]>
Commit | Line | Data |
---|---|---|
bf0f6f24 IM |
1 | /* |
2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) | |
3 | * | |
4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> | |
5 | * | |
6 | * Interactivity improvements by Mike Galbraith | |
7 | * (C) 2007 Mike Galbraith <efault@gmx.de> | |
8 | * | |
9 | * Various enhancements by Dmitry Adamushko. | |
10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> | |
11 | * | |
12 | * Group scheduling enhancements by Srivatsa Vaddagiri | |
13 | * Copyright IBM Corporation, 2007 | |
14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> | |
15 | * | |
16 | * Scaled math optimizations by Thomas Gleixner | |
17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> | |
21805085 PZ |
18 | * |
19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra | |
90eec103 | 20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra |
bf0f6f24 IM |
21 | */ |
22 | ||
589ee628 | 23 | #include <linux/sched/mm.h> |
105ab3d8 IM |
24 | #include <linux/sched/topology.h> |
25 | ||
cb251765 | 26 | #include <linux/latencytop.h> |
3436ae12 | 27 | #include <linux/cpumask.h> |
83a0a96a | 28 | #include <linux/cpuidle.h> |
029632fb PZ |
29 | #include <linux/slab.h> |
30 | #include <linux/profile.h> | |
31 | #include <linux/interrupt.h> | |
cbee9f88 | 32 | #include <linux/mempolicy.h> |
e14808b4 | 33 | #include <linux/migrate.h> |
cbee9f88 | 34 | #include <linux/task_work.h> |
029632fb PZ |
35 | |
36 | #include <trace/events/sched.h> | |
37 | ||
38 | #include "sched.h" | |
9745512c | 39 | |
bf0f6f24 | 40 | /* |
21805085 | 41 | * Targeted preemption latency for CPU-bound tasks: |
bf0f6f24 | 42 | * |
21805085 | 43 | * NOTE: this latency value is not the same as the concept of |
d274a4ce IM |
44 | * 'timeslice length' - timeslices in CFS are of variable length |
45 | * and have no persistent notion like in traditional, time-slice | |
46 | * based scheduling concepts. | |
bf0f6f24 | 47 | * |
d274a4ce IM |
48 | * (to see the precise effective timeslice length of your workload, |
49 | * run vmstat and monitor the context-switches (cs) field) | |
2b4d5b25 IM |
50 | * |
51 | * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 52 | */ |
2b4d5b25 IM |
53 | unsigned int sysctl_sched_latency = 6000000ULL; |
54 | unsigned int normalized_sysctl_sched_latency = 6000000ULL; | |
2bd8e6d4 | 55 | |
1983a922 CE |
56 | /* |
57 | * The initial- and re-scaling of tunables is configurable | |
1983a922 CE |
58 | * |
59 | * Options are: | |
2b4d5b25 IM |
60 | * |
61 | * SCHED_TUNABLESCALING_NONE - unscaled, always *1 | |
62 | * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) | |
63 | * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus | |
64 | * | |
65 | * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) | |
1983a922 | 66 | */ |
2b4d5b25 | 67 | enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG; |
1983a922 | 68 | |
2bd8e6d4 | 69 | /* |
b2be5e96 | 70 | * Minimal preemption granularity for CPU-bound tasks: |
2b4d5b25 | 71 | * |
864616ee | 72 | * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) |
2bd8e6d4 | 73 | */ |
2b4d5b25 IM |
74 | unsigned int sysctl_sched_min_granularity = 750000ULL; |
75 | unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; | |
21805085 PZ |
76 | |
77 | /* | |
2b4d5b25 | 78 | * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity |
b2be5e96 | 79 | */ |
0bf377bb | 80 | static unsigned int sched_nr_latency = 8; |
b2be5e96 PZ |
81 | |
82 | /* | |
2bba22c5 | 83 | * After fork, child runs first. If set to 0 (default) then |
b2be5e96 | 84 | * parent will (try to) run first. |
21805085 | 85 | */ |
2bba22c5 | 86 | unsigned int sysctl_sched_child_runs_first __read_mostly; |
bf0f6f24 | 87 | |
bf0f6f24 IM |
88 | /* |
89 | * SCHED_OTHER wake-up granularity. | |
bf0f6f24 IM |
90 | * |
91 | * This option delays the preemption effects of decoupled workloads | |
92 | * and reduces their over-scheduling. Synchronous workloads will still | |
93 | * have immediate wakeup/sleep latencies. | |
2b4d5b25 IM |
94 | * |
95 | * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) | |
bf0f6f24 | 96 | */ |
2b4d5b25 IM |
97 | unsigned int sysctl_sched_wakeup_granularity = 1000000UL; |
98 | unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; | |
bf0f6f24 | 99 | |
2b4d5b25 | 100 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
da84d961 | 101 | |
afe06efd TC |
102 | #ifdef CONFIG_SMP |
103 | /* | |
104 | * For asym packing, by default the lower numbered cpu has higher priority. | |
105 | */ | |
106 | int __weak arch_asym_cpu_priority(int cpu) | |
107 | { | |
108 | return -cpu; | |
109 | } | |
110 | #endif | |
111 | ||
ec12cb7f PT |
112 | #ifdef CONFIG_CFS_BANDWIDTH |
113 | /* | |
114 | * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool | |
115 | * each time a cfs_rq requests quota. | |
116 | * | |
117 | * Note: in the case that the slice exceeds the runtime remaining (either due | |
118 | * to consumption or the quota being specified to be smaller than the slice) | |
119 | * we will always only issue the remaining available time. | |
120 | * | |
2b4d5b25 IM |
121 | * (default: 5 msec, units: microseconds) |
122 | */ | |
123 | unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; | |
ec12cb7f PT |
124 | #endif |
125 | ||
3273163c MR |
126 | /* |
127 | * The margin used when comparing utilization with CPU capacity: | |
893c5d22 | 128 | * util * margin < capacity * 1024 |
2b4d5b25 IM |
129 | * |
130 | * (default: ~20%) | |
3273163c | 131 | */ |
2b4d5b25 | 132 | unsigned int capacity_margin = 1280; |
3273163c | 133 | |
8527632d PG |
134 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
135 | { | |
136 | lw->weight += inc; | |
137 | lw->inv_weight = 0; | |
138 | } | |
139 | ||
140 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) | |
141 | { | |
142 | lw->weight -= dec; | |
143 | lw->inv_weight = 0; | |
144 | } | |
145 | ||
146 | static inline void update_load_set(struct load_weight *lw, unsigned long w) | |
147 | { | |
148 | lw->weight = w; | |
149 | lw->inv_weight = 0; | |
150 | } | |
151 | ||
029632fb PZ |
152 | /* |
153 | * Increase the granularity value when there are more CPUs, | |
154 | * because with more CPUs the 'effective latency' as visible | |
155 | * to users decreases. But the relationship is not linear, | |
156 | * so pick a second-best guess by going with the log2 of the | |
157 | * number of CPUs. | |
158 | * | |
159 | * This idea comes from the SD scheduler of Con Kolivas: | |
160 | */ | |
58ac93e4 | 161 | static unsigned int get_update_sysctl_factor(void) |
029632fb | 162 | { |
58ac93e4 | 163 | unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); |
029632fb PZ |
164 | unsigned int factor; |
165 | ||
166 | switch (sysctl_sched_tunable_scaling) { | |
167 | case SCHED_TUNABLESCALING_NONE: | |
168 | factor = 1; | |
169 | break; | |
170 | case SCHED_TUNABLESCALING_LINEAR: | |
171 | factor = cpus; | |
172 | break; | |
173 | case SCHED_TUNABLESCALING_LOG: | |
174 | default: | |
175 | factor = 1 + ilog2(cpus); | |
176 | break; | |
177 | } | |
178 | ||
179 | return factor; | |
180 | } | |
181 | ||
182 | static void update_sysctl(void) | |
183 | { | |
184 | unsigned int factor = get_update_sysctl_factor(); | |
185 | ||
186 | #define SET_SYSCTL(name) \ | |
187 | (sysctl_##name = (factor) * normalized_sysctl_##name) | |
188 | SET_SYSCTL(sched_min_granularity); | |
189 | SET_SYSCTL(sched_latency); | |
190 | SET_SYSCTL(sched_wakeup_granularity); | |
191 | #undef SET_SYSCTL | |
192 | } | |
193 | ||
194 | void sched_init_granularity(void) | |
195 | { | |
196 | update_sysctl(); | |
197 | } | |
198 | ||
9dbdb155 | 199 | #define WMULT_CONST (~0U) |
029632fb PZ |
200 | #define WMULT_SHIFT 32 |
201 | ||
9dbdb155 PZ |
202 | static void __update_inv_weight(struct load_weight *lw) |
203 | { | |
204 | unsigned long w; | |
205 | ||
206 | if (likely(lw->inv_weight)) | |
207 | return; | |
208 | ||
209 | w = scale_load_down(lw->weight); | |
210 | ||
211 | if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) | |
212 | lw->inv_weight = 1; | |
213 | else if (unlikely(!w)) | |
214 | lw->inv_weight = WMULT_CONST; | |
215 | else | |
216 | lw->inv_weight = WMULT_CONST / w; | |
217 | } | |
029632fb PZ |
218 | |
219 | /* | |
9dbdb155 PZ |
220 | * delta_exec * weight / lw.weight |
221 | * OR | |
222 | * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT | |
223 | * | |
1c3de5e1 | 224 | * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case |
9dbdb155 PZ |
225 | * we're guaranteed shift stays positive because inv_weight is guaranteed to |
226 | * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. | |
227 | * | |
228 | * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus | |
229 | * weight/lw.weight <= 1, and therefore our shift will also be positive. | |
029632fb | 230 | */ |
9dbdb155 | 231 | static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) |
029632fb | 232 | { |
9dbdb155 PZ |
233 | u64 fact = scale_load_down(weight); |
234 | int shift = WMULT_SHIFT; | |
029632fb | 235 | |
9dbdb155 | 236 | __update_inv_weight(lw); |
029632fb | 237 | |
9dbdb155 PZ |
238 | if (unlikely(fact >> 32)) { |
239 | while (fact >> 32) { | |
240 | fact >>= 1; | |
241 | shift--; | |
242 | } | |
029632fb PZ |
243 | } |
244 | ||
9dbdb155 PZ |
245 | /* hint to use a 32x32->64 mul */ |
246 | fact = (u64)(u32)fact * lw->inv_weight; | |
029632fb | 247 | |
9dbdb155 PZ |
248 | while (fact >> 32) { |
249 | fact >>= 1; | |
250 | shift--; | |
251 | } | |
029632fb | 252 | |
9dbdb155 | 253 | return mul_u64_u32_shr(delta_exec, fact, shift); |
029632fb PZ |
254 | } |
255 | ||
256 | ||
257 | const struct sched_class fair_sched_class; | |
a4c2f00f | 258 | |
bf0f6f24 IM |
259 | /************************************************************** |
260 | * CFS operations on generic schedulable entities: | |
261 | */ | |
262 | ||
62160e3f | 263 | #ifdef CONFIG_FAIR_GROUP_SCHED |
bf0f6f24 | 264 | |
62160e3f | 265 | /* cpu runqueue to which this cfs_rq is attached */ |
bf0f6f24 IM |
266 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
267 | { | |
62160e3f | 268 | return cfs_rq->rq; |
bf0f6f24 IM |
269 | } |
270 | ||
62160e3f IM |
271 | /* An entity is a task if it doesn't "own" a runqueue */ |
272 | #define entity_is_task(se) (!se->my_q) | |
bf0f6f24 | 273 | |
8f48894f PZ |
274 | static inline struct task_struct *task_of(struct sched_entity *se) |
275 | { | |
9148a3a1 | 276 | SCHED_WARN_ON(!entity_is_task(se)); |
8f48894f PZ |
277 | return container_of(se, struct task_struct, se); |
278 | } | |
279 | ||
b758149c PZ |
280 | /* Walk up scheduling entities hierarchy */ |
281 | #define for_each_sched_entity(se) \ | |
282 | for (; se; se = se->parent) | |
283 | ||
284 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) | |
285 | { | |
286 | return p->se.cfs_rq; | |
287 | } | |
288 | ||
289 | /* runqueue on which this entity is (to be) queued */ | |
290 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) | |
291 | { | |
292 | return se->cfs_rq; | |
293 | } | |
294 | ||
295 | /* runqueue "owned" by this group */ | |
296 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
297 | { | |
298 | return grp->my_q; | |
299 | } | |
300 | ||
3d4b47b4 PZ |
301 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
302 | { | |
303 | if (!cfs_rq->on_list) { | |
9c2791f9 VG |
304 | struct rq *rq = rq_of(cfs_rq); |
305 | int cpu = cpu_of(rq); | |
67e86250 PT |
306 | /* |
307 | * Ensure we either appear before our parent (if already | |
308 | * enqueued) or force our parent to appear after us when it is | |
9c2791f9 VG |
309 | * enqueued. The fact that we always enqueue bottom-up |
310 | * reduces this to two cases and a special case for the root | |
311 | * cfs_rq. Furthermore, it also means that we will always reset | |
312 | * tmp_alone_branch either when the branch is connected | |
313 | * to a tree or when we reach the beg of the tree | |
67e86250 PT |
314 | */ |
315 | if (cfs_rq->tg->parent && | |
9c2791f9 VG |
316 | cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { |
317 | /* | |
318 | * If parent is already on the list, we add the child | |
319 | * just before. Thanks to circular linked property of | |
320 | * the list, this means to put the child at the tail | |
321 | * of the list that starts by parent. | |
322 | */ | |
323 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, | |
324 | &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); | |
325 | /* | |
326 | * The branch is now connected to its tree so we can | |
327 | * reset tmp_alone_branch to the beginning of the | |
328 | * list. | |
329 | */ | |
330 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
331 | } else if (!cfs_rq->tg->parent) { | |
332 | /* | |
333 | * cfs rq without parent should be put | |
334 | * at the tail of the list. | |
335 | */ | |
67e86250 | 336 | list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, |
9c2791f9 VG |
337 | &rq->leaf_cfs_rq_list); |
338 | /* | |
339 | * We have reach the beg of a tree so we can reset | |
340 | * tmp_alone_branch to the beginning of the list. | |
341 | */ | |
342 | rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; | |
343 | } else { | |
344 | /* | |
345 | * The parent has not already been added so we want to | |
346 | * make sure that it will be put after us. | |
347 | * tmp_alone_branch points to the beg of the branch | |
348 | * where we will add parent. | |
349 | */ | |
350 | list_add_rcu(&cfs_rq->leaf_cfs_rq_list, | |
351 | rq->tmp_alone_branch); | |
352 | /* | |
353 | * update tmp_alone_branch to points to the new beg | |
354 | * of the branch | |
355 | */ | |
356 | rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; | |
67e86250 | 357 | } |
3d4b47b4 PZ |
358 | |
359 | cfs_rq->on_list = 1; | |
360 | } | |
361 | } | |
362 | ||
363 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
364 | { | |
365 | if (cfs_rq->on_list) { | |
366 | list_del_rcu(&cfs_rq->leaf_cfs_rq_list); | |
367 | cfs_rq->on_list = 0; | |
368 | } | |
369 | } | |
370 | ||
b758149c | 371 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
a9e7f654 TH |
372 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ |
373 | list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \ | |
374 | leaf_cfs_rq_list) | |
b758149c PZ |
375 | |
376 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 377 | static inline struct cfs_rq * |
b758149c PZ |
378 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
379 | { | |
380 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 381 | return se->cfs_rq; |
b758149c | 382 | |
fed14d45 | 383 | return NULL; |
b758149c PZ |
384 | } |
385 | ||
386 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
387 | { | |
388 | return se->parent; | |
389 | } | |
390 | ||
464b7527 PZ |
391 | static void |
392 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
393 | { | |
394 | int se_depth, pse_depth; | |
395 | ||
396 | /* | |
397 | * preemption test can be made between sibling entities who are in the | |
398 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
399 | * both tasks until we find their ancestors who are siblings of common | |
400 | * parent. | |
401 | */ | |
402 | ||
403 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
404 | se_depth = (*se)->depth; |
405 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
406 | |
407 | while (se_depth > pse_depth) { | |
408 | se_depth--; | |
409 | *se = parent_entity(*se); | |
410 | } | |
411 | ||
412 | while (pse_depth > se_depth) { | |
413 | pse_depth--; | |
414 | *pse = parent_entity(*pse); | |
415 | } | |
416 | ||
417 | while (!is_same_group(*se, *pse)) { | |
418 | *se = parent_entity(*se); | |
419 | *pse = parent_entity(*pse); | |
420 | } | |
421 | } | |
422 | ||
8f48894f PZ |
423 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
424 | ||
425 | static inline struct task_struct *task_of(struct sched_entity *se) | |
426 | { | |
427 | return container_of(se, struct task_struct, se); | |
428 | } | |
bf0f6f24 | 429 | |
62160e3f IM |
430 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
431 | { | |
432 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
433 | } |
434 | ||
435 | #define entity_is_task(se) 1 | |
436 | ||
b758149c PZ |
437 | #define for_each_sched_entity(se) \ |
438 | for (; se; se = NULL) | |
bf0f6f24 | 439 | |
b758149c | 440 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 441 | { |
b758149c | 442 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
443 | } |
444 | ||
b758149c PZ |
445 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
446 | { | |
447 | struct task_struct *p = task_of(se); | |
448 | struct rq *rq = task_rq(p); | |
449 | ||
450 | return &rq->cfs; | |
451 | } | |
452 | ||
453 | /* runqueue "owned" by this group */ | |
454 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
455 | { | |
456 | return NULL; | |
457 | } | |
458 | ||
3d4b47b4 PZ |
459 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
460 | { | |
461 | } | |
462 | ||
463 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
464 | { | |
465 | } | |
466 | ||
a9e7f654 TH |
467 | #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ |
468 | for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos) | |
b758149c | 469 | |
b758149c PZ |
470 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
471 | { | |
472 | return NULL; | |
473 | } | |
474 | ||
464b7527 PZ |
475 | static inline void |
476 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
477 | { | |
478 | } | |
479 | ||
b758149c PZ |
480 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
481 | ||
6c16a6dc | 482 | static __always_inline |
9dbdb155 | 483 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
484 | |
485 | /************************************************************** | |
486 | * Scheduling class tree data structure manipulation methods: | |
487 | */ | |
488 | ||
1bf08230 | 489 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 490 | { |
1bf08230 | 491 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 492 | if (delta > 0) |
1bf08230 | 493 | max_vruntime = vruntime; |
02e0431a | 494 | |
1bf08230 | 495 | return max_vruntime; |
02e0431a PZ |
496 | } |
497 | ||
0702e3eb | 498 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
499 | { |
500 | s64 delta = (s64)(vruntime - min_vruntime); | |
501 | if (delta < 0) | |
502 | min_vruntime = vruntime; | |
503 | ||
504 | return min_vruntime; | |
505 | } | |
506 | ||
54fdc581 FC |
507 | static inline int entity_before(struct sched_entity *a, |
508 | struct sched_entity *b) | |
509 | { | |
510 | return (s64)(a->vruntime - b->vruntime) < 0; | |
511 | } | |
512 | ||
1af5f730 PZ |
513 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
514 | { | |
b60205c7 | 515 | struct sched_entity *curr = cfs_rq->curr; |
bfb06889 | 516 | struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); |
b60205c7 | 517 | |
1af5f730 PZ |
518 | u64 vruntime = cfs_rq->min_vruntime; |
519 | ||
b60205c7 PZ |
520 | if (curr) { |
521 | if (curr->on_rq) | |
522 | vruntime = curr->vruntime; | |
523 | else | |
524 | curr = NULL; | |
525 | } | |
1af5f730 | 526 | |
bfb06889 DB |
527 | if (leftmost) { /* non-empty tree */ |
528 | struct sched_entity *se; | |
529 | se = rb_entry(leftmost, struct sched_entity, run_node); | |
1af5f730 | 530 | |
b60205c7 | 531 | if (!curr) |
1af5f730 PZ |
532 | vruntime = se->vruntime; |
533 | else | |
534 | vruntime = min_vruntime(vruntime, se->vruntime); | |
535 | } | |
536 | ||
1bf08230 | 537 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 538 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
539 | #ifndef CONFIG_64BIT |
540 | smp_wmb(); | |
541 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
542 | #endif | |
1af5f730 PZ |
543 | } |
544 | ||
bf0f6f24 IM |
545 | /* |
546 | * Enqueue an entity into the rb-tree: | |
547 | */ | |
0702e3eb | 548 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 549 | { |
bfb06889 | 550 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node; |
bf0f6f24 IM |
551 | struct rb_node *parent = NULL; |
552 | struct sched_entity *entry; | |
bfb06889 | 553 | bool leftmost = true; |
bf0f6f24 IM |
554 | |
555 | /* | |
556 | * Find the right place in the rbtree: | |
557 | */ | |
558 | while (*link) { | |
559 | parent = *link; | |
560 | entry = rb_entry(parent, struct sched_entity, run_node); | |
561 | /* | |
562 | * We dont care about collisions. Nodes with | |
563 | * the same key stay together. | |
564 | */ | |
2bd2d6f2 | 565 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
566 | link = &parent->rb_left; |
567 | } else { | |
568 | link = &parent->rb_right; | |
bfb06889 | 569 | leftmost = false; |
bf0f6f24 IM |
570 | } |
571 | } | |
572 | ||
bf0f6f24 | 573 | rb_link_node(&se->run_node, parent, link); |
bfb06889 DB |
574 | rb_insert_color_cached(&se->run_node, |
575 | &cfs_rq->tasks_timeline, leftmost); | |
bf0f6f24 IM |
576 | } |
577 | ||
0702e3eb | 578 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 579 | { |
bfb06889 | 580 | rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
581 | } |
582 | ||
029632fb | 583 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 584 | { |
bfb06889 | 585 | struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); |
f4b6755f PZ |
586 | |
587 | if (!left) | |
588 | return NULL; | |
589 | ||
590 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
591 | } |
592 | ||
ac53db59 RR |
593 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
594 | { | |
595 | struct rb_node *next = rb_next(&se->run_node); | |
596 | ||
597 | if (!next) | |
598 | return NULL; | |
599 | ||
600 | return rb_entry(next, struct sched_entity, run_node); | |
601 | } | |
602 | ||
603 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 604 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 605 | { |
bfb06889 | 606 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); |
aeb73b04 | 607 | |
70eee74b BS |
608 | if (!last) |
609 | return NULL; | |
7eee3e67 IM |
610 | |
611 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
612 | } |
613 | ||
bf0f6f24 IM |
614 | /************************************************************** |
615 | * Scheduling class statistics methods: | |
616 | */ | |
617 | ||
acb4a848 | 618 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 619 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
620 | loff_t *ppos) |
621 | { | |
8d65af78 | 622 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
58ac93e4 | 623 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
624 | |
625 | if (ret || !write) | |
626 | return ret; | |
627 | ||
628 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
629 | sysctl_sched_min_granularity); | |
630 | ||
acb4a848 CE |
631 | #define WRT_SYSCTL(name) \ |
632 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
633 | WRT_SYSCTL(sched_min_granularity); | |
634 | WRT_SYSCTL(sched_latency); | |
635 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
636 | #undef WRT_SYSCTL |
637 | ||
b2be5e96 PZ |
638 | return 0; |
639 | } | |
640 | #endif | |
647e7cac | 641 | |
a7be37ac | 642 | /* |
f9c0b095 | 643 | * delta /= w |
a7be37ac | 644 | */ |
9dbdb155 | 645 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 646 | { |
f9c0b095 | 647 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 648 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
649 | |
650 | return delta; | |
651 | } | |
652 | ||
647e7cac IM |
653 | /* |
654 | * The idea is to set a period in which each task runs once. | |
655 | * | |
532b1858 | 656 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
657 | * this period because otherwise the slices get too small. |
658 | * | |
659 | * p = (nr <= nl) ? l : l*nr/nl | |
660 | */ | |
4d78e7b6 PZ |
661 | static u64 __sched_period(unsigned long nr_running) |
662 | { | |
8e2b0bf3 BF |
663 | if (unlikely(nr_running > sched_nr_latency)) |
664 | return nr_running * sysctl_sched_min_granularity; | |
665 | else | |
666 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
667 | } |
668 | ||
647e7cac IM |
669 | /* |
670 | * We calculate the wall-time slice from the period by taking a part | |
671 | * proportional to the weight. | |
672 | * | |
f9c0b095 | 673 | * s = p*P[w/rw] |
647e7cac | 674 | */ |
6d0f0ebd | 675 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 676 | { |
0a582440 | 677 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 678 | |
0a582440 | 679 | for_each_sched_entity(se) { |
6272d68c | 680 | struct load_weight *load; |
3104bf03 | 681 | struct load_weight lw; |
6272d68c LM |
682 | |
683 | cfs_rq = cfs_rq_of(se); | |
684 | load = &cfs_rq->load; | |
f9c0b095 | 685 | |
0a582440 | 686 | if (unlikely(!se->on_rq)) { |
3104bf03 | 687 | lw = cfs_rq->load; |
0a582440 MG |
688 | |
689 | update_load_add(&lw, se->load.weight); | |
690 | load = &lw; | |
691 | } | |
9dbdb155 | 692 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
693 | } |
694 | return slice; | |
bf0f6f24 IM |
695 | } |
696 | ||
647e7cac | 697 | /* |
660cc00f | 698 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 699 | * |
f9c0b095 | 700 | * vs = s/w |
647e7cac | 701 | */ |
f9c0b095 | 702 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 703 | { |
f9c0b095 | 704 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
705 | } |
706 | ||
a75cdaa9 | 707 | #ifdef CONFIG_SMP |
283e2ed3 PZ |
708 | |
709 | #include "sched-pelt.h" | |
710 | ||
772bd008 | 711 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee MG |
712 | static unsigned long task_h_load(struct task_struct *p); |
713 | ||
540247fb YD |
714 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
715 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 716 | { |
540247fb | 717 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 718 | |
f207934f PZ |
719 | memset(sa, 0, sizeof(*sa)); |
720 | ||
b5a9b340 VG |
721 | /* |
722 | * Tasks are intialized with full load to be seen as heavy tasks until | |
723 | * they get a chance to stabilize to their real load level. | |
724 | * Group entities are intialized with zero load to reflect the fact that | |
725 | * nothing has been attached to the task group yet. | |
726 | */ | |
727 | if (entity_is_task(se)) | |
1ea6c46a | 728 | sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight); |
1ea6c46a | 729 | |
f207934f PZ |
730 | se->runnable_weight = se->load.weight; |
731 | ||
9d89c257 | 732 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 733 | } |
7ea241af | 734 | |
7dc603c9 | 735 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
df217913 | 736 | static void attach_entity_cfs_rq(struct sched_entity *se); |
7dc603c9 | 737 | |
2b8c41da YD |
738 | /* |
739 | * With new tasks being created, their initial util_avgs are extrapolated | |
740 | * based on the cfs_rq's current util_avg: | |
741 | * | |
742 | * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight | |
743 | * | |
744 | * However, in many cases, the above util_avg does not give a desired | |
745 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
746 | * as when the series is a harmonic series. | |
747 | * | |
748 | * To solve this problem, we also cap the util_avg of successive tasks to | |
749 | * only 1/2 of the left utilization budget: | |
750 | * | |
751 | * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n | |
752 | * | |
753 | * where n denotes the nth task. | |
754 | * | |
755 | * For example, a simplest series from the beginning would be like: | |
756 | * | |
757 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
758 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
759 | * | |
760 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
761 | * if util_avg > util_avg_cap. | |
762 | */ | |
763 | void post_init_entity_util_avg(struct sched_entity *se) | |
764 | { | |
765 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
766 | struct sched_avg *sa = &se->avg; | |
172895e6 | 767 | long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2; |
2b8c41da YD |
768 | |
769 | if (cap > 0) { | |
770 | if (cfs_rq->avg.util_avg != 0) { | |
771 | sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; | |
772 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); | |
773 | ||
774 | if (sa->util_avg > cap) | |
775 | sa->util_avg = cap; | |
776 | } else { | |
777 | sa->util_avg = cap; | |
778 | } | |
2b8c41da | 779 | } |
7dc603c9 PZ |
780 | |
781 | if (entity_is_task(se)) { | |
782 | struct task_struct *p = task_of(se); | |
783 | if (p->sched_class != &fair_sched_class) { | |
784 | /* | |
785 | * For !fair tasks do: | |
786 | * | |
3a123bbb | 787 | update_cfs_rq_load_avg(now, cfs_rq); |
7dc603c9 PZ |
788 | attach_entity_load_avg(cfs_rq, se); |
789 | switched_from_fair(rq, p); | |
790 | * | |
791 | * such that the next switched_to_fair() has the | |
792 | * expected state. | |
793 | */ | |
df217913 | 794 | se->avg.last_update_time = cfs_rq_clock_task(cfs_rq); |
7dc603c9 PZ |
795 | return; |
796 | } | |
797 | } | |
798 | ||
df217913 | 799 | attach_entity_cfs_rq(se); |
2b8c41da YD |
800 | } |
801 | ||
7dc603c9 | 802 | #else /* !CONFIG_SMP */ |
540247fb | 803 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
804 | { |
805 | } | |
2b8c41da YD |
806 | void post_init_entity_util_avg(struct sched_entity *se) |
807 | { | |
808 | } | |
3d30544f PZ |
809 | static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
810 | { | |
811 | } | |
7dc603c9 | 812 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 813 | |
bf0f6f24 | 814 | /* |
9dbdb155 | 815 | * Update the current task's runtime statistics. |
bf0f6f24 | 816 | */ |
b7cc0896 | 817 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 818 | { |
429d43bc | 819 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 820 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 821 | u64 delta_exec; |
bf0f6f24 IM |
822 | |
823 | if (unlikely(!curr)) | |
824 | return; | |
825 | ||
9dbdb155 PZ |
826 | delta_exec = now - curr->exec_start; |
827 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 828 | return; |
bf0f6f24 | 829 | |
8ebc91d9 | 830 | curr->exec_start = now; |
d842de87 | 831 | |
9dbdb155 PZ |
832 | schedstat_set(curr->statistics.exec_max, |
833 | max(delta_exec, curr->statistics.exec_max)); | |
834 | ||
835 | curr->sum_exec_runtime += delta_exec; | |
ae92882e | 836 | schedstat_add(cfs_rq->exec_clock, delta_exec); |
9dbdb155 PZ |
837 | |
838 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
839 | update_min_vruntime(cfs_rq); | |
840 | ||
d842de87 SV |
841 | if (entity_is_task(curr)) { |
842 | struct task_struct *curtask = task_of(curr); | |
843 | ||
f977bb49 | 844 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 845 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 846 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 847 | } |
ec12cb7f PT |
848 | |
849 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
850 | } |
851 | ||
6e998916 SG |
852 | static void update_curr_fair(struct rq *rq) |
853 | { | |
854 | update_curr(cfs_rq_of(&rq->curr->se)); | |
855 | } | |
856 | ||
bf0f6f24 | 857 | static inline void |
5870db5b | 858 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 859 | { |
4fa8d299 JP |
860 | u64 wait_start, prev_wait_start; |
861 | ||
862 | if (!schedstat_enabled()) | |
863 | return; | |
864 | ||
865 | wait_start = rq_clock(rq_of(cfs_rq)); | |
866 | prev_wait_start = schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
867 | |
868 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && | |
4fa8d299 JP |
869 | likely(wait_start > prev_wait_start)) |
870 | wait_start -= prev_wait_start; | |
3ea94de1 | 871 | |
4fa8d299 | 872 | schedstat_set(se->statistics.wait_start, wait_start); |
bf0f6f24 IM |
873 | } |
874 | ||
4fa8d299 | 875 | static inline void |
3ea94de1 JP |
876 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
877 | { | |
878 | struct task_struct *p; | |
cb251765 MG |
879 | u64 delta; |
880 | ||
4fa8d299 JP |
881 | if (!schedstat_enabled()) |
882 | return; | |
883 | ||
884 | delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
885 | |
886 | if (entity_is_task(se)) { | |
887 | p = task_of(se); | |
888 | if (task_on_rq_migrating(p)) { | |
889 | /* | |
890 | * Preserve migrating task's wait time so wait_start | |
891 | * time stamp can be adjusted to accumulate wait time | |
892 | * prior to migration. | |
893 | */ | |
4fa8d299 | 894 | schedstat_set(se->statistics.wait_start, delta); |
3ea94de1 JP |
895 | return; |
896 | } | |
897 | trace_sched_stat_wait(p, delta); | |
898 | } | |
899 | ||
4fa8d299 JP |
900 | schedstat_set(se->statistics.wait_max, |
901 | max(schedstat_val(se->statistics.wait_max), delta)); | |
902 | schedstat_inc(se->statistics.wait_count); | |
903 | schedstat_add(se->statistics.wait_sum, delta); | |
904 | schedstat_set(se->statistics.wait_start, 0); | |
3ea94de1 | 905 | } |
3ea94de1 | 906 | |
4fa8d299 | 907 | static inline void |
1a3d027c JP |
908 | update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
909 | { | |
910 | struct task_struct *tsk = NULL; | |
4fa8d299 JP |
911 | u64 sleep_start, block_start; |
912 | ||
913 | if (!schedstat_enabled()) | |
914 | return; | |
915 | ||
916 | sleep_start = schedstat_val(se->statistics.sleep_start); | |
917 | block_start = schedstat_val(se->statistics.block_start); | |
1a3d027c JP |
918 | |
919 | if (entity_is_task(se)) | |
920 | tsk = task_of(se); | |
921 | ||
4fa8d299 JP |
922 | if (sleep_start) { |
923 | u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; | |
1a3d027c JP |
924 | |
925 | if ((s64)delta < 0) | |
926 | delta = 0; | |
927 | ||
4fa8d299 JP |
928 | if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) |
929 | schedstat_set(se->statistics.sleep_max, delta); | |
1a3d027c | 930 | |
4fa8d299 JP |
931 | schedstat_set(se->statistics.sleep_start, 0); |
932 | schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
933 | |
934 | if (tsk) { | |
935 | account_scheduler_latency(tsk, delta >> 10, 1); | |
936 | trace_sched_stat_sleep(tsk, delta); | |
937 | } | |
938 | } | |
4fa8d299 JP |
939 | if (block_start) { |
940 | u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; | |
1a3d027c JP |
941 | |
942 | if ((s64)delta < 0) | |
943 | delta = 0; | |
944 | ||
4fa8d299 JP |
945 | if (unlikely(delta > schedstat_val(se->statistics.block_max))) |
946 | schedstat_set(se->statistics.block_max, delta); | |
1a3d027c | 947 | |
4fa8d299 JP |
948 | schedstat_set(se->statistics.block_start, 0); |
949 | schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
950 | |
951 | if (tsk) { | |
952 | if (tsk->in_iowait) { | |
4fa8d299 JP |
953 | schedstat_add(se->statistics.iowait_sum, delta); |
954 | schedstat_inc(se->statistics.iowait_count); | |
1a3d027c JP |
955 | trace_sched_stat_iowait(tsk, delta); |
956 | } | |
957 | ||
958 | trace_sched_stat_blocked(tsk, delta); | |
959 | ||
960 | /* | |
961 | * Blocking time is in units of nanosecs, so shift by | |
962 | * 20 to get a milliseconds-range estimation of the | |
963 | * amount of time that the task spent sleeping: | |
964 | */ | |
965 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
966 | profile_hits(SLEEP_PROFILING, | |
967 | (void *)get_wchan(tsk), | |
968 | delta >> 20); | |
969 | } | |
970 | account_scheduler_latency(tsk, delta >> 10, 0); | |
971 | } | |
972 | } | |
3ea94de1 | 973 | } |
3ea94de1 | 974 | |
bf0f6f24 IM |
975 | /* |
976 | * Task is being enqueued - update stats: | |
977 | */ | |
cb251765 | 978 | static inline void |
1a3d027c | 979 | update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 980 | { |
4fa8d299 JP |
981 | if (!schedstat_enabled()) |
982 | return; | |
983 | ||
bf0f6f24 IM |
984 | /* |
985 | * Are we enqueueing a waiting task? (for current tasks | |
986 | * a dequeue/enqueue event is a NOP) | |
987 | */ | |
429d43bc | 988 | if (se != cfs_rq->curr) |
5870db5b | 989 | update_stats_wait_start(cfs_rq, se); |
1a3d027c JP |
990 | |
991 | if (flags & ENQUEUE_WAKEUP) | |
992 | update_stats_enqueue_sleeper(cfs_rq, se); | |
bf0f6f24 IM |
993 | } |
994 | ||
bf0f6f24 | 995 | static inline void |
cb251765 | 996 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 997 | { |
4fa8d299 JP |
998 | |
999 | if (!schedstat_enabled()) | |
1000 | return; | |
1001 | ||
bf0f6f24 IM |
1002 | /* |
1003 | * Mark the end of the wait period if dequeueing a | |
1004 | * waiting task: | |
1005 | */ | |
429d43bc | 1006 | if (se != cfs_rq->curr) |
9ef0a961 | 1007 | update_stats_wait_end(cfs_rq, se); |
cb251765 | 1008 | |
4fa8d299 JP |
1009 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
1010 | struct task_struct *tsk = task_of(se); | |
cb251765 | 1011 | |
4fa8d299 JP |
1012 | if (tsk->state & TASK_INTERRUPTIBLE) |
1013 | schedstat_set(se->statistics.sleep_start, | |
1014 | rq_clock(rq_of(cfs_rq))); | |
1015 | if (tsk->state & TASK_UNINTERRUPTIBLE) | |
1016 | schedstat_set(se->statistics.block_start, | |
1017 | rq_clock(rq_of(cfs_rq))); | |
cb251765 | 1018 | } |
cb251765 MG |
1019 | } |
1020 | ||
bf0f6f24 IM |
1021 | /* |
1022 | * We are picking a new current task - update its stats: | |
1023 | */ | |
1024 | static inline void | |
79303e9e | 1025 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
1026 | { |
1027 | /* | |
1028 | * We are starting a new run period: | |
1029 | */ | |
78becc27 | 1030 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1031 | } |
1032 | ||
bf0f6f24 IM |
1033 | /************************************************** |
1034 | * Scheduling class queueing methods: | |
1035 | */ | |
1036 | ||
cbee9f88 PZ |
1037 | #ifdef CONFIG_NUMA_BALANCING |
1038 | /* | |
598f0ec0 MG |
1039 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1040 | * calculated based on the tasks virtual memory size and | |
1041 | * numa_balancing_scan_size. | |
cbee9f88 | 1042 | */ |
598f0ec0 MG |
1043 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1044 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1045 | |
1046 | /* Portion of address space to scan in MB */ | |
1047 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1048 | |
4b96a29b PZ |
1049 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1050 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1051 | ||
b5dd77c8 RR |
1052 | struct numa_group { |
1053 | atomic_t refcount; | |
1054 | ||
1055 | spinlock_t lock; /* nr_tasks, tasks */ | |
1056 | int nr_tasks; | |
1057 | pid_t gid; | |
1058 | int active_nodes; | |
1059 | ||
1060 | struct rcu_head rcu; | |
1061 | unsigned long total_faults; | |
1062 | unsigned long max_faults_cpu; | |
1063 | /* | |
1064 | * Faults_cpu is used to decide whether memory should move | |
1065 | * towards the CPU. As a consequence, these stats are weighted | |
1066 | * more by CPU use than by memory faults. | |
1067 | */ | |
1068 | unsigned long *faults_cpu; | |
1069 | unsigned long faults[0]; | |
1070 | }; | |
1071 | ||
1072 | static inline unsigned long group_faults_priv(struct numa_group *ng); | |
1073 | static inline unsigned long group_faults_shared(struct numa_group *ng); | |
1074 | ||
598f0ec0 MG |
1075 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1076 | { | |
1077 | unsigned long rss = 0; | |
1078 | unsigned long nr_scan_pages; | |
1079 | ||
1080 | /* | |
1081 | * Calculations based on RSS as non-present and empty pages are skipped | |
1082 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1083 | * on resident pages | |
1084 | */ | |
1085 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1086 | rss = get_mm_rss(p->mm); | |
1087 | if (!rss) | |
1088 | rss = nr_scan_pages; | |
1089 | ||
1090 | rss = round_up(rss, nr_scan_pages); | |
1091 | return rss / nr_scan_pages; | |
1092 | } | |
1093 | ||
1094 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
1095 | #define MAX_SCAN_WINDOW 2560 | |
1096 | ||
1097 | static unsigned int task_scan_min(struct task_struct *p) | |
1098 | { | |
316c1608 | 1099 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1100 | unsigned int scan, floor; |
1101 | unsigned int windows = 1; | |
1102 | ||
64192658 KT |
1103 | if (scan_size < MAX_SCAN_WINDOW) |
1104 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1105 | floor = 1000 / windows; |
1106 | ||
1107 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1108 | return max_t(unsigned int, floor, scan); | |
1109 | } | |
1110 | ||
b5dd77c8 RR |
1111 | static unsigned int task_scan_start(struct task_struct *p) |
1112 | { | |
1113 | unsigned long smin = task_scan_min(p); | |
1114 | unsigned long period = smin; | |
1115 | ||
1116 | /* Scale the maximum scan period with the amount of shared memory. */ | |
1117 | if (p->numa_group) { | |
1118 | struct numa_group *ng = p->numa_group; | |
1119 | unsigned long shared = group_faults_shared(ng); | |
1120 | unsigned long private = group_faults_priv(ng); | |
1121 | ||
1122 | period *= atomic_read(&ng->refcount); | |
1123 | period *= shared + 1; | |
1124 | period /= private + shared + 1; | |
1125 | } | |
1126 | ||
1127 | return max(smin, period); | |
1128 | } | |
1129 | ||
598f0ec0 MG |
1130 | static unsigned int task_scan_max(struct task_struct *p) |
1131 | { | |
b5dd77c8 RR |
1132 | unsigned long smin = task_scan_min(p); |
1133 | unsigned long smax; | |
598f0ec0 MG |
1134 | |
1135 | /* Watch for min being lower than max due to floor calculations */ | |
1136 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
b5dd77c8 RR |
1137 | |
1138 | /* Scale the maximum scan period with the amount of shared memory. */ | |
1139 | if (p->numa_group) { | |
1140 | struct numa_group *ng = p->numa_group; | |
1141 | unsigned long shared = group_faults_shared(ng); | |
1142 | unsigned long private = group_faults_priv(ng); | |
1143 | unsigned long period = smax; | |
1144 | ||
1145 | period *= atomic_read(&ng->refcount); | |
1146 | period *= shared + 1; | |
1147 | period /= private + shared + 1; | |
1148 | ||
1149 | smax = max(smax, period); | |
1150 | } | |
1151 | ||
598f0ec0 MG |
1152 | return max(smin, smax); |
1153 | } | |
1154 | ||
0ec8aa00 PZ |
1155 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1156 | { | |
1157 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | |
1158 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | |
1159 | } | |
1160 | ||
1161 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1162 | { | |
1163 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | |
1164 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | |
1165 | } | |
1166 | ||
be1e4e76 RR |
1167 | /* Shared or private faults. */ |
1168 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1169 | ||
1170 | /* Memory and CPU locality */ | |
1171 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1172 | ||
1173 | /* Averaged statistics, and temporary buffers. */ | |
1174 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1175 | ||
e29cf08b MG |
1176 | pid_t task_numa_group_id(struct task_struct *p) |
1177 | { | |
1178 | return p->numa_group ? p->numa_group->gid : 0; | |
1179 | } | |
1180 | ||
44dba3d5 IM |
1181 | /* |
1182 | * The averaged statistics, shared & private, memory & cpu, | |
1183 | * occupy the first half of the array. The second half of the | |
1184 | * array is for current counters, which are averaged into the | |
1185 | * first set by task_numa_placement. | |
1186 | */ | |
1187 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1188 | { |
44dba3d5 | 1189 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1190 | } |
1191 | ||
1192 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1193 | { | |
44dba3d5 | 1194 | if (!p->numa_faults) |
ac8e895b MG |
1195 | return 0; |
1196 | ||
44dba3d5 IM |
1197 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1198 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1199 | } |
1200 | ||
83e1d2cd MG |
1201 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1202 | { | |
1203 | if (!p->numa_group) | |
1204 | return 0; | |
1205 | ||
44dba3d5 IM |
1206 | return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1207 | p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1208 | } |
1209 | ||
20e07dea RR |
1210 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1211 | { | |
44dba3d5 IM |
1212 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
1213 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | |
20e07dea RR |
1214 | } |
1215 | ||
b5dd77c8 RR |
1216 | static inline unsigned long group_faults_priv(struct numa_group *ng) |
1217 | { | |
1218 | unsigned long faults = 0; | |
1219 | int node; | |
1220 | ||
1221 | for_each_online_node(node) { | |
1222 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
1223 | } | |
1224 | ||
1225 | return faults; | |
1226 | } | |
1227 | ||
1228 | static inline unsigned long group_faults_shared(struct numa_group *ng) | |
1229 | { | |
1230 | unsigned long faults = 0; | |
1231 | int node; | |
1232 | ||
1233 | for_each_online_node(node) { | |
1234 | faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
1235 | } | |
1236 | ||
1237 | return faults; | |
1238 | } | |
1239 | ||
4142c3eb RR |
1240 | /* |
1241 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1242 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1243 | * between these nodes are slowed down, to allow things to settle down. | |
1244 | */ | |
1245 | #define ACTIVE_NODE_FRACTION 3 | |
1246 | ||
1247 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1248 | { | |
1249 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1250 | } | |
1251 | ||
6c6b1193 RR |
1252 | /* Handle placement on systems where not all nodes are directly connected. */ |
1253 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
1254 | int maxdist, bool task) | |
1255 | { | |
1256 | unsigned long score = 0; | |
1257 | int node; | |
1258 | ||
1259 | /* | |
1260 | * All nodes are directly connected, and the same distance | |
1261 | * from each other. No need for fancy placement algorithms. | |
1262 | */ | |
1263 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1264 | return 0; | |
1265 | ||
1266 | /* | |
1267 | * This code is called for each node, introducing N^2 complexity, | |
1268 | * which should be ok given the number of nodes rarely exceeds 8. | |
1269 | */ | |
1270 | for_each_online_node(node) { | |
1271 | unsigned long faults; | |
1272 | int dist = node_distance(nid, node); | |
1273 | ||
1274 | /* | |
1275 | * The furthest away nodes in the system are not interesting | |
1276 | * for placement; nid was already counted. | |
1277 | */ | |
1278 | if (dist == sched_max_numa_distance || node == nid) | |
1279 | continue; | |
1280 | ||
1281 | /* | |
1282 | * On systems with a backplane NUMA topology, compare groups | |
1283 | * of nodes, and move tasks towards the group with the most | |
1284 | * memory accesses. When comparing two nodes at distance | |
1285 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1286 | * of each group. Skip other nodes. | |
1287 | */ | |
1288 | if (sched_numa_topology_type == NUMA_BACKPLANE && | |
1289 | dist > maxdist) | |
1290 | continue; | |
1291 | ||
1292 | /* Add up the faults from nearby nodes. */ | |
1293 | if (task) | |
1294 | faults = task_faults(p, node); | |
1295 | else | |
1296 | faults = group_faults(p, node); | |
1297 | ||
1298 | /* | |
1299 | * On systems with a glueless mesh NUMA topology, there are | |
1300 | * no fixed "groups of nodes". Instead, nodes that are not | |
1301 | * directly connected bounce traffic through intermediate | |
1302 | * nodes; a numa_group can occupy any set of nodes. | |
1303 | * The further away a node is, the less the faults count. | |
1304 | * This seems to result in good task placement. | |
1305 | */ | |
1306 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1307 | faults *= (sched_max_numa_distance - dist); | |
1308 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | |
1309 | } | |
1310 | ||
1311 | score += faults; | |
1312 | } | |
1313 | ||
1314 | return score; | |
1315 | } | |
1316 | ||
83e1d2cd MG |
1317 | /* |
1318 | * These return the fraction of accesses done by a particular task, or | |
1319 | * task group, on a particular numa node. The group weight is given a | |
1320 | * larger multiplier, in order to group tasks together that are almost | |
1321 | * evenly spread out between numa nodes. | |
1322 | */ | |
7bd95320 RR |
1323 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1324 | int dist) | |
83e1d2cd | 1325 | { |
7bd95320 | 1326 | unsigned long faults, total_faults; |
83e1d2cd | 1327 | |
44dba3d5 | 1328 | if (!p->numa_faults) |
83e1d2cd MG |
1329 | return 0; |
1330 | ||
1331 | total_faults = p->total_numa_faults; | |
1332 | ||
1333 | if (!total_faults) | |
1334 | return 0; | |
1335 | ||
7bd95320 | 1336 | faults = task_faults(p, nid); |
6c6b1193 RR |
1337 | faults += score_nearby_nodes(p, nid, dist, true); |
1338 | ||
7bd95320 | 1339 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1340 | } |
1341 | ||
7bd95320 RR |
1342 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1343 | int dist) | |
83e1d2cd | 1344 | { |
7bd95320 RR |
1345 | unsigned long faults, total_faults; |
1346 | ||
1347 | if (!p->numa_group) | |
1348 | return 0; | |
1349 | ||
1350 | total_faults = p->numa_group->total_faults; | |
1351 | ||
1352 | if (!total_faults) | |
83e1d2cd MG |
1353 | return 0; |
1354 | ||
7bd95320 | 1355 | faults = group_faults(p, nid); |
6c6b1193 RR |
1356 | faults += score_nearby_nodes(p, nid, dist, false); |
1357 | ||
7bd95320 | 1358 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1359 | } |
1360 | ||
10f39042 RR |
1361 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1362 | int src_nid, int dst_cpu) | |
1363 | { | |
1364 | struct numa_group *ng = p->numa_group; | |
1365 | int dst_nid = cpu_to_node(dst_cpu); | |
1366 | int last_cpupid, this_cpupid; | |
1367 | ||
1368 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
1369 | ||
1370 | /* | |
1371 | * Multi-stage node selection is used in conjunction with a periodic | |
1372 | * migration fault to build a temporal task<->page relation. By using | |
1373 | * a two-stage filter we remove short/unlikely relations. | |
1374 | * | |
1375 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1376 | * a task's usage of a particular page (n_p) per total usage of this | |
1377 | * page (n_t) (in a given time-span) to a probability. | |
1378 | * | |
1379 | * Our periodic faults will sample this probability and getting the | |
1380 | * same result twice in a row, given these samples are fully | |
1381 | * independent, is then given by P(n)^2, provided our sample period | |
1382 | * is sufficiently short compared to the usage pattern. | |
1383 | * | |
1384 | * This quadric squishes small probabilities, making it less likely we | |
1385 | * act on an unlikely task<->page relation. | |
1386 | */ | |
1387 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); | |
1388 | if (!cpupid_pid_unset(last_cpupid) && | |
1389 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1390 | return false; | |
1391 | ||
1392 | /* Always allow migrate on private faults */ | |
1393 | if (cpupid_match_pid(p, last_cpupid)) | |
1394 | return true; | |
1395 | ||
1396 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1397 | if (!ng) | |
1398 | return true; | |
1399 | ||
1400 | /* | |
4142c3eb RR |
1401 | * Destination node is much more heavily used than the source |
1402 | * node? Allow migration. | |
10f39042 | 1403 | */ |
4142c3eb RR |
1404 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1405 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1406 | return true; |
1407 | ||
1408 | /* | |
4142c3eb RR |
1409 | * Distribute memory according to CPU & memory use on each node, |
1410 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1411 | * | |
1412 | * faults_cpu(dst) 3 faults_cpu(src) | |
1413 | * --------------- * - > --------------- | |
1414 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1415 | */ |
4142c3eb RR |
1416 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1417 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1418 | } |
1419 | ||
c7132dd6 | 1420 | static unsigned long weighted_cpuload(struct rq *rq); |
58d081b5 MG |
1421 | static unsigned long source_load(int cpu, int type); |
1422 | static unsigned long target_load(int cpu, int type); | |
ced549fa | 1423 | static unsigned long capacity_of(int cpu); |
58d081b5 | 1424 | |
fb13c7ee | 1425 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 1426 | struct numa_stats { |
fb13c7ee | 1427 | unsigned long nr_running; |
58d081b5 | 1428 | unsigned long load; |
fb13c7ee MG |
1429 | |
1430 | /* Total compute capacity of CPUs on a node */ | |
5ef20ca1 | 1431 | unsigned long compute_capacity; |
fb13c7ee MG |
1432 | |
1433 | /* Approximate capacity in terms of runnable tasks on a node */ | |
5ef20ca1 | 1434 | unsigned long task_capacity; |
1b6a7495 | 1435 | int has_free_capacity; |
58d081b5 | 1436 | }; |
e6628d5b | 1437 | |
fb13c7ee MG |
1438 | /* |
1439 | * XXX borrowed from update_sg_lb_stats | |
1440 | */ | |
1441 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
1442 | { | |
83d7f242 RR |
1443 | int smt, cpu, cpus = 0; |
1444 | unsigned long capacity; | |
fb13c7ee MG |
1445 | |
1446 | memset(ns, 0, sizeof(*ns)); | |
1447 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
1448 | struct rq *rq = cpu_rq(cpu); | |
1449 | ||
1450 | ns->nr_running += rq->nr_running; | |
c7132dd6 | 1451 | ns->load += weighted_cpuload(rq); |
ced549fa | 1452 | ns->compute_capacity += capacity_of(cpu); |
5eca82a9 PZ |
1453 | |
1454 | cpus++; | |
fb13c7ee MG |
1455 | } |
1456 | ||
5eca82a9 PZ |
1457 | /* |
1458 | * If we raced with hotplug and there are no CPUs left in our mask | |
1459 | * the @ns structure is NULL'ed and task_numa_compare() will | |
1460 | * not find this node attractive. | |
1461 | * | |
1b6a7495 NP |
1462 | * We'll either bail at !has_free_capacity, or we'll detect a huge |
1463 | * imbalance and bail there. | |
5eca82a9 PZ |
1464 | */ |
1465 | if (!cpus) | |
1466 | return; | |
1467 | ||
83d7f242 RR |
1468 | /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */ |
1469 | smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity); | |
1470 | capacity = cpus / smt; /* cores */ | |
1471 | ||
1472 | ns->task_capacity = min_t(unsigned, capacity, | |
1473 | DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE)); | |
1b6a7495 | 1474 | ns->has_free_capacity = (ns->nr_running < ns->task_capacity); |
fb13c7ee MG |
1475 | } |
1476 | ||
58d081b5 MG |
1477 | struct task_numa_env { |
1478 | struct task_struct *p; | |
e6628d5b | 1479 | |
58d081b5 MG |
1480 | int src_cpu, src_nid; |
1481 | int dst_cpu, dst_nid; | |
e6628d5b | 1482 | |
58d081b5 | 1483 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1484 | |
40ea2b42 | 1485 | int imbalance_pct; |
7bd95320 | 1486 | int dist; |
fb13c7ee MG |
1487 | |
1488 | struct task_struct *best_task; | |
1489 | long best_imp; | |
58d081b5 MG |
1490 | int best_cpu; |
1491 | }; | |
1492 | ||
fb13c7ee MG |
1493 | static void task_numa_assign(struct task_numa_env *env, |
1494 | struct task_struct *p, long imp) | |
1495 | { | |
1496 | if (env->best_task) | |
1497 | put_task_struct(env->best_task); | |
bac78573 ON |
1498 | if (p) |
1499 | get_task_struct(p); | |
fb13c7ee MG |
1500 | |
1501 | env->best_task = p; | |
1502 | env->best_imp = imp; | |
1503 | env->best_cpu = env->dst_cpu; | |
1504 | } | |
1505 | ||
28a21745 | 1506 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1507 | struct task_numa_env *env) |
1508 | { | |
e4991b24 RR |
1509 | long imb, old_imb; |
1510 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1511 | long src_capacity, dst_capacity; |
1512 | ||
1513 | /* | |
1514 | * The load is corrected for the CPU capacity available on each node. | |
1515 | * | |
1516 | * src_load dst_load | |
1517 | * ------------ vs --------- | |
1518 | * src_capacity dst_capacity | |
1519 | */ | |
1520 | src_capacity = env->src_stats.compute_capacity; | |
1521 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 RR |
1522 | |
1523 | /* We care about the slope of the imbalance, not the direction. */ | |
e4991b24 RR |
1524 | if (dst_load < src_load) |
1525 | swap(dst_load, src_load); | |
e63da036 RR |
1526 | |
1527 | /* Is the difference below the threshold? */ | |
e4991b24 RR |
1528 | imb = dst_load * src_capacity * 100 - |
1529 | src_load * dst_capacity * env->imbalance_pct; | |
e63da036 RR |
1530 | if (imb <= 0) |
1531 | return false; | |
1532 | ||
1533 | /* | |
1534 | * The imbalance is above the allowed threshold. | |
e4991b24 | 1535 | * Compare it with the old imbalance. |
e63da036 | 1536 | */ |
28a21745 | 1537 | orig_src_load = env->src_stats.load; |
e4991b24 | 1538 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1539 | |
e4991b24 RR |
1540 | if (orig_dst_load < orig_src_load) |
1541 | swap(orig_dst_load, orig_src_load); | |
e63da036 | 1542 | |
e4991b24 RR |
1543 | old_imb = orig_dst_load * src_capacity * 100 - |
1544 | orig_src_load * dst_capacity * env->imbalance_pct; | |
1545 | ||
1546 | /* Would this change make things worse? */ | |
1547 | return (imb > old_imb); | |
e63da036 RR |
1548 | } |
1549 | ||
fb13c7ee MG |
1550 | /* |
1551 | * This checks if the overall compute and NUMA accesses of the system would | |
1552 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1553 | * into account that it might be best if task running on the dst_cpu should | |
1554 | * be exchanged with the source task | |
1555 | */ | |
887c290e RR |
1556 | static void task_numa_compare(struct task_numa_env *env, |
1557 | long taskimp, long groupimp) | |
fb13c7ee MG |
1558 | { |
1559 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
1560 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
1561 | struct task_struct *cur; | |
28a21745 | 1562 | long src_load, dst_load; |
fb13c7ee | 1563 | long load; |
1c5d3eb3 | 1564 | long imp = env->p->numa_group ? groupimp : taskimp; |
0132c3e1 | 1565 | long moveimp = imp; |
7bd95320 | 1566 | int dist = env->dist; |
fb13c7ee MG |
1567 | |
1568 | rcu_read_lock(); | |
bac78573 ON |
1569 | cur = task_rcu_dereference(&dst_rq->curr); |
1570 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) | |
fb13c7ee MG |
1571 | cur = NULL; |
1572 | ||
7af68335 PZ |
1573 | /* |
1574 | * Because we have preemption enabled we can get migrated around and | |
1575 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1576 | */ | |
1577 | if (cur == env->p) | |
1578 | goto unlock; | |
1579 | ||
fb13c7ee MG |
1580 | /* |
1581 | * "imp" is the fault differential for the source task between the | |
1582 | * source and destination node. Calculate the total differential for | |
1583 | * the source task and potential destination task. The more negative | |
1584 | * the value is, the more rmeote accesses that would be expected to | |
1585 | * be incurred if the tasks were swapped. | |
1586 | */ | |
1587 | if (cur) { | |
1588 | /* Skip this swap candidate if cannot move to the source cpu */ | |
0c98d344 | 1589 | if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed)) |
fb13c7ee MG |
1590 | goto unlock; |
1591 | ||
887c290e RR |
1592 | /* |
1593 | * If dst and source tasks are in the same NUMA group, or not | |
ca28aa53 | 1594 | * in any group then look only at task weights. |
887c290e | 1595 | */ |
ca28aa53 | 1596 | if (cur->numa_group == env->p->numa_group) { |
7bd95320 RR |
1597 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1598 | task_weight(cur, env->dst_nid, dist); | |
ca28aa53 RR |
1599 | /* |
1600 | * Add some hysteresis to prevent swapping the | |
1601 | * tasks within a group over tiny differences. | |
1602 | */ | |
1603 | if (cur->numa_group) | |
1604 | imp -= imp/16; | |
887c290e | 1605 | } else { |
ca28aa53 RR |
1606 | /* |
1607 | * Compare the group weights. If a task is all by | |
1608 | * itself (not part of a group), use the task weight | |
1609 | * instead. | |
1610 | */ | |
ca28aa53 | 1611 | if (cur->numa_group) |
7bd95320 RR |
1612 | imp += group_weight(cur, env->src_nid, dist) - |
1613 | group_weight(cur, env->dst_nid, dist); | |
ca28aa53 | 1614 | else |
7bd95320 RR |
1615 | imp += task_weight(cur, env->src_nid, dist) - |
1616 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1617 | } |
fb13c7ee MG |
1618 | } |
1619 | ||
0132c3e1 | 1620 | if (imp <= env->best_imp && moveimp <= env->best_imp) |
fb13c7ee MG |
1621 | goto unlock; |
1622 | ||
1623 | if (!cur) { | |
1624 | /* Is there capacity at our destination? */ | |
b932c03c | 1625 | if (env->src_stats.nr_running <= env->src_stats.task_capacity && |
1b6a7495 | 1626 | !env->dst_stats.has_free_capacity) |
fb13c7ee MG |
1627 | goto unlock; |
1628 | ||
1629 | goto balance; | |
1630 | } | |
1631 | ||
1632 | /* Balance doesn't matter much if we're running a task per cpu */ | |
0132c3e1 RR |
1633 | if (imp > env->best_imp && src_rq->nr_running == 1 && |
1634 | dst_rq->nr_running == 1) | |
fb13c7ee MG |
1635 | goto assign; |
1636 | ||
1637 | /* | |
1638 | * In the overloaded case, try and keep the load balanced. | |
1639 | */ | |
1640 | balance: | |
e720fff6 PZ |
1641 | load = task_h_load(env->p); |
1642 | dst_load = env->dst_stats.load + load; | |
1643 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1644 | |
0132c3e1 RR |
1645 | if (moveimp > imp && moveimp > env->best_imp) { |
1646 | /* | |
1647 | * If the improvement from just moving env->p direction is | |
1648 | * better than swapping tasks around, check if a move is | |
1649 | * possible. Store a slightly smaller score than moveimp, | |
1650 | * so an actually idle CPU will win. | |
1651 | */ | |
1652 | if (!load_too_imbalanced(src_load, dst_load, env)) { | |
1653 | imp = moveimp - 1; | |
1654 | cur = NULL; | |
1655 | goto assign; | |
1656 | } | |
1657 | } | |
1658 | ||
1659 | if (imp <= env->best_imp) | |
1660 | goto unlock; | |
1661 | ||
fb13c7ee | 1662 | if (cur) { |
e720fff6 PZ |
1663 | load = task_h_load(cur); |
1664 | dst_load -= load; | |
1665 | src_load += load; | |
fb13c7ee MG |
1666 | } |
1667 | ||
28a21745 | 1668 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1669 | goto unlock; |
1670 | ||
ba7e5a27 RR |
1671 | /* |
1672 | * One idle CPU per node is evaluated for a task numa move. | |
1673 | * Call select_idle_sibling to maybe find a better one. | |
1674 | */ | |
10e2f1ac PZ |
1675 | if (!cur) { |
1676 | /* | |
1677 | * select_idle_siblings() uses an per-cpu cpumask that | |
1678 | * can be used from IRQ context. | |
1679 | */ | |
1680 | local_irq_disable(); | |
772bd008 MR |
1681 | env->dst_cpu = select_idle_sibling(env->p, env->src_cpu, |
1682 | env->dst_cpu); | |
10e2f1ac PZ |
1683 | local_irq_enable(); |
1684 | } | |
ba7e5a27 | 1685 | |
fb13c7ee MG |
1686 | assign: |
1687 | task_numa_assign(env, cur, imp); | |
1688 | unlock: | |
1689 | rcu_read_unlock(); | |
1690 | } | |
1691 | ||
887c290e RR |
1692 | static void task_numa_find_cpu(struct task_numa_env *env, |
1693 | long taskimp, long groupimp) | |
2c8a50aa MG |
1694 | { |
1695 | int cpu; | |
1696 | ||
1697 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1698 | /* Skip this CPU if the source task cannot migrate */ | |
0c98d344 | 1699 | if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed)) |
2c8a50aa MG |
1700 | continue; |
1701 | ||
1702 | env->dst_cpu = cpu; | |
887c290e | 1703 | task_numa_compare(env, taskimp, groupimp); |
2c8a50aa MG |
1704 | } |
1705 | } | |
1706 | ||
6f9aad0b RR |
1707 | /* Only move tasks to a NUMA node less busy than the current node. */ |
1708 | static bool numa_has_capacity(struct task_numa_env *env) | |
1709 | { | |
1710 | struct numa_stats *src = &env->src_stats; | |
1711 | struct numa_stats *dst = &env->dst_stats; | |
1712 | ||
1713 | if (src->has_free_capacity && !dst->has_free_capacity) | |
1714 | return false; | |
1715 | ||
1716 | /* | |
1717 | * Only consider a task move if the source has a higher load | |
1718 | * than the destination, corrected for CPU capacity on each node. | |
1719 | * | |
1720 | * src->load dst->load | |
1721 | * --------------------- vs --------------------- | |
1722 | * src->compute_capacity dst->compute_capacity | |
1723 | */ | |
44dcb04f SD |
1724 | if (src->load * dst->compute_capacity * env->imbalance_pct > |
1725 | ||
1726 | dst->load * src->compute_capacity * 100) | |
6f9aad0b RR |
1727 | return true; |
1728 | ||
1729 | return false; | |
1730 | } | |
1731 | ||
58d081b5 MG |
1732 | static int task_numa_migrate(struct task_struct *p) |
1733 | { | |
58d081b5 MG |
1734 | struct task_numa_env env = { |
1735 | .p = p, | |
fb13c7ee | 1736 | |
58d081b5 | 1737 | .src_cpu = task_cpu(p), |
b32e86b4 | 1738 | .src_nid = task_node(p), |
fb13c7ee MG |
1739 | |
1740 | .imbalance_pct = 112, | |
1741 | ||
1742 | .best_task = NULL, | |
1743 | .best_imp = 0, | |
4142c3eb | 1744 | .best_cpu = -1, |
58d081b5 MG |
1745 | }; |
1746 | struct sched_domain *sd; | |
887c290e | 1747 | unsigned long taskweight, groupweight; |
7bd95320 | 1748 | int nid, ret, dist; |
887c290e | 1749 | long taskimp, groupimp; |
e6628d5b | 1750 | |
58d081b5 | 1751 | /* |
fb13c7ee MG |
1752 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1753 | * imbalance and would be the first to start moving tasks about. | |
1754 | * | |
1755 | * And we want to avoid any moving of tasks about, as that would create | |
1756 | * random movement of tasks -- counter the numa conditions we're trying | |
1757 | * to satisfy here. | |
58d081b5 MG |
1758 | */ |
1759 | rcu_read_lock(); | |
fb13c7ee | 1760 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1761 | if (sd) |
1762 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1763 | rcu_read_unlock(); |
1764 | ||
46a73e8a RR |
1765 | /* |
1766 | * Cpusets can break the scheduler domain tree into smaller | |
1767 | * balance domains, some of which do not cross NUMA boundaries. | |
1768 | * Tasks that are "trapped" in such domains cannot be migrated | |
1769 | * elsewhere, so there is no point in (re)trying. | |
1770 | */ | |
1771 | if (unlikely(!sd)) { | |
de1b301a | 1772 | p->numa_preferred_nid = task_node(p); |
46a73e8a RR |
1773 | return -EINVAL; |
1774 | } | |
1775 | ||
2c8a50aa | 1776 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
1777 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
1778 | taskweight = task_weight(p, env.src_nid, dist); | |
1779 | groupweight = group_weight(p, env.src_nid, dist); | |
1780 | update_numa_stats(&env.src_stats, env.src_nid); | |
1781 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; | |
1782 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
2c8a50aa | 1783 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1784 | |
a43455a1 | 1785 | /* Try to find a spot on the preferred nid. */ |
6f9aad0b RR |
1786 | if (numa_has_capacity(&env)) |
1787 | task_numa_find_cpu(&env, taskimp, groupimp); | |
e1dda8a7 | 1788 | |
9de05d48 RR |
1789 | /* |
1790 | * Look at other nodes in these cases: | |
1791 | * - there is no space available on the preferred_nid | |
1792 | * - the task is part of a numa_group that is interleaved across | |
1793 | * multiple NUMA nodes; in order to better consolidate the group, | |
1794 | * we need to check other locations. | |
1795 | */ | |
4142c3eb | 1796 | if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) { |
2c8a50aa MG |
1797 | for_each_online_node(nid) { |
1798 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1799 | continue; | |
58d081b5 | 1800 | |
7bd95320 | 1801 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
1802 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
1803 | dist != env.dist) { | |
1804 | taskweight = task_weight(p, env.src_nid, dist); | |
1805 | groupweight = group_weight(p, env.src_nid, dist); | |
1806 | } | |
7bd95320 | 1807 | |
83e1d2cd | 1808 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
1809 | taskimp = task_weight(p, nid, dist) - taskweight; |
1810 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 1811 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
1812 | continue; |
1813 | ||
7bd95320 | 1814 | env.dist = dist; |
2c8a50aa MG |
1815 | env.dst_nid = nid; |
1816 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
6f9aad0b RR |
1817 | if (numa_has_capacity(&env)) |
1818 | task_numa_find_cpu(&env, taskimp, groupimp); | |
58d081b5 MG |
1819 | } |
1820 | } | |
1821 | ||
68d1b02a RR |
1822 | /* |
1823 | * If the task is part of a workload that spans multiple NUMA nodes, | |
1824 | * and is migrating into one of the workload's active nodes, remember | |
1825 | * this node as the task's preferred numa node, so the workload can | |
1826 | * settle down. | |
1827 | * A task that migrated to a second choice node will be better off | |
1828 | * trying for a better one later. Do not set the preferred node here. | |
1829 | */ | |
db015dae | 1830 | if (p->numa_group) { |
4142c3eb RR |
1831 | struct numa_group *ng = p->numa_group; |
1832 | ||
db015dae RR |
1833 | if (env.best_cpu == -1) |
1834 | nid = env.src_nid; | |
1835 | else | |
1836 | nid = env.dst_nid; | |
1837 | ||
4142c3eb | 1838 | if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng)) |
db015dae RR |
1839 | sched_setnuma(p, env.dst_nid); |
1840 | } | |
1841 | ||
1842 | /* No better CPU than the current one was found. */ | |
1843 | if (env.best_cpu == -1) | |
1844 | return -EAGAIN; | |
0ec8aa00 | 1845 | |
04bb2f94 RR |
1846 | /* |
1847 | * Reset the scan period if the task is being rescheduled on an | |
1848 | * alternative node to recheck if the tasks is now properly placed. | |
1849 | */ | |
b5dd77c8 | 1850 | p->numa_scan_period = task_scan_start(p); |
04bb2f94 | 1851 | |
fb13c7ee | 1852 | if (env.best_task == NULL) { |
286549dc MG |
1853 | ret = migrate_task_to(p, env.best_cpu); |
1854 | if (ret != 0) | |
1855 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); | |
fb13c7ee MG |
1856 | return ret; |
1857 | } | |
1858 | ||
1859 | ret = migrate_swap(p, env.best_task); | |
286549dc MG |
1860 | if (ret != 0) |
1861 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); | |
fb13c7ee MG |
1862 | put_task_struct(env.best_task); |
1863 | return ret; | |
e6628d5b MG |
1864 | } |
1865 | ||
6b9a7460 MG |
1866 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1867 | static void numa_migrate_preferred(struct task_struct *p) | |
1868 | { | |
5085e2a3 RR |
1869 | unsigned long interval = HZ; |
1870 | ||
2739d3ee | 1871 | /* This task has no NUMA fault statistics yet */ |
44dba3d5 | 1872 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults)) |
6b9a7460 MG |
1873 | return; |
1874 | ||
2739d3ee | 1875 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 RR |
1876 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
1877 | p->numa_migrate_retry = jiffies + interval; | |
2739d3ee RR |
1878 | |
1879 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 1880 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
1881 | return; |
1882 | ||
1883 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 1884 | task_numa_migrate(p); |
6b9a7460 MG |
1885 | } |
1886 | ||
20e07dea | 1887 | /* |
4142c3eb | 1888 | * Find out how many nodes on the workload is actively running on. Do this by |
20e07dea RR |
1889 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
1890 | * be different from the set of nodes where the workload's memory is currently | |
1891 | * located. | |
20e07dea | 1892 | */ |
4142c3eb | 1893 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
1894 | { |
1895 | unsigned long faults, max_faults = 0; | |
4142c3eb | 1896 | int nid, active_nodes = 0; |
20e07dea RR |
1897 | |
1898 | for_each_online_node(nid) { | |
1899 | faults = group_faults_cpu(numa_group, nid); | |
1900 | if (faults > max_faults) | |
1901 | max_faults = faults; | |
1902 | } | |
1903 | ||
1904 | for_each_online_node(nid) { | |
1905 | faults = group_faults_cpu(numa_group, nid); | |
4142c3eb RR |
1906 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
1907 | active_nodes++; | |
20e07dea | 1908 | } |
4142c3eb RR |
1909 | |
1910 | numa_group->max_faults_cpu = max_faults; | |
1911 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
1912 | } |
1913 | ||
04bb2f94 RR |
1914 | /* |
1915 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1916 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
1917 | * period will be for the next scan window. If local/(local+remote) ratio is |
1918 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
1919 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
1920 | */ |
1921 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 1922 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
1923 | |
1924 | /* | |
1925 | * Increase the scan period (slow down scanning) if the majority of | |
1926 | * our memory is already on our local node, or if the majority of | |
1927 | * the page accesses are shared with other processes. | |
1928 | * Otherwise, decrease the scan period. | |
1929 | */ | |
1930 | static void update_task_scan_period(struct task_struct *p, | |
1931 | unsigned long shared, unsigned long private) | |
1932 | { | |
1933 | unsigned int period_slot; | |
37ec97de | 1934 | int lr_ratio, ps_ratio; |
04bb2f94 RR |
1935 | int diff; |
1936 | ||
1937 | unsigned long remote = p->numa_faults_locality[0]; | |
1938 | unsigned long local = p->numa_faults_locality[1]; | |
1939 | ||
1940 | /* | |
1941 | * If there were no record hinting faults then either the task is | |
1942 | * completely idle or all activity is areas that are not of interest | |
074c2381 MG |
1943 | * to automatic numa balancing. Related to that, if there were failed |
1944 | * migration then it implies we are migrating too quickly or the local | |
1945 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 1946 | */ |
074c2381 | 1947 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
1948 | p->numa_scan_period = min(p->numa_scan_period_max, |
1949 | p->numa_scan_period << 1); | |
1950 | ||
1951 | p->mm->numa_next_scan = jiffies + | |
1952 | msecs_to_jiffies(p->numa_scan_period); | |
1953 | ||
1954 | return; | |
1955 | } | |
1956 | ||
1957 | /* | |
1958 | * Prepare to scale scan period relative to the current period. | |
1959 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1960 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1961 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1962 | */ | |
1963 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
37ec97de RR |
1964 | lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); |
1965 | ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); | |
1966 | ||
1967 | if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { | |
1968 | /* | |
1969 | * Most memory accesses are local. There is no need to | |
1970 | * do fast NUMA scanning, since memory is already local. | |
1971 | */ | |
1972 | int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; | |
1973 | if (!slot) | |
1974 | slot = 1; | |
1975 | diff = slot * period_slot; | |
1976 | } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { | |
1977 | /* | |
1978 | * Most memory accesses are shared with other tasks. | |
1979 | * There is no point in continuing fast NUMA scanning, | |
1980 | * since other tasks may just move the memory elsewhere. | |
1981 | */ | |
1982 | int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; | |
04bb2f94 RR |
1983 | if (!slot) |
1984 | slot = 1; | |
1985 | diff = slot * period_slot; | |
1986 | } else { | |
04bb2f94 | 1987 | /* |
37ec97de RR |
1988 | * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, |
1989 | * yet they are not on the local NUMA node. Speed up | |
1990 | * NUMA scanning to get the memory moved over. | |
04bb2f94 | 1991 | */ |
37ec97de RR |
1992 | int ratio = max(lr_ratio, ps_ratio); |
1993 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
04bb2f94 RR |
1994 | } |
1995 | ||
1996 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
1997 | task_scan_min(p), task_scan_max(p)); | |
1998 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
1999 | } | |
2000 | ||
7e2703e6 RR |
2001 | /* |
2002 | * Get the fraction of time the task has been running since the last | |
2003 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
2004 | * decays those on a 32ms period, which is orders of magnitude off | |
2005 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
2006 | * stats only if the task is so new there are no NUMA statistics yet. | |
2007 | */ | |
2008 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
2009 | { | |
2010 | u64 runtime, delta, now; | |
2011 | /* Use the start of this time slice to avoid calculations. */ | |
2012 | now = p->se.exec_start; | |
2013 | runtime = p->se.sum_exec_runtime; | |
2014 | ||
2015 | if (p->last_task_numa_placement) { | |
2016 | delta = runtime - p->last_sum_exec_runtime; | |
2017 | *period = now - p->last_task_numa_placement; | |
2018 | } else { | |
c7b50216 | 2019 | delta = p->se.avg.load_sum; |
9d89c257 | 2020 | *period = LOAD_AVG_MAX; |
7e2703e6 RR |
2021 | } |
2022 | ||
2023 | p->last_sum_exec_runtime = runtime; | |
2024 | p->last_task_numa_placement = now; | |
2025 | ||
2026 | return delta; | |
2027 | } | |
2028 | ||
54009416 RR |
2029 | /* |
2030 | * Determine the preferred nid for a task in a numa_group. This needs to | |
2031 | * be done in a way that produces consistent results with group_weight, | |
2032 | * otherwise workloads might not converge. | |
2033 | */ | |
2034 | static int preferred_group_nid(struct task_struct *p, int nid) | |
2035 | { | |
2036 | nodemask_t nodes; | |
2037 | int dist; | |
2038 | ||
2039 | /* Direct connections between all NUMA nodes. */ | |
2040 | if (sched_numa_topology_type == NUMA_DIRECT) | |
2041 | return nid; | |
2042 | ||
2043 | /* | |
2044 | * On a system with glueless mesh NUMA topology, group_weight | |
2045 | * scores nodes according to the number of NUMA hinting faults on | |
2046 | * both the node itself, and on nearby nodes. | |
2047 | */ | |
2048 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
2049 | unsigned long score, max_score = 0; | |
2050 | int node, max_node = nid; | |
2051 | ||
2052 | dist = sched_max_numa_distance; | |
2053 | ||
2054 | for_each_online_node(node) { | |
2055 | score = group_weight(p, node, dist); | |
2056 | if (score > max_score) { | |
2057 | max_score = score; | |
2058 | max_node = node; | |
2059 | } | |
2060 | } | |
2061 | return max_node; | |
2062 | } | |
2063 | ||
2064 | /* | |
2065 | * Finding the preferred nid in a system with NUMA backplane | |
2066 | * interconnect topology is more involved. The goal is to locate | |
2067 | * tasks from numa_groups near each other in the system, and | |
2068 | * untangle workloads from different sides of the system. This requires | |
2069 | * searching down the hierarchy of node groups, recursively searching | |
2070 | * inside the highest scoring group of nodes. The nodemask tricks | |
2071 | * keep the complexity of the search down. | |
2072 | */ | |
2073 | nodes = node_online_map; | |
2074 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | |
2075 | unsigned long max_faults = 0; | |
81907478 | 2076 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
2077 | int a, b; |
2078 | ||
2079 | /* Are there nodes at this distance from each other? */ | |
2080 | if (!find_numa_distance(dist)) | |
2081 | continue; | |
2082 | ||
2083 | for_each_node_mask(a, nodes) { | |
2084 | unsigned long faults = 0; | |
2085 | nodemask_t this_group; | |
2086 | nodes_clear(this_group); | |
2087 | ||
2088 | /* Sum group's NUMA faults; includes a==b case. */ | |
2089 | for_each_node_mask(b, nodes) { | |
2090 | if (node_distance(a, b) < dist) { | |
2091 | faults += group_faults(p, b); | |
2092 | node_set(b, this_group); | |
2093 | node_clear(b, nodes); | |
2094 | } | |
2095 | } | |
2096 | ||
2097 | /* Remember the top group. */ | |
2098 | if (faults > max_faults) { | |
2099 | max_faults = faults; | |
2100 | max_group = this_group; | |
2101 | /* | |
2102 | * subtle: at the smallest distance there is | |
2103 | * just one node left in each "group", the | |
2104 | * winner is the preferred nid. | |
2105 | */ | |
2106 | nid = a; | |
2107 | } | |
2108 | } | |
2109 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2110 | if (!max_faults) |
2111 | break; | |
54009416 RR |
2112 | nodes = max_group; |
2113 | } | |
2114 | return nid; | |
2115 | } | |
2116 | ||
cbee9f88 PZ |
2117 | static void task_numa_placement(struct task_struct *p) |
2118 | { | |
83e1d2cd MG |
2119 | int seq, nid, max_nid = -1, max_group_nid = -1; |
2120 | unsigned long max_faults = 0, max_group_faults = 0; | |
04bb2f94 | 2121 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2122 | unsigned long total_faults; |
2123 | u64 runtime, period; | |
7dbd13ed | 2124 | spinlock_t *group_lock = NULL; |
cbee9f88 | 2125 | |
7e5a2c17 JL |
2126 | /* |
2127 | * The p->mm->numa_scan_seq field gets updated without | |
2128 | * exclusive access. Use READ_ONCE() here to ensure | |
2129 | * that the field is read in a single access: | |
2130 | */ | |
316c1608 | 2131 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2132 | if (p->numa_scan_seq == seq) |
2133 | return; | |
2134 | p->numa_scan_seq = seq; | |
598f0ec0 | 2135 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2136 | |
7e2703e6 RR |
2137 | total_faults = p->numa_faults_locality[0] + |
2138 | p->numa_faults_locality[1]; | |
2139 | runtime = numa_get_avg_runtime(p, &period); | |
2140 | ||
7dbd13ed MG |
2141 | /* If the task is part of a group prevent parallel updates to group stats */ |
2142 | if (p->numa_group) { | |
2143 | group_lock = &p->numa_group->lock; | |
60e69eed | 2144 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2145 | } |
2146 | ||
688b7585 MG |
2147 | /* Find the node with the highest number of faults */ |
2148 | for_each_online_node(nid) { | |
44dba3d5 IM |
2149 | /* Keep track of the offsets in numa_faults array */ |
2150 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2151 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2152 | int priv; |
745d6147 | 2153 | |
be1e4e76 | 2154 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2155 | long diff, f_diff, f_weight; |
8c8a743c | 2156 | |
44dba3d5 IM |
2157 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2158 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2159 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2160 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2161 | |
ac8e895b | 2162 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2163 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2164 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2165 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2166 | |
7e2703e6 RR |
2167 | /* |
2168 | * Normalize the faults_from, so all tasks in a group | |
2169 | * count according to CPU use, instead of by the raw | |
2170 | * number of faults. Tasks with little runtime have | |
2171 | * little over-all impact on throughput, and thus their | |
2172 | * faults are less important. | |
2173 | */ | |
2174 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2175 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2176 | (total_faults + 1); |
44dba3d5 IM |
2177 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2178 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2179 | |
44dba3d5 IM |
2180 | p->numa_faults[mem_idx] += diff; |
2181 | p->numa_faults[cpu_idx] += f_diff; | |
2182 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2183 | p->total_numa_faults += diff; |
8c8a743c | 2184 | if (p->numa_group) { |
44dba3d5 IM |
2185 | /* |
2186 | * safe because we can only change our own group | |
2187 | * | |
2188 | * mem_idx represents the offset for a given | |
2189 | * nid and priv in a specific region because it | |
2190 | * is at the beginning of the numa_faults array. | |
2191 | */ | |
2192 | p->numa_group->faults[mem_idx] += diff; | |
2193 | p->numa_group->faults_cpu[mem_idx] += f_diff; | |
989348b5 | 2194 | p->numa_group->total_faults += diff; |
44dba3d5 | 2195 | group_faults += p->numa_group->faults[mem_idx]; |
8c8a743c | 2196 | } |
ac8e895b MG |
2197 | } |
2198 | ||
688b7585 MG |
2199 | if (faults > max_faults) { |
2200 | max_faults = faults; | |
2201 | max_nid = nid; | |
2202 | } | |
83e1d2cd MG |
2203 | |
2204 | if (group_faults > max_group_faults) { | |
2205 | max_group_faults = group_faults; | |
2206 | max_group_nid = nid; | |
2207 | } | |
2208 | } | |
2209 | ||
04bb2f94 RR |
2210 | update_task_scan_period(p, fault_types[0], fault_types[1]); |
2211 | ||
7dbd13ed | 2212 | if (p->numa_group) { |
4142c3eb | 2213 | numa_group_count_active_nodes(p->numa_group); |
60e69eed | 2214 | spin_unlock_irq(group_lock); |
54009416 | 2215 | max_nid = preferred_group_nid(p, max_group_nid); |
688b7585 MG |
2216 | } |
2217 | ||
bb97fc31 RR |
2218 | if (max_faults) { |
2219 | /* Set the new preferred node */ | |
2220 | if (max_nid != p->numa_preferred_nid) | |
2221 | sched_setnuma(p, max_nid); | |
2222 | ||
2223 | if (task_node(p) != p->numa_preferred_nid) | |
2224 | numa_migrate_preferred(p); | |
3a7053b3 | 2225 | } |
cbee9f88 PZ |
2226 | } |
2227 | ||
8c8a743c PZ |
2228 | static inline int get_numa_group(struct numa_group *grp) |
2229 | { | |
2230 | return atomic_inc_not_zero(&grp->refcount); | |
2231 | } | |
2232 | ||
2233 | static inline void put_numa_group(struct numa_group *grp) | |
2234 | { | |
2235 | if (atomic_dec_and_test(&grp->refcount)) | |
2236 | kfree_rcu(grp, rcu); | |
2237 | } | |
2238 | ||
3e6a9418 MG |
2239 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2240 | int *priv) | |
8c8a743c PZ |
2241 | { |
2242 | struct numa_group *grp, *my_grp; | |
2243 | struct task_struct *tsk; | |
2244 | bool join = false; | |
2245 | int cpu = cpupid_to_cpu(cpupid); | |
2246 | int i; | |
2247 | ||
2248 | if (unlikely(!p->numa_group)) { | |
2249 | unsigned int size = sizeof(struct numa_group) + | |
50ec8a40 | 2250 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
2251 | |
2252 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2253 | if (!grp) | |
2254 | return; | |
2255 | ||
2256 | atomic_set(&grp->refcount, 1); | |
4142c3eb RR |
2257 | grp->active_nodes = 1; |
2258 | grp->max_faults_cpu = 0; | |
8c8a743c | 2259 | spin_lock_init(&grp->lock); |
e29cf08b | 2260 | grp->gid = p->pid; |
50ec8a40 | 2261 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
2262 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
2263 | nr_node_ids; | |
8c8a743c | 2264 | |
be1e4e76 | 2265 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2266 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2267 | |
989348b5 | 2268 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2269 | |
8c8a743c PZ |
2270 | grp->nr_tasks++; |
2271 | rcu_assign_pointer(p->numa_group, grp); | |
2272 | } | |
2273 | ||
2274 | rcu_read_lock(); | |
316c1608 | 2275 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2276 | |
2277 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2278 | goto no_join; |
8c8a743c PZ |
2279 | |
2280 | grp = rcu_dereference(tsk->numa_group); | |
2281 | if (!grp) | |
3354781a | 2282 | goto no_join; |
8c8a743c PZ |
2283 | |
2284 | my_grp = p->numa_group; | |
2285 | if (grp == my_grp) | |
3354781a | 2286 | goto no_join; |
8c8a743c PZ |
2287 | |
2288 | /* | |
2289 | * Only join the other group if its bigger; if we're the bigger group, | |
2290 | * the other task will join us. | |
2291 | */ | |
2292 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2293 | goto no_join; |
8c8a743c PZ |
2294 | |
2295 | /* | |
2296 | * Tie-break on the grp address. | |
2297 | */ | |
2298 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2299 | goto no_join; |
8c8a743c | 2300 | |
dabe1d99 RR |
2301 | /* Always join threads in the same process. */ |
2302 | if (tsk->mm == current->mm) | |
2303 | join = true; | |
2304 | ||
2305 | /* Simple filter to avoid false positives due to PID collisions */ | |
2306 | if (flags & TNF_SHARED) | |
2307 | join = true; | |
8c8a743c | 2308 | |
3e6a9418 MG |
2309 | /* Update priv based on whether false sharing was detected */ |
2310 | *priv = !join; | |
2311 | ||
dabe1d99 | 2312 | if (join && !get_numa_group(grp)) |
3354781a | 2313 | goto no_join; |
8c8a743c | 2314 | |
8c8a743c PZ |
2315 | rcu_read_unlock(); |
2316 | ||
2317 | if (!join) | |
2318 | return; | |
2319 | ||
60e69eed MG |
2320 | BUG_ON(irqs_disabled()); |
2321 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 2322 | |
be1e4e76 | 2323 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2324 | my_grp->faults[i] -= p->numa_faults[i]; |
2325 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2326 | } |
989348b5 MG |
2327 | my_grp->total_faults -= p->total_numa_faults; |
2328 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2329 | |
8c8a743c PZ |
2330 | my_grp->nr_tasks--; |
2331 | grp->nr_tasks++; | |
2332 | ||
2333 | spin_unlock(&my_grp->lock); | |
60e69eed | 2334 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2335 | |
2336 | rcu_assign_pointer(p->numa_group, grp); | |
2337 | ||
2338 | put_numa_group(my_grp); | |
3354781a PZ |
2339 | return; |
2340 | ||
2341 | no_join: | |
2342 | rcu_read_unlock(); | |
2343 | return; | |
8c8a743c PZ |
2344 | } |
2345 | ||
2346 | void task_numa_free(struct task_struct *p) | |
2347 | { | |
2348 | struct numa_group *grp = p->numa_group; | |
44dba3d5 | 2349 | void *numa_faults = p->numa_faults; |
e9dd685c SR |
2350 | unsigned long flags; |
2351 | int i; | |
8c8a743c PZ |
2352 | |
2353 | if (grp) { | |
e9dd685c | 2354 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2355 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2356 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2357 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2358 | |
8c8a743c | 2359 | grp->nr_tasks--; |
e9dd685c | 2360 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2361 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2362 | put_numa_group(grp); |
2363 | } | |
2364 | ||
44dba3d5 | 2365 | p->numa_faults = NULL; |
82727018 | 2366 | kfree(numa_faults); |
8c8a743c PZ |
2367 | } |
2368 | ||
cbee9f88 PZ |
2369 | /* |
2370 | * Got a PROT_NONE fault for a page on @node. | |
2371 | */ | |
58b46da3 | 2372 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2373 | { |
2374 | struct task_struct *p = current; | |
6688cc05 | 2375 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2376 | int cpu_node = task_node(current); |
792568ec | 2377 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 2378 | struct numa_group *ng; |
ac8e895b | 2379 | int priv; |
cbee9f88 | 2380 | |
2a595721 | 2381 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2382 | return; |
2383 | ||
9ff1d9ff MG |
2384 | /* for example, ksmd faulting in a user's mm */ |
2385 | if (!p->mm) | |
2386 | return; | |
2387 | ||
f809ca9a | 2388 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2389 | if (unlikely(!p->numa_faults)) { |
2390 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2391 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2392 | |
44dba3d5 IM |
2393 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2394 | if (!p->numa_faults) | |
f809ca9a | 2395 | return; |
745d6147 | 2396 | |
83e1d2cd | 2397 | p->total_numa_faults = 0; |
04bb2f94 | 2398 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2399 | } |
cbee9f88 | 2400 | |
8c8a743c PZ |
2401 | /* |
2402 | * First accesses are treated as private, otherwise consider accesses | |
2403 | * to be private if the accessing pid has not changed | |
2404 | */ | |
2405 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2406 | priv = 1; | |
2407 | } else { | |
2408 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2409 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2410 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2411 | } |
2412 | ||
792568ec RR |
2413 | /* |
2414 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2415 | * occurs wholly within the set of nodes that the workload is | |
2416 | * actively using should be counted as local. This allows the | |
2417 | * scan rate to slow down when a workload has settled down. | |
2418 | */ | |
4142c3eb RR |
2419 | ng = p->numa_group; |
2420 | if (!priv && !local && ng && ng->active_nodes > 1 && | |
2421 | numa_is_active_node(cpu_node, ng) && | |
2422 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
2423 | local = 1; |
2424 | ||
cbee9f88 | 2425 | task_numa_placement(p); |
f809ca9a | 2426 | |
2739d3ee RR |
2427 | /* |
2428 | * Retry task to preferred node migration periodically, in case it | |
2429 | * case it previously failed, or the scheduler moved us. | |
2430 | */ | |
2431 | if (time_after(jiffies, p->numa_migrate_retry)) | |
6b9a7460 MG |
2432 | numa_migrate_preferred(p); |
2433 | ||
b32e86b4 IM |
2434 | if (migrated) |
2435 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2436 | if (flags & TNF_MIGRATE_FAIL) |
2437 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2438 | |
44dba3d5 IM |
2439 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2440 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2441 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2442 | } |
2443 | ||
6e5fb223 PZ |
2444 | static void reset_ptenuma_scan(struct task_struct *p) |
2445 | { | |
7e5a2c17 JL |
2446 | /* |
2447 | * We only did a read acquisition of the mmap sem, so | |
2448 | * p->mm->numa_scan_seq is written to without exclusive access | |
2449 | * and the update is not guaranteed to be atomic. That's not | |
2450 | * much of an issue though, since this is just used for | |
2451 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2452 | * expensive, to avoid any form of compiler optimizations: | |
2453 | */ | |
316c1608 | 2454 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2455 | p->mm->numa_scan_offset = 0; |
2456 | } | |
2457 | ||
cbee9f88 PZ |
2458 | /* |
2459 | * The expensive part of numa migration is done from task_work context. | |
2460 | * Triggered from task_tick_numa(). | |
2461 | */ | |
2462 | void task_numa_work(struct callback_head *work) | |
2463 | { | |
2464 | unsigned long migrate, next_scan, now = jiffies; | |
2465 | struct task_struct *p = current; | |
2466 | struct mm_struct *mm = p->mm; | |
51170840 | 2467 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 2468 | struct vm_area_struct *vma; |
9f40604c | 2469 | unsigned long start, end; |
598f0ec0 | 2470 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2471 | long pages, virtpages; |
cbee9f88 | 2472 | |
9148a3a1 | 2473 | SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); |
cbee9f88 PZ |
2474 | |
2475 | work->next = work; /* protect against double add */ | |
2476 | /* | |
2477 | * Who cares about NUMA placement when they're dying. | |
2478 | * | |
2479 | * NOTE: make sure not to dereference p->mm before this check, | |
2480 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2481 | * without p->mm even though we still had it when we enqueued this | |
2482 | * work. | |
2483 | */ | |
2484 | if (p->flags & PF_EXITING) | |
2485 | return; | |
2486 | ||
930aa174 | 2487 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2488 | mm->numa_next_scan = now + |
2489 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2490 | } |
2491 | ||
cbee9f88 PZ |
2492 | /* |
2493 | * Enforce maximal scan/migration frequency.. | |
2494 | */ | |
2495 | migrate = mm->numa_next_scan; | |
2496 | if (time_before(now, migrate)) | |
2497 | return; | |
2498 | ||
598f0ec0 MG |
2499 | if (p->numa_scan_period == 0) { |
2500 | p->numa_scan_period_max = task_scan_max(p); | |
b5dd77c8 | 2501 | p->numa_scan_period = task_scan_start(p); |
598f0ec0 | 2502 | } |
cbee9f88 | 2503 | |
fb003b80 | 2504 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2505 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2506 | return; | |
2507 | ||
19a78d11 PZ |
2508 | /* |
2509 | * Delay this task enough that another task of this mm will likely win | |
2510 | * the next time around. | |
2511 | */ | |
2512 | p->node_stamp += 2 * TICK_NSEC; | |
2513 | ||
9f40604c MG |
2514 | start = mm->numa_scan_offset; |
2515 | pages = sysctl_numa_balancing_scan_size; | |
2516 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2517 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2518 | if (!pages) |
2519 | return; | |
cbee9f88 | 2520 | |
4620f8c1 | 2521 | |
8655d549 VB |
2522 | if (!down_read_trylock(&mm->mmap_sem)) |
2523 | return; | |
9f40604c | 2524 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2525 | if (!vma) { |
2526 | reset_ptenuma_scan(p); | |
9f40604c | 2527 | start = 0; |
6e5fb223 PZ |
2528 | vma = mm->mmap; |
2529 | } | |
9f40604c | 2530 | for (; vma; vma = vma->vm_next) { |
6b79c57b | 2531 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 2532 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 2533 | continue; |
6b79c57b | 2534 | } |
6e5fb223 | 2535 | |
4591ce4f MG |
2536 | /* |
2537 | * Shared library pages mapped by multiple processes are not | |
2538 | * migrated as it is expected they are cache replicated. Avoid | |
2539 | * hinting faults in read-only file-backed mappings or the vdso | |
2540 | * as migrating the pages will be of marginal benefit. | |
2541 | */ | |
2542 | if (!vma->vm_mm || | |
2543 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2544 | continue; | |
2545 | ||
3c67f474 MG |
2546 | /* |
2547 | * Skip inaccessible VMAs to avoid any confusion between | |
2548 | * PROT_NONE and NUMA hinting ptes | |
2549 | */ | |
2550 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | |
2551 | continue; | |
4591ce4f | 2552 | |
9f40604c MG |
2553 | do { |
2554 | start = max(start, vma->vm_start); | |
2555 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2556 | end = min(end, vma->vm_end); | |
4620f8c1 | 2557 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
2558 | |
2559 | /* | |
4620f8c1 RR |
2560 | * Try to scan sysctl_numa_balancing_size worth of |
2561 | * hpages that have at least one present PTE that | |
2562 | * is not already pte-numa. If the VMA contains | |
2563 | * areas that are unused or already full of prot_numa | |
2564 | * PTEs, scan up to virtpages, to skip through those | |
2565 | * areas faster. | |
598f0ec0 MG |
2566 | */ |
2567 | if (nr_pte_updates) | |
2568 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 2569 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 2570 | |
9f40604c | 2571 | start = end; |
4620f8c1 | 2572 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 2573 | goto out; |
3cf1962c RR |
2574 | |
2575 | cond_resched(); | |
9f40604c | 2576 | } while (end != vma->vm_end); |
cbee9f88 | 2577 | } |
6e5fb223 | 2578 | |
9f40604c | 2579 | out: |
6e5fb223 | 2580 | /* |
c69307d5 PZ |
2581 | * It is possible to reach the end of the VMA list but the last few |
2582 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2583 | * would find the !migratable VMA on the next scan but not reset the | |
2584 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2585 | */ |
2586 | if (vma) | |
9f40604c | 2587 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2588 | else |
2589 | reset_ptenuma_scan(p); | |
2590 | up_read(&mm->mmap_sem); | |
51170840 RR |
2591 | |
2592 | /* | |
2593 | * Make sure tasks use at least 32x as much time to run other code | |
2594 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
2595 | * Usually update_task_scan_period slows down scanning enough; on an | |
2596 | * overloaded system we need to limit overhead on a per task basis. | |
2597 | */ | |
2598 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
2599 | u64 diff = p->se.sum_exec_runtime - runtime; | |
2600 | p->node_stamp += 32 * diff; | |
2601 | } | |
cbee9f88 PZ |
2602 | } |
2603 | ||
2604 | /* | |
2605 | * Drive the periodic memory faults.. | |
2606 | */ | |
2607 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2608 | { | |
2609 | struct callback_head *work = &curr->numa_work; | |
2610 | u64 period, now; | |
2611 | ||
2612 | /* | |
2613 | * We don't care about NUMA placement if we don't have memory. | |
2614 | */ | |
2615 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
2616 | return; | |
2617 | ||
2618 | /* | |
2619 | * Using runtime rather than walltime has the dual advantage that | |
2620 | * we (mostly) drive the selection from busy threads and that the | |
2621 | * task needs to have done some actual work before we bother with | |
2622 | * NUMA placement. | |
2623 | */ | |
2624 | now = curr->se.sum_exec_runtime; | |
2625 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2626 | ||
25b3e5a3 | 2627 | if (now > curr->node_stamp + period) { |
4b96a29b | 2628 | if (!curr->node_stamp) |
b5dd77c8 | 2629 | curr->numa_scan_period = task_scan_start(curr); |
19a78d11 | 2630 | curr->node_stamp += period; |
cbee9f88 PZ |
2631 | |
2632 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
2633 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
2634 | task_work_add(curr, work, true); | |
2635 | } | |
2636 | } | |
2637 | } | |
3fed382b | 2638 | |
cbee9f88 PZ |
2639 | #else |
2640 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2641 | { | |
2642 | } | |
0ec8aa00 PZ |
2643 | |
2644 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2645 | { | |
2646 | } | |
2647 | ||
2648 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2649 | { | |
2650 | } | |
3fed382b | 2651 | |
cbee9f88 PZ |
2652 | #endif /* CONFIG_NUMA_BALANCING */ |
2653 | ||
30cfdcfc DA |
2654 | static void |
2655 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2656 | { | |
2657 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2658 | if (!parent_entity(se)) |
029632fb | 2659 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 2660 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2661 | if (entity_is_task(se)) { |
2662 | struct rq *rq = rq_of(cfs_rq); | |
2663 | ||
2664 | account_numa_enqueue(rq, task_of(se)); | |
2665 | list_add(&se->group_node, &rq->cfs_tasks); | |
2666 | } | |
367456c7 | 2667 | #endif |
30cfdcfc | 2668 | cfs_rq->nr_running++; |
30cfdcfc DA |
2669 | } |
2670 | ||
2671 | static void | |
2672 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2673 | { | |
2674 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2675 | if (!parent_entity(se)) |
029632fb | 2676 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
bfdb198c | 2677 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2678 | if (entity_is_task(se)) { |
2679 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 2680 | list_del_init(&se->group_node); |
0ec8aa00 | 2681 | } |
bfdb198c | 2682 | #endif |
30cfdcfc | 2683 | cfs_rq->nr_running--; |
30cfdcfc DA |
2684 | } |
2685 | ||
8d5b9025 PZ |
2686 | /* |
2687 | * Signed add and clamp on underflow. | |
2688 | * | |
2689 | * Explicitly do a load-store to ensure the intermediate value never hits | |
2690 | * memory. This allows lockless observations without ever seeing the negative | |
2691 | * values. | |
2692 | */ | |
2693 | #define add_positive(_ptr, _val) do { \ | |
2694 | typeof(_ptr) ptr = (_ptr); \ | |
2695 | typeof(_val) val = (_val); \ | |
2696 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
2697 | \ | |
2698 | res = var + val; \ | |
2699 | \ | |
2700 | if (val < 0 && res > var) \ | |
2701 | res = 0; \ | |
2702 | \ | |
2703 | WRITE_ONCE(*ptr, res); \ | |
2704 | } while (0) | |
2705 | ||
2706 | /* | |
2707 | * Unsigned subtract and clamp on underflow. | |
2708 | * | |
2709 | * Explicitly do a load-store to ensure the intermediate value never hits | |
2710 | * memory. This allows lockless observations without ever seeing the negative | |
2711 | * values. | |
2712 | */ | |
2713 | #define sub_positive(_ptr, _val) do { \ | |
2714 | typeof(_ptr) ptr = (_ptr); \ | |
2715 | typeof(*ptr) val = (_val); \ | |
2716 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
2717 | res = var - val; \ | |
2718 | if (res > var) \ | |
2719 | res = 0; \ | |
2720 | WRITE_ONCE(*ptr, res); \ | |
2721 | } while (0) | |
2722 | ||
2723 | #ifdef CONFIG_SMP | |
2724 | /* | |
1ea6c46a | 2725 | * XXX we want to get rid of these helpers and use the full load resolution. |
8d5b9025 PZ |
2726 | */ |
2727 | static inline long se_weight(struct sched_entity *se) | |
2728 | { | |
2729 | return scale_load_down(se->load.weight); | |
2730 | } | |
2731 | ||
1ea6c46a PZ |
2732 | static inline long se_runnable(struct sched_entity *se) |
2733 | { | |
2734 | return scale_load_down(se->runnable_weight); | |
2735 | } | |
2736 | ||
8d5b9025 PZ |
2737 | static inline void |
2738 | enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2739 | { | |
1ea6c46a PZ |
2740 | cfs_rq->runnable_weight += se->runnable_weight; |
2741 | ||
2742 | cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg; | |
2743 | cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum; | |
8d5b9025 PZ |
2744 | } |
2745 | ||
2746 | static inline void | |
2747 | dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2748 | { | |
1ea6c46a PZ |
2749 | cfs_rq->runnable_weight -= se->runnable_weight; |
2750 | ||
2751 | sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg); | |
2752 | sub_positive(&cfs_rq->avg.runnable_load_sum, | |
2753 | se_runnable(se) * se->avg.runnable_load_sum); | |
8d5b9025 PZ |
2754 | } |
2755 | ||
2756 | static inline void | |
2757 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2758 | { | |
2759 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
2760 | cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; | |
2761 | } | |
2762 | ||
2763 | static inline void | |
2764 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2765 | { | |
2766 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); | |
2767 | sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); | |
2768 | } | |
2769 | #else | |
2770 | static inline void | |
2771 | enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2772 | static inline void | |
2773 | dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2774 | static inline void | |
2775 | enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2776 | static inline void | |
2777 | dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } | |
2778 | #endif | |
2779 | ||
9059393e | 2780 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
1ea6c46a | 2781 | unsigned long weight, unsigned long runnable) |
9059393e VG |
2782 | { |
2783 | if (se->on_rq) { | |
2784 | /* commit outstanding execution time */ | |
2785 | if (cfs_rq->curr == se) | |
2786 | update_curr(cfs_rq); | |
2787 | account_entity_dequeue(cfs_rq, se); | |
2788 | dequeue_runnable_load_avg(cfs_rq, se); | |
2789 | } | |
2790 | dequeue_load_avg(cfs_rq, se); | |
2791 | ||
1ea6c46a | 2792 | se->runnable_weight = runnable; |
9059393e VG |
2793 | update_load_set(&se->load, weight); |
2794 | ||
2795 | #ifdef CONFIG_SMP | |
1ea6c46a PZ |
2796 | do { |
2797 | u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib; | |
2798 | ||
2799 | se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); | |
2800 | se->avg.runnable_load_avg = | |
2801 | div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider); | |
2802 | } while (0); | |
9059393e VG |
2803 | #endif |
2804 | ||
2805 | enqueue_load_avg(cfs_rq, se); | |
2806 | if (se->on_rq) { | |
2807 | account_entity_enqueue(cfs_rq, se); | |
2808 | enqueue_runnable_load_avg(cfs_rq, se); | |
2809 | } | |
2810 | } | |
2811 | ||
2812 | void reweight_task(struct task_struct *p, int prio) | |
2813 | { | |
2814 | struct sched_entity *se = &p->se; | |
2815 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2816 | struct load_weight *load = &se->load; | |
2817 | unsigned long weight = scale_load(sched_prio_to_weight[prio]); | |
2818 | ||
1ea6c46a | 2819 | reweight_entity(cfs_rq, se, weight, weight); |
9059393e VG |
2820 | load->inv_weight = sched_prio_to_wmult[prio]; |
2821 | } | |
2822 | ||
3ff6dcac YZ |
2823 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2824 | # ifdef CONFIG_SMP | |
cef27403 PZ |
2825 | /* |
2826 | * All this does is approximate the hierarchical proportion which includes that | |
2827 | * global sum we all love to hate. | |
2828 | * | |
2829 | * That is, the weight of a group entity, is the proportional share of the | |
2830 | * group weight based on the group runqueue weights. That is: | |
2831 | * | |
2832 | * tg->weight * grq->load.weight | |
2833 | * ge->load.weight = ----------------------------- (1) | |
2834 | * \Sum grq->load.weight | |
2835 | * | |
2836 | * Now, because computing that sum is prohibitively expensive to compute (been | |
2837 | * there, done that) we approximate it with this average stuff. The average | |
2838 | * moves slower and therefore the approximation is cheaper and more stable. | |
2839 | * | |
2840 | * So instead of the above, we substitute: | |
2841 | * | |
2842 | * grq->load.weight -> grq->avg.load_avg (2) | |
2843 | * | |
2844 | * which yields the following: | |
2845 | * | |
2846 | * tg->weight * grq->avg.load_avg | |
2847 | * ge->load.weight = ------------------------------ (3) | |
2848 | * tg->load_avg | |
2849 | * | |
2850 | * Where: tg->load_avg ~= \Sum grq->avg.load_avg | |
2851 | * | |
2852 | * That is shares_avg, and it is right (given the approximation (2)). | |
2853 | * | |
2854 | * The problem with it is that because the average is slow -- it was designed | |
2855 | * to be exactly that of course -- this leads to transients in boundary | |
2856 | * conditions. In specific, the case where the group was idle and we start the | |
2857 | * one task. It takes time for our CPU's grq->avg.load_avg to build up, | |
2858 | * yielding bad latency etc.. | |
2859 | * | |
2860 | * Now, in that special case (1) reduces to: | |
2861 | * | |
2862 | * tg->weight * grq->load.weight | |
17de4ee0 | 2863 | * ge->load.weight = ----------------------------- = tg->weight (4) |
cef27403 PZ |
2864 | * grp->load.weight |
2865 | * | |
2866 | * That is, the sum collapses because all other CPUs are idle; the UP scenario. | |
2867 | * | |
2868 | * So what we do is modify our approximation (3) to approach (4) in the (near) | |
2869 | * UP case, like: | |
2870 | * | |
2871 | * ge->load.weight = | |
2872 | * | |
2873 | * tg->weight * grq->load.weight | |
2874 | * --------------------------------------------------- (5) | |
2875 | * tg->load_avg - grq->avg.load_avg + grq->load.weight | |
2876 | * | |
17de4ee0 PZ |
2877 | * But because grq->load.weight can drop to 0, resulting in a divide by zero, |
2878 | * we need to use grq->avg.load_avg as its lower bound, which then gives: | |
2879 | * | |
2880 | * | |
2881 | * tg->weight * grq->load.weight | |
2882 | * ge->load.weight = ----------------------------- (6) | |
2883 | * tg_load_avg' | |
2884 | * | |
2885 | * Where: | |
2886 | * | |
2887 | * tg_load_avg' = tg->load_avg - grq->avg.load_avg + | |
2888 | * max(grq->load.weight, grq->avg.load_avg) | |
cef27403 PZ |
2889 | * |
2890 | * And that is shares_weight and is icky. In the (near) UP case it approaches | |
2891 | * (4) while in the normal case it approaches (3). It consistently | |
2892 | * overestimates the ge->load.weight and therefore: | |
2893 | * | |
2894 | * \Sum ge->load.weight >= tg->weight | |
2895 | * | |
2896 | * hence icky! | |
2897 | */ | |
2c8e4dce | 2898 | static long calc_group_shares(struct cfs_rq *cfs_rq) |
cf5f0acf | 2899 | { |
7c80cfc9 PZ |
2900 | long tg_weight, tg_shares, load, shares; |
2901 | struct task_group *tg = cfs_rq->tg; | |
2902 | ||
2903 | tg_shares = READ_ONCE(tg->shares); | |
cf5f0acf | 2904 | |
3d4b60d3 | 2905 | load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); |
cf5f0acf | 2906 | |
ea1dc6fc | 2907 | tg_weight = atomic_long_read(&tg->load_avg); |
3ff6dcac | 2908 | |
ea1dc6fc PZ |
2909 | /* Ensure tg_weight >= load */ |
2910 | tg_weight -= cfs_rq->tg_load_avg_contrib; | |
2911 | tg_weight += load; | |
3ff6dcac | 2912 | |
7c80cfc9 | 2913 | shares = (tg_shares * load); |
cf5f0acf PZ |
2914 | if (tg_weight) |
2915 | shares /= tg_weight; | |
3ff6dcac | 2916 | |
b8fd8423 DE |
2917 | /* |
2918 | * MIN_SHARES has to be unscaled here to support per-CPU partitioning | |
2919 | * of a group with small tg->shares value. It is a floor value which is | |
2920 | * assigned as a minimum load.weight to the sched_entity representing | |
2921 | * the group on a CPU. | |
2922 | * | |
2923 | * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 | |
2924 | * on an 8-core system with 8 tasks each runnable on one CPU shares has | |
2925 | * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In | |
2926 | * case no task is runnable on a CPU MIN_SHARES=2 should be returned | |
2927 | * instead of 0. | |
2928 | */ | |
7c80cfc9 | 2929 | return clamp_t(long, shares, MIN_SHARES, tg_shares); |
3ff6dcac | 2930 | } |
2c8e4dce JB |
2931 | |
2932 | /* | |
17de4ee0 PZ |
2933 | * This calculates the effective runnable weight for a group entity based on |
2934 | * the group entity weight calculated above. | |
2935 | * | |
2936 | * Because of the above approximation (2), our group entity weight is | |
2937 | * an load_avg based ratio (3). This means that it includes blocked load and | |
2938 | * does not represent the runnable weight. | |
2939 | * | |
2940 | * Approximate the group entity's runnable weight per ratio from the group | |
2941 | * runqueue: | |
2942 | * | |
2943 | * grq->avg.runnable_load_avg | |
2944 | * ge->runnable_weight = ge->load.weight * -------------------------- (7) | |
2945 | * grq->avg.load_avg | |
2946 | * | |
2947 | * However, analogous to above, since the avg numbers are slow, this leads to | |
2948 | * transients in the from-idle case. Instead we use: | |
2949 | * | |
2950 | * ge->runnable_weight = ge->load.weight * | |
2951 | * | |
2952 | * max(grq->avg.runnable_load_avg, grq->runnable_weight) | |
2953 | * ----------------------------------------------------- (8) | |
2954 | * max(grq->avg.load_avg, grq->load.weight) | |
2955 | * | |
2956 | * Where these max() serve both to use the 'instant' values to fix the slow | |
2957 | * from-idle and avoid the /0 on to-idle, similar to (6). | |
2c8e4dce JB |
2958 | */ |
2959 | static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares) | |
2960 | { | |
17de4ee0 PZ |
2961 | long runnable, load_avg; |
2962 | ||
2963 | load_avg = max(cfs_rq->avg.load_avg, | |
2964 | scale_load_down(cfs_rq->load.weight)); | |
2965 | ||
2966 | runnable = max(cfs_rq->avg.runnable_load_avg, | |
2967 | scale_load_down(cfs_rq->runnable_weight)); | |
2c8e4dce JB |
2968 | |
2969 | runnable *= shares; | |
2970 | if (load_avg) | |
2971 | runnable /= load_avg; | |
17de4ee0 | 2972 | |
2c8e4dce JB |
2973 | return clamp_t(long, runnable, MIN_SHARES, shares); |
2974 | } | |
3ff6dcac | 2975 | # endif /* CONFIG_SMP */ |
ea1dc6fc | 2976 | |
82958366 PT |
2977 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
2978 | ||
1ea6c46a PZ |
2979 | /* |
2980 | * Recomputes the group entity based on the current state of its group | |
2981 | * runqueue. | |
2982 | */ | |
2983 | static void update_cfs_group(struct sched_entity *se) | |
2069dd75 | 2984 | { |
1ea6c46a PZ |
2985 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); |
2986 | long shares, runnable; | |
2069dd75 | 2987 | |
1ea6c46a | 2988 | if (!gcfs_rq) |
89ee048f VG |
2989 | return; |
2990 | ||
1ea6c46a | 2991 | if (throttled_hierarchy(gcfs_rq)) |
2069dd75 | 2992 | return; |
89ee048f | 2993 | |
3ff6dcac | 2994 | #ifndef CONFIG_SMP |
1ea6c46a | 2995 | runnable = shares = READ_ONCE(gcfs_rq->tg->shares); |
7c80cfc9 PZ |
2996 | |
2997 | if (likely(se->load.weight == shares)) | |
3ff6dcac | 2998 | return; |
7c80cfc9 | 2999 | #else |
2c8e4dce JB |
3000 | shares = calc_group_shares(gcfs_rq); |
3001 | runnable = calc_group_runnable(gcfs_rq, shares); | |
3ff6dcac | 3002 | #endif |
2069dd75 | 3003 | |
1ea6c46a | 3004 | reweight_entity(cfs_rq_of(se), se, shares, runnable); |
2069dd75 | 3005 | } |
89ee048f | 3006 | |
2069dd75 | 3007 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
1ea6c46a | 3008 | static inline void update_cfs_group(struct sched_entity *se) |
2069dd75 PZ |
3009 | { |
3010 | } | |
3011 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
3012 | ||
a030d738 VK |
3013 | static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq) |
3014 | { | |
43964409 LT |
3015 | struct rq *rq = rq_of(cfs_rq); |
3016 | ||
3017 | if (&rq->cfs == cfs_rq) { | |
a030d738 VK |
3018 | /* |
3019 | * There are a few boundary cases this might miss but it should | |
3020 | * get called often enough that that should (hopefully) not be | |
3021 | * a real problem -- added to that it only calls on the local | |
3022 | * CPU, so if we enqueue remotely we'll miss an update, but | |
3023 | * the next tick/schedule should update. | |
3024 | * | |
3025 | * It will not get called when we go idle, because the idle | |
3026 | * thread is a different class (!fair), nor will the utilization | |
3027 | * number include things like RT tasks. | |
3028 | * | |
3029 | * As is, the util number is not freq-invariant (we'd have to | |
3030 | * implement arch_scale_freq_capacity() for that). | |
3031 | * | |
3032 | * See cpu_util(). | |
3033 | */ | |
43964409 | 3034 | cpufreq_update_util(rq, 0); |
a030d738 VK |
3035 | } |
3036 | } | |
3037 | ||
141965c7 | 3038 | #ifdef CONFIG_SMP |
9d85f21c PT |
3039 | /* |
3040 | * Approximate: | |
3041 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
3042 | */ | |
a481db34 | 3043 | static u64 decay_load(u64 val, u64 n) |
9d85f21c | 3044 | { |
5b51f2f8 PT |
3045 | unsigned int local_n; |
3046 | ||
05296e75 | 3047 | if (unlikely(n > LOAD_AVG_PERIOD * 63)) |
5b51f2f8 PT |
3048 | return 0; |
3049 | ||
3050 | /* after bounds checking we can collapse to 32-bit */ | |
3051 | local_n = n; | |
3052 | ||
3053 | /* | |
3054 | * As y^PERIOD = 1/2, we can combine | |
9c58c79a ZZ |
3055 | * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD) |
3056 | * With a look-up table which covers y^n (n<PERIOD) | |
5b51f2f8 PT |
3057 | * |
3058 | * To achieve constant time decay_load. | |
3059 | */ | |
3060 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
3061 | val >>= local_n / LOAD_AVG_PERIOD; | |
3062 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
3063 | } |
3064 | ||
9d89c257 YD |
3065 | val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32); |
3066 | return val; | |
5b51f2f8 PT |
3067 | } |
3068 | ||
05296e75 | 3069 | static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3) |
5b51f2f8 | 3070 | { |
05296e75 | 3071 | u32 c1, c2, c3 = d3; /* y^0 == 1 */ |
5b51f2f8 | 3072 | |
a481db34 | 3073 | /* |
3841cdc3 | 3074 | * c1 = d1 y^p |
a481db34 | 3075 | */ |
05296e75 | 3076 | c1 = decay_load((u64)d1, periods); |
a481db34 | 3077 | |
a481db34 | 3078 | /* |
3841cdc3 | 3079 | * p-1 |
05296e75 PZ |
3080 | * c2 = 1024 \Sum y^n |
3081 | * n=1 | |
a481db34 | 3082 | * |
05296e75 PZ |
3083 | * inf inf |
3084 | * = 1024 ( \Sum y^n - \Sum y^n - y^0 ) | |
3841cdc3 | 3085 | * n=0 n=p |
a481db34 | 3086 | */ |
05296e75 | 3087 | c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024; |
a481db34 YD |
3088 | |
3089 | return c1 + c2 + c3; | |
9d85f21c PT |
3090 | } |
3091 | ||
54a21385 | 3092 | #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT) |
e0f5f3af | 3093 | |
a481db34 YD |
3094 | /* |
3095 | * Accumulate the three separate parts of the sum; d1 the remainder | |
3096 | * of the last (incomplete) period, d2 the span of full periods and d3 | |
3097 | * the remainder of the (incomplete) current period. | |
3098 | * | |
3099 | * d1 d2 d3 | |
3100 | * ^ ^ ^ | |
3101 | * | | | | |
3102 | * |<->|<----------------->|<--->| | |
3103 | * ... |---x---|------| ... |------|-----x (now) | |
3104 | * | |
3841cdc3 PZ |
3105 | * p-1 |
3106 | * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0 | |
3107 | * n=1 | |
a481db34 | 3108 | * |
3841cdc3 | 3109 | * = u y^p + (Step 1) |
a481db34 | 3110 | * |
3841cdc3 PZ |
3111 | * p-1 |
3112 | * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2) | |
3113 | * n=1 | |
a481db34 YD |
3114 | */ |
3115 | static __always_inline u32 | |
3116 | accumulate_sum(u64 delta, int cpu, struct sched_avg *sa, | |
1ea6c46a | 3117 | unsigned long load, unsigned long runnable, int running) |
a481db34 YD |
3118 | { |
3119 | unsigned long scale_freq, scale_cpu; | |
05296e75 | 3120 | u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */ |
a481db34 | 3121 | u64 periods; |
a481db34 YD |
3122 | |
3123 | scale_freq = arch_scale_freq_capacity(NULL, cpu); | |
3124 | scale_cpu = arch_scale_cpu_capacity(NULL, cpu); | |
3125 | ||
3126 | delta += sa->period_contrib; | |
3127 | periods = delta / 1024; /* A period is 1024us (~1ms) */ | |
3128 | ||
3129 | /* | |
3130 | * Step 1: decay old *_sum if we crossed period boundaries. | |
3131 | */ | |
3132 | if (periods) { | |
3133 | sa->load_sum = decay_load(sa->load_sum, periods); | |
1ea6c46a PZ |
3134 | sa->runnable_load_sum = |
3135 | decay_load(sa->runnable_load_sum, periods); | |
a481db34 | 3136 | sa->util_sum = decay_load((u64)(sa->util_sum), periods); |
a481db34 | 3137 | |
05296e75 PZ |
3138 | /* |
3139 | * Step 2 | |
3140 | */ | |
3141 | delta %= 1024; | |
3142 | contrib = __accumulate_pelt_segments(periods, | |
3143 | 1024 - sa->period_contrib, delta); | |
3144 | } | |
a481db34 YD |
3145 | sa->period_contrib = delta; |
3146 | ||
3147 | contrib = cap_scale(contrib, scale_freq); | |
1ea6c46a PZ |
3148 | if (load) |
3149 | sa->load_sum += load * contrib; | |
3150 | if (runnable) | |
3151 | sa->runnable_load_sum += runnable * contrib; | |
a481db34 YD |
3152 | if (running) |
3153 | sa->util_sum += contrib * scale_cpu; | |
3154 | ||
3155 | return periods; | |
3156 | } | |
3157 | ||
9d85f21c PT |
3158 | /* |
3159 | * We can represent the historical contribution to runnable average as the | |
3160 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
3161 | * history into segments of approximately 1ms (1024us); label the segment that | |
3162 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
3163 | * | |
3164 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
3165 | * p0 p1 p2 | |
3166 | * (now) (~1ms ago) (~2ms ago) | |
3167 | * | |
3168 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
3169 | * | |
3170 | * We then designate the fractions u_i as our co-efficients, yielding the | |
3171 | * following representation of historical load: | |
3172 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
3173 | * | |
3174 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
3175 | * y^32 = 0.5 | |
3176 | * | |
3177 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
3178 | * approximately half as much as the contribution to load within the last ms | |
3179 | * (u_0). | |
3180 | * | |
3181 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
3182 | * sum again by y is sufficient to update: | |
3183 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
3184 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
3185 | */ | |
9d89c257 | 3186 | static __always_inline int |
c7b50216 | 3187 | ___update_load_sum(u64 now, int cpu, struct sched_avg *sa, |
1ea6c46a | 3188 | unsigned long load, unsigned long runnable, int running) |
9d85f21c | 3189 | { |
a481db34 | 3190 | u64 delta; |
9d85f21c | 3191 | |
9d89c257 | 3192 | delta = now - sa->last_update_time; |
9d85f21c PT |
3193 | /* |
3194 | * This should only happen when time goes backwards, which it | |
3195 | * unfortunately does during sched clock init when we swap over to TSC. | |
3196 | */ | |
3197 | if ((s64)delta < 0) { | |
9d89c257 | 3198 | sa->last_update_time = now; |
9d85f21c PT |
3199 | return 0; |
3200 | } | |
3201 | ||
3202 | /* | |
3203 | * Use 1024ns as the unit of measurement since it's a reasonable | |
3204 | * approximation of 1us and fast to compute. | |
3205 | */ | |
3206 | delta >>= 10; | |
3207 | if (!delta) | |
3208 | return 0; | |
bb0bd044 PZ |
3209 | |
3210 | sa->last_update_time += delta << 10; | |
9d85f21c | 3211 | |
f235a54f VG |
3212 | /* |
3213 | * running is a subset of runnable (weight) so running can't be set if | |
3214 | * runnable is clear. But there are some corner cases where the current | |
3215 | * se has been already dequeued but cfs_rq->curr still points to it. | |
3216 | * This means that weight will be 0 but not running for a sched_entity | |
3217 | * but also for a cfs_rq if the latter becomes idle. As an example, | |
3218 | * this happens during idle_balance() which calls | |
3219 | * update_blocked_averages() | |
3220 | */ | |
1ea6c46a PZ |
3221 | if (!load) |
3222 | runnable = running = 0; | |
f235a54f | 3223 | |
a481db34 YD |
3224 | /* |
3225 | * Now we know we crossed measurement unit boundaries. The *_avg | |
3226 | * accrues by two steps: | |
3227 | * | |
3228 | * Step 1: accumulate *_sum since last_update_time. If we haven't | |
3229 | * crossed period boundaries, finish. | |
3230 | */ | |
1ea6c46a | 3231 | if (!accumulate_sum(delta, cpu, sa, load, runnable, running)) |
a481db34 | 3232 | return 0; |
9ee474f5 | 3233 | |
c7b50216 PZ |
3234 | return 1; |
3235 | } | |
3236 | ||
3237 | static __always_inline void | |
1ea6c46a | 3238 | ___update_load_avg(struct sched_avg *sa, unsigned long load, unsigned long runnable) |
c7b50216 PZ |
3239 | { |
3240 | u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib; | |
3241 | ||
a481db34 YD |
3242 | /* |
3243 | * Step 2: update *_avg. | |
3244 | */ | |
1ea6c46a PZ |
3245 | sa->load_avg = div_u64(load * sa->load_sum, divider); |
3246 | sa->runnable_load_avg = div_u64(runnable * sa->runnable_load_sum, divider); | |
c7b50216 PZ |
3247 | sa->util_avg = sa->util_sum / divider; |
3248 | } | |
aff3e498 | 3249 | |
c7b50216 PZ |
3250 | /* |
3251 | * sched_entity: | |
3252 | * | |
1ea6c46a PZ |
3253 | * task: |
3254 | * se_runnable() == se_weight() | |
3255 | * | |
3256 | * group: [ see update_cfs_group() ] | |
3257 | * se_weight() = tg->weight * grq->load_avg / tg->load_avg | |
3258 | * se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg | |
3259 | * | |
c7b50216 PZ |
3260 | * load_sum := runnable_sum |
3261 | * load_avg = se_weight(se) * runnable_avg | |
3262 | * | |
1ea6c46a PZ |
3263 | * runnable_load_sum := runnable_sum |
3264 | * runnable_load_avg = se_runnable(se) * runnable_avg | |
3265 | * | |
3266 | * XXX collapse load_sum and runnable_load_sum | |
3267 | * | |
c7b50216 PZ |
3268 | * cfq_rs: |
3269 | * | |
3270 | * load_sum = \Sum se_weight(se) * se->avg.load_sum | |
3271 | * load_avg = \Sum se->avg.load_avg | |
1ea6c46a PZ |
3272 | * |
3273 | * runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum | |
3274 | * runnable_load_avg = \Sum se->avg.runable_load_avg | |
c7b50216 PZ |
3275 | */ |
3276 | ||
0ccb977f PZ |
3277 | static int |
3278 | __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se) | |
3279 | { | |
1ea6c46a PZ |
3280 | if (entity_is_task(se)) |
3281 | se->runnable_weight = se->load.weight; | |
3282 | ||
3283 | if (___update_load_sum(now, cpu, &se->avg, 0, 0, 0)) { | |
3284 | ___update_load_avg(&se->avg, se_weight(se), se_runnable(se)); | |
c7b50216 PZ |
3285 | return 1; |
3286 | } | |
3287 | ||
3288 | return 0; | |
0ccb977f PZ |
3289 | } |
3290 | ||
3291 | static int | |
3292 | __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3293 | { | |
1ea6c46a PZ |
3294 | if (entity_is_task(se)) |
3295 | se->runnable_weight = se->load.weight; | |
3296 | ||
3297 | if (___update_load_sum(now, cpu, &se->avg, !!se->on_rq, !!se->on_rq, | |
3298 | cfs_rq->curr == se)) { | |
c7b50216 | 3299 | |
1ea6c46a | 3300 | ___update_load_avg(&se->avg, se_weight(se), se_runnable(se)); |
c7b50216 PZ |
3301 | return 1; |
3302 | } | |
3303 | ||
3304 | return 0; | |
0ccb977f PZ |
3305 | } |
3306 | ||
3307 | static int | |
3308 | __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq) | |
3309 | { | |
c7b50216 PZ |
3310 | if (___update_load_sum(now, cpu, &cfs_rq->avg, |
3311 | scale_load_down(cfs_rq->load.weight), | |
1ea6c46a PZ |
3312 | scale_load_down(cfs_rq->runnable_weight), |
3313 | cfs_rq->curr != NULL)) { | |
3314 | ||
3315 | ___update_load_avg(&cfs_rq->avg, 1, 1); | |
c7b50216 PZ |
3316 | return 1; |
3317 | } | |
3318 | ||
3319 | return 0; | |
0ccb977f PZ |
3320 | } |
3321 | ||
c566e8e9 | 3322 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7c3edd2c PZ |
3323 | /** |
3324 | * update_tg_load_avg - update the tg's load avg | |
3325 | * @cfs_rq: the cfs_rq whose avg changed | |
3326 | * @force: update regardless of how small the difference | |
3327 | * | |
3328 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
3329 | * However, because tg->load_avg is a global value there are performance | |
3330 | * considerations. | |
3331 | * | |
3332 | * In order to avoid having to look at the other cfs_rq's, we use a | |
3333 | * differential update where we store the last value we propagated. This in | |
3334 | * turn allows skipping updates if the differential is 'small'. | |
3335 | * | |
815abf5a | 3336 | * Updating tg's load_avg is necessary before update_cfs_share(). |
bb17f655 | 3337 | */ |
9d89c257 | 3338 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
bb17f655 | 3339 | { |
9d89c257 | 3340 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 3341 | |
aa0b7ae0 WL |
3342 | /* |
3343 | * No need to update load_avg for root_task_group as it is not used. | |
3344 | */ | |
3345 | if (cfs_rq->tg == &root_task_group) | |
3346 | return; | |
3347 | ||
9d89c257 YD |
3348 | if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
3349 | atomic_long_add(delta, &cfs_rq->tg->load_avg); | |
3350 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 3351 | } |
8165e145 | 3352 | } |
f5f9739d | 3353 | |
ad936d86 BP |
3354 | /* |
3355 | * Called within set_task_rq() right before setting a task's cpu. The | |
3356 | * caller only guarantees p->pi_lock is held; no other assumptions, | |
3357 | * including the state of rq->lock, should be made. | |
3358 | */ | |
3359 | void set_task_rq_fair(struct sched_entity *se, | |
3360 | struct cfs_rq *prev, struct cfs_rq *next) | |
3361 | { | |
0ccb977f PZ |
3362 | u64 p_last_update_time; |
3363 | u64 n_last_update_time; | |
3364 | ||
ad936d86 BP |
3365 | if (!sched_feat(ATTACH_AGE_LOAD)) |
3366 | return; | |
3367 | ||
3368 | /* | |
3369 | * We are supposed to update the task to "current" time, then its up to | |
3370 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
3371 | * getting what current time is, so simply throw away the out-of-date | |
3372 | * time. This will result in the wakee task is less decayed, but giving | |
3373 | * the wakee more load sounds not bad. | |
3374 | */ | |
0ccb977f PZ |
3375 | if (!(se->avg.last_update_time && prev)) |
3376 | return; | |
ad936d86 BP |
3377 | |
3378 | #ifndef CONFIG_64BIT | |
0ccb977f | 3379 | { |
ad936d86 BP |
3380 | u64 p_last_update_time_copy; |
3381 | u64 n_last_update_time_copy; | |
3382 | ||
3383 | do { | |
3384 | p_last_update_time_copy = prev->load_last_update_time_copy; | |
3385 | n_last_update_time_copy = next->load_last_update_time_copy; | |
3386 | ||
3387 | smp_rmb(); | |
3388 | ||
3389 | p_last_update_time = prev->avg.last_update_time; | |
3390 | n_last_update_time = next->avg.last_update_time; | |
3391 | ||
3392 | } while (p_last_update_time != p_last_update_time_copy || | |
3393 | n_last_update_time != n_last_update_time_copy); | |
0ccb977f | 3394 | } |
ad936d86 | 3395 | #else |
0ccb977f PZ |
3396 | p_last_update_time = prev->avg.last_update_time; |
3397 | n_last_update_time = next->avg.last_update_time; | |
ad936d86 | 3398 | #endif |
0ccb977f PZ |
3399 | __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se); |
3400 | se->avg.last_update_time = n_last_update_time; | |
ad936d86 | 3401 | } |
09a43ace | 3402 | |
0e2d2aaa PZ |
3403 | |
3404 | /* | |
3405 | * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to | |
3406 | * propagate its contribution. The key to this propagation is the invariant | |
3407 | * that for each group: | |
3408 | * | |
3409 | * ge->avg == grq->avg (1) | |
3410 | * | |
3411 | * _IFF_ we look at the pure running and runnable sums. Because they | |
3412 | * represent the very same entity, just at different points in the hierarchy. | |
3413 | * | |
3414 | * | |
3415 | * Per the above update_tg_cfs_util() is trivial (and still 'wrong') and | |
3416 | * simply copies the running sum over. | |
3417 | * | |
3418 | * However, update_tg_cfs_runnable() is more complex. So we have: | |
3419 | * | |
3420 | * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) | |
3421 | * | |
3422 | * And since, like util, the runnable part should be directly transferable, | |
3423 | * the following would _appear_ to be the straight forward approach: | |
3424 | * | |
3425 | * grq->avg.load_avg = grq->load.weight * grq->avg.running_avg (3) | |
3426 | * | |
3427 | * And per (1) we have: | |
3428 | * | |
3429 | * ge->avg.running_avg == grq->avg.running_avg | |
3430 | * | |
3431 | * Which gives: | |
3432 | * | |
3433 | * ge->load.weight * grq->avg.load_avg | |
3434 | * ge->avg.load_avg = ----------------------------------- (4) | |
3435 | * grq->load.weight | |
3436 | * | |
3437 | * Except that is wrong! | |
3438 | * | |
3439 | * Because while for entities historical weight is not important and we | |
3440 | * really only care about our future and therefore can consider a pure | |
3441 | * runnable sum, runqueues can NOT do this. | |
3442 | * | |
3443 | * We specifically want runqueues to have a load_avg that includes | |
3444 | * historical weights. Those represent the blocked load, the load we expect | |
3445 | * to (shortly) return to us. This only works by keeping the weights as | |
3446 | * integral part of the sum. We therefore cannot decompose as per (3). | |
3447 | * | |
3448 | * OK, so what then? | |
3449 | * | |
3450 | * | |
3451 | * Another way to look at things is: | |
3452 | * | |
3453 | * grq->avg.load_avg = \Sum se->avg.load_avg | |
3454 | * | |
3455 | * Therefore, per (2): | |
3456 | * | |
3457 | * grq->avg.load_avg = \Sum se->load.weight * se->avg.runnable_avg | |
3458 | * | |
3459 | * And the very thing we're propagating is a change in that sum (someone | |
3460 | * joined/left). So we can easily know the runnable change, which would be, per | |
3461 | * (2) the already tracked se->load_avg divided by the corresponding | |
3462 | * se->weight. | |
3463 | * | |
3464 | * Basically (4) but in differential form: | |
3465 | * | |
3466 | * d(runnable_avg) += se->avg.load_avg / se->load.weight | |
3467 | * (5) | |
3468 | * ge->avg.load_avg += ge->load.weight * d(runnable_avg) | |
3469 | */ | |
3470 | ||
09a43ace | 3471 | static inline void |
0e2d2aaa | 3472 | update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3473 | { |
09a43ace VG |
3474 | long delta = gcfs_rq->avg.util_avg - se->avg.util_avg; |
3475 | ||
3476 | /* Nothing to update */ | |
3477 | if (!delta) | |
3478 | return; | |
3479 | ||
3480 | /* Set new sched_entity's utilization */ | |
3481 | se->avg.util_avg = gcfs_rq->avg.util_avg; | |
3482 | se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX; | |
3483 | ||
3484 | /* Update parent cfs_rq utilization */ | |
3485 | add_positive(&cfs_rq->avg.util_avg, delta); | |
3486 | cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX; | |
3487 | } | |
3488 | ||
09a43ace | 3489 | static inline void |
0e2d2aaa | 3490 | update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) |
09a43ace | 3491 | { |
0e2d2aaa | 3492 | long runnable_sum = gcfs_rq->prop_runnable_sum; |
1ea6c46a PZ |
3493 | long runnable_load_avg, load_avg; |
3494 | s64 runnable_load_sum, load_sum; | |
09a43ace | 3495 | |
0e2d2aaa PZ |
3496 | if (!runnable_sum) |
3497 | return; | |
09a43ace | 3498 | |
0e2d2aaa | 3499 | gcfs_rq->prop_runnable_sum = 0; |
09a43ace | 3500 | |
0e2d2aaa PZ |
3501 | load_sum = (s64)se_weight(se) * runnable_sum; |
3502 | load_avg = div_s64(load_sum, LOAD_AVG_MAX); | |
09a43ace | 3503 | |
0e2d2aaa PZ |
3504 | add_positive(&se->avg.load_sum, runnable_sum); |
3505 | add_positive(&se->avg.load_avg, load_avg); | |
09a43ace | 3506 | |
0e2d2aaa PZ |
3507 | add_positive(&cfs_rq->avg.load_avg, load_avg); |
3508 | add_positive(&cfs_rq->avg.load_sum, load_sum); | |
09a43ace | 3509 | |
1ea6c46a PZ |
3510 | runnable_load_sum = (s64)se_runnable(se) * runnable_sum; |
3511 | runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX); | |
3512 | ||
3513 | add_positive(&se->avg.runnable_load_sum, runnable_sum); | |
3514 | add_positive(&se->avg.runnable_load_avg, runnable_load_avg); | |
3515 | ||
09a43ace | 3516 | if (se->on_rq) { |
1ea6c46a PZ |
3517 | add_positive(&cfs_rq->avg.runnable_load_avg, runnable_load_avg); |
3518 | add_positive(&cfs_rq->avg.runnable_load_sum, runnable_load_sum); | |
09a43ace VG |
3519 | } |
3520 | } | |
3521 | ||
0e2d2aaa | 3522 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) |
09a43ace | 3523 | { |
0e2d2aaa PZ |
3524 | cfs_rq->propagate = 1; |
3525 | cfs_rq->prop_runnable_sum += runnable_sum; | |
09a43ace VG |
3526 | } |
3527 | ||
3528 | /* Update task and its cfs_rq load average */ | |
3529 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3530 | { | |
0e2d2aaa | 3531 | struct cfs_rq *cfs_rq, *gcfs_rq; |
09a43ace VG |
3532 | |
3533 | if (entity_is_task(se)) | |
3534 | return 0; | |
3535 | ||
0e2d2aaa PZ |
3536 | gcfs_rq = group_cfs_rq(se); |
3537 | if (!gcfs_rq->propagate) | |
09a43ace VG |
3538 | return 0; |
3539 | ||
0e2d2aaa PZ |
3540 | gcfs_rq->propagate = 0; |
3541 | ||
09a43ace VG |
3542 | cfs_rq = cfs_rq_of(se); |
3543 | ||
0e2d2aaa | 3544 | add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); |
09a43ace | 3545 | |
0e2d2aaa PZ |
3546 | update_tg_cfs_util(cfs_rq, se, gcfs_rq); |
3547 | update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); | |
09a43ace VG |
3548 | |
3549 | return 1; | |
3550 | } | |
3551 | ||
bc427898 VG |
3552 | /* |
3553 | * Check if we need to update the load and the utilization of a blocked | |
3554 | * group_entity: | |
3555 | */ | |
3556 | static inline bool skip_blocked_update(struct sched_entity *se) | |
3557 | { | |
3558 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3559 | ||
3560 | /* | |
3561 | * If sched_entity still have not zero load or utilization, we have to | |
3562 | * decay it: | |
3563 | */ | |
3564 | if (se->avg.load_avg || se->avg.util_avg) | |
3565 | return false; | |
3566 | ||
3567 | /* | |
3568 | * If there is a pending propagation, we have to update the load and | |
3569 | * the utilization of the sched_entity: | |
3570 | */ | |
0e2d2aaa | 3571 | if (gcfs_rq->propagate) |
bc427898 VG |
3572 | return false; |
3573 | ||
3574 | /* | |
3575 | * Otherwise, the load and the utilization of the sched_entity is | |
3576 | * already zero and there is no pending propagation, so it will be a | |
3577 | * waste of time to try to decay it: | |
3578 | */ | |
3579 | return true; | |
3580 | } | |
3581 | ||
6e83125c | 3582 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
09a43ace | 3583 | |
9d89c257 | 3584 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {} |
09a43ace VG |
3585 | |
3586 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3587 | { | |
3588 | return 0; | |
3589 | } | |
3590 | ||
0e2d2aaa | 3591 | static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} |
09a43ace | 3592 | |
6e83125c | 3593 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 3594 | |
3d30544f PZ |
3595 | /** |
3596 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
3597 | * @now: current time, as per cfs_rq_clock_task() | |
3598 | * @cfs_rq: cfs_rq to update | |
3d30544f PZ |
3599 | * |
3600 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
3601 | * avg. The immediate corollary is that all (fair) tasks must be attached, see | |
3602 | * post_init_entity_util_avg(). | |
3603 | * | |
3604 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
3605 | * | |
7c3edd2c PZ |
3606 | * Returns true if the load decayed or we removed load. |
3607 | * | |
3608 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
3609 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 3610 | */ |
a2c6c91f | 3611 | static inline int |
3a123bbb | 3612 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) |
2dac754e | 3613 | { |
0e2d2aaa | 3614 | unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0; |
9d89c257 | 3615 | struct sched_avg *sa = &cfs_rq->avg; |
2a2f5d4e | 3616 | int decayed = 0; |
2dac754e | 3617 | |
2a2f5d4e PZ |
3618 | if (cfs_rq->removed.nr) { |
3619 | unsigned long r; | |
9a2dd585 | 3620 | u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib; |
2a2f5d4e PZ |
3621 | |
3622 | raw_spin_lock(&cfs_rq->removed.lock); | |
3623 | swap(cfs_rq->removed.util_avg, removed_util); | |
3624 | swap(cfs_rq->removed.load_avg, removed_load); | |
0e2d2aaa | 3625 | swap(cfs_rq->removed.runnable_sum, removed_runnable_sum); |
2a2f5d4e PZ |
3626 | cfs_rq->removed.nr = 0; |
3627 | raw_spin_unlock(&cfs_rq->removed.lock); | |
3628 | ||
2a2f5d4e | 3629 | r = removed_load; |
89741892 | 3630 | sub_positive(&sa->load_avg, r); |
9a2dd585 | 3631 | sub_positive(&sa->load_sum, r * divider); |
2dac754e | 3632 | |
2a2f5d4e | 3633 | r = removed_util; |
89741892 | 3634 | sub_positive(&sa->util_avg, r); |
9a2dd585 | 3635 | sub_positive(&sa->util_sum, r * divider); |
2a2f5d4e | 3636 | |
0e2d2aaa | 3637 | add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum); |
2a2f5d4e PZ |
3638 | |
3639 | decayed = 1; | |
9d89c257 | 3640 | } |
36ee28e4 | 3641 | |
2a2f5d4e | 3642 | decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq); |
36ee28e4 | 3643 | |
9d89c257 YD |
3644 | #ifndef CONFIG_64BIT |
3645 | smp_wmb(); | |
3646 | cfs_rq->load_last_update_time_copy = sa->last_update_time; | |
3647 | #endif | |
36ee28e4 | 3648 | |
2a2f5d4e | 3649 | if (decayed) |
a2c6c91f | 3650 | cfs_rq_util_change(cfs_rq); |
21e96f88 | 3651 | |
2a2f5d4e | 3652 | return decayed; |
21e96f88 SM |
3653 | } |
3654 | ||
3d30544f PZ |
3655 | /** |
3656 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
3657 | * @cfs_rq: cfs_rq to attach to | |
3658 | * @se: sched_entity to attach | |
3659 | * | |
3660 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3661 | * cfs_rq->avg.last_update_time being current. | |
3662 | */ | |
a05e8c51 BP |
3663 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3664 | { | |
f207934f PZ |
3665 | u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib; |
3666 | ||
3667 | /* | |
3668 | * When we attach the @se to the @cfs_rq, we must align the decay | |
3669 | * window because without that, really weird and wonderful things can | |
3670 | * happen. | |
3671 | * | |
3672 | * XXX illustrate | |
3673 | */ | |
a05e8c51 | 3674 | se->avg.last_update_time = cfs_rq->avg.last_update_time; |
f207934f PZ |
3675 | se->avg.period_contrib = cfs_rq->avg.period_contrib; |
3676 | ||
3677 | /* | |
3678 | * Hell(o) Nasty stuff.. we need to recompute _sum based on the new | |
3679 | * period_contrib. This isn't strictly correct, but since we're | |
3680 | * entirely outside of the PELT hierarchy, nobody cares if we truncate | |
3681 | * _sum a little. | |
3682 | */ | |
3683 | se->avg.util_sum = se->avg.util_avg * divider; | |
3684 | ||
3685 | se->avg.load_sum = divider; | |
3686 | if (se_weight(se)) { | |
3687 | se->avg.load_sum = | |
3688 | div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se)); | |
3689 | } | |
3690 | ||
3691 | se->avg.runnable_load_sum = se->avg.load_sum; | |
3692 | ||
8d5b9025 | 3693 | enqueue_load_avg(cfs_rq, se); |
a05e8c51 BP |
3694 | cfs_rq->avg.util_avg += se->avg.util_avg; |
3695 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
0e2d2aaa PZ |
3696 | |
3697 | add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); | |
a2c6c91f SM |
3698 | |
3699 | cfs_rq_util_change(cfs_rq); | |
a05e8c51 BP |
3700 | } |
3701 | ||
3d30544f PZ |
3702 | /** |
3703 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
3704 | * @cfs_rq: cfs_rq to detach from | |
3705 | * @se: sched_entity to detach | |
3706 | * | |
3707 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3708 | * cfs_rq->avg.last_update_time being current. | |
3709 | */ | |
a05e8c51 BP |
3710 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3711 | { | |
8d5b9025 | 3712 | dequeue_load_avg(cfs_rq, se); |
89741892 PZ |
3713 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); |
3714 | sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); | |
0e2d2aaa PZ |
3715 | |
3716 | add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); | |
a2c6c91f SM |
3717 | |
3718 | cfs_rq_util_change(cfs_rq); | |
a05e8c51 BP |
3719 | } |
3720 | ||
b382a531 PZ |
3721 | /* |
3722 | * Optional action to be done while updating the load average | |
3723 | */ | |
3724 | #define UPDATE_TG 0x1 | |
3725 | #define SKIP_AGE_LOAD 0x2 | |
3726 | #define DO_ATTACH 0x4 | |
3727 | ||
3728 | /* Update task and its cfs_rq load average */ | |
3729 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) | |
3730 | { | |
3731 | u64 now = cfs_rq_clock_task(cfs_rq); | |
3732 | struct rq *rq = rq_of(cfs_rq); | |
3733 | int cpu = cpu_of(rq); | |
3734 | int decayed; | |
3735 | ||
3736 | /* | |
3737 | * Track task load average for carrying it to new CPU after migrated, and | |
3738 | * track group sched_entity load average for task_h_load calc in migration | |
3739 | */ | |
3740 | if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) | |
3741 | __update_load_avg_se(now, cpu, cfs_rq, se); | |
3742 | ||
3743 | decayed = update_cfs_rq_load_avg(now, cfs_rq); | |
3744 | decayed |= propagate_entity_load_avg(se); | |
3745 | ||
3746 | if (!se->avg.last_update_time && (flags & DO_ATTACH)) { | |
3747 | ||
3748 | attach_entity_load_avg(cfs_rq, se); | |
3749 | update_tg_load_avg(cfs_rq, 0); | |
3750 | ||
3751 | } else if (decayed && (flags & UPDATE_TG)) | |
3752 | update_tg_load_avg(cfs_rq, 0); | |
3753 | } | |
3754 | ||
9d89c257 | 3755 | #ifndef CONFIG_64BIT |
0905f04e YD |
3756 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3757 | { | |
9d89c257 | 3758 | u64 last_update_time_copy; |
0905f04e | 3759 | u64 last_update_time; |
9ee474f5 | 3760 | |
9d89c257 YD |
3761 | do { |
3762 | last_update_time_copy = cfs_rq->load_last_update_time_copy; | |
3763 | smp_rmb(); | |
3764 | last_update_time = cfs_rq->avg.last_update_time; | |
3765 | } while (last_update_time != last_update_time_copy); | |
0905f04e YD |
3766 | |
3767 | return last_update_time; | |
3768 | } | |
9d89c257 | 3769 | #else |
0905f04e YD |
3770 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3771 | { | |
3772 | return cfs_rq->avg.last_update_time; | |
3773 | } | |
9d89c257 YD |
3774 | #endif |
3775 | ||
104cb16d MR |
3776 | /* |
3777 | * Synchronize entity load avg of dequeued entity without locking | |
3778 | * the previous rq. | |
3779 | */ | |
3780 | void sync_entity_load_avg(struct sched_entity *se) | |
3781 | { | |
3782 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3783 | u64 last_update_time; | |
3784 | ||
3785 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
0ccb977f | 3786 | __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se); |
104cb16d MR |
3787 | } |
3788 | ||
0905f04e YD |
3789 | /* |
3790 | * Task first catches up with cfs_rq, and then subtract | |
3791 | * itself from the cfs_rq (task must be off the queue now). | |
3792 | */ | |
3793 | void remove_entity_load_avg(struct sched_entity *se) | |
3794 | { | |
3795 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
2a2f5d4e | 3796 | unsigned long flags; |
0905f04e YD |
3797 | |
3798 | /* | |
7dc603c9 PZ |
3799 | * tasks cannot exit without having gone through wake_up_new_task() -> |
3800 | * post_init_entity_util_avg() which will have added things to the | |
3801 | * cfs_rq, so we can remove unconditionally. | |
3802 | * | |
3803 | * Similarly for groups, they will have passed through | |
3804 | * post_init_entity_util_avg() before unregister_sched_fair_group() | |
3805 | * calls this. | |
0905f04e | 3806 | */ |
0905f04e | 3807 | |
104cb16d | 3808 | sync_entity_load_avg(se); |
2a2f5d4e PZ |
3809 | |
3810 | raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); | |
3811 | ++cfs_rq->removed.nr; | |
3812 | cfs_rq->removed.util_avg += se->avg.util_avg; | |
3813 | cfs_rq->removed.load_avg += se->avg.load_avg; | |
0e2d2aaa | 3814 | cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */ |
2a2f5d4e | 3815 | raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); |
2dac754e | 3816 | } |
642dbc39 | 3817 | |
7ea241af YD |
3818 | static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq) |
3819 | { | |
1ea6c46a | 3820 | return cfs_rq->avg.runnable_load_avg; |
7ea241af YD |
3821 | } |
3822 | ||
3823 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) | |
3824 | { | |
3825 | return cfs_rq->avg.load_avg; | |
3826 | } | |
3827 | ||
46f69fa3 | 3828 | static int idle_balance(struct rq *this_rq, struct rq_flags *rf); |
6e83125c | 3829 | |
38033c37 PZ |
3830 | #else /* CONFIG_SMP */ |
3831 | ||
01011473 | 3832 | static inline int |
3a123bbb | 3833 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) |
01011473 PZ |
3834 | { |
3835 | return 0; | |
3836 | } | |
3837 | ||
d31b1a66 VG |
3838 | #define UPDATE_TG 0x0 |
3839 | #define SKIP_AGE_LOAD 0x0 | |
b382a531 | 3840 | #define DO_ATTACH 0x0 |
d31b1a66 | 3841 | |
88c0616e | 3842 | static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) |
536bd00c | 3843 | { |
88c0616e | 3844 | cfs_rq_util_change(cfs_rq); |
536bd00c RW |
3845 | } |
3846 | ||
9d89c257 | 3847 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 3848 | |
a05e8c51 BP |
3849 | static inline void |
3850 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
3851 | static inline void | |
3852 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
3853 | ||
46f69fa3 | 3854 | static inline int idle_balance(struct rq *rq, struct rq_flags *rf) |
6e83125c PZ |
3855 | { |
3856 | return 0; | |
3857 | } | |
3858 | ||
38033c37 | 3859 | #endif /* CONFIG_SMP */ |
9d85f21c | 3860 | |
ddc97297 PZ |
3861 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3862 | { | |
3863 | #ifdef CONFIG_SCHED_DEBUG | |
3864 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
3865 | ||
3866 | if (d < 0) | |
3867 | d = -d; | |
3868 | ||
3869 | if (d > 3*sysctl_sched_latency) | |
ae92882e | 3870 | schedstat_inc(cfs_rq->nr_spread_over); |
ddc97297 PZ |
3871 | #endif |
3872 | } | |
3873 | ||
aeb73b04 PZ |
3874 | static void |
3875 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
3876 | { | |
1af5f730 | 3877 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 3878 | |
2cb8600e PZ |
3879 | /* |
3880 | * The 'current' period is already promised to the current tasks, | |
3881 | * however the extra weight of the new task will slow them down a | |
3882 | * little, place the new task so that it fits in the slot that | |
3883 | * stays open at the end. | |
3884 | */ | |
94dfb5e7 | 3885 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 3886 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 3887 | |
a2e7a7eb | 3888 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 3889 | if (!initial) { |
a2e7a7eb | 3890 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 3891 | |
a2e7a7eb MG |
3892 | /* |
3893 | * Halve their sleep time's effect, to allow | |
3894 | * for a gentler effect of sleepers: | |
3895 | */ | |
3896 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
3897 | thresh >>= 1; | |
51e0304c | 3898 | |
a2e7a7eb | 3899 | vruntime -= thresh; |
aeb73b04 PZ |
3900 | } |
3901 | ||
b5d9d734 | 3902 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 3903 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
3904 | } |
3905 | ||
d3d9dc33 PT |
3906 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
3907 | ||
cb251765 MG |
3908 | static inline void check_schedstat_required(void) |
3909 | { | |
3910 | #ifdef CONFIG_SCHEDSTATS | |
3911 | if (schedstat_enabled()) | |
3912 | return; | |
3913 | ||
3914 | /* Force schedstat enabled if a dependent tracepoint is active */ | |
3915 | if (trace_sched_stat_wait_enabled() || | |
3916 | trace_sched_stat_sleep_enabled() || | |
3917 | trace_sched_stat_iowait_enabled() || | |
3918 | trace_sched_stat_blocked_enabled() || | |
3919 | trace_sched_stat_runtime_enabled()) { | |
eda8dca5 | 3920 | printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " |
cb251765 | 3921 | "stat_blocked and stat_runtime require the " |
f67abed5 | 3922 | "kernel parameter schedstats=enable or " |
cb251765 MG |
3923 | "kernel.sched_schedstats=1\n"); |
3924 | } | |
3925 | #endif | |
3926 | } | |
3927 | ||
b5179ac7 PZ |
3928 | |
3929 | /* | |
3930 | * MIGRATION | |
3931 | * | |
3932 | * dequeue | |
3933 | * update_curr() | |
3934 | * update_min_vruntime() | |
3935 | * vruntime -= min_vruntime | |
3936 | * | |
3937 | * enqueue | |
3938 | * update_curr() | |
3939 | * update_min_vruntime() | |
3940 | * vruntime += min_vruntime | |
3941 | * | |
3942 | * this way the vruntime transition between RQs is done when both | |
3943 | * min_vruntime are up-to-date. | |
3944 | * | |
3945 | * WAKEUP (remote) | |
3946 | * | |
59efa0ba | 3947 | * ->migrate_task_rq_fair() (p->state == TASK_WAKING) |
b5179ac7 PZ |
3948 | * vruntime -= min_vruntime |
3949 | * | |
3950 | * enqueue | |
3951 | * update_curr() | |
3952 | * update_min_vruntime() | |
3953 | * vruntime += min_vruntime | |
3954 | * | |
3955 | * this way we don't have the most up-to-date min_vruntime on the originating | |
3956 | * CPU and an up-to-date min_vruntime on the destination CPU. | |
3957 | */ | |
3958 | ||
bf0f6f24 | 3959 | static void |
88ec22d3 | 3960 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3961 | { |
2f950354 PZ |
3962 | bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); |
3963 | bool curr = cfs_rq->curr == se; | |
3964 | ||
88ec22d3 | 3965 | /* |
2f950354 PZ |
3966 | * If we're the current task, we must renormalise before calling |
3967 | * update_curr(). | |
88ec22d3 | 3968 | */ |
2f950354 | 3969 | if (renorm && curr) |
88ec22d3 PZ |
3970 | se->vruntime += cfs_rq->min_vruntime; |
3971 | ||
2f950354 PZ |
3972 | update_curr(cfs_rq); |
3973 | ||
bf0f6f24 | 3974 | /* |
2f950354 PZ |
3975 | * Otherwise, renormalise after, such that we're placed at the current |
3976 | * moment in time, instead of some random moment in the past. Being | |
3977 | * placed in the past could significantly boost this task to the | |
3978 | * fairness detriment of existing tasks. | |
bf0f6f24 | 3979 | */ |
2f950354 PZ |
3980 | if (renorm && !curr) |
3981 | se->vruntime += cfs_rq->min_vruntime; | |
3982 | ||
89ee048f VG |
3983 | /* |
3984 | * When enqueuing a sched_entity, we must: | |
3985 | * - Update loads to have both entity and cfs_rq synced with now. | |
3986 | * - Add its load to cfs_rq->runnable_avg | |
3987 | * - For group_entity, update its weight to reflect the new share of | |
3988 | * its group cfs_rq | |
3989 | * - Add its new weight to cfs_rq->load.weight | |
3990 | */ | |
b382a531 | 3991 | update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); |
1ea6c46a | 3992 | update_cfs_group(se); |
b5b3e35f | 3993 | enqueue_runnable_load_avg(cfs_rq, se); |
17bc14b7 | 3994 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 3995 | |
1a3d027c | 3996 | if (flags & ENQUEUE_WAKEUP) |
aeb73b04 | 3997 | place_entity(cfs_rq, se, 0); |
bf0f6f24 | 3998 | |
cb251765 | 3999 | check_schedstat_required(); |
4fa8d299 JP |
4000 | update_stats_enqueue(cfs_rq, se, flags); |
4001 | check_spread(cfs_rq, se); | |
2f950354 | 4002 | if (!curr) |
83b699ed | 4003 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 4004 | se->on_rq = 1; |
3d4b47b4 | 4005 | |
d3d9dc33 | 4006 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 4007 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
4008 | check_enqueue_throttle(cfs_rq); |
4009 | } | |
bf0f6f24 IM |
4010 | } |
4011 | ||
2c13c919 | 4012 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 4013 | { |
2c13c919 RR |
4014 | for_each_sched_entity(se) { |
4015 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4016 | if (cfs_rq->last != se) |
2c13c919 | 4017 | break; |
f1044799 PZ |
4018 | |
4019 | cfs_rq->last = NULL; | |
2c13c919 RR |
4020 | } |
4021 | } | |
2002c695 | 4022 | |
2c13c919 RR |
4023 | static void __clear_buddies_next(struct sched_entity *se) |
4024 | { | |
4025 | for_each_sched_entity(se) { | |
4026 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4027 | if (cfs_rq->next != se) |
2c13c919 | 4028 | break; |
f1044799 PZ |
4029 | |
4030 | cfs_rq->next = NULL; | |
2c13c919 | 4031 | } |
2002c695 PZ |
4032 | } |
4033 | ||
ac53db59 RR |
4034 | static void __clear_buddies_skip(struct sched_entity *se) |
4035 | { | |
4036 | for_each_sched_entity(se) { | |
4037 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 4038 | if (cfs_rq->skip != se) |
ac53db59 | 4039 | break; |
f1044799 PZ |
4040 | |
4041 | cfs_rq->skip = NULL; | |
ac53db59 RR |
4042 | } |
4043 | } | |
4044 | ||
a571bbea PZ |
4045 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
4046 | { | |
2c13c919 RR |
4047 | if (cfs_rq->last == se) |
4048 | __clear_buddies_last(se); | |
4049 | ||
4050 | if (cfs_rq->next == se) | |
4051 | __clear_buddies_next(se); | |
ac53db59 RR |
4052 | |
4053 | if (cfs_rq->skip == se) | |
4054 | __clear_buddies_skip(se); | |
a571bbea PZ |
4055 | } |
4056 | ||
6c16a6dc | 4057 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 4058 | |
bf0f6f24 | 4059 | static void |
371fd7e7 | 4060 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 4061 | { |
a2a2d680 DA |
4062 | /* |
4063 | * Update run-time statistics of the 'current'. | |
4064 | */ | |
4065 | update_curr(cfs_rq); | |
89ee048f VG |
4066 | |
4067 | /* | |
4068 | * When dequeuing a sched_entity, we must: | |
4069 | * - Update loads to have both entity and cfs_rq synced with now. | |
4070 | * - Substract its load from the cfs_rq->runnable_avg. | |
4071 | * - Substract its previous weight from cfs_rq->load.weight. | |
4072 | * - For group entity, update its weight to reflect the new share | |
4073 | * of its group cfs_rq. | |
4074 | */ | |
88c0616e | 4075 | update_load_avg(cfs_rq, se, UPDATE_TG); |
b5b3e35f | 4076 | dequeue_runnable_load_avg(cfs_rq, se); |
a2a2d680 | 4077 | |
4fa8d299 | 4078 | update_stats_dequeue(cfs_rq, se, flags); |
67e9fb2a | 4079 | |
2002c695 | 4080 | clear_buddies(cfs_rq, se); |
4793241b | 4081 | |
83b699ed | 4082 | if (se != cfs_rq->curr) |
30cfdcfc | 4083 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 4084 | se->on_rq = 0; |
30cfdcfc | 4085 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
4086 | |
4087 | /* | |
b60205c7 PZ |
4088 | * Normalize after update_curr(); which will also have moved |
4089 | * min_vruntime if @se is the one holding it back. But before doing | |
4090 | * update_min_vruntime() again, which will discount @se's position and | |
4091 | * can move min_vruntime forward still more. | |
88ec22d3 | 4092 | */ |
371fd7e7 | 4093 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 4094 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 4095 | |
d8b4986d PT |
4096 | /* return excess runtime on last dequeue */ |
4097 | return_cfs_rq_runtime(cfs_rq); | |
4098 | ||
1ea6c46a | 4099 | update_cfs_group(se); |
b60205c7 PZ |
4100 | |
4101 | /* | |
4102 | * Now advance min_vruntime if @se was the entity holding it back, | |
4103 | * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be | |
4104 | * put back on, and if we advance min_vruntime, we'll be placed back | |
4105 | * further than we started -- ie. we'll be penalized. | |
4106 | */ | |
4107 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE) | |
4108 | update_min_vruntime(cfs_rq); | |
bf0f6f24 IM |
4109 | } |
4110 | ||
4111 | /* | |
4112 | * Preempt the current task with a newly woken task if needed: | |
4113 | */ | |
7c92e54f | 4114 | static void |
2e09bf55 | 4115 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 4116 | { |
11697830 | 4117 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
4118 | struct sched_entity *se; |
4119 | s64 delta; | |
11697830 | 4120 | |
6d0f0ebd | 4121 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 4122 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 4123 | if (delta_exec > ideal_runtime) { |
8875125e | 4124 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
4125 | /* |
4126 | * The current task ran long enough, ensure it doesn't get | |
4127 | * re-elected due to buddy favours. | |
4128 | */ | |
4129 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
4130 | return; |
4131 | } | |
4132 | ||
4133 | /* | |
4134 | * Ensure that a task that missed wakeup preemption by a | |
4135 | * narrow margin doesn't have to wait for a full slice. | |
4136 | * This also mitigates buddy induced latencies under load. | |
4137 | */ | |
f685ceac MG |
4138 | if (delta_exec < sysctl_sched_min_granularity) |
4139 | return; | |
4140 | ||
f4cfb33e WX |
4141 | se = __pick_first_entity(cfs_rq); |
4142 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 4143 | |
f4cfb33e WX |
4144 | if (delta < 0) |
4145 | return; | |
d7d82944 | 4146 | |
f4cfb33e | 4147 | if (delta > ideal_runtime) |
8875125e | 4148 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
4149 | } |
4150 | ||
83b699ed | 4151 | static void |
8494f412 | 4152 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 4153 | { |
83b699ed SV |
4154 | /* 'current' is not kept within the tree. */ |
4155 | if (se->on_rq) { | |
4156 | /* | |
4157 | * Any task has to be enqueued before it get to execute on | |
4158 | * a CPU. So account for the time it spent waiting on the | |
4159 | * runqueue. | |
4160 | */ | |
4fa8d299 | 4161 | update_stats_wait_end(cfs_rq, se); |
83b699ed | 4162 | __dequeue_entity(cfs_rq, se); |
88c0616e | 4163 | update_load_avg(cfs_rq, se, UPDATE_TG); |
83b699ed SV |
4164 | } |
4165 | ||
79303e9e | 4166 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 4167 | cfs_rq->curr = se; |
4fa8d299 | 4168 | |
eba1ed4b IM |
4169 | /* |
4170 | * Track our maximum slice length, if the CPU's load is at | |
4171 | * least twice that of our own weight (i.e. dont track it | |
4172 | * when there are only lesser-weight tasks around): | |
4173 | */ | |
cb251765 | 4174 | if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
4fa8d299 JP |
4175 | schedstat_set(se->statistics.slice_max, |
4176 | max((u64)schedstat_val(se->statistics.slice_max), | |
4177 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); | |
eba1ed4b | 4178 | } |
4fa8d299 | 4179 | |
4a55b450 | 4180 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
4181 | } |
4182 | ||
3f3a4904 PZ |
4183 | static int |
4184 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
4185 | ||
ac53db59 RR |
4186 | /* |
4187 | * Pick the next process, keeping these things in mind, in this order: | |
4188 | * 1) keep things fair between processes/task groups | |
4189 | * 2) pick the "next" process, since someone really wants that to run | |
4190 | * 3) pick the "last" process, for cache locality | |
4191 | * 4) do not run the "skip" process, if something else is available | |
4192 | */ | |
678d5718 PZ |
4193 | static struct sched_entity * |
4194 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 4195 | { |
678d5718 PZ |
4196 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
4197 | struct sched_entity *se; | |
4198 | ||
4199 | /* | |
4200 | * If curr is set we have to see if its left of the leftmost entity | |
4201 | * still in the tree, provided there was anything in the tree at all. | |
4202 | */ | |
4203 | if (!left || (curr && entity_before(curr, left))) | |
4204 | left = curr; | |
4205 | ||
4206 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 4207 | |
ac53db59 RR |
4208 | /* |
4209 | * Avoid running the skip buddy, if running something else can | |
4210 | * be done without getting too unfair. | |
4211 | */ | |
4212 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
4213 | struct sched_entity *second; |
4214 | ||
4215 | if (se == curr) { | |
4216 | second = __pick_first_entity(cfs_rq); | |
4217 | } else { | |
4218 | second = __pick_next_entity(se); | |
4219 | if (!second || (curr && entity_before(curr, second))) | |
4220 | second = curr; | |
4221 | } | |
4222 | ||
ac53db59 RR |
4223 | if (second && wakeup_preempt_entity(second, left) < 1) |
4224 | se = second; | |
4225 | } | |
aa2ac252 | 4226 | |
f685ceac MG |
4227 | /* |
4228 | * Prefer last buddy, try to return the CPU to a preempted task. | |
4229 | */ | |
4230 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
4231 | se = cfs_rq->last; | |
4232 | ||
ac53db59 RR |
4233 | /* |
4234 | * Someone really wants this to run. If it's not unfair, run it. | |
4235 | */ | |
4236 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
4237 | se = cfs_rq->next; | |
4238 | ||
f685ceac | 4239 | clear_buddies(cfs_rq, se); |
4793241b PZ |
4240 | |
4241 | return se; | |
aa2ac252 PZ |
4242 | } |
4243 | ||
678d5718 | 4244 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 4245 | |
ab6cde26 | 4246 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
4247 | { |
4248 | /* | |
4249 | * If still on the runqueue then deactivate_task() | |
4250 | * was not called and update_curr() has to be done: | |
4251 | */ | |
4252 | if (prev->on_rq) | |
b7cc0896 | 4253 | update_curr(cfs_rq); |
bf0f6f24 | 4254 | |
d3d9dc33 PT |
4255 | /* throttle cfs_rqs exceeding runtime */ |
4256 | check_cfs_rq_runtime(cfs_rq); | |
4257 | ||
4fa8d299 | 4258 | check_spread(cfs_rq, prev); |
cb251765 | 4259 | |
30cfdcfc | 4260 | if (prev->on_rq) { |
4fa8d299 | 4261 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
4262 | /* Put 'current' back into the tree. */ |
4263 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 4264 | /* in !on_rq case, update occurred at dequeue */ |
88c0616e | 4265 | update_load_avg(cfs_rq, prev, 0); |
30cfdcfc | 4266 | } |
429d43bc | 4267 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
4268 | } |
4269 | ||
8f4d37ec PZ |
4270 | static void |
4271 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 4272 | { |
bf0f6f24 | 4273 | /* |
30cfdcfc | 4274 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 4275 | */ |
30cfdcfc | 4276 | update_curr(cfs_rq); |
bf0f6f24 | 4277 | |
9d85f21c PT |
4278 | /* |
4279 | * Ensure that runnable average is periodically updated. | |
4280 | */ | |
88c0616e | 4281 | update_load_avg(cfs_rq, curr, UPDATE_TG); |
1ea6c46a | 4282 | update_cfs_group(curr); |
9d85f21c | 4283 | |
8f4d37ec PZ |
4284 | #ifdef CONFIG_SCHED_HRTICK |
4285 | /* | |
4286 | * queued ticks are scheduled to match the slice, so don't bother | |
4287 | * validating it and just reschedule. | |
4288 | */ | |
983ed7a6 | 4289 | if (queued) { |
8875125e | 4290 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
4291 | return; |
4292 | } | |
8f4d37ec PZ |
4293 | /* |
4294 | * don't let the period tick interfere with the hrtick preemption | |
4295 | */ | |
4296 | if (!sched_feat(DOUBLE_TICK) && | |
4297 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
4298 | return; | |
4299 | #endif | |
4300 | ||
2c2efaed | 4301 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 4302 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
4303 | } |
4304 | ||
ab84d31e PT |
4305 | |
4306 | /************************************************** | |
4307 | * CFS bandwidth control machinery | |
4308 | */ | |
4309 | ||
4310 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
4311 | |
4312 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 4313 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
4314 | |
4315 | static inline bool cfs_bandwidth_used(void) | |
4316 | { | |
c5905afb | 4317 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
4318 | } |
4319 | ||
1ee14e6c | 4320 | void cfs_bandwidth_usage_inc(void) |
029632fb | 4321 | { |
1ee14e6c BS |
4322 | static_key_slow_inc(&__cfs_bandwidth_used); |
4323 | } | |
4324 | ||
4325 | void cfs_bandwidth_usage_dec(void) | |
4326 | { | |
4327 | static_key_slow_dec(&__cfs_bandwidth_used); | |
029632fb PZ |
4328 | } |
4329 | #else /* HAVE_JUMP_LABEL */ | |
4330 | static bool cfs_bandwidth_used(void) | |
4331 | { | |
4332 | return true; | |
4333 | } | |
4334 | ||
1ee14e6c BS |
4335 | void cfs_bandwidth_usage_inc(void) {} |
4336 | void cfs_bandwidth_usage_dec(void) {} | |
029632fb PZ |
4337 | #endif /* HAVE_JUMP_LABEL */ |
4338 | ||
ab84d31e PT |
4339 | /* |
4340 | * default period for cfs group bandwidth. | |
4341 | * default: 0.1s, units: nanoseconds | |
4342 | */ | |
4343 | static inline u64 default_cfs_period(void) | |
4344 | { | |
4345 | return 100000000ULL; | |
4346 | } | |
ec12cb7f PT |
4347 | |
4348 | static inline u64 sched_cfs_bandwidth_slice(void) | |
4349 | { | |
4350 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
4351 | } | |
4352 | ||
a9cf55b2 PT |
4353 | /* |
4354 | * Replenish runtime according to assigned quota and update expiration time. | |
4355 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
4356 | * additional synchronization around rq->lock. | |
4357 | * | |
4358 | * requires cfs_b->lock | |
4359 | */ | |
029632fb | 4360 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
4361 | { |
4362 | u64 now; | |
4363 | ||
4364 | if (cfs_b->quota == RUNTIME_INF) | |
4365 | return; | |
4366 | ||
4367 | now = sched_clock_cpu(smp_processor_id()); | |
4368 | cfs_b->runtime = cfs_b->quota; | |
4369 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
4370 | } | |
4371 | ||
029632fb PZ |
4372 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4373 | { | |
4374 | return &tg->cfs_bandwidth; | |
4375 | } | |
4376 | ||
f1b17280 PT |
4377 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
4378 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
4379 | { | |
4380 | if (unlikely(cfs_rq->throttle_count)) | |
1a99ae3f | 4381 | return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time; |
f1b17280 | 4382 | |
78becc27 | 4383 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
4384 | } |
4385 | ||
85dac906 PT |
4386 | /* returns 0 on failure to allocate runtime */ |
4387 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
4388 | { |
4389 | struct task_group *tg = cfs_rq->tg; | |
4390 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 4391 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
4392 | |
4393 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
4394 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
4395 | ||
4396 | raw_spin_lock(&cfs_b->lock); | |
4397 | if (cfs_b->quota == RUNTIME_INF) | |
4398 | amount = min_amount; | |
58088ad0 | 4399 | else { |
77a4d1a1 | 4400 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
4401 | |
4402 | if (cfs_b->runtime > 0) { | |
4403 | amount = min(cfs_b->runtime, min_amount); | |
4404 | cfs_b->runtime -= amount; | |
4405 | cfs_b->idle = 0; | |
4406 | } | |
ec12cb7f | 4407 | } |
a9cf55b2 | 4408 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
4409 | raw_spin_unlock(&cfs_b->lock); |
4410 | ||
4411 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
4412 | /* |
4413 | * we may have advanced our local expiration to account for allowed | |
4414 | * spread between our sched_clock and the one on which runtime was | |
4415 | * issued. | |
4416 | */ | |
4417 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
4418 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
4419 | |
4420 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
4421 | } |
4422 | ||
a9cf55b2 PT |
4423 | /* |
4424 | * Note: This depends on the synchronization provided by sched_clock and the | |
4425 | * fact that rq->clock snapshots this value. | |
4426 | */ | |
4427 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 4428 | { |
a9cf55b2 | 4429 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
4430 | |
4431 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 4432 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
4433 | return; |
4434 | ||
a9cf55b2 PT |
4435 | if (cfs_rq->runtime_remaining < 0) |
4436 | return; | |
4437 | ||
4438 | /* | |
4439 | * If the local deadline has passed we have to consider the | |
4440 | * possibility that our sched_clock is 'fast' and the global deadline | |
4441 | * has not truly expired. | |
4442 | * | |
4443 | * Fortunately we can check determine whether this the case by checking | |
51f2176d BS |
4444 | * whether the global deadline has advanced. It is valid to compare |
4445 | * cfs_b->runtime_expires without any locks since we only care about | |
4446 | * exact equality, so a partial write will still work. | |
a9cf55b2 PT |
4447 | */ |
4448 | ||
51f2176d | 4449 | if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { |
a9cf55b2 PT |
4450 | /* extend local deadline, drift is bounded above by 2 ticks */ |
4451 | cfs_rq->runtime_expires += TICK_NSEC; | |
4452 | } else { | |
4453 | /* global deadline is ahead, expiration has passed */ | |
4454 | cfs_rq->runtime_remaining = 0; | |
4455 | } | |
4456 | } | |
4457 | ||
9dbdb155 | 4458 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
4459 | { |
4460 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 4461 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
4462 | expire_cfs_rq_runtime(cfs_rq); |
4463 | ||
4464 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
4465 | return; |
4466 | ||
85dac906 PT |
4467 | /* |
4468 | * if we're unable to extend our runtime we resched so that the active | |
4469 | * hierarchy can be throttled | |
4470 | */ | |
4471 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 4472 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
4473 | } |
4474 | ||
6c16a6dc | 4475 | static __always_inline |
9dbdb155 | 4476 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 4477 | { |
56f570e5 | 4478 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
4479 | return; |
4480 | ||
4481 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
4482 | } | |
4483 | ||
85dac906 PT |
4484 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
4485 | { | |
56f570e5 | 4486 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
4487 | } |
4488 | ||
64660c86 PT |
4489 | /* check whether cfs_rq, or any parent, is throttled */ |
4490 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4491 | { | |
56f570e5 | 4492 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
4493 | } |
4494 | ||
4495 | /* | |
4496 | * Ensure that neither of the group entities corresponding to src_cpu or | |
4497 | * dest_cpu are members of a throttled hierarchy when performing group | |
4498 | * load-balance operations. | |
4499 | */ | |
4500 | static inline int throttled_lb_pair(struct task_group *tg, | |
4501 | int src_cpu, int dest_cpu) | |
4502 | { | |
4503 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
4504 | ||
4505 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
4506 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
4507 | ||
4508 | return throttled_hierarchy(src_cfs_rq) || | |
4509 | throttled_hierarchy(dest_cfs_rq); | |
4510 | } | |
4511 | ||
4512 | /* updated child weight may affect parent so we have to do this bottom up */ | |
4513 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
4514 | { | |
4515 | struct rq *rq = data; | |
4516 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4517 | ||
4518 | cfs_rq->throttle_count--; | |
64660c86 | 4519 | if (!cfs_rq->throttle_count) { |
f1b17280 | 4520 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 4521 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 4522 | cfs_rq->throttled_clock_task; |
64660c86 | 4523 | } |
64660c86 PT |
4524 | |
4525 | return 0; | |
4526 | } | |
4527 | ||
4528 | static int tg_throttle_down(struct task_group *tg, void *data) | |
4529 | { | |
4530 | struct rq *rq = data; | |
4531 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4532 | ||
82958366 PT |
4533 | /* group is entering throttled state, stop time */ |
4534 | if (!cfs_rq->throttle_count) | |
78becc27 | 4535 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
4536 | cfs_rq->throttle_count++; |
4537 | ||
4538 | return 0; | |
4539 | } | |
4540 | ||
d3d9dc33 | 4541 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
4542 | { |
4543 | struct rq *rq = rq_of(cfs_rq); | |
4544 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4545 | struct sched_entity *se; | |
4546 | long task_delta, dequeue = 1; | |
77a4d1a1 | 4547 | bool empty; |
85dac906 PT |
4548 | |
4549 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
4550 | ||
f1b17280 | 4551 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
4552 | rcu_read_lock(); |
4553 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
4554 | rcu_read_unlock(); | |
85dac906 PT |
4555 | |
4556 | task_delta = cfs_rq->h_nr_running; | |
4557 | for_each_sched_entity(se) { | |
4558 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
4559 | /* throttled entity or throttle-on-deactivate */ | |
4560 | if (!se->on_rq) | |
4561 | break; | |
4562 | ||
4563 | if (dequeue) | |
4564 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
4565 | qcfs_rq->h_nr_running -= task_delta; | |
4566 | ||
4567 | if (qcfs_rq->load.weight) | |
4568 | dequeue = 0; | |
4569 | } | |
4570 | ||
4571 | if (!se) | |
72465447 | 4572 | sub_nr_running(rq, task_delta); |
85dac906 PT |
4573 | |
4574 | cfs_rq->throttled = 1; | |
78becc27 | 4575 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 | 4576 | raw_spin_lock(&cfs_b->lock); |
d49db342 | 4577 | empty = list_empty(&cfs_b->throttled_cfs_rq); |
77a4d1a1 | 4578 | |
c06f04c7 BS |
4579 | /* |
4580 | * Add to the _head_ of the list, so that an already-started | |
4581 | * distribute_cfs_runtime will not see us | |
4582 | */ | |
4583 | list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
77a4d1a1 PZ |
4584 | |
4585 | /* | |
4586 | * If we're the first throttled task, make sure the bandwidth | |
4587 | * timer is running. | |
4588 | */ | |
4589 | if (empty) | |
4590 | start_cfs_bandwidth(cfs_b); | |
4591 | ||
85dac906 PT |
4592 | raw_spin_unlock(&cfs_b->lock); |
4593 | } | |
4594 | ||
029632fb | 4595 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
4596 | { |
4597 | struct rq *rq = rq_of(cfs_rq); | |
4598 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4599 | struct sched_entity *se; | |
4600 | int enqueue = 1; | |
4601 | long task_delta; | |
4602 | ||
22b958d8 | 4603 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
4604 | |
4605 | cfs_rq->throttled = 0; | |
1a55af2e FW |
4606 | |
4607 | update_rq_clock(rq); | |
4608 | ||
671fd9da | 4609 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 4610 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
4611 | list_del_rcu(&cfs_rq->throttled_list); |
4612 | raw_spin_unlock(&cfs_b->lock); | |
4613 | ||
64660c86 PT |
4614 | /* update hierarchical throttle state */ |
4615 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
4616 | ||
671fd9da PT |
4617 | if (!cfs_rq->load.weight) |
4618 | return; | |
4619 | ||
4620 | task_delta = cfs_rq->h_nr_running; | |
4621 | for_each_sched_entity(se) { | |
4622 | if (se->on_rq) | |
4623 | enqueue = 0; | |
4624 | ||
4625 | cfs_rq = cfs_rq_of(se); | |
4626 | if (enqueue) | |
4627 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
4628 | cfs_rq->h_nr_running += task_delta; | |
4629 | ||
4630 | if (cfs_rq_throttled(cfs_rq)) | |
4631 | break; | |
4632 | } | |
4633 | ||
4634 | if (!se) | |
72465447 | 4635 | add_nr_running(rq, task_delta); |
671fd9da PT |
4636 | |
4637 | /* determine whether we need to wake up potentially idle cpu */ | |
4638 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
8875125e | 4639 | resched_curr(rq); |
671fd9da PT |
4640 | } |
4641 | ||
4642 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
4643 | u64 remaining, u64 expires) | |
4644 | { | |
4645 | struct cfs_rq *cfs_rq; | |
c06f04c7 BS |
4646 | u64 runtime; |
4647 | u64 starting_runtime = remaining; | |
671fd9da PT |
4648 | |
4649 | rcu_read_lock(); | |
4650 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
4651 | throttled_list) { | |
4652 | struct rq *rq = rq_of(cfs_rq); | |
8a8c69c3 | 4653 | struct rq_flags rf; |
671fd9da | 4654 | |
8a8c69c3 | 4655 | rq_lock(rq, &rf); |
671fd9da PT |
4656 | if (!cfs_rq_throttled(cfs_rq)) |
4657 | goto next; | |
4658 | ||
4659 | runtime = -cfs_rq->runtime_remaining + 1; | |
4660 | if (runtime > remaining) | |
4661 | runtime = remaining; | |
4662 | remaining -= runtime; | |
4663 | ||
4664 | cfs_rq->runtime_remaining += runtime; | |
4665 | cfs_rq->runtime_expires = expires; | |
4666 | ||
4667 | /* we check whether we're throttled above */ | |
4668 | if (cfs_rq->runtime_remaining > 0) | |
4669 | unthrottle_cfs_rq(cfs_rq); | |
4670 | ||
4671 | next: | |
8a8c69c3 | 4672 | rq_unlock(rq, &rf); |
671fd9da PT |
4673 | |
4674 | if (!remaining) | |
4675 | break; | |
4676 | } | |
4677 | rcu_read_unlock(); | |
4678 | ||
c06f04c7 | 4679 | return starting_runtime - remaining; |
671fd9da PT |
4680 | } |
4681 | ||
58088ad0 PT |
4682 | /* |
4683 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
4684 | * cfs_rqs as appropriate. If there has been no activity within the last | |
4685 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
4686 | * used to track this state. | |
4687 | */ | |
4688 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
4689 | { | |
671fd9da | 4690 | u64 runtime, runtime_expires; |
51f2176d | 4691 | int throttled; |
58088ad0 | 4692 | |
58088ad0 PT |
4693 | /* no need to continue the timer with no bandwidth constraint */ |
4694 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 4695 | goto out_deactivate; |
58088ad0 | 4696 | |
671fd9da | 4697 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 4698 | cfs_b->nr_periods += overrun; |
671fd9da | 4699 | |
51f2176d BS |
4700 | /* |
4701 | * idle depends on !throttled (for the case of a large deficit), and if | |
4702 | * we're going inactive then everything else can be deferred | |
4703 | */ | |
4704 | if (cfs_b->idle && !throttled) | |
4705 | goto out_deactivate; | |
a9cf55b2 PT |
4706 | |
4707 | __refill_cfs_bandwidth_runtime(cfs_b); | |
4708 | ||
671fd9da PT |
4709 | if (!throttled) { |
4710 | /* mark as potentially idle for the upcoming period */ | |
4711 | cfs_b->idle = 1; | |
51f2176d | 4712 | return 0; |
671fd9da PT |
4713 | } |
4714 | ||
e8da1b18 NR |
4715 | /* account preceding periods in which throttling occurred */ |
4716 | cfs_b->nr_throttled += overrun; | |
4717 | ||
671fd9da | 4718 | runtime_expires = cfs_b->runtime_expires; |
671fd9da PT |
4719 | |
4720 | /* | |
c06f04c7 BS |
4721 | * This check is repeated as we are holding onto the new bandwidth while |
4722 | * we unthrottle. This can potentially race with an unthrottled group | |
4723 | * trying to acquire new bandwidth from the global pool. This can result | |
4724 | * in us over-using our runtime if it is all used during this loop, but | |
4725 | * only by limited amounts in that extreme case. | |
671fd9da | 4726 | */ |
c06f04c7 BS |
4727 | while (throttled && cfs_b->runtime > 0) { |
4728 | runtime = cfs_b->runtime; | |
671fd9da PT |
4729 | raw_spin_unlock(&cfs_b->lock); |
4730 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
4731 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
4732 | runtime_expires); | |
4733 | raw_spin_lock(&cfs_b->lock); | |
4734 | ||
4735 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
c06f04c7 BS |
4736 | |
4737 | cfs_b->runtime -= min(runtime, cfs_b->runtime); | |
671fd9da | 4738 | } |
58088ad0 | 4739 | |
671fd9da PT |
4740 | /* |
4741 | * While we are ensured activity in the period following an | |
4742 | * unthrottle, this also covers the case in which the new bandwidth is | |
4743 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
4744 | * timer to remain active while there are any throttled entities.) | |
4745 | */ | |
4746 | cfs_b->idle = 0; | |
58088ad0 | 4747 | |
51f2176d BS |
4748 | return 0; |
4749 | ||
4750 | out_deactivate: | |
51f2176d | 4751 | return 1; |
58088ad0 | 4752 | } |
d3d9dc33 | 4753 | |
d8b4986d PT |
4754 | /* a cfs_rq won't donate quota below this amount */ |
4755 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
4756 | /* minimum remaining period time to redistribute slack quota */ | |
4757 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
4758 | /* how long we wait to gather additional slack before distributing */ | |
4759 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
4760 | ||
db06e78c BS |
4761 | /* |
4762 | * Are we near the end of the current quota period? | |
4763 | * | |
4764 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 4765 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
4766 | * migrate_hrtimers, base is never cleared, so we are fine. |
4767 | */ | |
d8b4986d PT |
4768 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
4769 | { | |
4770 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
4771 | u64 remaining; | |
4772 | ||
4773 | /* if the call-back is running a quota refresh is already occurring */ | |
4774 | if (hrtimer_callback_running(refresh_timer)) | |
4775 | return 1; | |
4776 | ||
4777 | /* is a quota refresh about to occur? */ | |
4778 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
4779 | if (remaining < min_expire) | |
4780 | return 1; | |
4781 | ||
4782 | return 0; | |
4783 | } | |
4784 | ||
4785 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
4786 | { | |
4787 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
4788 | ||
4789 | /* if there's a quota refresh soon don't bother with slack */ | |
4790 | if (runtime_refresh_within(cfs_b, min_left)) | |
4791 | return; | |
4792 | ||
4cfafd30 PZ |
4793 | hrtimer_start(&cfs_b->slack_timer, |
4794 | ns_to_ktime(cfs_bandwidth_slack_period), | |
4795 | HRTIMER_MODE_REL); | |
d8b4986d PT |
4796 | } |
4797 | ||
4798 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
4799 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4800 | { | |
4801 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4802 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
4803 | ||
4804 | if (slack_runtime <= 0) | |
4805 | return; | |
4806 | ||
4807 | raw_spin_lock(&cfs_b->lock); | |
4808 | if (cfs_b->quota != RUNTIME_INF && | |
4809 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
4810 | cfs_b->runtime += slack_runtime; | |
4811 | ||
4812 | /* we are under rq->lock, defer unthrottling using a timer */ | |
4813 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
4814 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
4815 | start_cfs_slack_bandwidth(cfs_b); | |
4816 | } | |
4817 | raw_spin_unlock(&cfs_b->lock); | |
4818 | ||
4819 | /* even if it's not valid for return we don't want to try again */ | |
4820 | cfs_rq->runtime_remaining -= slack_runtime; | |
4821 | } | |
4822 | ||
4823 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4824 | { | |
56f570e5 PT |
4825 | if (!cfs_bandwidth_used()) |
4826 | return; | |
4827 | ||
fccfdc6f | 4828 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
4829 | return; |
4830 | ||
4831 | __return_cfs_rq_runtime(cfs_rq); | |
4832 | } | |
4833 | ||
4834 | /* | |
4835 | * This is done with a timer (instead of inline with bandwidth return) since | |
4836 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
4837 | */ | |
4838 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
4839 | { | |
4840 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
4841 | u64 expires; | |
4842 | ||
4843 | /* confirm we're still not at a refresh boundary */ | |
db06e78c BS |
4844 | raw_spin_lock(&cfs_b->lock); |
4845 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { | |
4846 | raw_spin_unlock(&cfs_b->lock); | |
d8b4986d | 4847 | return; |
db06e78c | 4848 | } |
d8b4986d | 4849 | |
c06f04c7 | 4850 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 4851 | runtime = cfs_b->runtime; |
c06f04c7 | 4852 | |
d8b4986d PT |
4853 | expires = cfs_b->runtime_expires; |
4854 | raw_spin_unlock(&cfs_b->lock); | |
4855 | ||
4856 | if (!runtime) | |
4857 | return; | |
4858 | ||
4859 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
4860 | ||
4861 | raw_spin_lock(&cfs_b->lock); | |
4862 | if (expires == cfs_b->runtime_expires) | |
c06f04c7 | 4863 | cfs_b->runtime -= min(runtime, cfs_b->runtime); |
d8b4986d PT |
4864 | raw_spin_unlock(&cfs_b->lock); |
4865 | } | |
4866 | ||
d3d9dc33 PT |
4867 | /* |
4868 | * When a group wakes up we want to make sure that its quota is not already | |
4869 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
4870 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
4871 | */ | |
4872 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
4873 | { | |
56f570e5 PT |
4874 | if (!cfs_bandwidth_used()) |
4875 | return; | |
4876 | ||
d3d9dc33 PT |
4877 | /* an active group must be handled by the update_curr()->put() path */ |
4878 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
4879 | return; | |
4880 | ||
4881 | /* ensure the group is not already throttled */ | |
4882 | if (cfs_rq_throttled(cfs_rq)) | |
4883 | return; | |
4884 | ||
4885 | /* update runtime allocation */ | |
4886 | account_cfs_rq_runtime(cfs_rq, 0); | |
4887 | if (cfs_rq->runtime_remaining <= 0) | |
4888 | throttle_cfs_rq(cfs_rq); | |
4889 | } | |
4890 | ||
55e16d30 PZ |
4891 | static void sync_throttle(struct task_group *tg, int cpu) |
4892 | { | |
4893 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
4894 | ||
4895 | if (!cfs_bandwidth_used()) | |
4896 | return; | |
4897 | ||
4898 | if (!tg->parent) | |
4899 | return; | |
4900 | ||
4901 | cfs_rq = tg->cfs_rq[cpu]; | |
4902 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
4903 | ||
4904 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
b8922125 | 4905 | cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); |
55e16d30 PZ |
4906 | } |
4907 | ||
d3d9dc33 | 4908 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 4909 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 4910 | { |
56f570e5 | 4911 | if (!cfs_bandwidth_used()) |
678d5718 | 4912 | return false; |
56f570e5 | 4913 | |
d3d9dc33 | 4914 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 4915 | return false; |
d3d9dc33 PT |
4916 | |
4917 | /* | |
4918 | * it's possible for a throttled entity to be forced into a running | |
4919 | * state (e.g. set_curr_task), in this case we're finished. | |
4920 | */ | |
4921 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 4922 | return true; |
d3d9dc33 PT |
4923 | |
4924 | throttle_cfs_rq(cfs_rq); | |
678d5718 | 4925 | return true; |
d3d9dc33 | 4926 | } |
029632fb | 4927 | |
029632fb PZ |
4928 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
4929 | { | |
4930 | struct cfs_bandwidth *cfs_b = | |
4931 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 4932 | |
029632fb PZ |
4933 | do_sched_cfs_slack_timer(cfs_b); |
4934 | ||
4935 | return HRTIMER_NORESTART; | |
4936 | } | |
4937 | ||
4938 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
4939 | { | |
4940 | struct cfs_bandwidth *cfs_b = | |
4941 | container_of(timer, struct cfs_bandwidth, period_timer); | |
029632fb PZ |
4942 | int overrun; |
4943 | int idle = 0; | |
4944 | ||
51f2176d | 4945 | raw_spin_lock(&cfs_b->lock); |
029632fb | 4946 | for (;;) { |
77a4d1a1 | 4947 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
4948 | if (!overrun) |
4949 | break; | |
4950 | ||
4951 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
4952 | } | |
4cfafd30 PZ |
4953 | if (idle) |
4954 | cfs_b->period_active = 0; | |
51f2176d | 4955 | raw_spin_unlock(&cfs_b->lock); |
029632fb PZ |
4956 | |
4957 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
4958 | } | |
4959 | ||
4960 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4961 | { | |
4962 | raw_spin_lock_init(&cfs_b->lock); | |
4963 | cfs_b->runtime = 0; | |
4964 | cfs_b->quota = RUNTIME_INF; | |
4965 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
4966 | ||
4967 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 4968 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
4969 | cfs_b->period_timer.function = sched_cfs_period_timer; |
4970 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
4971 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
4972 | } | |
4973 | ||
4974 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4975 | { | |
4976 | cfs_rq->runtime_enabled = 0; | |
4977 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
4978 | } | |
4979 | ||
77a4d1a1 | 4980 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 4981 | { |
4cfafd30 | 4982 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 4983 | |
4cfafd30 PZ |
4984 | if (!cfs_b->period_active) { |
4985 | cfs_b->period_active = 1; | |
4986 | hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); | |
4987 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); | |
4988 | } | |
029632fb PZ |
4989 | } |
4990 | ||
4991 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4992 | { | |
7f1a169b TH |
4993 | /* init_cfs_bandwidth() was not called */ |
4994 | if (!cfs_b->throttled_cfs_rq.next) | |
4995 | return; | |
4996 | ||
029632fb PZ |
4997 | hrtimer_cancel(&cfs_b->period_timer); |
4998 | hrtimer_cancel(&cfs_b->slack_timer); | |
4999 | } | |
5000 | ||
502ce005 PZ |
5001 | /* |
5002 | * Both these cpu hotplug callbacks race against unregister_fair_sched_group() | |
5003 | * | |
5004 | * The race is harmless, since modifying bandwidth settings of unhooked group | |
5005 | * bits doesn't do much. | |
5006 | */ | |
5007 | ||
5008 | /* cpu online calback */ | |
0e59bdae KT |
5009 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
5010 | { | |
502ce005 | 5011 | struct task_group *tg; |
0e59bdae | 5012 | |
502ce005 PZ |
5013 | lockdep_assert_held(&rq->lock); |
5014 | ||
5015 | rcu_read_lock(); | |
5016 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5017 | struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; | |
5018 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
0e59bdae KT |
5019 | |
5020 | raw_spin_lock(&cfs_b->lock); | |
5021 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
5022 | raw_spin_unlock(&cfs_b->lock); | |
5023 | } | |
502ce005 | 5024 | rcu_read_unlock(); |
0e59bdae KT |
5025 | } |
5026 | ||
502ce005 | 5027 | /* cpu offline callback */ |
38dc3348 | 5028 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb | 5029 | { |
502ce005 PZ |
5030 | struct task_group *tg; |
5031 | ||
5032 | lockdep_assert_held(&rq->lock); | |
5033 | ||
5034 | rcu_read_lock(); | |
5035 | list_for_each_entry_rcu(tg, &task_groups, list) { | |
5036 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
029632fb | 5037 | |
029632fb PZ |
5038 | if (!cfs_rq->runtime_enabled) |
5039 | continue; | |
5040 | ||
5041 | /* | |
5042 | * clock_task is not advancing so we just need to make sure | |
5043 | * there's some valid quota amount | |
5044 | */ | |
51f2176d | 5045 | cfs_rq->runtime_remaining = 1; |
0e59bdae KT |
5046 | /* |
5047 | * Offline rq is schedulable till cpu is completely disabled | |
5048 | * in take_cpu_down(), so we prevent new cfs throttling here. | |
5049 | */ | |
5050 | cfs_rq->runtime_enabled = 0; | |
5051 | ||
029632fb PZ |
5052 | if (cfs_rq_throttled(cfs_rq)) |
5053 | unthrottle_cfs_rq(cfs_rq); | |
5054 | } | |
502ce005 | 5055 | rcu_read_unlock(); |
029632fb PZ |
5056 | } |
5057 | ||
5058 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
5059 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
5060 | { | |
78becc27 | 5061 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
5062 | } |
5063 | ||
9dbdb155 | 5064 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 5065 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 5066 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 5067 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 5068 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
5069 | |
5070 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
5071 | { | |
5072 | return 0; | |
5073 | } | |
64660c86 PT |
5074 | |
5075 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
5076 | { | |
5077 | return 0; | |
5078 | } | |
5079 | ||
5080 | static inline int throttled_lb_pair(struct task_group *tg, | |
5081 | int src_cpu, int dest_cpu) | |
5082 | { | |
5083 | return 0; | |
5084 | } | |
029632fb PZ |
5085 | |
5086 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
5087 | ||
5088 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
5089 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
5090 | #endif |
5091 | ||
029632fb PZ |
5092 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
5093 | { | |
5094 | return NULL; | |
5095 | } | |
5096 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 5097 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 5098 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
5099 | |
5100 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
5101 | ||
bf0f6f24 IM |
5102 | /************************************************** |
5103 | * CFS operations on tasks: | |
5104 | */ | |
5105 | ||
8f4d37ec PZ |
5106 | #ifdef CONFIG_SCHED_HRTICK |
5107 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5108 | { | |
8f4d37ec PZ |
5109 | struct sched_entity *se = &p->se; |
5110 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
5111 | ||
9148a3a1 | 5112 | SCHED_WARN_ON(task_rq(p) != rq); |
8f4d37ec | 5113 | |
8bf46a39 | 5114 | if (rq->cfs.h_nr_running > 1) { |
8f4d37ec PZ |
5115 | u64 slice = sched_slice(cfs_rq, se); |
5116 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
5117 | s64 delta = slice - ran; | |
5118 | ||
5119 | if (delta < 0) { | |
5120 | if (rq->curr == p) | |
8875125e | 5121 | resched_curr(rq); |
8f4d37ec PZ |
5122 | return; |
5123 | } | |
31656519 | 5124 | hrtick_start(rq, delta); |
8f4d37ec PZ |
5125 | } |
5126 | } | |
a4c2f00f PZ |
5127 | |
5128 | /* | |
5129 | * called from enqueue/dequeue and updates the hrtick when the | |
5130 | * current task is from our class and nr_running is low enough | |
5131 | * to matter. | |
5132 | */ | |
5133 | static void hrtick_update(struct rq *rq) | |
5134 | { | |
5135 | struct task_struct *curr = rq->curr; | |
5136 | ||
b39e66ea | 5137 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
5138 | return; |
5139 | ||
5140 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
5141 | hrtick_start_fair(rq, curr); | |
5142 | } | |
55e12e5e | 5143 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
5144 | static inline void |
5145 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
5146 | { | |
5147 | } | |
a4c2f00f PZ |
5148 | |
5149 | static inline void hrtick_update(struct rq *rq) | |
5150 | { | |
5151 | } | |
8f4d37ec PZ |
5152 | #endif |
5153 | ||
bf0f6f24 IM |
5154 | /* |
5155 | * The enqueue_task method is called before nr_running is | |
5156 | * increased. Here we update the fair scheduling stats and | |
5157 | * then put the task into the rbtree: | |
5158 | */ | |
ea87bb78 | 5159 | static void |
371fd7e7 | 5160 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5161 | { |
5162 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5163 | struct sched_entity *se = &p->se; |
bf0f6f24 | 5164 | |
8c34ab19 RW |
5165 | /* |
5166 | * If in_iowait is set, the code below may not trigger any cpufreq | |
5167 | * utilization updates, so do it here explicitly with the IOWAIT flag | |
5168 | * passed. | |
5169 | */ | |
5170 | if (p->in_iowait) | |
674e7541 | 5171 | cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); |
8c34ab19 | 5172 | |
bf0f6f24 | 5173 | for_each_sched_entity(se) { |
62fb1851 | 5174 | if (se->on_rq) |
bf0f6f24 IM |
5175 | break; |
5176 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 5177 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
5178 | |
5179 | /* | |
5180 | * end evaluation on encountering a throttled cfs_rq | |
5181 | * | |
5182 | * note: in the case of encountering a throttled cfs_rq we will | |
5183 | * post the final h_nr_running increment below. | |
e210bffd | 5184 | */ |
85dac906 PT |
5185 | if (cfs_rq_throttled(cfs_rq)) |
5186 | break; | |
953bfcd1 | 5187 | cfs_rq->h_nr_running++; |
85dac906 | 5188 | |
88ec22d3 | 5189 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 5190 | } |
8f4d37ec | 5191 | |
2069dd75 | 5192 | for_each_sched_entity(se) { |
0f317143 | 5193 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 5194 | cfs_rq->h_nr_running++; |
2069dd75 | 5195 | |
85dac906 PT |
5196 | if (cfs_rq_throttled(cfs_rq)) |
5197 | break; | |
5198 | ||
88c0616e | 5199 | update_load_avg(cfs_rq, se, UPDATE_TG); |
1ea6c46a | 5200 | update_cfs_group(se); |
2069dd75 PZ |
5201 | } |
5202 | ||
cd126afe | 5203 | if (!se) |
72465447 | 5204 | add_nr_running(rq, 1); |
cd126afe | 5205 | |
a4c2f00f | 5206 | hrtick_update(rq); |
bf0f6f24 IM |
5207 | } |
5208 | ||
2f36825b VP |
5209 | static void set_next_buddy(struct sched_entity *se); |
5210 | ||
bf0f6f24 IM |
5211 | /* |
5212 | * The dequeue_task method is called before nr_running is | |
5213 | * decreased. We remove the task from the rbtree and | |
5214 | * update the fair scheduling stats: | |
5215 | */ | |
371fd7e7 | 5216 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
5217 | { |
5218 | struct cfs_rq *cfs_rq; | |
62fb1851 | 5219 | struct sched_entity *se = &p->se; |
2f36825b | 5220 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
5221 | |
5222 | for_each_sched_entity(se) { | |
5223 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 5224 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
5225 | |
5226 | /* | |
5227 | * end evaluation on encountering a throttled cfs_rq | |
5228 | * | |
5229 | * note: in the case of encountering a throttled cfs_rq we will | |
5230 | * post the final h_nr_running decrement below. | |
5231 | */ | |
5232 | if (cfs_rq_throttled(cfs_rq)) | |
5233 | break; | |
953bfcd1 | 5234 | cfs_rq->h_nr_running--; |
2069dd75 | 5235 | |
bf0f6f24 | 5236 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 5237 | if (cfs_rq->load.weight) { |
754bd598 KK |
5238 | /* Avoid re-evaluating load for this entity: */ |
5239 | se = parent_entity(se); | |
2f36825b VP |
5240 | /* |
5241 | * Bias pick_next to pick a task from this cfs_rq, as | |
5242 | * p is sleeping when it is within its sched_slice. | |
5243 | */ | |
754bd598 KK |
5244 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
5245 | set_next_buddy(se); | |
bf0f6f24 | 5246 | break; |
2f36825b | 5247 | } |
371fd7e7 | 5248 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 5249 | } |
8f4d37ec | 5250 | |
2069dd75 | 5251 | for_each_sched_entity(se) { |
0f317143 | 5252 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 5253 | cfs_rq->h_nr_running--; |
2069dd75 | 5254 | |
85dac906 PT |
5255 | if (cfs_rq_throttled(cfs_rq)) |
5256 | break; | |
5257 | ||
88c0616e | 5258 | update_load_avg(cfs_rq, se, UPDATE_TG); |
1ea6c46a | 5259 | update_cfs_group(se); |
2069dd75 PZ |
5260 | } |
5261 | ||
cd126afe | 5262 | if (!se) |
72465447 | 5263 | sub_nr_running(rq, 1); |
cd126afe | 5264 | |
a4c2f00f | 5265 | hrtick_update(rq); |
bf0f6f24 IM |
5266 | } |
5267 | ||
e7693a36 | 5268 | #ifdef CONFIG_SMP |
10e2f1ac PZ |
5269 | |
5270 | /* Working cpumask for: load_balance, load_balance_newidle. */ | |
5271 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); | |
5272 | DEFINE_PER_CPU(cpumask_var_t, select_idle_mask); | |
5273 | ||
9fd81dd5 | 5274 | #ifdef CONFIG_NO_HZ_COMMON |
3289bdb4 PZ |
5275 | /* |
5276 | * per rq 'load' arrray crap; XXX kill this. | |
5277 | */ | |
5278 | ||
5279 | /* | |
d937cdc5 | 5280 | * The exact cpuload calculated at every tick would be: |
3289bdb4 | 5281 | * |
d937cdc5 PZ |
5282 | * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load |
5283 | * | |
5284 | * If a cpu misses updates for n ticks (as it was idle) and update gets | |
5285 | * called on the n+1-th tick when cpu may be busy, then we have: | |
5286 | * | |
5287 | * load_n = (1 - 1/2^i)^n * load_0 | |
5288 | * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load | |
3289bdb4 PZ |
5289 | * |
5290 | * decay_load_missed() below does efficient calculation of | |
3289bdb4 | 5291 | * |
d937cdc5 PZ |
5292 | * load' = (1 - 1/2^i)^n * load |
5293 | * | |
5294 | * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors. | |
5295 | * This allows us to precompute the above in said factors, thereby allowing the | |
5296 | * reduction of an arbitrary n in O(log_2 n) steps. (See also | |
5297 | * fixed_power_int()) | |
3289bdb4 | 5298 | * |
d937cdc5 | 5299 | * The calculation is approximated on a 128 point scale. |
3289bdb4 PZ |
5300 | */ |
5301 | #define DEGRADE_SHIFT 7 | |
d937cdc5 PZ |
5302 | |
5303 | static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; | |
5304 | static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { | |
5305 | { 0, 0, 0, 0, 0, 0, 0, 0 }, | |
5306 | { 64, 32, 8, 0, 0, 0, 0, 0 }, | |
5307 | { 96, 72, 40, 12, 1, 0, 0, 0 }, | |
5308 | { 112, 98, 75, 43, 15, 1, 0, 0 }, | |
5309 | { 120, 112, 98, 76, 45, 16, 2, 0 } | |
5310 | }; | |
3289bdb4 PZ |
5311 | |
5312 | /* | |
5313 | * Update cpu_load for any missed ticks, due to tickless idle. The backlog | |
5314 | * would be when CPU is idle and so we just decay the old load without | |
5315 | * adding any new load. | |
5316 | */ | |
5317 | static unsigned long | |
5318 | decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) | |
5319 | { | |
5320 | int j = 0; | |
5321 | ||
5322 | if (!missed_updates) | |
5323 | return load; | |
5324 | ||
5325 | if (missed_updates >= degrade_zero_ticks[idx]) | |
5326 | return 0; | |
5327 | ||
5328 | if (idx == 1) | |
5329 | return load >> missed_updates; | |
5330 | ||
5331 | while (missed_updates) { | |
5332 | if (missed_updates % 2) | |
5333 | load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; | |
5334 | ||
5335 | missed_updates >>= 1; | |
5336 | j++; | |
5337 | } | |
5338 | return load; | |
5339 | } | |
9fd81dd5 | 5340 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 5341 | |
59543275 | 5342 | /** |
cee1afce | 5343 | * __cpu_load_update - update the rq->cpu_load[] statistics |
59543275 BP |
5344 | * @this_rq: The rq to update statistics for |
5345 | * @this_load: The current load | |
5346 | * @pending_updates: The number of missed updates | |
59543275 | 5347 | * |
3289bdb4 | 5348 | * Update rq->cpu_load[] statistics. This function is usually called every |
59543275 BP |
5349 | * scheduler tick (TICK_NSEC). |
5350 | * | |
5351 | * This function computes a decaying average: | |
5352 | * | |
5353 | * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load | |
5354 | * | |
5355 | * Because of NOHZ it might not get called on every tick which gives need for | |
5356 | * the @pending_updates argument. | |
5357 | * | |
5358 | * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1 | |
5359 | * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load | |
5360 | * = A * (A * load[i]_n-2 + B) + B | |
5361 | * = A * (A * (A * load[i]_n-3 + B) + B) + B | |
5362 | * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B | |
5363 | * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B | |
5364 | * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B | |
5365 | * = (1 - 1/2^i)^n * (load[i]_0 - load) + load | |
5366 | * | |
5367 | * In the above we've assumed load_n := load, which is true for NOHZ_FULL as | |
5368 | * any change in load would have resulted in the tick being turned back on. | |
5369 | * | |
5370 | * For regular NOHZ, this reduces to: | |
5371 | * | |
5372 | * load[i]_n = (1 - 1/2^i)^n * load[i]_0 | |
5373 | * | |
5374 | * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra | |
1f41906a | 5375 | * term. |
3289bdb4 | 5376 | */ |
1f41906a FW |
5377 | static void cpu_load_update(struct rq *this_rq, unsigned long this_load, |
5378 | unsigned long pending_updates) | |
3289bdb4 | 5379 | { |
9fd81dd5 | 5380 | unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0]; |
3289bdb4 PZ |
5381 | int i, scale; |
5382 | ||
5383 | this_rq->nr_load_updates++; | |
5384 | ||
5385 | /* Update our load: */ | |
5386 | this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ | |
5387 | for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
5388 | unsigned long old_load, new_load; | |
5389 | ||
5390 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
5391 | ||
7400d3bb | 5392 | old_load = this_rq->cpu_load[i]; |
9fd81dd5 | 5393 | #ifdef CONFIG_NO_HZ_COMMON |
3289bdb4 | 5394 | old_load = decay_load_missed(old_load, pending_updates - 1, i); |
7400d3bb BP |
5395 | if (tickless_load) { |
5396 | old_load -= decay_load_missed(tickless_load, pending_updates - 1, i); | |
5397 | /* | |
5398 | * old_load can never be a negative value because a | |
5399 | * decayed tickless_load cannot be greater than the | |
5400 | * original tickless_load. | |
5401 | */ | |
5402 | old_load += tickless_load; | |
5403 | } | |
9fd81dd5 | 5404 | #endif |
3289bdb4 PZ |
5405 | new_load = this_load; |
5406 | /* | |
5407 | * Round up the averaging division if load is increasing. This | |
5408 | * prevents us from getting stuck on 9 if the load is 10, for | |
5409 | * example. | |
5410 | */ | |
5411 | if (new_load > old_load) | |
5412 | new_load += scale - 1; | |
5413 | ||
5414 | this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; | |
5415 | } | |
5416 | ||
5417 | sched_avg_update(this_rq); | |
5418 | } | |
5419 | ||
7ea241af | 5420 | /* Used instead of source_load when we know the type == 0 */ |
c7132dd6 | 5421 | static unsigned long weighted_cpuload(struct rq *rq) |
7ea241af | 5422 | { |
c7132dd6 | 5423 | return cfs_rq_runnable_load_avg(&rq->cfs); |
7ea241af YD |
5424 | } |
5425 | ||
3289bdb4 | 5426 | #ifdef CONFIG_NO_HZ_COMMON |
1f41906a FW |
5427 | /* |
5428 | * There is no sane way to deal with nohz on smp when using jiffies because the | |
5429 | * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading | |
5430 | * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. | |
5431 | * | |
5432 | * Therefore we need to avoid the delta approach from the regular tick when | |
5433 | * possible since that would seriously skew the load calculation. This is why we | |
5434 | * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on | |
5435 | * jiffies deltas for updates happening while in nohz mode (idle ticks, idle | |
5436 | * loop exit, nohz_idle_balance, nohz full exit...) | |
5437 | * | |
5438 | * This means we might still be one tick off for nohz periods. | |
5439 | */ | |
5440 | ||
5441 | static void cpu_load_update_nohz(struct rq *this_rq, | |
5442 | unsigned long curr_jiffies, | |
5443 | unsigned long load) | |
be68a682 FW |
5444 | { |
5445 | unsigned long pending_updates; | |
5446 | ||
5447 | pending_updates = curr_jiffies - this_rq->last_load_update_tick; | |
5448 | if (pending_updates) { | |
5449 | this_rq->last_load_update_tick = curr_jiffies; | |
5450 | /* | |
5451 | * In the regular NOHZ case, we were idle, this means load 0. | |
5452 | * In the NOHZ_FULL case, we were non-idle, we should consider | |
5453 | * its weighted load. | |
5454 | */ | |
1f41906a | 5455 | cpu_load_update(this_rq, load, pending_updates); |
be68a682 FW |
5456 | } |
5457 | } | |
5458 | ||
3289bdb4 PZ |
5459 | /* |
5460 | * Called from nohz_idle_balance() to update the load ratings before doing the | |
5461 | * idle balance. | |
5462 | */ | |
cee1afce | 5463 | static void cpu_load_update_idle(struct rq *this_rq) |
3289bdb4 | 5464 | { |
3289bdb4 PZ |
5465 | /* |
5466 | * bail if there's load or we're actually up-to-date. | |
5467 | */ | |
c7132dd6 | 5468 | if (weighted_cpuload(this_rq)) |
3289bdb4 PZ |
5469 | return; |
5470 | ||
1f41906a | 5471 | cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0); |
3289bdb4 PZ |
5472 | } |
5473 | ||
5474 | /* | |
1f41906a FW |
5475 | * Record CPU load on nohz entry so we know the tickless load to account |
5476 | * on nohz exit. cpu_load[0] happens then to be updated more frequently | |
5477 | * than other cpu_load[idx] but it should be fine as cpu_load readers | |
5478 | * shouldn't rely into synchronized cpu_load[*] updates. | |
3289bdb4 | 5479 | */ |
1f41906a | 5480 | void cpu_load_update_nohz_start(void) |
3289bdb4 PZ |
5481 | { |
5482 | struct rq *this_rq = this_rq(); | |
1f41906a FW |
5483 | |
5484 | /* | |
5485 | * This is all lockless but should be fine. If weighted_cpuload changes | |
5486 | * concurrently we'll exit nohz. And cpu_load write can race with | |
5487 | * cpu_load_update_idle() but both updater would be writing the same. | |
5488 | */ | |
c7132dd6 | 5489 | this_rq->cpu_load[0] = weighted_cpuload(this_rq); |
1f41906a FW |
5490 | } |
5491 | ||
5492 | /* | |
5493 | * Account the tickless load in the end of a nohz frame. | |
5494 | */ | |
5495 | void cpu_load_update_nohz_stop(void) | |
5496 | { | |
316c1608 | 5497 | unsigned long curr_jiffies = READ_ONCE(jiffies); |
1f41906a FW |
5498 | struct rq *this_rq = this_rq(); |
5499 | unsigned long load; | |
8a8c69c3 | 5500 | struct rq_flags rf; |
3289bdb4 PZ |
5501 | |
5502 | if (curr_jiffies == this_rq->last_load_update_tick) | |
5503 | return; | |
5504 | ||
c7132dd6 | 5505 | load = weighted_cpuload(this_rq); |
8a8c69c3 | 5506 | rq_lock(this_rq, &rf); |
b52fad2d | 5507 | update_rq_clock(this_rq); |
1f41906a | 5508 | cpu_load_update_nohz(this_rq, curr_jiffies, load); |
8a8c69c3 | 5509 | rq_unlock(this_rq, &rf); |
3289bdb4 | 5510 | } |
1f41906a FW |
5511 | #else /* !CONFIG_NO_HZ_COMMON */ |
5512 | static inline void cpu_load_update_nohz(struct rq *this_rq, | |
5513 | unsigned long curr_jiffies, | |
5514 | unsigned long load) { } | |
5515 | #endif /* CONFIG_NO_HZ_COMMON */ | |
5516 | ||
5517 | static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load) | |
5518 | { | |
9fd81dd5 | 5519 | #ifdef CONFIG_NO_HZ_COMMON |
1f41906a FW |
5520 | /* See the mess around cpu_load_update_nohz(). */ |
5521 | this_rq->last_load_update_tick = READ_ONCE(jiffies); | |
9fd81dd5 | 5522 | #endif |
1f41906a FW |
5523 | cpu_load_update(this_rq, load, 1); |
5524 | } | |
3289bdb4 PZ |
5525 | |
5526 | /* | |
5527 | * Called from scheduler_tick() | |
5528 | */ | |
cee1afce | 5529 | void cpu_load_update_active(struct rq *this_rq) |
3289bdb4 | 5530 | { |
c7132dd6 | 5531 | unsigned long load = weighted_cpuload(this_rq); |
1f41906a FW |
5532 | |
5533 | if (tick_nohz_tick_stopped()) | |
5534 | cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load); | |
5535 | else | |
5536 | cpu_load_update_periodic(this_rq, load); | |
3289bdb4 PZ |
5537 | } |
5538 | ||
029632fb PZ |
5539 | /* |
5540 | * Return a low guess at the load of a migration-source cpu weighted | |
5541 | * according to the scheduling class and "nice" value. | |
5542 | * | |
5543 | * We want to under-estimate the load of migration sources, to | |
5544 | * balance conservatively. | |
5545 | */ | |
5546 | static unsigned long source_load(int cpu, int type) | |
5547 | { | |
5548 | struct rq *rq = cpu_rq(cpu); | |
c7132dd6 | 5549 | unsigned long total = weighted_cpuload(rq); |
029632fb PZ |
5550 | |
5551 | if (type == 0 || !sched_feat(LB_BIAS)) | |
5552 | return total; | |
5553 | ||
5554 | return min(rq->cpu_load[type-1], total); | |
5555 | } | |
5556 | ||
5557 | /* | |
5558 | * Return a high guess at the load of a migration-target cpu weighted | |
5559 | * according to the scheduling class and "nice" value. | |
5560 | */ | |
5561 | static unsigned long target_load(int cpu, int type) | |
5562 | { | |
5563 | struct rq *rq = cpu_rq(cpu); | |
c7132dd6 | 5564 | unsigned long total = weighted_cpuload(rq); |
029632fb PZ |
5565 | |
5566 | if (type == 0 || !sched_feat(LB_BIAS)) | |
5567 | return total; | |
5568 | ||
5569 | return max(rq->cpu_load[type-1], total); | |
5570 | } | |
5571 | ||
ced549fa | 5572 | static unsigned long capacity_of(int cpu) |
029632fb | 5573 | { |
ced549fa | 5574 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
5575 | } |
5576 | ||
ca6d75e6 VG |
5577 | static unsigned long capacity_orig_of(int cpu) |
5578 | { | |
5579 | return cpu_rq(cpu)->cpu_capacity_orig; | |
5580 | } | |
5581 | ||
029632fb PZ |
5582 | static unsigned long cpu_avg_load_per_task(int cpu) |
5583 | { | |
5584 | struct rq *rq = cpu_rq(cpu); | |
316c1608 | 5585 | unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running); |
c7132dd6 | 5586 | unsigned long load_avg = weighted_cpuload(rq); |
029632fb PZ |
5587 | |
5588 | if (nr_running) | |
b92486cb | 5589 | return load_avg / nr_running; |
029632fb PZ |
5590 | |
5591 | return 0; | |
5592 | } | |
5593 | ||
c58d25f3 PZ |
5594 | static void record_wakee(struct task_struct *p) |
5595 | { | |
5596 | /* | |
5597 | * Only decay a single time; tasks that have less then 1 wakeup per | |
5598 | * jiffy will not have built up many flips. | |
5599 | */ | |
5600 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
5601 | current->wakee_flips >>= 1; | |
5602 | current->wakee_flip_decay_ts = jiffies; | |
5603 | } | |
5604 | ||
5605 | if (current->last_wakee != p) { | |
5606 | current->last_wakee = p; | |
5607 | current->wakee_flips++; | |
5608 | } | |
5609 | } | |
5610 | ||
63b0e9ed MG |
5611 | /* |
5612 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 5613 | * |
63b0e9ed | 5614 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
5615 | * at a frequency roughly N times higher than one of its wakees. |
5616 | * | |
5617 | * In order to determine whether we should let the load spread vs consolidating | |
5618 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
5619 | * partner, and a factor of lls_size higher frequency in the other. | |
5620 | * | |
5621 | * With both conditions met, we can be relatively sure that the relationship is | |
5622 | * non-monogamous, with partner count exceeding socket size. | |
5623 | * | |
5624 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
5625 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
5626 | * socket size. | |
63b0e9ed | 5627 | */ |
62470419 MW |
5628 | static int wake_wide(struct task_struct *p) |
5629 | { | |
63b0e9ed MG |
5630 | unsigned int master = current->wakee_flips; |
5631 | unsigned int slave = p->wakee_flips; | |
7d9ffa89 | 5632 | int factor = this_cpu_read(sd_llc_size); |
62470419 | 5633 | |
63b0e9ed MG |
5634 | if (master < slave) |
5635 | swap(master, slave); | |
5636 | if (slave < factor || master < slave * factor) | |
5637 | return 0; | |
5638 | return 1; | |
62470419 MW |
5639 | } |
5640 | ||
90001d67 | 5641 | /* |
d153b153 PZ |
5642 | * The purpose of wake_affine() is to quickly determine on which CPU we can run |
5643 | * soonest. For the purpose of speed we only consider the waking and previous | |
5644 | * CPU. | |
90001d67 | 5645 | * |
d153b153 PZ |
5646 | * wake_affine_idle() - only considers 'now', it check if the waking CPU is (or |
5647 | * will be) idle. | |
f2cdd9cc PZ |
5648 | * |
5649 | * wake_affine_weight() - considers the weight to reflect the average | |
5650 | * scheduling latency of the CPUs. This seems to work | |
5651 | * for the overloaded case. | |
90001d67 | 5652 | */ |
90001d67 | 5653 | |
90001d67 | 5654 | static bool |
d153b153 PZ |
5655 | wake_affine_idle(struct sched_domain *sd, struct task_struct *p, |
5656 | int this_cpu, int prev_cpu, int sync) | |
90001d67 | 5657 | { |
d153b153 | 5658 | if (idle_cpu(this_cpu)) |
90001d67 | 5659 | return true; |
90001d67 | 5660 | |
d153b153 PZ |
5661 | if (sync && cpu_rq(this_cpu)->nr_running == 1) |
5662 | return true; | |
90001d67 | 5663 | |
d153b153 | 5664 | return false; |
90001d67 PZ |
5665 | } |
5666 | ||
90001d67 | 5667 | static bool |
f2cdd9cc PZ |
5668 | wake_affine_weight(struct sched_domain *sd, struct task_struct *p, |
5669 | int this_cpu, int prev_cpu, int sync) | |
90001d67 | 5670 | { |
90001d67 PZ |
5671 | s64 this_eff_load, prev_eff_load; |
5672 | unsigned long task_load; | |
5673 | ||
f2cdd9cc PZ |
5674 | this_eff_load = target_load(this_cpu, sd->wake_idx); |
5675 | prev_eff_load = source_load(prev_cpu, sd->wake_idx); | |
90001d67 | 5676 | |
90001d67 PZ |
5677 | if (sync) { |
5678 | unsigned long current_load = task_h_load(current); | |
5679 | ||
f2cdd9cc | 5680 | if (current_load > this_eff_load) |
90001d67 PZ |
5681 | return true; |
5682 | ||
f2cdd9cc | 5683 | this_eff_load -= current_load; |
90001d67 PZ |
5684 | } |
5685 | ||
90001d67 PZ |
5686 | task_load = task_h_load(p); |
5687 | ||
f2cdd9cc PZ |
5688 | this_eff_load += task_load; |
5689 | if (sched_feat(WA_BIAS)) | |
5690 | this_eff_load *= 100; | |
5691 | this_eff_load *= capacity_of(prev_cpu); | |
90001d67 | 5692 | |
f2cdd9cc PZ |
5693 | prev_eff_load -= task_load; |
5694 | if (sched_feat(WA_BIAS)) | |
5695 | prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; | |
5696 | prev_eff_load *= capacity_of(this_cpu); | |
90001d67 PZ |
5697 | |
5698 | return this_eff_load <= prev_eff_load; | |
5699 | } | |
5700 | ||
772bd008 MR |
5701 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
5702 | int prev_cpu, int sync) | |
098fb9db | 5703 | { |
3fed382b | 5704 | int this_cpu = smp_processor_id(); |
d153b153 | 5705 | bool affine = false; |
098fb9db | 5706 | |
d153b153 PZ |
5707 | if (sched_feat(WA_IDLE) && !affine) |
5708 | affine = wake_affine_idle(sd, p, this_cpu, prev_cpu, sync); | |
90001d67 | 5709 | |
f2cdd9cc PZ |
5710 | if (sched_feat(WA_WEIGHT) && !affine) |
5711 | affine = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); | |
098fb9db | 5712 | |
ae92882e | 5713 | schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); |
3fed382b RR |
5714 | if (affine) { |
5715 | schedstat_inc(sd->ttwu_move_affine); | |
5716 | schedstat_inc(p->se.statistics.nr_wakeups_affine); | |
5717 | } | |
098fb9db | 5718 | |
3fed382b | 5719 | return affine; |
098fb9db IM |
5720 | } |
5721 | ||
6a0b19c0 MR |
5722 | static inline int task_util(struct task_struct *p); |
5723 | static int cpu_util_wake(int cpu, struct task_struct *p); | |
5724 | ||
5725 | static unsigned long capacity_spare_wake(int cpu, struct task_struct *p) | |
5726 | { | |
5727 | return capacity_orig_of(cpu) - cpu_util_wake(cpu, p); | |
5728 | } | |
5729 | ||
aaee1203 PZ |
5730 | /* |
5731 | * find_idlest_group finds and returns the least busy CPU group within the | |
5732 | * domain. | |
5733 | */ | |
5734 | static struct sched_group * | |
78e7ed53 | 5735 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
c44f2a02 | 5736 | int this_cpu, int sd_flag) |
e7693a36 | 5737 | { |
b3bd3de6 | 5738 | struct sched_group *idlest = NULL, *group = sd->groups; |
6a0b19c0 | 5739 | struct sched_group *most_spare_sg = NULL; |
6b94780e VG |
5740 | unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0; |
5741 | unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0; | |
6a0b19c0 | 5742 | unsigned long most_spare = 0, this_spare = 0; |
c44f2a02 | 5743 | int load_idx = sd->forkexec_idx; |
6b94780e VG |
5744 | int imbalance_scale = 100 + (sd->imbalance_pct-100)/2; |
5745 | unsigned long imbalance = scale_load_down(NICE_0_LOAD) * | |
5746 | (sd->imbalance_pct-100) / 100; | |
e7693a36 | 5747 | |
c44f2a02 VG |
5748 | if (sd_flag & SD_BALANCE_WAKE) |
5749 | load_idx = sd->wake_idx; | |
5750 | ||
aaee1203 | 5751 | do { |
6b94780e VG |
5752 | unsigned long load, avg_load, runnable_load; |
5753 | unsigned long spare_cap, max_spare_cap; | |
aaee1203 PZ |
5754 | int local_group; |
5755 | int i; | |
e7693a36 | 5756 | |
aaee1203 | 5757 | /* Skip over this group if it has no CPUs allowed */ |
ae4df9d6 | 5758 | if (!cpumask_intersects(sched_group_span(group), |
0c98d344 | 5759 | &p->cpus_allowed)) |
aaee1203 PZ |
5760 | continue; |
5761 | ||
5762 | local_group = cpumask_test_cpu(this_cpu, | |
ae4df9d6 | 5763 | sched_group_span(group)); |
aaee1203 | 5764 | |
6a0b19c0 MR |
5765 | /* |
5766 | * Tally up the load of all CPUs in the group and find | |
5767 | * the group containing the CPU with most spare capacity. | |
5768 | */ | |
aaee1203 | 5769 | avg_load = 0; |
6b94780e | 5770 | runnable_load = 0; |
6a0b19c0 | 5771 | max_spare_cap = 0; |
aaee1203 | 5772 | |
ae4df9d6 | 5773 | for_each_cpu(i, sched_group_span(group)) { |
aaee1203 PZ |
5774 | /* Bias balancing toward cpus of our domain */ |
5775 | if (local_group) | |
5776 | load = source_load(i, load_idx); | |
5777 | else | |
5778 | load = target_load(i, load_idx); | |
5779 | ||
6b94780e VG |
5780 | runnable_load += load; |
5781 | ||
5782 | avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs); | |
6a0b19c0 MR |
5783 | |
5784 | spare_cap = capacity_spare_wake(i, p); | |
5785 | ||
5786 | if (spare_cap > max_spare_cap) | |
5787 | max_spare_cap = spare_cap; | |
aaee1203 PZ |
5788 | } |
5789 | ||
63b2ca30 | 5790 | /* Adjust by relative CPU capacity of the group */ |
6b94780e VG |
5791 | avg_load = (avg_load * SCHED_CAPACITY_SCALE) / |
5792 | group->sgc->capacity; | |
5793 | runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) / | |
5794 | group->sgc->capacity; | |
aaee1203 PZ |
5795 | |
5796 | if (local_group) { | |
6b94780e VG |
5797 | this_runnable_load = runnable_load; |
5798 | this_avg_load = avg_load; | |
6a0b19c0 MR |
5799 | this_spare = max_spare_cap; |
5800 | } else { | |
6b94780e VG |
5801 | if (min_runnable_load > (runnable_load + imbalance)) { |
5802 | /* | |
5803 | * The runnable load is significantly smaller | |
5804 | * so we can pick this new cpu | |
5805 | */ | |
5806 | min_runnable_load = runnable_load; | |
5807 | min_avg_load = avg_load; | |
5808 | idlest = group; | |
5809 | } else if ((runnable_load < (min_runnable_load + imbalance)) && | |
5810 | (100*min_avg_load > imbalance_scale*avg_load)) { | |
5811 | /* | |
5812 | * The runnable loads are close so take the | |
5813 | * blocked load into account through avg_load. | |
5814 | */ | |
5815 | min_avg_load = avg_load; | |
6a0b19c0 MR |
5816 | idlest = group; |
5817 | } | |
5818 | ||
5819 | if (most_spare < max_spare_cap) { | |
5820 | most_spare = max_spare_cap; | |
5821 | most_spare_sg = group; | |
5822 | } | |
aaee1203 PZ |
5823 | } |
5824 | } while (group = group->next, group != sd->groups); | |
5825 | ||
6a0b19c0 MR |
5826 | /* |
5827 | * The cross-over point between using spare capacity or least load | |
5828 | * is too conservative for high utilization tasks on partially | |
5829 | * utilized systems if we require spare_capacity > task_util(p), | |
5830 | * so we allow for some task stuffing by using | |
5831 | * spare_capacity > task_util(p)/2. | |
f519a3f1 VG |
5832 | * |
5833 | * Spare capacity can't be used for fork because the utilization has | |
5834 | * not been set yet, we must first select a rq to compute the initial | |
5835 | * utilization. | |
6a0b19c0 | 5836 | */ |
f519a3f1 VG |
5837 | if (sd_flag & SD_BALANCE_FORK) |
5838 | goto skip_spare; | |
5839 | ||
6a0b19c0 | 5840 | if (this_spare > task_util(p) / 2 && |
6b94780e | 5841 | imbalance_scale*this_spare > 100*most_spare) |
6a0b19c0 | 5842 | return NULL; |
6b94780e VG |
5843 | |
5844 | if (most_spare > task_util(p) / 2) | |
6a0b19c0 MR |
5845 | return most_spare_sg; |
5846 | ||
f519a3f1 | 5847 | skip_spare: |
6b94780e VG |
5848 | if (!idlest) |
5849 | return NULL; | |
5850 | ||
5851 | if (min_runnable_load > (this_runnable_load + imbalance)) | |
aaee1203 | 5852 | return NULL; |
6b94780e VG |
5853 | |
5854 | if ((this_runnable_load < (min_runnable_load + imbalance)) && | |
5855 | (100*this_avg_load < imbalance_scale*min_avg_load)) | |
5856 | return NULL; | |
5857 | ||
aaee1203 PZ |
5858 | return idlest; |
5859 | } | |
5860 | ||
5861 | /* | |
5862 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
5863 | */ | |
5864 | static int | |
5865 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
5866 | { | |
5867 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
5868 | unsigned int min_exit_latency = UINT_MAX; |
5869 | u64 latest_idle_timestamp = 0; | |
5870 | int least_loaded_cpu = this_cpu; | |
5871 | int shallowest_idle_cpu = -1; | |
aaee1203 PZ |
5872 | int i; |
5873 | ||
eaecf41f MR |
5874 | /* Check if we have any choice: */ |
5875 | if (group->group_weight == 1) | |
ae4df9d6 | 5876 | return cpumask_first(sched_group_span(group)); |
eaecf41f | 5877 | |
aaee1203 | 5878 | /* Traverse only the allowed CPUs */ |
ae4df9d6 | 5879 | for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) { |
83a0a96a NP |
5880 | if (idle_cpu(i)) { |
5881 | struct rq *rq = cpu_rq(i); | |
5882 | struct cpuidle_state *idle = idle_get_state(rq); | |
5883 | if (idle && idle->exit_latency < min_exit_latency) { | |
5884 | /* | |
5885 | * We give priority to a CPU whose idle state | |
5886 | * has the smallest exit latency irrespective | |
5887 | * of any idle timestamp. | |
5888 | */ | |
5889 | min_exit_latency = idle->exit_latency; | |
5890 | latest_idle_timestamp = rq->idle_stamp; | |
5891 | shallowest_idle_cpu = i; | |
5892 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
5893 | rq->idle_stamp > latest_idle_timestamp) { | |
5894 | /* | |
5895 | * If equal or no active idle state, then | |
5896 | * the most recently idled CPU might have | |
5897 | * a warmer cache. | |
5898 | */ | |
5899 | latest_idle_timestamp = rq->idle_stamp; | |
5900 | shallowest_idle_cpu = i; | |
5901 | } | |
9f96742a | 5902 | } else if (shallowest_idle_cpu == -1) { |
c7132dd6 | 5903 | load = weighted_cpuload(cpu_rq(i)); |
83a0a96a NP |
5904 | if (load < min_load || (load == min_load && i == this_cpu)) { |
5905 | min_load = load; | |
5906 | least_loaded_cpu = i; | |
5907 | } | |
e7693a36 GH |
5908 | } |
5909 | } | |
5910 | ||
83a0a96a | 5911 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 5912 | } |
e7693a36 | 5913 | |
10e2f1ac PZ |
5914 | #ifdef CONFIG_SCHED_SMT |
5915 | ||
5916 | static inline void set_idle_cores(int cpu, int val) | |
5917 | { | |
5918 | struct sched_domain_shared *sds; | |
5919 | ||
5920 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
5921 | if (sds) | |
5922 | WRITE_ONCE(sds->has_idle_cores, val); | |
5923 | } | |
5924 | ||
5925 | static inline bool test_idle_cores(int cpu, bool def) | |
5926 | { | |
5927 | struct sched_domain_shared *sds; | |
5928 | ||
5929 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
5930 | if (sds) | |
5931 | return READ_ONCE(sds->has_idle_cores); | |
5932 | ||
5933 | return def; | |
5934 | } | |
5935 | ||
5936 | /* | |
5937 | * Scans the local SMT mask to see if the entire core is idle, and records this | |
5938 | * information in sd_llc_shared->has_idle_cores. | |
5939 | * | |
5940 | * Since SMT siblings share all cache levels, inspecting this limited remote | |
5941 | * state should be fairly cheap. | |
5942 | */ | |
1b568f0a | 5943 | void __update_idle_core(struct rq *rq) |
10e2f1ac PZ |
5944 | { |
5945 | int core = cpu_of(rq); | |
5946 | int cpu; | |
5947 | ||
5948 | rcu_read_lock(); | |
5949 | if (test_idle_cores(core, true)) | |
5950 | goto unlock; | |
5951 | ||
5952 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
5953 | if (cpu == core) | |
5954 | continue; | |
5955 | ||
5956 | if (!idle_cpu(cpu)) | |
5957 | goto unlock; | |
5958 | } | |
5959 | ||
5960 | set_idle_cores(core, 1); | |
5961 | unlock: | |
5962 | rcu_read_unlock(); | |
5963 | } | |
5964 | ||
5965 | /* | |
5966 | * Scan the entire LLC domain for idle cores; this dynamically switches off if | |
5967 | * there are no idle cores left in the system; tracked through | |
5968 | * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. | |
5969 | */ | |
5970 | static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) | |
5971 | { | |
5972 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); | |
c743f0a5 | 5973 | int core, cpu; |
10e2f1ac | 5974 | |
1b568f0a PZ |
5975 | if (!static_branch_likely(&sched_smt_present)) |
5976 | return -1; | |
5977 | ||
10e2f1ac PZ |
5978 | if (!test_idle_cores(target, false)) |
5979 | return -1; | |
5980 | ||
0c98d344 | 5981 | cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed); |
10e2f1ac | 5982 | |
c743f0a5 | 5983 | for_each_cpu_wrap(core, cpus, target) { |
10e2f1ac PZ |
5984 | bool idle = true; |
5985 | ||
5986 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
5987 | cpumask_clear_cpu(cpu, cpus); | |
5988 | if (!idle_cpu(cpu)) | |
5989 | idle = false; | |
5990 | } | |
5991 | ||
5992 | if (idle) | |
5993 | return core; | |
5994 | } | |
5995 | ||
5996 | /* | |
5997 | * Failed to find an idle core; stop looking for one. | |
5998 | */ | |
5999 | set_idle_cores(target, 0); | |
6000 | ||
6001 | return -1; | |
6002 | } | |
6003 | ||
6004 | /* | |
6005 | * Scan the local SMT mask for idle CPUs. | |
6006 | */ | |
6007 | static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) | |
6008 | { | |
6009 | int cpu; | |
6010 | ||
1b568f0a PZ |
6011 | if (!static_branch_likely(&sched_smt_present)) |
6012 | return -1; | |
6013 | ||
10e2f1ac | 6014 | for_each_cpu(cpu, cpu_smt_mask(target)) { |
0c98d344 | 6015 | if (!cpumask_test_cpu(cpu, &p->cpus_allowed)) |
10e2f1ac PZ |
6016 | continue; |
6017 | if (idle_cpu(cpu)) | |
6018 | return cpu; | |
6019 | } | |
6020 | ||
6021 | return -1; | |
6022 | } | |
6023 | ||
6024 | #else /* CONFIG_SCHED_SMT */ | |
6025 | ||
6026 | static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) | |
6027 | { | |
6028 | return -1; | |
6029 | } | |
6030 | ||
6031 | static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) | |
6032 | { | |
6033 | return -1; | |
6034 | } | |
6035 | ||
6036 | #endif /* CONFIG_SCHED_SMT */ | |
6037 | ||
6038 | /* | |
6039 | * Scan the LLC domain for idle CPUs; this is dynamically regulated by | |
6040 | * comparing the average scan cost (tracked in sd->avg_scan_cost) against the | |
6041 | * average idle time for this rq (as found in rq->avg_idle). | |
a50bde51 | 6042 | */ |
10e2f1ac PZ |
6043 | static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target) |
6044 | { | |
9cfb38a7 | 6045 | struct sched_domain *this_sd; |
1ad3aaf3 | 6046 | u64 avg_cost, avg_idle; |
10e2f1ac PZ |
6047 | u64 time, cost; |
6048 | s64 delta; | |
1ad3aaf3 | 6049 | int cpu, nr = INT_MAX; |
10e2f1ac | 6050 | |
9cfb38a7 WL |
6051 | this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); |
6052 | if (!this_sd) | |
6053 | return -1; | |
6054 | ||
10e2f1ac PZ |
6055 | /* |
6056 | * Due to large variance we need a large fuzz factor; hackbench in | |
6057 | * particularly is sensitive here. | |
6058 | */ | |
1ad3aaf3 PZ |
6059 | avg_idle = this_rq()->avg_idle / 512; |
6060 | avg_cost = this_sd->avg_scan_cost + 1; | |
6061 | ||
6062 | if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost) | |
10e2f1ac PZ |
6063 | return -1; |
6064 | ||
1ad3aaf3 PZ |
6065 | if (sched_feat(SIS_PROP)) { |
6066 | u64 span_avg = sd->span_weight * avg_idle; | |
6067 | if (span_avg > 4*avg_cost) | |
6068 | nr = div_u64(span_avg, avg_cost); | |
6069 | else | |
6070 | nr = 4; | |
6071 | } | |
6072 | ||
10e2f1ac PZ |
6073 | time = local_clock(); |
6074 | ||
c743f0a5 | 6075 | for_each_cpu_wrap(cpu, sched_domain_span(sd), target) { |
1ad3aaf3 PZ |
6076 | if (!--nr) |
6077 | return -1; | |
0c98d344 | 6078 | if (!cpumask_test_cpu(cpu, &p->cpus_allowed)) |
10e2f1ac PZ |
6079 | continue; |
6080 | if (idle_cpu(cpu)) | |
6081 | break; | |
6082 | } | |
6083 | ||
6084 | time = local_clock() - time; | |
6085 | cost = this_sd->avg_scan_cost; | |
6086 | delta = (s64)(time - cost) / 8; | |
6087 | this_sd->avg_scan_cost += delta; | |
6088 | ||
6089 | return cpu; | |
6090 | } | |
6091 | ||
6092 | /* | |
6093 | * Try and locate an idle core/thread in the LLC cache domain. | |
a50bde51 | 6094 | */ |
772bd008 | 6095 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 6096 | { |
99bd5e2f | 6097 | struct sched_domain *sd; |
10e2f1ac | 6098 | int i; |
a50bde51 | 6099 | |
e0a79f52 MG |
6100 | if (idle_cpu(target)) |
6101 | return target; | |
99bd5e2f SS |
6102 | |
6103 | /* | |
10e2f1ac | 6104 | * If the previous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 6105 | */ |
772bd008 MR |
6106 | if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev)) |
6107 | return prev; | |
a50bde51 | 6108 | |
518cd623 | 6109 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
10e2f1ac PZ |
6110 | if (!sd) |
6111 | return target; | |
772bd008 | 6112 | |
10e2f1ac PZ |
6113 | i = select_idle_core(p, sd, target); |
6114 | if ((unsigned)i < nr_cpumask_bits) | |
6115 | return i; | |
37407ea7 | 6116 | |
10e2f1ac PZ |
6117 | i = select_idle_cpu(p, sd, target); |
6118 | if ((unsigned)i < nr_cpumask_bits) | |
6119 | return i; | |
6120 | ||
6121 | i = select_idle_smt(p, sd, target); | |
6122 | if ((unsigned)i < nr_cpumask_bits) | |
6123 | return i; | |
970e1789 | 6124 | |
a50bde51 PZ |
6125 | return target; |
6126 | } | |
231678b7 | 6127 | |
8bb5b00c | 6128 | /* |
9e91d61d | 6129 | * cpu_util returns the amount of capacity of a CPU that is used by CFS |
8bb5b00c | 6130 | * tasks. The unit of the return value must be the one of capacity so we can |
9e91d61d DE |
6131 | * compare the utilization with the capacity of the CPU that is available for |
6132 | * CFS task (ie cpu_capacity). | |
231678b7 DE |
6133 | * |
6134 | * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the | |
6135 | * recent utilization of currently non-runnable tasks on a CPU. It represents | |
6136 | * the amount of utilization of a CPU in the range [0..capacity_orig] where | |
6137 | * capacity_orig is the cpu_capacity available at the highest frequency | |
6138 | * (arch_scale_freq_capacity()). | |
6139 | * The utilization of a CPU converges towards a sum equal to or less than the | |
6140 | * current capacity (capacity_curr <= capacity_orig) of the CPU because it is | |
6141 | * the running time on this CPU scaled by capacity_curr. | |
6142 | * | |
6143 | * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even | |
6144 | * higher than capacity_orig because of unfortunate rounding in | |
6145 | * cfs.avg.util_avg or just after migrating tasks and new task wakeups until | |
6146 | * the average stabilizes with the new running time. We need to check that the | |
6147 | * utilization stays within the range of [0..capacity_orig] and cap it if | |
6148 | * necessary. Without utilization capping, a group could be seen as overloaded | |
6149 | * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of | |
6150 | * available capacity. We allow utilization to overshoot capacity_curr (but not | |
6151 | * capacity_orig) as it useful for predicting the capacity required after task | |
6152 | * migrations (scheduler-driven DVFS). | |
8bb5b00c | 6153 | */ |
9e91d61d | 6154 | static int cpu_util(int cpu) |
8bb5b00c | 6155 | { |
9e91d61d | 6156 | unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg; |
8bb5b00c VG |
6157 | unsigned long capacity = capacity_orig_of(cpu); |
6158 | ||
231678b7 | 6159 | return (util >= capacity) ? capacity : util; |
8bb5b00c | 6160 | } |
a50bde51 | 6161 | |
3273163c MR |
6162 | static inline int task_util(struct task_struct *p) |
6163 | { | |
6164 | return p->se.avg.util_avg; | |
6165 | } | |
6166 | ||
104cb16d MR |
6167 | /* |
6168 | * cpu_util_wake: Compute cpu utilization with any contributions from | |
6169 | * the waking task p removed. | |
6170 | */ | |
6171 | static int cpu_util_wake(int cpu, struct task_struct *p) | |
6172 | { | |
6173 | unsigned long util, capacity; | |
6174 | ||
6175 | /* Task has no contribution or is new */ | |
6176 | if (cpu != task_cpu(p) || !p->se.avg.last_update_time) | |
6177 | return cpu_util(cpu); | |
6178 | ||
6179 | capacity = capacity_orig_of(cpu); | |
6180 | util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0); | |
6181 | ||
6182 | return (util >= capacity) ? capacity : util; | |
6183 | } | |
6184 | ||
3273163c MR |
6185 | /* |
6186 | * Disable WAKE_AFFINE in the case where task @p doesn't fit in the | |
6187 | * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu. | |
6188 | * | |
6189 | * In that case WAKE_AFFINE doesn't make sense and we'll let | |
6190 | * BALANCE_WAKE sort things out. | |
6191 | */ | |
6192 | static int wake_cap(struct task_struct *p, int cpu, int prev_cpu) | |
6193 | { | |
6194 | long min_cap, max_cap; | |
6195 | ||
6196 | min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu)); | |
6197 | max_cap = cpu_rq(cpu)->rd->max_cpu_capacity; | |
6198 | ||
6199 | /* Minimum capacity is close to max, no need to abort wake_affine */ | |
6200 | if (max_cap - min_cap < max_cap >> 3) | |
6201 | return 0; | |
6202 | ||
104cb16d MR |
6203 | /* Bring task utilization in sync with prev_cpu */ |
6204 | sync_entity_load_avg(&p->se); | |
6205 | ||
3273163c MR |
6206 | return min_cap * 1024 < task_util(p) * capacity_margin; |
6207 | } | |
6208 | ||
aaee1203 | 6209 | /* |
de91b9cb MR |
6210 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
6211 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, | |
6212 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. | |
aaee1203 | 6213 | * |
de91b9cb MR |
6214 | * Balances load by selecting the idlest cpu in the idlest group, or under |
6215 | * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 6216 | * |
de91b9cb | 6217 | * Returns the target cpu number. |
aaee1203 PZ |
6218 | * |
6219 | * preempt must be disabled. | |
6220 | */ | |
0017d735 | 6221 | static int |
ac66f547 | 6222 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 6223 | { |
29cd8bae | 6224 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 6225 | int cpu = smp_processor_id(); |
63b0e9ed | 6226 | int new_cpu = prev_cpu; |
99bd5e2f | 6227 | int want_affine = 0; |
5158f4e4 | 6228 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 6229 | |
c58d25f3 PZ |
6230 | if (sd_flag & SD_BALANCE_WAKE) { |
6231 | record_wakee(p); | |
3273163c | 6232 | want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) |
0c98d344 | 6233 | && cpumask_test_cpu(cpu, &p->cpus_allowed); |
c58d25f3 | 6234 | } |
aaee1203 | 6235 | |
dce840a0 | 6236 | rcu_read_lock(); |
aaee1203 | 6237 | for_each_domain(cpu, tmp) { |
e4f42888 | 6238 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
63b0e9ed | 6239 | break; |
e4f42888 | 6240 | |
fe3bcfe1 | 6241 | /* |
99bd5e2f SS |
6242 | * If both cpu and prev_cpu are part of this domain, |
6243 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 6244 | */ |
99bd5e2f SS |
6245 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
6246 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
6247 | affine_sd = tmp; | |
29cd8bae | 6248 | break; |
f03542a7 | 6249 | } |
29cd8bae | 6250 | |
f03542a7 | 6251 | if (tmp->flags & sd_flag) |
29cd8bae | 6252 | sd = tmp; |
63b0e9ed MG |
6253 | else if (!want_affine) |
6254 | break; | |
29cd8bae PZ |
6255 | } |
6256 | ||
63b0e9ed MG |
6257 | if (affine_sd) { |
6258 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
7d894e6e RR |
6259 | if (cpu == prev_cpu) |
6260 | goto pick_cpu; | |
6261 | ||
6262 | if (wake_affine(affine_sd, p, prev_cpu, sync)) | |
63b0e9ed | 6263 | new_cpu = cpu; |
8b911acd | 6264 | } |
e7693a36 | 6265 | |
63b0e9ed | 6266 | if (!sd) { |
7d894e6e | 6267 | pick_cpu: |
63b0e9ed | 6268 | if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */ |
772bd008 | 6269 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); |
63b0e9ed MG |
6270 | |
6271 | } else while (sd) { | |
aaee1203 | 6272 | struct sched_group *group; |
c88d5910 | 6273 | int weight; |
098fb9db | 6274 | |
0763a660 | 6275 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
6276 | sd = sd->child; |
6277 | continue; | |
6278 | } | |
098fb9db | 6279 | |
c44f2a02 | 6280 | group = find_idlest_group(sd, p, cpu, sd_flag); |
aaee1203 PZ |
6281 | if (!group) { |
6282 | sd = sd->child; | |
6283 | continue; | |
6284 | } | |
4ae7d5ce | 6285 | |
d7c33c49 | 6286 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
6287 | if (new_cpu == -1 || new_cpu == cpu) { |
6288 | /* Now try balancing at a lower domain level of cpu */ | |
6289 | sd = sd->child; | |
6290 | continue; | |
e7693a36 | 6291 | } |
aaee1203 PZ |
6292 | |
6293 | /* Now try balancing at a lower domain level of new_cpu */ | |
6294 | cpu = new_cpu; | |
669c55e9 | 6295 | weight = sd->span_weight; |
aaee1203 PZ |
6296 | sd = NULL; |
6297 | for_each_domain(cpu, tmp) { | |
669c55e9 | 6298 | if (weight <= tmp->span_weight) |
aaee1203 | 6299 | break; |
0763a660 | 6300 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
6301 | sd = tmp; |
6302 | } | |
6303 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 6304 | } |
dce840a0 | 6305 | rcu_read_unlock(); |
e7693a36 | 6306 | |
c88d5910 | 6307 | return new_cpu; |
e7693a36 | 6308 | } |
0a74bef8 | 6309 | |
144d8487 PZ |
6310 | static void detach_entity_cfs_rq(struct sched_entity *se); |
6311 | ||
0a74bef8 PT |
6312 | /* |
6313 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
6314 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
525628c7 | 6315 | * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 6316 | */ |
5a4fd036 | 6317 | static void migrate_task_rq_fair(struct task_struct *p) |
0a74bef8 | 6318 | { |
59efa0ba PZ |
6319 | /* |
6320 | * As blocked tasks retain absolute vruntime the migration needs to | |
6321 | * deal with this by subtracting the old and adding the new | |
6322 | * min_vruntime -- the latter is done by enqueue_entity() when placing | |
6323 | * the task on the new runqueue. | |
6324 | */ | |
6325 | if (p->state == TASK_WAKING) { | |
6326 | struct sched_entity *se = &p->se; | |
6327 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6328 | u64 min_vruntime; | |
6329 | ||
6330 | #ifndef CONFIG_64BIT | |
6331 | u64 min_vruntime_copy; | |
6332 | ||
6333 | do { | |
6334 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
6335 | smp_rmb(); | |
6336 | min_vruntime = cfs_rq->min_vruntime; | |
6337 | } while (min_vruntime != min_vruntime_copy); | |
6338 | #else | |
6339 | min_vruntime = cfs_rq->min_vruntime; | |
6340 | #endif | |
6341 | ||
6342 | se->vruntime -= min_vruntime; | |
6343 | } | |
6344 | ||
144d8487 PZ |
6345 | if (p->on_rq == TASK_ON_RQ_MIGRATING) { |
6346 | /* | |
6347 | * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old' | |
6348 | * rq->lock and can modify state directly. | |
6349 | */ | |
6350 | lockdep_assert_held(&task_rq(p)->lock); | |
6351 | detach_entity_cfs_rq(&p->se); | |
6352 | ||
6353 | } else { | |
6354 | /* | |
6355 | * We are supposed to update the task to "current" time, then | |
6356 | * its up to date and ready to go to new CPU/cfs_rq. But we | |
6357 | * have difficulty in getting what current time is, so simply | |
6358 | * throw away the out-of-date time. This will result in the | |
6359 | * wakee task is less decayed, but giving the wakee more load | |
6360 | * sounds not bad. | |
6361 | */ | |
6362 | remove_entity_load_avg(&p->se); | |
6363 | } | |
9d89c257 YD |
6364 | |
6365 | /* Tell new CPU we are migrated */ | |
6366 | p->se.avg.last_update_time = 0; | |
3944a927 BS |
6367 | |
6368 | /* We have migrated, no longer consider this task hot */ | |
9d89c257 | 6369 | p->se.exec_start = 0; |
0a74bef8 | 6370 | } |
12695578 YD |
6371 | |
6372 | static void task_dead_fair(struct task_struct *p) | |
6373 | { | |
6374 | remove_entity_load_avg(&p->se); | |
6375 | } | |
e7693a36 GH |
6376 | #endif /* CONFIG_SMP */ |
6377 | ||
e52fb7c0 PZ |
6378 | static unsigned long |
6379 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
6380 | { |
6381 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
6382 | ||
6383 | /* | |
e52fb7c0 PZ |
6384 | * Since its curr running now, convert the gran from real-time |
6385 | * to virtual-time in his units. | |
13814d42 MG |
6386 | * |
6387 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
6388 | * they get preempted easier. That is, if 'se' < 'curr' then | |
6389 | * the resulting gran will be larger, therefore penalizing the | |
6390 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
6391 | * be smaller, again penalizing the lighter task. | |
6392 | * | |
6393 | * This is especially important for buddies when the leftmost | |
6394 | * task is higher priority than the buddy. | |
0bbd3336 | 6395 | */ |
f4ad9bd2 | 6396 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
6397 | } |
6398 | ||
464b7527 PZ |
6399 | /* |
6400 | * Should 'se' preempt 'curr'. | |
6401 | * | |
6402 | * |s1 | |
6403 | * |s2 | |
6404 | * |s3 | |
6405 | * g | |
6406 | * |<--->|c | |
6407 | * | |
6408 | * w(c, s1) = -1 | |
6409 | * w(c, s2) = 0 | |
6410 | * w(c, s3) = 1 | |
6411 | * | |
6412 | */ | |
6413 | static int | |
6414 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
6415 | { | |
6416 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
6417 | ||
6418 | if (vdiff <= 0) | |
6419 | return -1; | |
6420 | ||
e52fb7c0 | 6421 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
6422 | if (vdiff > gran) |
6423 | return 1; | |
6424 | ||
6425 | return 0; | |
6426 | } | |
6427 | ||
02479099 PZ |
6428 | static void set_last_buddy(struct sched_entity *se) |
6429 | { | |
69c80f3e VP |
6430 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
6431 | return; | |
6432 | ||
c5ae366e DA |
6433 | for_each_sched_entity(se) { |
6434 | if (SCHED_WARN_ON(!se->on_rq)) | |
6435 | return; | |
69c80f3e | 6436 | cfs_rq_of(se)->last = se; |
c5ae366e | 6437 | } |
02479099 PZ |
6438 | } |
6439 | ||
6440 | static void set_next_buddy(struct sched_entity *se) | |
6441 | { | |
69c80f3e VP |
6442 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
6443 | return; | |
6444 | ||
c5ae366e DA |
6445 | for_each_sched_entity(se) { |
6446 | if (SCHED_WARN_ON(!se->on_rq)) | |
6447 | return; | |
69c80f3e | 6448 | cfs_rq_of(se)->next = se; |
c5ae366e | 6449 | } |
02479099 PZ |
6450 | } |
6451 | ||
ac53db59 RR |
6452 | static void set_skip_buddy(struct sched_entity *se) |
6453 | { | |
69c80f3e VP |
6454 | for_each_sched_entity(se) |
6455 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
6456 | } |
6457 | ||
bf0f6f24 IM |
6458 | /* |
6459 | * Preempt the current task with a newly woken task if needed: | |
6460 | */ | |
5a9b86f6 | 6461 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
6462 | { |
6463 | struct task_struct *curr = rq->curr; | |
8651a86c | 6464 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 6465 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 6466 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 6467 | int next_buddy_marked = 0; |
bf0f6f24 | 6468 | |
4ae7d5ce IM |
6469 | if (unlikely(se == pse)) |
6470 | return; | |
6471 | ||
5238cdd3 | 6472 | /* |
163122b7 | 6473 | * This is possible from callers such as attach_tasks(), in which we |
5238cdd3 PT |
6474 | * unconditionally check_prempt_curr() after an enqueue (which may have |
6475 | * lead to a throttle). This both saves work and prevents false | |
6476 | * next-buddy nomination below. | |
6477 | */ | |
6478 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
6479 | return; | |
6480 | ||
2f36825b | 6481 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 6482 | set_next_buddy(pse); |
2f36825b VP |
6483 | next_buddy_marked = 1; |
6484 | } | |
57fdc26d | 6485 | |
aec0a514 BR |
6486 | /* |
6487 | * We can come here with TIF_NEED_RESCHED already set from new task | |
6488 | * wake up path. | |
5238cdd3 PT |
6489 | * |
6490 | * Note: this also catches the edge-case of curr being in a throttled | |
6491 | * group (e.g. via set_curr_task), since update_curr() (in the | |
6492 | * enqueue of curr) will have resulted in resched being set. This | |
6493 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
6494 | * below. | |
aec0a514 BR |
6495 | */ |
6496 | if (test_tsk_need_resched(curr)) | |
6497 | return; | |
6498 | ||
a2f5c9ab DH |
6499 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
6500 | if (unlikely(curr->policy == SCHED_IDLE) && | |
6501 | likely(p->policy != SCHED_IDLE)) | |
6502 | goto preempt; | |
6503 | ||
91c234b4 | 6504 | /* |
a2f5c9ab DH |
6505 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
6506 | * is driven by the tick): | |
91c234b4 | 6507 | */ |
8ed92e51 | 6508 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 6509 | return; |
bf0f6f24 | 6510 | |
464b7527 | 6511 | find_matching_se(&se, &pse); |
9bbd7374 | 6512 | update_curr(cfs_rq_of(se)); |
002f128b | 6513 | BUG_ON(!pse); |
2f36825b VP |
6514 | if (wakeup_preempt_entity(se, pse) == 1) { |
6515 | /* | |
6516 | * Bias pick_next to pick the sched entity that is | |
6517 | * triggering this preemption. | |
6518 | */ | |
6519 | if (!next_buddy_marked) | |
6520 | set_next_buddy(pse); | |
3a7e73a2 | 6521 | goto preempt; |
2f36825b | 6522 | } |
464b7527 | 6523 | |
3a7e73a2 | 6524 | return; |
a65ac745 | 6525 | |
3a7e73a2 | 6526 | preempt: |
8875125e | 6527 | resched_curr(rq); |
3a7e73a2 PZ |
6528 | /* |
6529 | * Only set the backward buddy when the current task is still | |
6530 | * on the rq. This can happen when a wakeup gets interleaved | |
6531 | * with schedule on the ->pre_schedule() or idle_balance() | |
6532 | * point, either of which can * drop the rq lock. | |
6533 | * | |
6534 | * Also, during early boot the idle thread is in the fair class, | |
6535 | * for obvious reasons its a bad idea to schedule back to it. | |
6536 | */ | |
6537 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
6538 | return; | |
6539 | ||
6540 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
6541 | set_last_buddy(se); | |
bf0f6f24 IM |
6542 | } |
6543 | ||
606dba2e | 6544 | static struct task_struct * |
d8ac8971 | 6545 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
bf0f6f24 IM |
6546 | { |
6547 | struct cfs_rq *cfs_rq = &rq->cfs; | |
6548 | struct sched_entity *se; | |
678d5718 | 6549 | struct task_struct *p; |
37e117c0 | 6550 | int new_tasks; |
678d5718 | 6551 | |
6e83125c | 6552 | again: |
678d5718 | 6553 | if (!cfs_rq->nr_running) |
38033c37 | 6554 | goto idle; |
678d5718 | 6555 | |
9674f5ca | 6556 | #ifdef CONFIG_FAIR_GROUP_SCHED |
3f1d2a31 | 6557 | if (prev->sched_class != &fair_sched_class) |
678d5718 PZ |
6558 | goto simple; |
6559 | ||
6560 | /* | |
6561 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
6562 | * likely that a next task is from the same cgroup as the current. | |
6563 | * | |
6564 | * Therefore attempt to avoid putting and setting the entire cgroup | |
6565 | * hierarchy, only change the part that actually changes. | |
6566 | */ | |
6567 | ||
6568 | do { | |
6569 | struct sched_entity *curr = cfs_rq->curr; | |
6570 | ||
6571 | /* | |
6572 | * Since we got here without doing put_prev_entity() we also | |
6573 | * have to consider cfs_rq->curr. If it is still a runnable | |
6574 | * entity, update_curr() will update its vruntime, otherwise | |
6575 | * forget we've ever seen it. | |
6576 | */ | |
54d27365 BS |
6577 | if (curr) { |
6578 | if (curr->on_rq) | |
6579 | update_curr(cfs_rq); | |
6580 | else | |
6581 | curr = NULL; | |
678d5718 | 6582 | |
54d27365 BS |
6583 | /* |
6584 | * This call to check_cfs_rq_runtime() will do the | |
6585 | * throttle and dequeue its entity in the parent(s). | |
9674f5ca | 6586 | * Therefore the nr_running test will indeed |
54d27365 BS |
6587 | * be correct. |
6588 | */ | |
9674f5ca VK |
6589 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) { |
6590 | cfs_rq = &rq->cfs; | |
6591 | ||
6592 | if (!cfs_rq->nr_running) | |
6593 | goto idle; | |
6594 | ||
54d27365 | 6595 | goto simple; |
9674f5ca | 6596 | } |
54d27365 | 6597 | } |
678d5718 PZ |
6598 | |
6599 | se = pick_next_entity(cfs_rq, curr); | |
6600 | cfs_rq = group_cfs_rq(se); | |
6601 | } while (cfs_rq); | |
6602 | ||
6603 | p = task_of(se); | |
6604 | ||
6605 | /* | |
6606 | * Since we haven't yet done put_prev_entity and if the selected task | |
6607 | * is a different task than we started out with, try and touch the | |
6608 | * least amount of cfs_rqs. | |
6609 | */ | |
6610 | if (prev != p) { | |
6611 | struct sched_entity *pse = &prev->se; | |
6612 | ||
6613 | while (!(cfs_rq = is_same_group(se, pse))) { | |
6614 | int se_depth = se->depth; | |
6615 | int pse_depth = pse->depth; | |
6616 | ||
6617 | if (se_depth <= pse_depth) { | |
6618 | put_prev_entity(cfs_rq_of(pse), pse); | |
6619 | pse = parent_entity(pse); | |
6620 | } | |
6621 | if (se_depth >= pse_depth) { | |
6622 | set_next_entity(cfs_rq_of(se), se); | |
6623 | se = parent_entity(se); | |
6624 | } | |
6625 | } | |
6626 | ||
6627 | put_prev_entity(cfs_rq, pse); | |
6628 | set_next_entity(cfs_rq, se); | |
6629 | } | |
6630 | ||
93824900 | 6631 | goto done; |
678d5718 | 6632 | simple: |
678d5718 | 6633 | #endif |
bf0f6f24 | 6634 | |
3f1d2a31 | 6635 | put_prev_task(rq, prev); |
606dba2e | 6636 | |
bf0f6f24 | 6637 | do { |
678d5718 | 6638 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 6639 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
6640 | cfs_rq = group_cfs_rq(se); |
6641 | } while (cfs_rq); | |
6642 | ||
8f4d37ec | 6643 | p = task_of(se); |
678d5718 | 6644 | |
93824900 UR |
6645 | done: __maybe_unused |
6646 | #ifdef CONFIG_SMP | |
6647 | /* | |
6648 | * Move the next running task to the front of | |
6649 | * the list, so our cfs_tasks list becomes MRU | |
6650 | * one. | |
6651 | */ | |
6652 | list_move(&p->se.group_node, &rq->cfs_tasks); | |
6653 | #endif | |
6654 | ||
b39e66ea MG |
6655 | if (hrtick_enabled(rq)) |
6656 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
6657 | |
6658 | return p; | |
38033c37 PZ |
6659 | |
6660 | idle: | |
46f69fa3 MF |
6661 | new_tasks = idle_balance(rq, rf); |
6662 | ||
37e117c0 PZ |
6663 | /* |
6664 | * Because idle_balance() releases (and re-acquires) rq->lock, it is | |
6665 | * possible for any higher priority task to appear. In that case we | |
6666 | * must re-start the pick_next_entity() loop. | |
6667 | */ | |
e4aa358b | 6668 | if (new_tasks < 0) |
37e117c0 PZ |
6669 | return RETRY_TASK; |
6670 | ||
e4aa358b | 6671 | if (new_tasks > 0) |
38033c37 | 6672 | goto again; |
38033c37 PZ |
6673 | |
6674 | return NULL; | |
bf0f6f24 IM |
6675 | } |
6676 | ||
6677 | /* | |
6678 | * Account for a descheduled task: | |
6679 | */ | |
31ee529c | 6680 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
6681 | { |
6682 | struct sched_entity *se = &prev->se; | |
6683 | struct cfs_rq *cfs_rq; | |
6684 | ||
6685 | for_each_sched_entity(se) { | |
6686 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 6687 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
6688 | } |
6689 | } | |
6690 | ||
ac53db59 RR |
6691 | /* |
6692 | * sched_yield() is very simple | |
6693 | * | |
6694 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
6695 | */ | |
6696 | static void yield_task_fair(struct rq *rq) | |
6697 | { | |
6698 | struct task_struct *curr = rq->curr; | |
6699 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
6700 | struct sched_entity *se = &curr->se; | |
6701 | ||
6702 | /* | |
6703 | * Are we the only task in the tree? | |
6704 | */ | |
6705 | if (unlikely(rq->nr_running == 1)) | |
6706 | return; | |
6707 | ||
6708 | clear_buddies(cfs_rq, se); | |
6709 | ||
6710 | if (curr->policy != SCHED_BATCH) { | |
6711 | update_rq_clock(rq); | |
6712 | /* | |
6713 | * Update run-time statistics of the 'current'. | |
6714 | */ | |
6715 | update_curr(cfs_rq); | |
916671c0 MG |
6716 | /* |
6717 | * Tell update_rq_clock() that we've just updated, | |
6718 | * so we don't do microscopic update in schedule() | |
6719 | * and double the fastpath cost. | |
6720 | */ | |
9edfbfed | 6721 | rq_clock_skip_update(rq, true); |
ac53db59 RR |
6722 | } |
6723 | ||
6724 | set_skip_buddy(se); | |
6725 | } | |
6726 | ||
d95f4122 MG |
6727 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
6728 | { | |
6729 | struct sched_entity *se = &p->se; | |
6730 | ||
5238cdd3 PT |
6731 | /* throttled hierarchies are not runnable */ |
6732 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
6733 | return false; |
6734 | ||
6735 | /* Tell the scheduler that we'd really like pse to run next. */ | |
6736 | set_next_buddy(se); | |
6737 | ||
d95f4122 MG |
6738 | yield_task_fair(rq); |
6739 | ||
6740 | return true; | |
6741 | } | |
6742 | ||
681f3e68 | 6743 | #ifdef CONFIG_SMP |
bf0f6f24 | 6744 | /************************************************** |
e9c84cb8 PZ |
6745 | * Fair scheduling class load-balancing methods. |
6746 | * | |
6747 | * BASICS | |
6748 | * | |
6749 | * The purpose of load-balancing is to achieve the same basic fairness the | |
6750 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
6751 | * time to each task. This is expressed in the following equation: | |
6752 | * | |
6753 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
6754 | * | |
6755 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
6756 | * W_i,0 is defined as: | |
6757 | * | |
6758 | * W_i,0 = \Sum_j w_i,j (2) | |
6759 | * | |
6760 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
1c3de5e1 | 6761 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
6762 | * |
6763 | * The weight average is an exponential decay average of the instantaneous | |
6764 | * weight: | |
6765 | * | |
6766 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
6767 | * | |
ced549fa | 6768 | * C_i is the compute capacity of cpu i, typically it is the |
e9c84cb8 PZ |
6769 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
6770 | * can also include other factors [XXX]. | |
6771 | * | |
6772 | * To achieve this balance we define a measure of imbalance which follows | |
6773 | * directly from (1): | |
6774 | * | |
ced549fa | 6775 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
6776 | * |
6777 | * We them move tasks around to minimize the imbalance. In the continuous | |
6778 | * function space it is obvious this converges, in the discrete case we get | |
6779 | * a few fun cases generally called infeasible weight scenarios. | |
6780 | * | |
6781 | * [XXX expand on: | |
6782 | * - infeasible weights; | |
6783 | * - local vs global optima in the discrete case. ] | |
6784 | * | |
6785 | * | |
6786 | * SCHED DOMAINS | |
6787 | * | |
6788 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
6789 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
6790 | * topology where each level pairs two lower groups (or better). This results | |
6791 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
6792 | * tree to only the first of the previous level and we decrease the frequency | |
6793 | * of load-balance at each level inv. proportional to the number of cpus in | |
6794 | * the groups. | |
6795 | * | |
6796 | * This yields: | |
6797 | * | |
6798 | * log_2 n 1 n | |
6799 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
6800 | * i = 0 2^i 2^i | |
6801 | * `- size of each group | |
6802 | * | | `- number of cpus doing load-balance | |
6803 | * | `- freq | |
6804 | * `- sum over all levels | |
6805 | * | |
6806 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
6807 | * this makes (5) the runtime complexity of the balancer. | |
6808 | * | |
6809 | * An important property here is that each CPU is still (indirectly) connected | |
6810 | * to every other cpu in at most O(log n) steps: | |
6811 | * | |
6812 | * The adjacency matrix of the resulting graph is given by: | |
6813 | * | |
97a7142f | 6814 | * log_2 n |
e9c84cb8 PZ |
6815 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
6816 | * k = 0 | |
6817 | * | |
6818 | * And you'll find that: | |
6819 | * | |
6820 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
6821 | * | |
6822 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
6823 | * The task movement gives a factor of O(m), giving a convergence complexity | |
6824 | * of: | |
6825 | * | |
6826 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
6827 | * | |
6828 | * | |
6829 | * WORK CONSERVING | |
6830 | * | |
6831 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
6832 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
6833 | * tree itself instead of relying on other CPUs to bring it work. | |
6834 | * | |
6835 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
6836 | * time. | |
6837 | * | |
6838 | * [XXX more?] | |
6839 | * | |
6840 | * | |
6841 | * CGROUPS | |
6842 | * | |
6843 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
6844 | * | |
6845 | * s_k,i | |
6846 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
6847 | * S_k | |
6848 | * | |
6849 | * Where | |
6850 | * | |
6851 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
6852 | * | |
6853 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
6854 | * | |
6855 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
6856 | * property. | |
6857 | * | |
6858 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
6859 | * rewrite all of this once again.] | |
97a7142f | 6860 | */ |
bf0f6f24 | 6861 | |
ed387b78 HS |
6862 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
6863 | ||
0ec8aa00 PZ |
6864 | enum fbq_type { regular, remote, all }; |
6865 | ||
ddcdf6e7 | 6866 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 6867 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
6868 | #define LBF_DST_PINNED 0x04 |
6869 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
6870 | |
6871 | struct lb_env { | |
6872 | struct sched_domain *sd; | |
6873 | ||
ddcdf6e7 | 6874 | struct rq *src_rq; |
85c1e7da | 6875 | int src_cpu; |
ddcdf6e7 PZ |
6876 | |
6877 | int dst_cpu; | |
6878 | struct rq *dst_rq; | |
6879 | ||
88b8dac0 SV |
6880 | struct cpumask *dst_grpmask; |
6881 | int new_dst_cpu; | |
ddcdf6e7 | 6882 | enum cpu_idle_type idle; |
bd939f45 | 6883 | long imbalance; |
b9403130 MW |
6884 | /* The set of CPUs under consideration for load-balancing */ |
6885 | struct cpumask *cpus; | |
6886 | ||
ddcdf6e7 | 6887 | unsigned int flags; |
367456c7 PZ |
6888 | |
6889 | unsigned int loop; | |
6890 | unsigned int loop_break; | |
6891 | unsigned int loop_max; | |
0ec8aa00 PZ |
6892 | |
6893 | enum fbq_type fbq_type; | |
163122b7 | 6894 | struct list_head tasks; |
ddcdf6e7 PZ |
6895 | }; |
6896 | ||
029632fb PZ |
6897 | /* |
6898 | * Is this task likely cache-hot: | |
6899 | */ | |
5d5e2b1b | 6900 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
6901 | { |
6902 | s64 delta; | |
6903 | ||
e5673f28 KT |
6904 | lockdep_assert_held(&env->src_rq->lock); |
6905 | ||
029632fb PZ |
6906 | if (p->sched_class != &fair_sched_class) |
6907 | return 0; | |
6908 | ||
6909 | if (unlikely(p->policy == SCHED_IDLE)) | |
6910 | return 0; | |
6911 | ||
6912 | /* | |
6913 | * Buddy candidates are cache hot: | |
6914 | */ | |
5d5e2b1b | 6915 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
6916 | (&p->se == cfs_rq_of(&p->se)->next || |
6917 | &p->se == cfs_rq_of(&p->se)->last)) | |
6918 | return 1; | |
6919 | ||
6920 | if (sysctl_sched_migration_cost == -1) | |
6921 | return 1; | |
6922 | if (sysctl_sched_migration_cost == 0) | |
6923 | return 0; | |
6924 | ||
5d5e2b1b | 6925 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
6926 | |
6927 | return delta < (s64)sysctl_sched_migration_cost; | |
6928 | } | |
6929 | ||
3a7053b3 | 6930 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 6931 | /* |
2a1ed24c SD |
6932 | * Returns 1, if task migration degrades locality |
6933 | * Returns 0, if task migration improves locality i.e migration preferred. | |
6934 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 6935 | */ |
2a1ed24c | 6936 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 6937 | { |
b1ad065e | 6938 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
c1ceac62 | 6939 | unsigned long src_faults, dst_faults; |
3a7053b3 MG |
6940 | int src_nid, dst_nid; |
6941 | ||
2a595721 | 6942 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
6943 | return -1; |
6944 | ||
c3b9bc5b | 6945 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 6946 | return -1; |
7a0f3083 MG |
6947 | |
6948 | src_nid = cpu_to_node(env->src_cpu); | |
6949 | dst_nid = cpu_to_node(env->dst_cpu); | |
6950 | ||
83e1d2cd | 6951 | if (src_nid == dst_nid) |
2a1ed24c | 6952 | return -1; |
7a0f3083 | 6953 | |
2a1ed24c SD |
6954 | /* Migrating away from the preferred node is always bad. */ |
6955 | if (src_nid == p->numa_preferred_nid) { | |
6956 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
6957 | return 1; | |
6958 | else | |
6959 | return -1; | |
6960 | } | |
b1ad065e | 6961 | |
c1ceac62 RR |
6962 | /* Encourage migration to the preferred node. */ |
6963 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 6964 | return 0; |
b1ad065e | 6965 | |
739294fb RR |
6966 | /* Leaving a core idle is often worse than degrading locality. */ |
6967 | if (env->idle != CPU_NOT_IDLE) | |
6968 | return -1; | |
6969 | ||
c1ceac62 RR |
6970 | if (numa_group) { |
6971 | src_faults = group_faults(p, src_nid); | |
6972 | dst_faults = group_faults(p, dst_nid); | |
6973 | } else { | |
6974 | src_faults = task_faults(p, src_nid); | |
6975 | dst_faults = task_faults(p, dst_nid); | |
b1ad065e RR |
6976 | } |
6977 | ||
c1ceac62 | 6978 | return dst_faults < src_faults; |
7a0f3083 MG |
6979 | } |
6980 | ||
3a7053b3 | 6981 | #else |
2a1ed24c | 6982 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
6983 | struct lb_env *env) |
6984 | { | |
2a1ed24c | 6985 | return -1; |
7a0f3083 | 6986 | } |
3a7053b3 MG |
6987 | #endif |
6988 | ||
1e3c88bd PZ |
6989 | /* |
6990 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
6991 | */ | |
6992 | static | |
8e45cb54 | 6993 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 6994 | { |
2a1ed24c | 6995 | int tsk_cache_hot; |
e5673f28 KT |
6996 | |
6997 | lockdep_assert_held(&env->src_rq->lock); | |
6998 | ||
1e3c88bd PZ |
6999 | /* |
7000 | * We do not migrate tasks that are: | |
d3198084 | 7001 | * 1) throttled_lb_pair, or |
1e3c88bd | 7002 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
7003 | * 3) running (obviously), or |
7004 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 7005 | */ |
d3198084 JK |
7006 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
7007 | return 0; | |
7008 | ||
0c98d344 | 7009 | if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) { |
e02e60c1 | 7010 | int cpu; |
88b8dac0 | 7011 | |
ae92882e | 7012 | schedstat_inc(p->se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 7013 | |
6263322c PZ |
7014 | env->flags |= LBF_SOME_PINNED; |
7015 | ||
88b8dac0 SV |
7016 | /* |
7017 | * Remember if this task can be migrated to any other cpu in | |
7018 | * our sched_group. We may want to revisit it if we couldn't | |
7019 | * meet load balance goals by pulling other tasks on src_cpu. | |
7020 | * | |
65a4433a JH |
7021 | * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have |
7022 | * already computed one in current iteration. | |
88b8dac0 | 7023 | */ |
65a4433a | 7024 | if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
7025 | return 0; |
7026 | ||
e02e60c1 JK |
7027 | /* Prevent to re-select dst_cpu via env's cpus */ |
7028 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
0c98d344 | 7029 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) { |
6263322c | 7030 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
7031 | env->new_dst_cpu = cpu; |
7032 | break; | |
7033 | } | |
88b8dac0 | 7034 | } |
e02e60c1 | 7035 | |
1e3c88bd PZ |
7036 | return 0; |
7037 | } | |
88b8dac0 SV |
7038 | |
7039 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 7040 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 7041 | |
ddcdf6e7 | 7042 | if (task_running(env->src_rq, p)) { |
ae92882e | 7043 | schedstat_inc(p->se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
7044 | return 0; |
7045 | } | |
7046 | ||
7047 | /* | |
7048 | * Aggressive migration if: | |
3a7053b3 MG |
7049 | * 1) destination numa is preferred |
7050 | * 2) task is cache cold, or | |
7051 | * 3) too many balance attempts have failed. | |
1e3c88bd | 7052 | */ |
2a1ed24c SD |
7053 | tsk_cache_hot = migrate_degrades_locality(p, env); |
7054 | if (tsk_cache_hot == -1) | |
7055 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 7056 | |
2a1ed24c | 7057 | if (tsk_cache_hot <= 0 || |
7a96c231 | 7058 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 7059 | if (tsk_cache_hot == 1) { |
ae92882e JP |
7060 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
7061 | schedstat_inc(p->se.statistics.nr_forced_migrations); | |
3a7053b3 | 7062 | } |
1e3c88bd PZ |
7063 | return 1; |
7064 | } | |
7065 | ||
ae92882e | 7066 | schedstat_inc(p->se.statistics.nr_failed_migrations_hot); |
4e2dcb73 | 7067 | return 0; |
1e3c88bd PZ |
7068 | } |
7069 | ||
897c395f | 7070 | /* |
163122b7 KT |
7071 | * detach_task() -- detach the task for the migration specified in env |
7072 | */ | |
7073 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
7074 | { | |
7075 | lockdep_assert_held(&env->src_rq->lock); | |
7076 | ||
163122b7 | 7077 | p->on_rq = TASK_ON_RQ_MIGRATING; |
5704ac0a | 7078 | deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); |
163122b7 KT |
7079 | set_task_cpu(p, env->dst_cpu); |
7080 | } | |
7081 | ||
897c395f | 7082 | /* |
e5673f28 | 7083 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 7084 | * part of active balancing operations within "domain". |
897c395f | 7085 | * |
e5673f28 | 7086 | * Returns a task if successful and NULL otherwise. |
897c395f | 7087 | */ |
e5673f28 | 7088 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f | 7089 | { |
93824900 | 7090 | struct task_struct *p; |
897c395f | 7091 | |
e5673f28 KT |
7092 | lockdep_assert_held(&env->src_rq->lock); |
7093 | ||
93824900 UR |
7094 | list_for_each_entry_reverse(p, |
7095 | &env->src_rq->cfs_tasks, se.group_node) { | |
367456c7 PZ |
7096 | if (!can_migrate_task(p, env)) |
7097 | continue; | |
897c395f | 7098 | |
163122b7 | 7099 | detach_task(p, env); |
e5673f28 | 7100 | |
367456c7 | 7101 | /* |
e5673f28 | 7102 | * Right now, this is only the second place where |
163122b7 | 7103 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 7104 | * so we can safely collect stats here rather than |
163122b7 | 7105 | * inside detach_tasks(). |
367456c7 | 7106 | */ |
ae92882e | 7107 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 7108 | return p; |
897c395f | 7109 | } |
e5673f28 | 7110 | return NULL; |
897c395f PZ |
7111 | } |
7112 | ||
eb95308e PZ |
7113 | static const unsigned int sched_nr_migrate_break = 32; |
7114 | ||
5d6523eb | 7115 | /* |
163122b7 KT |
7116 | * detach_tasks() -- tries to detach up to imbalance weighted load from |
7117 | * busiest_rq, as part of a balancing operation within domain "sd". | |
5d6523eb | 7118 | * |
163122b7 | 7119 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 7120 | */ |
163122b7 | 7121 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 7122 | { |
5d6523eb PZ |
7123 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
7124 | struct task_struct *p; | |
367456c7 | 7125 | unsigned long load; |
163122b7 KT |
7126 | int detached = 0; |
7127 | ||
7128 | lockdep_assert_held(&env->src_rq->lock); | |
1e3c88bd | 7129 | |
bd939f45 | 7130 | if (env->imbalance <= 0) |
5d6523eb | 7131 | return 0; |
1e3c88bd | 7132 | |
5d6523eb | 7133 | while (!list_empty(tasks)) { |
985d3a4c YD |
7134 | /* |
7135 | * We don't want to steal all, otherwise we may be treated likewise, | |
7136 | * which could at worst lead to a livelock crash. | |
7137 | */ | |
7138 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
7139 | break; | |
7140 | ||
93824900 | 7141 | p = list_last_entry(tasks, struct task_struct, se.group_node); |
1e3c88bd | 7142 | |
367456c7 PZ |
7143 | env->loop++; |
7144 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 7145 | if (env->loop > env->loop_max) |
367456c7 | 7146 | break; |
5d6523eb PZ |
7147 | |
7148 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 7149 | if (env->loop > env->loop_break) { |
eb95308e | 7150 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 7151 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 7152 | break; |
a195f004 | 7153 | } |
1e3c88bd | 7154 | |
d3198084 | 7155 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
7156 | goto next; |
7157 | ||
7158 | load = task_h_load(p); | |
5d6523eb | 7159 | |
eb95308e | 7160 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
7161 | goto next; |
7162 | ||
bd939f45 | 7163 | if ((load / 2) > env->imbalance) |
367456c7 | 7164 | goto next; |
1e3c88bd | 7165 | |
163122b7 KT |
7166 | detach_task(p, env); |
7167 | list_add(&p->se.group_node, &env->tasks); | |
7168 | ||
7169 | detached++; | |
bd939f45 | 7170 | env->imbalance -= load; |
1e3c88bd PZ |
7171 | |
7172 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
7173 | /* |
7174 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 7175 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
7176 | * the critical section. |
7177 | */ | |
5d6523eb | 7178 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 7179 | break; |
1e3c88bd PZ |
7180 | #endif |
7181 | ||
ee00e66f PZ |
7182 | /* |
7183 | * We only want to steal up to the prescribed amount of | |
7184 | * weighted load. | |
7185 | */ | |
bd939f45 | 7186 | if (env->imbalance <= 0) |
ee00e66f | 7187 | break; |
367456c7 PZ |
7188 | |
7189 | continue; | |
7190 | next: | |
93824900 | 7191 | list_move(&p->se.group_node, tasks); |
1e3c88bd | 7192 | } |
5d6523eb | 7193 | |
1e3c88bd | 7194 | /* |
163122b7 KT |
7195 | * Right now, this is one of only two places we collect this stat |
7196 | * so we can safely collect detach_one_task() stats here rather | |
7197 | * than inside detach_one_task(). | |
1e3c88bd | 7198 | */ |
ae92882e | 7199 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 7200 | |
163122b7 KT |
7201 | return detached; |
7202 | } | |
7203 | ||
7204 | /* | |
7205 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
7206 | */ | |
7207 | static void attach_task(struct rq *rq, struct task_struct *p) | |
7208 | { | |
7209 | lockdep_assert_held(&rq->lock); | |
7210 | ||
7211 | BUG_ON(task_rq(p) != rq); | |
5704ac0a | 7212 | activate_task(rq, p, ENQUEUE_NOCLOCK); |
3ea94de1 | 7213 | p->on_rq = TASK_ON_RQ_QUEUED; |
163122b7 KT |
7214 | check_preempt_curr(rq, p, 0); |
7215 | } | |
7216 | ||
7217 | /* | |
7218 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
7219 | * its new rq. | |
7220 | */ | |
7221 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
7222 | { | |
8a8c69c3 PZ |
7223 | struct rq_flags rf; |
7224 | ||
7225 | rq_lock(rq, &rf); | |
5704ac0a | 7226 | update_rq_clock(rq); |
163122b7 | 7227 | attach_task(rq, p); |
8a8c69c3 | 7228 | rq_unlock(rq, &rf); |
163122b7 KT |
7229 | } |
7230 | ||
7231 | /* | |
7232 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
7233 | * new rq. | |
7234 | */ | |
7235 | static void attach_tasks(struct lb_env *env) | |
7236 | { | |
7237 | struct list_head *tasks = &env->tasks; | |
7238 | struct task_struct *p; | |
8a8c69c3 | 7239 | struct rq_flags rf; |
163122b7 | 7240 | |
8a8c69c3 | 7241 | rq_lock(env->dst_rq, &rf); |
5704ac0a | 7242 | update_rq_clock(env->dst_rq); |
163122b7 KT |
7243 | |
7244 | while (!list_empty(tasks)) { | |
7245 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
7246 | list_del_init(&p->se.group_node); | |
1e3c88bd | 7247 | |
163122b7 KT |
7248 | attach_task(env->dst_rq, p); |
7249 | } | |
7250 | ||
8a8c69c3 | 7251 | rq_unlock(env->dst_rq, &rf); |
1e3c88bd PZ |
7252 | } |
7253 | ||
230059de | 7254 | #ifdef CONFIG_FAIR_GROUP_SCHED |
a9e7f654 TH |
7255 | |
7256 | static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) | |
7257 | { | |
7258 | if (cfs_rq->load.weight) | |
7259 | return false; | |
7260 | ||
7261 | if (cfs_rq->avg.load_sum) | |
7262 | return false; | |
7263 | ||
7264 | if (cfs_rq->avg.util_sum) | |
7265 | return false; | |
7266 | ||
1ea6c46a | 7267 | if (cfs_rq->avg.runnable_load_sum) |
a9e7f654 TH |
7268 | return false; |
7269 | ||
7270 | return true; | |
7271 | } | |
7272 | ||
48a16753 | 7273 | static void update_blocked_averages(int cpu) |
9e3081ca | 7274 | { |
9e3081ca | 7275 | struct rq *rq = cpu_rq(cpu); |
a9e7f654 | 7276 | struct cfs_rq *cfs_rq, *pos; |
8a8c69c3 | 7277 | struct rq_flags rf; |
9e3081ca | 7278 | |
8a8c69c3 | 7279 | rq_lock_irqsave(rq, &rf); |
48a16753 | 7280 | update_rq_clock(rq); |
9d89c257 | 7281 | |
9763b67f PZ |
7282 | /* |
7283 | * Iterates the task_group tree in a bottom up fashion, see | |
7284 | * list_add_leaf_cfs_rq() for details. | |
7285 | */ | |
a9e7f654 | 7286 | for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { |
bc427898 VG |
7287 | struct sched_entity *se; |
7288 | ||
9d89c257 YD |
7289 | /* throttled entities do not contribute to load */ |
7290 | if (throttled_hierarchy(cfs_rq)) | |
7291 | continue; | |
48a16753 | 7292 | |
3a123bbb | 7293 | if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq)) |
9d89c257 | 7294 | update_tg_load_avg(cfs_rq, 0); |
4e516076 | 7295 | |
bc427898 VG |
7296 | /* Propagate pending load changes to the parent, if any: */ |
7297 | se = cfs_rq->tg->se[cpu]; | |
7298 | if (se && !skip_blocked_update(se)) | |
88c0616e | 7299 | update_load_avg(cfs_rq_of(se), se, 0); |
a9e7f654 TH |
7300 | |
7301 | /* | |
7302 | * There can be a lot of idle CPU cgroups. Don't let fully | |
7303 | * decayed cfs_rqs linger on the list. | |
7304 | */ | |
7305 | if (cfs_rq_is_decayed(cfs_rq)) | |
7306 | list_del_leaf_cfs_rq(cfs_rq); | |
9d89c257 | 7307 | } |
8a8c69c3 | 7308 | rq_unlock_irqrestore(rq, &rf); |
9e3081ca PZ |
7309 | } |
7310 | ||
9763b67f | 7311 | /* |
68520796 | 7312 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
7313 | * This needs to be done in a top-down fashion because the load of a child |
7314 | * group is a fraction of its parents load. | |
7315 | */ | |
68520796 | 7316 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 7317 | { |
68520796 VD |
7318 | struct rq *rq = rq_of(cfs_rq); |
7319 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 7320 | unsigned long now = jiffies; |
68520796 | 7321 | unsigned long load; |
a35b6466 | 7322 | |
68520796 | 7323 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
7324 | return; |
7325 | ||
68520796 VD |
7326 | cfs_rq->h_load_next = NULL; |
7327 | for_each_sched_entity(se) { | |
7328 | cfs_rq = cfs_rq_of(se); | |
7329 | cfs_rq->h_load_next = se; | |
7330 | if (cfs_rq->last_h_load_update == now) | |
7331 | break; | |
7332 | } | |
a35b6466 | 7333 | |
68520796 | 7334 | if (!se) { |
7ea241af | 7335 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
7336 | cfs_rq->last_h_load_update = now; |
7337 | } | |
7338 | ||
7339 | while ((se = cfs_rq->h_load_next) != NULL) { | |
7340 | load = cfs_rq->h_load; | |
7ea241af YD |
7341 | load = div64_ul(load * se->avg.load_avg, |
7342 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
7343 | cfs_rq = group_cfs_rq(se); |
7344 | cfs_rq->h_load = load; | |
7345 | cfs_rq->last_h_load_update = now; | |
7346 | } | |
9763b67f PZ |
7347 | } |
7348 | ||
367456c7 | 7349 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 7350 | { |
367456c7 | 7351 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 7352 | |
68520796 | 7353 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 7354 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 7355 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
7356 | } |
7357 | #else | |
48a16753 | 7358 | static inline void update_blocked_averages(int cpu) |
9e3081ca | 7359 | { |
6c1d47c0 VG |
7360 | struct rq *rq = cpu_rq(cpu); |
7361 | struct cfs_rq *cfs_rq = &rq->cfs; | |
8a8c69c3 | 7362 | struct rq_flags rf; |
6c1d47c0 | 7363 | |
8a8c69c3 | 7364 | rq_lock_irqsave(rq, &rf); |
6c1d47c0 | 7365 | update_rq_clock(rq); |
3a123bbb | 7366 | update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq); |
8a8c69c3 | 7367 | rq_unlock_irqrestore(rq, &rf); |
9e3081ca PZ |
7368 | } |
7369 | ||
367456c7 | 7370 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 7371 | { |
9d89c257 | 7372 | return p->se.avg.load_avg; |
1e3c88bd | 7373 | } |
230059de | 7374 | #endif |
1e3c88bd | 7375 | |
1e3c88bd | 7376 | /********** Helpers for find_busiest_group ************************/ |
caeb178c RR |
7377 | |
7378 | enum group_type { | |
7379 | group_other = 0, | |
7380 | group_imbalanced, | |
7381 | group_overloaded, | |
7382 | }; | |
7383 | ||
1e3c88bd PZ |
7384 | /* |
7385 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
7386 | */ | |
7387 | struct sg_lb_stats { | |
7388 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
7389 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 7390 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 7391 | unsigned long load_per_task; |
63b2ca30 | 7392 | unsigned long group_capacity; |
9e91d61d | 7393 | unsigned long group_util; /* Total utilization of the group */ |
147c5fc2 | 7394 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
147c5fc2 PZ |
7395 | unsigned int idle_cpus; |
7396 | unsigned int group_weight; | |
caeb178c | 7397 | enum group_type group_type; |
ea67821b | 7398 | int group_no_capacity; |
0ec8aa00 PZ |
7399 | #ifdef CONFIG_NUMA_BALANCING |
7400 | unsigned int nr_numa_running; | |
7401 | unsigned int nr_preferred_running; | |
7402 | #endif | |
1e3c88bd PZ |
7403 | }; |
7404 | ||
56cf515b JK |
7405 | /* |
7406 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
7407 | * during load balancing. | |
7408 | */ | |
7409 | struct sd_lb_stats { | |
7410 | struct sched_group *busiest; /* Busiest group in this sd */ | |
7411 | struct sched_group *local; /* Local group in this sd */ | |
90001d67 | 7412 | unsigned long total_running; |
56cf515b | 7413 | unsigned long total_load; /* Total load of all groups in sd */ |
63b2ca30 | 7414 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b JK |
7415 | unsigned long avg_load; /* Average load across all groups in sd */ |
7416 | ||
56cf515b | 7417 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 7418 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
7419 | }; |
7420 | ||
147c5fc2 PZ |
7421 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
7422 | { | |
7423 | /* | |
7424 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
7425 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
7426 | * We must however clear busiest_stat::avg_load because | |
7427 | * update_sd_pick_busiest() reads this before assignment. | |
7428 | */ | |
7429 | *sds = (struct sd_lb_stats){ | |
7430 | .busiest = NULL, | |
7431 | .local = NULL, | |
90001d67 | 7432 | .total_running = 0UL, |
147c5fc2 | 7433 | .total_load = 0UL, |
63b2ca30 | 7434 | .total_capacity = 0UL, |
147c5fc2 PZ |
7435 | .busiest_stat = { |
7436 | .avg_load = 0UL, | |
caeb178c RR |
7437 | .sum_nr_running = 0, |
7438 | .group_type = group_other, | |
147c5fc2 PZ |
7439 | }, |
7440 | }; | |
7441 | } | |
7442 | ||
1e3c88bd PZ |
7443 | /** |
7444 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
7445 | * @sd: The sched_domain whose load_idx is to be obtained. | |
ed1b7732 | 7446 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. |
e69f6186 YB |
7447 | * |
7448 | * Return: The load index. | |
1e3c88bd PZ |
7449 | */ |
7450 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
7451 | enum cpu_idle_type idle) | |
7452 | { | |
7453 | int load_idx; | |
7454 | ||
7455 | switch (idle) { | |
7456 | case CPU_NOT_IDLE: | |
7457 | load_idx = sd->busy_idx; | |
7458 | break; | |
7459 | ||
7460 | case CPU_NEWLY_IDLE: | |
7461 | load_idx = sd->newidle_idx; | |
7462 | break; | |
7463 | default: | |
7464 | load_idx = sd->idle_idx; | |
7465 | break; | |
7466 | } | |
7467 | ||
7468 | return load_idx; | |
7469 | } | |
7470 | ||
ced549fa | 7471 | static unsigned long scale_rt_capacity(int cpu) |
1e3c88bd PZ |
7472 | { |
7473 | struct rq *rq = cpu_rq(cpu); | |
b5b4860d | 7474 | u64 total, used, age_stamp, avg; |
cadefd3d | 7475 | s64 delta; |
1e3c88bd | 7476 | |
b654f7de PZ |
7477 | /* |
7478 | * Since we're reading these variables without serialization make sure | |
7479 | * we read them once before doing sanity checks on them. | |
7480 | */ | |
316c1608 JL |
7481 | age_stamp = READ_ONCE(rq->age_stamp); |
7482 | avg = READ_ONCE(rq->rt_avg); | |
cebde6d6 | 7483 | delta = __rq_clock_broken(rq) - age_stamp; |
b654f7de | 7484 | |
cadefd3d PZ |
7485 | if (unlikely(delta < 0)) |
7486 | delta = 0; | |
7487 | ||
7488 | total = sched_avg_period() + delta; | |
aa483808 | 7489 | |
b5b4860d | 7490 | used = div_u64(avg, total); |
1e3c88bd | 7491 | |
b5b4860d VG |
7492 | if (likely(used < SCHED_CAPACITY_SCALE)) |
7493 | return SCHED_CAPACITY_SCALE - used; | |
1e3c88bd | 7494 | |
b5b4860d | 7495 | return 1; |
1e3c88bd PZ |
7496 | } |
7497 | ||
ced549fa | 7498 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 7499 | { |
8cd5601c | 7500 | unsigned long capacity = arch_scale_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
7501 | struct sched_group *sdg = sd->groups; |
7502 | ||
ca6d75e6 | 7503 | cpu_rq(cpu)->cpu_capacity_orig = capacity; |
9d5efe05 | 7504 | |
ced549fa | 7505 | capacity *= scale_rt_capacity(cpu); |
ca8ce3d0 | 7506 | capacity >>= SCHED_CAPACITY_SHIFT; |
1e3c88bd | 7507 | |
ced549fa NP |
7508 | if (!capacity) |
7509 | capacity = 1; | |
1e3c88bd | 7510 | |
ced549fa NP |
7511 | cpu_rq(cpu)->cpu_capacity = capacity; |
7512 | sdg->sgc->capacity = capacity; | |
bf475ce0 | 7513 | sdg->sgc->min_capacity = capacity; |
1e3c88bd PZ |
7514 | } |
7515 | ||
63b2ca30 | 7516 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
7517 | { |
7518 | struct sched_domain *child = sd->child; | |
7519 | struct sched_group *group, *sdg = sd->groups; | |
bf475ce0 | 7520 | unsigned long capacity, min_capacity; |
4ec4412e VG |
7521 | unsigned long interval; |
7522 | ||
7523 | interval = msecs_to_jiffies(sd->balance_interval); | |
7524 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 7525 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
7526 | |
7527 | if (!child) { | |
ced549fa | 7528 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
7529 | return; |
7530 | } | |
7531 | ||
dc7ff76e | 7532 | capacity = 0; |
bf475ce0 | 7533 | min_capacity = ULONG_MAX; |
1e3c88bd | 7534 | |
74a5ce20 PZ |
7535 | if (child->flags & SD_OVERLAP) { |
7536 | /* | |
7537 | * SD_OVERLAP domains cannot assume that child groups | |
7538 | * span the current group. | |
7539 | */ | |
7540 | ||
ae4df9d6 | 7541 | for_each_cpu(cpu, sched_group_span(sdg)) { |
63b2ca30 | 7542 | struct sched_group_capacity *sgc; |
9abf24d4 | 7543 | struct rq *rq = cpu_rq(cpu); |
863bffc8 | 7544 | |
9abf24d4 | 7545 | /* |
63b2ca30 | 7546 | * build_sched_domains() -> init_sched_groups_capacity() |
9abf24d4 SD |
7547 | * gets here before we've attached the domains to the |
7548 | * runqueues. | |
7549 | * | |
ced549fa NP |
7550 | * Use capacity_of(), which is set irrespective of domains |
7551 | * in update_cpu_capacity(). | |
9abf24d4 | 7552 | * |
dc7ff76e | 7553 | * This avoids capacity from being 0 and |
9abf24d4 | 7554 | * causing divide-by-zero issues on boot. |
9abf24d4 SD |
7555 | */ |
7556 | if (unlikely(!rq->sd)) { | |
ced549fa | 7557 | capacity += capacity_of(cpu); |
bf475ce0 MR |
7558 | } else { |
7559 | sgc = rq->sd->groups->sgc; | |
7560 | capacity += sgc->capacity; | |
9abf24d4 | 7561 | } |
863bffc8 | 7562 | |
bf475ce0 | 7563 | min_capacity = min(capacity, min_capacity); |
863bffc8 | 7564 | } |
74a5ce20 PZ |
7565 | } else { |
7566 | /* | |
7567 | * !SD_OVERLAP domains can assume that child groups | |
7568 | * span the current group. | |
97a7142f | 7569 | */ |
74a5ce20 PZ |
7570 | |
7571 | group = child->groups; | |
7572 | do { | |
bf475ce0 MR |
7573 | struct sched_group_capacity *sgc = group->sgc; |
7574 | ||
7575 | capacity += sgc->capacity; | |
7576 | min_capacity = min(sgc->min_capacity, min_capacity); | |
74a5ce20 PZ |
7577 | group = group->next; |
7578 | } while (group != child->groups); | |
7579 | } | |
1e3c88bd | 7580 | |
63b2ca30 | 7581 | sdg->sgc->capacity = capacity; |
bf475ce0 | 7582 | sdg->sgc->min_capacity = min_capacity; |
1e3c88bd PZ |
7583 | } |
7584 | ||
9d5efe05 | 7585 | /* |
ea67821b VG |
7586 | * Check whether the capacity of the rq has been noticeably reduced by side |
7587 | * activity. The imbalance_pct is used for the threshold. | |
7588 | * Return true is the capacity is reduced | |
9d5efe05 SV |
7589 | */ |
7590 | static inline int | |
ea67821b | 7591 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 7592 | { |
ea67821b VG |
7593 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
7594 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
7595 | } |
7596 | ||
30ce5dab PZ |
7597 | /* |
7598 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
0c98d344 | 7599 | * groups is inadequate due to ->cpus_allowed constraints. |
30ce5dab PZ |
7600 | * |
7601 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
7602 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
7603 | * Something like: | |
7604 | * | |
2b4d5b25 IM |
7605 | * { 0 1 2 3 } { 4 5 6 7 } |
7606 | * * * * * | |
30ce5dab PZ |
7607 | * |
7608 | * If we were to balance group-wise we'd place two tasks in the first group and | |
7609 | * two tasks in the second group. Clearly this is undesired as it will overload | |
7610 | * cpu 3 and leave one of the cpus in the second group unused. | |
7611 | * | |
7612 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
7613 | * by noticing the lower domain failed to reach balance and had difficulty |
7614 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
7615 | * |
7616 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 7617 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 7618 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
7619 | * to create an effective group imbalance. |
7620 | * | |
7621 | * This is a somewhat tricky proposition since the next run might not find the | |
7622 | * group imbalance and decide the groups need to be balanced again. A most | |
7623 | * subtle and fragile situation. | |
7624 | */ | |
7625 | ||
6263322c | 7626 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 7627 | { |
63b2ca30 | 7628 | return group->sgc->imbalance; |
30ce5dab PZ |
7629 | } |
7630 | ||
b37d9316 | 7631 | /* |
ea67821b VG |
7632 | * group_has_capacity returns true if the group has spare capacity that could |
7633 | * be used by some tasks. | |
7634 | * We consider that a group has spare capacity if the * number of task is | |
9e91d61d DE |
7635 | * smaller than the number of CPUs or if the utilization is lower than the |
7636 | * available capacity for CFS tasks. | |
ea67821b VG |
7637 | * For the latter, we use a threshold to stabilize the state, to take into |
7638 | * account the variance of the tasks' load and to return true if the available | |
7639 | * capacity in meaningful for the load balancer. | |
7640 | * As an example, an available capacity of 1% can appear but it doesn't make | |
7641 | * any benefit for the load balance. | |
b37d9316 | 7642 | */ |
ea67821b VG |
7643 | static inline bool |
7644 | group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs) | |
b37d9316 | 7645 | { |
ea67821b VG |
7646 | if (sgs->sum_nr_running < sgs->group_weight) |
7647 | return true; | |
c61037e9 | 7648 | |
ea67821b | 7649 | if ((sgs->group_capacity * 100) > |
9e91d61d | 7650 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 7651 | return true; |
b37d9316 | 7652 | |
ea67821b VG |
7653 | return false; |
7654 | } | |
7655 | ||
7656 | /* | |
7657 | * group_is_overloaded returns true if the group has more tasks than it can | |
7658 | * handle. | |
7659 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
7660 | * with the exact right number of tasks, has no more spare capacity but is not | |
7661 | * overloaded so both group_has_capacity and group_is_overloaded return | |
7662 | * false. | |
7663 | */ | |
7664 | static inline bool | |
7665 | group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs) | |
7666 | { | |
7667 | if (sgs->sum_nr_running <= sgs->group_weight) | |
7668 | return false; | |
b37d9316 | 7669 | |
ea67821b | 7670 | if ((sgs->group_capacity * 100) < |
9e91d61d | 7671 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 7672 | return true; |
b37d9316 | 7673 | |
ea67821b | 7674 | return false; |
b37d9316 PZ |
7675 | } |
7676 | ||
9e0994c0 MR |
7677 | /* |
7678 | * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller | |
7679 | * per-CPU capacity than sched_group ref. | |
7680 | */ | |
7681 | static inline bool | |
7682 | group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref) | |
7683 | { | |
7684 | return sg->sgc->min_capacity * capacity_margin < | |
7685 | ref->sgc->min_capacity * 1024; | |
7686 | } | |
7687 | ||
79a89f92 LY |
7688 | static inline enum |
7689 | group_type group_classify(struct sched_group *group, | |
7690 | struct sg_lb_stats *sgs) | |
caeb178c | 7691 | { |
ea67821b | 7692 | if (sgs->group_no_capacity) |
caeb178c RR |
7693 | return group_overloaded; |
7694 | ||
7695 | if (sg_imbalanced(group)) | |
7696 | return group_imbalanced; | |
7697 | ||
7698 | return group_other; | |
7699 | } | |
7700 | ||
1e3c88bd PZ |
7701 | /** |
7702 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 7703 | * @env: The load balancing environment. |
1e3c88bd | 7704 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 7705 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 7706 | * @local_group: Does group contain this_cpu. |
1e3c88bd | 7707 | * @sgs: variable to hold the statistics for this group. |
cd3bd4e6 | 7708 | * @overload: Indicate more than one runnable task for any CPU. |
1e3c88bd | 7709 | */ |
bd939f45 PZ |
7710 | static inline void update_sg_lb_stats(struct lb_env *env, |
7711 | struct sched_group *group, int load_idx, | |
4486edd1 TC |
7712 | int local_group, struct sg_lb_stats *sgs, |
7713 | bool *overload) | |
1e3c88bd | 7714 | { |
30ce5dab | 7715 | unsigned long load; |
a426f99c | 7716 | int i, nr_running; |
1e3c88bd | 7717 | |
b72ff13c PZ |
7718 | memset(sgs, 0, sizeof(*sgs)); |
7719 | ||
ae4df9d6 | 7720 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
1e3c88bd PZ |
7721 | struct rq *rq = cpu_rq(i); |
7722 | ||
1e3c88bd | 7723 | /* Bias balancing toward cpus of our domain */ |
6263322c | 7724 | if (local_group) |
04f733b4 | 7725 | load = target_load(i, load_idx); |
6263322c | 7726 | else |
1e3c88bd | 7727 | load = source_load(i, load_idx); |
1e3c88bd PZ |
7728 | |
7729 | sgs->group_load += load; | |
9e91d61d | 7730 | sgs->group_util += cpu_util(i); |
65fdac08 | 7731 | sgs->sum_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 7732 | |
a426f99c WL |
7733 | nr_running = rq->nr_running; |
7734 | if (nr_running > 1) | |
4486edd1 TC |
7735 | *overload = true; |
7736 | ||
0ec8aa00 PZ |
7737 | #ifdef CONFIG_NUMA_BALANCING |
7738 | sgs->nr_numa_running += rq->nr_numa_running; | |
7739 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
7740 | #endif | |
c7132dd6 | 7741 | sgs->sum_weighted_load += weighted_cpuload(rq); |
a426f99c WL |
7742 | /* |
7743 | * No need to call idle_cpu() if nr_running is not 0 | |
7744 | */ | |
7745 | if (!nr_running && idle_cpu(i)) | |
aae6d3dd | 7746 | sgs->idle_cpus++; |
1e3c88bd PZ |
7747 | } |
7748 | ||
63b2ca30 NP |
7749 | /* Adjust by relative CPU capacity of the group */ |
7750 | sgs->group_capacity = group->sgc->capacity; | |
ca8ce3d0 | 7751 | sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; |
1e3c88bd | 7752 | |
dd5feea1 | 7753 | if (sgs->sum_nr_running) |
38d0f770 | 7754 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 7755 | |
aae6d3dd | 7756 | sgs->group_weight = group->group_weight; |
b37d9316 | 7757 | |
ea67821b | 7758 | sgs->group_no_capacity = group_is_overloaded(env, sgs); |
79a89f92 | 7759 | sgs->group_type = group_classify(group, sgs); |
1e3c88bd PZ |
7760 | } |
7761 | ||
532cb4c4 MN |
7762 | /** |
7763 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 7764 | * @env: The load balancing environment. |
532cb4c4 MN |
7765 | * @sds: sched_domain statistics |
7766 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 7767 | * @sgs: sched_group statistics |
532cb4c4 MN |
7768 | * |
7769 | * Determine if @sg is a busier group than the previously selected | |
7770 | * busiest group. | |
e69f6186 YB |
7771 | * |
7772 | * Return: %true if @sg is a busier group than the previously selected | |
7773 | * busiest group. %false otherwise. | |
532cb4c4 | 7774 | */ |
bd939f45 | 7775 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
7776 | struct sd_lb_stats *sds, |
7777 | struct sched_group *sg, | |
bd939f45 | 7778 | struct sg_lb_stats *sgs) |
532cb4c4 | 7779 | { |
caeb178c | 7780 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 7781 | |
caeb178c | 7782 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
7783 | return true; |
7784 | ||
caeb178c RR |
7785 | if (sgs->group_type < busiest->group_type) |
7786 | return false; | |
7787 | ||
7788 | if (sgs->avg_load <= busiest->avg_load) | |
7789 | return false; | |
7790 | ||
9e0994c0 MR |
7791 | if (!(env->sd->flags & SD_ASYM_CPUCAPACITY)) |
7792 | goto asym_packing; | |
7793 | ||
7794 | /* | |
7795 | * Candidate sg has no more than one task per CPU and | |
7796 | * has higher per-CPU capacity. Migrating tasks to less | |
7797 | * capable CPUs may harm throughput. Maximize throughput, | |
7798 | * power/energy consequences are not considered. | |
7799 | */ | |
7800 | if (sgs->sum_nr_running <= sgs->group_weight && | |
7801 | group_smaller_cpu_capacity(sds->local, sg)) | |
7802 | return false; | |
7803 | ||
7804 | asym_packing: | |
caeb178c RR |
7805 | /* This is the busiest node in its class. */ |
7806 | if (!(env->sd->flags & SD_ASYM_PACKING)) | |
532cb4c4 MN |
7807 | return true; |
7808 | ||
1f621e02 SD |
7809 | /* No ASYM_PACKING if target cpu is already busy */ |
7810 | if (env->idle == CPU_NOT_IDLE) | |
7811 | return true; | |
532cb4c4 | 7812 | /* |
afe06efd TC |
7813 | * ASYM_PACKING needs to move all the work to the highest |
7814 | * prority CPUs in the group, therefore mark all groups | |
7815 | * of lower priority than ourself as busy. | |
532cb4c4 | 7816 | */ |
afe06efd TC |
7817 | if (sgs->sum_nr_running && |
7818 | sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) { | |
532cb4c4 MN |
7819 | if (!sds->busiest) |
7820 | return true; | |
7821 | ||
afe06efd TC |
7822 | /* Prefer to move from lowest priority cpu's work */ |
7823 | if (sched_asym_prefer(sds->busiest->asym_prefer_cpu, | |
7824 | sg->asym_prefer_cpu)) | |
532cb4c4 MN |
7825 | return true; |
7826 | } | |
7827 | ||
7828 | return false; | |
7829 | } | |
7830 | ||
0ec8aa00 PZ |
7831 | #ifdef CONFIG_NUMA_BALANCING |
7832 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
7833 | { | |
7834 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
7835 | return regular; | |
7836 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
7837 | return remote; | |
7838 | return all; | |
7839 | } | |
7840 | ||
7841 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
7842 | { | |
7843 | if (rq->nr_running > rq->nr_numa_running) | |
7844 | return regular; | |
7845 | if (rq->nr_running > rq->nr_preferred_running) | |
7846 | return remote; | |
7847 | return all; | |
7848 | } | |
7849 | #else | |
7850 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
7851 | { | |
7852 | return all; | |
7853 | } | |
7854 | ||
7855 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
7856 | { | |
7857 | return regular; | |
7858 | } | |
7859 | #endif /* CONFIG_NUMA_BALANCING */ | |
7860 | ||
1e3c88bd | 7861 | /** |
461819ac | 7862 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 7863 | * @env: The load balancing environment. |
1e3c88bd PZ |
7864 | * @sds: variable to hold the statistics for this sched_domain. |
7865 | */ | |
0ec8aa00 | 7866 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 7867 | { |
bd939f45 PZ |
7868 | struct sched_domain *child = env->sd->child; |
7869 | struct sched_group *sg = env->sd->groups; | |
05b40e05 | 7870 | struct sg_lb_stats *local = &sds->local_stat; |
56cf515b | 7871 | struct sg_lb_stats tmp_sgs; |
1e3c88bd | 7872 | int load_idx, prefer_sibling = 0; |
4486edd1 | 7873 | bool overload = false; |
1e3c88bd PZ |
7874 | |
7875 | if (child && child->flags & SD_PREFER_SIBLING) | |
7876 | prefer_sibling = 1; | |
7877 | ||
bd939f45 | 7878 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
7879 | |
7880 | do { | |
56cf515b | 7881 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
7882 | int local_group; |
7883 | ||
ae4df9d6 | 7884 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); |
56cf515b JK |
7885 | if (local_group) { |
7886 | sds->local = sg; | |
05b40e05 | 7887 | sgs = local; |
b72ff13c PZ |
7888 | |
7889 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
7890 | time_after_eq(jiffies, sg->sgc->next_update)) |
7891 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 7892 | } |
1e3c88bd | 7893 | |
4486edd1 TC |
7894 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs, |
7895 | &overload); | |
1e3c88bd | 7896 | |
b72ff13c PZ |
7897 | if (local_group) |
7898 | goto next_group; | |
7899 | ||
1e3c88bd PZ |
7900 | /* |
7901 | * In case the child domain prefers tasks go to siblings | |
ea67821b | 7902 | * first, lower the sg capacity so that we'll try |
75dd321d NR |
7903 | * and move all the excess tasks away. We lower the capacity |
7904 | * of a group only if the local group has the capacity to fit | |
ea67821b VG |
7905 | * these excess tasks. The extra check prevents the case where |
7906 | * you always pull from the heaviest group when it is already | |
7907 | * under-utilized (possible with a large weight task outweighs | |
7908 | * the tasks on the system). | |
1e3c88bd | 7909 | */ |
b72ff13c | 7910 | if (prefer_sibling && sds->local && |
05b40e05 SD |
7911 | group_has_capacity(env, local) && |
7912 | (sgs->sum_nr_running > local->sum_nr_running + 1)) { | |
ea67821b | 7913 | sgs->group_no_capacity = 1; |
79a89f92 | 7914 | sgs->group_type = group_classify(sg, sgs); |
cb0b9f24 | 7915 | } |
1e3c88bd | 7916 | |
b72ff13c | 7917 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 7918 | sds->busiest = sg; |
56cf515b | 7919 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
7920 | } |
7921 | ||
b72ff13c PZ |
7922 | next_group: |
7923 | /* Now, start updating sd_lb_stats */ | |
90001d67 | 7924 | sds->total_running += sgs->sum_nr_running; |
b72ff13c | 7925 | sds->total_load += sgs->group_load; |
63b2ca30 | 7926 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 7927 | |
532cb4c4 | 7928 | sg = sg->next; |
bd939f45 | 7929 | } while (sg != env->sd->groups); |
0ec8aa00 PZ |
7930 | |
7931 | if (env->sd->flags & SD_NUMA) | |
7932 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
7933 | |
7934 | if (!env->sd->parent) { | |
7935 | /* update overload indicator if we are at root domain */ | |
7936 | if (env->dst_rq->rd->overload != overload) | |
7937 | env->dst_rq->rd->overload = overload; | |
7938 | } | |
532cb4c4 MN |
7939 | } |
7940 | ||
532cb4c4 MN |
7941 | /** |
7942 | * check_asym_packing - Check to see if the group is packed into the | |
0ba42a59 | 7943 | * sched domain. |
532cb4c4 MN |
7944 | * |
7945 | * This is primarily intended to used at the sibling level. Some | |
7946 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
7947 | * case of POWER7, it can move to lower SMT modes only when higher | |
7948 | * threads are idle. When in lower SMT modes, the threads will | |
7949 | * perform better since they share less core resources. Hence when we | |
7950 | * have idle threads, we want them to be the higher ones. | |
7951 | * | |
7952 | * This packing function is run on idle threads. It checks to see if | |
7953 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
7954 | * CPU number than the packing function is being run on. Here we are | |
7955 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
7956 | * number. | |
7957 | * | |
e69f6186 | 7958 | * Return: 1 when packing is required and a task should be moved to |
46123355 | 7959 | * this CPU. The amount of the imbalance is returned in env->imbalance. |
b6b12294 | 7960 | * |
cd96891d | 7961 | * @env: The load balancing environment. |
532cb4c4 | 7962 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 7963 | */ |
bd939f45 | 7964 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
7965 | { |
7966 | int busiest_cpu; | |
7967 | ||
bd939f45 | 7968 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
7969 | return 0; |
7970 | ||
1f621e02 SD |
7971 | if (env->idle == CPU_NOT_IDLE) |
7972 | return 0; | |
7973 | ||
532cb4c4 MN |
7974 | if (!sds->busiest) |
7975 | return 0; | |
7976 | ||
afe06efd TC |
7977 | busiest_cpu = sds->busiest->asym_prefer_cpu; |
7978 | if (sched_asym_prefer(busiest_cpu, env->dst_cpu)) | |
532cb4c4 MN |
7979 | return 0; |
7980 | ||
bd939f45 | 7981 | env->imbalance = DIV_ROUND_CLOSEST( |
63b2ca30 | 7982 | sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity, |
ca8ce3d0 | 7983 | SCHED_CAPACITY_SCALE); |
bd939f45 | 7984 | |
532cb4c4 | 7985 | return 1; |
1e3c88bd PZ |
7986 | } |
7987 | ||
7988 | /** | |
7989 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
7990 | * amongst the groups of a sched_domain, during | |
7991 | * load balancing. | |
cd96891d | 7992 | * @env: The load balancing environment. |
1e3c88bd | 7993 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 7994 | */ |
bd939f45 PZ |
7995 | static inline |
7996 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd | 7997 | { |
63b2ca30 | 7998 | unsigned long tmp, capa_now = 0, capa_move = 0; |
1e3c88bd | 7999 | unsigned int imbn = 2; |
dd5feea1 | 8000 | unsigned long scaled_busy_load_per_task; |
56cf515b | 8001 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 8002 | |
56cf515b JK |
8003 | local = &sds->local_stat; |
8004 | busiest = &sds->busiest_stat; | |
1e3c88bd | 8005 | |
56cf515b JK |
8006 | if (!local->sum_nr_running) |
8007 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
8008 | else if (busiest->load_per_task > local->load_per_task) | |
8009 | imbn = 1; | |
dd5feea1 | 8010 | |
56cf515b | 8011 | scaled_busy_load_per_task = |
ca8ce3d0 | 8012 | (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 8013 | busiest->group_capacity; |
56cf515b | 8014 | |
3029ede3 VD |
8015 | if (busiest->avg_load + scaled_busy_load_per_task >= |
8016 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 8017 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
8018 | return; |
8019 | } | |
8020 | ||
8021 | /* | |
8022 | * OK, we don't have enough imbalance to justify moving tasks, | |
ced549fa | 8023 | * however we may be able to increase total CPU capacity used by |
1e3c88bd PZ |
8024 | * moving them. |
8025 | */ | |
8026 | ||
63b2ca30 | 8027 | capa_now += busiest->group_capacity * |
56cf515b | 8028 | min(busiest->load_per_task, busiest->avg_load); |
63b2ca30 | 8029 | capa_now += local->group_capacity * |
56cf515b | 8030 | min(local->load_per_task, local->avg_load); |
ca8ce3d0 | 8031 | capa_now /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
8032 | |
8033 | /* Amount of load we'd subtract */ | |
a2cd4260 | 8034 | if (busiest->avg_load > scaled_busy_load_per_task) { |
63b2ca30 | 8035 | capa_move += busiest->group_capacity * |
56cf515b | 8036 | min(busiest->load_per_task, |
a2cd4260 | 8037 | busiest->avg_load - scaled_busy_load_per_task); |
56cf515b | 8038 | } |
1e3c88bd PZ |
8039 | |
8040 | /* Amount of load we'd add */ | |
63b2ca30 | 8041 | if (busiest->avg_load * busiest->group_capacity < |
ca8ce3d0 | 8042 | busiest->load_per_task * SCHED_CAPACITY_SCALE) { |
63b2ca30 NP |
8043 | tmp = (busiest->avg_load * busiest->group_capacity) / |
8044 | local->group_capacity; | |
56cf515b | 8045 | } else { |
ca8ce3d0 | 8046 | tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 8047 | local->group_capacity; |
56cf515b | 8048 | } |
63b2ca30 | 8049 | capa_move += local->group_capacity * |
3ae11c90 | 8050 | min(local->load_per_task, local->avg_load + tmp); |
ca8ce3d0 | 8051 | capa_move /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
8052 | |
8053 | /* Move if we gain throughput */ | |
63b2ca30 | 8054 | if (capa_move > capa_now) |
56cf515b | 8055 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
8056 | } |
8057 | ||
8058 | /** | |
8059 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
8060 | * groups of a given sched_domain during load balance. | |
bd939f45 | 8061 | * @env: load balance environment |
1e3c88bd | 8062 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 8063 | */ |
bd939f45 | 8064 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 8065 | { |
dd5feea1 | 8066 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
8067 | struct sg_lb_stats *local, *busiest; |
8068 | ||
8069 | local = &sds->local_stat; | |
56cf515b | 8070 | busiest = &sds->busiest_stat; |
dd5feea1 | 8071 | |
caeb178c | 8072 | if (busiest->group_type == group_imbalanced) { |
30ce5dab PZ |
8073 | /* |
8074 | * In the group_imb case we cannot rely on group-wide averages | |
8075 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
8076 | */ | |
56cf515b JK |
8077 | busiest->load_per_task = |
8078 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
8079 | } |
8080 | ||
1e3c88bd | 8081 | /* |
885e542c DE |
8082 | * Avg load of busiest sg can be less and avg load of local sg can |
8083 | * be greater than avg load across all sgs of sd because avg load | |
8084 | * factors in sg capacity and sgs with smaller group_type are | |
8085 | * skipped when updating the busiest sg: | |
1e3c88bd | 8086 | */ |
b1885550 VD |
8087 | if (busiest->avg_load <= sds->avg_load || |
8088 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
8089 | env->imbalance = 0; |
8090 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
8091 | } |
8092 | ||
9a5d9ba6 PZ |
8093 | /* |
8094 | * If there aren't any idle cpus, avoid creating some. | |
8095 | */ | |
8096 | if (busiest->group_type == group_overloaded && | |
8097 | local->group_type == group_overloaded) { | |
1be0eb2a | 8098 | load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE; |
cfa10334 | 8099 | if (load_above_capacity > busiest->group_capacity) { |
ea67821b | 8100 | load_above_capacity -= busiest->group_capacity; |
26656215 | 8101 | load_above_capacity *= scale_load_down(NICE_0_LOAD); |
cfa10334 MR |
8102 | load_above_capacity /= busiest->group_capacity; |
8103 | } else | |
ea67821b | 8104 | load_above_capacity = ~0UL; |
dd5feea1 SS |
8105 | } |
8106 | ||
8107 | /* | |
8108 | * We're trying to get all the cpus to the average_load, so we don't | |
8109 | * want to push ourselves above the average load, nor do we wish to | |
8110 | * reduce the max loaded cpu below the average load. At the same time, | |
0a9b23ce DE |
8111 | * we also don't want to reduce the group load below the group |
8112 | * capacity. Thus we look for the minimum possible imbalance. | |
dd5feea1 | 8113 | */ |
30ce5dab | 8114 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
8115 | |
8116 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 8117 | env->imbalance = min( |
63b2ca30 NP |
8118 | max_pull * busiest->group_capacity, |
8119 | (sds->avg_load - local->avg_load) * local->group_capacity | |
ca8ce3d0 | 8120 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
8121 | |
8122 | /* | |
8123 | * if *imbalance is less than the average load per runnable task | |
25985edc | 8124 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
8125 | * a think about bumping its value to force at least one task to be |
8126 | * moved | |
8127 | */ | |
56cf515b | 8128 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 8129 | return fix_small_imbalance(env, sds); |
1e3c88bd | 8130 | } |
fab47622 | 8131 | |
1e3c88bd PZ |
8132 | /******* find_busiest_group() helpers end here *********************/ |
8133 | ||
8134 | /** | |
8135 | * find_busiest_group - Returns the busiest group within the sched_domain | |
0a9b23ce | 8136 | * if there is an imbalance. |
1e3c88bd PZ |
8137 | * |
8138 | * Also calculates the amount of weighted load which should be moved | |
8139 | * to restore balance. | |
8140 | * | |
cd96891d | 8141 | * @env: The load balancing environment. |
1e3c88bd | 8142 | * |
e69f6186 | 8143 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 8144 | */ |
56cf515b | 8145 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 8146 | { |
56cf515b | 8147 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
8148 | struct sd_lb_stats sds; |
8149 | ||
147c5fc2 | 8150 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
8151 | |
8152 | /* | |
8153 | * Compute the various statistics relavent for load balancing at | |
8154 | * this level. | |
8155 | */ | |
23f0d209 | 8156 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
8157 | local = &sds.local_stat; |
8158 | busiest = &sds.busiest_stat; | |
1e3c88bd | 8159 | |
ea67821b | 8160 | /* ASYM feature bypasses nice load balance check */ |
1f621e02 | 8161 | if (check_asym_packing(env, &sds)) |
532cb4c4 MN |
8162 | return sds.busiest; |
8163 | ||
cc57aa8f | 8164 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 8165 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
8166 | goto out_balanced; |
8167 | ||
90001d67 | 8168 | /* XXX broken for overlapping NUMA groups */ |
ca8ce3d0 NP |
8169 | sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) |
8170 | / sds.total_capacity; | |
b0432d8f | 8171 | |
866ab43e PZ |
8172 | /* |
8173 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 8174 | * work because they assume all things are equal, which typically |
866ab43e PZ |
8175 | * isn't true due to cpus_allowed constraints and the like. |
8176 | */ | |
caeb178c | 8177 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
8178 | goto force_balance; |
8179 | ||
cc57aa8f | 8180 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
ea67821b VG |
8181 | if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) && |
8182 | busiest->group_no_capacity) | |
fab47622 NR |
8183 | goto force_balance; |
8184 | ||
cc57aa8f | 8185 | /* |
9c58c79a | 8186 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
8187 | * don't try and pull any tasks. |
8188 | */ | |
56cf515b | 8189 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
8190 | goto out_balanced; |
8191 | ||
cc57aa8f PZ |
8192 | /* |
8193 | * Don't pull any tasks if this group is already above the domain | |
8194 | * average load. | |
8195 | */ | |
56cf515b | 8196 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
8197 | goto out_balanced; |
8198 | ||
bd939f45 | 8199 | if (env->idle == CPU_IDLE) { |
aae6d3dd | 8200 | /* |
43f4d666 VG |
8201 | * This cpu is idle. If the busiest group is not overloaded |
8202 | * and there is no imbalance between this and busiest group | |
8203 | * wrt idle cpus, it is balanced. The imbalance becomes | |
8204 | * significant if the diff is greater than 1 otherwise we | |
8205 | * might end up to just move the imbalance on another group | |
aae6d3dd | 8206 | */ |
43f4d666 VG |
8207 | if ((busiest->group_type != group_overloaded) && |
8208 | (local->idle_cpus <= (busiest->idle_cpus + 1))) | |
aae6d3dd | 8209 | goto out_balanced; |
c186fafe PZ |
8210 | } else { |
8211 | /* | |
8212 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
8213 | * imbalance_pct to be conservative. | |
8214 | */ | |
56cf515b JK |
8215 | if (100 * busiest->avg_load <= |
8216 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 8217 | goto out_balanced; |
aae6d3dd | 8218 | } |
1e3c88bd | 8219 | |
fab47622 | 8220 | force_balance: |
1e3c88bd | 8221 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 8222 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
8223 | return sds.busiest; |
8224 | ||
8225 | out_balanced: | |
bd939f45 | 8226 | env->imbalance = 0; |
1e3c88bd PZ |
8227 | return NULL; |
8228 | } | |
8229 | ||
8230 | /* | |
8231 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
8232 | */ | |
bd939f45 | 8233 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 8234 | struct sched_group *group) |
1e3c88bd PZ |
8235 | { |
8236 | struct rq *busiest = NULL, *rq; | |
ced549fa | 8237 | unsigned long busiest_load = 0, busiest_capacity = 1; |
1e3c88bd PZ |
8238 | int i; |
8239 | ||
ae4df9d6 | 8240 | for_each_cpu_and(i, sched_group_span(group), env->cpus) { |
ea67821b | 8241 | unsigned long capacity, wl; |
0ec8aa00 PZ |
8242 | enum fbq_type rt; |
8243 | ||
8244 | rq = cpu_rq(i); | |
8245 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 8246 | |
0ec8aa00 PZ |
8247 | /* |
8248 | * We classify groups/runqueues into three groups: | |
8249 | * - regular: there are !numa tasks | |
8250 | * - remote: there are numa tasks that run on the 'wrong' node | |
8251 | * - all: there is no distinction | |
8252 | * | |
8253 | * In order to avoid migrating ideally placed numa tasks, | |
8254 | * ignore those when there's better options. | |
8255 | * | |
8256 | * If we ignore the actual busiest queue to migrate another | |
8257 | * task, the next balance pass can still reduce the busiest | |
8258 | * queue by moving tasks around inside the node. | |
8259 | * | |
8260 | * If we cannot move enough load due to this classification | |
8261 | * the next pass will adjust the group classification and | |
8262 | * allow migration of more tasks. | |
8263 | * | |
8264 | * Both cases only affect the total convergence complexity. | |
8265 | */ | |
8266 | if (rt > env->fbq_type) | |
8267 | continue; | |
8268 | ||
ced549fa | 8269 | capacity = capacity_of(i); |
9d5efe05 | 8270 | |
c7132dd6 | 8271 | wl = weighted_cpuload(rq); |
1e3c88bd | 8272 | |
6e40f5bb TG |
8273 | /* |
8274 | * When comparing with imbalance, use weighted_cpuload() | |
ced549fa | 8275 | * which is not scaled with the cpu capacity. |
6e40f5bb | 8276 | */ |
ea67821b VG |
8277 | |
8278 | if (rq->nr_running == 1 && wl > env->imbalance && | |
8279 | !check_cpu_capacity(rq, env->sd)) | |
1e3c88bd PZ |
8280 | continue; |
8281 | ||
6e40f5bb TG |
8282 | /* |
8283 | * For the load comparisons with the other cpu's, consider | |
ced549fa NP |
8284 | * the weighted_cpuload() scaled with the cpu capacity, so |
8285 | * that the load can be moved away from the cpu that is | |
8286 | * potentially running at a lower capacity. | |
95a79b80 | 8287 | * |
ced549fa | 8288 | * Thus we're looking for max(wl_i / capacity_i), crosswise |
95a79b80 | 8289 | * multiplication to rid ourselves of the division works out |
ced549fa NP |
8290 | * to: wl_i * capacity_j > wl_j * capacity_i; where j is |
8291 | * our previous maximum. | |
6e40f5bb | 8292 | */ |
ced549fa | 8293 | if (wl * busiest_capacity > busiest_load * capacity) { |
95a79b80 | 8294 | busiest_load = wl; |
ced549fa | 8295 | busiest_capacity = capacity; |
1e3c88bd PZ |
8296 | busiest = rq; |
8297 | } | |
8298 | } | |
8299 | ||
8300 | return busiest; | |
8301 | } | |
8302 | ||
8303 | /* | |
8304 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
8305 | * so long as it is large enough. | |
8306 | */ | |
8307 | #define MAX_PINNED_INTERVAL 512 | |
8308 | ||
bd939f45 | 8309 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 8310 | { |
bd939f45 PZ |
8311 | struct sched_domain *sd = env->sd; |
8312 | ||
8313 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
8314 | |
8315 | /* | |
8316 | * ASYM_PACKING needs to force migrate tasks from busy but | |
afe06efd TC |
8317 | * lower priority CPUs in order to pack all tasks in the |
8318 | * highest priority CPUs. | |
532cb4c4 | 8319 | */ |
afe06efd TC |
8320 | if ((sd->flags & SD_ASYM_PACKING) && |
8321 | sched_asym_prefer(env->dst_cpu, env->src_cpu)) | |
532cb4c4 | 8322 | return 1; |
1af3ed3d PZ |
8323 | } |
8324 | ||
1aaf90a4 VG |
8325 | /* |
8326 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
8327 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
8328 | * because of other sched_class or IRQs if more capacity stays | |
8329 | * available on dst_cpu. | |
8330 | */ | |
8331 | if ((env->idle != CPU_NOT_IDLE) && | |
8332 | (env->src_rq->cfs.h_nr_running == 1)) { | |
8333 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
8334 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
8335 | return 1; | |
8336 | } | |
8337 | ||
1af3ed3d PZ |
8338 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); |
8339 | } | |
8340 | ||
969c7921 TH |
8341 | static int active_load_balance_cpu_stop(void *data); |
8342 | ||
23f0d209 JK |
8343 | static int should_we_balance(struct lb_env *env) |
8344 | { | |
8345 | struct sched_group *sg = env->sd->groups; | |
23f0d209 JK |
8346 | int cpu, balance_cpu = -1; |
8347 | ||
024c9d2f PZ |
8348 | /* |
8349 | * Ensure the balancing environment is consistent; can happen | |
8350 | * when the softirq triggers 'during' hotplug. | |
8351 | */ | |
8352 | if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) | |
8353 | return 0; | |
8354 | ||
23f0d209 JK |
8355 | /* |
8356 | * In the newly idle case, we will allow all the cpu's | |
8357 | * to do the newly idle load balance. | |
8358 | */ | |
8359 | if (env->idle == CPU_NEWLY_IDLE) | |
8360 | return 1; | |
8361 | ||
23f0d209 | 8362 | /* Try to find first idle cpu */ |
e5c14b1f | 8363 | for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { |
af218122 | 8364 | if (!idle_cpu(cpu)) |
23f0d209 JK |
8365 | continue; |
8366 | ||
8367 | balance_cpu = cpu; | |
8368 | break; | |
8369 | } | |
8370 | ||
8371 | if (balance_cpu == -1) | |
8372 | balance_cpu = group_balance_cpu(sg); | |
8373 | ||
8374 | /* | |
8375 | * First idle cpu or the first cpu(busiest) in this sched group | |
8376 | * is eligible for doing load balancing at this and above domains. | |
8377 | */ | |
b0cff9d8 | 8378 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
8379 | } |
8380 | ||
1e3c88bd PZ |
8381 | /* |
8382 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
8383 | * tasks if there is an imbalance. | |
8384 | */ | |
8385 | static int load_balance(int this_cpu, struct rq *this_rq, | |
8386 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 8387 | int *continue_balancing) |
1e3c88bd | 8388 | { |
88b8dac0 | 8389 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 8390 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 8391 | struct sched_group *group; |
1e3c88bd | 8392 | struct rq *busiest; |
8a8c69c3 | 8393 | struct rq_flags rf; |
4ba29684 | 8394 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 8395 | |
8e45cb54 PZ |
8396 | struct lb_env env = { |
8397 | .sd = sd, | |
ddcdf6e7 PZ |
8398 | .dst_cpu = this_cpu, |
8399 | .dst_rq = this_rq, | |
ae4df9d6 | 8400 | .dst_grpmask = sched_group_span(sd->groups), |
8e45cb54 | 8401 | .idle = idle, |
eb95308e | 8402 | .loop_break = sched_nr_migrate_break, |
b9403130 | 8403 | .cpus = cpus, |
0ec8aa00 | 8404 | .fbq_type = all, |
163122b7 | 8405 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
8406 | }; |
8407 | ||
65a4433a | 8408 | cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); |
1e3c88bd | 8409 | |
ae92882e | 8410 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
8411 | |
8412 | redo: | |
23f0d209 JK |
8413 | if (!should_we_balance(&env)) { |
8414 | *continue_balancing = 0; | |
1e3c88bd | 8415 | goto out_balanced; |
23f0d209 | 8416 | } |
1e3c88bd | 8417 | |
23f0d209 | 8418 | group = find_busiest_group(&env); |
1e3c88bd | 8419 | if (!group) { |
ae92882e | 8420 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
8421 | goto out_balanced; |
8422 | } | |
8423 | ||
b9403130 | 8424 | busiest = find_busiest_queue(&env, group); |
1e3c88bd | 8425 | if (!busiest) { |
ae92882e | 8426 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
8427 | goto out_balanced; |
8428 | } | |
8429 | ||
78feefc5 | 8430 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 8431 | |
ae92882e | 8432 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 8433 | |
1aaf90a4 VG |
8434 | env.src_cpu = busiest->cpu; |
8435 | env.src_rq = busiest; | |
8436 | ||
1e3c88bd PZ |
8437 | ld_moved = 0; |
8438 | if (busiest->nr_running > 1) { | |
8439 | /* | |
8440 | * Attempt to move tasks. If find_busiest_group has found | |
8441 | * an imbalance but busiest->nr_running <= 1, the group is | |
8442 | * still unbalanced. ld_moved simply stays zero, so it is | |
8443 | * correctly treated as an imbalance. | |
8444 | */ | |
8e45cb54 | 8445 | env.flags |= LBF_ALL_PINNED; |
c82513e5 | 8446 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 8447 | |
5d6523eb | 8448 | more_balance: |
8a8c69c3 | 8449 | rq_lock_irqsave(busiest, &rf); |
3bed5e21 | 8450 | update_rq_clock(busiest); |
88b8dac0 SV |
8451 | |
8452 | /* | |
8453 | * cur_ld_moved - load moved in current iteration | |
8454 | * ld_moved - cumulative load moved across iterations | |
8455 | */ | |
163122b7 | 8456 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
8457 | |
8458 | /* | |
163122b7 KT |
8459 | * We've detached some tasks from busiest_rq. Every |
8460 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
8461 | * unlock busiest->lock, and we are able to be sure | |
8462 | * that nobody can manipulate the tasks in parallel. | |
8463 | * See task_rq_lock() family for the details. | |
1e3c88bd | 8464 | */ |
163122b7 | 8465 | |
8a8c69c3 | 8466 | rq_unlock(busiest, &rf); |
163122b7 KT |
8467 | |
8468 | if (cur_ld_moved) { | |
8469 | attach_tasks(&env); | |
8470 | ld_moved += cur_ld_moved; | |
8471 | } | |
8472 | ||
8a8c69c3 | 8473 | local_irq_restore(rf.flags); |
88b8dac0 | 8474 | |
f1cd0858 JK |
8475 | if (env.flags & LBF_NEED_BREAK) { |
8476 | env.flags &= ~LBF_NEED_BREAK; | |
8477 | goto more_balance; | |
8478 | } | |
8479 | ||
88b8dac0 SV |
8480 | /* |
8481 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
8482 | * us and move them to an alternate dst_cpu in our sched_group | |
8483 | * where they can run. The upper limit on how many times we | |
8484 | * iterate on same src_cpu is dependent on number of cpus in our | |
8485 | * sched_group. | |
8486 | * | |
8487 | * This changes load balance semantics a bit on who can move | |
8488 | * load to a given_cpu. In addition to the given_cpu itself | |
8489 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
8490 | * nohz-idle), we now have balance_cpu in a position to move | |
8491 | * load to given_cpu. In rare situations, this may cause | |
8492 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
8493 | * _independently_ and at _same_ time to move some load to | |
8494 | * given_cpu) causing exceess load to be moved to given_cpu. | |
8495 | * This however should not happen so much in practice and | |
8496 | * moreover subsequent load balance cycles should correct the | |
8497 | * excess load moved. | |
8498 | */ | |
6263322c | 8499 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 8500 | |
7aff2e3a VD |
8501 | /* Prevent to re-select dst_cpu via env's cpus */ |
8502 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
8503 | ||
78feefc5 | 8504 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 8505 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 8506 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
8507 | env.loop = 0; |
8508 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 8509 | |
88b8dac0 SV |
8510 | /* |
8511 | * Go back to "more_balance" rather than "redo" since we | |
8512 | * need to continue with same src_cpu. | |
8513 | */ | |
8514 | goto more_balance; | |
8515 | } | |
1e3c88bd | 8516 | |
6263322c PZ |
8517 | /* |
8518 | * We failed to reach balance because of affinity. | |
8519 | */ | |
8520 | if (sd_parent) { | |
63b2ca30 | 8521 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 8522 | |
afdeee05 | 8523 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 8524 | *group_imbalance = 1; |
6263322c PZ |
8525 | } |
8526 | ||
1e3c88bd | 8527 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 8528 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 8529 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
65a4433a JH |
8530 | /* |
8531 | * Attempting to continue load balancing at the current | |
8532 | * sched_domain level only makes sense if there are | |
8533 | * active CPUs remaining as possible busiest CPUs to | |
8534 | * pull load from which are not contained within the | |
8535 | * destination group that is receiving any migrated | |
8536 | * load. | |
8537 | */ | |
8538 | if (!cpumask_subset(cpus, env.dst_grpmask)) { | |
bbf18b19 PN |
8539 | env.loop = 0; |
8540 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 8541 | goto redo; |
bbf18b19 | 8542 | } |
afdeee05 | 8543 | goto out_all_pinned; |
1e3c88bd PZ |
8544 | } |
8545 | } | |
8546 | ||
8547 | if (!ld_moved) { | |
ae92882e | 8548 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
8549 | /* |
8550 | * Increment the failure counter only on periodic balance. | |
8551 | * We do not want newidle balance, which can be very | |
8552 | * frequent, pollute the failure counter causing | |
8553 | * excessive cache_hot migrations and active balances. | |
8554 | */ | |
8555 | if (idle != CPU_NEWLY_IDLE) | |
8556 | sd->nr_balance_failed++; | |
1e3c88bd | 8557 | |
bd939f45 | 8558 | if (need_active_balance(&env)) { |
8a8c69c3 PZ |
8559 | unsigned long flags; |
8560 | ||
1e3c88bd PZ |
8561 | raw_spin_lock_irqsave(&busiest->lock, flags); |
8562 | ||
969c7921 TH |
8563 | /* don't kick the active_load_balance_cpu_stop, |
8564 | * if the curr task on busiest cpu can't be | |
8565 | * moved to this_cpu | |
1e3c88bd | 8566 | */ |
0c98d344 | 8567 | if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) { |
1e3c88bd PZ |
8568 | raw_spin_unlock_irqrestore(&busiest->lock, |
8569 | flags); | |
8e45cb54 | 8570 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
8571 | goto out_one_pinned; |
8572 | } | |
8573 | ||
969c7921 TH |
8574 | /* |
8575 | * ->active_balance synchronizes accesses to | |
8576 | * ->active_balance_work. Once set, it's cleared | |
8577 | * only after active load balance is finished. | |
8578 | */ | |
1e3c88bd PZ |
8579 | if (!busiest->active_balance) { |
8580 | busiest->active_balance = 1; | |
8581 | busiest->push_cpu = this_cpu; | |
8582 | active_balance = 1; | |
8583 | } | |
8584 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 8585 | |
bd939f45 | 8586 | if (active_balance) { |
969c7921 TH |
8587 | stop_one_cpu_nowait(cpu_of(busiest), |
8588 | active_load_balance_cpu_stop, busiest, | |
8589 | &busiest->active_balance_work); | |
bd939f45 | 8590 | } |
1e3c88bd | 8591 | |
d02c0711 | 8592 | /* We've kicked active balancing, force task migration. */ |
1e3c88bd PZ |
8593 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
8594 | } | |
8595 | } else | |
8596 | sd->nr_balance_failed = 0; | |
8597 | ||
8598 | if (likely(!active_balance)) { | |
8599 | /* We were unbalanced, so reset the balancing interval */ | |
8600 | sd->balance_interval = sd->min_interval; | |
8601 | } else { | |
8602 | /* | |
8603 | * If we've begun active balancing, start to back off. This | |
8604 | * case may not be covered by the all_pinned logic if there | |
8605 | * is only 1 task on the busy runqueue (because we don't call | |
163122b7 | 8606 | * detach_tasks). |
1e3c88bd PZ |
8607 | */ |
8608 | if (sd->balance_interval < sd->max_interval) | |
8609 | sd->balance_interval *= 2; | |
8610 | } | |
8611 | ||
1e3c88bd PZ |
8612 | goto out; |
8613 | ||
8614 | out_balanced: | |
afdeee05 VG |
8615 | /* |
8616 | * We reach balance although we may have faced some affinity | |
8617 | * constraints. Clear the imbalance flag if it was set. | |
8618 | */ | |
8619 | if (sd_parent) { | |
8620 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; | |
8621 | ||
8622 | if (*group_imbalance) | |
8623 | *group_imbalance = 0; | |
8624 | } | |
8625 | ||
8626 | out_all_pinned: | |
8627 | /* | |
8628 | * We reach balance because all tasks are pinned at this level so | |
8629 | * we can't migrate them. Let the imbalance flag set so parent level | |
8630 | * can try to migrate them. | |
8631 | */ | |
ae92882e | 8632 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
8633 | |
8634 | sd->nr_balance_failed = 0; | |
8635 | ||
8636 | out_one_pinned: | |
8637 | /* tune up the balancing interval */ | |
8e45cb54 | 8638 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 8639 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
8640 | (sd->balance_interval < sd->max_interval)) |
8641 | sd->balance_interval *= 2; | |
8642 | ||
46e49b38 | 8643 | ld_moved = 0; |
1e3c88bd | 8644 | out: |
1e3c88bd PZ |
8645 | return ld_moved; |
8646 | } | |
8647 | ||
52a08ef1 JL |
8648 | static inline unsigned long |
8649 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
8650 | { | |
8651 | unsigned long interval = sd->balance_interval; | |
8652 | ||
8653 | if (cpu_busy) | |
8654 | interval *= sd->busy_factor; | |
8655 | ||
8656 | /* scale ms to jiffies */ | |
8657 | interval = msecs_to_jiffies(interval); | |
8658 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
8659 | ||
8660 | return interval; | |
8661 | } | |
8662 | ||
8663 | static inline void | |
31851a98 | 8664 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
8665 | { |
8666 | unsigned long interval, next; | |
8667 | ||
31851a98 LY |
8668 | /* used by idle balance, so cpu_busy = 0 */ |
8669 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
8670 | next = sd->last_balance + interval; |
8671 | ||
8672 | if (time_after(*next_balance, next)) | |
8673 | *next_balance = next; | |
8674 | } | |
8675 | ||
1e3c88bd PZ |
8676 | /* |
8677 | * idle_balance is called by schedule() if this_cpu is about to become | |
8678 | * idle. Attempts to pull tasks from other CPUs. | |
8679 | */ | |
46f69fa3 | 8680 | static int idle_balance(struct rq *this_rq, struct rq_flags *rf) |
1e3c88bd | 8681 | { |
52a08ef1 JL |
8682 | unsigned long next_balance = jiffies + HZ; |
8683 | int this_cpu = this_rq->cpu; | |
1e3c88bd PZ |
8684 | struct sched_domain *sd; |
8685 | int pulled_task = 0; | |
9bd721c5 | 8686 | u64 curr_cost = 0; |
1e3c88bd | 8687 | |
6e83125c PZ |
8688 | /* |
8689 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
8690 | * measure the duration of idle_balance() as idle time. | |
8691 | */ | |
8692 | this_rq->idle_stamp = rq_clock(this_rq); | |
8693 | ||
2800486e PZ |
8694 | /* |
8695 | * Do not pull tasks towards !active CPUs... | |
8696 | */ | |
8697 | if (!cpu_active(this_cpu)) | |
8698 | return 0; | |
8699 | ||
46f69fa3 MF |
8700 | /* |
8701 | * This is OK, because current is on_cpu, which avoids it being picked | |
8702 | * for load-balance and preemption/IRQs are still disabled avoiding | |
8703 | * further scheduler activity on it and we're being very careful to | |
8704 | * re-start the picking loop. | |
8705 | */ | |
8706 | rq_unpin_lock(this_rq, rf); | |
8707 | ||
4486edd1 TC |
8708 | if (this_rq->avg_idle < sysctl_sched_migration_cost || |
8709 | !this_rq->rd->overload) { | |
52a08ef1 JL |
8710 | rcu_read_lock(); |
8711 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
8712 | if (sd) | |
31851a98 | 8713 | update_next_balance(sd, &next_balance); |
52a08ef1 JL |
8714 | rcu_read_unlock(); |
8715 | ||
6e83125c | 8716 | goto out; |
52a08ef1 | 8717 | } |
1e3c88bd | 8718 | |
f492e12e PZ |
8719 | raw_spin_unlock(&this_rq->lock); |
8720 | ||
48a16753 | 8721 | update_blocked_averages(this_cpu); |
dce840a0 | 8722 | rcu_read_lock(); |
1e3c88bd | 8723 | for_each_domain(this_cpu, sd) { |
23f0d209 | 8724 | int continue_balancing = 1; |
9bd721c5 | 8725 | u64 t0, domain_cost; |
1e3c88bd PZ |
8726 | |
8727 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
8728 | continue; | |
8729 | ||
52a08ef1 | 8730 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { |
31851a98 | 8731 | update_next_balance(sd, &next_balance); |
9bd721c5 | 8732 | break; |
52a08ef1 | 8733 | } |
9bd721c5 | 8734 | |
f492e12e | 8735 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
8736 | t0 = sched_clock_cpu(this_cpu); |
8737 | ||
f492e12e | 8738 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
8739 | sd, CPU_NEWLY_IDLE, |
8740 | &continue_balancing); | |
9bd721c5 JL |
8741 | |
8742 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
8743 | if (domain_cost > sd->max_newidle_lb_cost) | |
8744 | sd->max_newidle_lb_cost = domain_cost; | |
8745 | ||
8746 | curr_cost += domain_cost; | |
f492e12e | 8747 | } |
1e3c88bd | 8748 | |
31851a98 | 8749 | update_next_balance(sd, &next_balance); |
39a4d9ca JL |
8750 | |
8751 | /* | |
8752 | * Stop searching for tasks to pull if there are | |
8753 | * now runnable tasks on this rq. | |
8754 | */ | |
8755 | if (pulled_task || this_rq->nr_running > 0) | |
1e3c88bd | 8756 | break; |
1e3c88bd | 8757 | } |
dce840a0 | 8758 | rcu_read_unlock(); |
f492e12e PZ |
8759 | |
8760 | raw_spin_lock(&this_rq->lock); | |
8761 | ||
0e5b5337 JL |
8762 | if (curr_cost > this_rq->max_idle_balance_cost) |
8763 | this_rq->max_idle_balance_cost = curr_cost; | |
8764 | ||
e5fc6611 | 8765 | /* |
0e5b5337 JL |
8766 | * While browsing the domains, we released the rq lock, a task could |
8767 | * have been enqueued in the meantime. Since we're not going idle, | |
8768 | * pretend we pulled a task. | |
e5fc6611 | 8769 | */ |
0e5b5337 | 8770 | if (this_rq->cfs.h_nr_running && !pulled_task) |
6e83125c | 8771 | pulled_task = 1; |
e5fc6611 | 8772 | |
52a08ef1 JL |
8773 | out: |
8774 | /* Move the next balance forward */ | |
8775 | if (time_after(this_rq->next_balance, next_balance)) | |
1e3c88bd | 8776 | this_rq->next_balance = next_balance; |
9bd721c5 | 8777 | |
e4aa358b | 8778 | /* Is there a task of a high priority class? */ |
46383648 | 8779 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) |
e4aa358b KT |
8780 | pulled_task = -1; |
8781 | ||
38c6ade2 | 8782 | if (pulled_task) |
6e83125c PZ |
8783 | this_rq->idle_stamp = 0; |
8784 | ||
46f69fa3 MF |
8785 | rq_repin_lock(this_rq, rf); |
8786 | ||
3c4017c1 | 8787 | return pulled_task; |
1e3c88bd PZ |
8788 | } |
8789 | ||
8790 | /* | |
969c7921 TH |
8791 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
8792 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
8793 | * least 1 task to be running on each physical CPU where possible, and | |
8794 | * avoids physical / logical imbalances. | |
1e3c88bd | 8795 | */ |
969c7921 | 8796 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 8797 | { |
969c7921 TH |
8798 | struct rq *busiest_rq = data; |
8799 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 8800 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 8801 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 8802 | struct sched_domain *sd; |
e5673f28 | 8803 | struct task_struct *p = NULL; |
8a8c69c3 | 8804 | struct rq_flags rf; |
969c7921 | 8805 | |
8a8c69c3 | 8806 | rq_lock_irq(busiest_rq, &rf); |
edd8e41d PZ |
8807 | /* |
8808 | * Between queueing the stop-work and running it is a hole in which | |
8809 | * CPUs can become inactive. We should not move tasks from or to | |
8810 | * inactive CPUs. | |
8811 | */ | |
8812 | if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) | |
8813 | goto out_unlock; | |
969c7921 TH |
8814 | |
8815 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
8816 | if (unlikely(busiest_cpu != smp_processor_id() || | |
8817 | !busiest_rq->active_balance)) | |
8818 | goto out_unlock; | |
1e3c88bd PZ |
8819 | |
8820 | /* Is there any task to move? */ | |
8821 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 8822 | goto out_unlock; |
1e3c88bd PZ |
8823 | |
8824 | /* | |
8825 | * This condition is "impossible", if it occurs | |
8826 | * we need to fix it. Originally reported by | |
8827 | * Bjorn Helgaas on a 128-cpu setup. | |
8828 | */ | |
8829 | BUG_ON(busiest_rq == target_rq); | |
8830 | ||
1e3c88bd | 8831 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 8832 | rcu_read_lock(); |
1e3c88bd PZ |
8833 | for_each_domain(target_cpu, sd) { |
8834 | if ((sd->flags & SD_LOAD_BALANCE) && | |
8835 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
8836 | break; | |
8837 | } | |
8838 | ||
8839 | if (likely(sd)) { | |
8e45cb54 PZ |
8840 | struct lb_env env = { |
8841 | .sd = sd, | |
ddcdf6e7 PZ |
8842 | .dst_cpu = target_cpu, |
8843 | .dst_rq = target_rq, | |
8844 | .src_cpu = busiest_rq->cpu, | |
8845 | .src_rq = busiest_rq, | |
8e45cb54 | 8846 | .idle = CPU_IDLE, |
65a4433a JH |
8847 | /* |
8848 | * can_migrate_task() doesn't need to compute new_dst_cpu | |
8849 | * for active balancing. Since we have CPU_IDLE, but no | |
8850 | * @dst_grpmask we need to make that test go away with lying | |
8851 | * about DST_PINNED. | |
8852 | */ | |
8853 | .flags = LBF_DST_PINNED, | |
8e45cb54 PZ |
8854 | }; |
8855 | ||
ae92882e | 8856 | schedstat_inc(sd->alb_count); |
3bed5e21 | 8857 | update_rq_clock(busiest_rq); |
1e3c88bd | 8858 | |
e5673f28 | 8859 | p = detach_one_task(&env); |
d02c0711 | 8860 | if (p) { |
ae92882e | 8861 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
8862 | /* Active balancing done, reset the failure counter. */ |
8863 | sd->nr_balance_failed = 0; | |
8864 | } else { | |
ae92882e | 8865 | schedstat_inc(sd->alb_failed); |
d02c0711 | 8866 | } |
1e3c88bd | 8867 | } |
dce840a0 | 8868 | rcu_read_unlock(); |
969c7921 TH |
8869 | out_unlock: |
8870 | busiest_rq->active_balance = 0; | |
8a8c69c3 | 8871 | rq_unlock(busiest_rq, &rf); |
e5673f28 KT |
8872 | |
8873 | if (p) | |
8874 | attach_one_task(target_rq, p); | |
8875 | ||
8876 | local_irq_enable(); | |
8877 | ||
969c7921 | 8878 | return 0; |
1e3c88bd PZ |
8879 | } |
8880 | ||
d987fc7f MG |
8881 | static inline int on_null_domain(struct rq *rq) |
8882 | { | |
8883 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
8884 | } | |
8885 | ||
3451d024 | 8886 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
8887 | /* |
8888 | * idle load balancing details | |
83cd4fe2 VP |
8889 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
8890 | * needed, they will kick the idle load balancer, which then does idle | |
8891 | * load balancing for all the idle CPUs. | |
8892 | */ | |
1e3c88bd | 8893 | static struct { |
83cd4fe2 | 8894 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 8895 | atomic_t nr_cpus; |
83cd4fe2 VP |
8896 | unsigned long next_balance; /* in jiffy units */ |
8897 | } nohz ____cacheline_aligned; | |
1e3c88bd | 8898 | |
3dd0337d | 8899 | static inline int find_new_ilb(void) |
1e3c88bd | 8900 | { |
0b005cf5 | 8901 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 8902 | |
786d6dc7 SS |
8903 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
8904 | return ilb; | |
8905 | ||
8906 | return nr_cpu_ids; | |
1e3c88bd | 8907 | } |
1e3c88bd | 8908 | |
83cd4fe2 VP |
8909 | /* |
8910 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
8911 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
8912 | * CPU (if there is one). | |
8913 | */ | |
0aeeeeba | 8914 | static void nohz_balancer_kick(void) |
83cd4fe2 VP |
8915 | { |
8916 | int ilb_cpu; | |
8917 | ||
8918 | nohz.next_balance++; | |
8919 | ||
3dd0337d | 8920 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 8921 | |
0b005cf5 SS |
8922 | if (ilb_cpu >= nr_cpu_ids) |
8923 | return; | |
83cd4fe2 | 8924 | |
cd490c5b | 8925 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
8926 | return; |
8927 | /* | |
8928 | * Use smp_send_reschedule() instead of resched_cpu(). | |
8929 | * This way we generate a sched IPI on the target cpu which | |
8930 | * is idle. And the softirq performing nohz idle load balance | |
8931 | * will be run before returning from the IPI. | |
8932 | */ | |
8933 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
8934 | return; |
8935 | } | |
8936 | ||
20a5c8cc | 8937 | void nohz_balance_exit_idle(unsigned int cpu) |
71325960 SS |
8938 | { |
8939 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
d987fc7f MG |
8940 | /* |
8941 | * Completely isolated CPUs don't ever set, so we must test. | |
8942 | */ | |
8943 | if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) { | |
8944 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
8945 | atomic_dec(&nohz.nr_cpus); | |
8946 | } | |
71325960 SS |
8947 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); |
8948 | } | |
8949 | } | |
8950 | ||
69e1e811 SS |
8951 | static inline void set_cpu_sd_state_busy(void) |
8952 | { | |
8953 | struct sched_domain *sd; | |
37dc6b50 | 8954 | int cpu = smp_processor_id(); |
69e1e811 | 8955 | |
69e1e811 | 8956 | rcu_read_lock(); |
0e369d75 | 8957 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
8958 | |
8959 | if (!sd || !sd->nohz_idle) | |
8960 | goto unlock; | |
8961 | sd->nohz_idle = 0; | |
8962 | ||
0e369d75 | 8963 | atomic_inc(&sd->shared->nr_busy_cpus); |
25f55d9d | 8964 | unlock: |
69e1e811 SS |
8965 | rcu_read_unlock(); |
8966 | } | |
8967 | ||
8968 | void set_cpu_sd_state_idle(void) | |
8969 | { | |
8970 | struct sched_domain *sd; | |
37dc6b50 | 8971 | int cpu = smp_processor_id(); |
69e1e811 | 8972 | |
69e1e811 | 8973 | rcu_read_lock(); |
0e369d75 | 8974 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
8975 | |
8976 | if (!sd || sd->nohz_idle) | |
8977 | goto unlock; | |
8978 | sd->nohz_idle = 1; | |
8979 | ||
0e369d75 | 8980 | atomic_dec(&sd->shared->nr_busy_cpus); |
25f55d9d | 8981 | unlock: |
69e1e811 SS |
8982 | rcu_read_unlock(); |
8983 | } | |
8984 | ||
1e3c88bd | 8985 | /* |
c1cc017c | 8986 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 8987 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 8988 | */ |
c1cc017c | 8989 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 8990 | { |
71325960 SS |
8991 | /* |
8992 | * If this cpu is going down, then nothing needs to be done. | |
8993 | */ | |
8994 | if (!cpu_active(cpu)) | |
8995 | return; | |
8996 | ||
387bc8b5 FW |
8997 | /* Spare idle load balancing on CPUs that don't want to be disturbed: */ |
8998 | if (!is_housekeeping_cpu(cpu)) | |
8999 | return; | |
9000 | ||
c1cc017c AS |
9001 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
9002 | return; | |
1e3c88bd | 9003 | |
d987fc7f MG |
9004 | /* |
9005 | * If we're a completely isolated CPU, we don't play. | |
9006 | */ | |
9007 | if (on_null_domain(cpu_rq(cpu))) | |
9008 | return; | |
9009 | ||
c1cc017c AS |
9010 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
9011 | atomic_inc(&nohz.nr_cpus); | |
9012 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd PZ |
9013 | } |
9014 | #endif | |
9015 | ||
9016 | static DEFINE_SPINLOCK(balancing); | |
9017 | ||
49c022e6 PZ |
9018 | /* |
9019 | * Scale the max load_balance interval with the number of CPUs in the system. | |
9020 | * This trades load-balance latency on larger machines for less cross talk. | |
9021 | */ | |
029632fb | 9022 | void update_max_interval(void) |
49c022e6 PZ |
9023 | { |
9024 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
9025 | } | |
9026 | ||
1e3c88bd PZ |
9027 | /* |
9028 | * It checks each scheduling domain to see if it is due to be balanced, | |
9029 | * and initiates a balancing operation if so. | |
9030 | * | |
b9b0853a | 9031 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd | 9032 | */ |
f7ed0a89 | 9033 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) |
1e3c88bd | 9034 | { |
23f0d209 | 9035 | int continue_balancing = 1; |
f7ed0a89 | 9036 | int cpu = rq->cpu; |
1e3c88bd | 9037 | unsigned long interval; |
04f733b4 | 9038 | struct sched_domain *sd; |
1e3c88bd PZ |
9039 | /* Earliest time when we have to do rebalance again */ |
9040 | unsigned long next_balance = jiffies + 60*HZ; | |
9041 | int update_next_balance = 0; | |
f48627e6 JL |
9042 | int need_serialize, need_decay = 0; |
9043 | u64 max_cost = 0; | |
1e3c88bd | 9044 | |
48a16753 | 9045 | update_blocked_averages(cpu); |
2069dd75 | 9046 | |
dce840a0 | 9047 | rcu_read_lock(); |
1e3c88bd | 9048 | for_each_domain(cpu, sd) { |
f48627e6 JL |
9049 | /* |
9050 | * Decay the newidle max times here because this is a regular | |
9051 | * visit to all the domains. Decay ~1% per second. | |
9052 | */ | |
9053 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
9054 | sd->max_newidle_lb_cost = | |
9055 | (sd->max_newidle_lb_cost * 253) / 256; | |
9056 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
9057 | need_decay = 1; | |
9058 | } | |
9059 | max_cost += sd->max_newidle_lb_cost; | |
9060 | ||
1e3c88bd PZ |
9061 | if (!(sd->flags & SD_LOAD_BALANCE)) |
9062 | continue; | |
9063 | ||
f48627e6 JL |
9064 | /* |
9065 | * Stop the load balance at this level. There is another | |
9066 | * CPU in our sched group which is doing load balancing more | |
9067 | * actively. | |
9068 | */ | |
9069 | if (!continue_balancing) { | |
9070 | if (need_decay) | |
9071 | continue; | |
9072 | break; | |
9073 | } | |
9074 | ||
52a08ef1 | 9075 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
9076 | |
9077 | need_serialize = sd->flags & SD_SERIALIZE; | |
1e3c88bd PZ |
9078 | if (need_serialize) { |
9079 | if (!spin_trylock(&balancing)) | |
9080 | goto out; | |
9081 | } | |
9082 | ||
9083 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 9084 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 9085 | /* |
6263322c | 9086 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
9087 | * env->dst_cpu, so we can't know our idle |
9088 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 9089 | */ |
de5eb2dd | 9090 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
9091 | } |
9092 | sd->last_balance = jiffies; | |
52a08ef1 | 9093 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
9094 | } |
9095 | if (need_serialize) | |
9096 | spin_unlock(&balancing); | |
9097 | out: | |
9098 | if (time_after(next_balance, sd->last_balance + interval)) { | |
9099 | next_balance = sd->last_balance + interval; | |
9100 | update_next_balance = 1; | |
9101 | } | |
f48627e6 JL |
9102 | } |
9103 | if (need_decay) { | |
1e3c88bd | 9104 | /* |
f48627e6 JL |
9105 | * Ensure the rq-wide value also decays but keep it at a |
9106 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 9107 | */ |
f48627e6 JL |
9108 | rq->max_idle_balance_cost = |
9109 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 9110 | } |
dce840a0 | 9111 | rcu_read_unlock(); |
1e3c88bd PZ |
9112 | |
9113 | /* | |
9114 | * next_balance will be updated only when there is a need. | |
9115 | * When the cpu is attached to null domain for ex, it will not be | |
9116 | * updated. | |
9117 | */ | |
c5afb6a8 | 9118 | if (likely(update_next_balance)) { |
1e3c88bd | 9119 | rq->next_balance = next_balance; |
c5afb6a8 VG |
9120 | |
9121 | #ifdef CONFIG_NO_HZ_COMMON | |
9122 | /* | |
9123 | * If this CPU has been elected to perform the nohz idle | |
9124 | * balance. Other idle CPUs have already rebalanced with | |
9125 | * nohz_idle_balance() and nohz.next_balance has been | |
9126 | * updated accordingly. This CPU is now running the idle load | |
9127 | * balance for itself and we need to update the | |
9128 | * nohz.next_balance accordingly. | |
9129 | */ | |
9130 | if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance)) | |
9131 | nohz.next_balance = rq->next_balance; | |
9132 | #endif | |
9133 | } | |
1e3c88bd PZ |
9134 | } |
9135 | ||
3451d024 | 9136 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 9137 | /* |
3451d024 | 9138 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
9139 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
9140 | */ | |
208cb16b | 9141 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
83cd4fe2 | 9142 | { |
208cb16b | 9143 | int this_cpu = this_rq->cpu; |
83cd4fe2 VP |
9144 | struct rq *rq; |
9145 | int balance_cpu; | |
c5afb6a8 VG |
9146 | /* Earliest time when we have to do rebalance again */ |
9147 | unsigned long next_balance = jiffies + 60*HZ; | |
9148 | int update_next_balance = 0; | |
83cd4fe2 | 9149 | |
1c792db7 SS |
9150 | if (idle != CPU_IDLE || |
9151 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
9152 | goto end; | |
83cd4fe2 VP |
9153 | |
9154 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 9155 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
9156 | continue; |
9157 | ||
9158 | /* | |
9159 | * If this cpu gets work to do, stop the load balancing | |
9160 | * work being done for other cpus. Next load | |
9161 | * balancing owner will pick it up. | |
9162 | */ | |
1c792db7 | 9163 | if (need_resched()) |
83cd4fe2 | 9164 | break; |
83cd4fe2 | 9165 | |
5ed4f1d9 VG |
9166 | rq = cpu_rq(balance_cpu); |
9167 | ||
ed61bbc6 TC |
9168 | /* |
9169 | * If time for next balance is due, | |
9170 | * do the balance. | |
9171 | */ | |
9172 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8a8c69c3 PZ |
9173 | struct rq_flags rf; |
9174 | ||
9175 | rq_lock_irq(rq, &rf); | |
ed61bbc6 | 9176 | update_rq_clock(rq); |
cee1afce | 9177 | cpu_load_update_idle(rq); |
8a8c69c3 PZ |
9178 | rq_unlock_irq(rq, &rf); |
9179 | ||
ed61bbc6 TC |
9180 | rebalance_domains(rq, CPU_IDLE); |
9181 | } | |
83cd4fe2 | 9182 | |
c5afb6a8 VG |
9183 | if (time_after(next_balance, rq->next_balance)) { |
9184 | next_balance = rq->next_balance; | |
9185 | update_next_balance = 1; | |
9186 | } | |
83cd4fe2 | 9187 | } |
c5afb6a8 VG |
9188 | |
9189 | /* | |
9190 | * next_balance will be updated only when there is a need. | |
9191 | * When the CPU is attached to null domain for ex, it will not be | |
9192 | * updated. | |
9193 | */ | |
9194 | if (likely(update_next_balance)) | |
9195 | nohz.next_balance = next_balance; | |
1c792db7 SS |
9196 | end: |
9197 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
9198 | } |
9199 | ||
9200 | /* | |
0b005cf5 | 9201 | * Current heuristic for kicking the idle load balancer in the presence |
1aaf90a4 | 9202 | * of an idle cpu in the system. |
0b005cf5 | 9203 | * - This rq has more than one task. |
1aaf90a4 VG |
9204 | * - This rq has at least one CFS task and the capacity of the CPU is |
9205 | * significantly reduced because of RT tasks or IRQs. | |
9206 | * - At parent of LLC scheduler domain level, this cpu's scheduler group has | |
9207 | * multiple busy cpu. | |
0b005cf5 SS |
9208 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler |
9209 | * domain span are idle. | |
83cd4fe2 | 9210 | */ |
1aaf90a4 | 9211 | static inline bool nohz_kick_needed(struct rq *rq) |
83cd4fe2 VP |
9212 | { |
9213 | unsigned long now = jiffies; | |
0e369d75 | 9214 | struct sched_domain_shared *sds; |
0b005cf5 | 9215 | struct sched_domain *sd; |
afe06efd | 9216 | int nr_busy, i, cpu = rq->cpu; |
1aaf90a4 | 9217 | bool kick = false; |
83cd4fe2 | 9218 | |
4a725627 | 9219 | if (unlikely(rq->idle_balance)) |
1aaf90a4 | 9220 | return false; |
83cd4fe2 | 9221 | |
1c792db7 SS |
9222 | /* |
9223 | * We may be recently in ticked or tickless idle mode. At the first | |
9224 | * busy tick after returning from idle, we will update the busy stats. | |
9225 | */ | |
69e1e811 | 9226 | set_cpu_sd_state_busy(); |
c1cc017c | 9227 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
9228 | |
9229 | /* | |
9230 | * None are in tickless mode and hence no need for NOHZ idle load | |
9231 | * balancing. | |
9232 | */ | |
9233 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
1aaf90a4 | 9234 | return false; |
1c792db7 SS |
9235 | |
9236 | if (time_before(now, nohz.next_balance)) | |
1aaf90a4 | 9237 | return false; |
83cd4fe2 | 9238 | |
0b005cf5 | 9239 | if (rq->nr_running >= 2) |
1aaf90a4 | 9240 | return true; |
83cd4fe2 | 9241 | |
067491b7 | 9242 | rcu_read_lock(); |
0e369d75 PZ |
9243 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
9244 | if (sds) { | |
9245 | /* | |
9246 | * XXX: write a coherent comment on why we do this. | |
9247 | * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com | |
9248 | */ | |
9249 | nr_busy = atomic_read(&sds->nr_busy_cpus); | |
1aaf90a4 VG |
9250 | if (nr_busy > 1) { |
9251 | kick = true; | |
9252 | goto unlock; | |
9253 | } | |
9254 | ||
83cd4fe2 | 9255 | } |
37dc6b50 | 9256 | |
1aaf90a4 VG |
9257 | sd = rcu_dereference(rq->sd); |
9258 | if (sd) { | |
9259 | if ((rq->cfs.h_nr_running >= 1) && | |
9260 | check_cpu_capacity(rq, sd)) { | |
9261 | kick = true; | |
9262 | goto unlock; | |
9263 | } | |
9264 | } | |
37dc6b50 | 9265 | |
1aaf90a4 | 9266 | sd = rcu_dereference(per_cpu(sd_asym, cpu)); |
afe06efd TC |
9267 | if (sd) { |
9268 | for_each_cpu(i, sched_domain_span(sd)) { | |
9269 | if (i == cpu || | |
9270 | !cpumask_test_cpu(i, nohz.idle_cpus_mask)) | |
9271 | continue; | |
067491b7 | 9272 | |
afe06efd TC |
9273 | if (sched_asym_prefer(i, cpu)) { |
9274 | kick = true; | |
9275 | goto unlock; | |
9276 | } | |
9277 | } | |
9278 | } | |
1aaf90a4 | 9279 | unlock: |
067491b7 | 9280 | rcu_read_unlock(); |
1aaf90a4 | 9281 | return kick; |
83cd4fe2 VP |
9282 | } |
9283 | #else | |
208cb16b | 9284 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } |
83cd4fe2 VP |
9285 | #endif |
9286 | ||
9287 | /* | |
9288 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
9289 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
9290 | */ | |
0766f788 | 9291 | static __latent_entropy void run_rebalance_domains(struct softirq_action *h) |
1e3c88bd | 9292 | { |
208cb16b | 9293 | struct rq *this_rq = this_rq(); |
6eb57e0d | 9294 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
9295 | CPU_IDLE : CPU_NOT_IDLE; |
9296 | ||
1e3c88bd | 9297 | /* |
83cd4fe2 | 9298 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd | 9299 | * balancing on behalf of the other idle cpus whose ticks are |
d4573c3e PM |
9300 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
9301 | * give the idle cpus a chance to load balance. Else we may | |
9302 | * load balance only within the local sched_domain hierarchy | |
9303 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 9304 | */ |
208cb16b | 9305 | nohz_idle_balance(this_rq, idle); |
d4573c3e | 9306 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
9307 | } |
9308 | ||
1e3c88bd PZ |
9309 | /* |
9310 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 9311 | */ |
7caff66f | 9312 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 9313 | { |
1e3c88bd | 9314 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
9315 | if (unlikely(on_null_domain(rq))) |
9316 | return; | |
9317 | ||
9318 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 9319 | raise_softirq(SCHED_SOFTIRQ); |
3451d024 | 9320 | #ifdef CONFIG_NO_HZ_COMMON |
c726099e | 9321 | if (nohz_kick_needed(rq)) |
0aeeeeba | 9322 | nohz_balancer_kick(); |
83cd4fe2 | 9323 | #endif |
1e3c88bd PZ |
9324 | } |
9325 | ||
0bcdcf28 CE |
9326 | static void rq_online_fair(struct rq *rq) |
9327 | { | |
9328 | update_sysctl(); | |
0e59bdae KT |
9329 | |
9330 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
9331 | } |
9332 | ||
9333 | static void rq_offline_fair(struct rq *rq) | |
9334 | { | |
9335 | update_sysctl(); | |
a4c96ae3 PB |
9336 | |
9337 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
9338 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
9339 | } |
9340 | ||
55e12e5e | 9341 | #endif /* CONFIG_SMP */ |
e1d1484f | 9342 | |
bf0f6f24 IM |
9343 | /* |
9344 | * scheduler tick hitting a task of our scheduling class: | |
9345 | */ | |
8f4d37ec | 9346 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
9347 | { |
9348 | struct cfs_rq *cfs_rq; | |
9349 | struct sched_entity *se = &curr->se; | |
9350 | ||
9351 | for_each_sched_entity(se) { | |
9352 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 9353 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 9354 | } |
18bf2805 | 9355 | |
b52da86e | 9356 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 9357 | task_tick_numa(rq, curr); |
bf0f6f24 IM |
9358 | } |
9359 | ||
9360 | /* | |
cd29fe6f PZ |
9361 | * called on fork with the child task as argument from the parent's context |
9362 | * - child not yet on the tasklist | |
9363 | * - preemption disabled | |
bf0f6f24 | 9364 | */ |
cd29fe6f | 9365 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 9366 | { |
4fc420c9 DN |
9367 | struct cfs_rq *cfs_rq; |
9368 | struct sched_entity *se = &p->se, *curr; | |
cd29fe6f | 9369 | struct rq *rq = this_rq(); |
8a8c69c3 | 9370 | struct rq_flags rf; |
bf0f6f24 | 9371 | |
8a8c69c3 | 9372 | rq_lock(rq, &rf); |
861d034e PZ |
9373 | update_rq_clock(rq); |
9374 | ||
4fc420c9 DN |
9375 | cfs_rq = task_cfs_rq(current); |
9376 | curr = cfs_rq->curr; | |
e210bffd PZ |
9377 | if (curr) { |
9378 | update_curr(cfs_rq); | |
b5d9d734 | 9379 | se->vruntime = curr->vruntime; |
e210bffd | 9380 | } |
aeb73b04 | 9381 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 9382 | |
cd29fe6f | 9383 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 9384 | /* |
edcb60a3 IM |
9385 | * Upon rescheduling, sched_class::put_prev_task() will place |
9386 | * 'current' within the tree based on its new key value. | |
9387 | */ | |
4d78e7b6 | 9388 | swap(curr->vruntime, se->vruntime); |
8875125e | 9389 | resched_curr(rq); |
4d78e7b6 | 9390 | } |
bf0f6f24 | 9391 | |
88ec22d3 | 9392 | se->vruntime -= cfs_rq->min_vruntime; |
8a8c69c3 | 9393 | rq_unlock(rq, &rf); |
bf0f6f24 IM |
9394 | } |
9395 | ||
cb469845 SR |
9396 | /* |
9397 | * Priority of the task has changed. Check to see if we preempt | |
9398 | * the current task. | |
9399 | */ | |
da7a735e PZ |
9400 | static void |
9401 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 9402 | { |
da0c1e65 | 9403 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
9404 | return; |
9405 | ||
cb469845 SR |
9406 | /* |
9407 | * Reschedule if we are currently running on this runqueue and | |
9408 | * our priority decreased, or if we are not currently running on | |
9409 | * this runqueue and our priority is higher than the current's | |
9410 | */ | |
da7a735e | 9411 | if (rq->curr == p) { |
cb469845 | 9412 | if (p->prio > oldprio) |
8875125e | 9413 | resched_curr(rq); |
cb469845 | 9414 | } else |
15afe09b | 9415 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
9416 | } |
9417 | ||
daa59407 | 9418 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
9419 | { |
9420 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
9421 | |
9422 | /* | |
daa59407 BP |
9423 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
9424 | * the dequeue_entity(.flags=0) will already have normalized the | |
9425 | * vruntime. | |
9426 | */ | |
9427 | if (p->on_rq) | |
9428 | return true; | |
9429 | ||
9430 | /* | |
9431 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
9432 | * But there are some cases where it has already been normalized: | |
da7a735e | 9433 | * |
daa59407 BP |
9434 | * - A forked child which is waiting for being woken up by |
9435 | * wake_up_new_task(). | |
9436 | * - A task which has been woken up by try_to_wake_up() and | |
9437 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 9438 | */ |
daa59407 BP |
9439 | if (!se->sum_exec_runtime || p->state == TASK_WAKING) |
9440 | return true; | |
9441 | ||
9442 | return false; | |
9443 | } | |
9444 | ||
09a43ace VG |
9445 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9446 | /* | |
9447 | * Propagate the changes of the sched_entity across the tg tree to make it | |
9448 | * visible to the root | |
9449 | */ | |
9450 | static void propagate_entity_cfs_rq(struct sched_entity *se) | |
9451 | { | |
9452 | struct cfs_rq *cfs_rq; | |
9453 | ||
9454 | /* Start to propagate at parent */ | |
9455 | se = se->parent; | |
9456 | ||
9457 | for_each_sched_entity(se) { | |
9458 | cfs_rq = cfs_rq_of(se); | |
9459 | ||
9460 | if (cfs_rq_throttled(cfs_rq)) | |
9461 | break; | |
9462 | ||
88c0616e | 9463 | update_load_avg(cfs_rq, se, UPDATE_TG); |
09a43ace VG |
9464 | } |
9465 | } | |
9466 | #else | |
9467 | static void propagate_entity_cfs_rq(struct sched_entity *se) { } | |
9468 | #endif | |
9469 | ||
df217913 | 9470 | static void detach_entity_cfs_rq(struct sched_entity *se) |
daa59407 | 9471 | { |
daa59407 BP |
9472 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
9473 | ||
9d89c257 | 9474 | /* Catch up with the cfs_rq and remove our load when we leave */ |
88c0616e | 9475 | update_load_avg(cfs_rq, se, 0); |
a05e8c51 | 9476 | detach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 9477 | update_tg_load_avg(cfs_rq, false); |
09a43ace | 9478 | propagate_entity_cfs_rq(se); |
da7a735e PZ |
9479 | } |
9480 | ||
df217913 | 9481 | static void attach_entity_cfs_rq(struct sched_entity *se) |
cb469845 | 9482 | { |
daa59407 | 9483 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a BP |
9484 | |
9485 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
eb7a59b2 M |
9486 | /* |
9487 | * Since the real-depth could have been changed (only FAIR | |
9488 | * class maintain depth value), reset depth properly. | |
9489 | */ | |
9490 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
9491 | #endif | |
7855a35a | 9492 | |
df217913 | 9493 | /* Synchronize entity with its cfs_rq */ |
88c0616e | 9494 | update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); |
daa59407 | 9495 | attach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 9496 | update_tg_load_avg(cfs_rq, false); |
09a43ace | 9497 | propagate_entity_cfs_rq(se); |
df217913 VG |
9498 | } |
9499 | ||
9500 | static void detach_task_cfs_rq(struct task_struct *p) | |
9501 | { | |
9502 | struct sched_entity *se = &p->se; | |
9503 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
9504 | ||
9505 | if (!vruntime_normalized(p)) { | |
9506 | /* | |
9507 | * Fix up our vruntime so that the current sleep doesn't | |
9508 | * cause 'unlimited' sleep bonus. | |
9509 | */ | |
9510 | place_entity(cfs_rq, se, 0); | |
9511 | se->vruntime -= cfs_rq->min_vruntime; | |
9512 | } | |
9513 | ||
9514 | detach_entity_cfs_rq(se); | |
9515 | } | |
9516 | ||
9517 | static void attach_task_cfs_rq(struct task_struct *p) | |
9518 | { | |
9519 | struct sched_entity *se = &p->se; | |
9520 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
9521 | ||
9522 | attach_entity_cfs_rq(se); | |
daa59407 BP |
9523 | |
9524 | if (!vruntime_normalized(p)) | |
9525 | se->vruntime += cfs_rq->min_vruntime; | |
9526 | } | |
6efdb105 | 9527 | |
daa59407 BP |
9528 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
9529 | { | |
9530 | detach_task_cfs_rq(p); | |
9531 | } | |
9532 | ||
9533 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
9534 | { | |
9535 | attach_task_cfs_rq(p); | |
7855a35a | 9536 | |
daa59407 | 9537 | if (task_on_rq_queued(p)) { |
7855a35a | 9538 | /* |
daa59407 BP |
9539 | * We were most likely switched from sched_rt, so |
9540 | * kick off the schedule if running, otherwise just see | |
9541 | * if we can still preempt the current task. | |
7855a35a | 9542 | */ |
daa59407 BP |
9543 | if (rq->curr == p) |
9544 | resched_curr(rq); | |
9545 | else | |
9546 | check_preempt_curr(rq, p, 0); | |
7855a35a | 9547 | } |
cb469845 SR |
9548 | } |
9549 | ||
83b699ed SV |
9550 | /* Account for a task changing its policy or group. |
9551 | * | |
9552 | * This routine is mostly called to set cfs_rq->curr field when a task | |
9553 | * migrates between groups/classes. | |
9554 | */ | |
9555 | static void set_curr_task_fair(struct rq *rq) | |
9556 | { | |
9557 | struct sched_entity *se = &rq->curr->se; | |
9558 | ||
ec12cb7f PT |
9559 | for_each_sched_entity(se) { |
9560 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
9561 | ||
9562 | set_next_entity(cfs_rq, se); | |
9563 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
9564 | account_cfs_rq_runtime(cfs_rq, 0); | |
9565 | } | |
83b699ed SV |
9566 | } |
9567 | ||
029632fb PZ |
9568 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
9569 | { | |
bfb06889 | 9570 | cfs_rq->tasks_timeline = RB_ROOT_CACHED; |
029632fb PZ |
9571 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
9572 | #ifndef CONFIG_64BIT | |
9573 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
9574 | #endif | |
141965c7 | 9575 | #ifdef CONFIG_SMP |
2a2f5d4e | 9576 | raw_spin_lock_init(&cfs_rq->removed.lock); |
9ee474f5 | 9577 | #endif |
029632fb PZ |
9578 | } |
9579 | ||
810b3817 | 9580 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b VG |
9581 | static void task_set_group_fair(struct task_struct *p) |
9582 | { | |
9583 | struct sched_entity *se = &p->se; | |
9584 | ||
9585 | set_task_rq(p, task_cpu(p)); | |
9586 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
9587 | } | |
9588 | ||
bc54da21 | 9589 | static void task_move_group_fair(struct task_struct *p) |
810b3817 | 9590 | { |
daa59407 | 9591 | detach_task_cfs_rq(p); |
b2b5ce02 | 9592 | set_task_rq(p, task_cpu(p)); |
6efdb105 BP |
9593 | |
9594 | #ifdef CONFIG_SMP | |
9595 | /* Tell se's cfs_rq has been changed -- migrated */ | |
9596 | p->se.avg.last_update_time = 0; | |
9597 | #endif | |
daa59407 | 9598 | attach_task_cfs_rq(p); |
810b3817 | 9599 | } |
029632fb | 9600 | |
ea86cb4b VG |
9601 | static void task_change_group_fair(struct task_struct *p, int type) |
9602 | { | |
9603 | switch (type) { | |
9604 | case TASK_SET_GROUP: | |
9605 | task_set_group_fair(p); | |
9606 | break; | |
9607 | ||
9608 | case TASK_MOVE_GROUP: | |
9609 | task_move_group_fair(p); | |
9610 | break; | |
9611 | } | |
9612 | } | |
9613 | ||
029632fb PZ |
9614 | void free_fair_sched_group(struct task_group *tg) |
9615 | { | |
9616 | int i; | |
9617 | ||
9618 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
9619 | ||
9620 | for_each_possible_cpu(i) { | |
9621 | if (tg->cfs_rq) | |
9622 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 9623 | if (tg->se) |
029632fb PZ |
9624 | kfree(tg->se[i]); |
9625 | } | |
9626 | ||
9627 | kfree(tg->cfs_rq); | |
9628 | kfree(tg->se); | |
9629 | } | |
9630 | ||
9631 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
9632 | { | |
029632fb | 9633 | struct sched_entity *se; |
b7fa30c9 | 9634 | struct cfs_rq *cfs_rq; |
029632fb PZ |
9635 | int i; |
9636 | ||
9637 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
9638 | if (!tg->cfs_rq) | |
9639 | goto err; | |
9640 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
9641 | if (!tg->se) | |
9642 | goto err; | |
9643 | ||
9644 | tg->shares = NICE_0_LOAD; | |
9645 | ||
9646 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
9647 | ||
9648 | for_each_possible_cpu(i) { | |
9649 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
9650 | GFP_KERNEL, cpu_to_node(i)); | |
9651 | if (!cfs_rq) | |
9652 | goto err; | |
9653 | ||
9654 | se = kzalloc_node(sizeof(struct sched_entity), | |
9655 | GFP_KERNEL, cpu_to_node(i)); | |
9656 | if (!se) | |
9657 | goto err_free_rq; | |
9658 | ||
9659 | init_cfs_rq(cfs_rq); | |
9660 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 9661 | init_entity_runnable_average(se); |
029632fb PZ |
9662 | } |
9663 | ||
9664 | return 1; | |
9665 | ||
9666 | err_free_rq: | |
9667 | kfree(cfs_rq); | |
9668 | err: | |
9669 | return 0; | |
9670 | } | |
9671 | ||
8663e24d PZ |
9672 | void online_fair_sched_group(struct task_group *tg) |
9673 | { | |
9674 | struct sched_entity *se; | |
9675 | struct rq *rq; | |
9676 | int i; | |
9677 | ||
9678 | for_each_possible_cpu(i) { | |
9679 | rq = cpu_rq(i); | |
9680 | se = tg->se[i]; | |
9681 | ||
9682 | raw_spin_lock_irq(&rq->lock); | |
4126bad6 | 9683 | update_rq_clock(rq); |
d0326691 | 9684 | attach_entity_cfs_rq(se); |
55e16d30 | 9685 | sync_throttle(tg, i); |
8663e24d PZ |
9686 | raw_spin_unlock_irq(&rq->lock); |
9687 | } | |
9688 | } | |
9689 | ||
6fe1f348 | 9690 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 9691 | { |
029632fb | 9692 | unsigned long flags; |
6fe1f348 PZ |
9693 | struct rq *rq; |
9694 | int cpu; | |
029632fb | 9695 | |
6fe1f348 PZ |
9696 | for_each_possible_cpu(cpu) { |
9697 | if (tg->se[cpu]) | |
9698 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 9699 | |
6fe1f348 PZ |
9700 | /* |
9701 | * Only empty task groups can be destroyed; so we can speculatively | |
9702 | * check on_list without danger of it being re-added. | |
9703 | */ | |
9704 | if (!tg->cfs_rq[cpu]->on_list) | |
9705 | continue; | |
9706 | ||
9707 | rq = cpu_rq(cpu); | |
9708 | ||
9709 | raw_spin_lock_irqsave(&rq->lock, flags); | |
9710 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
9711 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9712 | } | |
029632fb PZ |
9713 | } |
9714 | ||
9715 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
9716 | struct sched_entity *se, int cpu, | |
9717 | struct sched_entity *parent) | |
9718 | { | |
9719 | struct rq *rq = cpu_rq(cpu); | |
9720 | ||
9721 | cfs_rq->tg = tg; | |
9722 | cfs_rq->rq = rq; | |
029632fb PZ |
9723 | init_cfs_rq_runtime(cfs_rq); |
9724 | ||
9725 | tg->cfs_rq[cpu] = cfs_rq; | |
9726 | tg->se[cpu] = se; | |
9727 | ||
9728 | /* se could be NULL for root_task_group */ | |
9729 | if (!se) | |
9730 | return; | |
9731 | ||
fed14d45 | 9732 | if (!parent) { |
029632fb | 9733 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
9734 | se->depth = 0; |
9735 | } else { | |
029632fb | 9736 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
9737 | se->depth = parent->depth + 1; |
9738 | } | |
029632fb PZ |
9739 | |
9740 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
9741 | /* guarantee group entities always have weight */ |
9742 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
9743 | se->parent = parent; |
9744 | } | |
9745 | ||
9746 | static DEFINE_MUTEX(shares_mutex); | |
9747 | ||
9748 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
9749 | { | |
9750 | int i; | |
029632fb PZ |
9751 | |
9752 | /* | |
9753 | * We can't change the weight of the root cgroup. | |
9754 | */ | |
9755 | if (!tg->se[0]) | |
9756 | return -EINVAL; | |
9757 | ||
9758 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
9759 | ||
9760 | mutex_lock(&shares_mutex); | |
9761 | if (tg->shares == shares) | |
9762 | goto done; | |
9763 | ||
9764 | tg->shares = shares; | |
9765 | for_each_possible_cpu(i) { | |
9766 | struct rq *rq = cpu_rq(i); | |
8a8c69c3 PZ |
9767 | struct sched_entity *se = tg->se[i]; |
9768 | struct rq_flags rf; | |
029632fb | 9769 | |
029632fb | 9770 | /* Propagate contribution to hierarchy */ |
8a8c69c3 | 9771 | rq_lock_irqsave(rq, &rf); |
71b1da46 | 9772 | update_rq_clock(rq); |
89ee048f | 9773 | for_each_sched_entity(se) { |
88c0616e | 9774 | update_load_avg(cfs_rq_of(se), se, UPDATE_TG); |
1ea6c46a | 9775 | update_cfs_group(se); |
89ee048f | 9776 | } |
8a8c69c3 | 9777 | rq_unlock_irqrestore(rq, &rf); |
029632fb PZ |
9778 | } |
9779 | ||
9780 | done: | |
9781 | mutex_unlock(&shares_mutex); | |
9782 | return 0; | |
9783 | } | |
9784 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
9785 | ||
9786 | void free_fair_sched_group(struct task_group *tg) { } | |
9787 | ||
9788 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
9789 | { | |
9790 | return 1; | |
9791 | } | |
9792 | ||
8663e24d PZ |
9793 | void online_fair_sched_group(struct task_group *tg) { } |
9794 | ||
6fe1f348 | 9795 | void unregister_fair_sched_group(struct task_group *tg) { } |
029632fb PZ |
9796 | |
9797 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
9798 | ||
810b3817 | 9799 | |
6d686f45 | 9800 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
9801 | { |
9802 | struct sched_entity *se = &task->se; | |
0d721cea PW |
9803 | unsigned int rr_interval = 0; |
9804 | ||
9805 | /* | |
9806 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
9807 | * idle runqueue: | |
9808 | */ | |
0d721cea | 9809 | if (rq->cfs.load.weight) |
a59f4e07 | 9810 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
9811 | |
9812 | return rr_interval; | |
9813 | } | |
9814 | ||
bf0f6f24 IM |
9815 | /* |
9816 | * All the scheduling class methods: | |
9817 | */ | |
029632fb | 9818 | const struct sched_class fair_sched_class = { |
5522d5d5 | 9819 | .next = &idle_sched_class, |
bf0f6f24 IM |
9820 | .enqueue_task = enqueue_task_fair, |
9821 | .dequeue_task = dequeue_task_fair, | |
9822 | .yield_task = yield_task_fair, | |
d95f4122 | 9823 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 9824 | |
2e09bf55 | 9825 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
9826 | |
9827 | .pick_next_task = pick_next_task_fair, | |
9828 | .put_prev_task = put_prev_task_fair, | |
9829 | ||
681f3e68 | 9830 | #ifdef CONFIG_SMP |
4ce72a2c | 9831 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 9832 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 9833 | |
0bcdcf28 CE |
9834 | .rq_online = rq_online_fair, |
9835 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 9836 | |
12695578 | 9837 | .task_dead = task_dead_fair, |
c5b28038 | 9838 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 9839 | #endif |
bf0f6f24 | 9840 | |
83b699ed | 9841 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 9842 | .task_tick = task_tick_fair, |
cd29fe6f | 9843 | .task_fork = task_fork_fair, |
cb469845 SR |
9844 | |
9845 | .prio_changed = prio_changed_fair, | |
da7a735e | 9846 | .switched_from = switched_from_fair, |
cb469845 | 9847 | .switched_to = switched_to_fair, |
810b3817 | 9848 | |
0d721cea PW |
9849 | .get_rr_interval = get_rr_interval_fair, |
9850 | ||
6e998916 SG |
9851 | .update_curr = update_curr_fair, |
9852 | ||
810b3817 | 9853 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 9854 | .task_change_group = task_change_group_fair, |
810b3817 | 9855 | #endif |
bf0f6f24 IM |
9856 | }; |
9857 | ||
9858 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 9859 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 9860 | { |
a9e7f654 | 9861 | struct cfs_rq *cfs_rq, *pos; |
bf0f6f24 | 9862 | |
5973e5b9 | 9863 | rcu_read_lock(); |
a9e7f654 | 9864 | for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos) |
5cef9eca | 9865 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 9866 | rcu_read_unlock(); |
bf0f6f24 | 9867 | } |
397f2378 SD |
9868 | |
9869 | #ifdef CONFIG_NUMA_BALANCING | |
9870 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
9871 | { | |
9872 | int node; | |
9873 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
9874 | ||
9875 | for_each_online_node(node) { | |
9876 | if (p->numa_faults) { | |
9877 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
9878 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
9879 | } | |
9880 | if (p->numa_group) { | |
9881 | gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
9882 | gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
9883 | } | |
9884 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
9885 | } | |
9886 | } | |
9887 | #endif /* CONFIG_NUMA_BALANCING */ | |
9888 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
9889 | |
9890 | __init void init_sched_fair_class(void) | |
9891 | { | |
9892 | #ifdef CONFIG_SMP | |
9893 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
9894 | ||
3451d024 | 9895 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 9896 | nohz.next_balance = jiffies; |
029632fb | 9897 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
9898 | #endif |
9899 | #endif /* SMP */ | |
9900 | ||
9901 | } |