3419 | if (decayed) |
ea14b57e | 3420 | cfs_rq_util_change(cfs_rq, 0); |
21e96f88 | 3421 | |
2a2f5d4e | 3422 | return decayed; |
21e96f88 SM |
3423 | } |
3424 | ||
3d30544f PZ |
3425 | /** |
3426 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
3427 | * @cfs_rq: cfs_rq to attach to | |
3428 | * @se: sched_entity to attach | |
882a78a9 | 3429 | * @flags: migration hints |
3d30544f PZ |
3430 | * |
3431 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3432 | * cfs_rq->avg.last_update_time being current. | |
3433 | */ | |
ea14b57e | 3434 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
a05e8c51 | 3435 | { |
f207934f PZ |
3436 | u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib; |
3437 | ||
3438 | /* | |
3439 | * When we attach the @se to the @cfs_rq, we must align the decay | |
3440 | * window because without that, really weird and wonderful things can | |
3441 | * happen. | |
3442 | * | |
3443 | * XXX illustrate | |
3444 | */ | |
a05e8c51 | 3445 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
f207934f PZ |
3446 | se->avg.period_contrib = cfs_rq->avg.period_contrib; |
3447 | ||
3448 | /* | |
3449 | * Hell(o) Nasty stuff.. we need to recompute _sum based on the new | |
3450 | * period_contrib. This isn't strictly correct, but since we're | |
3451 | * entirely outside of the PELT hierarchy, nobody cares if we truncate | |
3452 | * _sum a little. | |
3453 | */ | |
3454 | se->avg.util_sum = se->avg.util_avg * divider; | |
3455 | ||
3456 | se->avg.load_sum = divider; | |
3457 | if (se_weight(se)) { | |
3458 | se->avg.load_sum = | |
3459 | div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se)); | |
3460 | } | |
3461 | ||
3462 | se->avg.runnable_load_sum = se->avg.load_sum; | |
3463 | ||
8d5b9025 | 3464 | enqueue_load_avg(cfs_rq, se); |
a05e8c51 BP |
3465 | cfs_rq->avg.util_avg += se->avg.util_avg; |
3466 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
0e2d2aaa PZ |
3467 | |
3468 | add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); | |
a2c6c91f | 3469 | |
ea14b57e | 3470 | cfs_rq_util_change(cfs_rq, flags); |
a05e8c51 BP |
3471 | } |
3472 | ||
3d30544f PZ |
3473 | /** |
3474 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
3475 | * @cfs_rq: cfs_rq to detach from | |
3476 | * @se: sched_entity to detach | |
3477 | * | |
3478 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3479 | * cfs_rq->avg.last_update_time being current. | |
3480 | */ | |
a05e8c51 BP |
3481 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3482 | { | |
8d5b9025 | 3483 | dequeue_load_avg(cfs_rq, se); |
89741892 PZ |
3484 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); |
3485 | sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); | |
0e2d2aaa PZ |
3486 | |
3487 | add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); | |
a2c6c91f | 3488 | |
ea14b57e | 3489 | cfs_rq_util_change(cfs_rq, 0); |
a05e8c51 BP |
3490 | } |
3491 | ||
b382a531 PZ |
3492 | /* |
3493 | * Optional action to be done while updating the load average | |
3494 | */ | |
3495 | #define UPDATE_TG 0x1 | |
3496 | #define SKIP_AGE_LOAD 0x2 | |
3497 | #define DO_ATTACH 0x4 | |
3498 | ||
3499 | /* Update task and its cfs_rq load average */ | |
3500 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
3501 | { | |
3502 | u64 now = cfs_rq_clock_task(cfs_rq); | |
3503 | struct rq *rq = rq_of(cfs_rq); | |
3504 | int cpu = cpu_of(rq); | |
3505 | int decayed; | |
3506 | ||
3507 | /* | |
3508 | * Track task load average for carrying it to new CPU after migrated, and | |
3509 | * track group sched_entity load average for task_h_load calc in migration | |
3510 | */ | |
3511 | if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) | |
3512 | __update_load_avg_se(now, cpu, cfs_rq, se); | |
3513 | ||
3514 | decayed = update_cfs_rq_load_avg(now, cfs_rq); | |
3515 | decayed |= propagate_entity_load_avg(se); | |
3516 | ||
3517 | if (!se->avg.last_update_time && (flags & DO_ATTACH)) { | |
3518 | ||
ea14b57e PZ |
3519 | /* |
3520 | * DO_ATTACH means we're here from enqueue_entity(). | |
3521 | * !last_update_time means we've passed through | |
3522 | * migrate_task_rq_fair() indicating we migrated. | |
3523 | * | |
3524 | * IOW we're enqueueing a task on a new CPU. | |
3525 | */ | |
3526 | attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION); | |
b382a531 PZ |
3527 | update_tg_load_avg(cfs_rq, 0); |
3528 | ||
3529 | } else if (decayed && (flags & UPDATE_TG)) | |
3530 | update_tg_load_avg(cfs_rq, 0); | |
3531 | } | |
3532 | ||
9d89c257 | 3533 | #ifndef CONFIG_64BIT |
0905f04e YD |
3534 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3535 | { | |
9d89c257 | 3536 | u64 last_update_time_copy; |
0905f04e | 3537 | u64 last_update_time; |
9ee474f5 | 3538 | |
9d89c257 YD |
3539 | do { |
3540 | last_update_time_copy = cfs_rq->load_last_update_time_copy; | |
3541 | smp_rmb(); | |
3542 | last_update_time = cfs_rq->avg.last_update_time; | |
3543 | } while (last_update_time != last_update_time_copy); | |
0905f04e YD |
3544 | |
3545 | return last_update_time; | |
3546 | } | |
9d89c257 | 3547 | #else |
0905f04e YD |
3548 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3549 | { | |
3550 | return cfs_rq->avg.last_update_time; | |
3551 | } | |
9d89c257 YD |
3552 | #endif |
3553 | ||
104cb16d MR |
3554 | /* |
3555 | * Synchronize entity load avg of dequeued entity without locking | |
3556 | * the previous rq. | |
3557 | */ | |
3558 | void sync_entity_load_avg(struct sched_entity *se) | |
3559 | { | |
3560 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3561 | u64 last_update_time; | |
3562 | ||
3563 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
0ccb977f | 3564 | __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se); |
104cb16d MR |
3565 | } |
3566 | ||
0905f04e YD |
3567 | /* |
3568 | * Task first catches up with cfs_rq, and then subtract | |
3569 | * itself from the cfs_rq (task must be off the queue now). | |
3570 | */ | |
3571 | void remove_entity_load_avg(struct sched_entity *se) | |
3572 | { | |
3573 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2a2f5d4e | 3574 | unsigned long flags; |
0905f04e YD |
3575 | |
3576 | /* | |
7dc603c9 PZ |
3577 | * tasks cannot exit without having gone through wake_up_new_task() -> |
3578 | * post_init_entity_util_avg() which will have added things to the | |
3579 | * cfs_rq, so we can remove unconditionally. | |
3580 | * | |
3581 | * Similarly for groups, they will have passed through | |
3582 | * post_init_entity_util_avg() before unregister_sched_fair_group() | |
3583 | * calls this. | |
0905f04e | 3584 | */ |
0905f04e | 3585 | |
104cb16d | 3586 | sync_entity_load_avg(se); |
2a2f5d4e PZ |
3587 | |
3588 | raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); | |
3589 | ++cfs_rq->removed.nr; | |
3590 | cfs_rq->removed.util_avg += se->avg.util_avg; | |
3591 | cfs_rq->removed.load_avg += se->avg.load_avg; | |
0e2d2aaa | 3592 | cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */ |
2a2f5d4e | 3593 | raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); |
2dac754e | 3594 | } |
642dbc39 | 3595 | |
7ea241af YD |
3596 | static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq) |
3597 | { | |
1ea6c46a | 3598 | return cfs_rq->avg.runnable_load_avg; |
7ea241af YD |
3599 | } |
3600 | ||
3601 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) | |
3602 | { | |
3603 | return cfs_rq->avg.load_avg; | |
3604 | } | |
3605 | ||
46f69fa3 | 3606 | static int idle_balance(struct rq *this_rq, struct rq_flags *rf); |
6e83125c | 3607 | |
7f65ea42 PB |
3608 | static inline unsigned long task_util(struct task_struct *p) |
3609 | { | |
3610 | return READ_ONCE(p->se.avg.util_avg); | |
3611 | } | |
3612 | ||
3613 | static inline unsigned long _task_util_est(struct task_struct *p) | |
3614 | { | |
3615 | struct util_est ue = READ_ONCE(p->se.avg.util_est); | |
3616 | ||
92a801e5 | 3617 | return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED); |
7f65ea42 PB |
3618 | } |
3619 | ||
3620 | static inline unsigned long task_util_est(struct task_struct *p) | |
3621 | { | |
3622 | return max(task_util(p), _task_util_est(p)); | |
3623 | } | |
3624 | ||
3625 | static inline void util_est_enqueue(struct cfs_rq *cfs_rq, | |
3626 | struct task_struct *p) | |
3627 | { | |
3628 | unsigned int enqueued; | |
3629 | ||
3630 | if (!sched_feat(UTIL_EST)) | |
3631 | return; | |
3632 | ||
3633 | /* Update root cfs_rq's estimated utilization */ | |
3634 | enqueued = cfs_rq->avg.util_est.enqueued; | |
92a801e5 | 3635 | enqueued += _task_util_est(p); |
7f65ea42 PB |
3636 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); |
3637 | } | |
3638 | ||
3639 | /* | |
3640 | * Check if a (signed) value is within a specified (unsigned) margin, | |
3641 | * based on the observation that: | |
3642 | * | |
3643 | * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1) | |
3644 | * | |
3645 | * NOTE: this only works when value + maring < INT_MAX. | |
3646 | */ | |
3647 | static inline bool within_margin(int value, int margin) | |
3648 | { | |
3649 | return ((unsigned int)(value + margin - 1) < (2 * margin - 1)); | |
3650 | } | |
3651 | ||
3652 | static void | |
3653 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep) | |
3654 | { | |
3655 | long last_ewma_diff; | |
3656 | struct util_est ue; | |
3657 | ||
3658 | if (!sched_feat(UTIL_EST)) | |
3659 | return; | |
3660 | ||
3482d98b VG |
3661 | /* Update root cfs_rq's estimated utilization */ |
3662 | ue.enqueued = cfs_rq->avg.util_est.enqueued; | |
92a801e5 | 3663 | ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p)); |
7f65ea42 PB |
3664 | WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued); |
3665 | ||
3666 | /* | |
3667 | * Skip update of task's estimated utilization when the task has not | |
3668 | * yet completed an activation, e.g. being migrated. | |
3669 | */ | |
3670 | if (!task_sleep) | |
3671 | return; | |
3672 | ||
d519329f PB |
3673 | /* |
3674 | * If the PELT values haven't changed since enqueue time, | |
3675 | * skip the util_est update. | |
3676 | */ | |
3677 | ue = p->se.avg.util_est; | |
3678 | if (ue.enqueued & UTIL_AVG_UNCHANGED) | |
3679 | return; | |
3680 | ||
7f65ea42 PB |
3681 | /* |
3682 | * Skip update of task's estimated utilization when its EWMA is | |
3683 | * already ~1% close to its last activation value. | |
3684 | */ | |
d519329f | 3685 | ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED); |
7f65ea42 PB |
3686 | last_ewma_diff = ue.enqueued - ue.ewma; |
3687 | if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100))) | |
3688 | return; | |
3689 | ||
3690 | /* | |
3691 | * Update Task's estimated utilization | |
3692 | * | |
3693 | * When *p completes an activation we can consolidate another sample | |
3694 | * of the task size. This is done by storing the current PELT value | |
3695 | * as ue.enqueued and by using this value to update the Exponential | |
3696 | * Weighted Moving Average (EWMA): | |
3697 | * | |
3698 | * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) | |
3699 | * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) | |
3700 | * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) | |
3701 | * = w * ( last_ewma_diff ) + ewma(t-1) | |
3702 | * = w * (last_ewma_diff + ewma(t-1) / w) | |
3703 | * | |
3704 | * Where 'w' is the weight of new samples, which is configured to be | |
3705 | * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT) | |
3706 | */ | |
3707 | ue.ewma <<= UTIL_EST_WEIGHT_SHIFT; | |
3708 | ue.ewma += last_ewma_diff; | |
3709 | ue.ewma >>= UTIL_EST_WEIGHT_SHIFT; | |
3710 | WRITE_ONCE(p->se.avg.util_est, ue); | |
3711 | } | |
3712 | ||
3b1baa64 MR |
3713 | static inline int task_fits_capacity(struct task_struct *p, long capacity) |
3714 | { | |
3715 | return capacity * 1024 > task_util_est(p) * capacity_margin; | |
3716 | } | |
3717 | ||
3718 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) | |
3719 | { | |
3720 | if (!static_branch_unlikely(&sched_asym_cpucapacity)) | |
3721 | return; | |
3722 | ||
3723 | if (!p) { | |
3724 | rq->misfit_task_load = 0; | |
3725 | return; | |
3726 | } | |
3727 | ||
3728 | if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) { | |
3729 | rq->misfit_task_load = 0; | |
3730 | return; | |
3731 | } | |
3732 | ||
3733 | rq->misfit_task_load = task_h_load(p); | |
3734 | } | |
3735 | ||
38033c37 PZ |
3736 | #else /* CONFIG_SMP */ |
3737 | ||
d31b1a66 VG |
3738 | #define UPDATE_TG 0x0 |
3739 | #define SKIP_AGE_LOAD 0x0 | |
b382a531 | 3740 | #define DO_ATTACH 0x0 |
d31b1a66 | 3741 | |
88c0616e | 3742 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) |
536bd00c | 3743 | { |
ea14b57e | 3744 | cfs_rq_util_change(cfs_rq, 0); |
536bd00c RW |
3745 | } |
3746 | ||
9d89c257 | 3747 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 3748 | |
a05e8c51 | 3749 | static inline void |
ea14b57e | 3750 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {} |
a05e8c51 BP |
3751 | static inline void |
3752 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
3753 | ||
46f69fa3 | 3754 | static inline int idle_balance(struct rq *rq, struct rq_flags *rf) |
6e83125c PZ |
3755 | { |
3756 | return 0; | |
3757 | } | |
3758 | ||
7f65ea42 PB |
3759 | static inline void |
3760 | util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {} | |
3761 | ||
3762 | static inline void | |
3763 | util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, | |
3764 | bool task_sleep) {} | |
3b1baa64 | 3765 | static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} |
7f65ea42 | 3766 | |
38033c37 | 3767 | #endif /* CONFIG_SMP */ |
9d85f21c | 3768 | |
ddc97297 PZ |
3769 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3770 | { | |
3771 | #ifdef CONFIG_SCHED_DEBUG | |
3772 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
3773 | ||
3774 | if (d < 0) | |
3775 | d = -d; | |
3776 | ||
3777 | if (d > 3*sysctl_sched_latency) | |
ae92882e | 3778 | schedstat_inc(cfs_rq->nr_spread_over); |
ddc97297 PZ |
3779 | #endif |
3780 | } | |
3781 | ||
aeb73b04 PZ |
3782 | static void |
3783 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
3784 | { | |
1af5f730 | 3785 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 3786 | |
2cb8600e PZ |
3787 | /* |
3788 | * The 'current' period is already promised to the current tasks, | |
3789 | * however the extra weight of the new task will slow them down a | |
3790 | * little, place the new task so that it fits in the slot that | |
3791 | * stays open at the end. | |
3792 | */ | |
94dfb5e7 | 3793 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 3794 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 3795 | |
a2e7a7eb | 3796 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 3797 | if (!initial) { |
a2e7a7eb | 3798 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 3799 | |
a2e7a7eb MG |
3800 | /* |
3801 | * Halve their sleep time's effect, to allow | |
3802 | * for a gentler effect of sleepers: | |
3803 | */ | |
3804 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
3805 | thresh >>= 1; | |
51e0304c | 3806 | |
a2e7a7eb | 3807 | vruntime -= thresh; |
aeb73b04 PZ |
3808 | } |
3809 | ||
b5d9d734 | 3810 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 3811 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
3812 | } |
3813 | ||
d3d9dc33 PT |
3814 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
3815 | ||
cb251765 MG |
3816 | static inline void check_schedstat_required(void) |
3817 | { | |
3818 | #ifdef CONFIG_SCHEDSTATS | |
3819 | if (schedstat_enabled()) | |
3820 | return; | |
3821 | ||
3822 | /* Force schedstat enabled if a dependent tracepoint is active */ | |
3823 | if (trace_sched_stat_wait_enabled() || | |
3824 | trace_sched_stat_sleep_enabled() || | |
3825 | trace_sched_stat_iowait_enabled() || | |
3826 | trace_sched_stat_blocked_enabled() || | |
3827 | trace_sched_stat_runtime_enabled()) { | |
eda8dca5 | 3828 | printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " |
cb251765 | 3829 | "stat_blocked and stat_runtime require the " |
f67abed5 | 3830 | "kernel parameter schedstats=enable or " |
cb251765 MG |
3831 | "kernel.sched_schedstats=1\n"); |
3832 | } | |
3833 | #endif | |
3834 | } | |
3835 | ||
b5179ac7 PZ |
3836 | |
3837 | /* | |
3838 | * MIGRATION | |
3839 | * | |
3840 | * dequeue | |
3841 | * update_curr() | |
3842 | * update_min_vruntime() | |
3843 | * vruntime -= min_vruntime | |
3844 | * | |
3845 | * enqueue | |
3846 | * update_curr() | |
3847 | * update_min_vruntime() | |
3848 | * vruntime += min_vruntime | |
3849 | * | |
3850 | * this way the vruntime transition between RQs is done when both | |
3851 | * min_vruntime are up-to-date. | |
3852 | * | |
3853 | * WAKEUP (remote) | |
3854 | * | |
59efa0ba | 3855 | * ->migrate_task_rq_fair() (p->state == TASK_WAKING) |
b5179ac7 PZ |
3856 | * vruntime -= min_vruntime |
3857 | * | |
3858 | * enqueue | |
3859 | * update_curr() | |
3860 | * update_min_vruntime() | |
3861 | * vruntime += min_vruntime | |
3862 | * | |
3863 | * this way we don't have the most up-to-date min_vruntime on the originating | |
3864 | * CPU and an up-to-date min_vruntime on the destination CPU. | |
3865 | */ | |
3866 | ||
bf0f6f24 | 3867 | static void |
88ec22d3 | 3868 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3869 | { |
2f950354 PZ |
3870 | bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); |
3871 | bool curr = cfs_rq->curr == se; | |
3872 | ||
88ec22d3 | 3873 | /* |
2f950354 PZ |
3874 | * If we're the current task, we must renormalise before calling |
3875 | * update_curr(). | |
88ec22d3 | 3876 | */ |
2f950354 | 3877 | if (renorm && curr) |
88ec22d3 PZ |
3878 | se->vruntime += cfs_rq->min_vruntime; |
3879 | ||
2f950354 PZ |
3880 | update_curr(cfs_rq); |
3881 | ||
bf0f6f24 | 3882 | /* |
2f950354 PZ |
3883 | * Otherwise, renormalise after, such that we're placed at the current |
3884 | * moment in time, instead of some random moment in the past. Being | |
3885 | * placed in the past could significantly boost this task to the | |
3886 | * fairness detriment of existing tasks. | |
bf0f6f24 | 3887 | */ |
2f950354 PZ |
3888 | if (renorm && !curr) |
3889 | se->vruntime += cfs_rq->min_vruntime; | |
3890 | ||
89ee048f VG |
3891 | /* |
3892 | * When enqueuing a sched_entity, we must: | |
3893 | * - Update loads to have both entity and cfs_rq synced with now. | |
3894 | * - Add its load to cfs_rq->runnable_avg | |
3895 | * - For group_entity, update its weight to reflect the new share of | |
3896 | * its group cfs_rq | |
3897 | * - Add its new weight to cfs_rq->load.weight | |
3898 | */ | |
b382a531 | 3899 | update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); |
1ea6c46a | 3900 | update_cfs_group(se); |
b5b3e35f | 3901 | enqueue_runnable_load_avg(cfs_rq, se); |
17bc14b7 | 3902 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 3903 | |
1a3d027c | 3904 | if (flags & ENQUEUE_WAKEUP) |
aeb73b04 | 3905 | place_entity(cfs_rq, se, 0); |
bf0f6f24 | 3906 | |
cb251765 | 3907 | check_schedstat_required(); |
4fa8d299 JP |
3908 | update_stats_enqueue(cfs_rq, se, flags); |
3909 | check_spread(cfs_rq, se); | |
2f950354 | 3910 | if (!curr) |
83b699ed | 3911 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 3912 | se->on_rq = 1; |
3d4b47b4 | 3913 | |
d3d9dc33 | 3914 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 3915 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
3916 | check_enqueue_throttle(cfs_rq); |
3917 | } | |
bf0f6f24 IM |
3918 | } |
3919 | ||
2c13c919 | 3920 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 3921 | { |
2c13c919 RR |
3922 | for_each_sched_entity(se) { |
3923 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3924 | if (cfs_rq->last != se) |
2c13c919 | 3925 | break; |
f1044799 PZ |
3926 | |
3927 | cfs_rq->last = NULL; | |
2c13c919 RR |
3928 | } |
3929 | } | |
2002c695 | 3930 | |
2c13c919 RR |
3931 | static void __clear_buddies_next(struct sched_entity *se) |
3932 | { | |
3933 | for_each_sched_entity(se) { | |
3934 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3935 | if (cfs_rq->next != se) |
2c13c919 | 3936 | break; |
f1044799 PZ |
3937 | |
3938 | cfs_rq->next = NULL; | |
2c13c919 | 3939 | } |
2002c695 PZ |
3940 | } |
3941 | ||
ac53db59 RR |
3942 | static void __clear_buddies_skip(struct sched_entity *se) |
3943 | { | |
3944 | for_each_sched_entity(se) { | |
3945 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3946 | if (cfs_rq->skip != se) |
ac53db59 | 3947 | break; |
f1044799 PZ |
3948 | |
3949 | cfs_rq->skip = NULL; | |
ac53db59 RR |
3950 | } |
3951 | } | |
3952 | ||
a571bbea PZ |
3953 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3954 | { | |
2c13c919 RR |
3955 | if (cfs_rq->last == se) |
3956 | __clear_buddies_last(se); | |
3957 | ||
3958 | if (cfs_rq->next == se) | |
3959 | __clear_buddies_next(se); | |
ac53db59 RR |
3960 | |
3961 | if (cfs_rq->skip == se) | |
3962 | __clear_buddies_skip(se); | |
a571bbea PZ |
3963 | } |
3964 | ||
6c16a6dc | 3965 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 3966 | |
bf0f6f24 | 3967 | static void |
371fd7e7 | 3968 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3969 | { |
a2a2d680 DA |
3970 | /* |
3971 | * Update run-time statistics of the 'current'. | |
3972 | */ | |
3973 | update_curr(cfs_rq); | |
89ee048f VG |
3974 | |
3975 | /* | |
3976 | * When dequeuing a sched_entity, we must: | |
3977 | * - Update loads to have both entity and cfs_rq synced with now. | |
dfcb245e IM |
3978 | * - Subtract its load from the cfs_rq->runnable_avg. |
3979 | * - Subtract its previous weight from cfs_rq->load.weight. | |
89ee048f VG |
3980 | * - For group entity, update its weight to reflect the new share |
3981 | * of its group cfs_rq. | |
3982 | */ | |
88c0616e | 3983 | update_load_avg(cfs_rq, se, UPDATE_TG); |
b5b3e35f | 3984 | dequeue_runnable_load_avg(cfs_rq, se); |
a2a2d680 | 3985 | |
4fa8d299 | 3986 | update_stats_dequeue(cfs_rq, se, flags); |
67e9fb2a | 3987 | |
2002c695 | 3988 | clear_buddies(cfs_rq, se); |
4793241b | 3989 | |
83b699ed | 3990 | if (se != cfs_rq->curr) |
30cfdcfc | 3991 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 3992 | se->on_rq = 0; |
30cfdcfc | 3993 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
3994 | |
3995 | /* | |
b60205c7 PZ |
3996 | * Normalize after update_curr(); which will also have moved |
3997 | * min_vruntime if @se is the one holding it back. But before doing | |
3998 | * update_min_vruntime() again, which will discount @se's position and | |
3999 | * can move min_vruntime forward still more. | |
88ec22d3 | 4000 | */ |
371fd7e7 | 4001 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 4002 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 4003 | |
d8b4986d PT |
4004 | /* return excess runtime on last dequeue */ |
4005 | return_cfs_rq_runtime(cfs_rq); | |
4006 | ||
1ea6c46a | 4007 | update_cfs_group(se); |
b60205c7 PZ |
4008 | |
4009 | /* | |
4010 | * Now advance min_vruntime if @se was the entity holding it back, | |
4011 | * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be | |
4012 | * put back on, and if we advance min_vruntime, we'll be placed back | |
4013 | * further than we started -- ie. we'll be penalized. | |
4014 | */ | |
9845c49c | 4015 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE) |
b60205c7 | 4016 | update_min_vruntime(cfs_rq); |
bf0f6f24 IM |
4017 | } |
4018 | ||
4019 | /* | |
4020 | * Preempt the current task with a newly woken task if needed: | |
4021 | */ | |
7c92e54f | 4022 | static void |
2e09bf55 | 4023 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 4024 | { |
11697830 | 4025 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
4026 | struct sched_entity *se; |
4027 | s64 delta; | |
11697830 | 4028 | |
6d0f0ebd | 4029 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 4030 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 4031 | if (delta_exec > ideal_runtime) { |
8875125e | 4032 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
4033 | /* |
4034 | * The current task ran long enough, ensure it doesn't get | |
4035 | * re-elected due to buddy favours. | |
4036 | */ | |
4037 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
4038 | return; |
4039 | } | |
4040 | ||
4041 | /* | |
4042 | * Ensure that a task that missed wakeup preemption by a | |
4043 | * narrow margin doesn't have to wait for a full slice. | |
4044 | * This also mitigates buddy induced latencies under load. | |
4045 | */ | |
f685ceac MG |
4046 | if (delta_exec < sysctl_sched_min_granularity) |
4047 | return; | |
4048 | ||
f4cfb33e WX |
4049 | se = __pick_first_entity(cfs_rq); |
4050 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 4051 | |
f4cfb33e WX |
4052 | if (delta < 0) |
4053 | return; | |
d7d82944 | 4054 | |
f4cfb33e | 4055 | if (delta > ideal_runtime) |
8875125e | 4056 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
4057 | } |
4058 | ||
83b699ed | 4059 | static void |
8494f412 | 4060 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 4061 | { |
83b699ed SV |
4062 | /* 'current' is not kept within the tree. */ |
4063 | if (se->on_rq) { | |
4064 | /* | |
4065 | * Any task has to be enqueued before it get to execute on | |
4066 | * a CPU. So account for the time it spent waiting on the | |
4067 | * runqueue. | |
4068 | */ | |
4fa8d299 | 4069 | update_stats_wait_end(cfs_rq, se); |
83b699ed | 4070 | __dequeue_entity(cfs_rq, se); |
88c0616e | 4071 | update_load_avg(cfs_rq, se, UPDATE_TG); |
83b699ed SV |
4072 | } |
4073 | ||
79303e9e | 4074 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 4075 | cfs_rq->curr = se; |
4fa8d299 | 4076 | |
eba1ed4b IM |
4077 | /* |
4078 | * Track our maximum slice length, if the CPU's load is at | |
4079 | * least twice that of our own weight (i.e. dont track it | |
4080 | * when there are only lesser-weight tasks around): | |
4081 | */ | |
cb251765 | 4082 | if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
4fa8d299 JP |
4083 | schedstat_set(se->statistics.slice_max, |
4084 | max((u64)schedstat_val(se->statistics.slice_max), | |
4085 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); | |
eba1ed4b | 4086 | } |
4fa8d299 | 4087 | |
4a55b450 | 4088 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
4089 | } |
4090 | ||
3f3a4904 PZ |
4091 | static int |
4092 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
4093 | ||
ac53db59 RR |
4094 | /* |
4095 | * Pick the next process, keeping these things in mind, in this order: | |
4096 | * 1) keep things fair between processes/task groups | |
4097 | * 2) pick the "next" process, since someone really wants that to run | |
4098 | * 3) pick the "last" process, for cache locality | |
4099 | * 4) do not run the "skip" process, if something else is available | |
4100 | */ | |
678d5718 PZ |
4101 | static struct sched_entity * |
4102 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 4103 | { |
678d5718 PZ |
4104 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
4105 | struct sched_entity *se; | |
4106 | ||
4107 | /* | |
4108 | * If curr is set we have to see if its left of the leftmost entity | |
4109 | * still in the tree, provided there was anything in the tree at all. | |
4110 | */ | |
4111 | if (!left || (curr && entity_before(curr, left))) | |
4112 | left = curr; | |
4113 | ||
4114 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 4115 | |
ac53db59 RR |
4116 | /* |
4117 | * Avoid running the skip buddy, if running something else can | |
4118 | * be done without getting too unfair. | |
4119 | */ | |
4120 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
4121 | struct sched_entity *second; |
4122 | ||
4123 | if (se == curr) { | |
4124 | second = __pick_first_entity(cfs_rq); | |
4125 | } else { | |
4126 | second = __pick_next_entity(se); | |
4127 | if (!second || (curr && entity_before(curr, second))) | |
4128 | second = curr; | |
4129 | } | |
4130 | ||
ac53db59 RR |
4131 | if (second && wakeup_preempt_entity(second, left) < 1) |
4132 | se = second; | |
4133 | } | |
aa2ac252 | 4134 | |
f685ceac MG |
4135 | /* |
4136 | * Prefer last buddy, try to return the CPU to a preempted task. | |
4137 | */ | |
4138 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
4139 | se = cfs_rq->last; | |
4140 | ||
ac53db59 RR |
4141 | /* |
4142 | * Someone really wants this to run. If it's not unfair, run it. | |
4143 | */ | |
4144 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
4145 | se = cfs_rq->next; | |
4146 | ||
f685ceac | 4147 | clear_buddies(cfs_rq, se); |
4793241b PZ |
4148 | |
4149 | return se; | |
aa2ac252 PZ |
4150 | } |
4151 | ||
678d5718 | 4152 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 4153 | |
ab6cde26 | 4154 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
4155 | { |
4156 | /* | |
4157 | * If still on the runqueue then deactivate_task() | |
4158 | * was not called and update_curr() has to be done: | |
4159 | */ | |
4160 | if (prev->on_rq) | |
b7cc0896 | 4161 | update_curr(cfs_rq); |
bf0f6f24 | 4162 | |
d3d9dc33 PT |
4163 | /* throttle cfs_rqs exceeding runtime */ |
4164 | check_cfs_rq_runtime(cfs_rq); | |
4165 | ||
4fa8d299 | 4166 | check_spread(cfs_rq, prev); |
cb251765 | 4167 | |
30cfdcfc | 4168 | if (prev->on_rq) { |
4fa8d299 | 4169 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
4170 | /* Put 'current' back into the tree. */ |
4171 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 4172 | /* in !on_rq case, update occurred at dequeue */ |
88c0616e | 4173 | update_load_avg(cfs_rq, prev, 0); |
30cfdcfc | 4174 | } |
429d43bc | 4175 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
4176 | } |
4177 | ||
8f4d37ec PZ |
4178 | static void |
4179 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 4180 | { |
bf0f6f24 | 4181 | /* |
30cfdcfc | 4182 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 4183 | */ |
30cfdcfc | 4184 | update_curr(cfs_rq); |
bf0f6f24 | 4185 | |
9d85f21c PT |
4186 | /* |
4187 | * Ensure that runnable average is periodically updated. | |
4188 | */ | |
88c0616e | 4189 | update_load_avg(cfs_rq, curr, UPDATE_TG); |
1ea6c46a | 4190 | update_cfs_group(curr); |
9d85f21c | 4191 | |
8f4d37ec PZ |
4192 | #ifdef CONFIG_SCHED_HRTICK |
4193 | /* | |
4194 | * queued ticks are scheduled to match the slice, so don't bother | |
4195 | * validating it and just reschedule. | |
4196 | */ | |
983ed7a6 | 4197 | if (queued) { |
8875125e | 4198 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
4199 | return; |
4200 | } | |
8f4d37ec PZ |
4201 | /* |
4202 | * don't let the period tick interfere with the hrtick preemption | |
4203 | */ | |
4204 | if (!sched_feat(DOUBLE_TICK) && | |
4205 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
4206 | return; | |
4207 | #endif | |
4208 | ||
2c2efaed | 4209 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 4210 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
4211 | } |
4212 | ||
ab84d31e PT |
4213 | |
4214 | /************************************************** | |
4215 | * CFS bandwidth control machinery | |
4216 | */ | |
4217 | ||
4218 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb | 4219 | |
e9666d10 | 4220 | #ifdef CONFIG_JUMP_LABEL |
c5905afb | 4221 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
4222 | |
4223 | static inline bool cfs_bandwidth_used(void) | |
4224 | { | |
c5905afb | 4225 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
4226 | } |
4227 | ||
1ee14e6c | 4228 | void cfs_bandwidth_usage_inc(void) |
029632fb | 4229 | { |
ce48c146 | 4230 | static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used); |
1ee14e6c BS |
4231 | } |
4232 | ||
4233 | void cfs_bandwidth_usage_dec(void) | |
4234 | { | |
ce48c146 | 4235 | static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used); |
029632fb | 4236 | } |
e9666d10 | 4237 | #else /* CONFIG_JUMP_LABEL */ |
029632fb PZ |
4238 | static bool cfs_bandwidth_used(void) |
4239 | { | |
4240 | return true; | |
4241 | } | |
4242 | ||
1ee14e6c BS |
4243 | void cfs_bandwidth_usage_inc(void) {} |
4244 | void cfs_bandwidth_usage_dec(void) {} | |
e9666d10 | 4245 | #endif /* CONFIG_JUMP_LABEL */ |
029632fb | 4246 | |
ab84d31e PT |
4247 | /* |
4248 | * default period for cfs group bandwidth. | |
4249 | * default: 0.1s, units: nanoseconds | |
4250 | */ | |
4251 | static inline u64 default_cfs_period(void) | |
4252 | { | |
4253 | return 100000000ULL; | |
4254 | } | |
ec12cb7f PT |
4255 | |
4256 | static inline u64 sched_cfs_bandwidth_slice(void) | |
4257 | { | |
4258 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
4259 | } | |
4260 | ||
a9cf55b2 PT |
4261 | /* |
4262 | * Replenish runtime according to assigned quota and update expiration time. | |
4263 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
4264 | * additional synchronization around rq->lock. | |
4265 | * | |
4266 | * requires cfs_b->lock | |
4267 | */ | |
029632fb | 4268 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
4269 | { |
4270 | u64 now; | |
4271 | ||
4272 | if (cfs_b->quota == RUNTIME_INF) | |
4273 | return; | |
4274 | ||
4275 | now = sched_clock_cpu(smp_processor_id()); | |
4276 | cfs_b->runtime = cfs_b->quota; | |
4277 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
512ac999 | 4278 | cfs_b->expires_seq++; |
a9cf55b2 PT |
4279 | } |
4280 | ||
029632fb PZ |
4281 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4282 | { | |
4283 | return &tg->cfs_bandwidth; | |
4284 | } | |
4285 | ||
f1b17280 PT |
4286 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
4287 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
4288 | { | |
4289 | if (unlikely(cfs_rq->throttle_count)) | |
1a99ae3f | 4290 | return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time; |
f1b17280 | 4291 | |
78becc27 | 4292 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
4293 | } |
4294 | ||
85dac906 PT |
4295 | /* returns 0 on failure to allocate runtime */ |
4296 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
4297 | { |
4298 | struct task_group *tg = cfs_rq->tg; | |
4299 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 4300 | u64 amount = 0, min_amount, expires; |
512ac999 | 4301 | int expires_seq; |
ec12cb7f PT |
4302 | |
4303 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
4304 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
4305 | ||
4306 | raw_spin_lock(&cfs_b->lock); | |
4307 | if (cfs_b->quota == RUNTIME_INF) | |
4308 | amount = min_amount; | |
58088ad0 | 4309 | else { |
77a4d1a1 | 4310 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
4311 | |
4312 | if (cfs_b->runtime > 0) { | |
4313 | amount = min(cfs_b->runtime, min_amount); | |
4314 | cfs_b->runtime -= amount; | |
4315 | cfs_b->idle = 0; | |
4316 | } | |
ec12cb7f | 4317 | } |
512ac999 | 4318 | expires_seq = cfs_b->expires_seq; |
a9cf55b2 | 4319 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
4320 | raw_spin_unlock(&cfs_b->lock); |
4321 | ||
4322 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
4323 | /* |
4324 | * we may have advanced our local expiration to account for allowed | |
4325 | * spread between our sched_clock and the one on which runtime was | |
4326 | * issued. | |
4327 | */ | |
512ac999 XP |
4328 | if (cfs_rq->expires_seq != expires_seq) { |
4329 | cfs_rq->expires_seq = expires_seq; | |
a9cf55b2 | 4330 | cfs_rq->runtime_expires = expires; |
512ac999 | 4331 | } |
85dac906 PT |
4332 | |
4333 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
4334 | } |
4335 | ||
a9cf55b2 PT |
4336 | /* |
4337 | * Note: This depends on the synchronization provided by sched_clock and the | |
4338 | * fact that rq->clock snapshots this value. | |
4339 | */ | |
4340 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 4341 | { |
a9cf55b2 | 4342 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
4343 | |
4344 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 4345 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
4346 | return; |
4347 | ||
a9cf55b2 PT |
4348 | if (cfs_rq->runtime_remaining < 0) |
4349 | return; | |
4350 | ||
4351 | /* | |
4352 | * If the local deadline has passed we have to consider the | |
4353 | * possibility that our sched_clock is 'fast' and the global deadline | |
4354 | * has not truly expired. | |
4355 | * | |
4356 | * Fortunately we can check determine whether this the case by checking | |
512ac999 | 4357 | * whether the global deadline(cfs_b->expires_seq) has advanced. |
a9cf55b2 | 4358 | */ |
512ac999 | 4359 | if (cfs_rq->expires_seq == cfs_b->expires_seq) { |
a9cf55b2 PT |
4360 | /* extend local deadline, drift is bounded above by 2 ticks */ |
4361 | cfs_rq->runtime_expires += TICK_NSEC; | |
4362 | } else { | |
4363 | /* global deadline is ahead, expiration has passed */ | |
4364 | cfs_rq->runtime_remaining = 0; | |
4365 | } | |
4366 | } | |
4367 | ||
9dbdb155 | 4368 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
4369 | { |
4370 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 4371 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
4372 | expire_cfs_rq_runtime(cfs_rq); |
4373 | ||
4374 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
4375 | return; |
4376 | ||
85dac906 PT |
4377 | /* |
4378 | * if we're unable to extend our runtime we resched so that the active | |
4379 | * hierarchy can be throttled | |
4380 | */ | |
4381 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 4382 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
4383 | } |
4384 | ||
6c16a6dc | 4385 | static __always_inline |
9dbdb155 | 4386 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 4387 | { |
56f570e5 | 4388 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
4389 | return; |
4390 | ||
4391 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
4392 | } | |
4393 | ||
85dac906 PT |
4394 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
4395 | { | |
56f570e5 | 4396 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
4397 | } |
4398 | ||
64660c86 PT |
4399 | /* check whether cfs_rq, or any parent, is throttled */ |
4400 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4401 | { | |
56f570e5 | 4402 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
4403 | } |
4404 | ||
4405 | /* | |
4406 | * Ensure that neither of the group entities corresponding to src_cpu or | |
4407 | * dest_cpu are members of a throttled hierarchy when performing group | |
4408 | * load-balance operations. | |
4409 | */ | |
4410 | static inline int throttled_lb_pair(struct task_group *tg, | |
4411 | int src_cpu, int dest_cpu) | |
4412 | { | |
4413 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
4414 | ||
4415 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
4416 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
4417 | ||
4418 | return throttled_hierarchy(src_cfs_rq) || | |
4419 | throttled_hierarchy(dest_cfs_rq); | |
4420 | } | |
4421 | ||
64660c86 PT |
4422 | static int tg_unthrottle_up(struct task_group *tg, void *data) |
4423 | { | |
4424 | struct rq *rq = data; | |
4425 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4426 | ||
4427 | cfs_rq->throttle_count--; | |
64660c86 | 4428 | if (!cfs_rq->throttle_count) { |
f1b17280 | 4429 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 4430 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 4431 | cfs_rq->throttled_clock_task; |
64660c86 | 4432 | } |
64660c86 PT |
4433 | |
4434 | return 0; | |
4435 | } | |
4436 | ||
4437 | static int tg_throttle_down(struct task_group *tg, void *data) | |
4438 | { | |
4439 | struct rq *rq = data; | |
4440 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4441 | ||
82958366 PT |
4442 | /* group is entering throttled state, stop time */ |
4443 | if (!cfs_rq->throttle_count) | |
78becc27 | 4444 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
4445 | cfs_rq->throttle_count++; |
4446 | ||
4447 | return 0; | |
4448 | } | |
4449 | ||
d3d9dc33 | 4450 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
4451 | { |
4452 | struct rq *rq = rq_of(cfs_rq); | |
4453 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4454 | struct sched_entity *se; | |
4455 | long task_delta, dequeue = 1; | |
77a4d1a1 | 4456 | bool empty; |
85dac906 PT |
4457 | |
4458 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
4459 | ||
f1b17280 | 4460 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
4461 | rcu_read_lock(); |
4462 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
4463 | rcu_read_unlock(); | |
85dac906 PT |
4464 | |
4465 | task_delta = cfs_rq->h_nr_running; | |
4466 | for_each_sched_entity(se) { | |
4467 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
4468 | /* throttled entity or throttle-on-deactivate */ | |
4469 | if (!se->on_rq) | |
4470 | break; | |
4471 | ||
4472 | if (dequeue) | |
4473 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
4474 | qcfs_rq->h_nr_running -= task_delta; | |
4475 | ||
4476 | if (qcfs_rq->load.weight) | |
4477 | dequeue = 0; | |
4478 | } | |
4479 | ||
4480 | if (!se) | |
72465447 | 4481 | sub_nr_running(rq, task_delta); |
85dac906 PT |
4482 | |
4483 | cfs_rq->throttled = 1; | |
78becc27 | 4484 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 | 4485 | raw_spin_lock(&cfs_b->lock); |
d49db342 | 4486 | empty = list_empty(&cfs_b->throttled_cfs_rq); |
77a4d1a1 | 4487 | |
c06f04c7 BS |
4488 | /* |
4489 | * Add to the _head_ of the list, so that an already-started | |
baa9be4f PA |
4490 | * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is |
4491 | * not running add to the tail so that later runqueues don't get starved. | |
c06f04c7 | 4492 | */ |
baa9be4f PA |
4493 | if (cfs_b->distribute_running) |
4494 | list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
4495 | else | |
4496 | list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
77a4d1a1 PZ |
4497 | |
4498 | /* | |
4499 | * If we're the first throttled task, make sure the bandwidth | |
4500 | * timer is running. | |
4501 | */ | |
4502 | if (empty) | |
4503 | start_cfs_bandwidth(cfs_b); | |
4504 | ||
85dac906 PT |
4505 | raw_spin_unlock(&cfs_b->lock); |
4506 | } | |
4507 | ||
029632fb | 4508 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
4509 | { |
4510 | struct rq *rq = rq_of(cfs_rq); | |
4511 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4512 | struct sched_entity *se; | |
4513 | int enqueue = 1; | |
4514 | long task_delta; | |
4515 | ||
22b958d8 | 4516 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
4517 | |
4518 | cfs_rq->throttled = 0; | |
1a55af2e FW |
4519 | |
4520 | update_rq_clock(rq); | |
4521 | ||
671fd9da | 4522 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 4523 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
4524 | list_del_rcu(&cfs_rq->throttled_list); |
4525 | raw_spin_unlock(&cfs_b->lock); | |
4526 | ||
64660c86 PT |
4527 | /* update hierarchical throttle state */ |
4528 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
4529 | ||
671fd9da PT |
4530 | if (!cfs_rq->load.weight) |
4531 | return; | |
4532 | ||
4533 | task_delta = cfs_rq->h_nr_running; | |
4534 | for_each_sched_entity(se) { | |
4535 | if (se->on_rq) | |
4536 | enqueue = 0; | |
4537 | ||
4538 | cfs_rq = cfs_rq_of(se); | |
4539 | if (enqueue) | |
4540 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
4541 | cfs_rq->h_nr_running += task_delta; | |
4542 | ||
4543 | if (cfs_rq_throttled(cfs_rq)) | |
4544 | break; | |
4545 | } | |
4546 | ||
4547 | if (!se) | |
72465447 | 4548 | add_nr_running(rq, task_delta); |
671fd9da | 4549 | |
97fb7a0a | 4550 | /* Determine whether we need to wake up potentially idle CPU: */ |
671fd9da | 4551 | if (rq->curr == rq->idle && rq->cfs.nr_running) |
8875125e | 4552 | resched_curr(rq); |
671fd9da PT |
4553 | } |
4554 | ||
4555 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
4556 | u64 remaining, u64 expires) | |
4557 | { | |
4558 | struct cfs_rq *cfs_rq; | |
c06f04c7 BS |
4559 | u64 runtime; |
4560 | u64 starting_runtime = remaining; | |
671fd9da PT |
4561 | |
4562 | rcu_read_lock(); | |
4563 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
4564 | throttled_list) { | |
4565 | struct rq *rq = rq_of(cfs_rq); | |
8a8c69c3 | 4566 | struct rq_flags rf; |
671fd9da | 4567 | |
c0ad4aa4 | 4568 | rq_lock_irqsave(rq, &rf); |
671fd9da PT |
4569 | if (!cfs_rq_throttled(cfs_rq)) |
4570 | goto next; | |
4571 | ||
4572 | runtime = -cfs_rq->runtime_remaining + 1; | |
4573 | if (runtime > remaining) | |
4574 | runtime = remaining; | |
4575 | remaining -= runtime; | |
4576 | ||
4577 | cfs_rq->runtime_remaining += runtime; | |
4578 | cfs_rq->runtime_expires = expires; | |
4579 | ||
4580 | /* we check whether we're throttled above */ | |
4581 | if (cfs_rq->runtime_remaining > 0) | |
4582 | unthrottle_cfs_rq(cfs_rq); | |
4583 | ||
4584 | next: | |
c0ad4aa4 | 4585 | rq_unlock_irqrestore(rq, &rf); |
671fd9da PT |
4586 | |
4587 | if (!remaining) | |
4588 | break; | |
4589 | } | |
4590 | rcu_read_unlock(); | |
4591 | ||
c06f04c7 | 4592 | return starting_runtime - remaining; |
671fd9da PT |
4593 | } |
4594 | ||
58088ad0 PT |
4595 | /* |
4596 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
4597 | * cfs_rqs as appropriate. If there has been no activity within the last | |
4598 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
4599 | * used to track this state. | |
4600 | */ | |
c0ad4aa4 | 4601 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) |
58088ad0 | 4602 | { |
671fd9da | 4603 | u64 runtime, runtime_expires; |
51f2176d | 4604 | int throttled; |
58088ad0 | 4605 | |
58088ad0 PT |
4606 | /* no need to continue the timer with no bandwidth constraint */ |
4607 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 4608 | goto out_deactivate; |
58088ad0 | 4609 | |
671fd9da | 4610 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 4611 | cfs_b->nr_periods += overrun; |
671fd9da | 4612 | |
51f2176d BS |
4613 | /* |
4614 | * idle depends on !throttled (for the case of a large deficit), and if | |
4615 | * we're going inactive then everything else can be deferred | |
4616 | */ | |
4617 | if (cfs_b->idle && !throttled) | |
4618 | goto out_deactivate; | |
a9cf55b2 PT |
4619 | |
4620 | __refill_cfs_bandwidth_runtime(cfs_b); | |
4621 | ||
671fd9da PT |
4622 | if (!throttled) { |
4623 | /* mark as potentially idle for the upcoming period */ | |
4624 | cfs_b->idle = 1; | |
51f2176d | 4625 | return 0; |
671fd9da PT |
4626 | } |
4627 | ||
e8da1b18 NR |
4628 | /* account preceding periods in which throttling occurred */ |
4629 | cfs_b->nr_throttled += overrun; | |
4630 | ||
671fd9da | 4631 | runtime_expires = cfs_b->runtime_expires; |
671fd9da PT |
4632 | |
4633 | /* | |
c06f04c7 BS |
4634 | * This check is repeated as we are holding onto the new bandwidth while |
4635 | * we unthrottle. This can potentially race with an unthrottled group | |
4636 | * trying to acquire new bandwidth from the global pool. This can result | |
4637 | * in us over-using our runtime if it is all used during this loop, but | |
4638 | * only by limited amounts in that extreme case. | |
671fd9da | 4639 | */ |
baa9be4f | 4640 | while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) { |
c06f04c7 | 4641 | runtime = cfs_b->runtime; |
baa9be4f | 4642 | cfs_b->distribute_running = 1; |
c0ad4aa4 | 4643 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
671fd9da PT |
4644 | /* we can't nest cfs_b->lock while distributing bandwidth */ |
4645 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
4646 | runtime_expires); | |
c0ad4aa4 | 4647 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
671fd9da | 4648 | |
baa9be4f | 4649 | cfs_b->distribute_running = 0; |
671fd9da | 4650 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
c06f04c7 | 4651 | |
b5c0ce7b | 4652 | lsub_positive(&cfs_b->runtime, runtime); |
671fd9da | 4653 | } |
58088ad0 | 4654 | |
671fd9da PT |
4655 | /* |
4656 | * While we are ensured activity in the period following an | |
4657 | * unthrottle, this also covers the case in which the new bandwidth is | |
4658 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
4659 | * timer to remain active while there are any throttled entities.) | |
4660 | */ | |
4661 | cfs_b->idle = 0; | |
58088ad0 | 4662 | |
51f2176d BS |
4663 | return 0; |
4664 | ||
4665 | out_deactivate: | |
51f2176d | 4666 | return 1; |
58088ad0 | 4667 | } |
d3d9dc33 | 4668 | |
d8b4986d PT |
4669 | /* a cfs_rq won't donate quota below this amount */ |
4670 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
4671 | /* minimum remaining period time to redistribute slack quota */ | |
4672 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
4673 | /* how long we wait to gather additional slack before distributing */ | |
4674 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
4675 | ||
db06e78c BS |
4676 | /* |
4677 | * Are we near the end of the current quota period? | |
4678 | * | |
4679 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 4680 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
4681 | * migrate_hrtimers, base is never cleared, so we are fine. |
4682 | */ | |
d8b4986d PT |
4683 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
4684 | { | |
4685 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
4686 | u64 remaining; | |
4687 | ||
4688 | /* if the call-back is running a quota refresh is already occurring */ | |
4689 | if (hrtimer_callback_running(refresh_timer)) | |
4690 | return 1; | |
4691 | ||
4692 | /* is a quota refresh about to occur? */ | |
4693 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
4694 | if (remaining < min_expire) | |
4695 | return 1; | |
4696 | ||
4697 | return 0; | |
4698 | } | |
4699 | ||
4700 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
4701 | { | |
4702 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
4703 | ||
4704 | /* if there's a quota refresh soon don't bother with slack */ | |
4705 | if (runtime_refresh_within(cfs_b, min_left)) | |
4706 | return; | |
4707 | ||
4cfafd30 PZ |
4708 | hrtimer_start(&cfs_b->slack_timer, |
4709 | ns_to_ktime(cfs_bandwidth_slack_period), | |
4710 | HRTIMER_MODE_REL); | |
d8b4986d PT |
4711 | } |
4712 | ||
4713 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
4714 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4715 | { | |
4716 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4717 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
4718 | ||
4719 | if (slack_runtime <= 0) | |
4720 | return; | |
4721 | ||
4722 | raw_spin_lock(&cfs_b->lock); | |
4723 | if (cfs_b->quota != RUNTIME_INF && | |
4724 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
4725 | cfs_b->runtime += slack_runtime; | |
4726 | ||
4727 | /* we are under rq->lock, defer unthrottling using a timer */ | |
4728 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
4729 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
4730 | start_cfs_slack_bandwidth(cfs_b); | |
4731 | } | |
4732 | raw_spin_unlock(&cfs_b->lock); | |
4733 | ||
4734 | /* even if it's not valid for return we don't want to try again */ | |
4735 | cfs_rq->runtime_remaining -= slack_runtime; | |
4736 | } | |
4737 | ||
4738 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4739 | { | |
56f570e5 PT |
4740 | if (!cfs_bandwidth_used()) |
4741 | return; | |
4742 | ||
fccfdc6f | 4743 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
4744 | return; |
4745 | ||
4746 | __return_cfs_rq_runtime(cfs_rq); | |
4747 | } | |
4748 | ||
4749 | /* | |
4750 | * This is done with a timer (instead of inline with bandwidth return) since | |
4751 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
4752 | */ | |
4753 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
4754 | { | |
4755 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
c0ad4aa4 | 4756 | unsigned long flags; |
d8b4986d PT |
4757 | u64 expires; |
4758 | ||
4759 | /* confirm we're still not at a refresh boundary */ | |
c0ad4aa4 | 4760 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
baa9be4f | 4761 | if (cfs_b->distribute_running) { |
c0ad4aa4 | 4762 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
baa9be4f PA |
4763 | return; |
4764 | } | |
4765 | ||
db06e78c | 4766 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { |
c0ad4aa4 | 4767 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d | 4768 | return; |
db06e78c | 4769 | } |
d8b4986d | 4770 | |
c06f04c7 | 4771 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 4772 | runtime = cfs_b->runtime; |
c06f04c7 | 4773 | |
d8b4986d | 4774 | expires = cfs_b->runtime_expires; |
baa9be4f PA |
4775 | if (runtime) |
4776 | cfs_b->distribute_running = 1; | |
4777 | ||
c0ad4aa4 | 4778 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
4779 | |
4780 | if (!runtime) | |
4781 | return; | |
4782 | ||
4783 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
4784 | ||
c0ad4aa4 | 4785 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
d8b4986d | 4786 | if (expires == cfs_b->runtime_expires) |
b5c0ce7b | 4787 | lsub_positive(&cfs_b->runtime, runtime); |
baa9be4f | 4788 | cfs_b->distribute_running = 0; |
c0ad4aa4 | 4789 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
d8b4986d PT |
4790 | } |
4791 | ||
d3d9dc33 PT |
4792 | /* |
4793 | * When a group wakes up we want to make sure that its quota is not already | |
4794 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
4795 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
4796 | */ | |
4797 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
4798 | { | |
56f570e5 PT |
4799 | if (!cfs_bandwidth_used()) |
4800 | return; | |
4801 | ||
d3d9dc33 PT |
4802 | /* an active group must be handled by the update_curr()->put() path */ |
4803 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
4804 | return; | |
4805 | ||
4806 | /* ensure the group is not already throttled */ | |
4807 | if (cfs_rq_throttled(cfs_rq)) | |
4808 | return; | |
4809 | ||
4810 | /* update runtime allocation */ | |
4811 | account_cfs_rq_runtime(cfs_rq, 0); | |
4812 | if (cfs_rq->runtime_remaining <= 0) | |
4813 | throttle_cfs_rq(cfs_rq); | |
4814 | } | |
4815 | ||
55e16d30 PZ |
4816 | static void sync_throttle(struct task_group *tg, int cpu) |
4817 | { | |
4818 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
4819 | ||
4820 | if (!cfs_bandwidth_used()) | |
4821 | return; | |
4822 | ||
4823 | if (!tg->parent) | |
4824 | return; | |
4825 | ||
4826 | cfs_rq = tg->cfs_rq[cpu]; | |
4827 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
4828 | ||
4829 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
b8922125 | 4830 | cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); |
55e16d30 PZ |
4831 | } |
4832 | ||
d3d9dc33 | 4833 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 4834 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 4835 | { |
56f570e5 | 4836 | if (!cfs_bandwidth_used()) |
678d5718 | 4837 | return false; |
56f570e5 | 4838 | |
d3d9dc33 | 4839 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 4840 | return false; |
d3d9dc33 PT |
4841 | |
4842 | /* | |
4843 | * it's possible for a throttled entity to be forced into a running | |
4844 | * state (e.g. set_curr_task), in this case we're finished. | |
4845 | */ | |
4846 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 4847 | return true; |
d3d9dc33 PT |
4848 | |
4849 | throttle_cfs_rq(cfs_rq); | |
678d5718 | 4850 | return true; |
d3d9dc33 | 4851 | } |
029632fb | 4852 | |
029632fb PZ |
4853 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
4854 | { | |
4855 | struct cfs_bandwidth *cfs_b = | |
4856 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 4857 | |
029632fb PZ |
4858 | do_sched_cfs_slack_timer(cfs_b); |
4859 | ||
4860 | return HRTIMER_NORESTART; | |
4861 | } | |
4862 | ||
4863 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
4864 | { | |
4865 | struct cfs_bandwidth *cfs_b = | |
4866 | container_of(timer, struct cfs_bandwidth, period_timer); | |
c0ad4aa4 | 4867 | unsigned long flags; |
029632fb PZ |
4868 | int overrun; |
4869 | int idle = 0; | |
4870 | ||
c0ad4aa4 | 4871 | raw_spin_lock_irqsave(&cfs_b->lock, flags); |
029632fb | 4872 | for (;;) { |
77a4d1a1 | 4873 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
4874 | if (!overrun) |
4875 | break; | |
4876 | ||
c0ad4aa4 | 4877 | idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); |
029632fb | 4878 | } |
4cfafd30 PZ |
4879 | if (idle) |
4880 | cfs_b->period_active = 0; | |
c0ad4aa4 | 4881 | raw_spin_unlock_irqrestore(&cfs_b->lock, flags); |
029632fb PZ |
4882 | |
4883 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
4884 | } | |
4885 | ||
4886 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4887 | { | |
4888 | raw_spin_lock_init(&cfs_b->lock); | |
4889 | cfs_b->runtime = 0; | |
4890 | cfs_b->quota = RUNTIME_INF; | |
4891 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
4892 | ||
4893 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 4894 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
4895 | cfs_b->period_timer.function = sched_cfs_period_timer; |
4896 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
4897 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
baa9be4f | 4898 | cfs_b->distribute_running = 0; |
029632fb PZ |
4899 | } |
4900 | ||
4901 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4902 | { | |
4903 | cfs_rq->runtime_enabled = 0; | |
4904 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
4905 | } | |
4906 | ||
77a4d1a1 | 4907 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 4908 | { |
f1d1be8a XP |
4909 | u64 overrun; |
4910 | ||
4cfafd30 | 4911 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 4912 | |
f1d1be8a XP |
4913 | if (cfs_b->period_active) |
4914 | return; | |
4915 | ||
4916 | cfs_b->period_active = 1; | |
4917 | overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); | |
4918 | cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period); | |
4919 | cfs_b->expires_seq++; | |
4920 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); | |
029632fb PZ |
4921 | } |
4922 | ||
4923 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4924 | { | |
7f1a169b TH |
4925 | /* init_cfs_bandwidth() was not called */ |
4926 | if (!cfs_b->throttled_cfs_rq.next) | |
4927 | return; | |
4928 | ||
029632fb PZ |
4929 | hrtimer_cancel(&cfs_b->period_timer); |
4930 | hrtimer_cancel(&cfs_b->slack_timer); | |
4931 | } | |
4932 | ||
502ce005 | 4933 | /* |
97fb7a0a | 4934 | * Both these CPU hotplug callbacks race against unregister_fair_sched_group() |
502ce005 PZ |
4935 | * |
4936 | * The race is harmless, since modifying bandwidth settings of unhooked group | |
4937 | * bits doesn't do much. | |
4938 | */ | |
4939 | ||
4940 | /* cpu online calback */ | |
0e59bdae KT |
4941 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
4942 | { | |
502ce005 | 4943 | struct task_group *tg; |
0e59bdae | 4944 | |
502ce005 PZ |
4945 | lockdep_assert_held(&rq->lock); |
4946 | ||
4947 | rcu_read_lock(); | |
4948 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
4949 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
4950 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
0e59bdae KT |
4951 | |
4952 | raw_spin_lock(&cfs_b->lock); | |
4953 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
4954 | raw_spin_unlock(&cfs_b->lock); | |
4955 | } | |
502ce005 | 4956 | rcu_read_unlock(); |
0e59bdae KT |
4957 | } |
4958 | ||
502ce005 | 4959 | /* cpu offline callback */ |
38dc3348 | 4960 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb | 4961 | { |
502ce005 PZ |
4962 | struct task_group *tg; |
4963 | ||
4964 | lockdep_assert_held(&rq->lock); | |
4965 | ||
4966 | rcu_read_lock(); | |
4967 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
4968 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
029632fb | 4969 | |
029632fb PZ |
4970 | if (!cfs_rq->runtime_enabled) |
4971 | continue; | |
4972 | ||
4973 | /* | |
4974 | * clock_task is not advancing so we just need to make sure | |
4975 | * there's some valid quota amount | |
4976 | */ | |
51f2176d | 4977 | cfs_rq->runtime_remaining = 1; |
0e59bdae | 4978 | /* |
97fb7a0a | 4979 | * Offline rq is schedulable till CPU is completely disabled |
0e59bdae KT |
4980 | * in take_cpu_down(), so we prevent new cfs throttling here. |
4981 | */ | |
4982 | cfs_rq->runtime_enabled = 0; | |
4983 | ||
029632fb PZ |
4984 | if (cfs_rq_throttled(cfs_rq)) |
4985 | unthrottle_cfs_rq(cfs_rq); | |
4986 | } | |
502ce005 | 4987 | rcu_read_unlock(); |
029632fb PZ |
4988 | } |
4989 | ||
4990 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
4991 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
4992 | { | |
78becc27 | 4993 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
4994 | } |
4995 | ||
9dbdb155 | 4996 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 4997 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 4998 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 4999 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 5000 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
5001 | |
5002 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
5003 | { | |
5004 | return 0; | |
5005 | } | |
64660c86 PT |
5006 | |
5007 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
5008 | { | |
5009 | return 0; | |
5010 | } | |
5011 | ||
5012 | static inline int throttled_lb_pair(struct task_group *tg, | |
5013 | int src_cpu, int dest_cpu) | |
5014 | { | |
5015 | return 0; | |
5016 | } | |
029632fb PZ |
5017 | |
5018 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
5019 | ||
5020 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5021 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
5022 | #endif |
5023 | ||
029632fb PZ |
5024 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
5025 | { | |
5026 | return NULL; | |
5027 | } | |
5028 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 5029 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 5030 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
5031 | |
5032 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
5033 | ||
bf0f6f24 IM |
5034 | /************************************************** |
5035 | * CFS operations on tasks: | |
5036 | */ | |
5037 | ||
8f4d37ec PZ |
5038 | #ifdef CONFIG_SCHED_HRTICK |
5039 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5040 | { | |
8f4d37ec PZ |
5041 | struct sched_entity *se = &p->se; |
5042 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5043 | ||
9148a3a1 | 5044 | SCHED_WARN_ON(task_rq(p) != rq); |
8f4d37ec | 5045 | |
8bf46a39 | 5046 | if (rq->cfs.h_nr_running > 1) { |
8f4d37ec PZ |
5047 | u64 slice = sched_slice(cfs_rq, se); |
5048 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
5049 | s64 delta = slice - ran; | |
5050 | ||
5051 | if (delta < 0) { | |
5052 | if (rq->curr == p) | |
8875125e | 5053 | resched_curr(rq); |
8f4d37ec PZ |
5054 | return; |
5055 | } | |
31656519 | 5056 | hrtick_start(rq, delta); |
8f4d37ec PZ |
5057 | } |
5058 | } | |
a4c2f00f PZ |
5059 | |
5060 | /* | |
5061 | * called from enqueue/dequeue and updates the hrtick when the | |
5062 | * current task is from our class and nr_running is low enough | |
5063 | * to matter. | |
5064 | */ | |
5065 | static void hrtick_update(struct rq *rq) | |
5066 | { | |
5067 | struct task_struct *curr = rq->curr; | |
5068 | ||
b39e66ea | 5069 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
5070 | return; |
5071 | ||
5072 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
5073 | hrtick_start_fair(rq, curr); | |
5074 | } | |
55e12e5e | 5075 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
5076 | static inline void |
5077 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5078 | { | |
5079 | } | |
a4c2f00f PZ |
5080 | |
5081 | static inline void hrtick_update(struct rq *rq) | |
5082 | { | |
5083 | } | |
8f4d37ec PZ |
5084 | #endif |
5085 | ||
2802bf3c MR |
5086 | #ifdef CONFIG_SMP |
5087 | static inline unsigned long cpu_util(int cpu); | |
5088 | static unsigned long capacity_of(int cpu); | |
5089 | ||
5090 | static inline bool cpu_overutilized(int cpu) | |
5091 | { | |
5092 | return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin); | |
5093 | } | |
5094 | ||
5095 | static inline void update_overutilized_status(struct rq *rq) | |
5096 | { | |
5097 | if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) | |
5098 | WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED); | |
5099 | } | |
5100 | #else | |
5101 | static inline void update_overutilized_status(struct rq *rq) { } | |
5102 | #endif | |
5103 | ||
bf0f6f24 IM |
5104 | /* |
5105 | * The enqueue_task method is called before nr_running is | |
5106 | * increased. Here we update the fair scheduling stats and | |
5107 | * then put the task into the rbtree: | |
5108 | */ | |
ea87bb78 | 5109 | static void |
371fd7e7 | 5110 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5111 | { |
5112 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5113 | struct sched_entity *se = &p->se; |
bf0f6f24 | 5114 | |
2539fc82 PB |
5115 | /* |
5116 | * The code below (indirectly) updates schedutil which looks at | |
5117 | * the cfs_rq utilization to select a frequency. | |
5118 | * Let's add the task's estimated utilization to the cfs_rq's | |
5119 | * estimated utilization, before we update schedutil. | |
5120 | */ | |
5121 | util_est_enqueue(&rq->cfs, p); | |
5122 | ||
8c34ab19 RW |
5123 | /* |
5124 | * If in_iowait is set, the code below may not trigger any cpufreq | |
5125 | * utilization updates, so do it here explicitly with the IOWAIT flag | |
5126 | * passed. | |
5127 | */ | |
5128 | if (p->in_iowait) | |
674e7541 | 5129 | cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); |
8c34ab19 | 5130 | |
bf0f6f24 | 5131 | for_each_sched_entity(se) { |
62fb1851 | 5132 | if (se->on_rq) |
bf0f6f24 IM |
5133 | break; |
5134 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 5135 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
5136 | |
5137 | /* | |
5138 | * end evaluation on encountering a throttled cfs_rq | |
5139 | * | |
5140 | * note: in the case of encountering a throttled cfs_rq we will | |
5141 | * post the final h_nr_running increment below. | |
e210bffd | 5142 | */ |
85dac906 PT |
5143 | if (cfs_rq_throttled(cfs_rq)) |
5144 | break; | |
953bfcd1 | 5145 | cfs_rq->h_nr_running++; |
85dac906 | 5146 | |
88ec22d3 | 5147 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 5148 | } |
8f4d37ec | 5149 | |
2069dd75 | 5150 | for_each_sched_entity(se) { |
0f317143 | 5151 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 5152 | cfs_rq->h_nr_running++; |
2069dd75 | 5153 | |
85dac906 PT |
5154 | if (cfs_rq_throttled(cfs_rq)) |
5155 | break; | |
5156 | ||
88c0616e | 5157 | update_load_avg(cfs_rq, se, UPDATE_TG); |
1ea6c46a | 5158 | update_cfs_group(se); |
2069dd75 PZ |
5159 | } |
5160 | ||
2802bf3c | 5161 | if (!se) { |
72465447 | 5162 | add_nr_running(rq, 1); |
2802bf3c MR |
5163 | /* |
5164 | * Since new tasks are assigned an initial util_avg equal to | |
5165 | * half of the spare capacity of their CPU, tiny tasks have the | |
5166 | * ability to cross the overutilized threshold, which will | |
5167 | * result in the load balancer ruining all the task placement | |
5168 | * done by EAS. As a way to mitigate that effect, do not account | |
5169 | * for the first enqueue operation of new tasks during the | |
5170 | * overutilized flag detection. | |
5171 | * | |
5172 | * A better way of solving this problem would be to wait for | |
5173 | * the PELT signals of tasks to converge before taking them | |
5174 | * into account, but that is not straightforward to implement, | |
5175 | * and the following generally works well enough in practice. | |
5176 | */ | |
5177 | if (flags & ENQUEUE_WAKEUP) | |
5178 | update_overutilized_status(rq); | |
5179 | ||
5180 | } | |
cd126afe | 5181 | |
a4c2f00f | 5182 | hrtick_update(rq); |
bf0f6f24 IM |
5183 | } |
5184 | ||
2f36825b VP |
5185 | static void set_next_buddy(struct sched_entity *se); |
5186 | ||
bf0f6f24 IM |
5187 | /* |
5188 | * The dequeue_task method is called before nr_running is | |
5189 | * decreased. We remove the task from the rbtree and | |
5190 | * update the fair scheduling stats: | |
5191 | */ | |
371fd7e7 | 5192 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5193 | { |
5194 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5195 | struct sched_entity *se = &p->se; |
2f36825b | 5196 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
5197 | |
5198 | for_each_sched_entity(se) { | |
5199 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 5200 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
5201 | |
5202 | /* | |
5203 | * end evaluation on encountering a throttled cfs_rq | |
5204 | * | |
5205 | * note: in the case of encountering a throttled cfs_rq we will | |
5206 | * post the final h_nr_running decrement below. | |
5207 | */ | |
5208 | if (cfs_rq_throttled(cfs_rq)) | |
5209 | break; | |
953bfcd1 | 5210 | cfs_rq->h_nr_running--; |
2069dd75 | 5211 | |
bf0f6f24 | 5212 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 5213 | if (cfs_rq->load.weight) { |
754bd598 KK |
5214 | /* Avoid re-evaluating load for this entity: */ |
5215 | se = parent_entity(se); | |
2f36825b VP |
5216 | /* |
5217 | * Bias pick_next to pick a task from this cfs_rq, as | |
5218 | * p is sleeping when it is within its sched_slice. | |
5219 | */ | |
754bd598 KK |
5220 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
5221 | set_next_buddy(se); | |
bf0f6f24 | 5222 | break; |
2f36825b | 5223 | } |
371fd7e7 | 5224 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 5225 | } |
8f4d37ec | 5226 | |
2069dd75 | 5227 | for_each_sched_entity(se) { |
0f317143 | 5228 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 5229 | cfs_rq->h_nr_running--; |
2069dd75 | 5230 | |
85dac906 PT |
5231 | if (cfs_rq_throttled(cfs_rq)) |
5232 | break; | |
5233 | ||
88c0616e | 5234 | update_load_avg(cfs_rq, se, UPDATE_TG); |
1ea6c46a | 5235 | update_cfs_group(se); |
2069dd75 PZ |
5236 | } |
5237 | ||
cd126afe | 5238 | if (!se) |
72465447 | 5239 | sub_nr_running(rq, 1); |
cd126afe | 5240 | |
7f65ea42 | 5241 | util_est_dequeue(&rq->cfs, p, task_sleep); |
a4c2f00f | 5242 | hrtick_update(rq); |
bf0f6f24 IM |
5243 | } |
5244 | ||
e7693a36 | 5245 | #ifdef CONFIG_SMP |
10e2f1ac PZ |
5246 | |
5247 | /* Working cpumask for: load_balance, load_balance_newidle. */ | |
5248 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); | |
5249 | DEFINE_PER_CPU(cpumask_var_t, select_idle_mask); | |
5250 | ||
9fd81dd5 | 5251 | #ifdef CONFIG_NO_HZ_COMMON |
3289bdb4 PZ |
5252 | /* |
5253 | * per rq 'load' arrray crap; XXX kill this. | |
5254 | */ | |
5255 | ||
5256 | /* | |
d937cdc5 | 5257 | * The exact cpuload calculated at every tick would be: |
3289bdb4 | 5258 | * |
d937cdc5 PZ |
5259 | * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load |
5260 | * | |
97fb7a0a IM |
5261 | * If a CPU misses updates for n ticks (as it was idle) and update gets |
5262 | * called on the n+1-th tick when CPU may be busy, then we have: | |
d937cdc5 PZ |
5263 | * |
5264 | * load_n = (1 - 1/2^i)^n * load_0 | |
5265 | * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load | |
3289bdb4 PZ |
5266 | * |
5267 | * decay_load_missed() below does efficient calculation of | |
3289bdb4 | 5268 | * |
d937cdc5 PZ |
5269 | * load' = (1 - 1/2^i)^n * load |
5270 | * | |
5271 | * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors. | |
5272 | * This allows us to precompute the above in said factors, thereby allowing the | |
5273 | * reduction of an arbitrary n in O(log_2 n) steps. (See also | |
5274 | * fixed_power_int()) | |
3289bdb4 | 5275 | * |
d937cdc5 | 5276 | * The calculation is approximated on a 128 point scale. |
3289bdb4 PZ |
5277 | */ |
5278 | #define DEGRADE_SHIFT 7 | |
d937cdc5 PZ |
5279 | |
5280 | static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; | |
5281 | static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { | |
5282 | { 0, 0, 0, 0, 0, 0, 0, 0 }, | |
5283 | { 64, 32, 8, 0, 0, 0, 0, 0 }, | |
5284 | { 96, 72, 40, 12, 1, 0, 0, 0 }, | |
5285 | { 112, 98, 75, 43, 15, 1, 0, 0 }, | |
5286 | { 120, 112, 98, 76, 45, 16, 2, 0 } | |
5287 | }; | |
3289bdb4 PZ |
5288 | |
5289 | /* | |
5290 | * Update cpu_load for any missed ticks, due to tickless idle. The backlog | |
5291 | * would be when CPU is idle and so we just decay the old load without | |
5292 | * adding any new load. | |
5293 | */ | |
5294 | static unsigned long | |
5295 | decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) | |
5296 | { | |
5297 | int j = 0; | |
5298 | ||
5299 | if (!missed_updates) | |
5300 | return load; | |
5301 | ||
5302 | if (missed_updates >= degrade_zero_ticks[idx]) | |
5303 | return 0; | |
5304 | ||
5305 | if (idx == 1) | |
5306 | return load >> missed_updates; | |
5307 | ||
5308 | while (missed_updates) { | |
5309 | if (missed_updates % 2) | |
5310 | load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; | |
5311 | ||
5312 | missed_updates >>= 1; | |
5313 | j++; | |
5314 | } | |
5315 | return load; | |
5316 | } | |
e022e0d3 PZ |
5317 | |
5318 | static struct { | |
5319 | cpumask_var_t idle_cpus_mask; | |
5320 | atomic_t nr_cpus; | |
f643ea22 | 5321 | int has_blocked; /* Idle CPUS has blocked load */ |
e022e0d3 | 5322 | unsigned long next_balance; /* in jiffy units */ |
f643ea22 | 5323 | unsigned long next_blocked; /* Next update of blocked load in jiffies */ |
e022e0d3 PZ |
5324 | } nohz ____cacheline_aligned; |
5325 | ||
9fd81dd5 | 5326 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 5327 | |
59543275 | 5328 | /** |
cee1afce | 5329 | * __cpu_load_update - update the rq->cpu_load[] statistics |
59543275 BP |
5330 | * @this_rq: The rq to update statistics for |
5331 | * @this_load: The current load | |
5332 | * @pending_updates: The number of missed updates | |
59543275 | 5333 | * |
3289bdb4 | 5334 | * Update rq->cpu_load[] statistics. This function is usually called every |
59543275 BP |
5335 | * scheduler tick (TICK_NSEC). |
5336 | * | |
5337 | * This function computes a decaying average: | |
5338 | * | |
5339 | * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load | |
5340 | * | |
5341 | * Because of NOHZ it might not get called on every tick which gives need for | |
5342 | * the @pending_updates argument. | |
5343 | * | |
5344 | * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1 | |
5345 | * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load | |
5346 | * = A * (A * load[i]_n-2 + B) + B | |
5347 | * = A * (A * (A * load[i]_n-3 + B) + B) + B | |
5348 | * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B | |
5349 | * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B | |
5350 | * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B | |
5351 | * = (1 - 1/2^i)^n * (load[i]_0 - load) + load | |
5352 | * | |
5353 | * In the above we've assumed load_n := load, which is true for NOHZ_FULL as | |
5354 | * any change in load would have resulted in the tick being turned back on. | |
5355 | * | |
5356 | * For regular NOHZ, this reduces to: | |
5357 | * | |
5358 | * load[i]_n = (1 - 1/2^i)^n * load[i]_0 | |
5359 | * | |
5360 | * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra | |
1f41906a | 5361 | * term. |
3289bdb4 | 5362 | */ |
1f41906a FW |
5363 | static void cpu_load_update(struct rq *this_rq, unsigned long this_load, |
5364 | unsigned long pending_updates) | |
3289bdb4 | 5365 | { |
9fd81dd5 | 5366 | unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0]; |
3289bdb4 PZ |
5367 | int i, scale; |
5368 | ||
5369 | this_rq->nr_load_updates++; | |
5370 | ||
5371 | /* Update our load: */ | |
5372 | this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ | |
5373 | for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
5374 | unsigned long old_load, new_load; | |
5375 | ||
5376 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
5377 | ||
7400d3bb | 5378 | old_load = this_rq->cpu_load[i]; |
9fd81dd5 | 5379 | #ifdef CONFIG_NO_HZ_COMMON |
3289bdb4 | 5380 | old_load = decay_load_missed(old_load, pending_updates - 1, i); |
7400d3bb BP |
5381 | if (tickless_load) { |
5382 | old_load -= decay_load_missed(tickless_load, pending_updates - 1, i); | |
5383 | /* | |
5384 | * old_load can never be a negative value because a | |
5385 | * decayed tickless_load cannot be greater than the | |
5386 | * original tickless_load. | |
5387 | */ | |
5388 | old_load += tickless_load; | |
5389 | } | |
9fd81dd5 | 5390 | #endif |
3289bdb4 PZ |
5391 | new_load = this_load; |
5392 | /* | |
5393 | * Round up the averaging division if load is increasing. This | |
5394 | * prevents us from getting stuck on 9 if the load is 10, for | |
5395 | * example. | |
5396 | */ | |
5397 | if (new_load > old_load) | |
5398 | new_load += scale - 1; | |
5399 | ||
5400 | this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; | |
5401 | } | |
3289bdb4 PZ |
5402 | } |
5403 | ||
7ea241af | 5404 | /* Used instead of source_load when we know the type == 0 */ |
c7132dd6 | 5405 | static unsigned long weighted_cpuload(struct rq *rq) |
7ea241af | 5406 | { |
c7132dd6 | 5407 | return cfs_rq_runnable_load_avg(&rq->cfs); |
7ea241af YD |
5408 | } |
5409 | ||
3289bdb4 | 5410 | #ifdef CONFIG_NO_HZ_COMMON |
1f41906a FW |
5411 | /* |
5412 | * There is no sane way to deal with nohz on smp when using jiffies because the | |
97fb7a0a | 5413 | * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading |
1f41906a FW |
5414 | * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. |
5415 | * | |
5416 | * Therefore we need to avoid the delta approach from the regular tick when | |
5417 | * possible since that would seriously skew the load calculation. This is why we | |
5418 | * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on | |
5419 | * jiffies deltas for updates happening while in nohz mode (idle ticks, idle | |
5420 | * loop exit, nohz_idle_balance, nohz full exit...) | |
5421 | * | |
5422 | * This means we might still be one tick off for nohz periods. | |
5423 | */ | |
5424 | ||
5425 | static void cpu_load_update_nohz(struct rq *this_rq, | |
5426 | unsigned long curr_jiffies, | |
5427 | unsigned long load) | |
be68a682 FW |
5428 | { |
5429 | unsigned long pending_updates; | |
5430 | ||
5431 | pending_updates = curr_jiffies - this_rq->last_load_update_tick; | |
5432 | if (pending_updates) { | |
5433 | this_rq->last_load_update_tick = curr_jiffies; | |
5434 | /* | |
5435 | * In the regular NOHZ case, we were idle, this means load 0. | |
5436 | * In the NOHZ_FULL case, we were non-idle, we should consider | |
5437 | * its weighted load. | |
5438 | */ | |
1f41906a | 5439 | cpu_load_update(this_rq, load, pending_updates); |
be68a682 FW |
5440 | } |
5441 | } | |
5442 | ||
3289bdb4 PZ |
5443 | /* |
5444 | * Called from nohz_idle_balance() to update the load ratings before doing the | |
5445 | * idle balance. | |
5446 | */ | |
cee1afce | 5447 | static void cpu_load_update_idle(struct rq *this_rq) |
3289bdb4 | 5448 | { |
3289bdb4 PZ |
5449 | /* |
5450 | * bail if there's load or we're actually up-to-date. | |
5451 | */ | |
c7132dd6 | 5452 | if (weighted_cpuload(this_rq)) |
3289bdb4 PZ |
5453 | return; |
5454 | ||
1f41906a | 5455 | cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0); |
3289bdb4 PZ |
5456 | } |
5457 | ||
5458 | /* | |
1f41906a FW |
5459 | * Record CPU load on nohz entry so we know the tickless load to account |
5460 | * on nohz exit. cpu_load[0] happens then to be updated more frequently | |
5461 | * than other cpu_load[idx] but it should be fine as cpu_load readers | |
5462 | * shouldn't rely into synchronized cpu_load[*] updates. | |
3289bdb4 | 5463 | */ |
1f41906a | 5464 | void cpu_load_update_nohz_start(void) |
3289bdb4 PZ |
5465 | { |
5466 | struct rq *this_rq = this_rq(); | |
1f41906a FW |
5467 | |
5468 | /* | |
5469 | * This is all lockless but should be fine. If weighted_cpuload changes | |
5470 | * concurrently we'll exit nohz. And cpu_load write can race with | |
5471 | * cpu_load_update_idle() but both updater would be writing the same. | |
5472 | */ | |
c7132dd6 | 5473 | this_rq->cpu_load[0] = weighted_cpuload(this_rq); |
1f41906a FW |
5474 | } |
5475 | ||
5476 | /* | |
5477 | * Account the tickless load in the end of a nohz frame. | |
5478 | */ | |
5479 | void cpu_load_update_nohz_stop(void) | |
5480 | { | |
316c1608 | 5481 | unsigned long curr_jiffies = READ_ONCE(jiffies); |
1f41906a FW |
5482 | struct rq *this_rq = this_rq(); |
5483 | unsigned long load; | |
8a8c69c3 | 5484 | struct rq_flags rf; |
3289bdb4 PZ |
5485 | |
5486 | if (curr_jiffies == this_rq->last_load_update_tick) | |
5487 | return; | |
5488 | ||
c7132dd6 | 5489 | load = weighted_cpuload(this_rq); |
8a8c69c3 | 5490 | rq_lock(this_rq, &rf); |
b52fad2d | 5491 | update_rq_clock(this_rq); |
1f41906a | 5492 | cpu_load_update_nohz(this_rq, curr_jiffies, load); |
8a8c69c3 | 5493 | rq_unlock(this_rq, &rf); |
3289bdb4 | 5494 | } |
1f41906a FW |
5495 | #else /* !CONFIG_NO_HZ_COMMON */ |
5496 | static inline void cpu_load_update_nohz(struct rq *this_rq, | |
5497 | unsigned long curr_jiffies, | |
5498 | unsigned long load) { } | |
5499 | #endif /* CONFIG_NO_HZ_COMMON */ | |
5500 | ||
5501 | static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load) | |
5502 | { | |
9fd81dd5 | 5503 | #ifdef CONFIG_NO_HZ_COMMON |
1f41906a FW |
5504 | /* See the mess around cpu_load_update_nohz(). */ |
5505 | this_rq->last_load_update_tick = READ_ONCE(jiffies); | |
9fd81dd5 | 5506 | #endif |
1f41906a FW |
5507 | cpu_load_update(this_rq, load, 1); |
5508 | } | |
3289bdb4 PZ |
5509 | |
5510 | /* | |
5511 | * Called from scheduler_tick() | |
5512 | */ | |
cee1afce | 5513 | void cpu_load_update_active(struct rq *this_rq) |
3289bdb4 | 5514 | { |
c7132dd6 | 5515 | unsigned long load = weighted_cpuload(this_rq); |
1f41906a FW |
5516 | |
5517 | if (tick_nohz_tick_stopped()) | |
5518 | cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load); | |
5519 | else | |
5520 | cpu_load_update_periodic(this_rq, load); | |
3289bdb4 PZ |
5521 | } |
5522 | ||
029632fb | 5523 | /* |
97fb7a0a | 5524 | * Return a low guess at the load of a migration-source CPU weighted |
029632fb PZ |
5525 | * according to the scheduling class and "nice" value. |
5526 | * | |
5527 | * We want to under-estimate the load of migration sources, to | |
5528 | * balance conservatively. | |
5529 | */ | |
5530 | static unsigned long source_load(int cpu, int type) | |
5531 | { | |
5532 | struct rq *rq = cpu_rq(cpu); | |
c7132dd6 | 5533 | unsigned long total = weighted_cpuload(rq); |
029632fb PZ |
5534 | |
5535 | if (type == 0 || !sched_feat(LB_BIAS)) | |
5536 | return total; | |
5537 | ||
5538 | return min(rq->cpu_load[type-1], total); | |
5539 | } | |
5540 | ||
5541 | /* | |
97fb7a0a | 5542 | * Return a high guess at the load of a migration-target CPU weighted |
029632fb PZ |
5543 | * according to the scheduling class and "nice" value. |
5544 | */ | |
5545 | static unsigned long target_load(int cpu, int type) | |
5546 | { | |
5547 | struct rq *rq = cpu_rq(cpu); | |
c7132dd6 | 5548 | unsigned long total = weighted_cpuload(rq); |
029632fb PZ |
5549 | |
5550 | if (type == 0 || !sched_feat(LB_BIAS)) | |
5551 | return total; | |
5552 | ||
5553 | return max(rq->cpu_load[type-1], total); | |
5554 | } | |
5555 | ||
ced549fa | 5556 | static unsigned long capacity_of(int cpu) |
029632fb | 5557 | { |
ced549fa | 5558 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
5559 | } |
5560 | ||
ca6d75e6 VG |
5561 | static unsigned long capacity_orig_of(int cpu) |
5562 | { | |
5563 | return cpu_rq(cpu)->cpu_capacity_orig; | |
5564 | } | |
5565 | ||
029632fb PZ |
5566 | static unsigned long cpu_avg_load_per_task(int cpu) |
5567 | { | |
5568 | struct rq *rq = cpu_rq(cpu); | |
316c1608 | 5569 | unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running); |
c7132dd6 | 5570 | unsigned long load_avg = weighted_cpuload(rq); |
029632fb PZ |
5571 | |
5572 | if (nr_running) | |
b92486cb | 5573 | return load_avg / nr_running; |
029632fb PZ |
5574 | |
5575 | return 0; | |
5576 | } | |
5577 | ||
c58d25f3 PZ |
5578 | static void record_wakee(struct task_struct *p) |
5579 | { | |
5580 | /* | |
5581 | * Only decay a single time; tasks that have less then 1 wakeup per | |
5582 | * jiffy will not have built up many flips. | |
5583 | */ | |
5584 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
5585 | current->wakee_flips >>= 1; | |
5586 | current->wakee_flip_decay_ts = jiffies; | |
5587 | } | |
5588 | ||
5589 | if (current->last_wakee != p) { | |
5590 | current->last_wakee = p; | |
5591 | current->wakee_flips++; | |
5592 | } | |
5593 | } | |
5594 | ||
63b0e9ed MG |
5595 | /* |
5596 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 5597 | * |
63b0e9ed | 5598 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
5599 | * at a frequency roughly N times higher than one of its wakees. |
5600 | * | |
5601 | * In order to determine whether we should let the load spread vs consolidating | |
5602 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
5603 | * partner, and a factor of lls_size higher frequency in the other. | |
5604 | * | |
5605 | * With both conditions met, we can be relatively sure that the relationship is | |
5606 | * non-monogamous, with partner count exceeding socket size. | |
5607 | * | |
5608 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
5609 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
5610 | * socket size. | |
63b0e9ed | 5611 | */ |
62470419 MW |
5612 | static int wake_wide(struct task_struct *p) |
5613 | { | |
63b0e9ed MG |
5614 | unsigned int master = current->wakee_flips; |
5615 | unsigned int slave = p->wakee_flips; | |
7d9ffa89 | 5616 | int factor = this_cpu_read(sd_llc_size); |
62470419 | 5617 | |
63b0e9ed MG |
5618 | if (master < slave) |
5619 | swap(master, slave); | |
5620 | if (slave < factor || master < slave * factor) | |
5621 | return 0; | |
5622 | return 1; | |
62470419 MW |
5623 | } |
5624 | ||
90001d67 | 5625 | /* |
d153b153 PZ |
5626 | * The purpose of wake_affine() is to quickly determine on which CPU we can run |
5627 | * soonest. For the purpose of speed we only consider the waking and previous | |
5628 | * CPU. | |
90001d67 | 5629 | * |
7332dec0 MG |
5630 | * wake_affine_idle() - only considers 'now', it check if the waking CPU is |
5631 | * cache-affine and is (or will be) idle. | |
f2cdd9cc PZ |
5632 | * |
5633 | * wake_affine_weight() - considers the weight to reflect the average | |
5634 | * scheduling latency of the CPUs. This seems to work | |
5635 | * for the overloaded case. | |
90001d67 | 5636 | */ |
3b76c4a3 | 5637 | static int |
89a55f56 | 5638 | wake_affine_idle(int this_cpu, int prev_cpu, int sync) |
90001d67 | 5639 | { |
7332dec0 MG |
5640 | /* |
5641 | * If this_cpu is idle, it implies the wakeup is from interrupt | |
5642 | * context. Only allow the move if cache is shared. Otherwise an | |
5643 | * interrupt intensive workload could force all tasks onto one | |
5644 | * node depending on the IO topology or IRQ affinity settings. | |
806486c3 MG |
5645 | * |
5646 | * If the prev_cpu is idle and cache affine then avoid a migration. | |
5647 | * There is no guarantee that the cache hot data from an interrupt | |
5648 | * is more important than cache hot data on the prev_cpu and from | |
5649 | * a cpufreq perspective, it's better to have higher utilisation | |
5650 | * on one CPU. | |
7332dec0 | 5651 | */ |
943d355d RJ |
5652 | if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu)) |
5653 | return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu; | |
90001d67 | 5654 | |
d153b153 | 5655 | if (sync && cpu_rq(this_cpu)->nr_running == 1) |
3b76c4a3 | 5656 | return this_cpu; |
90001d67 | 5657 | |
3b76c4a3 | 5658 | return nr_cpumask_bits; |
90001d67 PZ |
5659 | } |
5660 | ||
3b76c4a3 | 5661 | static int |
f2cdd9cc PZ |
5662 | wake_affine_weight(struct sched_domain *sd, struct task_struct *p, |
5663 | int this_cpu, int prev_cpu, int sync) | |
90001d67 | 5664 | { |
90001d67 PZ |
5665 | s64 this_eff_load, prev_eff_load; |
5666 | unsigned long task_load; | |
5667 | ||
f2cdd9cc | 5668 | this_eff_load = target_load(this_cpu, sd->wake_idx); |
90001d67 | 5669 | |
90001d67 PZ |
5670 | if (sync) { |
5671 | unsigned long current_load = task_h_load(current); | |
5672 | ||
f2cdd9cc | 5673 | if (current_load > this_eff_load) |
3b76c4a3 | 5674 | return this_cpu; |
90001d67 | 5675 | |
f2cdd9cc | 5676 | this_eff_load -= current_load; |
90001d67 PZ |
5677 | } |
5678 | ||
90001d67 PZ |
5679 | task_load = task_h_load(p); |
5680 | ||
f2cdd9cc PZ |
5681 | this_eff_load += task_load; |
5682 | if (sched_feat(WA_BIAS)) | |
5683 | this_eff_load *= 100; | |
5684 | this_eff_load *= capacity_of(prev_cpu); | |
90001d67 | 5685 | |
eeb60398 | 5686 | prev_eff_load = source_load(prev_cpu, sd->wake_idx); |
f2cdd9cc PZ |
5687 | prev_eff_load -= task_load; |
5688 | if (sched_feat(WA_BIAS)) | |
5689 | prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; | |
5690 | prev_eff_load *= capacity_of(this_cpu); | |
90001d67 | 5691 | |
082f764a MG |
5692 | /* |
5693 | * If sync, adjust the weight of prev_eff_load such that if | |
5694 | * prev_eff == this_eff that select_idle_sibling() will consider | |
5695 | * stacking the wakee on top of the waker if no other CPU is | |
5696 | * idle. | |
5697 | */ | |
5698 | if (sync) | |
5699 | prev_eff_load += 1; | |
5700 | ||
5701 | return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits; | |
90001d67 PZ |
5702 | } |
5703 | ||
772bd008 | 5704 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
7ebb66a1 | 5705 | int this_cpu, int prev_cpu, int sync) |
098fb9db | 5706 | { |
3b76c4a3 | 5707 | int target = nr_cpumask_bits; |
098fb9db | 5708 | |
89a55f56 | 5709 | if (sched_feat(WA_IDLE)) |
3b76c4a3 | 5710 | target = wake_affine_idle(this_cpu, prev_cpu, sync); |
90001d67 | 5711 | |
3b76c4a3 MG |
5712 | if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits) |
5713 | target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); | |
098fb9db | 5714 | |
ae92882e | 5715 | schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); |
3b76c4a3 MG |
5716 | if (target == nr_cpumask_bits) |
5717 | return prev_cpu; | |
098fb9db | 5718 | |
3b76c4a3 MG |
5719 | schedstat_inc(sd->ttwu_move_affine); |
5720 | schedstat_inc(p->se.statistics.nr_wakeups_affine); | |
5721 | return target; | |
098fb9db IM |
5722 | } |
5723 | ||
c469933e | 5724 | static unsigned long cpu_util_without(int cpu, struct task_struct *p); |
6a0b19c0 | 5725 | |
c469933e | 5726 | static unsigned long capacity_spare_without(int cpu, struct task_struct *p) |
6a0b19c0 | 5727 | { |
c469933e | 5728 | return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0); |
6a0b19c0 MR |
5729 | } |
5730 | ||
aaee1203 PZ |
5731 | /* |
5732 | * find_idlest_group finds and returns the least busy CPU group within the | |
5733 | * domain. | |
6fee85cc BJ |
5734 | * |
5735 | * Assumes p is allowed on at least one CPU in sd. | |
aaee1203 PZ |
5736 | */ |
5737 | static struct sched_group * | |
78e7ed53 | 5738 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
c44f2a02 | 5739 | int this_cpu, int sd_flag) |
e7693a36 | 5740 | { |
b3bd3de6 | 5741 | struct sched_group *idlest = NULL, *group = sd->groups; |
6a0b19c0 | 5742 | struct sched_group *most_spare_sg = NULL; |
0d10ab95 BJ |
5743 | unsigned long min_runnable_load = ULONG_MAX; |
5744 | unsigned long this_runnable_load = ULONG_MAX; | |
5745 | unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX; | |
6a0b19c0 | 5746 | unsigned long most_spare = 0, this_spare = 0; |
c44f2a02 | 5747 | int load_idx = sd->forkexec_idx; |
6b94780e VG |
5748 | int imbalance_scale = 100 + (sd->imbalance_pct-100)/2; |
5749 | unsigned long imbalance = scale_load_down(NICE_0_LOAD) * | |
5750 | (sd->imbalance_pct-100) / 100; | |
e7693a36 | 5751 | |
c44f2a02 VG |
5752 | if (sd_flag & SD_BALANCE_WAKE) |
5753 | load_idx = sd->wake_idx; | |
5754 | ||
aaee1203 | 5755 | do { |
6b94780e VG |
5756 | unsigned long load, avg_load, runnable_load; |
5757 | unsigned long spare_cap, max_spare_cap; | |
aaee1203 PZ |
5758 | int local_group; |
5759 | int i; | |
e7693a36 | 5760 | |
aaee1203 | 5761 | /* Skip over this group if it has no CPUs allowed */ |
ae4df9d6 | 5762 | if (!cpumask_intersects(sched_group_span(group), |
0c98d344 | 5763 | &p->cpus_allowed)) |
aaee1203 PZ |
5764 | continue; |
5765 | ||
5766 | local_group = cpumask_test_cpu(this_cpu, | |
ae4df9d6 | 5767 | sched_group_span(group)); |
aaee1203 | 5768 | |
6a0b19c0 MR |
5769 | /* |
5770 | * Tally up the load of all CPUs in the group and find | |
5771 | * the group containing the CPU with most spare capacity. | |
5772 | */ | |
aaee1203 | 5773 | avg_load = 0; |
6b94780e | 5774 | runnable_load = 0; |
6a0b19c0 | 5775 | max_spare_cap = 0; |
aaee1203 | 5776 | |
ae4df9d6 | 5777 | for_each_cpu(i, sched_group_span(group)) { |
97fb7a0a | 5778 | /* Bias balancing toward CPUs of our domain */ |
aaee1203 PZ |
5779 | if (local_group) |
5780 | load = source_load(i, load_idx); | |
5781 | else | |
5782 | load = target_load(i, load_idx); | |
5783 | ||
6b94780e VG |
5784 | runnable_load += load; |
5785 | ||
5786 | avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs); | |
6a0b19c0 | 5787 | |
c469933e | 5788 | spare_cap = capacity_spare_without(i, p); |
6a0b19c0 MR |
5789 | |
5790 | if (spare_cap > max_spare_cap) | |
5791 | max_spare_cap = spare_cap; | |
aaee1203 PZ |
5792 | } |
5793 | ||
63b2ca30 | 5794 | /* Adjust by relative CPU capacity of the group */ |
6b94780e VG |
5795 | avg_load = (avg_load * SCHED_CAPACITY_SCALE) / |
5796 | group->sgc->capacity; | |
5797 | runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) / | |
5798 | group->sgc->capacity; | |
aaee1203 PZ |
5799 | |
5800 | if (local_group) { | |
6b94780e VG |
5801 | this_runnable_load = runnable_load; |
5802 | this_avg_load = avg_load; | |
6a0b19c0 MR |
5803 | this_spare = max_spare_cap; |
5804 | } else { | |
6b94780e VG |
5805 | if (min_runnable_load > (runnable_load + imbalance)) { |
5806 | /* | |
5807 | * The runnable load is significantly smaller | |
97fb7a0a | 5808 | * so we can pick this new CPU: |
6b94780e VG |
5809 | */ |
5810 | min_runnable_load = runnable_load; | |
5811 | min_avg_load = avg_load; | |
5812 | idlest = group; | |
5813 | } else if ((runnable_load < (min_runnable_load + imbalance)) && | |
5814 | (100*min_avg_load > imbalance_scale*avg_load)) { | |
5815 | /* | |
5816 | * The runnable loads are close so take the | |
97fb7a0a | 5817 | * blocked load into account through avg_load: |
6b94780e VG |
5818 | */ |
5819 | min_avg_load = avg_load; | |
6a0b19c0 MR |
5820 | idlest = group; |
5821 | } | |
5822 | ||
5823 | if (most_spare < max_spare_cap) { | |
5824 | most_spare = max_spare_cap; | |
5825 | most_spare_sg = group; | |
5826 | } | |
aaee1203 PZ |
5827 | } |
5828 | } while (group = group->next, group != sd->groups); | |
5829 | ||
6a0b19c0 MR |
5830 | /* |
5831 | * The cross-over point between using spare capacity or least load | |
5832 | * is too conservative for high utilization tasks on partially | |
5833 | * utilized systems if we require spare_capacity > task_util(p), | |
5834 | * so we allow for some task stuffing by using | |
5835 | * spare_capacity > task_util(p)/2. | |
f519a3f1 VG |
5836 | * |
5837 | * Spare capacity can't be used for fork because the utilization has | |
5838 | * not been set yet, we must first select a rq to compute the initial | |
5839 | * utilization. | |
6a0b19c0 | 5840 | */ |
f519a3f1 VG |
5841 | if (sd_flag & SD_BALANCE_FORK) |
5842 | goto skip_spare; | |
5843 | ||
6a0b19c0 | 5844 | if (this_spare > task_util(p) / 2 && |
6b94780e | 5845 | imbalance_scale*this_spare > 100*most_spare) |
6a0b19c0 | 5846 | return NULL; |
6b94780e VG |
5847 | |
5848 | if (most_spare > task_util(p) / 2) | |
6a0b19c0 MR |
5849 | return most_spare_sg; |
5850 | ||
f519a3f1 | 5851 | skip_spare: |
6b94780e VG |
5852 | if (!idlest) |
5853 | return NULL; | |
5854 | ||
2c833627 MG |
5855 | /* |
5856 | * When comparing groups across NUMA domains, it's possible for the | |
5857 | * local domain to be very lightly loaded relative to the remote | |
5858 | * domains but "imbalance" skews the comparison making remote CPUs | |
5859 | * look much more favourable. When considering cross-domain, add | |
5860 | * imbalance to the runnable load on the remote node and consider | |
5861 | * staying local. | |
5862 | */ | |
5863 | if ((sd->flags & SD_NUMA) && | |
5864 | min_runnable_load + imbalance >= this_runnable_load) | |
5865 | return NULL; | |
5866 | ||
6b94780e | 5867 | if (min_runnable_load > (this_runnable_load + imbalance)) |
aaee1203 | 5868 | return NULL; |
6b94780e VG |
5869 | |
5870 | if ((this_runnable_load < (min_runnable_load + imbalance)) && | |
5871 | (100*this_avg_load < imbalance_scale*min_avg_load)) | |
5872 | return NULL; | |
5873 | ||
aaee1203 PZ |
5874 | return idlest; |
5875 | } | |
5876 | ||
5877 | /* | |
97fb7a0a | 5878 | * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group. |
aaee1203 PZ |
5879 | */ |
5880 | static int | |
18bd1b4b | 5881 | find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) |
aaee1203 PZ |
5882 | { |
5883 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
5884 | unsigned int min_exit_latency = UINT_MAX; |
5885 | u64 latest_idle_timestamp = 0; | |
5886 | int least_loaded_cpu = this_cpu; | |
5887 | int shallowest_idle_cpu = -1; | |
aaee1203 PZ |
5888 | int i; |
5889 | ||
eaecf41f MR |
5890 | /* Check if we have any choice: */ |
5891 | if (group->group_weight == 1) | |
ae4df9d6 | 5892 | return cpumask_first(sched_group_span(group)); |
eaecf41f | 5893 | |
aaee1203 | 5894 | /* Traverse only the allowed CPUs */ |
ae4df9d6 | 5895 | for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) { |
943d355d | 5896 | if (available_idle_cpu(i)) { |
83a0a96a NP |
5897 | struct rq *rq = cpu_rq(i); |
5898 | struct cpuidle_state *idle = idle_get_state(rq); | |
5899 | if (idle && idle->exit_latency < min_exit_latency) { | |
5900 | /* | |
5901 | * We give priority to a CPU whose idle state | |
5902 | * has the smallest exit latency irrespective | |
5903 | * of any idle timestamp. | |
5904 | */ | |
5905 | min_exit_latency = idle->exit_latency; | |
5906 | latest_idle_timestamp = rq->idle_stamp; | |
5907 | shallowest_idle_cpu = i; | |
5908 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
5909 | rq->idle_stamp > latest_idle_timestamp) { | |
5910 | /* | |
5911 | * If equal or no active idle state, then | |
5912 | * the most recently idled CPU might have | |
5913 | * a warmer cache. | |
5914 | */ | |
5915 | latest_idle_timestamp = rq->idle_stamp; | |
5916 | shallowest_idle_cpu = i; | |
5917 | } | |
9f96742a | 5918 | } else if (shallowest_idle_cpu == -1) { |
c7132dd6 | 5919 | load = weighted_cpuload(cpu_rq(i)); |
18cec7e0 | 5920 | if (load < min_load) { |
83a0a96a NP |
5921 | min_load = load; |
5922 | least_loaded_cpu = i; | |
5923 | } | |
e7693a36 GH |
5924 | } |
5925 | } | |
5926 | ||
83a0a96a | 5927 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 5928 | } |
e7693a36 | 5929 | |
18bd1b4b BJ |
5930 | static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p, |
5931 | int cpu, int prev_cpu, int sd_flag) | |
5932 | { | |
93f50f90 | 5933 | int new_cpu = cpu; |
18bd1b4b | 5934 | |
6fee85cc BJ |
5935 | if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed)) |
5936 | return prev_cpu; | |
5937 | ||
c976a862 | 5938 | /* |
c469933e PB |
5939 | * We need task's util for capacity_spare_without, sync it up to |
5940 | * prev_cpu's last_update_time. | |
c976a862 VK |
5941 | */ |
5942 | if (!(sd_flag & SD_BALANCE_FORK)) | |
5943 | sync_entity_load_avg(&p->se); | |
5944 | ||
18bd1b4b BJ |
5945 | while (sd) { |
5946 | struct sched_group *group; | |
5947 | struct sched_domain *tmp; | |
5948 | int weight; | |
5949 | ||
5950 | if (!(sd->flags & sd_flag)) { | |
5951 | sd = sd->child; | |
5952 | continue; | |
5953 | } | |
5954 | ||
5955 | group = find_idlest_group(sd, p, cpu, sd_flag); | |
5956 | if (!group) { | |
5957 | sd = sd->child; | |
5958 | continue; | |
5959 | } | |
5960 | ||
5961 | new_cpu = find_idlest_group_cpu(group, p, cpu); | |
e90381ea | 5962 | if (new_cpu == cpu) { |
97fb7a0a | 5963 | /* Now try balancing at a lower domain level of 'cpu': */ |
18bd1b4b BJ |
5964 | sd = sd->child; |
5965 | continue; | |
5966 | } | |
5967 | ||
97fb7a0a | 5968 | /* Now try balancing at a lower domain level of 'new_cpu': */ |
18bd1b4b BJ |
5969 | cpu = new_cpu; |
5970 | weight = sd->span_weight; | |
5971 | sd = NULL; | |
5972 | for_each_domain(cpu, tmp) { | |
5973 | if (weight <= tmp->span_weight) | |
5974 | break; | |
5975 | if (tmp->flags & sd_flag) | |
5976 | sd = tmp; | |
5977 | } | |
18bd1b4b BJ |
5978 | } |
5979 | ||
5980 | return new_cpu; | |
5981 | } | |
5982 | ||
10e2f1ac | 5983 | #ifdef CONFIG_SCHED_SMT |
ba2591a5 | 5984 | DEFINE_STATIC_KEY_FALSE(sched_smt_present); |
10e2f1ac PZ |
5985 | |
5986 | static inline void set_idle_cores(int cpu, int val) | |
5987 | { | |
5988 | struct sched_domain_shared *sds; | |
5989 | ||
5990 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
5991 | if (sds) | |
5992 | WRITE_ONCE(sds->has_idle_cores, val); | |
5993 | } | |
5994 | ||
5995 | static inline bool test_idle_cores(int cpu, bool def) | |
5996 | { | |
5997 | struct sched_domain_shared *sds; | |
5998 | ||
5999 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
6000 | if (sds) | |
6001 | return READ_ONCE(sds->has_idle_cores); | |
6002 | ||
6003 | return def; | |
6004 | } | |
6005 | ||
6006 | /* | |
6007 | * Scans the local SMT mask to see if the entire core is idle, and records this | |
6008 | * information in sd_llc_shared->has_idle_cores. | |
6009 | * | |
6010 | * Since SMT siblings share all cache levels, inspecting this limited remote | |
6011 | * state should be fairly cheap. | |
6012 | */ | |
1b568f0a | 6013 | void __update_idle_core(struct rq *rq) |
10e2f1ac PZ |
6014 | { |
6015 | int core = cpu_of(rq); | |
6016 | int cpu; | |
6017 | ||
6018 | rcu_read_lock(); | |
6019 | if (test_idle_cores(core, true)) | |
6020 | goto unlock; | |
6021 | ||
6022 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
6023 | if (cpu == core) | |
6024 | continue; | |
6025 | ||
943d355d | 6026 | if (!available_idle_cpu(cpu)) |
10e2f1ac PZ |
6027 | goto unlock; |
6028 | } | |
6029 | ||
6030 | set_idle_cores(core, 1); | |
6031 | unlock: | |
6032 | rcu_read_unlock(); | |
6033 | } | |
6034 | ||
6035 | /* | |
6036 | * Scan the entire LLC domain for idle cores; this dynamically switches off if | |
6037 | * there are no idle cores left in the system; tracked through | |
6038 | * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. | |
6039 | */ | |
6040 | static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) | |
6041 | { | |
6042 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); | |
c743f0a5 | 6043 | int core, cpu; |
10e2f1ac | 6044 | |
1b568f0a PZ |
6045 | if (!static_branch_likely(&sched_smt_present)) |
6046 | return -1; | |
6047 | ||
10e2f1ac PZ |
6048 | if (!test_idle_cores(target, false)) |
6049 | return -1; | |
6050 | ||
0c98d344 | 6051 | cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed); |
10e2f1ac | 6052 | |
c743f0a5 | 6053 | for_each_cpu_wrap(core, cpus, target) { |
10e2f1ac PZ |
6054 | bool idle = true; |
6055 | ||
6056 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
6057 | cpumask_clear_cpu(cpu, cpus); | |
943d355d | 6058 | if (!available_idle_cpu(cpu)) |
10e2f1ac PZ |
6059 | idle = false; |
6060 | } | |
6061 | ||
6062 | if (idle) | |
6063 | return core; | |
6064 | } | |
6065 | ||
6066 | /* | |
6067 | * Failed to find an idle core; stop looking for one. | |
6068 | */ | |
6069 | set_idle_cores(target, 0); | |
6070 | ||
6071 | return -1; | |
6072 | } | |
6073 | ||
6074 | /* | |
6075 | * Scan the local SMT mask for idle CPUs. | |
6076 | */ | |
6077 | static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) | |
6078 | { | |
6079 | int cpu; | |
6080 | ||
1b568f0a PZ |
6081 | if (!static_branch_likely(&sched_smt_present)) |
6082 | return -1; | |
6083 | ||
10e2f1ac | 6084 | for_each_cpu(cpu, cpu_smt_mask(target)) { |
0c98d344 | 6085 | if (!cpumask_test_cpu(cpu, &p->cpus_allowed)) |
10e2f1ac | 6086 | continue; |
943d355d | 6087 | if (available_idle_cpu(cpu)) |
10e2f1ac PZ |
6088 | return cpu; |
6089 | } | |
6090 | ||
6091 | return -1; | |
6092 | } | |
6093 | ||
6094 | #else /* CONFIG_SCHED_SMT */ | |
6095 | ||
6096 | static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) | |
6097 | { | |
6098 | return -1; | |
6099 | } | |
6100 | ||
6101 | static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) | |
6102 | { | |
6103 | return -1; | |
6104 | } | |
6105 | ||
6106 | #endif /* CONFIG_SCHED_SMT */ | |
6107 | ||
6108 | /* | |
6109 | * Scan the LLC domain for idle CPUs; this is dynamically regulated by | |
6110 | * comparing the average scan cost (tracked in sd->avg_scan_cost) against the | |
6111 | * average idle time for this rq (as found in rq->avg_idle). | |
a50bde51 | 6112 | */ |
10e2f1ac PZ |
6113 | static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target) |
6114 | { | |
9cfb38a7 | 6115 | struct sched_domain *this_sd; |
1ad3aaf3 | 6116 | u64 avg_cost, avg_idle; |
10e2f1ac PZ |
6117 | u64 time, cost; |
6118 | s64 delta; | |
1ad3aaf3 | 6119 | int cpu, nr = INT_MAX; |
10e2f1ac | 6120 | |
9cfb38a7 WL |
6121 | this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); |
6122 | if (!this_sd) | |
6123 | return -1; | |
6124 | ||
10e2f1ac PZ |
6125 | /* |
6126 | * Due to large variance we need a large fuzz factor; hackbench in | |
6127 | * particularly is sensitive here. | |
6128 | */ | |
1ad3aaf3 PZ |
6129 | avg_idle = this_rq()->avg_idle / 512; |
6130 | avg_cost = this_sd->avg_scan_cost + 1; | |
6131 | ||
6132 | if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost) | |
10e2f1ac PZ |
6133 | return -1; |
6134 | ||
1ad3aaf3 PZ |
6135 | if (sched_feat(SIS_PROP)) { |
6136 | u64 span_avg = sd->span_weight * avg_idle; | |
6137 | if (span_avg > 4*avg_cost) | |
6138 | nr = div_u64(span_avg, avg_cost); | |
6139 | else | |
6140 | nr = 4; | |
6141 | } | |
6142 | ||
10e2f1ac PZ |
6143 | time = local_clock(); |
6144 | ||
c743f0a5 | 6145 | for_each_cpu_wrap(cpu, sched_domain_span(sd), target) { |
1ad3aaf3 PZ |
6146 | if (!--nr) |
6147 | return -1; | |
0c98d344 | 6148 | if (!cpumask_test_cpu(cpu, &p->cpus_allowed)) |
10e2f1ac | 6149 | continue; |
943d355d | 6150 | if (available_idle_cpu(cpu)) |
10e2f1ac PZ |
6151 | break; |
6152 | } | |
6153 | ||
6154 | time = local_clock() - time; | |
6155 | cost = this_sd->avg_scan_cost; | |
6156 | delta = (s64)(time - cost) / 8; | |
6157 | this_sd->avg_scan_cost += delta; | |
6158 | ||
6159 | return cpu; | |
6160 | } | |
6161 | ||
6162 | /* | |
6163 | * Try and locate an idle core/thread in the LLC cache domain. | |
a50bde51 | 6164 | */ |
772bd008 | 6165 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 6166 | { |
99bd5e2f | 6167 | struct sched_domain *sd; |
32e839dd | 6168 | int i, recent_used_cpu; |
a50bde51 | 6169 | |
943d355d | 6170 | if (available_idle_cpu(target)) |
e0a79f52 | 6171 | return target; |
99bd5e2f SS |
6172 | |
6173 | /* | |
97fb7a0a | 6174 | * If the previous CPU is cache affine and idle, don't be stupid: |
99bd5e2f | 6175 | */ |
943d355d | 6176 | if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev)) |
772bd008 | 6177 | return prev; |
a50bde51 | 6178 | |
97fb7a0a | 6179 | /* Check a recently used CPU as a potential idle candidate: */ |
32e839dd MG |
6180 | recent_used_cpu = p->recent_used_cpu; |
6181 | if (recent_used_cpu != prev && | |
6182 | recent_used_cpu != target && | |
6183 | cpus_share_cache(recent_used_cpu, target) && | |
943d355d | 6184 | available_idle_cpu(recent_used_cpu) && |
32e839dd MG |
6185 | cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) { |
6186 | /* | |
6187 | * Replace recent_used_cpu with prev as it is a potential | |
97fb7a0a | 6188 | * candidate for the next wake: |
32e839dd MG |
6189 | */ |
6190 | p->recent_used_cpu = prev; | |
6191 | return recent_used_cpu; | |
6192 | } | |
6193 | ||
518cd623 | 6194 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
10e2f1ac PZ |
6195 | if (!sd) |
6196 | return target; | |
772bd008 | 6197 | |
10e2f1ac PZ |
6198 | i = select_idle_core(p, sd, target); |
6199 | if ((unsigned)i < nr_cpumask_bits) | |
6200 | return i; | |
37407ea7 | 6201 | |
10e2f1ac PZ |
6202 | i = select_idle_cpu(p, sd, target); |
6203 | if ((unsigned)i < nr_cpumask_bits) | |
6204 | return i; | |
6205 | ||
6206 | i = select_idle_smt(p, sd, target); | |
6207 | if ((unsigned)i < nr_cpumask_bits) | |
6208 | return i; | |
970e1789 | 6209 | |
a50bde51 PZ |
6210 | return target; |
6211 | } | |
231678b7 | 6212 | |
f9be3e59 PB |
6213 | /** |
6214 | * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks | |
6215 | * @cpu: the CPU to get the utilization of | |
6216 | * | |
6217 | * The unit of the return value must be the one of capacity so we can compare | |
6218 | * the utilization with the capacity of the CPU that is available for CFS task | |
6219 | * (ie cpu_capacity). | |
231678b7 DE |
6220 | * |
6221 | * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the | |
6222 | * recent utilization of currently non-runnable tasks on a CPU. It represents | |
6223 | * the amount of utilization of a CPU in the range [0..capacity_orig] where | |
6224 | * capacity_orig is the cpu_capacity available at the highest frequency | |
6225 | * (arch_scale_freq_capacity()). | |
6226 | * The utilization of a CPU converges towards a sum equal to or less than the | |
6227 | * current capacity (capacity_curr <= capacity_orig) of the CPU because it is | |
6228 | * the running time on this CPU scaled by capacity_curr. | |
6229 | * | |
f9be3e59 PB |
6230 | * The estimated utilization of a CPU is defined to be the maximum between its |
6231 | * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks | |
6232 | * currently RUNNABLE on that CPU. | |
6233 | * This allows to properly represent the expected utilization of a CPU which | |
6234 | * has just got a big task running since a long sleep period. At the same time | |
6235 | * however it preserves the benefits of the "blocked utilization" in | |
6236 | * describing the potential for other tasks waking up on the same CPU. | |
6237 | * | |
231678b7 DE |
6238 | * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even |
6239 | * higher than capacity_orig because of unfortunate rounding in | |
6240 | * cfs.avg.util_avg or just after migrating tasks and new task wakeups until | |
6241 | * the average stabilizes with the new running time. We need to check that the | |
6242 | * utilization stays within the range of [0..capacity_orig] and cap it if | |
6243 | * necessary. Without utilization capping, a group could be seen as overloaded | |
6244 | * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of | |
6245 | * available capacity. We allow utilization to overshoot capacity_curr (but not | |
6246 | * capacity_orig) as it useful for predicting the capacity required after task | |
6247 | * migrations (scheduler-driven DVFS). | |
f9be3e59 PB |
6248 | * |
6249 | * Return: the (estimated) utilization for the specified CPU | |
8bb5b00c | 6250 | */ |
f9be3e59 | 6251 | static inline unsigned long cpu_util(int cpu) |
8bb5b00c | 6252 | { |
f9be3e59 PB |
6253 | struct cfs_rq *cfs_rq; |
6254 | unsigned int util; | |
6255 | ||
6256 | cfs_rq = &cpu_rq(cpu)->cfs; | |
6257 | util = READ_ONCE(cfs_rq->avg.util_avg); | |
6258 | ||
6259 | if (sched_feat(UTIL_EST)) | |
6260 | util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued)); | |
8bb5b00c | 6261 | |
f9be3e59 | 6262 | return min_t(unsigned long, util, capacity_orig_of(cpu)); |
8bb5b00c | 6263 | } |
a50bde51 | 6264 | |
104cb16d | 6265 | /* |
c469933e PB |
6266 | * cpu_util_without: compute cpu utilization without any contributions from *p |
6267 | * @cpu: the CPU which utilization is requested | |
6268 | * @p: the task which utilization should be discounted | |
6269 | * | |
6270 | * The utilization of a CPU is defined by the utilization of tasks currently | |
6271 | * enqueued on that CPU as well as tasks which are currently sleeping after an | |
6272 | * execution on that CPU. | |
6273 | * | |
6274 | * This method returns the utilization of the specified CPU by discounting the | |
6275 | * utilization of the specified task, whenever the task is currently | |
6276 | * contributing to the CPU utilization. | |
104cb16d | 6277 | */ |
c469933e | 6278 | static unsigned long cpu_util_without(int cpu, struct task_struct *p) |
104cb16d | 6279 | { |
f9be3e59 PB |
6280 | struct cfs_rq *cfs_rq; |
6281 | unsigned int util; | |
104cb16d MR |
6282 | |
6283 | /* Task has no contribution or is new */ | |
f9be3e59 | 6284 | if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) |
104cb16d MR |
6285 | return cpu_util(cpu); |
6286 | ||
f9be3e59 PB |
6287 | cfs_rq = &cpu_rq(cpu)->cfs; |
6288 | util = READ_ONCE(cfs_rq->avg.util_avg); | |
6289 | ||
c469933e | 6290 | /* Discount task's util from CPU's util */ |
b5c0ce7b | 6291 | lsub_positive(&util, task_util(p)); |
104cb16d | 6292 | |
f9be3e59 PB |
6293 | /* |
6294 | * Covered cases: | |
6295 | * | |
6296 | * a) if *p is the only task sleeping on this CPU, then: | |
6297 | * cpu_util (== task_util) > util_est (== 0) | |
6298 | * and thus we return: | |
c469933e | 6299 | * cpu_util_without = (cpu_util - task_util) = 0 |
f9be3e59 PB |
6300 | * |
6301 | * b) if other tasks are SLEEPING on this CPU, which is now exiting | |
6302 | * IDLE, then: | |
6303 | * cpu_util >= task_util | |
6304 | * cpu_util > util_est (== 0) | |
6305 | * and thus we discount *p's blocked utilization to return: | |
c469933e | 6306 | * cpu_util_without = (cpu_util - task_util) >= 0 |
f9be3e59 PB |
6307 | * |
6308 | * c) if other tasks are RUNNABLE on that CPU and | |
6309 | * util_est > cpu_util | |
6310 | * then we use util_est since it returns a more restrictive | |
6311 | * estimation of the spare capacity on that CPU, by just | |
6312 | * considering the expected utilization of tasks already | |
6313 | * runnable on that CPU. | |
6314 | * | |
6315 | * Cases a) and b) are covered by the above code, while case c) is | |
6316 | * covered by the following code when estimated utilization is | |
6317 | * enabled. | |
6318 | */ | |
c469933e PB |
6319 | if (sched_feat(UTIL_EST)) { |
6320 | unsigned int estimated = | |
6321 | READ_ONCE(cfs_rq->avg.util_est.enqueued); | |
6322 | ||
6323 | /* | |
6324 | * Despite the following checks we still have a small window | |
6325 | * for a possible race, when an execl's select_task_rq_fair() | |
6326 | * races with LB's detach_task(): | |
6327 | * | |
6328 | * detach_task() | |
6329 | * p->on_rq = TASK_ON_RQ_MIGRATING; | |
6330 | * ---------------------------------- A | |
6331 | * deactivate_task() \ | |
6332 | * dequeue_task() + RaceTime | |
6333 | * util_est_dequeue() / | |
6334 | * ---------------------------------- B | |
6335 | * | |
6336 | * The additional check on "current == p" it's required to | |
6337 | * properly fix the execl regression and it helps in further | |
6338 | * reducing the chances for the above race. | |
6339 | */ | |
b5c0ce7b PB |
6340 | if (unlikely(task_on_rq_queued(p) || current == p)) |
6341 | lsub_positive(&estimated, _task_util_est(p)); | |
6342 | ||
c469933e PB |
6343 | util = max(util, estimated); |
6344 | } | |
f9be3e59 PB |
6345 | |
6346 | /* | |
6347 | * Utilization (estimated) can exceed the CPU capacity, thus let's | |
6348 | * clamp to the maximum CPU capacity to ensure consistency with | |
6349 | * the cpu_util call. | |
6350 | */ | |
6351 | return min_t(unsigned long, util, capacity_orig_of(cpu)); | |
104cb16d MR |
6352 | } |
6353 | ||
3273163c MR |
6354 | /* |
6355 | * Disable WAKE_AFFINE in the case where task @p doesn't fit in the | |
6356 | * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu. | |
6357 | * | |
6358 | * In that case WAKE_AFFINE doesn't make sense and we'll let | |
6359 | * BALANCE_WAKE sort things out. | |
6360 | */ | |
6361 | static int wake_cap(struct task_struct *p, int cpu, int prev_cpu) | |
6362 | { | |
6363 | long min_cap, max_cap; | |
6364 | ||
df054e84 MR |
6365 | if (!static_branch_unlikely(&sched_asym_cpucapacity)) |
6366 | return 0; | |
6367 | ||
3273163c MR |
6368 | min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu)); |
6369 | max_cap = cpu_rq(cpu)->rd->max_cpu_capacity; | |
6370 | ||
6371 | /* Minimum capacity is close to max, no need to abort wake_affine */ | |
6372 | if (max_cap - min_cap < max_cap >> 3) | |
6373 | return 0; | |
6374 | ||
104cb16d MR |
6375 | /* Bring task utilization in sync with prev_cpu */ |
6376 | sync_entity_load_avg(&p->se); | |
6377 | ||
3b1baa64 | 6378 | return !task_fits_capacity(p, min_cap); |
3273163c MR |
6379 | } |
6380 | ||
390031e4 QP |
6381 | /* |
6382 | * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued) | |
6383 | * to @dst_cpu. | |
6384 | */ | |
6385 | static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) | |
6386 | { | |
6387 | struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; | |
6388 | unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg); | |
6389 | ||
6390 | /* | |
6391 | * If @p migrates from @cpu to another, remove its contribution. Or, | |
6392 | * if @p migrates from another CPU to @cpu, add its contribution. In | |
6393 | * the other cases, @cpu is not impacted by the migration, so the | |
6394 | * util_avg should already be correct. | |
6395 | */ | |
6396 | if (task_cpu(p) == cpu && dst_cpu != cpu) | |
6397 | sub_positive(&util, task_util(p)); | |
6398 | else if (task_cpu(p) != cpu && dst_cpu == cpu) | |
6399 | util += task_util(p); | |
6400 | ||
6401 | if (sched_feat(UTIL_EST)) { | |
6402 | util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued); | |
6403 | ||
6404 | /* | |
6405 | * During wake-up, the task isn't enqueued yet and doesn't | |
6406 | * appear in the cfs_rq->avg.util_est.enqueued of any rq, | |
6407 | * so just add it (if needed) to "simulate" what will be | |
6408 | * cpu_util() after the task has been enqueued. | |
6409 | */ | |
6410 | if (dst_cpu == cpu) | |
6411 | util_est += _task_util_est(p); | |
6412 | ||
6413 | util = max(util, util_est); | |
6414 | } | |
6415 | ||
6416 | return min(util, capacity_orig_of(cpu)); | |
6417 | } | |
6418 | ||
6419 | /* | |
6420 | * compute_energy(): Estimates the energy that would be consumed if @p was | |
6421 | * migrated to @dst_cpu. compute_energy() predicts what will be the utilization | |
6422 | * landscape of the * CPUs after the task migration, and uses the Energy Model | |
6423 | * to compute what would be the energy if we decided to actually migrate that | |
6424 | * task. | |
6425 | */ | |
6426 | static long | |
6427 | compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) | |
6428 | { | |
6429 | long util, max_util, sum_util, energy = 0; | |
6430 | int cpu; | |
6431 | ||
6432 | for (; pd; pd = pd->next) { | |
6433 | max_util = sum_util = 0; | |
6434 | /* | |
6435 | * The capacity state of CPUs of the current rd can be driven by | |
6436 | * CPUs of another rd if they belong to the same performance | |
6437 | * domain. So, account for the utilization of these CPUs too | |
6438 | * by masking pd with cpu_online_mask instead of the rd span. | |
6439 | * | |
6440 | * If an entire performance domain is outside of the current rd, | |
6441 | * it will not appear in its pd list and will not be accounted | |
6442 | * by compute_energy(). | |
6443 | */ | |
6444 | for_each_cpu_and(cpu, perf_domain_span(pd), cpu_online_mask) { | |
6445 | util = cpu_util_next(cpu, p, dst_cpu); | |
6446 | util = schedutil_energy_util(cpu, util); | |
6447 | max_util = max(util, max_util); | |
6448 | sum_util += util; | |
6449 | } | |
6450 | ||
6451 | energy += em_pd_energy(pd->em_pd, max_util, sum_util); | |
6452 | } | |
6453 | ||
6454 | return energy; | |
6455 | } | |
6456 | ||
732cd75b QP |
6457 | /* |
6458 | * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the | |
6459 | * waking task. find_energy_efficient_cpu() looks for the CPU with maximum | |
6460 | * spare capacity in each performance domain and uses it as a potential | |
6461 | * candidate to execute the task. Then, it uses the Energy Model to figure | |
6462 | * out which of the CPU candidates is the most energy-efficient. | |
6463 | * | |
6464 | * The rationale for this heuristic is as follows. In a performance domain, | |
6465 | * all the most energy efficient CPU candidates (according to the Energy | |
6466 | * Model) are those for which we'll request a low frequency. When there are | |
6467 | * several CPUs for which the frequency request will be the same, we don't | |
6468 | * have enough data to break the tie between them, because the Energy Model | |
6469 | * only includes active power costs. With this model, if we assume that | |
6470 | * frequency requests follow utilization (e.g. using schedutil), the CPU with | |
6471 | * the maximum spare capacity in a performance domain is guaranteed to be among | |
6472 | * the best candidates of the performance domain. | |
6473 | * | |
6474 | * In practice, it could be preferable from an energy standpoint to pack | |
6475 | * small tasks on a CPU in order to let other CPUs go in deeper idle states, | |
6476 | * but that could also hurt our chances to go cluster idle, and we have no | |
6477 | * ways to tell with the current Energy Model if this is actually a good | |
6478 | * idea or not. So, find_energy_efficient_cpu() basically favors | |
6479 | * cluster-packing, and spreading inside a cluster. That should at least be | |
6480 | * a good thing for latency, and this is consistent with the idea that most | |
6481 | * of the energy savings of EAS come from the asymmetry of the system, and | |
6482 | * not so much from breaking the tie between identical CPUs. That's also the | |
6483 | * reason why EAS is enabled in the topology code only for systems where | |
6484 | * SD_ASYM_CPUCAPACITY is set. | |
6485 | * | |
6486 | * NOTE: Forkees are not accepted in the energy-aware wake-up path because | |
6487 | * they don't have any useful utilization data yet and it's not possible to | |
6488 | * forecast their impact on energy consumption. Consequently, they will be | |
6489 | * placed by find_idlest_cpu() on the least loaded CPU, which might turn out | |
6490 | * to be energy-inefficient in some use-cases. The alternative would be to | |
6491 | * bias new tasks towards specific types of CPUs first, or to try to infer | |
6492 | * their util_avg from the parent task, but those heuristics could hurt | |
6493 | * other use-cases too. So, until someone finds a better way to solve this, | |
6494 | * let's keep things simple by re-using the existing slow path. | |
6495 | */ | |
6496 | ||
6497 | static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) | |
6498 | { | |
6499 | unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX; | |
6500 | struct root_domain *rd = cpu_rq(smp_processor_id())->rd; | |
6501 | int cpu, best_energy_cpu = prev_cpu; | |
6502 | struct perf_domain *head, *pd; | |
6503 | unsigned long cpu_cap, util; | |
6504 | struct sched_domain *sd; | |
6505 | ||
6506 | rcu_read_lock(); | |
6507 | pd = rcu_dereference(rd->pd); | |
6508 | if (!pd || READ_ONCE(rd->overutilized)) | |
6509 | goto fail; | |
6510 | head = pd; | |
6511 | ||
6512 | /* | |
6513 | * Energy-aware wake-up happens on the lowest sched_domain starting | |
6514 | * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. | |
6515 | */ | |
6516 | sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); | |
6517 | while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) | |
6518 | sd = sd->parent; | |
6519 | if (!sd) | |
6520 | goto fail; | |
6521 | ||
6522 | sync_entity_load_avg(&p->se); | |
6523 | if (!task_util_est(p)) | |
6524 | goto unlock; | |
6525 | ||
6526 | for (; pd; pd = pd->next) { | |
6527 | unsigned long cur_energy, spare_cap, max_spare_cap = 0; | |
6528 | int max_spare_cap_cpu = -1; | |
6529 | ||
6530 | for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) { | |
6531 | if (!cpumask_test_cpu(cpu, &p->cpus_allowed)) | |
6532 | continue; | |
6533 | ||
6534 | /* Skip CPUs that will be overutilized. */ | |
6535 | util = cpu_util_next(cpu, p, cpu); | |
6536 | cpu_cap = capacity_of(cpu); | |
6537 | if (cpu_cap * 1024 < util * capacity_margin) | |
6538 | continue; | |
6539 | ||
6540 | /* Always use prev_cpu as a candidate. */ | |
6541 | if (cpu == prev_cpu) { | |
6542 | prev_energy = compute_energy(p, prev_cpu, head); | |
6543 | best_energy = min(best_energy, prev_energy); | |
6544 | continue; | |
6545 | } | |
6546 | ||
6547 | /* | |
6548 | * Find the CPU with the maximum spare capacity in | |
6549 | * the performance domain | |
6550 | */ | |
6551 | spare_cap = cpu_cap - util; | |
6552 | if (spare_cap > max_spare_cap) { | |
6553 | max_spare_cap = spare_cap; | |
6554 | max_spare_cap_cpu = cpu; | |
6555 | } | |
6556 | } | |
6557 | ||
6558 | /* Evaluate the energy impact of using this CPU. */ | |
6559 | if (max_spare_cap_cpu >= 0) { | |
6560 | cur_energy = compute_energy(p, max_spare_cap_cpu, head); | |
6561 | if (cur_energy < best_energy) { | |
6562 | best_energy = cur_energy; | |
6563 | best_energy_cpu = max_spare_cap_cpu; | |
6564 | } | |
6565 | } | |
6566 | } | |
6567 | unlock: | |
6568 | rcu_read_unlock(); | |
6569 | ||
6570 | /* | |
6571 | * Pick the best CPU if prev_cpu cannot be used, or if it saves at | |
6572 | * least 6% of the energy used by prev_cpu. | |
6573 | */ | |
6574 | if (prev_energy == ULONG_MAX) | |
6575 | return best_energy_cpu; | |
6576 | ||
6577 | if ((prev_energy - best_energy) > (prev_energy >> 4)) | |
6578 | return best_energy_cpu; | |
6579 | ||
6580 | return prev_cpu; | |
6581 | ||
6582 | fail: | |
6583 | rcu_read_unlock(); | |
6584 | ||
6585 | return -1; | |
6586 | } | |
6587 | ||
aaee1203 | 6588 | /* |
de91b9cb MR |
6589 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
6590 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, | |
6591 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. | |
aaee1203 | 6592 | * |
97fb7a0a IM |
6593 | * Balances load by selecting the idlest CPU in the idlest group, or under |
6594 | * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 6595 | * |
97fb7a0a | 6596 | * Returns the target CPU number. |
aaee1203 PZ |
6597 | * |
6598 | * preempt must be disabled. | |
6599 | */ | |
0017d735 | 6600 | static int |
ac66f547 | 6601 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 6602 | { |
f1d88b44 | 6603 | struct sched_domain *tmp, *sd = NULL; |
c88d5910 | 6604 | int cpu = smp_processor_id(); |
63b0e9ed | 6605 | int new_cpu = prev_cpu; |
99bd5e2f | 6606 | int want_affine = 0; |
24d0c1d6 | 6607 | int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); |
c88d5910 | 6608 | |
c58d25f3 PZ |
6609 | if (sd_flag & SD_BALANCE_WAKE) { |
6610 | record_wakee(p); | |
732cd75b | 6611 | |
f8a696f2 | 6612 | if (sched_energy_enabled()) { |
732cd75b QP |
6613 | new_cpu = find_energy_efficient_cpu(p, prev_cpu); |
6614 | if (new_cpu >= 0) | |
6615 | return new_cpu; | |
6616 | new_cpu = prev_cpu; | |
6617 | } | |
6618 | ||
6619 | want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) && | |
6620 | cpumask_test_cpu(cpu, &p->cpus_allowed); | |
c58d25f3 | 6621 | } |
aaee1203 | 6622 | |
dce840a0 | 6623 | rcu_read_lock(); |
aaee1203 | 6624 | for_each_domain(cpu, tmp) { |
e4f42888 | 6625 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
63b0e9ed | 6626 | break; |
e4f42888 | 6627 | |
fe3bcfe1 | 6628 | /* |
97fb7a0a | 6629 | * If both 'cpu' and 'prev_cpu' are part of this domain, |
99bd5e2f | 6630 | * cpu is a valid SD_WAKE_AFFINE target. |
fe3bcfe1 | 6631 | */ |
99bd5e2f SS |
6632 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
6633 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
f1d88b44 VK |
6634 | if (cpu != prev_cpu) |
6635 | new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync); | |
6636 | ||
6637 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
29cd8bae | 6638 | break; |
f03542a7 | 6639 | } |
29cd8bae | 6640 | |
f03542a7 | 6641 | if (tmp->flags & sd_flag) |
29cd8bae | 6642 | sd = tmp; |
63b0e9ed MG |
6643 | else if (!want_affine) |
6644 | break; | |
29cd8bae PZ |
6645 | } |
6646 | ||
f1d88b44 VK |
6647 | if (unlikely(sd)) { |
6648 | /* Slow path */ | |
18bd1b4b | 6649 | new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag); |
f1d88b44 VK |
6650 | } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */ |
6651 | /* Fast path */ | |
6652 | ||
6653 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); | |
6654 | ||
6655 | if (want_affine) | |
6656 | current->recent_used_cpu = cpu; | |
e7693a36 | 6657 | } |
dce840a0 | 6658 | rcu_read_unlock(); |
e7693a36 | 6659 | |
c88d5910 | 6660 | return new_cpu; |
e7693a36 | 6661 | } |
0a74bef8 | 6662 | |
144d8487 PZ |
6663 | static void detach_entity_cfs_rq(struct sched_entity *se); |
6664 | ||
0a74bef8 | 6665 | /* |
97fb7a0a | 6666 | * Called immediately before a task is migrated to a new CPU; task_cpu(p) and |
0a74bef8 | 6667 | * cfs_rq_of(p) references at time of call are still valid and identify the |
97fb7a0a | 6668 | * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 6669 | */ |
3f9672ba | 6670 | static void migrate_task_rq_fair(struct task_struct *p, int new_cpu) |
0a74bef8 | 6671 | { |
59efa0ba PZ |
6672 | /* |
6673 | * As blocked tasks retain absolute vruntime the migration needs to | |
6674 | * deal with this by subtracting the old and adding the new | |
6675 | * min_vruntime -- the latter is done by enqueue_entity() when placing | |
6676 | * the task on the new runqueue. | |
6677 | */ | |
6678 | if (p->state == TASK_WAKING) { | |
6679 | struct sched_entity *se = &p->se; | |
6680 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6681 | u64 min_vruntime; | |
6682 | ||
6683 | #ifndef CONFIG_64BIT | |
6684 | u64 min_vruntime_copy; | |
6685 | ||
6686 | do { | |
6687 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
6688 | smp_rmb(); | |
6689 | min_vruntime = cfs_rq->min_vruntime; | |
6690 | } while (min_vruntime != min_vruntime_copy); | |
6691 | #else | |
6692 | min_vruntime = cfs_rq->min_vruntime; | |
6693 | #endif | |
6694 | ||
6695 | se->vruntime -= min_vruntime; | |
6696 | } | |
6697 | ||
144d8487 PZ |
6698 | if (p->on_rq == TASK_ON_RQ_MIGRATING) { |
6699 | /* | |
6700 | * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old' | |
6701 | * rq->lock and can modify state directly. | |
6702 | */ | |
6703 | lockdep_assert_held(&task_rq(p)->lock); | |
6704 | detach_entity_cfs_rq(&p->se); | |
6705 | ||
6706 | } else { | |
6707 | /* | |
6708 | * We are supposed to update the task to "current" time, then | |
6709 | * its up to date and ready to go to new CPU/cfs_rq. But we | |
6710 | * have difficulty in getting what current time is, so simply | |
6711 | * throw away the out-of-date time. This will result in the | |
6712 | * wakee task is less decayed, but giving the wakee more load | |
6713 | * sounds not bad. | |
6714 | */ | |
6715 | remove_entity_load_avg(&p->se); | |
6716 | } | |
9d89c257 YD |
6717 | |
6718 | /* Tell new CPU we are migrated */ | |
6719 | p->se.avg.last_update_time = 0; | |
3944a927 BS |
6720 | |
6721 | /* We have migrated, no longer consider this task hot */ | |
9d89c257 | 6722 | p->se.exec_start = 0; |
3f9672ba SD |
6723 | |
6724 | update_scan_period(p, new_cpu); | |
0a74bef8 | 6725 | } |
12695578 YD |
6726 | |
6727 | static void task_dead_fair(struct task_struct *p) | |
6728 | { | |
6729 | remove_entity_load_avg(&p->se); | |
6730 | } | |
e7693a36 GH |
6731 | #endif /* CONFIG_SMP */ |
6732 | ||
a555e9d8 | 6733 | static unsigned long wakeup_gran(struct sched_entity *se) |
0bbd3336 PZ |
6734 | { |
6735 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
6736 | ||
6737 | /* | |
e52fb7c0 PZ |
6738 | * Since its curr running now, convert the gran from real-time |
6739 | * to virtual-time in his units. | |
13814d42 MG |
6740 | * |
6741 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
6742 | * they get preempted easier. That is, if 'se' < 'curr' then | |
6743 | * the resulting gran will be larger, therefore penalizing the | |
6744 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
6745 | * be smaller, again penalizing the lighter task. | |
6746 | * | |
6747 | * This is especially important for buddies when the leftmost | |
6748 | * task is higher priority than the buddy. | |
0bbd3336 | 6749 | */ |
f4ad9bd2 | 6750 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
6751 | } |
6752 | ||
464b7527 PZ |
6753 | /* |
6754 | * Should 'se' preempt 'curr'. | |
6755 | * | |
6756 | * |s1 | |
6757 | * |s2 | |
6758 | * |s3 | |
6759 | * g | |
6760 | * |<--->|c | |
6761 | * | |
6762 | * w(c, s1) = -1 | |
6763 | * w(c, s2) = 0 | |
6764 | * w(c, s3) = 1 | |
6765 | * | |
6766 | */ | |
6767 | static int | |
6768 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
6769 | { | |
6770 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
6771 | ||
6772 | if (vdiff <= 0) | |
6773 | return -1; | |
6774 | ||
a555e9d8 | 6775 | gran = wakeup_gran(se); |
464b7527 PZ |
6776 | if (vdiff > gran) |
6777 | return 1; | |
6778 | ||
6779 | return 0; | |
6780 | } | |
6781 | ||
02479099 PZ |
6782 | static void set_last_buddy(struct sched_entity *se) |
6783 | { | |
1da1843f | 6784 | if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) |
69c80f3e VP |
6785 | return; |
6786 | ||
c5ae366e DA |
6787 | for_each_sched_entity(se) { |
6788 | if (SCHED_WARN_ON(!se->on_rq)) | |
6789 | return; | |
69c80f3e | 6790 | cfs_rq_of(se)->last = se; |
c5ae366e | 6791 | } |
02479099 PZ |
6792 | } |
6793 | ||
6794 | static void set_next_buddy(struct sched_entity *se) | |
6795 | { | |
1da1843f | 6796 | if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) |
69c80f3e VP |
6797 | return; |
6798 | ||
c5ae366e DA |
6799 | for_each_sched_entity(se) { |
6800 | if (SCHED_WARN_ON(!se->on_rq)) | |
6801 | return; | |
69c80f3e | 6802 | cfs_rq_of(se)->next = se; |
c5ae366e | 6803 | } |
02479099 PZ |
6804 | } |
6805 | ||
ac53db59 RR |
6806 | static void set_skip_buddy(struct sched_entity *se) |
6807 | { | |
69c80f3e VP |
6808 | for_each_sched_entity(se) |
6809 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
6810 | } |
6811 | ||
bf0f6f24 IM |
6812 | /* |
6813 | * Preempt the current task with a newly woken task if needed: | |
6814 | */ | |
5a9b86f6 | 6815 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
6816 | { |
6817 | struct task_struct *curr = rq->curr; | |
8651a86c | 6818 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 6819 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 6820 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 6821 | int next_buddy_marked = 0; |
bf0f6f24 | 6822 | |
4ae7d5ce IM |
6823 | if (unlikely(se == pse)) |
6824 | return; | |
6825 | ||
5238cdd3 | 6826 | /* |
163122b7 | 6827 | * This is possible from callers such as attach_tasks(), in which we |
5238cdd3 PT |
6828 | * unconditionally check_prempt_curr() after an enqueue (which may have |
6829 | * lead to a throttle). This both saves work and prevents false | |
6830 | * next-buddy nomination below. | |
6831 | */ | |
6832 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
6833 | return; | |
6834 | ||
2f36825b | 6835 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 6836 | set_next_buddy(pse); |
2f36825b VP |
6837 | next_buddy_marked = 1; |
6838 | } | |
57fdc26d | 6839 | |
aec0a514 BR |
6840 | /* |
6841 | * We can come here with TIF_NEED_RESCHED already set from new task | |
6842 | * wake up path. | |
5238cdd3 PT |
6843 | * |
6844 | * Note: this also catches the edge-case of curr being in a throttled | |
6845 | * group (e.g. via set_curr_task), since update_curr() (in the | |
6846 | * enqueue of curr) will have resulted in resched being set. This | |
6847 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
6848 | * below. | |
aec0a514 BR |
6849 | */ |
6850 | if (test_tsk_need_resched(curr)) | |
6851 | return; | |
6852 | ||
a2f5c9ab | 6853 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
1da1843f VK |
6854 | if (unlikely(task_has_idle_policy(curr)) && |
6855 | likely(!task_has_idle_policy(p))) | |
a2f5c9ab DH |
6856 | goto preempt; |
6857 | ||
91c234b4 | 6858 | /* |
a2f5c9ab DH |
6859 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
6860 | * is driven by the tick): | |
91c234b4 | 6861 | */ |
8ed92e51 | 6862 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 6863 | return; |
bf0f6f24 | 6864 | |
464b7527 | 6865 | find_matching_se(&se, &pse); |
9bbd7374 | 6866 | update_curr(cfs_rq_of(se)); |
002f128b | 6867 | BUG_ON(!pse); |
2f36825b VP |
6868 | if (wakeup_preempt_entity(se, pse) == 1) { |
6869 | /* | |
6870 | * Bias pick_next to pick the sched entity that is | |
6871 | * triggering this preemption. | |
6872 | */ | |
6873 | if (!next_buddy_marked) | |
6874 | set_next_buddy(pse); | |
3a7e73a2 | 6875 | goto preempt; |
2f36825b | 6876 | } |
464b7527 | 6877 | |
3a7e73a2 | 6878 | return; |
a65ac745 | 6879 | |
3a7e73a2 | 6880 | preempt: |
8875125e | 6881 | resched_curr(rq); |
3a7e73a2 PZ |
6882 | /* |
6883 | * Only set the backward buddy when the current task is still | |
6884 | * on the rq. This can happen when a wakeup gets interleaved | |
6885 | * with schedule on the ->pre_schedule() or idle_balance() | |
6886 | * point, either of which can * drop the rq lock. | |
6887 | * | |
6888 | * Also, during early boot the idle thread is in the fair class, | |
6889 | * for obvious reasons its a bad idea to schedule back to it. | |
6890 | */ | |
6891 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
6892 | return; | |
6893 | ||
6894 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
6895 | set_last_buddy(se); | |
bf0f6f24 IM |
6896 | } |
6897 | ||
606dba2e | 6898 | static struct task_struct * |
d8ac8971 | 6899 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
bf0f6f24 IM |
6900 | { |
6901 | struct cfs_rq *cfs_rq = &rq->cfs; | |
6902 | struct sched_entity *se; | |
678d5718 | 6903 | struct task_struct *p; |
37e117c0 | 6904 | int new_tasks; |
678d5718 | 6905 | |
6e83125c | 6906 | again: |
678d5718 | 6907 | if (!cfs_rq->nr_running) |
38033c37 | 6908 | goto idle; |
678d5718 | 6909 | |
9674f5ca | 6910 | #ifdef CONFIG_FAIR_GROUP_SCHED |
3f1d2a31 | 6911 | if (prev->sched_class != &fair_sched_class) |
678d5718 PZ |
6912 | goto simple; |
6913 | ||
6914 | /* | |
6915 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
6916 | * likely that a next task is from the same cgroup as the current. | |
6917 | * | |
6918 | * Therefore attempt to avoid putting and setting the entire cgroup | |
6919 | * hierarchy, only change the part that actually changes. | |
6920 | */ | |
6921 | ||
6922 | do { | |
6923 | struct sched_entity *curr = cfs_rq->curr; | |
6924 | ||
6925 | /* | |
6926 | * Since we got here without doing put_prev_entity() we also | |
6927 | * have to consider cfs_rq->curr. If it is still a runnable | |
6928 | * entity, update_curr() will update its vruntime, otherwise | |
6929 | * forget we've ever seen it. | |
6930 | */ | |
54d27365 BS |
6931 | if (curr) { |
6932 | if (curr->on_rq) | |
6933 | update_curr(cfs_rq); | |
6934 | else | |
6935 | curr = NULL; | |
678d5718 | 6936 | |
54d27365 BS |
6937 | /* |
6938 | * This call to check_cfs_rq_runtime() will do the | |
6939 | * throttle and dequeue its entity in the parent(s). | |
9674f5ca | 6940 | * Therefore the nr_running test will indeed |
54d27365 BS |
6941 | * be correct. |
6942 | */ | |
9674f5ca VK |
6943 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) { |
6944 | cfs_rq = &rq->cfs; | |
6945 | ||
6946 | if (!cfs_rq->nr_running) | |
6947 | goto idle; | |
6948 | ||
54d27365 | 6949 | goto simple; |
9674f5ca | 6950 | } |
54d27365 | 6951 | } |
678d5718 PZ |
6952 | |
6953 | se = pick_next_entity(cfs_rq, curr); | |
6954 | cfs_rq = group_cfs_rq(se); | |
6955 | } while (cfs_rq); | |
6956 | ||
6957 | p = task_of(se); | |
6958 | ||
6959 | /* | |
6960 | * Since we haven't yet done put_prev_entity and if the selected task | |
6961 | * is a different task than we started out with, try and touch the | |
6962 | * least amount of cfs_rqs. | |
6963 | */ | |
6964 | if (prev != p) { | |
6965 | struct sched_entity *pse = &prev->se; | |
6966 | ||
6967 | while (!(cfs_rq = is_same_group(se, pse))) { | |
6968 | int se_depth = se->depth; | |
6969 | int pse_depth = pse->depth; | |
6970 | ||
6971 | if (se_depth <= pse_depth) { | |
6972 | put_prev_entity(cfs_rq_of(pse), pse); | |
6973 | pse = parent_entity(pse); | |
6974 | } | |
6975 | if (se_depth >= pse_depth) { | |
6976 | set_next_entity(cfs_rq_of(se), se); | |
6977 | se = parent_entity(se); | |
6978 | } | |
6979 | } | |
6980 | ||
6981 | put_prev_entity(cfs_rq, pse); | |
6982 | set_next_entity(cfs_rq, se); | |
6983 | } | |
6984 | ||
93824900 | 6985 | goto done; |
678d5718 | 6986 | simple: |
678d5718 | 6987 | #endif |
bf0f6f24 | 6988 | |
3f1d2a31 | 6989 | put_prev_task(rq, prev); |
606dba2e | 6990 | |
bf0f6f24 | 6991 | do { |
678d5718 | 6992 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 6993 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
6994 | cfs_rq = group_cfs_rq(se); |
6995 | } while (cfs_rq); | |
6996 | ||
8f4d37ec | 6997 | p = task_of(se); |
678d5718 | 6998 | |
13a453c2 | 6999 | done: __maybe_unused; |
93824900 UR |
7000 | #ifdef CONFIG_SMP |
7001 | /* | |
7002 | * Move the next running task to the front of | |
7003 | * the list, so our cfs_tasks list becomes MRU | |
7004 | * one. | |
7005 | */ | |
7006 | list_move(&p->se.group_node, &rq->cfs_tasks); | |
7007 | #endif | |
7008 | ||
b39e66ea MG |
7009 | if (hrtick_enabled(rq)) |
7010 | hrtick_start_fair(rq, p); | |
8f4d37ec | 7011 | |
3b1baa64 MR |
7012 | update_misfit_status(p, rq); |
7013 | ||
8f4d37ec | 7014 | return p; |
38033c37 PZ |
7015 | |
7016 | idle: | |
3b1baa64 | 7017 | update_misfit_status(NULL, rq); |
46f69fa3 MF |
7018 | new_tasks = idle_balance(rq, rf); |
7019 | ||
37e117c0 PZ |
7020 | /* |
7021 | * Because idle_balance() releases (and re-acquires) rq->lock, it is | |
7022 | * possible for any higher priority task to appear. In that case we | |
7023 | * must re-start the pick_next_entity() loop. | |
7024 | */ | |
e4aa358b | 7025 | if (new_tasks < 0) |
37e117c0 PZ |
7026 | return RETRY_TASK; |
7027 | ||
e4aa358b | 7028 | if (new_tasks > 0) |
38033c37 | 7029 | goto again; |
38033c37 PZ |
7030 | |
7031 | return NULL; | |
bf0f6f24 IM |
7032 | } |
7033 | ||
7034 | /* | |
7035 | * Account for a descheduled task: | |
7036 | */ | |
31ee529c | 7037 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
7038 | { |
7039 | struct sched_entity *se = &prev->se; | |
7040 | struct cfs_rq *cfs_rq; | |
7041 | ||
7042 | for_each_sched_entity(se) { | |
7043 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 7044 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
7045 | } |
7046 | } | |
7047 | ||
ac53db59 RR |
7048 | /* |
7049 | * sched_yield() is very simple | |
7050 | * | |
7051 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
7052 | */ | |
7053 | static void yield_task_fair(struct rq *rq) | |
7054 | { | |
7055 | struct task_struct *curr = rq->curr; | |
7056 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
7057 | struct sched_entity *se = &curr->se; | |
7058 | ||
7059 | /* | |
7060 | * Are we the only task in the tree? | |
7061 | */ | |
7062 | if (unlikely(rq->nr_running == 1)) | |
7063 | return; | |
7064 | ||
7065 | clear_buddies(cfs_rq, se); | |
7066 | ||
7067 | if (curr->policy != SCHED_BATCH) { | |
7068 | update_rq_clock(rq); | |
7069 | /* | |
7070 | * Update run-time statistics of the 'current'. | |
7071 | */ | |
7072 | update_curr(cfs_rq); | |
916671c0 MG |
7073 | /* |
7074 | * Tell update_rq_clock() that we've just updated, | |
7075 | * so we don't do microscopic update in schedule() | |
7076 | * and double the fastpath cost. | |
7077 | */ | |
adcc8da8 | 7078 | rq_clock_skip_update(rq); |
ac53db59 RR |
7079 | } |
7080 | ||
7081 | set_skip_buddy(se); | |
7082 | } | |
7083 | ||
d95f4122 MG |
7084 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
7085 | { | |
7086 | struct sched_entity *se = &p->se; | |
7087 | ||
5238cdd3 PT |
7088 | /* throttled hierarchies are not runnable */ |
7089 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
7090 | return false; |
7091 | ||
7092 | /* Tell the scheduler that we'd really like pse to run next. */ | |
7093 | set_next_buddy(se); | |
7094 | ||
d95f4122 MG |
7095 | yield_task_fair(rq); |
7096 | ||
7097 | return true; | |
7098 | } | |
7099 | ||
681f3e68 | 7100 | #ifdef CONFIG_SMP |
bf0f6f24 | 7101 | /************************************************** |
e9c84cb8 PZ |
7102 | * Fair scheduling class load-balancing methods. |
7103 | * | |
7104 | * BASICS | |
7105 | * | |
7106 | * The purpose of load-balancing is to achieve the same basic fairness the | |
97fb7a0a | 7107 | * per-CPU scheduler provides, namely provide a proportional amount of compute |
e9c84cb8 PZ |
7108 | * time to each task. This is expressed in the following equation: |
7109 | * | |
7110 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
7111 | * | |
97fb7a0a | 7112 | * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight |
e9c84cb8 PZ |
7113 | * W_i,0 is defined as: |
7114 | * | |
7115 | * W_i,0 = \Sum_j w_i,j (2) | |
7116 | * | |
97fb7a0a | 7117 | * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight |
1c3de5e1 | 7118 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
7119 | * |
7120 | * The weight average is an exponential decay average of the instantaneous | |
7121 | * weight: | |
7122 | * | |
7123 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
7124 | * | |
97fb7a0a | 7125 | * C_i is the compute capacity of CPU i, typically it is the |
e9c84cb8 PZ |
7126 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
7127 | * can also include other factors [XXX]. | |
7128 | * | |
7129 | * To achieve this balance we define a measure of imbalance which follows | |
7130 | * directly from (1): | |
7131 | * | |
ced549fa | 7132 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
7133 | * |
7134 | * We them move tasks around to minimize the imbalance. In the continuous | |
7135 | * function space it is obvious this converges, in the discrete case we get | |
7136 | * a few fun cases generally called infeasible weight scenarios. | |
7137 | * | |
7138 | * [XXX expand on: | |
7139 | * - infeasible weights; | |
7140 | * - local vs global optima in the discrete case. ] | |
7141 | * | |
7142 | * | |
7143 | * SCHED DOMAINS | |
7144 | * | |
7145 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
97fb7a0a | 7146 | * for all i,j solution, we create a tree of CPUs that follows the hardware |
e9c84cb8 | 7147 | * topology where each level pairs two lower groups (or better). This results |
97fb7a0a | 7148 | * in O(log n) layers. Furthermore we reduce the number of CPUs going up the |
e9c84cb8 | 7149 | * tree to only the first of the previous level and we decrease the frequency |
97fb7a0a | 7150 | * of load-balance at each level inv. proportional to the number of CPUs in |
e9c84cb8 PZ |
7151 | * the groups. |
7152 | * | |
7153 | * This yields: | |
7154 | * | |
7155 | * log_2 n 1 n | |
7156 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
7157 | * i = 0 2^i 2^i | |
7158 | * `- size of each group | |
97fb7a0a | 7159 | * | | `- number of CPUs doing load-balance |
e9c84cb8 PZ |
7160 | * | `- freq |
7161 | * `- sum over all levels | |
7162 | * | |
7163 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
7164 | * this makes (5) the runtime complexity of the balancer. | |
7165 | * | |
7166 | * An important property here is that each CPU is still (indirectly) connected | |
97fb7a0a | 7167 | * to every other CPU in at most O(log n) steps: |
e9c84cb8 PZ |
7168 | * |
7169 | * The adjacency matrix of the resulting graph is given by: | |
7170 | * | |
97a7142f | 7171 | * log_2 n |
e9c84cb8 PZ |
7172 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
7173 | * k = 0 | |
7174 | * | |
7175 | * And you'll find that: | |
7176 | * | |
7177 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
7178 | * | |
97fb7a0a | 7179 | * Showing there's indeed a path between every CPU in at most O(log n) steps. |
e9c84cb8 PZ |
7180 | * The task movement gives a factor of O(m), giving a convergence complexity |
7181 | * of: | |
7182 | * | |
7183 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
7184 | * | |
7185 | * | |
7186 | * WORK CONSERVING | |
7187 | * | |
7188 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
97fb7a0a | 7189 | * balancing is more aggressive and has the newly idle CPU iterate up the domain |
e9c84cb8 PZ |
7190 | * tree itself instead of relying on other CPUs to bring it work. |
7191 | * | |
7192 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
7193 | * time. | |
7194 | * | |
7195 | * [XXX more?] | |
7196 | * | |
7197 | * | |
7198 | * CGROUPS | |
7199 | * | |
7200 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
7201 | * | |
7202 | * s_k,i | |
7203 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
7204 | * S_k | |
7205 | * | |
7206 | * Where | |
7207 | * | |
7208 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
7209 | * | |
97fb7a0a | 7210 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i. |
e9c84cb8 PZ |
7211 | * |
7212 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
7213 | * property. | |
7214 | * | |
7215 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
7216 | * rewrite all of this once again.] | |
97a7142f | 7217 | */ |
bf0f6f24 | 7218 | |
ed387b78 HS |
7219 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
7220 | ||
0ec8aa00 PZ |
7221 | enum fbq_type { regular, remote, all }; |
7222 | ||
3b1baa64 MR |
7223 | enum group_type { |
7224 | group_other = 0, | |
7225 | group_misfit_task, | |
7226 | group_imbalanced, | |
7227 | group_overloaded, | |
7228 | }; | |
7229 | ||
ddcdf6e7 | 7230 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 7231 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
7232 | #define LBF_DST_PINNED 0x04 |
7233 | #define LBF_SOME_PINNED 0x08 | |
e022e0d3 | 7234 | #define LBF_NOHZ_STATS 0x10 |
f643ea22 | 7235 | #define LBF_NOHZ_AGAIN 0x20 |
ddcdf6e7 PZ |
7236 | |
7237 | struct lb_env { | |
7238 | struct sched_domain *sd; | |
7239 | ||
ddcdf6e7 | 7240 | struct rq *src_rq; |
85c1e7da | 7241 | int src_cpu; |
ddcdf6e7 PZ |
7242 | |
7243 | int dst_cpu; | |
7244 | struct rq *dst_rq; | |
7245 | ||
88b8dac0 SV |
7246 | struct cpumask *dst_grpmask; |
7247 | int new_dst_cpu; | |
ddcdf6e7 | 7248 | enum cpu_idle_type idle; |
bd939f45 | 7249 | long imbalance; |
b9403130 MW |
7250 | /* The set of CPUs under consideration for load-balancing */ |
7251 | struct cpumask *cpus; | |
7252 | ||
ddcdf6e7 | 7253 | unsigned int flags; |
367456c7 PZ |
7254 | |
7255 | unsigned int loop; | |
7256 | unsigned int loop_break; | |
7257 | unsigned int loop_max; | |
0ec8aa00 PZ |
7258 | |
7259 | enum fbq_type fbq_type; | |
cad68e55 | 7260 | enum group_type src_grp_type; |
163122b7 | 7261 | struct list_head tasks; |
ddcdf6e7 PZ |
7262 | }; |
7263 | ||
029632fb PZ |
7264 | /* |
7265 | * Is this task likely cache-hot: | |
7266 | */ | |
5d5e2b1b | 7267 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
7268 | { |
7269 | s64 delta; | |
7270 | ||
e5673f28 KT |
7271 | lockdep_assert_held(&env->src_rq->lock); |
7272 | ||
029632fb PZ |
7273 | if (p->sched_class != &fair_sched_class) |
7274 | return 0; | |
7275 | ||
1da1843f | 7276 | if (unlikely(task_has_idle_policy(p))) |
029632fb PZ |
7277 | return 0; |
7278 | ||
7279 | /* | |
7280 | * Buddy candidates are cache hot: | |
7281 | */ | |
5d5e2b1b | 7282 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
7283 | (&p->se == cfs_rq_of(&p->se)->next || |
7284 | &p->se == cfs_rq_of(&p->se)->last)) | |
7285 | return 1; | |
7286 | ||
7287 | if (sysctl_sched_migration_cost == -1) | |
7288 | return 1; | |
7289 | if (sysctl_sched_migration_cost == 0) | |
7290 | return 0; | |
7291 | ||
5d5e2b1b | 7292 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
7293 | |
7294 | return delta < (s64)sysctl_sched_migration_cost; | |
7295 | } | |
7296 | ||
3a7053b3 | 7297 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 7298 | /* |
2a1ed24c SD |
7299 | * Returns 1, if task migration degrades locality |
7300 | * Returns 0, if task migration improves locality i.e migration preferred. | |
7301 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 7302 | */ |
2a1ed24c | 7303 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 7304 | { |
b1ad065e | 7305 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
f35678b6 SD |
7306 | unsigned long src_weight, dst_weight; |
7307 | int src_nid, dst_nid, dist; | |
3a7053b3 | 7308 | |
2a595721 | 7309 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
7310 | return -1; |
7311 | ||
c3b9bc5b | 7312 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 7313 | return -1; |
7a0f3083 MG |
7314 | |
7315 | src_nid = cpu_to_node(env->src_cpu); | |
7316 | dst_nid = cpu_to_node(env->dst_cpu); | |
7317 | ||
83e1d2cd | 7318 | if (src_nid == dst_nid) |
2a1ed24c | 7319 | return -1; |
7a0f3083 | 7320 | |
2a1ed24c SD |
7321 | /* Migrating away from the preferred node is always bad. */ |
7322 | if (src_nid == p->numa_preferred_nid) { | |
7323 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
7324 | return 1; | |
7325 | else | |
7326 | return -1; | |
7327 | } | |
b1ad065e | 7328 | |
c1ceac62 RR |
7329 | /* Encourage migration to the preferred node. */ |
7330 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 7331 | return 0; |
b1ad065e | 7332 | |
739294fb | 7333 | /* Leaving a core idle is often worse than degrading locality. */ |
f35678b6 | 7334 | if (env->idle == CPU_IDLE) |
739294fb RR |
7335 | return -1; |
7336 | ||
f35678b6 | 7337 | dist = node_distance(src_nid, dst_nid); |
c1ceac62 | 7338 | if (numa_group) { |
f35678b6 SD |
7339 | src_weight = group_weight(p, src_nid, dist); |
7340 | dst_weight = group_weight(p, dst_nid, dist); | |
c1ceac62 | 7341 | } else { |
f35678b6 SD |
7342 | src_weight = task_weight(p, src_nid, dist); |
7343 | dst_weight = task_weight(p, dst_nid, dist); | |
b1ad065e RR |
7344 | } |
7345 | ||
f35678b6 | 7346 | return dst_weight < src_weight; |
7a0f3083 MG |
7347 | } |
7348 | ||
3a7053b3 | 7349 | #else |
2a1ed24c | 7350 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
7351 | struct lb_env *env) |
7352 | { | |
2a1ed24c | 7353 | return -1; |
7a0f3083 | 7354 | } |
3a7053b3 MG |
7355 | #endif |
7356 | ||
1e3c88bd PZ |
7357 | /* |
7358 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
7359 | */ | |
7360 | static | |
8e45cb54 | 7361 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 7362 | { |
2a1ed24c | 7363 | int tsk_cache_hot; |
e5673f28 KT |
7364 | |
7365 | lockdep_assert_held(&env->src_rq->lock); | |
7366 | ||
1e3c88bd PZ |
7367 | /* |
7368 | * We do not migrate tasks that are: | |
d3198084 | 7369 | * 1) throttled_lb_pair, or |
1e3c88bd | 7370 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
7371 | * 3) running (obviously), or |
7372 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 7373 | */ |
d3198084 JK |
7374 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
7375 | return 0; | |
7376 | ||
0c98d344 | 7377 | if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) { |
e02e60c1 | 7378 | int cpu; |
88b8dac0 | 7379 | |
ae92882e | 7380 | schedstat_inc(p->se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 7381 | |
6263322c PZ |
7382 | env->flags |= LBF_SOME_PINNED; |
7383 | ||
88b8dac0 | 7384 | /* |
97fb7a0a | 7385 | * Remember if this task can be migrated to any other CPU in |
88b8dac0 SV |
7386 | * our sched_group. We may want to revisit it if we couldn't |
7387 | * meet load balance goals by pulling other tasks on src_cpu. | |
7388 | * | |
65a4433a JH |
7389 | * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have |
7390 | * already computed one in current iteration. | |
88b8dac0 | 7391 | */ |
65a4433a | 7392 | if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
7393 | return 0; |
7394 | ||
97fb7a0a | 7395 | /* Prevent to re-select dst_cpu via env's CPUs: */ |
e02e60c1 | 7396 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { |
0c98d344 | 7397 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) { |
6263322c | 7398 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
7399 | env->new_dst_cpu = cpu; |
7400 | break; | |
7401 | } | |
88b8dac0 | 7402 | } |
e02e60c1 | 7403 | |
1e3c88bd PZ |
7404 | return 0; |
7405 | } | |
88b8dac0 SV |
7406 | |
7407 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 7408 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 7409 | |
ddcdf6e7 | 7410 | if (task_running(env->src_rq, p)) { |
ae92882e | 7411 | schedstat_inc(p->se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
7412 | return 0; |
7413 | } | |
7414 | ||
7415 | /* | |
7416 | * Aggressive migration if: | |
3a7053b3 MG |
7417 | * 1) destination numa is preferred |
7418 | * 2) task is cache cold, or | |
7419 | * 3) too many balance attempts have failed. | |
1e3c88bd | 7420 | */ |
2a1ed24c SD |
7421 | tsk_cache_hot = migrate_degrades_locality(p, env); |
7422 | if (tsk_cache_hot == -1) | |
7423 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 7424 | |
2a1ed24c | 7425 | if (tsk_cache_hot <= 0 || |
7a96c231 | 7426 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 7427 | if (tsk_cache_hot == 1) { |
ae92882e JP |
7428 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
7429 | schedstat_inc(p->se.statistics.nr_forced_migrations); | |
3a7053b3 | 7430 | } |
1e3c88bd PZ |
7431 | return 1; |
7432 | } | |
7433 | ||
ae92882e | 7434 | schedstat_inc(p->se.statistics.nr_failed_migrations_hot); |
4e2dcb73 | 7435 | return 0; |
1e3c88bd PZ |
7436 | } |
7437 | ||
897c395f | 7438 | /* |
163122b7 KT |
7439 | * detach_task() -- detach the task for the migration specified in env |
7440 | */ | |
7441 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
7442 | { | |
7443 | lockdep_assert_held(&env->src_rq->lock); | |
7444 | ||
163122b7 | 7445 | p->on_rq = TASK_ON_RQ_MIGRATING; |
5704ac0a | 7446 | deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); |
163122b7 KT |
7447 | set_task_cpu(p, env->dst_cpu); |
7448 | } | |
7449 | ||
897c395f | 7450 | /* |
e5673f28 | 7451 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 7452 | * part of active balancing operations within "domain". |
897c395f | 7453 | * |
e5673f28 | 7454 | * Returns a task if successful and NULL otherwise. |
897c395f | 7455 | */ |
e5673f28 | 7456 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f | 7457 | { |
93824900 | 7458 | struct task_struct *p; |
897c395f | 7459 | |
e5673f28 KT |
7460 | lockdep_assert_held(&env->src_rq->lock); |
7461 | ||
93824900 UR |
7462 | list_for_each_entry_reverse(p, |
7463 | &env->src_rq->cfs_tasks, se.group_node) { | |
367456c7 PZ |
7464 | if (!can_migrate_task(p, env)) |
7465 | continue; | |
897c395f | 7466 | |
163122b7 | 7467 | detach_task(p, env); |
e5673f28 | 7468 | |
367456c7 | 7469 | /* |
e5673f28 | 7470 | * Right now, this is only the second place where |
163122b7 | 7471 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 7472 | * so we can safely collect stats here rather than |
163122b7 | 7473 | * inside detach_tasks(). |
367456c7 | 7474 | */ |
ae92882e | 7475 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 7476 | return p; |
897c395f | 7477 | } |
e5673f28 | 7478 | return NULL; |
897c395f PZ |
7479 | } |
7480 | ||
eb95308e PZ |
7481 | static const unsigned int sched_nr_migrate_break = 32; |
7482 | ||
5d6523eb | 7483 | /* |
163122b7 KT |
7484 | * detach_tasks() -- tries to detach up to imbalance weighted load from |
7485 | * busiest_rq, as part of a balancing operation within domain "sd". | |
5d6523eb | 7486 | * |
163122b7 | 7487 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 7488 | */ |
163122b7 | 7489 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 7490 | { |
5d6523eb PZ |
7491 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
7492 | struct task_struct *p; | |
367456c7 | 7493 | unsigned long load; |
163122b7 KT |
7494 | int detached = 0; |
7495 | ||
7496 | lockdep_assert_held(&env->src_rq->lock); | |
1e3c88bd | 7497 | |
bd939f45 | 7498 | if (env->imbalance <= 0) |
5d6523eb | 7499 | return 0; |
1e3c88bd | 7500 | |
5d6523eb | 7501 | while (!list_empty(tasks)) { |
985d3a4c YD |
7502 | /* |
7503 | * We don't want to steal all, otherwise we may be treated likewise, | |
7504 | * which could at worst lead to a livelock crash. | |
7505 | */ | |
7506 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
7507 | break; | |
7508 | ||
93824900 | 7509 | p = list_last_entry(tasks, struct task_struct, se.group_node); |
1e3c88bd | 7510 | |
367456c7 PZ |
7511 | env->loop++; |
7512 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 7513 | if (env->loop > env->loop_max) |
367456c7 | 7514 | break; |
5d6523eb PZ |
7515 | |
7516 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 7517 | if (env->loop > env->loop_break) { |
eb95308e | 7518 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 7519 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 7520 | break; |
a195f004 | 7521 | } |
1e3c88bd | 7522 | |
d3198084 | 7523 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
7524 | goto next; |
7525 | ||
7526 | load = task_h_load(p); | |
5d6523eb | 7527 | |
eb95308e | 7528 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
7529 | goto next; |
7530 | ||
bd939f45 | 7531 | if ((load / 2) > env->imbalance) |
367456c7 | 7532 | goto next; |
1e3c88bd | 7533 | |
163122b7 KT |
7534 | detach_task(p, env); |
7535 | list_add(&p->se.group_node, &env->tasks); | |
7536 | ||
7537 | detached++; | |
bd939f45 | 7538 | env->imbalance -= load; |
1e3c88bd PZ |
7539 | |
7540 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
7541 | /* |
7542 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 7543 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
7544 | * the critical section. |
7545 | */ | |
5d6523eb | 7546 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 7547 | break; |
1e3c88bd PZ |
7548 | #endif |
7549 | ||
ee00e66f PZ |
7550 | /* |
7551 | * We only want to steal up to the prescribed amount of | |
7552 | * weighted load. | |
7553 | */ | |
bd939f45 | 7554 | if (env->imbalance <= 0) |
ee00e66f | 7555 | break; |
367456c7 PZ |
7556 | |
7557 | continue; | |
7558 | next: | |
93824900 | 7559 | list_move(&p->se.group_node, tasks); |
1e3c88bd | 7560 | } |
5d6523eb | 7561 | |
1e3c88bd | 7562 | /* |
163122b7 KT |
7563 | * Right now, this is one of only two places we collect this stat |
7564 | * so we can safely collect detach_one_task() stats here rather | |
7565 | * than inside detach_one_task(). | |
1e3c88bd | 7566 | */ |
ae92882e | 7567 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 7568 | |
163122b7 KT |
7569 | return detached; |
7570 | } | |
7571 | ||
7572 | /* | |
7573 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
7574 | */ | |
7575 | static void attach_task(struct rq *rq, struct task_struct *p) | |
7576 | { | |
7577 | lockdep_assert_held(&rq->lock); | |
7578 | ||
7579 | BUG_ON(task_rq(p) != rq); | |
5704ac0a | 7580 | activate_task(rq, p, ENQUEUE_NOCLOCK); |
3ea94de1 | 7581 | p->on_rq = TASK_ON_RQ_QUEUED; |
163122b7 KT |
7582 | check_preempt_curr(rq, p, 0); |
7583 | } | |
7584 | ||
7585 | /* | |
7586 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
7587 | * its new rq. | |
7588 | */ | |
7589 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
7590 | { | |
8a8c69c3 PZ |
7591 | struct rq_flags rf; |
7592 | ||
7593 | rq_lock(rq, &rf); | |
5704ac0a | 7594 | update_rq_clock(rq); |
163122b7 | 7595 | attach_task(rq, p); |
8a8c69c3 | 7596 | rq_unlock(rq, &rf); |
163122b7 KT |
7597 | } |
7598 | ||
7599 | /* | |
7600 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
7601 | * new rq. | |
7602 | */ | |
7603 | static void attach_tasks(struct lb_env *env) | |
7604 | { | |
7605 | struct list_head *tasks = &env->tasks; | |
7606 | struct task_struct *p; | |
8a8c69c3 | 7607 | struct rq_flags rf; |
163122b7 | 7608 | |
8a8c69c3 | 7609 | rq_lock(env->dst_rq, &rf); |
5704ac0a | 7610 | update_rq_clock(env->dst_rq); |
163122b7 KT |
7611 | |
7612 | while (!list_empty(tasks)) { | |
7613 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
7614 | list_del_init(&p->se.group_node); | |
1e3c88bd | 7615 | |
163122b7 KT |
7616 | attach_task(env->dst_rq, p); |
7617 | } | |
7618 | ||
8a8c69c3 | 7619 | rq_unlock(env->dst_rq, &rf); |
1e3c88bd PZ |
7620 | } |
7621 | ||
1936c53c VG |
7622 | static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) |
7623 | { | |
7624 | if (cfs_rq->avg.load_avg) | |
7625 | return true; | |
7626 | ||
7627 | if (cfs_rq->avg.util_avg) | |
7628 | return true; | |
7629 | ||
7630 | return false; | |
7631 | } | |
7632 | ||
91c27493 | 7633 | static inline bool others_have_blocked(struct rq *rq) |
371bf427 VG |
7634 | { |
7635 | if (READ_ONCE(rq->avg_rt.util_avg)) | |
7636 | return true; | |
7637 | ||
3727e0e1 VG |
7638 | if (READ_ONCE(rq->avg_dl.util_avg)) |
7639 | return true; | |
7640 | ||
11d4afd4 | 7641 | #ifdef CONFIG_HAVE_SCHED_AVG_IRQ |
91c27493 VG |
7642 | if (READ_ONCE(rq->avg_irq.util_avg)) |
7643 | return true; | |
7644 | #endif | |
7645 | ||
371bf427 VG |
7646 | return false; |
7647 | } | |
7648 | ||
1936c53c VG |
7649 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7650 | ||
48a16753 | 7651 | static void update_blocked_averages(int cpu) |
9e3081ca | 7652 | { |
9e3081ca | 7653 | struct rq *rq = cpu_rq(cpu); |
c40f7d74 | 7654 | struct cfs_rq *cfs_rq; |
12b04875 | 7655 | const struct sched_class *curr_class; |
8a8c69c3 | 7656 | struct rq_flags rf; |
f643ea22 | 7657 | bool done = true; |
9e3081ca | 7658 | |
8a8c69c3 | 7659 | rq_lock_irqsave(rq, &rf); |
48a16753 | 7660 | update_rq_clock(rq); |
9d89c257 | 7661 | |
9763b67f PZ |
7662 | /* |
7663 | * Iterates the task_group tree in a bottom up fashion, see | |
7664 | * list_add_leaf_cfs_rq() for details. | |
7665 | */ | |
c40f7d74 | 7666 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
bc427898 VG |
7667 | struct sched_entity *se; |
7668 | ||
9d89c257 YD |
7669 | /* throttled entities do not contribute to load */ |
7670 | if (throttled_hierarchy(cfs_rq)) | |
7671 | continue; | |
48a16753 | 7672 | |
3a123bbb | 7673 | if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq)) |
9d89c257 | 7674 | update_tg_load_avg(cfs_rq, 0); |
4e516076 | 7675 | |
bc427898 VG |
7676 | /* Propagate pending load changes to the parent, if any: */ |
7677 | se = cfs_rq->tg->se[cpu]; | |
7678 | if (se && !skip_blocked_update(se)) | |
88c0616e | 7679 | update_load_avg(cfs_rq_of(se), se, 0); |
a9e7f654 | 7680 | |
1936c53c VG |
7681 | /* Don't need periodic decay once load/util_avg are null */ |
7682 | if (cfs_rq_has_blocked(cfs_rq)) | |
f643ea22 | 7683 | done = false; |
9d89c257 | 7684 | } |
12b04875 VG |
7685 | |
7686 | curr_class = rq->curr->sched_class; | |
7687 | update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class); | |
7688 | update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class); | |
91c27493 | 7689 | update_irq_load_avg(rq, 0); |
371bf427 | 7690 | /* Don't need periodic decay once load/util_avg are null */ |
91c27493 | 7691 | if (others_have_blocked(rq)) |
371bf427 | 7692 | done = false; |
e022e0d3 PZ |
7693 | |
7694 | #ifdef CONFIG_NO_HZ_COMMON | |
7695 | rq->last_blocked_load_update_tick = jiffies; | |
f643ea22 VG |
7696 | if (done) |
7697 | rq->has_blocked_load = 0; | |
e022e0d3 | 7698 | #endif |
8a8c69c3 | 7699 | rq_unlock_irqrestore(rq, &rf); |
9e3081ca PZ |
7700 | } |
7701 | ||
9763b67f | 7702 | /* |
68520796 | 7703 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
7704 | * This needs to be done in a top-down fashion because the load of a child |
7705 | * group is a fraction of its parents load. | |
7706 | */ | |
68520796 | 7707 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 7708 | { |
68520796 VD |
7709 | struct rq *rq = rq_of(cfs_rq); |
7710 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 7711 | unsigned long now = jiffies; |
68520796 | 7712 | unsigned long load; |
a35b6466 | 7713 | |
68520796 | 7714 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
7715 | return; |
7716 | ||
68520796 VD |
7717 | cfs_rq->h_load_next = NULL; |
7718 | for_each_sched_entity(se) { | |
7719 | cfs_rq = cfs_rq_of(se); | |
7720 | cfs_rq->h_load_next = se; | |
7721 | if (cfs_rq->last_h_load_update == now) | |
7722 | break; | |
7723 | } | |
a35b6466 | 7724 | |
68520796 | 7725 | if (!se) { |
7ea241af | 7726 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
7727 | cfs_rq->last_h_load_update = now; |
7728 | } | |
7729 | ||
7730 | while ((se = cfs_rq->h_load_next) != NULL) { | |
7731 | load = cfs_rq->h_load; | |
7ea241af YD |
7732 | load = div64_ul(load * se->avg.load_avg, |
7733 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
7734 | cfs_rq = group_cfs_rq(se); |
7735 | cfs_rq->h_load = load; | |
7736 | cfs_rq->last_h_load_update = now; | |
7737 | } | |
9763b67f PZ |
7738 | } |
7739 | ||
367456c7 | 7740 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 7741 | { |
367456c7 | 7742 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 7743 | |
68520796 | 7744 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 7745 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 7746 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
7747 | } |
7748 | #else | |
48a16753 | 7749 | static inline void update_blocked_averages(int cpu) |
9e3081ca | 7750 | { |
6c1d47c0 VG |
7751 | struct rq *rq = cpu_rq(cpu); |
7752 | struct cfs_rq *cfs_rq = &rq->cfs; | |
12b04875 | 7753 | const struct sched_class *curr_class; |
8a8c69c3 | 7754 | struct rq_flags rf; |
6c1d47c0 | 7755 | |
8a8c69c3 | 7756 | rq_lock_irqsave(rq, &rf); |
6c1d47c0 | 7757 | update_rq_clock(rq); |
3a123bbb | 7758 | update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq); |
12b04875 VG |
7759 | |
7760 | curr_class = rq->curr->sched_class; | |
7761 | update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class); | |
7762 | update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class); | |
91c27493 | 7763 | update_irq_load_avg(rq, 0); |
e022e0d3 PZ |
7764 | #ifdef CONFIG_NO_HZ_COMMON |
7765 | rq->last_blocked_load_update_tick = jiffies; | |
91c27493 | 7766 | if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq)) |
f643ea22 | 7767 | rq->has_blocked_load = 0; |
e022e0d3 | 7768 | #endif |
8a8c69c3 | 7769 | rq_unlock_irqrestore(rq, &rf); |
9e3081ca PZ |
7770 | } |
7771 | ||
367456c7 | 7772 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 7773 | { |
9d89c257 | 7774 | return p->se.avg.load_avg; |
1e3c88bd | 7775 | } |
230059de | 7776 | #endif |
1e3c88bd | 7777 | |
1e3c88bd | 7778 | /********** Helpers for find_busiest_group ************************/ |
caeb178c | 7779 | |
1e3c88bd PZ |
7780 | /* |
7781 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
7782 | */ | |
7783 | struct sg_lb_stats { | |
7784 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
7785 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 7786 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 7787 | unsigned long load_per_task; |
63b2ca30 | 7788 | unsigned long group_capacity; |
9e91d61d | 7789 | unsigned long group_util; /* Total utilization of the group */ |
147c5fc2 | 7790 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
147c5fc2 PZ |
7791 | unsigned int idle_cpus; |
7792 | unsigned int group_weight; | |
caeb178c | 7793 | enum group_type group_type; |
ea67821b | 7794 | int group_no_capacity; |
3b1baa64 | 7795 | unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */ |
0ec8aa00 PZ |
7796 | #ifdef CONFIG_NUMA_BALANCING |
7797 | unsigned int nr_numa_running; | |
7798 | unsigned int nr_preferred_running; | |
7799 | #endif | |
1e3c88bd PZ |
7800 | }; |
7801 | ||
56cf515b JK |
7802 | /* |
7803 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
7804 | * during load balancing. | |
7805 | */ | |
7806 | struct sd_lb_stats { | |
7807 | struct sched_group *busiest; /* Busiest group in this sd */ | |
7808 | struct sched_group *local; /* Local group in this sd */ | |
90001d67 | 7809 | unsigned long total_running; |
56cf515b | 7810 | unsigned long total_load; /* Total load of all groups in sd */ |
63b2ca30 | 7811 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b JK |
7812 | unsigned long avg_load; /* Average load across all groups in sd */ |
7813 | ||
56cf515b | 7814 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 7815 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
7816 | }; |
7817 | ||
147c5fc2 PZ |
7818 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
7819 | { | |
7820 | /* | |
7821 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
7822 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
7823 | * We must however clear busiest_stat::avg_load because | |
7824 | * update_sd_pick_busiest() reads this before assignment. | |
7825 | */ | |
7826 | *sds = (struct sd_lb_stats){ | |
7827 | .busiest = NULL, | |
7828 | .local = NULL, | |
90001d67 | 7829 | .total_running = 0UL, |
147c5fc2 | 7830 | .total_load = 0UL, |
63b2ca30 | 7831 | .total_capacity = 0UL, |
147c5fc2 PZ |
7832 | .busiest_stat = { |
7833 | .avg_load = 0UL, | |
caeb178c RR |
7834 | .sum_nr_running = 0, |
7835 | .group_type = group_other, | |
147c5fc2 PZ |
7836 | }, |
7837 | }; | |
7838 | } | |
7839 | ||
1e3c88bd PZ |
7840 | /** |
7841 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
7842 | * @sd: The sched_domain whose load_idx is to be obtained. | |
ed1b7732 | 7843 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. |
e69f6186 YB |
7844 | * |
7845 | * Return: The load index. | |
1e3c88bd PZ |
7846 | */ |
7847 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
7848 | enum cpu_idle_type idle) | |
7849 | { | |
7850 | int load_idx; | |
7851 | ||
7852 | switch (idle) { | |
7853 | case CPU_NOT_IDLE: | |
7854 | load_idx = sd->busy_idx; | |
7855 | break; | |
7856 | ||
7857 | case CPU_NEWLY_IDLE: | |
7858 | load_idx = sd->newidle_idx; | |
7859 | break; | |
7860 | default: | |
7861 | load_idx = sd->idle_idx; | |
7862 | break; | |
7863 | } | |
7864 | ||
7865 | return load_idx; | |
7866 | } | |
7867 | ||
287cdaac | 7868 | static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
7869 | { |
7870 | struct rq *rq = cpu_rq(cpu); | |
287cdaac | 7871 | unsigned long max = arch_scale_cpu_capacity(sd, cpu); |
523e979d | 7872 | unsigned long used, free; |
523e979d | 7873 | unsigned long irq; |
b654f7de | 7874 | |
2e62c474 | 7875 | irq = cpu_util_irq(rq); |
cadefd3d | 7876 | |
523e979d VG |
7877 | if (unlikely(irq >= max)) |
7878 | return 1; | |
aa483808 | 7879 | |
523e979d VG |
7880 | used = READ_ONCE(rq->avg_rt.util_avg); |
7881 | used += READ_ONCE(rq->avg_dl.util_avg); | |
1e3c88bd | 7882 | |
523e979d VG |
7883 | if (unlikely(used >= max)) |
7884 | return 1; | |
1e3c88bd | 7885 | |
523e979d | 7886 | free = max - used; |
2e62c474 VG |
7887 | |
7888 | return scale_irq_capacity(free, irq, max); | |
1e3c88bd PZ |
7889 | } |
7890 | ||
ced549fa | 7891 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 7892 | { |
287cdaac | 7893 | unsigned long capacity = scale_rt_capacity(sd, cpu); |
1e3c88bd PZ |
7894 | struct sched_group *sdg = sd->groups; |
7895 | ||
523e979d | 7896 | cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu); |
1e3c88bd | 7897 | |
ced549fa NP |
7898 | if (!capacity) |
7899 | capacity = 1; | |
1e3c88bd | 7900 | |
ced549fa NP |
7901 | cpu_rq(cpu)->cpu_capacity = capacity; |
7902 | sdg->sgc->capacity = capacity; | |
bf475ce0 | 7903 | sdg->sgc->min_capacity = capacity; |
e3d6d0cb | 7904 | sdg->sgc->max_capacity = capacity; |
1e3c88bd PZ |
7905 | } |
7906 | ||
63b2ca30 | 7907 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
7908 | { |
7909 | struct sched_domain *child = sd->child; | |
7910 | struct sched_group *group, *sdg = sd->groups; | |
e3d6d0cb | 7911 | unsigned long capacity, min_capacity, max_capacity; |
4ec4412e VG |
7912 | unsigned long interval; |
7913 | ||
7914 | interval = msecs_to_jiffies(sd->balance_interval); | |
7915 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 7916 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
7917 | |
7918 | if (!child) { | |
ced549fa | 7919 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
7920 | return; |
7921 | } | |
7922 | ||
dc7ff76e | 7923 | capacity = 0; |
bf475ce0 | 7924 | min_capacity = ULONG_MAX; |
e3d6d0cb | 7925 | max_capacity = 0; |
1e3c88bd | 7926 | |
74a5ce20 PZ |
7927 | if (child->flags & SD_OVERLAP) { |
7928 | /* | |
7929 | * SD_OVERLAP domains cannot assume that child groups | |
7930 | * span the current group. | |
7931 | */ | |
7932 | ||
ae4df9d6 | 7933 | for_each_cpu(cpu, sched_group_span(sdg)) { |
63b2ca30 | 7934 | struct sched_group_capacity *sgc; |
9abf24d4 | 7935 | struct rq *rq = cpu_rq(cpu); |
863bffc8 | 7936 | |
9abf24d4 | 7937 | /* |
63b2ca30 | 7938 | * build_sched_domains() -> init_sched_groups_capacity() |
9abf24d4 SD |
7939 | * gets here before we've attached the domains to the |
7940 | * runqueues. | |
7941 | * | |
ced549fa NP |
7942 | * Use capacity_of(), which is set irrespective of domains |
7943 | * in update_cpu_capacity(). | |
9abf24d4 | 7944 | * |
dc7ff76e | 7945 | * This avoids capacity from being 0 and |
9abf24d4 | 7946 | * causing divide-by-zero issues on boot. |
9abf24d4 SD |
7947 | */ |
7948 | if (unlikely(!rq->sd)) { | |
ced549fa | 7949 | capacity += capacity_of(cpu); |
bf475ce0 MR |
7950 | } else { |
7951 | sgc = rq->sd->groups->sgc; | |
7952 | capacity += sgc->capacity; | |
9abf24d4 | 7953 | } |
863bffc8 | 7954 | |
bf475ce0 | 7955 | min_capacity = min(capacity, min_capacity); |
e3d6d0cb | 7956 | max_capacity = max(capacity, max_capacity); |
863bffc8 | 7957 | } |
74a5ce20 PZ |
7958 | } else { |
7959 | /* | |
7960 | * !SD_OVERLAP domains can assume that child groups | |
7961 | * span the current group. | |
97a7142f | 7962 | */ |
74a5ce20 PZ |
7963 | |
7964 | group = child->groups; | |
7965 | do { | |
bf475ce0 MR |
7966 | struct sched_group_capacity *sgc = group->sgc; |
7967 | ||
7968 | capacity += sgc->capacity; | |
7969 | min_capacity = min(sgc->min_capacity, min_capacity); | |
e3d6d0cb | 7970 | max_capacity = max(sgc->max_capacity, max_capacity); |
74a5ce20 PZ |
7971 | group = group->next; |
7972 | } while (group != child->groups); | |
7973 | } | |
1e3c88bd | 7974 | |
63b2ca30 | 7975 | sdg->sgc->capacity = capacity; |
bf475ce0 | 7976 | sdg->sgc->min_capacity = min_capacity; |
e3d6d0cb | 7977 | sdg->sgc->max_capacity = max_capacity; |
1e3c88bd PZ |
7978 | } |
7979 | ||
9d5efe05 | 7980 | /* |
ea67821b VG |
7981 | * Check whether the capacity of the rq has been noticeably reduced by side |
7982 | * activity. The imbalance_pct is used for the threshold. | |
7983 | * Return true is the capacity is reduced | |
9d5efe05 SV |
7984 | */ |
7985 | static inline int | |
ea67821b | 7986 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 7987 | { |
ea67821b VG |
7988 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
7989 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
7990 | } |
7991 | ||
30ce5dab PZ |
7992 | /* |
7993 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
0c98d344 | 7994 | * groups is inadequate due to ->cpus_allowed constraints. |
30ce5dab | 7995 | * |
97fb7a0a IM |
7996 | * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a |
7997 | * cpumask covering 1 CPU of the first group and 3 CPUs of the second group. | |
30ce5dab PZ |
7998 | * Something like: |
7999 | * | |
2b4d5b25 IM |
8000 | * { 0 1 2 3 } { 4 5 6 7 } |
8001 | * * * * * | |
30ce5dab PZ |
8002 | * |
8003 | * If we were to balance group-wise we'd place two tasks in the first group and | |
8004 | * two tasks in the second group. Clearly this is undesired as it will overload | |
97fb7a0a | 8005 | * cpu 3 and leave one of the CPUs in the second group unused. |
30ce5dab PZ |
8006 | * |
8007 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
8008 | * by noticing the lower domain failed to reach balance and had difficulty |
8009 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
8010 | * |
8011 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 8012 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 8013 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
8014 | * to create an effective group imbalance. |
8015 | * | |
8016 | * This is a somewhat tricky proposition since the next run might not find the | |
8017 | * group imbalance and decide the groups need to be balanced again. A most | |
8018 | * subtle and fragile situation. | |
8019 | */ | |
8020 | ||
6263322c | 8021 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 8022 | { |
63b2ca30 | 8023 | return group->sgc->imbalance; |
30ce5dab PZ |
8024 | } |
8025 | ||
b37d9316 | 8026 | /* |
ea67821b VG |
8027 | * group_has_capacity returns true if the group has spare capacity that could |
8028 | * be used by some tasks. | |
8029 | * We consider that a group has spare capacity if the * number of task is | |
9e91d61d DE |
8030 | * smaller than the number of CPUs or if the utilization is lower than the |
8031 | * available capacity for CFS tasks. | |
ea67821b VG |
8032 | * For the latter, we use a threshold to stabilize the state, to take into |
8033 | * account the variance of the tasks' load and to return true if the available | |
8034 | * capacity in meaningful for the load balancer. | |
8035 | * As an example, an available capacity of 1% can appear but it doesn't make | |
8036 | * any benefit for the load balance. | |
b37d9316 | 8037 | */ |
ea67821b VG |
8038 | static inline bool |
8039 | group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs) | |
b37d9316 | 8040 | { |
ea67821b VG |
8041 | if (sgs->sum_nr_running < sgs->group_weight) |
8042 | return true; | |
c61037e9 | 8043 | |
ea67821b | 8044 | if ((sgs->group_capacity * 100) > |
9e91d61d | 8045 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 8046 | return true; |
b37d9316 | 8047 | |
ea67821b VG |
8048 | return false; |
8049 | } | |
8050 | ||
8051 | /* | |
8052 | * group_is_overloaded returns true if the group has more tasks than it can | |
8053 | * handle. | |
8054 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
8055 | * with the exact right number of tasks, has no more spare capacity but is not | |
8056 | * overloaded so both group_has_capacity and group_is_overloaded return | |
8057 | * false. | |
8058 | */ | |
8059 | static inline bool | |
8060 | group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs) | |
8061 | { | |
8062 | if (sgs->sum_nr_running <= sgs->group_weight) | |
8063 | return false; | |
b37d9316 | 8064 | |
ea67821b | 8065 | if ((sgs->group_capacity * 100) < |
9e91d61d | 8066 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 8067 | return true; |
b37d9316 | 8068 | |
ea67821b | 8069 | return false; |
b37d9316 PZ |
8070 | } |
8071 | ||
9e0994c0 | 8072 | /* |
e3d6d0cb | 8073 | * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller |
9e0994c0 MR |
8074 | * per-CPU capacity than sched_group ref. |
8075 | */ | |
8076 | static inline bool | |
e3d6d0cb | 8077 | group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref) |
9e0994c0 MR |
8078 | { |
8079 | return sg->sgc->min_capacity * capacity_margin < | |
8080 | ref->sgc->min_capacity * 1024; | |
8081 | } | |
8082 | ||
e3d6d0cb MR |
8083 | /* |
8084 | * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller | |
8085 | * per-CPU capacity_orig than sched_group ref. | |
8086 | */ | |
8087 | static inline bool | |
8088 | group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref) | |
8089 | { | |
8090 | return sg->sgc->max_capacity * capacity_margin < | |
8091 | ref->sgc->max_capacity * 1024; | |
8092 | } | |
8093 | ||
79a89f92 LY |
8094 | static inline enum |
8095 | group_type group_classify(struct sched_group *group, | |
8096 | struct sg_lb_stats *sgs) | |
caeb178c | 8097 | { |
ea67821b | 8098 | if (sgs->group_no_capacity) |
caeb178c RR |
8099 | return group_overloaded; |
8100 | ||
8101 | if (sg_imbalanced(group)) | |
8102 | return group_imbalanced; | |
8103 | ||
3b1baa64 MR |
8104 | if (sgs->group_misfit_task_load) |
8105 | return group_misfit_task; | |
8106 | ||
caeb178c RR |
8107 | return group_other; |
8108 | } | |
8109 | ||
63928384 | 8110 | static bool update_nohz_stats(struct rq *rq, bool force) |
e022e0d3 PZ |
8111 | { |
8112 | #ifdef CONFIG_NO_HZ_COMMON | |
8113 | unsigned int cpu = rq->cpu; | |
8114 | ||
f643ea22 VG |
8115 | if (!rq->has_blocked_load) |
8116 | return false; | |
8117 | ||
e022e0d3 | 8118 | if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) |
f643ea22 | 8119 | return false; |
e022e0d3 | 8120 | |
63928384 | 8121 | if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick)) |
f643ea22 | 8122 | return true; |
e022e0d3 PZ |
8123 | |
8124 | update_blocked_averages(cpu); | |
f643ea22 VG |
8125 | |
8126 | return rq->has_blocked_load; | |
8127 | #else | |
8128 | return false; | |
e022e0d3 PZ |
8129 | #endif |
8130 | } | |
8131 | ||
1e3c88bd PZ |
8132 | /** |
8133 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 8134 | * @env: The load balancing environment. |
1e3c88bd | 8135 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 8136 | * @sgs: variable to hold the statistics for this group. |
630246a0 | 8137 | * @sg_status: Holds flag indicating the status of the sched_group |
1e3c88bd | 8138 | */ |
bd939f45 | 8139 | static inline void update_sg_lb_stats(struct lb_env *env, |
630246a0 QP |
8140 | struct sched_group *group, |
8141 | struct sg_lb_stats *sgs, | |
8142 | int *sg_status) | |
1e3c88bd | 8143 | { |
630246a0 QP |
8144 | int local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group)); |
8145 | int load_idx = get_sd_load_idx(env->sd, env->idle); | |
30ce5dab | 8146 | unsigned long load; |
a426f99c | 8147 | int i, nr_running; |
1e3c88bd | 8148 | |
b72ff13c PZ |
8149 | memset(sgs, 0, sizeof(*sgs)); |
8150 | ||
ae4df9d6 | 8151 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
1e3c88bd PZ |
8152 | struct rq *rq = cpu_rq(i); |
8153 | ||
63928384 | 8154 | if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false)) |
f643ea22 | 8155 | env->flags |= LBF_NOHZ_AGAIN; |
e022e0d3 | 8156 | |
97fb7a0a | 8157 | /* Bias balancing toward CPUs of our domain: */ |
6263322c | 8158 | if (local_group) |
04f733b4 | 8159 | load = target_load(i, load_idx); |
6263322c | 8160 | else |
1e3c88bd | 8161 | load = source_load(i, load_idx); |
1e3c88bd PZ |
8162 | |
8163 | sgs->group_load += load; | |
9e91d61d | 8164 | sgs->group_util += cpu_util(i); |
65fdac08 | 8165 | sgs->sum_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 8166 | |
a426f99c WL |
8167 | nr_running = rq->nr_running; |
8168 | if (nr_running > 1) | |
630246a0 | 8169 | *sg_status |= SG_OVERLOAD; |
4486edd1 | 8170 | |
2802bf3c MR |
8171 | if (cpu_overutilized(i)) |
8172 | *sg_status |= SG_OVERUTILIZED; | |
4486edd1 | 8173 | |
0ec8aa00 PZ |
8174 | #ifdef CONFIG_NUMA_BALANCING |
8175 | sgs->nr_numa_running += rq->nr_numa_running; | |
8176 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
8177 | #endif | |
c7132dd6 | 8178 | sgs->sum_weighted_load += weighted_cpuload(rq); |
a426f99c WL |
8179 | /* |
8180 | * No need to call idle_cpu() if nr_running is not 0 | |
8181 | */ | |
8182 | if (!nr_running && idle_cpu(i)) | |
aae6d3dd | 8183 | sgs->idle_cpus++; |
3b1baa64 MR |
8184 | |
8185 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && | |
757ffdd7 | 8186 | sgs->group_misfit_task_load < rq->misfit_task_load) { |
3b1baa64 | 8187 | sgs->group_misfit_task_load = rq->misfit_task_load; |
630246a0 | 8188 | *sg_status |= SG_OVERLOAD; |
757ffdd7 | 8189 | } |
1e3c88bd PZ |
8190 | } |
8191 | ||
63b2ca30 NP |
8192 | /* Adjust by relative CPU capacity of the group */ |
8193 | sgs->group_capacity = group->sgc->capacity; | |
ca8ce3d0 | 8194 | sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; |
1e3c88bd | 8195 | |
dd5feea1 | 8196 | if (sgs->sum_nr_running) |
38d0f770 | 8197 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 8198 | |
aae6d3dd | 8199 | sgs->group_weight = group->group_weight; |
b37d9316 | 8200 | |
ea67821b | 8201 | sgs->group_no_capacity = group_is_overloaded(env, sgs); |
79a89f92 | 8202 | sgs->group_type = group_classify(group, sgs); |
1e3c88bd PZ |
8203 | } |
8204 | ||
532cb4c4 MN |
8205 | /** |
8206 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 8207 | * @env: The load balancing environment. |
532cb4c4 MN |
8208 | * @sds: sched_domain statistics |
8209 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 8210 | * @sgs: sched_group statistics |
532cb4c4 MN |
8211 | * |
8212 | * Determine if @sg is a busier group than the previously selected | |
8213 | * busiest group. | |
e69f6186 YB |
8214 | * |
8215 | * Return: %true if @sg is a busier group than the previously selected | |
8216 | * busiest group. %false otherwise. | |
532cb4c4 | 8217 | */ |
bd939f45 | 8218 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
8219 | struct sd_lb_stats *sds, |
8220 | struct sched_group *sg, | |
bd939f45 | 8221 | struct sg_lb_stats *sgs) |
532cb4c4 | 8222 | { |
caeb178c | 8223 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 8224 | |
cad68e55 MR |
8225 | /* |
8226 | * Don't try to pull misfit tasks we can't help. | |
8227 | * We can use max_capacity here as reduction in capacity on some | |
8228 | * CPUs in the group should either be possible to resolve | |
8229 | * internally or be covered by avg_load imbalance (eventually). | |
8230 | */ | |
8231 | if (sgs->group_type == group_misfit_task && | |
8232 | (!group_smaller_max_cpu_capacity(sg, sds->local) || | |
8233 | !group_has_capacity(env, &sds->local_stat))) | |
8234 | return false; | |
8235 | ||
caeb178c | 8236 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
8237 | return true; |
8238 | ||
caeb178c RR |
8239 | if (sgs->group_type < busiest->group_type) |
8240 | return false; | |
8241 | ||
8242 | if (sgs->avg_load <= busiest->avg_load) | |
8243 | return false; | |
8244 | ||
9e0994c0 MR |
8245 | if (!(env->sd->flags & SD_ASYM_CPUCAPACITY)) |
8246 | goto asym_packing; | |
8247 | ||
8248 | /* | |
8249 | * Candidate sg has no more than one task per CPU and | |
8250 | * has higher per-CPU capacity. Migrating tasks to less | |
8251 | * capable CPUs may harm throughput. Maximize throughput, | |
8252 | * power/energy consequences are not considered. | |
8253 | */ | |
8254 | if (sgs->sum_nr_running <= sgs->group_weight && | |
e3d6d0cb | 8255 | group_smaller_min_cpu_capacity(sds->local, sg)) |
9e0994c0 MR |
8256 | return false; |
8257 | ||
cad68e55 MR |
8258 | /* |
8259 | * If we have more than one misfit sg go with the biggest misfit. | |
8260 | */ | |
8261 | if (sgs->group_type == group_misfit_task && | |
8262 | sgs->group_misfit_task_load < busiest->group_misfit_task_load) | |
9e0994c0 MR |
8263 | return false; |
8264 | ||
8265 | asym_packing: | |
caeb178c RR |
8266 | /* This is the busiest node in its class. */ |
8267 | if (!(env->sd->flags & SD_ASYM_PACKING)) | |
532cb4c4 MN |
8268 | return true; |
8269 | ||
97fb7a0a | 8270 | /* No ASYM_PACKING if target CPU is already busy */ |
1f621e02 SD |
8271 | if (env->idle == CPU_NOT_IDLE) |
8272 | return true; | |
532cb4c4 | 8273 | /* |
afe06efd TC |
8274 | * ASYM_PACKING needs to move all the work to the highest |
8275 | * prority CPUs in the group, therefore mark all groups | |
8276 | * of lower priority than ourself as busy. | |
532cb4c4 | 8277 | */ |
afe06efd TC |
8278 | if (sgs->sum_nr_running && |
8279 | sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) { | |
532cb4c4 MN |
8280 | if (!sds->busiest) |
8281 | return true; | |
8282 | ||
97fb7a0a | 8283 | /* Prefer to move from lowest priority CPU's work */ |
afe06efd TC |
8284 | if (sched_asym_prefer(sds->busiest->asym_prefer_cpu, |
8285 | sg->asym_prefer_cpu)) | |
532cb4c4 MN |
8286 | return true; |
8287 | } | |
8288 | ||
8289 | return false; | |
8290 | } | |
8291 | ||
0ec8aa00 PZ |
8292 | #ifdef CONFIG_NUMA_BALANCING |
8293 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
8294 | { | |
8295 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
8296 | return regular; | |
8297 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
8298 | return remote; | |
8299 | return all; | |
8300 | } | |
8301 | ||
8302 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
8303 | { | |
8304 | if (rq->nr_running > rq->nr_numa_running) | |
8305 | return regular; | |
8306 | if (rq->nr_running > rq->nr_preferred_running) | |
8307 | return remote; | |
8308 | return all; | |
8309 | } | |
8310 | #else | |
8311 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
8312 | { | |
8313 | return all; | |
8314 | } | |
8315 | ||
8316 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
8317 | { | |
8318 | return regular; | |
8319 | } | |
8320 | #endif /* CONFIG_NUMA_BALANCING */ | |
8321 | ||
1e3c88bd | 8322 | /** |
461819ac | 8323 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 8324 | * @env: The load balancing environment. |
1e3c88bd PZ |
8325 | * @sds: variable to hold the statistics for this sched_domain. |
8326 | */ | |
0ec8aa00 | 8327 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 8328 | { |
bd939f45 PZ |
8329 | struct sched_domain *child = env->sd->child; |
8330 | struct sched_group *sg = env->sd->groups; | |
05b40e05 | 8331 | struct sg_lb_stats *local = &sds->local_stat; |
56cf515b | 8332 | struct sg_lb_stats tmp_sgs; |
dbbad719 | 8333 | bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING; |
630246a0 | 8334 | int sg_status = 0; |
1e3c88bd | 8335 | |
e022e0d3 | 8336 | #ifdef CONFIG_NO_HZ_COMMON |
f643ea22 | 8337 | if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked)) |
e022e0d3 | 8338 | env->flags |= LBF_NOHZ_STATS; |
e022e0d3 PZ |
8339 | #endif |
8340 | ||
1e3c88bd | 8341 | do { |
56cf515b | 8342 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
8343 | int local_group; |
8344 | ||
ae4df9d6 | 8345 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); |
56cf515b JK |
8346 | if (local_group) { |
8347 | sds->local = sg; | |
05b40e05 | 8348 | sgs = local; |
b72ff13c PZ |
8349 | |
8350 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
8351 | time_after_eq(jiffies, sg->sgc->next_update)) |
8352 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 8353 | } |
1e3c88bd | 8354 | |
630246a0 | 8355 | update_sg_lb_stats(env, sg, sgs, &sg_status); |
1e3c88bd | 8356 | |
b72ff13c PZ |
8357 | if (local_group) |
8358 | goto next_group; | |
8359 | ||
1e3c88bd PZ |
8360 | /* |
8361 | * In case the child domain prefers tasks go to siblings | |
ea67821b | 8362 | * first, lower the sg capacity so that we'll try |
75dd321d NR |
8363 | * and move all the excess tasks away. We lower the capacity |
8364 | * of a group only if the local group has the capacity to fit | |
ea67821b VG |
8365 | * these excess tasks. The extra check prevents the case where |
8366 | * you always pull from the heaviest group when it is already | |
8367 | * under-utilized (possible with a large weight task outweighs | |
8368 | * the tasks on the system). | |
1e3c88bd | 8369 | */ |
b72ff13c | 8370 | if (prefer_sibling && sds->local && |
05b40e05 SD |
8371 | group_has_capacity(env, local) && |
8372 | (sgs->sum_nr_running > local->sum_nr_running + 1)) { | |
ea67821b | 8373 | sgs->group_no_capacity = 1; |
79a89f92 | 8374 | sgs->group_type = group_classify(sg, sgs); |
cb0b9f24 | 8375 | } |
1e3c88bd | 8376 | |
b72ff13c | 8377 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 8378 | sds->busiest = sg; |
56cf515b | 8379 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
8380 | } |
8381 | ||
b72ff13c PZ |
8382 | next_group: |
8383 | /* Now, start updating sd_lb_stats */ | |
90001d67 | 8384 | sds->total_running += sgs->sum_nr_running; |
b72ff13c | 8385 | sds->total_load += sgs->group_load; |
63b2ca30 | 8386 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 8387 | |
532cb4c4 | 8388 | sg = sg->next; |
bd939f45 | 8389 | } while (sg != env->sd->groups); |
0ec8aa00 | 8390 | |
f643ea22 VG |
8391 | #ifdef CONFIG_NO_HZ_COMMON |
8392 | if ((env->flags & LBF_NOHZ_AGAIN) && | |
8393 | cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) { | |
8394 | ||
8395 | WRITE_ONCE(nohz.next_blocked, | |
8396 | jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
8397 | } | |
8398 | #endif | |
8399 | ||
0ec8aa00 PZ |
8400 | if (env->sd->flags & SD_NUMA) |
8401 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
8402 | |
8403 | if (!env->sd->parent) { | |
2802bf3c MR |
8404 | struct root_domain *rd = env->dst_rq->rd; |
8405 | ||
4486edd1 | 8406 | /* update overload indicator if we are at root domain */ |
2802bf3c MR |
8407 | WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD); |
8408 | ||
8409 | /* Update over-utilization (tipping point, U >= 0) indicator */ | |
8410 | WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED); | |
8411 | } else if (sg_status & SG_OVERUTILIZED) { | |
8412 | WRITE_ONCE(env->dst_rq->rd->overutilized, SG_OVERUTILIZED); | |
4486edd1 | 8413 | } |
532cb4c4 MN |
8414 | } |
8415 | ||
532cb4c4 MN |
8416 | /** |
8417 | * check_asym_packing - Check to see if the group is packed into the | |
0ba42a59 | 8418 | * sched domain. |
532cb4c4 MN |
8419 | * |
8420 | * This is primarily intended to used at the sibling level. Some | |
8421 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
8422 | * case of POWER7, it can move to lower SMT modes only when higher | |
8423 | * threads are idle. When in lower SMT modes, the threads will | |
8424 | * perform better since they share less core resources. Hence when we | |
8425 | * have idle threads, we want them to be the higher ones. | |
8426 | * | |
8427 | * This packing function is run on idle threads. It checks to see if | |
8428 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
8429 | * CPU number than the packing function is being run on. Here we are | |
8430 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
8431 | * number. | |
8432 | * | |
e69f6186 | 8433 | * Return: 1 when packing is required and a task should be moved to |
46123355 | 8434 | * this CPU. The amount of the imbalance is returned in env->imbalance. |
b6b12294 | 8435 | * |
cd96891d | 8436 | * @env: The load balancing environment. |
532cb4c4 | 8437 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 8438 | */ |
bd939f45 | 8439 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
8440 | { |
8441 | int busiest_cpu; | |
8442 | ||
bd939f45 | 8443 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
8444 | return 0; |
8445 | ||
1f621e02 SD |
8446 | if (env->idle == CPU_NOT_IDLE) |
8447 | return 0; | |
8448 | ||
532cb4c4 MN |
8449 | if (!sds->busiest) |
8450 | return 0; | |
8451 | ||
afe06efd TC |
8452 | busiest_cpu = sds->busiest->asym_prefer_cpu; |
8453 | if (sched_asym_prefer(busiest_cpu, env->dst_cpu)) | |
532cb4c4 MN |
8454 | return 0; |
8455 | ||
4ad4e481 | 8456 | env->imbalance = sds->busiest_stat.group_load; |
bd939f45 | 8457 | |
532cb4c4 | 8458 | return 1; |
1e3c88bd PZ |
8459 | } |
8460 | ||
8461 | /** | |
8462 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
8463 | * amongst the groups of a sched_domain, during | |
8464 | * load balancing. | |
cd96891d | 8465 | * @env: The load balancing environment. |
1e3c88bd | 8466 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 8467 | */ |
bd939f45 PZ |
8468 | static inline |
8469 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd | 8470 | { |
63b2ca30 | 8471 | unsigned long tmp, capa_now = 0, capa_move = 0; |
1e3c88bd | 8472 | unsigned int imbn = 2; |
dd5feea1 | 8473 | unsigned long scaled_busy_load_per_task; |
56cf515b | 8474 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 8475 | |
56cf515b JK |
8476 | local = &sds->local_stat; |
8477 | busiest = &sds->busiest_stat; | |
1e3c88bd | 8478 | |
56cf515b JK |
8479 | if (!local->sum_nr_running) |
8480 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
8481 | else if (busiest->load_per_task > local->load_per_task) | |
8482 | imbn = 1; | |
dd5feea1 | 8483 | |
56cf515b | 8484 | scaled_busy_load_per_task = |
ca8ce3d0 | 8485 | (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 8486 | busiest->group_capacity; |
56cf515b | 8487 | |
3029ede3 VD |
8488 | if (busiest->avg_load + scaled_busy_load_per_task >= |
8489 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 8490 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
8491 | return; |
8492 | } | |
8493 | ||
8494 | /* | |
8495 | * OK, we don't have enough imbalance to justify moving tasks, | |
ced549fa | 8496 | * however we may be able to increase total CPU capacity used by |
1e3c88bd PZ |
8497 | * moving them. |
8498 | */ | |
8499 | ||
63b2ca30 | 8500 | capa_now += busiest->group_capacity * |
56cf515b | 8501 | min(busiest->load_per_task, busiest->avg_load); |
63b2ca30 | 8502 | capa_now += local->group_capacity * |
56cf515b | 8503 | min(local->load_per_task, local->avg_load); |
ca8ce3d0 | 8504 | capa_now /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
8505 | |
8506 | /* Amount of load we'd subtract */ | |
a2cd4260 | 8507 | if (busiest->avg_load > scaled_busy_load_per_task) { |
63b2ca30 | 8508 | capa_move += busiest->group_capacity * |
56cf515b | 8509 | min(busiest->load_per_task, |
a2cd4260 | 8510 | busiest->avg_load - scaled_busy_load_per_task); |
56cf515b | 8511 | } |
1e3c88bd PZ |
8512 | |
8513 | /* Amount of load we'd add */ | |
63b2ca30 | 8514 | if (busiest->avg_load * busiest->group_capacity < |
ca8ce3d0 | 8515 | busiest->load_per_task * SCHED_CAPACITY_SCALE) { |
63b2ca30 NP |
8516 | tmp = (busiest->avg_load * busiest->group_capacity) / |
8517 | local->group_capacity; | |
56cf515b | 8518 | } else { |
ca8ce3d0 | 8519 | tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 8520 | local->group_capacity; |
56cf515b | 8521 | } |
63b2ca30 | 8522 | capa_move += local->group_capacity * |
3ae11c90 | 8523 | min(local->load_per_task, local->avg_load + tmp); |
ca8ce3d0 | 8524 | capa_move /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
8525 | |
8526 | /* Move if we gain throughput */ | |
63b2ca30 | 8527 | if (capa_move > capa_now) |
56cf515b | 8528 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
8529 | } |
8530 | ||
8531 | /** | |
8532 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
8533 | * groups of a given sched_domain during load balance. | |
bd939f45 | 8534 | * @env: load balance environment |
1e3c88bd | 8535 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 8536 | */ |
bd939f45 | 8537 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 8538 | { |
dd5feea1 | 8539 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
8540 | struct sg_lb_stats *local, *busiest; |
8541 | ||
8542 | local = &sds->local_stat; | |
56cf515b | 8543 | busiest = &sds->busiest_stat; |
dd5feea1 | 8544 | |
caeb178c | 8545 | if (busiest->group_type == group_imbalanced) { |
30ce5dab PZ |
8546 | /* |
8547 | * In the group_imb case we cannot rely on group-wide averages | |
97fb7a0a | 8548 | * to ensure CPU-load equilibrium, look at wider averages. XXX |
30ce5dab | 8549 | */ |
56cf515b JK |
8550 | busiest->load_per_task = |
8551 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
8552 | } |
8553 | ||
1e3c88bd | 8554 | /* |
885e542c DE |
8555 | * Avg load of busiest sg can be less and avg load of local sg can |
8556 | * be greater than avg load across all sgs of sd because avg load | |
8557 | * factors in sg capacity and sgs with smaller group_type are | |
8558 | * skipped when updating the busiest sg: | |
1e3c88bd | 8559 | */ |
cad68e55 MR |
8560 | if (busiest->group_type != group_misfit_task && |
8561 | (busiest->avg_load <= sds->avg_load || | |
8562 | local->avg_load >= sds->avg_load)) { | |
bd939f45 PZ |
8563 | env->imbalance = 0; |
8564 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
8565 | } |
8566 | ||
9a5d9ba6 | 8567 | /* |
97fb7a0a | 8568 | * If there aren't any idle CPUs, avoid creating some. |
9a5d9ba6 PZ |
8569 | */ |
8570 | if (busiest->group_type == group_overloaded && | |
8571 | local->group_type == group_overloaded) { | |
1be0eb2a | 8572 | load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE; |
cfa10334 | 8573 | if (load_above_capacity > busiest->group_capacity) { |
ea67821b | 8574 | load_above_capacity -= busiest->group_capacity; |
26656215 | 8575 | load_above_capacity *= scale_load_down(NICE_0_LOAD); |
cfa10334 MR |
8576 | load_above_capacity /= busiest->group_capacity; |
8577 | } else | |
ea67821b | 8578 | load_above_capacity = ~0UL; |
dd5feea1 SS |
8579 | } |
8580 | ||
8581 | /* | |
97fb7a0a | 8582 | * We're trying to get all the CPUs to the average_load, so we don't |
dd5feea1 | 8583 | * want to push ourselves above the average load, nor do we wish to |
97fb7a0a | 8584 | * reduce the max loaded CPU below the average load. At the same time, |
0a9b23ce DE |
8585 | * we also don't want to reduce the group load below the group |
8586 | * capacity. Thus we look for the minimum possible imbalance. | |
dd5feea1 | 8587 | */ |
30ce5dab | 8588 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
8589 | |
8590 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 8591 | env->imbalance = min( |
63b2ca30 NP |
8592 | max_pull * busiest->group_capacity, |
8593 | (sds->avg_load - local->avg_load) * local->group_capacity | |
ca8ce3d0 | 8594 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd | 8595 | |
cad68e55 MR |
8596 | /* Boost imbalance to allow misfit task to be balanced. */ |
8597 | if (busiest->group_type == group_misfit_task) { | |
8598 | env->imbalance = max_t(long, env->imbalance, | |
8599 | busiest->group_misfit_task_load); | |
8600 | } | |
8601 | ||
1e3c88bd PZ |
8602 | /* |
8603 | * if *imbalance is less than the average load per runnable task | |
25985edc | 8604 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
8605 | * a think about bumping its value to force at least one task to be |
8606 | * moved | |
8607 | */ | |
56cf515b | 8608 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 8609 | return fix_small_imbalance(env, sds); |
1e3c88bd | 8610 | } |
fab47622 | 8611 | |
1e3c88bd PZ |
8612 | /******* find_busiest_group() helpers end here *********************/ |
8613 | ||
8614 | /** | |
8615 | * find_busiest_group - Returns the busiest group within the sched_domain | |
0a9b23ce | 8616 | * if there is an imbalance. |
1e3c88bd PZ |
8617 | * |
8618 | * Also calculates the amount of weighted load which should be moved | |
8619 | * to restore balance. | |
8620 | * | |
cd96891d | 8621 | * @env: The load balancing environment. |
1e3c88bd | 8622 | * |
e69f6186 | 8623 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 8624 | */ |
56cf515b | 8625 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 8626 | { |
56cf515b | 8627 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
8628 | struct sd_lb_stats sds; |
8629 | ||
147c5fc2 | 8630 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
8631 | |
8632 | /* | |
8633 | * Compute the various statistics relavent for load balancing at | |
8634 | * this level. | |
8635 | */ | |
23f0d209 | 8636 | update_sd_lb_stats(env, &sds); |
2802bf3c | 8637 | |
f8a696f2 | 8638 | if (sched_energy_enabled()) { |
2802bf3c MR |
8639 | struct root_domain *rd = env->dst_rq->rd; |
8640 | ||
8641 | if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)) | |
8642 | goto out_balanced; | |
8643 | } | |
8644 | ||
56cf515b JK |
8645 | local = &sds.local_stat; |
8646 | busiest = &sds.busiest_stat; | |
1e3c88bd | 8647 | |
ea67821b | 8648 | /* ASYM feature bypasses nice load balance check */ |
1f621e02 | 8649 | if (check_asym_packing(env, &sds)) |
532cb4c4 MN |
8650 | return sds.busiest; |
8651 | ||
cc57aa8f | 8652 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 8653 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
8654 | goto out_balanced; |
8655 | ||
90001d67 | 8656 | /* XXX broken for overlapping NUMA groups */ |
ca8ce3d0 NP |
8657 | sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) |
8658 | / sds.total_capacity; | |
b0432d8f | 8659 | |
866ab43e PZ |
8660 | /* |
8661 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 8662 | * work because they assume all things are equal, which typically |
866ab43e PZ |
8663 | * isn't true due to cpus_allowed constraints and the like. |
8664 | */ | |
caeb178c | 8665 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
8666 | goto force_balance; |
8667 | ||
583ffd99 BJ |
8668 | /* |
8669 | * When dst_cpu is idle, prevent SMP nice and/or asymmetric group | |
8670 | * capacities from resulting in underutilization due to avg_load. | |
8671 | */ | |
8672 | if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) && | |
ea67821b | 8673 | busiest->group_no_capacity) |
fab47622 NR |
8674 | goto force_balance; |
8675 | ||
cad68e55 MR |
8676 | /* Misfit tasks should be dealt with regardless of the avg load */ |
8677 | if (busiest->group_type == group_misfit_task) | |
8678 | goto force_balance; | |
8679 | ||
cc57aa8f | 8680 | /* |
9c58c79a | 8681 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
8682 | * don't try and pull any tasks. |
8683 | */ | |
56cf515b | 8684 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
8685 | goto out_balanced; |
8686 | ||
cc57aa8f PZ |
8687 | /* |
8688 | * Don't pull any tasks if this group is already above the domain | |
8689 | * average load. | |
8690 | */ | |
56cf515b | 8691 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
8692 | goto out_balanced; |
8693 | ||
bd939f45 | 8694 | if (env->idle == CPU_IDLE) { |
aae6d3dd | 8695 | /* |
97fb7a0a | 8696 | * This CPU is idle. If the busiest group is not overloaded |
43f4d666 | 8697 | * and there is no imbalance between this and busiest group |
97fb7a0a | 8698 | * wrt idle CPUs, it is balanced. The imbalance becomes |
43f4d666 VG |
8699 | * significant if the diff is greater than 1 otherwise we |
8700 | * might end up to just move the imbalance on another group | |
aae6d3dd | 8701 | */ |
43f4d666 VG |
8702 | if ((busiest->group_type != group_overloaded) && |
8703 | (local->idle_cpus <= (busiest->idle_cpus + 1))) | |
aae6d3dd | 8704 | goto out_balanced; |
c186fafe PZ |
8705 | } else { |
8706 | /* | |
8707 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
8708 | * imbalance_pct to be conservative. | |
8709 | */ | |
56cf515b JK |
8710 | if (100 * busiest->avg_load <= |
8711 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 8712 | goto out_balanced; |
aae6d3dd | 8713 | } |
1e3c88bd | 8714 | |
fab47622 | 8715 | force_balance: |
1e3c88bd | 8716 | /* Looks like there is an imbalance. Compute it */ |
cad68e55 | 8717 | env->src_grp_type = busiest->group_type; |
bd939f45 | 8718 | calculate_imbalance(env, &sds); |
bb3485c8 | 8719 | return env->imbalance ? sds.busiest : NULL; |
1e3c88bd PZ |
8720 | |
8721 | out_balanced: | |
bd939f45 | 8722 | env->imbalance = 0; |
1e3c88bd PZ |
8723 | return NULL; |
8724 | } | |
8725 | ||
8726 | /* | |
97fb7a0a | 8727 | * find_busiest_queue - find the busiest runqueue among the CPUs in the group. |
1e3c88bd | 8728 | */ |
bd939f45 | 8729 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 8730 | struct sched_group *group) |
1e3c88bd PZ |
8731 | { |
8732 | struct rq *busiest = NULL, *rq; | |
ced549fa | 8733 | unsigned long busiest_load = 0, busiest_capacity = 1; |
1e3c88bd PZ |
8734 | int i; |
8735 | ||
ae4df9d6 | 8736 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
ea67821b | 8737 | unsigned long capacity, wl; |
0ec8aa00 PZ |
8738 | enum fbq_type rt; |
8739 | ||
8740 | rq = cpu_rq(i); | |
8741 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 8742 | |
0ec8aa00 PZ |
8743 | /* |
8744 | * We classify groups/runqueues into three groups: | |
8745 | * - regular: there are !numa tasks | |
8746 | * - remote: there are numa tasks that run on the 'wrong' node | |
8747 | * - all: there is no distinction | |
8748 | * | |
8749 | * In order to avoid migrating ideally placed numa tasks, | |
8750 | * ignore those when there's better options. | |
8751 | * | |
8752 | * If we ignore the actual busiest queue to migrate another | |
8753 | * task, the next balance pass can still reduce the busiest | |
8754 | * queue by moving tasks around inside the node. | |
8755 | * | |
8756 | * If we cannot move enough load due to this classification | |
8757 | * the next pass will adjust the group classification and | |
8758 | * allow migration of more tasks. | |
8759 | * | |
8760 | * Both cases only affect the total convergence complexity. | |
8761 | */ | |
8762 | if (rt > env->fbq_type) | |
8763 | continue; | |
8764 | ||
cad68e55 MR |
8765 | /* |
8766 | * For ASYM_CPUCAPACITY domains with misfit tasks we simply | |
8767 | * seek the "biggest" misfit task. | |
8768 | */ | |
8769 | if (env->src_grp_type == group_misfit_task) { | |
8770 | if (rq->misfit_task_load > busiest_load) { | |
8771 | busiest_load = rq->misfit_task_load; | |
8772 | busiest = rq; | |
8773 | } | |
8774 | ||
8775 | continue; | |
8776 | } | |
8777 | ||
ced549fa | 8778 | capacity = capacity_of(i); |
9d5efe05 | 8779 | |
4ad3831a CR |
8780 | /* |
8781 | * For ASYM_CPUCAPACITY domains, don't pick a CPU that could | |
8782 | * eventually lead to active_balancing high->low capacity. | |
8783 | * Higher per-CPU capacity is considered better than balancing | |
8784 | * average load. | |
8785 | */ | |
8786 | if (env->sd->flags & SD_ASYM_CPUCAPACITY && | |
8787 | capacity_of(env->dst_cpu) < capacity && | |
8788 | rq->nr_running == 1) | |
8789 | continue; | |
8790 | ||
c7132dd6 | 8791 | wl = weighted_cpuload(rq); |
1e3c88bd | 8792 | |
6e40f5bb TG |
8793 | /* |
8794 | * When comparing with imbalance, use weighted_cpuload() | |
97fb7a0a | 8795 | * which is not scaled with the CPU capacity. |
6e40f5bb | 8796 | */ |
ea67821b VG |
8797 | |
8798 | if (rq->nr_running == 1 && wl > env->imbalance && | |
8799 | !check_cpu_capacity(rq, env->sd)) | |
1e3c88bd PZ |
8800 | continue; |
8801 | ||
6e40f5bb | 8802 | /* |
97fb7a0a IM |
8803 | * For the load comparisons with the other CPU's, consider |
8804 | * the weighted_cpuload() scaled with the CPU capacity, so | |
8805 | * that the load can be moved away from the CPU that is | |
ced549fa | 8806 | * potentially running at a lower capacity. |
95a79b80 | 8807 | * |
ced549fa | 8808 | * Thus we're looking for max(wl_i / capacity_i), crosswise |
95a79b80 | 8809 | * multiplication to rid ourselves of the division works out |
ced549fa NP |
8810 | * to: wl_i * capacity_j > wl_j * capacity_i; where j is |
8811 | * our previous maximum. | |
6e40f5bb | 8812 | */ |
ced549fa | 8813 | if (wl * busiest_capacity > busiest_load * capacity) { |
95a79b80 | 8814 | busiest_load = wl; |
ced549fa | 8815 | busiest_capacity = capacity; |
1e3c88bd PZ |
8816 | busiest = rq; |
8817 | } | |
8818 | } | |
8819 | ||
8820 | return busiest; | |
8821 | } | |
8822 | ||
8823 | /* | |
8824 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
8825 | * so long as it is large enough. | |
8826 | */ | |
8827 | #define MAX_PINNED_INTERVAL 512 | |
8828 | ||
46a745d9 VG |
8829 | static inline bool |
8830 | asym_active_balance(struct lb_env *env) | |
1af3ed3d | 8831 | { |
46a745d9 VG |
8832 | /* |
8833 | * ASYM_PACKING needs to force migrate tasks from busy but | |
8834 | * lower priority CPUs in order to pack all tasks in the | |
8835 | * highest priority CPUs. | |
8836 | */ | |
8837 | return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) && | |
8838 | sched_asym_prefer(env->dst_cpu, env->src_cpu); | |
8839 | } | |
bd939f45 | 8840 | |
46a745d9 VG |
8841 | static inline bool |
8842 | voluntary_active_balance(struct lb_env *env) | |
8843 | { | |
8844 | struct sched_domain *sd = env->sd; | |
532cb4c4 | 8845 | |
46a745d9 VG |
8846 | if (asym_active_balance(env)) |
8847 | return 1; | |
1af3ed3d | 8848 | |
1aaf90a4 VG |
8849 | /* |
8850 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
8851 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
8852 | * because of other sched_class or IRQs if more capacity stays | |
8853 | * available on dst_cpu. | |
8854 | */ | |
8855 | if ((env->idle != CPU_NOT_IDLE) && | |
8856 | (env->src_rq->cfs.h_nr_running == 1)) { | |
8857 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
8858 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
8859 | return 1; | |
8860 | } | |
8861 | ||
cad68e55 MR |
8862 | if (env->src_grp_type == group_misfit_task) |
8863 | return 1; | |
8864 | ||
46a745d9 VG |
8865 | return 0; |
8866 | } | |
8867 | ||
8868 | static int need_active_balance(struct lb_env *env) | |
8869 | { | |
8870 | struct sched_domain *sd = env->sd; | |
8871 | ||
8872 | if (voluntary_active_balance(env)) | |
8873 | return 1; | |
8874 | ||
1af3ed3d PZ |
8875 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); |
8876 | } | |
8877 | ||
969c7921 TH |
8878 | static int active_load_balance_cpu_stop(void *data); |
8879 | ||
23f0d209 JK |
8880 | static int should_we_balance(struct lb_env *env) |
8881 | { | |
8882 | struct sched_group *sg = env->sd->groups; | |
23f0d209 JK |
8883 | int cpu, balance_cpu = -1; |
8884 | ||
024c9d2f PZ |
8885 | /* |
8886 | * Ensure the balancing environment is consistent; can happen | |
8887 | * when the softirq triggers 'during' hotplug. | |
8888 | */ | |
8889 | if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) | |
8890 | return 0; | |
8891 | ||
23f0d209 | 8892 | /* |
97fb7a0a | 8893 | * In the newly idle case, we will allow all the CPUs |
23f0d209 JK |
8894 | * to do the newly idle load balance. |
8895 | */ | |
8896 | if (env->idle == CPU_NEWLY_IDLE) | |
8897 | return 1; | |
8898 | ||
97fb7a0a | 8899 | /* Try to find first idle CPU */ |
e5c14b1f | 8900 | for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { |
af218122 | 8901 | if (!idle_cpu(cpu)) |
23f0d209 JK |
8902 | continue; |
8903 | ||
8904 | balance_cpu = cpu; | |
8905 | break; | |
8906 | } | |
8907 | ||
8908 | if (balance_cpu == -1) | |
8909 | balance_cpu = group_balance_cpu(sg); | |
8910 | ||
8911 | /* | |
97fb7a0a | 8912 | * First idle CPU or the first CPU(busiest) in this sched group |
23f0d209 JK |
8913 | * is eligible for doing load balancing at this and above domains. |
8914 | */ | |
b0cff9d8 | 8915 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
8916 | } |
8917 | ||
1e3c88bd PZ |
8918 | /* |
8919 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
8920 | * tasks if there is an imbalance. | |
8921 | */ | |
8922 | static int load_balance(int this_cpu, struct rq *this_rq, | |
8923 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 8924 | int *continue_balancing) |
1e3c88bd | 8925 | { |
88b8dac0 | 8926 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 8927 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 8928 | struct sched_group *group; |
1e3c88bd | 8929 | struct rq *busiest; |
8a8c69c3 | 8930 | struct rq_flags rf; |
4ba29684 | 8931 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 8932 | |
8e45cb54 PZ |
8933 | struct lb_env env = { |
8934 | .sd = sd, | |
ddcdf6e7 PZ |
8935 | .dst_cpu = this_cpu, |
8936 | .dst_rq = this_rq, | |
ae4df9d6 | 8937 | .dst_grpmask = sched_group_span(sd->groups), |
8e45cb54 | 8938 | .idle = idle, |
eb95308e | 8939 | .loop_break = sched_nr_migrate_break, |
b9403130 | 8940 | .cpus = cpus, |
0ec8aa00 | 8941 | .fbq_type = all, |
163122b7 | 8942 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
8943 | }; |
8944 | ||
65a4433a | 8945 | cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); |
1e3c88bd | 8946 | |
ae92882e | 8947 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
8948 | |
8949 | redo: | |
23f0d209 JK |
8950 | if (!should_we_balance(&env)) { |
8951 | *continue_balancing = 0; | |
1e3c88bd | 8952 | goto out_balanced; |
23f0d209 | 8953 | } |
1e3c88bd | 8954 | |
23f0d209 | 8955 | group = find_busiest_group(&env); |
1e3c88bd | 8956 | if (!group) { |
ae92882e | 8957 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
8958 | goto out_balanced; |
8959 | } | |
8960 | ||
b9403130 | 8961 | busiest = find_busiest_queue(&env, group); |
1e3c88bd | 8962 | if (!busiest) { |
ae92882e | 8963 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
8964 | goto out_balanced; |
8965 | } | |
8966 | ||
78feefc5 | 8967 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 8968 | |
ae92882e | 8969 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 8970 | |
1aaf90a4 VG |
8971 | env.src_cpu = busiest->cpu; |
8972 | env.src_rq = busiest; | |
8973 | ||
1e3c88bd PZ |
8974 | ld_moved = 0; |
8975 | if (busiest->nr_running > 1) { | |
8976 | /* | |
8977 | * Attempt to move tasks. If find_busiest_group has found | |
8978 | * an imbalance but busiest->nr_running <= 1, the group is | |
8979 | * still unbalanced. ld_moved simply stays zero, so it is | |
8980 | * correctly treated as an imbalance. | |
8981 | */ | |
8e45cb54 | 8982 | env.flags |= LBF_ALL_PINNED; |
c82513e5 | 8983 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 8984 | |
5d6523eb | 8985 | more_balance: |
8a8c69c3 | 8986 | rq_lock_irqsave(busiest, &rf); |
3bed5e21 | 8987 | update_rq_clock(busiest); |
88b8dac0 SV |
8988 | |
8989 | /* | |
8990 | * cur_ld_moved - load moved in current iteration | |
8991 | * ld_moved - cumulative load moved across iterations | |
8992 | */ | |
163122b7 | 8993 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
8994 | |
8995 | /* | |
163122b7 KT |
8996 | * We've detached some tasks from busiest_rq. Every |
8997 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
8998 | * unlock busiest->lock, and we are able to be sure | |
8999 | * that nobody can manipulate the tasks in parallel. | |
9000 | * See task_rq_lock() family for the details. | |
1e3c88bd | 9001 | */ |
163122b7 | 9002 | |
8a8c69c3 | 9003 | rq_unlock(busiest, &rf); |
163122b7 KT |
9004 | |
9005 | if (cur_ld_moved) { | |
9006 | attach_tasks(&env); | |
9007 | ld_moved += cur_ld_moved; | |
9008 | } | |
9009 | ||
8a8c69c3 | 9010 | local_irq_restore(rf.flags); |
88b8dac0 | 9011 | |
f1cd0858 JK |
9012 | if (env.flags & LBF_NEED_BREAK) { |
9013 | env.flags &= ~LBF_NEED_BREAK; | |
9014 | goto more_balance; | |
9015 | } | |
9016 | ||
88b8dac0 SV |
9017 | /* |
9018 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
9019 | * us and move them to an alternate dst_cpu in our sched_group | |
9020 | * where they can run. The upper limit on how many times we | |
97fb7a0a | 9021 | * iterate on same src_cpu is dependent on number of CPUs in our |
88b8dac0 SV |
9022 | * sched_group. |
9023 | * | |
9024 | * This changes load balance semantics a bit on who can move | |
9025 | * load to a given_cpu. In addition to the given_cpu itself | |
9026 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
9027 | * nohz-idle), we now have balance_cpu in a position to move | |
9028 | * load to given_cpu. In rare situations, this may cause | |
9029 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
9030 | * _independently_ and at _same_ time to move some load to | |
9031 | * given_cpu) causing exceess load to be moved to given_cpu. | |
9032 | * This however should not happen so much in practice and | |
9033 | * moreover subsequent load balance cycles should correct the | |
9034 | * excess load moved. | |
9035 | */ | |
6263322c | 9036 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 9037 | |
97fb7a0a | 9038 | /* Prevent to re-select dst_cpu via env's CPUs */ |
7aff2e3a VD |
9039 | cpumask_clear_cpu(env.dst_cpu, env.cpus); |
9040 | ||
78feefc5 | 9041 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 9042 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 9043 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
9044 | env.loop = 0; |
9045 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 9046 | |
88b8dac0 SV |
9047 | /* |
9048 | * Go back to "more_balance" rather than "redo" since we | |
9049 | * need to continue with same src_cpu. | |
9050 | */ | |
9051 | goto more_balance; | |
9052 | } | |
1e3c88bd | 9053 | |
6263322c PZ |
9054 | /* |
9055 | * We failed to reach balance because of affinity. | |
9056 | */ | |
9057 | if (sd_parent) { | |
63b2ca30 | 9058 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 9059 | |
afdeee05 | 9060 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 9061 | *group_imbalance = 1; |
6263322c PZ |
9062 | } |
9063 | ||
1e3c88bd | 9064 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 9065 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 9066 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
65a4433a JH |
9067 | /* |
9068 | * Attempting to continue load balancing at the current | |
9069 | * sched_domain level only makes sense if there are | |
9070 | * active CPUs remaining as possible busiest CPUs to | |
9071 | * pull load from which are not contained within the | |
9072 | * destination group that is receiving any migrated | |
9073 | * load. | |
9074 | */ | |
9075 | if (!cpumask_subset(cpus, env.dst_grpmask)) { | |
bbf18b19 PN |
9076 | env.loop = 0; |
9077 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 9078 | goto redo; |
bbf18b19 | 9079 | } |
afdeee05 | 9080 | goto out_all_pinned; |
1e3c88bd PZ |
9081 | } |
9082 | } | |
9083 | ||
9084 | if (!ld_moved) { | |
ae92882e | 9085 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
9086 | /* |
9087 | * Increment the failure counter only on periodic balance. | |
9088 | * We do not want newidle balance, which can be very | |
9089 | * frequent, pollute the failure counter causing | |
9090 | * excessive cache_hot migrations and active balances. | |
9091 | */ | |
9092 | if (idle != CPU_NEWLY_IDLE) | |
9093 | sd->nr_balance_failed++; | |
1e3c88bd | 9094 | |
bd939f45 | 9095 | if (need_active_balance(&env)) { |
8a8c69c3 PZ |
9096 | unsigned long flags; |
9097 | ||
1e3c88bd PZ |
9098 | raw_spin_lock_irqsave(&busiest->lock, flags); |
9099 | ||
97fb7a0a IM |
9100 | /* |
9101 | * Don't kick the active_load_balance_cpu_stop, | |
9102 | * if the curr task on busiest CPU can't be | |
9103 | * moved to this_cpu: | |
1e3c88bd | 9104 | */ |
0c98d344 | 9105 | if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) { |
1e3c88bd PZ |
9106 | raw_spin_unlock_irqrestore(&busiest->lock, |
9107 | flags); | |
8e45cb54 | 9108 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
9109 | goto out_one_pinned; |
9110 | } | |
9111 | ||
969c7921 TH |
9112 | /* |
9113 | * ->active_balance synchronizes accesses to | |
9114 | * ->active_balance_work. Once set, it's cleared | |
9115 | * only after active load balance is finished. | |
9116 | */ | |
1e3c88bd PZ |
9117 | if (!busiest->active_balance) { |
9118 | busiest->active_balance = 1; | |
9119 | busiest->push_cpu = this_cpu; | |
9120 | active_balance = 1; | |
9121 | } | |
9122 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 9123 | |
bd939f45 | 9124 | if (active_balance) { |
969c7921 TH |
9125 | stop_one_cpu_nowait(cpu_of(busiest), |
9126 | active_load_balance_cpu_stop, busiest, | |
9127 | &busiest->active_balance_work); | |
bd939f45 | 9128 | } |
1e3c88bd | 9129 | |
d02c0711 | 9130 | /* We've kicked active balancing, force task migration. */ |
1e3c88bd PZ |
9131 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
9132 | } | |
9133 | } else | |
9134 | sd->nr_balance_failed = 0; | |
9135 | ||
46a745d9 | 9136 | if (likely(!active_balance) || voluntary_active_balance(&env)) { |
1e3c88bd PZ |
9137 | /* We were unbalanced, so reset the balancing interval */ |
9138 | sd->balance_interval = sd->min_interval; | |
9139 | } else { | |
9140 | /* | |
9141 | * If we've begun active balancing, start to back off. This | |
9142 | * case may not be covered by the all_pinned logic if there | |
9143 | * is only 1 task on the busy runqueue (because we don't call | |
163122b7 | 9144 | * detach_tasks). |
1e3c88bd PZ |
9145 | */ |
9146 | if (sd->balance_interval < sd->max_interval) | |
9147 | sd->balance_interval *= 2; | |
9148 | } | |
9149 | ||
1e3c88bd PZ |
9150 | goto out; |
9151 | ||
9152 | out_balanced: | |
afdeee05 VG |
9153 | /* |
9154 | * We reach balance although we may have faced some affinity | |
9155 | * constraints. Clear the imbalance flag if it was set. | |
9156 | */ | |
9157 | if (sd_parent) { | |
9158 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; | |
9159 | ||
9160 | if (*group_imbalance) | |
9161 | *group_imbalance = 0; | |
9162 | } | |
9163 | ||
9164 | out_all_pinned: | |
9165 | /* | |
9166 | * We reach balance because all tasks are pinned at this level so | |
9167 | * we can't migrate them. Let the imbalance flag set so parent level | |
9168 | * can try to migrate them. | |
9169 | */ | |
ae92882e | 9170 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
9171 | |
9172 | sd->nr_balance_failed = 0; | |
9173 | ||
9174 | out_one_pinned: | |
3f130a37 VS |
9175 | ld_moved = 0; |
9176 | ||
9177 | /* | |
9178 | * idle_balance() disregards balance intervals, so we could repeatedly | |
9179 | * reach this code, which would lead to balance_interval skyrocketting | |
9180 | * in a short amount of time. Skip the balance_interval increase logic | |
9181 | * to avoid that. | |
9182 | */ | |
9183 | if (env.idle == CPU_NEWLY_IDLE) | |
9184 | goto out; | |
9185 | ||
1e3c88bd | 9186 | /* tune up the balancing interval */ |
47b7aee1 VS |
9187 | if ((env.flags & LBF_ALL_PINNED && |
9188 | sd->balance_interval < MAX_PINNED_INTERVAL) || | |
9189 | sd->balance_interval < sd->max_interval) | |
1e3c88bd | 9190 | sd->balance_interval *= 2; |
1e3c88bd | 9191 | out: |
1e3c88bd PZ |
9192 | return ld_moved; |
9193 | } | |
9194 | ||
52a08ef1 JL |
9195 | static inline unsigned long |
9196 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
9197 | { | |
9198 | unsigned long interval = sd->balance_interval; | |
9199 | ||
9200 | if (cpu_busy) | |
9201 | interval *= sd->busy_factor; | |
9202 | ||
9203 | /* scale ms to jiffies */ | |
9204 | interval = msecs_to_jiffies(interval); | |
9205 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
9206 | ||
9207 | return interval; | |
9208 | } | |
9209 | ||
9210 | static inline void | |
31851a98 | 9211 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
9212 | { |
9213 | unsigned long interval, next; | |
9214 | ||
31851a98 LY |
9215 | /* used by idle balance, so cpu_busy = 0 */ |
9216 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
9217 | next = sd->last_balance + interval; |
9218 | ||
9219 | if (time_after(*next_balance, next)) | |
9220 | *next_balance = next; | |
9221 | } | |
9222 | ||
1e3c88bd | 9223 | /* |
97fb7a0a | 9224 | * active_load_balance_cpu_stop is run by the CPU stopper. It pushes |
969c7921 TH |
9225 | * running tasks off the busiest CPU onto idle CPUs. It requires at |
9226 | * least 1 task to be running on each physical CPU where possible, and | |
9227 | * avoids physical / logical imbalances. | |
1e3c88bd | 9228 | */ |
969c7921 | 9229 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 9230 | { |
969c7921 TH |
9231 | struct rq *busiest_rq = data; |
9232 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 9233 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 9234 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 9235 | struct sched_domain *sd; |
e5673f28 | 9236 | struct task_struct *p = NULL; |
8a8c69c3 | 9237 | struct rq_flags rf; |
969c7921 | 9238 | |
8a8c69c3 | 9239 | rq_lock_irq(busiest_rq, &rf); |
edd8e41d PZ |
9240 | /* |
9241 | * Between queueing the stop-work and running it is a hole in which | |
9242 | * CPUs can become inactive. We should not move tasks from or to | |
9243 | * inactive CPUs. | |
9244 | */ | |
9245 | if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) | |
9246 | goto out_unlock; | |
969c7921 | 9247 | |
97fb7a0a | 9248 | /* Make sure the requested CPU hasn't gone down in the meantime: */ |
969c7921 TH |
9249 | if (unlikely(busiest_cpu != smp_processor_id() || |
9250 | !busiest_rq->active_balance)) | |
9251 | goto out_unlock; | |
1e3c88bd PZ |
9252 | |
9253 | /* Is there any task to move? */ | |
9254 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 9255 | goto out_unlock; |
1e3c88bd PZ |
9256 | |
9257 | /* | |
9258 | * This condition is "impossible", if it occurs | |
9259 | * we need to fix it. Originally reported by | |
97fb7a0a | 9260 | * Bjorn Helgaas on a 128-CPU setup. |
1e3c88bd PZ |
9261 | */ |
9262 | BUG_ON(busiest_rq == target_rq); | |
9263 | ||
1e3c88bd | 9264 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 9265 | rcu_read_lock(); |
1e3c88bd PZ |
9266 | for_each_domain(target_cpu, sd) { |
9267 | if ((sd->flags & SD_LOAD_BALANCE) && | |
9268 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
9269 | break; | |
9270 | } | |
9271 | ||
9272 | if (likely(sd)) { | |
8e45cb54 PZ |
9273 | struct lb_env env = { |
9274 | .sd = sd, | |
ddcdf6e7 PZ |
9275 | .dst_cpu = target_cpu, |
9276 | .dst_rq = target_rq, | |
9277 | .src_cpu = busiest_rq->cpu, | |
9278 | .src_rq = busiest_rq, | |
8e45cb54 | 9279 | .idle = CPU_IDLE, |
65a4433a JH |
9280 | /* |
9281 | * can_migrate_task() doesn't need to compute new_dst_cpu | |
9282 | * for active balancing. Since we have CPU_IDLE, but no | |
9283 | * @dst_grpmask we need to make that test go away with lying | |
9284 | * about DST_PINNED. | |
9285 | */ | |
9286 | .flags = LBF_DST_PINNED, | |
8e45cb54 PZ |
9287 | }; |
9288 | ||
ae92882e | 9289 | schedstat_inc(sd->alb_count); |
3bed5e21 | 9290 | update_rq_clock(busiest_rq); |
1e3c88bd | 9291 | |
e5673f28 | 9292 | p = detach_one_task(&env); |
d02c0711 | 9293 | if (p) { |
ae92882e | 9294 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
9295 | /* Active balancing done, reset the failure counter. */ |
9296 | sd->nr_balance_failed = 0; | |
9297 | } else { | |
ae92882e | 9298 | schedstat_inc(sd->alb_failed); |
d02c0711 | 9299 | } |
1e3c88bd | 9300 | } |
dce840a0 | 9301 | rcu_read_unlock(); |
969c7921 TH |
9302 | out_unlock: |
9303 | busiest_rq->active_balance = 0; | |
8a8c69c3 | 9304 | rq_unlock(busiest_rq, &rf); |
e5673f28 KT |
9305 | |
9306 | if (p) | |
9307 | attach_one_task(target_rq, p); | |
9308 | ||
9309 | local_irq_enable(); | |
9310 | ||
969c7921 | 9311 | return 0; |
1e3c88bd PZ |
9312 | } |
9313 | ||
af3fe03c PZ |
9314 | static DEFINE_SPINLOCK(balancing); |
9315 | ||
9316 | /* | |
9317 | * Scale the max load_balance interval with the number of CPUs in the system. | |
9318 | * This trades load-balance latency on larger machines for less cross talk. | |
9319 | */ | |
9320 | void update_max_interval(void) | |
9321 | { | |
9322 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
9323 | } | |
9324 | ||
9325 | /* | |
9326 | * It checks each scheduling domain to see if it is due to be balanced, | |
9327 | * and initiates a balancing operation if so. | |
9328 | * | |
9329 | * Balancing parameters are set up in init_sched_domains. | |
9330 | */ | |
9331 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) | |
9332 | { | |
9333 | int continue_balancing = 1; | |
9334 | int cpu = rq->cpu; | |
9335 | unsigned long interval; | |
9336 | struct sched_domain *sd; | |
9337 | /* Earliest time when we have to do rebalance again */ | |
9338 | unsigned long next_balance = jiffies + 60*HZ; | |
9339 | int update_next_balance = 0; | |
9340 | int need_serialize, need_decay = 0; | |
9341 | u64 max_cost = 0; | |
9342 | ||
9343 | rcu_read_lock(); | |
9344 | for_each_domain(cpu, sd) { | |
9345 | /* | |
9346 | * Decay the newidle max times here because this is a regular | |
9347 | * visit to all the domains. Decay ~1% per second. | |
9348 | */ | |
9349 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
9350 | sd->max_newidle_lb_cost = | |
9351 | (sd->max_newidle_lb_cost * 253) / 256; | |
9352 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
9353 | need_decay = 1; | |
9354 | } | |
9355 | max_cost += sd->max_newidle_lb_cost; | |
9356 | ||
9357 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
9358 | continue; | |
9359 | ||
9360 | /* | |
9361 | * Stop the load balance at this level. There is another | |
9362 | * CPU in our sched group which is doing load balancing more | |
9363 | * actively. | |
9364 | */ | |
9365 | if (!continue_balancing) { | |
9366 | if (need_decay) | |
9367 | continue; | |
9368 | break; | |
9369 | } | |
9370 | ||
9371 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); | |
9372 | ||
9373 | need_serialize = sd->flags & SD_SERIALIZE; | |
9374 | if (need_serialize) { | |
9375 | if (!spin_trylock(&balancing)) | |
9376 | goto out; | |
9377 | } | |
9378 | ||
9379 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
9380 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { | |
9381 | /* | |
9382 | * The LBF_DST_PINNED logic could have changed | |
9383 | * env->dst_cpu, so we can't know our idle | |
9384 | * state even if we migrated tasks. Update it. | |
9385 | */ | |
9386 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; | |
9387 | } | |
9388 | sd->last_balance = jiffies; | |
9389 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); | |
9390 | } | |
9391 | if (need_serialize) | |
9392 | spin_unlock(&balancing); | |
9393 | out: | |
9394 | if (time_after(next_balance, sd->last_balance + interval)) { | |
9395 | next_balance = sd->last_balance + interval; | |
9396 | update_next_balance = 1; | |
9397 | } | |
9398 | } | |
9399 | if (need_decay) { | |
9400 | /* | |
9401 | * Ensure the rq-wide value also decays but keep it at a | |
9402 | * reasonable floor to avoid funnies with rq->avg_idle. | |
9403 | */ | |
9404 | rq->max_idle_balance_cost = | |
9405 | max((u64)sysctl_sched_migration_cost, max_cost); | |
9406 | } | |
9407 | rcu_read_unlock(); | |
9408 | ||
9409 | /* | |
9410 | * next_balance will be updated only when there is a need. | |
9411 | * When the cpu is attached to null domain for ex, it will not be | |
9412 | * updated. | |
9413 | */ | |
9414 | if (likely(update_next_balance)) { | |
9415 | rq->next_balance = next_balance; | |
9416 | ||
9417 | #ifdef CONFIG_NO_HZ_COMMON | |
9418 | /* | |
9419 | * If this CPU has been elected to perform the nohz idle | |
9420 | * balance. Other idle CPUs have already rebalanced with | |
9421 | * nohz_idle_balance() and nohz.next_balance has been | |
9422 | * updated accordingly. This CPU is now running the idle load | |
9423 | * balance for itself and we need to update the | |
9424 | * nohz.next_balance accordingly. | |
9425 | */ | |
9426 | if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance)) | |
9427 | nohz.next_balance = rq->next_balance; | |
9428 | #endif | |
9429 | } | |
9430 | } | |
9431 | ||
d987fc7f MG |
9432 | static inline int on_null_domain(struct rq *rq) |
9433 | { | |
9434 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
9435 | } | |
9436 | ||
3451d024 | 9437 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
9438 | /* |
9439 | * idle load balancing details | |
83cd4fe2 VP |
9440 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
9441 | * needed, they will kick the idle load balancer, which then does idle | |
9442 | * load balancing for all the idle CPUs. | |
9443 | */ | |
1e3c88bd | 9444 | |
3dd0337d | 9445 | static inline int find_new_ilb(void) |
1e3c88bd | 9446 | { |
0b005cf5 | 9447 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 9448 | |
786d6dc7 SS |
9449 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
9450 | return ilb; | |
9451 | ||
9452 | return nr_cpu_ids; | |
1e3c88bd | 9453 | } |
1e3c88bd | 9454 | |
83cd4fe2 VP |
9455 | /* |
9456 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
9457 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
9458 | * CPU (if there is one). | |
9459 | */ | |
a4064fb6 | 9460 | static void kick_ilb(unsigned int flags) |
83cd4fe2 VP |
9461 | { |
9462 | int ilb_cpu; | |
9463 | ||
9464 | nohz.next_balance++; | |
9465 | ||
3dd0337d | 9466 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 9467 | |
0b005cf5 SS |
9468 | if (ilb_cpu >= nr_cpu_ids) |
9469 | return; | |
83cd4fe2 | 9470 | |
a4064fb6 | 9471 | flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu)); |
b7031a02 | 9472 | if (flags & NOHZ_KICK_MASK) |
1c792db7 | 9473 | return; |
4550487a | 9474 | |
1c792db7 SS |
9475 | /* |
9476 | * Use smp_send_reschedule() instead of resched_cpu(). | |
97fb7a0a | 9477 | * This way we generate a sched IPI on the target CPU which |
1c792db7 SS |
9478 | * is idle. And the softirq performing nohz idle load balance |
9479 | * will be run before returning from the IPI. | |
9480 | */ | |
9481 | smp_send_reschedule(ilb_cpu); | |
4550487a PZ |
9482 | } |
9483 | ||
9484 | /* | |
9485 | * Current heuristic for kicking the idle load balancer in the presence | |
9486 | * of an idle cpu in the system. | |
9487 | * - This rq has more than one task. | |
9488 | * - This rq has at least one CFS task and the capacity of the CPU is | |
9489 | * significantly reduced because of RT tasks or IRQs. | |
9490 | * - At parent of LLC scheduler domain level, this cpu's scheduler group has | |
9491 | * multiple busy cpu. | |
9492 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler | |
9493 | * domain span are idle. | |
9494 | */ | |
9495 | static void nohz_balancer_kick(struct rq *rq) | |
9496 | { | |
9497 | unsigned long now = jiffies; | |
9498 | struct sched_domain_shared *sds; | |
9499 | struct sched_domain *sd; | |
9500 | int nr_busy, i, cpu = rq->cpu; | |
a4064fb6 | 9501 | unsigned int flags = 0; |
4550487a PZ |
9502 | |
9503 | if (unlikely(rq->idle_balance)) | |
9504 | return; | |
9505 | ||
9506 | /* | |
9507 | * We may be recently in ticked or tickless idle mode. At the first | |
9508 | * busy tick after returning from idle, we will update the busy stats. | |
9509 | */ | |
00357f5e | 9510 | nohz_balance_exit_idle(rq); |
4550487a PZ |
9511 | |
9512 | /* | |
9513 | * None are in tickless mode and hence no need for NOHZ idle load | |
9514 | * balancing. | |
9515 | */ | |
9516 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
9517 | return; | |
9518 | ||
f643ea22 VG |
9519 | if (READ_ONCE(nohz.has_blocked) && |
9520 | time_after(now, READ_ONCE(nohz.next_blocked))) | |
a4064fb6 PZ |
9521 | flags = NOHZ_STATS_KICK; |
9522 | ||
4550487a | 9523 | if (time_before(now, nohz.next_balance)) |
a4064fb6 | 9524 | goto out; |
4550487a | 9525 | |
5fbdfae5 | 9526 | if (rq->nr_running >= 2 || rq->misfit_task_load) { |
a4064fb6 | 9527 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
9528 | goto out; |
9529 | } | |
9530 | ||
9531 | rcu_read_lock(); | |
9532 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
9533 | if (sds) { | |
9534 | /* | |
9535 | * XXX: write a coherent comment on why we do this. | |
9536 | * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com | |
9537 | */ | |
9538 | nr_busy = atomic_read(&sds->nr_busy_cpus); | |
9539 | if (nr_busy > 1) { | |
a4064fb6 | 9540 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
9541 | goto unlock; |
9542 | } | |
9543 | ||
9544 | } | |
9545 | ||
9546 | sd = rcu_dereference(rq->sd); | |
9547 | if (sd) { | |
9548 | if ((rq->cfs.h_nr_running >= 1) && | |
9549 | check_cpu_capacity(rq, sd)) { | |
a4064fb6 | 9550 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
9551 | goto unlock; |
9552 | } | |
9553 | } | |
9554 | ||
011b27bb | 9555 | sd = rcu_dereference(per_cpu(sd_asym_packing, cpu)); |
4550487a PZ |
9556 | if (sd) { |
9557 | for_each_cpu(i, sched_domain_span(sd)) { | |
9558 | if (i == cpu || | |
9559 | !cpumask_test_cpu(i, nohz.idle_cpus_mask)) | |
9560 | continue; | |
9561 | ||
9562 | if (sched_asym_prefer(i, cpu)) { | |
a4064fb6 | 9563 | flags = NOHZ_KICK_MASK; |
4550487a PZ |
9564 | goto unlock; |
9565 | } | |
9566 | } | |
9567 | } | |
9568 | unlock: | |
9569 | rcu_read_unlock(); | |
9570 | out: | |
a4064fb6 PZ |
9571 | if (flags) |
9572 | kick_ilb(flags); | |
83cd4fe2 VP |
9573 | } |
9574 | ||
00357f5e | 9575 | static void set_cpu_sd_state_busy(int cpu) |
71325960 | 9576 | { |
00357f5e | 9577 | struct sched_domain *sd; |
a22e47a4 | 9578 | |
00357f5e PZ |
9579 | rcu_read_lock(); |
9580 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); | |
a22e47a4 | 9581 | |
00357f5e PZ |
9582 | if (!sd || !sd->nohz_idle) |
9583 | goto unlock; | |
9584 | sd->nohz_idle = 0; | |
9585 | ||
9586 | atomic_inc(&sd->shared->nr_busy_cpus); | |
9587 | unlock: | |
9588 | rcu_read_unlock(); | |
71325960 SS |
9589 | } |
9590 | ||
00357f5e PZ |
9591 | void nohz_balance_exit_idle(struct rq *rq) |
9592 | { | |
9593 | SCHED_WARN_ON(rq != this_rq()); | |
9594 | ||
9595 | if (likely(!rq->nohz_tick_stopped)) | |
9596 | return; | |
9597 | ||
9598 | rq->nohz_tick_stopped = 0; | |
9599 | cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); | |
9600 | atomic_dec(&nohz.nr_cpus); | |
9601 | ||
9602 | set_cpu_sd_state_busy(rq->cpu); | |
9603 | } | |
9604 | ||
9605 | static void set_cpu_sd_state_idle(int cpu) | |
69e1e811 SS |
9606 | { |
9607 | struct sched_domain *sd; | |
69e1e811 | 9608 | |
69e1e811 | 9609 | rcu_read_lock(); |
0e369d75 | 9610 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
9611 | |
9612 | if (!sd || sd->nohz_idle) | |
9613 | goto unlock; | |
9614 | sd->nohz_idle = 1; | |
9615 | ||
0e369d75 | 9616 | atomic_dec(&sd->shared->nr_busy_cpus); |
25f55d9d | 9617 | unlock: |
69e1e811 SS |
9618 | rcu_read_unlock(); |
9619 | } | |
9620 | ||
1e3c88bd | 9621 | /* |
97fb7a0a | 9622 | * This routine will record that the CPU is going idle with tick stopped. |
0b005cf5 | 9623 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 9624 | */ |
c1cc017c | 9625 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 9626 | { |
00357f5e PZ |
9627 | struct rq *rq = cpu_rq(cpu); |
9628 | ||
9629 | SCHED_WARN_ON(cpu != smp_processor_id()); | |
9630 | ||
97fb7a0a | 9631 | /* If this CPU is going down, then nothing needs to be done: */ |
71325960 SS |
9632 | if (!cpu_active(cpu)) |
9633 | return; | |
9634 | ||
387bc8b5 | 9635 | /* Spare idle load balancing on CPUs that don't want to be disturbed: */ |
de201559 | 9636 | if (!housekeeping_cpu(cpu, HK_FLAG_SCHED)) |
387bc8b5 FW |
9637 | return; |
9638 | ||
f643ea22 VG |
9639 | /* |
9640 | * Can be set safely without rq->lock held | |
9641 | * If a clear happens, it will have evaluated last additions because | |
9642 | * rq->lock is held during the check and the clear | |
9643 | */ | |
9644 | rq->has_blocked_load = 1; | |
9645 | ||
9646 | /* | |
9647 | * The tick is still stopped but load could have been added in the | |
9648 | * meantime. We set the nohz.has_blocked flag to trig a check of the | |
9649 | * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear | |
9650 | * of nohz.has_blocked can only happen after checking the new load | |
9651 | */ | |
00357f5e | 9652 | if (rq->nohz_tick_stopped) |
f643ea22 | 9653 | goto out; |
1e3c88bd | 9654 | |
97fb7a0a | 9655 | /* If we're a completely isolated CPU, we don't play: */ |
00357f5e | 9656 | if (on_null_domain(rq)) |
d987fc7f MG |
9657 | return; |
9658 | ||
00357f5e PZ |
9659 | rq->nohz_tick_stopped = 1; |
9660 | ||
c1cc017c AS |
9661 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
9662 | atomic_inc(&nohz.nr_cpus); | |
00357f5e | 9663 | |
f643ea22 VG |
9664 | /* |
9665 | * Ensures that if nohz_idle_balance() fails to observe our | |
9666 | * @idle_cpus_mask store, it must observe the @has_blocked | |
9667 | * store. | |
9668 | */ | |
9669 | smp_mb__after_atomic(); | |
9670 | ||
00357f5e | 9671 | set_cpu_sd_state_idle(cpu); |
f643ea22 VG |
9672 | |
9673 | out: | |
9674 | /* | |
9675 | * Each time a cpu enter idle, we assume that it has blocked load and | |
9676 | * enable the periodic update of the load of idle cpus | |
9677 | */ | |
9678 | WRITE_ONCE(nohz.has_blocked, 1); | |
1e3c88bd | 9679 | } |
1e3c88bd | 9680 | |
1e3c88bd | 9681 | /* |
31e77c93 VG |
9682 | * Internal function that runs load balance for all idle cpus. The load balance |
9683 | * can be a simple update of blocked load or a complete load balance with | |
9684 | * tasks movement depending of flags. | |
9685 | * The function returns false if the loop has stopped before running | |
9686 | * through all idle CPUs. | |
1e3c88bd | 9687 | */ |
31e77c93 VG |
9688 | static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags, |
9689 | enum cpu_idle_type idle) | |
83cd4fe2 | 9690 | { |
c5afb6a8 | 9691 | /* Earliest time when we have to do rebalance again */ |
a4064fb6 PZ |
9692 | unsigned long now = jiffies; |
9693 | unsigned long next_balance = now + 60*HZ; | |
f643ea22 | 9694 | bool has_blocked_load = false; |
c5afb6a8 | 9695 | int update_next_balance = 0; |
b7031a02 | 9696 | int this_cpu = this_rq->cpu; |
b7031a02 | 9697 | int balance_cpu; |
31e77c93 | 9698 | int ret = false; |
b7031a02 | 9699 | struct rq *rq; |
83cd4fe2 | 9700 | |
b7031a02 | 9701 | SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK); |
83cd4fe2 | 9702 | |
f643ea22 VG |
9703 | /* |
9704 | * We assume there will be no idle load after this update and clear | |
9705 | * the has_blocked flag. If a cpu enters idle in the mean time, it will | |
9706 | * set the has_blocked flag and trig another update of idle load. | |
9707 | * Because a cpu that becomes idle, is added to idle_cpus_mask before | |
9708 | * setting the flag, we are sure to not clear the state and not | |
9709 | * check the load of an idle cpu. | |
9710 | */ | |
9711 | WRITE_ONCE(nohz.has_blocked, 0); | |
9712 | ||
9713 | /* | |
9714 | * Ensures that if we miss the CPU, we must see the has_blocked | |
9715 | * store from nohz_balance_enter_idle(). | |
9716 | */ | |
9717 | smp_mb(); | |
9718 | ||
83cd4fe2 | 9719 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { |
8a6d42d1 | 9720 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
9721 | continue; |
9722 | ||
9723 | /* | |
97fb7a0a IM |
9724 | * If this CPU gets work to do, stop the load balancing |
9725 | * work being done for other CPUs. Next load | |
83cd4fe2 VP |
9726 | * balancing owner will pick it up. |
9727 | */ | |
f643ea22 VG |
9728 | if (need_resched()) { |
9729 | has_blocked_load = true; | |
9730 | goto abort; | |
9731 | } | |
83cd4fe2 | 9732 | |
5ed4f1d9 VG |
9733 | rq = cpu_rq(balance_cpu); |
9734 | ||
63928384 | 9735 | has_blocked_load |= update_nohz_stats(rq, true); |
f643ea22 | 9736 | |
ed61bbc6 TC |
9737 | /* |
9738 | * If time for next balance is due, | |
9739 | * do the balance. | |
9740 | */ | |
9741 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8a8c69c3 PZ |
9742 | struct rq_flags rf; |
9743 | ||
31e77c93 | 9744 | rq_lock_irqsave(rq, &rf); |
ed61bbc6 | 9745 | update_rq_clock(rq); |
cee1afce | 9746 | cpu_load_update_idle(rq); |
31e77c93 | 9747 | rq_unlock_irqrestore(rq, &rf); |
8a8c69c3 | 9748 | |
b7031a02 PZ |
9749 | if (flags & NOHZ_BALANCE_KICK) |
9750 | rebalance_domains(rq, CPU_IDLE); | |
ed61bbc6 | 9751 | } |
83cd4fe2 | 9752 | |
c5afb6a8 VG |
9753 | if (time_after(next_balance, rq->next_balance)) { |
9754 | next_balance = rq->next_balance; | |
9755 | update_next_balance = 1; | |
9756 | } | |
83cd4fe2 | 9757 | } |
c5afb6a8 | 9758 | |
31e77c93 VG |
9759 | /* Newly idle CPU doesn't need an update */ |
9760 | if (idle != CPU_NEWLY_IDLE) { | |
9761 | update_blocked_averages(this_cpu); | |
9762 | has_blocked_load |= this_rq->has_blocked_load; | |
9763 | } | |
9764 | ||
b7031a02 PZ |
9765 | if (flags & NOHZ_BALANCE_KICK) |
9766 | rebalance_domains(this_rq, CPU_IDLE); | |
9767 | ||
f643ea22 VG |
9768 | WRITE_ONCE(nohz.next_blocked, |
9769 | now + msecs_to_jiffies(LOAD_AVG_PERIOD)); | |
9770 | ||
31e77c93 VG |
9771 | /* The full idle balance loop has been done */ |
9772 | ret = true; | |
9773 | ||
f643ea22 VG |
9774 | abort: |
9775 | /* There is still blocked load, enable periodic update */ | |
9776 | if (has_blocked_load) | |
9777 | WRITE_ONCE(nohz.has_blocked, 1); | |
a4064fb6 | 9778 | |
c5afb6a8 VG |
9779 | /* |
9780 | * next_balance will be updated only when there is a need. | |
9781 | * When the CPU is attached to null domain for ex, it will not be | |
9782 | * updated. | |
9783 | */ | |
9784 | if (likely(update_next_balance)) | |
9785 | nohz.next_balance = next_balance; | |
b7031a02 | 9786 | |
31e77c93 VG |
9787 | return ret; |
9788 | } | |
9789 | ||
9790 | /* | |
9791 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the | |
9792 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
9793 | */ | |
9794 | static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) | |
9795 | { | |
9796 | int this_cpu = this_rq->cpu; | |
9797 | unsigned int flags; | |
9798 | ||
9799 | if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK)) | |
9800 | return false; | |
9801 | ||
9802 | if (idle != CPU_IDLE) { | |
9803 | atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu)); | |
9804 | return false; | |
9805 | } | |
9806 | ||
80eb8657 | 9807 | /* could be _relaxed() */ |
31e77c93 VG |
9808 | flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu)); |
9809 | if (!(flags & NOHZ_KICK_MASK)) | |
9810 | return false; | |
9811 | ||
9812 | _nohz_idle_balance(this_rq, flags, idle); | |
9813 | ||
b7031a02 | 9814 | return true; |
83cd4fe2 | 9815 | } |
31e77c93 VG |
9816 | |
9817 | static void nohz_newidle_balance(struct rq *this_rq) | |
9818 | { | |
9819 | int this_cpu = this_rq->cpu; | |
9820 | ||
9821 | /* | |
9822 | * This CPU doesn't want to be disturbed by scheduler | |
9823 | * housekeeping | |
9824 | */ | |
9825 | if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED)) | |
9826 | return; | |
9827 | ||
9828 | /* Will wake up very soon. No time for doing anything else*/ | |
9829 | if (this_rq->avg_idle < sysctl_sched_migration_cost) | |
9830 | return; | |
9831 | ||
9832 | /* Don't need to update blocked load of idle CPUs*/ | |
9833 | if (!READ_ONCE(nohz.has_blocked) || | |
9834 | time_before(jiffies, READ_ONCE(nohz.next_blocked))) | |
9835 | return; | |
9836 | ||
9837 | raw_spin_unlock(&this_rq->lock); | |
9838 | /* | |
9839 | * This CPU is going to be idle and blocked load of idle CPUs | |
9840 | * need to be updated. Run the ilb locally as it is a good | |
9841 | * candidate for ilb instead of waking up another idle CPU. | |
9842 | * Kick an normal ilb if we failed to do the update. | |
9843 | */ | |
9844 | if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE)) | |
9845 | kick_ilb(NOHZ_STATS_KICK); | |
9846 | raw_spin_lock(&this_rq->lock); | |
9847 | } | |
9848 | ||
dd707247 PZ |
9849 | #else /* !CONFIG_NO_HZ_COMMON */ |
9850 | static inline void nohz_balancer_kick(struct rq *rq) { } | |
9851 | ||
31e77c93 | 9852 | static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
b7031a02 PZ |
9853 | { |
9854 | return false; | |
9855 | } | |
31e77c93 VG |
9856 | |
9857 | static inline void nohz_newidle_balance(struct rq *this_rq) { } | |
dd707247 | 9858 | #endif /* CONFIG_NO_HZ_COMMON */ |
83cd4fe2 | 9859 | |
47ea5412 PZ |
9860 | /* |
9861 | * idle_balance is called by schedule() if this_cpu is about to become | |
9862 | * idle. Attempts to pull tasks from other CPUs. | |
9863 | */ | |
9864 | static int idle_balance(struct rq *this_rq, struct rq_flags *rf) | |
9865 | { | |
9866 | unsigned long next_balance = jiffies + HZ; | |
9867 | int this_cpu = this_rq->cpu; | |
9868 | struct sched_domain *sd; | |
9869 | int pulled_task = 0; | |
9870 | u64 curr_cost = 0; | |
9871 | ||
9872 | /* | |
9873 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
9874 | * measure the duration of idle_balance() as idle time. | |
9875 | */ | |
9876 | this_rq->idle_stamp = rq_clock(this_rq); | |
9877 | ||
9878 | /* | |
9879 | * Do not pull tasks towards !active CPUs... | |
9880 | */ | |
9881 | if (!cpu_active(this_cpu)) | |
9882 | return 0; | |
9883 | ||
9884 | /* | |
9885 | * This is OK, because current is on_cpu, which avoids it being picked | |
9886 | * for load-balance and preemption/IRQs are still disabled avoiding | |
9887 | * further scheduler activity on it and we're being very careful to | |
9888 | * re-start the picking loop. | |
9889 | */ | |
9890 | rq_unpin_lock(this_rq, rf); | |
9891 | ||
9892 | if (this_rq->avg_idle < sysctl_sched_migration_cost || | |
e90c8fe1 | 9893 | !READ_ONCE(this_rq->rd->overload)) { |
31e77c93 | 9894 | |
47ea5412 PZ |
9895 | rcu_read_lock(); |
9896 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
9897 | if (sd) | |
9898 | update_next_balance(sd, &next_balance); | |
9899 | rcu_read_unlock(); | |
9900 | ||
31e77c93 VG |
9901 | nohz_newidle_balance(this_rq); |
9902 | ||
47ea5412 PZ |
9903 | goto out; |
9904 | } | |
9905 | ||
9906 | raw_spin_unlock(&this_rq->lock); | |
9907 | ||
9908 | update_blocked_averages(this_cpu); | |
9909 | rcu_read_lock(); | |
9910 | for_each_domain(this_cpu, sd) { | |
9911 | int continue_balancing = 1; | |
9912 | u64 t0, domain_cost; | |
9913 | ||
9914 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
9915 | continue; | |
9916 | ||
9917 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { | |
9918 | update_next_balance(sd, &next_balance); | |
9919 | break; | |
9920 | } | |
9921 | ||
9922 | if (sd->flags & SD_BALANCE_NEWIDLE) { | |
9923 | t0 = sched_clock_cpu(this_cpu); | |
9924 | ||
9925 | pulled_task = load_balance(this_cpu, this_rq, | |
9926 | sd, CPU_NEWLY_IDLE, | |
9927 | &continue_balancing); | |
9928 | ||
9929 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
9930 | if (domain_cost > sd->max_newidle_lb_cost) | |
9931 | sd->max_newidle_lb_cost = domain_cost; | |
9932 | ||
9933 | curr_cost += domain_cost; | |
9934 | } | |
9935 | ||
9936 | update_next_balance(sd, &next_balance); | |
9937 | ||
9938 | /* | |
9939 | * Stop searching for tasks to pull if there are | |
9940 | * now runnable tasks on this rq. | |
9941 | */ | |
9942 | if (pulled_task || this_rq->nr_running > 0) | |
9943 | break; | |
9944 | } | |
9945 | rcu_read_unlock(); | |
9946 | ||
9947 | raw_spin_lock(&this_rq->lock); | |
9948 | ||
9949 | if (curr_cost > this_rq->max_idle_balance_cost) | |
9950 | this_rq->max_idle_balance_cost = curr_cost; | |
9951 | ||
457be908 | 9952 | out: |
47ea5412 PZ |
9953 | /* |
9954 | * While browsing the domains, we released the rq lock, a task could | |
9955 | * have been enqueued in the meantime. Since we're not going idle, | |
9956 | * pretend we pulled a task. | |
9957 | */ | |
9958 | if (this_rq->cfs.h_nr_running && !pulled_task) | |
9959 | pulled_task = 1; | |
9960 | ||
47ea5412 PZ |
9961 | /* Move the next balance forward */ |
9962 | if (time_after(this_rq->next_balance, next_balance)) | |
9963 | this_rq->next_balance = next_balance; | |
9964 | ||
9965 | /* Is there a task of a high priority class? */ | |
9966 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) | |
9967 | pulled_task = -1; | |
9968 | ||
9969 | if (pulled_task) | |
9970 | this_rq->idle_stamp = 0; | |
9971 | ||
9972 | rq_repin_lock(this_rq, rf); | |
9973 | ||
9974 | return pulled_task; | |
9975 | } | |
9976 | ||
83cd4fe2 VP |
9977 | /* |
9978 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
9979 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
9980 | */ | |
0766f788 | 9981 | static __latent_entropy void run_rebalance_domains(struct softirq_action *h) |
1e3c88bd | 9982 | { |
208cb16b | 9983 | struct rq *this_rq = this_rq(); |
6eb57e0d | 9984 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
9985 | CPU_IDLE : CPU_NOT_IDLE; |
9986 | ||
1e3c88bd | 9987 | /* |
97fb7a0a IM |
9988 | * If this CPU has a pending nohz_balance_kick, then do the |
9989 | * balancing on behalf of the other idle CPUs whose ticks are | |
d4573c3e | 9990 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
97fb7a0a | 9991 | * give the idle CPUs a chance to load balance. Else we may |
d4573c3e PM |
9992 | * load balance only within the local sched_domain hierarchy |
9993 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 9994 | */ |
b7031a02 PZ |
9995 | if (nohz_idle_balance(this_rq, idle)) |
9996 | return; | |
9997 | ||
9998 | /* normal load balance */ | |
9999 | update_blocked_averages(this_rq->cpu); | |
d4573c3e | 10000 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
10001 | } |
10002 | ||
1e3c88bd PZ |
10003 | /* |
10004 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 10005 | */ |
7caff66f | 10006 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 10007 | { |
1e3c88bd | 10008 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
10009 | if (unlikely(on_null_domain(rq))) |
10010 | return; | |
10011 | ||
10012 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 10013 | raise_softirq(SCHED_SOFTIRQ); |
4550487a PZ |
10014 | |
10015 | nohz_balancer_kick(rq); | |
1e3c88bd PZ |
10016 | } |
10017 | ||
0bcdcf28 CE |
10018 | static void rq_online_fair(struct rq *rq) |
10019 | { | |
10020 | update_sysctl(); | |
0e59bdae KT |
10021 | |
10022 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
10023 | } |
10024 | ||
10025 | static void rq_offline_fair(struct rq *rq) | |
10026 | { | |
10027 | update_sysctl(); | |
a4c96ae3 PB |
10028 | |
10029 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
10030 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
10031 | } |
10032 | ||
55e12e5e | 10033 | #endif /* CONFIG_SMP */ |
e1d1484f | 10034 | |
bf0f6f24 | 10035 | /* |
d84b3131 FW |
10036 | * scheduler tick hitting a task of our scheduling class. |
10037 | * | |
10038 | * NOTE: This function can be called remotely by the tick offload that | |
10039 | * goes along full dynticks. Therefore no local assumption can be made | |
10040 | * and everything must be accessed through the @rq and @curr passed in | |
10041 | * parameters. | |
bf0f6f24 | 10042 | */ |
8f4d37ec | 10043 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
10044 | { |
10045 | struct cfs_rq *cfs_rq; | |
10046 | struct sched_entity *se = &curr->se; | |
10047 | ||
10048 | for_each_sched_entity(se) { | |
10049 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 10050 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 10051 | } |
18bf2805 | 10052 | |
b52da86e | 10053 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 10054 | task_tick_numa(rq, curr); |
3b1baa64 MR |
10055 | |
10056 | update_misfit_status(curr, rq); | |
2802bf3c | 10057 | update_overutilized_status(task_rq(curr)); |
bf0f6f24 IM |
10058 | } |
10059 | ||
10060 | /* | |
cd29fe6f PZ |
10061 | * called on fork with the child task as argument from the parent's context |
10062 | * - child not yet on the tasklist | |
10063 | * - preemption disabled | |
bf0f6f24 | 10064 | */ |
cd29fe6f | 10065 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 10066 | { |
4fc420c9 DN |
10067 | struct cfs_rq *cfs_rq; |
10068 | struct sched_entity *se = &p->se, *curr; | |
cd29fe6f | 10069 | struct rq *rq = this_rq(); |
8a8c69c3 | 10070 | struct rq_flags rf; |
bf0f6f24 | 10071 | |
8a8c69c3 | 10072 | rq_lock(rq, &rf); |
861d034e PZ |
10073 | update_rq_clock(rq); |
10074 | ||
4fc420c9 DN |
10075 | cfs_rq = task_cfs_rq(current); |
10076 | curr = cfs_rq->curr; | |
e210bffd PZ |
10077 | if (curr) { |
10078 | update_curr(cfs_rq); | |
b5d9d734 | 10079 | se->vruntime = curr->vruntime; |
e210bffd | 10080 | } |
aeb73b04 | 10081 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 10082 | |
cd29fe6f | 10083 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 10084 | /* |
edcb60a3 IM |
10085 | * Upon rescheduling, sched_class::put_prev_task() will place |
10086 | * 'current' within the tree based on its new key value. | |
10087 | */ | |
4d78e7b6 | 10088 | swap(curr->vruntime, se->vruntime); |
8875125e | 10089 | resched_curr(rq); |
4d78e7b6 | 10090 | } |
bf0f6f24 | 10091 | |
88ec22d3 | 10092 | se->vruntime -= cfs_rq->min_vruntime; |
8a8c69c3 | 10093 | rq_unlock(rq, &rf); |
bf0f6f24 IM |
10094 | } |
10095 | ||
cb469845 SR |
10096 | /* |
10097 | * Priority of the task has changed. Check to see if we preempt | |
10098 | * the current task. | |
10099 | */ | |
da7a735e PZ |
10100 | static void |
10101 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 10102 | { |
da0c1e65 | 10103 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
10104 | return; |
10105 | ||
cb469845 SR |
10106 | /* |
10107 | * Reschedule if we are currently running on this runqueue and | |
10108 | * our priority decreased, or if we are not currently running on | |
10109 | * this runqueue and our priority is higher than the current's | |
10110 | */ | |
da7a735e | 10111 | if (rq->curr == p) { |
cb469845 | 10112 | if (p->prio > oldprio) |
8875125e | 10113 | resched_curr(rq); |
cb469845 | 10114 | } else |
15afe09b | 10115 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
10116 | } |
10117 | ||
daa59407 | 10118 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
10119 | { |
10120 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
10121 | |
10122 | /* | |
daa59407 BP |
10123 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
10124 | * the dequeue_entity(.flags=0) will already have normalized the | |
10125 | * vruntime. | |
10126 | */ | |
10127 | if (p->on_rq) | |
10128 | return true; | |
10129 | ||
10130 | /* | |
10131 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
10132 | * But there are some cases where it has already been normalized: | |
da7a735e | 10133 | * |
daa59407 BP |
10134 | * - A forked child which is waiting for being woken up by |
10135 | * wake_up_new_task(). | |
10136 | * - A task which has been woken up by try_to_wake_up() and | |
10137 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 10138 | */ |
d0cdb3ce SM |
10139 | if (!se->sum_exec_runtime || |
10140 | (p->state == TASK_WAKING && p->sched_remote_wakeup)) | |
daa59407 BP |
10141 | return true; |
10142 | ||
10143 | return false; | |
10144 | } | |
10145 | ||
09a43ace VG |
10146 | #ifdef CONFIG_FAIR_GROUP_SCHED |
10147 | /* | |
10148 | * Propagate the changes of the sched_entity across the tg tree to make it | |
10149 | * visible to the root | |
10150 | */ | |
10151 | static void propagate_entity_cfs_rq(struct sched_entity *se) | |
10152 | { | |
10153 | struct cfs_rq *cfs_rq; | |
10154 | ||
10155 | /* Start to propagate at parent */ | |
10156 | se = se->parent; | |
10157 | ||
10158 | for_each_sched_entity(se) { | |
10159 | cfs_rq = cfs_rq_of(se); | |
10160 | ||
10161 | if (cfs_rq_throttled(cfs_rq)) | |
10162 | break; | |
10163 | ||
88c0616e | 10164 | update_load_avg(cfs_rq, se, UPDATE_TG); |
09a43ace VG |
10165 | } |
10166 | } | |
10167 | #else | |
10168 | static void propagate_entity_cfs_rq(struct sched_entity *se) { } | |
10169 | #endif | |
10170 | ||
df217913 | 10171 | static void detach_entity_cfs_rq(struct sched_entity *se) |
daa59407 | 10172 | { |
daa59407 BP |
10173 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
10174 | ||
9d89c257 | 10175 | /* Catch up with the cfs_rq and remove our load when we leave */ |
88c0616e | 10176 | update_load_avg(cfs_rq, se, 0); |
a05e8c51 | 10177 | detach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 10178 | update_tg_load_avg(cfs_rq, false); |
09a43ace | 10179 | propagate_entity_cfs_rq(se); |
da7a735e PZ |
10180 | } |
10181 | ||
df217913 | 10182 | static void attach_entity_cfs_rq(struct sched_entity *se) |
cb469845 | 10183 | { |
daa59407 | 10184 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a BP |
10185 | |
10186 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
eb7a59b2 M |
10187 | /* |
10188 | * Since the real-depth could have been changed (only FAIR | |
10189 | * class maintain depth value), reset depth properly. | |
10190 | */ | |
10191 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
10192 | #endif | |
7855a35a | 10193 | |
df217913 | 10194 | /* Synchronize entity with its cfs_rq */ |
88c0616e | 10195 | update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); |
ea14b57e | 10196 | attach_entity_load_avg(cfs_rq, se, 0); |
7c3edd2c | 10197 | update_tg_load_avg(cfs_rq, false); |
09a43ace | 10198 | propagate_entity_cfs_rq(se); |
df217913 VG |
10199 | } |
10200 | ||
10201 | static void detach_task_cfs_rq(struct task_struct *p) | |
10202 | { | |
10203 | struct sched_entity *se = &p->se; | |
10204 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10205 | ||
10206 | if (!vruntime_normalized(p)) { | |
10207 | /* | |
10208 | * Fix up our vruntime so that the current sleep doesn't | |
10209 | * cause 'unlimited' sleep bonus. | |
10210 | */ | |
10211 | place_entity(cfs_rq, se, 0); | |
10212 | se->vruntime -= cfs_rq->min_vruntime; | |
10213 | } | |
10214 | ||
10215 | detach_entity_cfs_rq(se); | |
10216 | } | |
10217 | ||
10218 | static void attach_task_cfs_rq(struct task_struct *p) | |
10219 | { | |
10220 | struct sched_entity *se = &p->se; | |
10221 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10222 | ||
10223 | attach_entity_cfs_rq(se); | |
daa59407 BP |
10224 | |
10225 | if (!vruntime_normalized(p)) | |
10226 | se->vruntime += cfs_rq->min_vruntime; | |
10227 | } | |
6efdb105 | 10228 | |
daa59407 BP |
10229 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
10230 | { | |
10231 | detach_task_cfs_rq(p); | |
10232 | } | |
10233 | ||
10234 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
10235 | { | |
10236 | attach_task_cfs_rq(p); | |
7855a35a | 10237 | |
daa59407 | 10238 | if (task_on_rq_queued(p)) { |
7855a35a | 10239 | /* |
daa59407 BP |
10240 | * We were most likely switched from sched_rt, so |
10241 | * kick off the schedule if running, otherwise just see | |
10242 | * if we can still preempt the current task. | |
7855a35a | 10243 | */ |
daa59407 BP |
10244 | if (rq->curr == p) |
10245 | resched_curr(rq); | |
10246 | else | |
10247 | check_preempt_curr(rq, p, 0); | |
7855a35a | 10248 | } |
cb469845 SR |
10249 | } |
10250 | ||
83b699ed SV |
10251 | /* Account for a task changing its policy or group. |
10252 | * | |
10253 | * This routine is mostly called to set cfs_rq->curr field when a task | |
10254 | * migrates between groups/classes. | |
10255 | */ | |
10256 | static void set_curr_task_fair(struct rq *rq) | |
10257 | { | |
10258 | struct sched_entity *se = &rq->curr->se; | |
10259 | ||
ec12cb7f PT |
10260 | for_each_sched_entity(se) { |
10261 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
10262 | ||
10263 | set_next_entity(cfs_rq, se); | |
10264 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
10265 | account_cfs_rq_runtime(cfs_rq, 0); | |
10266 | } | |
83b699ed SV |
10267 | } |
10268 | ||
029632fb PZ |
10269 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
10270 | { | |
bfb06889 | 10271 | cfs_rq->tasks_timeline = RB_ROOT_CACHED; |
029632fb PZ |
10272 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
10273 | #ifndef CONFIG_64BIT | |
10274 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
10275 | #endif | |
141965c7 | 10276 | #ifdef CONFIG_SMP |
2a2f5d4e | 10277 | raw_spin_lock_init(&cfs_rq->removed.lock); |
9ee474f5 | 10278 | #endif |
029632fb PZ |
10279 | } |
10280 | ||
810b3817 | 10281 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b VG |
10282 | static void task_set_group_fair(struct task_struct *p) |
10283 | { | |
10284 | struct sched_entity *se = &p->se; | |
10285 | ||
10286 | set_task_rq(p, task_cpu(p)); | |
10287 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
10288 | } | |
10289 | ||
bc54da21 | 10290 | static void task_move_group_fair(struct task_struct *p) |
810b3817 | 10291 | { |
daa59407 | 10292 | detach_task_cfs_rq(p); |
b2b5ce02 | 10293 | set_task_rq(p, task_cpu(p)); |
6efdb105 BP |
10294 | |
10295 | #ifdef CONFIG_SMP | |
10296 | /* Tell se's cfs_rq has been changed -- migrated */ | |
10297 | p->se.avg.last_update_time = 0; | |
10298 | #endif | |
daa59407 | 10299 | attach_task_cfs_rq(p); |
810b3817 | 10300 | } |
029632fb | 10301 | |
ea86cb4b VG |
10302 | static void task_change_group_fair(struct task_struct *p, int type) |
10303 | { | |
10304 | switch (type) { | |
10305 | case TASK_SET_GROUP: | |
10306 | task_set_group_fair(p); | |
10307 | break; | |
10308 | ||
10309 | case TASK_MOVE_GROUP: | |
10310 | task_move_group_fair(p); | |
10311 | break; | |
10312 | } | |
10313 | } | |
10314 | ||
029632fb PZ |
10315 | void free_fair_sched_group(struct task_group *tg) |
10316 | { | |
10317 | int i; | |
10318 | ||
10319 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
10320 | ||
10321 | for_each_possible_cpu(i) { | |
10322 | if (tg->cfs_rq) | |
10323 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 10324 | if (tg->se) |
029632fb PZ |
10325 | kfree(tg->se[i]); |
10326 | } | |
10327 | ||
10328 | kfree(tg->cfs_rq); | |
10329 | kfree(tg->se); | |
10330 | } | |
10331 | ||
10332 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
10333 | { | |
029632fb | 10334 | struct sched_entity *se; |
b7fa30c9 | 10335 | struct cfs_rq *cfs_rq; |
029632fb PZ |
10336 | int i; |
10337 | ||
6396bb22 | 10338 | tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); |
029632fb PZ |
10339 | if (!tg->cfs_rq) |
10340 | goto err; | |
6396bb22 | 10341 | tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); |
029632fb PZ |
10342 | if (!tg->se) |
10343 | goto err; | |
10344 | ||
10345 | tg->shares = NICE_0_LOAD; | |
10346 | ||
10347 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
10348 | ||
10349 | for_each_possible_cpu(i) { | |
10350 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
10351 | GFP_KERNEL, cpu_to_node(i)); | |
10352 | if (!cfs_rq) | |
10353 | goto err; | |
10354 | ||
10355 | se = kzalloc_node(sizeof(struct sched_entity), | |
10356 | GFP_KERNEL, cpu_to_node(i)); | |
10357 | if (!se) | |
10358 | goto err_free_rq; | |
10359 | ||
10360 | init_cfs_rq(cfs_rq); | |
10361 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 10362 | init_entity_runnable_average(se); |
029632fb PZ |
10363 | } |
10364 | ||
10365 | return 1; | |
10366 | ||
10367 | err_free_rq: | |
10368 | kfree(cfs_rq); | |
10369 | err: | |
10370 | return 0; | |
10371 | } | |
10372 | ||
8663e24d PZ |
10373 | void online_fair_sched_group(struct task_group *tg) |
10374 | { | |
10375 | struct sched_entity *se; | |
10376 | struct rq *rq; | |
10377 | int i; | |
10378 | ||
10379 | for_each_possible_cpu(i) { | |
10380 | rq = cpu_rq(i); | |
10381 | se = tg->se[i]; | |
10382 | ||
10383 | raw_spin_lock_irq(&rq->lock); | |
4126bad6 | 10384 | update_rq_clock(rq); |
d0326691 | 10385 | attach_entity_cfs_rq(se); |
55e16d30 | 10386 | sync_throttle(tg, i); |
8663e24d PZ |
10387 | raw_spin_unlock_irq(&rq->lock); |
10388 | } | |
10389 | } | |
10390 | ||
6fe1f348 | 10391 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 10392 | { |
029632fb | 10393 | unsigned long flags; |
6fe1f348 PZ |
10394 | struct rq *rq; |
10395 | int cpu; | |
029632fb | 10396 | |
6fe1f348 PZ |
10397 | for_each_possible_cpu(cpu) { |
10398 | if (tg->se[cpu]) | |
10399 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 10400 | |
6fe1f348 PZ |
10401 | /* |
10402 | * Only empty task groups can be destroyed; so we can speculatively | |
10403 | * check on_list without danger of it being re-added. | |
10404 | */ | |
10405 | if (!tg->cfs_rq[cpu]->on_list) | |
10406 | continue; | |
10407 | ||
10408 | rq = cpu_rq(cpu); | |
10409 | ||
10410 | raw_spin_lock_irqsave(&rq->lock, flags); | |
10411 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
10412 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
10413 | } | |
029632fb PZ |
10414 | } |
10415 | ||
10416 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
10417 | struct sched_entity *se, int cpu, | |
10418 | struct sched_entity *parent) | |
10419 | { | |
10420 | struct rq *rq = cpu_rq(cpu); | |
10421 | ||
10422 | cfs_rq->tg = tg; | |
10423 | cfs_rq->rq = rq; | |
029632fb PZ |
10424 | init_cfs_rq_runtime(cfs_rq); |
10425 | ||
10426 | tg->cfs_rq[cpu] = cfs_rq; | |
10427 | tg->se[cpu] = se; | |
10428 | ||
10429 | /* se could be NULL for root_task_group */ | |
10430 | if (!se) | |
10431 | return; | |
10432 | ||
fed14d45 | 10433 | if (!parent) { |
029632fb | 10434 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
10435 | se->depth = 0; |
10436 | } else { | |
029632fb | 10437 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
10438 | se->depth = parent->depth + 1; |
10439 | } | |
029632fb PZ |
10440 | |
10441 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
10442 | /* guarantee group entities always have weight */ |
10443 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
10444 | se->parent = parent; |
10445 | } | |
10446 | ||
10447 | static DEFINE_MUTEX(shares_mutex); | |
10448 | ||
10449 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
10450 | { | |
10451 | int i; | |
029632fb PZ |
10452 | |
10453 | /* | |
10454 | * We can't change the weight of the root cgroup. | |
10455 | */ | |
10456 | if (!tg->se[0]) | |
10457 | return -EINVAL; | |
10458 | ||
10459 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
10460 | ||
10461 | mutex_lock(&shares_mutex); | |
10462 | if (tg->shares == shares) | |
10463 | goto done; | |
10464 | ||
10465 | tg->shares = shares; | |
10466 | for_each_possible_cpu(i) { | |
10467 | struct rq *rq = cpu_rq(i); | |
8a8c69c3 PZ |
10468 | struct sched_entity *se = tg->se[i]; |
10469 | struct rq_flags rf; | |
029632fb | 10470 | |
029632fb | 10471 | /* Propagate contribution to hierarchy */ |
8a8c69c3 | 10472 | rq_lock_irqsave(rq, &rf); |
71b1da46 | 10473 | update_rq_clock(rq); |
89ee048f | 10474 | for_each_sched_entity(se) { |
88c0616e | 10475 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
1ea6c46a | 10476 | update_cfs_group(se); |
89ee048f | 10477 | } |
8a8c69c3 | 10478 | rq_unlock_irqrestore(rq, &rf); |
029632fb PZ |
10479 | } |
10480 | ||
10481 | done: | |
10482 | mutex_unlock(&shares_mutex); | |
10483 | return 0; | |
10484 | } | |
10485 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
10486 | ||
10487 | void free_fair_sched_group(struct task_group *tg) { } | |
10488 | ||
10489 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
10490 | { | |
10491 | return 1; | |
10492 | } | |
10493 | ||
8663e24d PZ |
10494 | void online_fair_sched_group(struct task_group *tg) { } |
10495 | ||
6fe1f348 | 10496 | void unregister_fair_sched_group(struct task_group *tg) { } |
029632fb PZ |
10497 | |
10498 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
10499 | ||
810b3817 | 10500 | |
6d686f45 | 10501 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
10502 | { |
10503 | struct sched_entity *se = &task->se; | |
0d721cea PW |
10504 | unsigned int rr_interval = 0; |
10505 | ||
10506 | /* | |
10507 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
10508 | * idle runqueue: | |
10509 | */ | |
0d721cea | 10510 | if (rq->cfs.load.weight) |
a59f4e07 | 10511 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
10512 | |
10513 | return rr_interval; | |
10514 | } | |
10515 | ||
bf0f6f24 IM |
10516 | /* |
10517 | * All the scheduling class methods: | |
10518 | */ | |
029632fb | 10519 | const struct sched_class fair_sched_class = { |
5522d5d5 | 10520 | .next = &idle_sched_class, |
bf0f6f24 IM |
10521 | .enqueue_task = enqueue_task_fair, |
10522 | .dequeue_task = dequeue_task_fair, | |
10523 | .yield_task = yield_task_fair, | |
d95f4122 | 10524 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 10525 | |
2e09bf55 | 10526 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
10527 | |
10528 | .pick_next_task = pick_next_task_fair, | |
10529 | .put_prev_task = put_prev_task_fair, | |
10530 | ||
681f3e68 | 10531 | #ifdef CONFIG_SMP |
4ce72a2c | 10532 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 10533 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 10534 | |
0bcdcf28 CE |
10535 | .rq_online = rq_online_fair, |
10536 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 10537 | |
12695578 | 10538 | .task_dead = task_dead_fair, |
c5b28038 | 10539 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 10540 | #endif |
bf0f6f24 | 10541 | |
83b699ed | 10542 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 10543 | .task_tick = task_tick_fair, |
cd29fe6f | 10544 | .task_fork = task_fork_fair, |
cb469845 SR |
10545 | |
10546 | .prio_changed = prio_changed_fair, | |
da7a735e | 10547 | .switched_from = switched_from_fair, |
cb469845 | 10548 | .switched_to = switched_to_fair, |
810b3817 | 10549 | |
0d721cea PW |
10550 | .get_rr_interval = get_rr_interval_fair, |
10551 | ||
6e998916 SG |
10552 | .update_curr = update_curr_fair, |
10553 | ||
810b3817 | 10554 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 10555 | .task_change_group = task_change_group_fair, |
810b3817 | 10556 | #endif |
bf0f6f24 IM |
10557 | }; |
10558 | ||
10559 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 10560 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 10561 | { |
c40f7d74 | 10562 | struct cfs_rq *cfs_rq; |
bf0f6f24 | 10563 | |
5973e5b9 | 10564 | rcu_read_lock(); |
c40f7d74 | 10565 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 10566 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 10567 | rcu_read_unlock(); |
bf0f6f24 | 10568 | } |
397f2378 SD |
10569 | |
10570 | #ifdef CONFIG_NUMA_BALANCING | |
10571 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
10572 | { | |
10573 | int node; | |
10574 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
10575 | ||
10576 | for_each_online_node(node) { | |
10577 | if (p->numa_faults) { | |
10578 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
10579 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
10580 | } | |
10581 | if (p->numa_group) { | |
10582 | gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
10583 | gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
10584 | } | |
10585 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
10586 | } | |
10587 | } | |
10588 | #endif /* CONFIG_NUMA_BALANCING */ | |
10589 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
10590 | |
10591 | __init void init_sched_fair_class(void) | |
10592 | { | |
10593 | #ifdef CONFIG_SMP | |
10594 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
10595 | ||
3451d024 | 10596 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 10597 | nohz.next_balance = jiffies; |
f643ea22 | 10598 | nohz.next_blocked = jiffies; |
029632fb | 10599 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
10600 | #endif |
10601 | #endif /* SMP */ | |
10602 | ||
10603 | } |