]>
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 PZ |
371 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
372 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ | |
373 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) | |
374 | ||
375 | /* Do the two (enqueued) entities belong to the same group ? */ | |
fed14d45 | 376 | static inline struct cfs_rq * |
b758149c PZ |
377 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
378 | { | |
379 | if (se->cfs_rq == pse->cfs_rq) | |
fed14d45 | 380 | return se->cfs_rq; |
b758149c | 381 | |
fed14d45 | 382 | return NULL; |
b758149c PZ |
383 | } |
384 | ||
385 | static inline struct sched_entity *parent_entity(struct sched_entity *se) | |
386 | { | |
387 | return se->parent; | |
388 | } | |
389 | ||
464b7527 PZ |
390 | static void |
391 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
392 | { | |
393 | int se_depth, pse_depth; | |
394 | ||
395 | /* | |
396 | * preemption test can be made between sibling entities who are in the | |
397 | * same cfs_rq i.e who have a common parent. Walk up the hierarchy of | |
398 | * both tasks until we find their ancestors who are siblings of common | |
399 | * parent. | |
400 | */ | |
401 | ||
402 | /* First walk up until both entities are at same depth */ | |
fed14d45 PZ |
403 | se_depth = (*se)->depth; |
404 | pse_depth = (*pse)->depth; | |
464b7527 PZ |
405 | |
406 | while (se_depth > pse_depth) { | |
407 | se_depth--; | |
408 | *se = parent_entity(*se); | |
409 | } | |
410 | ||
411 | while (pse_depth > se_depth) { | |
412 | pse_depth--; | |
413 | *pse = parent_entity(*pse); | |
414 | } | |
415 | ||
416 | while (!is_same_group(*se, *pse)) { | |
417 | *se = parent_entity(*se); | |
418 | *pse = parent_entity(*pse); | |
419 | } | |
420 | } | |
421 | ||
8f48894f PZ |
422 | #else /* !CONFIG_FAIR_GROUP_SCHED */ |
423 | ||
424 | static inline struct task_struct *task_of(struct sched_entity *se) | |
425 | { | |
426 | return container_of(se, struct task_struct, se); | |
427 | } | |
bf0f6f24 | 428 | |
62160e3f IM |
429 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
430 | { | |
431 | return container_of(cfs_rq, struct rq, cfs); | |
bf0f6f24 IM |
432 | } |
433 | ||
434 | #define entity_is_task(se) 1 | |
435 | ||
b758149c PZ |
436 | #define for_each_sched_entity(se) \ |
437 | for (; se; se = NULL) | |
bf0f6f24 | 438 | |
b758149c | 439 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
bf0f6f24 | 440 | { |
b758149c | 441 | return &task_rq(p)->cfs; |
bf0f6f24 IM |
442 | } |
443 | ||
b758149c PZ |
444 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
445 | { | |
446 | struct task_struct *p = task_of(se); | |
447 | struct rq *rq = task_rq(p); | |
448 | ||
449 | return &rq->cfs; | |
450 | } | |
451 | ||
452 | /* runqueue "owned" by this group */ | |
453 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) | |
454 | { | |
455 | return NULL; | |
456 | } | |
457 | ||
3d4b47b4 PZ |
458 | static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) |
459 | { | |
460 | } | |
461 | ||
462 | static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) | |
463 | { | |
464 | } | |
465 | ||
b758149c PZ |
466 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
467 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) | |
468 | ||
b758149c PZ |
469 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
470 | { | |
471 | return NULL; | |
472 | } | |
473 | ||
464b7527 PZ |
474 | static inline void |
475 | find_matching_se(struct sched_entity **se, struct sched_entity **pse) | |
476 | { | |
477 | } | |
478 | ||
b758149c PZ |
479 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
480 | ||
6c16a6dc | 481 | static __always_inline |
9dbdb155 | 482 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); |
bf0f6f24 IM |
483 | |
484 | /************************************************************** | |
485 | * Scheduling class tree data structure manipulation methods: | |
486 | */ | |
487 | ||
1bf08230 | 488 | static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) |
02e0431a | 489 | { |
1bf08230 | 490 | s64 delta = (s64)(vruntime - max_vruntime); |
368059a9 | 491 | if (delta > 0) |
1bf08230 | 492 | max_vruntime = vruntime; |
02e0431a | 493 | |
1bf08230 | 494 | return max_vruntime; |
02e0431a PZ |
495 | } |
496 | ||
0702e3eb | 497 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
b0ffd246 PZ |
498 | { |
499 | s64 delta = (s64)(vruntime - min_vruntime); | |
500 | if (delta < 0) | |
501 | min_vruntime = vruntime; | |
502 | ||
503 | return min_vruntime; | |
504 | } | |
505 | ||
54fdc581 FC |
506 | static inline int entity_before(struct sched_entity *a, |
507 | struct sched_entity *b) | |
508 | { | |
509 | return (s64)(a->vruntime - b->vruntime) < 0; | |
510 | } | |
511 | ||
1af5f730 PZ |
512 | static void update_min_vruntime(struct cfs_rq *cfs_rq) |
513 | { | |
b60205c7 PZ |
514 | struct sched_entity *curr = cfs_rq->curr; |
515 | ||
1af5f730 PZ |
516 | u64 vruntime = cfs_rq->min_vruntime; |
517 | ||
b60205c7 PZ |
518 | if (curr) { |
519 | if (curr->on_rq) | |
520 | vruntime = curr->vruntime; | |
521 | else | |
522 | curr = NULL; | |
523 | } | |
1af5f730 PZ |
524 | |
525 | if (cfs_rq->rb_leftmost) { | |
526 | struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, | |
527 | struct sched_entity, | |
528 | run_node); | |
529 | ||
b60205c7 | 530 | if (!curr) |
1af5f730 PZ |
531 | vruntime = se->vruntime; |
532 | else | |
533 | vruntime = min_vruntime(vruntime, se->vruntime); | |
534 | } | |
535 | ||
1bf08230 | 536 | /* ensure we never gain time by being placed backwards. */ |
1af5f730 | 537 | cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); |
3fe1698b PZ |
538 | #ifndef CONFIG_64BIT |
539 | smp_wmb(); | |
540 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
541 | #endif | |
1af5f730 PZ |
542 | } |
543 | ||
bf0f6f24 IM |
544 | /* |
545 | * Enqueue an entity into the rb-tree: | |
546 | */ | |
0702e3eb | 547 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
548 | { |
549 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; | |
550 | struct rb_node *parent = NULL; | |
551 | struct sched_entity *entry; | |
bf0f6f24 IM |
552 | int leftmost = 1; |
553 | ||
554 | /* | |
555 | * Find the right place in the rbtree: | |
556 | */ | |
557 | while (*link) { | |
558 | parent = *link; | |
559 | entry = rb_entry(parent, struct sched_entity, run_node); | |
560 | /* | |
561 | * We dont care about collisions. Nodes with | |
562 | * the same key stay together. | |
563 | */ | |
2bd2d6f2 | 564 | if (entity_before(se, entry)) { |
bf0f6f24 IM |
565 | link = &parent->rb_left; |
566 | } else { | |
567 | link = &parent->rb_right; | |
568 | leftmost = 0; | |
569 | } | |
570 | } | |
571 | ||
572 | /* | |
573 | * Maintain a cache of leftmost tree entries (it is frequently | |
574 | * used): | |
575 | */ | |
1af5f730 | 576 | if (leftmost) |
57cb499d | 577 | cfs_rq->rb_leftmost = &se->run_node; |
bf0f6f24 IM |
578 | |
579 | rb_link_node(&se->run_node, parent, link); | |
580 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); | |
bf0f6f24 IM |
581 | } |
582 | ||
0702e3eb | 583 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 584 | { |
3fe69747 PZ |
585 | if (cfs_rq->rb_leftmost == &se->run_node) { |
586 | struct rb_node *next_node; | |
3fe69747 PZ |
587 | |
588 | next_node = rb_next(&se->run_node); | |
589 | cfs_rq->rb_leftmost = next_node; | |
3fe69747 | 590 | } |
e9acbff6 | 591 | |
bf0f6f24 | 592 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
bf0f6f24 IM |
593 | } |
594 | ||
029632fb | 595 | struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) |
bf0f6f24 | 596 | { |
f4b6755f PZ |
597 | struct rb_node *left = cfs_rq->rb_leftmost; |
598 | ||
599 | if (!left) | |
600 | return NULL; | |
601 | ||
602 | return rb_entry(left, struct sched_entity, run_node); | |
bf0f6f24 IM |
603 | } |
604 | ||
ac53db59 RR |
605 | static struct sched_entity *__pick_next_entity(struct sched_entity *se) |
606 | { | |
607 | struct rb_node *next = rb_next(&se->run_node); | |
608 | ||
609 | if (!next) | |
610 | return NULL; | |
611 | ||
612 | return rb_entry(next, struct sched_entity, run_node); | |
613 | } | |
614 | ||
615 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 616 | struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
aeb73b04 | 617 | { |
7eee3e67 | 618 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
aeb73b04 | 619 | |
70eee74b BS |
620 | if (!last) |
621 | return NULL; | |
7eee3e67 IM |
622 | |
623 | return rb_entry(last, struct sched_entity, run_node); | |
aeb73b04 PZ |
624 | } |
625 | ||
bf0f6f24 IM |
626 | /************************************************************** |
627 | * Scheduling class statistics methods: | |
628 | */ | |
629 | ||
acb4a848 | 630 | int sched_proc_update_handler(struct ctl_table *table, int write, |
8d65af78 | 631 | void __user *buffer, size_t *lenp, |
b2be5e96 PZ |
632 | loff_t *ppos) |
633 | { | |
8d65af78 | 634 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
58ac93e4 | 635 | unsigned int factor = get_update_sysctl_factor(); |
b2be5e96 PZ |
636 | |
637 | if (ret || !write) | |
638 | return ret; | |
639 | ||
640 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, | |
641 | sysctl_sched_min_granularity); | |
642 | ||
acb4a848 CE |
643 | #define WRT_SYSCTL(name) \ |
644 | (normalized_sysctl_##name = sysctl_##name / (factor)) | |
645 | WRT_SYSCTL(sched_min_granularity); | |
646 | WRT_SYSCTL(sched_latency); | |
647 | WRT_SYSCTL(sched_wakeup_granularity); | |
acb4a848 CE |
648 | #undef WRT_SYSCTL |
649 | ||
b2be5e96 PZ |
650 | return 0; |
651 | } | |
652 | #endif | |
647e7cac | 653 | |
a7be37ac | 654 | /* |
f9c0b095 | 655 | * delta /= w |
a7be37ac | 656 | */ |
9dbdb155 | 657 | static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) |
a7be37ac | 658 | { |
f9c0b095 | 659 | if (unlikely(se->load.weight != NICE_0_LOAD)) |
9dbdb155 | 660 | delta = __calc_delta(delta, NICE_0_LOAD, &se->load); |
a7be37ac PZ |
661 | |
662 | return delta; | |
663 | } | |
664 | ||
647e7cac IM |
665 | /* |
666 | * The idea is to set a period in which each task runs once. | |
667 | * | |
532b1858 | 668 | * When there are too many tasks (sched_nr_latency) we have to stretch |
647e7cac IM |
669 | * this period because otherwise the slices get too small. |
670 | * | |
671 | * p = (nr <= nl) ? l : l*nr/nl | |
672 | */ | |
4d78e7b6 PZ |
673 | static u64 __sched_period(unsigned long nr_running) |
674 | { | |
8e2b0bf3 BF |
675 | if (unlikely(nr_running > sched_nr_latency)) |
676 | return nr_running * sysctl_sched_min_granularity; | |
677 | else | |
678 | return sysctl_sched_latency; | |
4d78e7b6 PZ |
679 | } |
680 | ||
647e7cac IM |
681 | /* |
682 | * We calculate the wall-time slice from the period by taking a part | |
683 | * proportional to the weight. | |
684 | * | |
f9c0b095 | 685 | * s = p*P[w/rw] |
647e7cac | 686 | */ |
6d0f0ebd | 687 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
21805085 | 688 | { |
0a582440 | 689 | u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); |
f9c0b095 | 690 | |
0a582440 | 691 | for_each_sched_entity(se) { |
6272d68c | 692 | struct load_weight *load; |
3104bf03 | 693 | struct load_weight lw; |
6272d68c LM |
694 | |
695 | cfs_rq = cfs_rq_of(se); | |
696 | load = &cfs_rq->load; | |
f9c0b095 | 697 | |
0a582440 | 698 | if (unlikely(!se->on_rq)) { |
3104bf03 | 699 | lw = cfs_rq->load; |
0a582440 MG |
700 | |
701 | update_load_add(&lw, se->load.weight); | |
702 | load = &lw; | |
703 | } | |
9dbdb155 | 704 | slice = __calc_delta(slice, se->load.weight, load); |
0a582440 MG |
705 | } |
706 | return slice; | |
bf0f6f24 IM |
707 | } |
708 | ||
647e7cac | 709 | /* |
660cc00f | 710 | * We calculate the vruntime slice of a to-be-inserted task. |
647e7cac | 711 | * |
f9c0b095 | 712 | * vs = s/w |
647e7cac | 713 | */ |
f9c0b095 | 714 | static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
67e9fb2a | 715 | { |
f9c0b095 | 716 | return calc_delta_fair(sched_slice(cfs_rq, se), se); |
a7be37ac PZ |
717 | } |
718 | ||
a75cdaa9 | 719 | #ifdef CONFIG_SMP |
772bd008 | 720 | static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); |
fb13c7ee MG |
721 | static unsigned long task_h_load(struct task_struct *p); |
722 | ||
9d89c257 YD |
723 | /* |
724 | * We choose a half-life close to 1 scheduling period. | |
84fb5a18 LY |
725 | * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are |
726 | * dependent on this value. | |
9d89c257 YD |
727 | */ |
728 | #define LOAD_AVG_PERIOD 32 | |
729 | #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ | |
a75cdaa9 | 730 | |
540247fb YD |
731 | /* Give new sched_entity start runnable values to heavy its load in infant time */ |
732 | void init_entity_runnable_average(struct sched_entity *se) | |
a75cdaa9 | 733 | { |
540247fb | 734 | struct sched_avg *sa = &se->avg; |
a75cdaa9 | 735 | |
9d89c257 YD |
736 | sa->last_update_time = 0; |
737 | /* | |
738 | * sched_avg's period_contrib should be strictly less then 1024, so | |
739 | * we give it 1023 to make sure it is almost a period (1024us), and | |
740 | * will definitely be update (after enqueue). | |
741 | */ | |
742 | sa->period_contrib = 1023; | |
b5a9b340 VG |
743 | /* |
744 | * Tasks are intialized with full load to be seen as heavy tasks until | |
745 | * they get a chance to stabilize to their real load level. | |
746 | * Group entities are intialized with zero load to reflect the fact that | |
747 | * nothing has been attached to the task group yet. | |
748 | */ | |
749 | if (entity_is_task(se)) | |
750 | sa->load_avg = scale_load_down(se->load.weight); | |
9d89c257 | 751 | sa->load_sum = sa->load_avg * LOAD_AVG_MAX; |
2b8c41da YD |
752 | /* |
753 | * At this point, util_avg won't be used in select_task_rq_fair anyway | |
754 | */ | |
755 | sa->util_avg = 0; | |
756 | sa->util_sum = 0; | |
9d89c257 | 757 | /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ |
a75cdaa9 | 758 | } |
7ea241af | 759 | |
7dc603c9 | 760 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); |
df217913 | 761 | static void attach_entity_cfs_rq(struct sched_entity *se); |
7dc603c9 | 762 | |
2b8c41da YD |
763 | /* |
764 | * With new tasks being created, their initial util_avgs are extrapolated | |
765 | * based on the cfs_rq's current util_avg: | |
766 | * | |
767 | * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight | |
768 | * | |
769 | * However, in many cases, the above util_avg does not give a desired | |
770 | * value. Moreover, the sum of the util_avgs may be divergent, such | |
771 | * as when the series is a harmonic series. | |
772 | * | |
773 | * To solve this problem, we also cap the util_avg of successive tasks to | |
774 | * only 1/2 of the left utilization budget: | |
775 | * | |
776 | * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n | |
777 | * | |
778 | * where n denotes the nth task. | |
779 | * | |
780 | * For example, a simplest series from the beginning would be like: | |
781 | * | |
782 | * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... | |
783 | * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... | |
784 | * | |
785 | * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) | |
786 | * if util_avg > util_avg_cap. | |
787 | */ | |
788 | void post_init_entity_util_avg(struct sched_entity *se) | |
789 | { | |
790 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
791 | struct sched_avg *sa = &se->avg; | |
172895e6 | 792 | long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2; |
2b8c41da YD |
793 | |
794 | if (cap > 0) { | |
795 | if (cfs_rq->avg.util_avg != 0) { | |
796 | sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; | |
797 | sa->util_avg /= (cfs_rq->avg.load_avg + 1); | |
798 | ||
799 | if (sa->util_avg > cap) | |
800 | sa->util_avg = cap; | |
801 | } else { | |
802 | sa->util_avg = cap; | |
803 | } | |
804 | sa->util_sum = sa->util_avg * LOAD_AVG_MAX; | |
805 | } | |
7dc603c9 PZ |
806 | |
807 | if (entity_is_task(se)) { | |
808 | struct task_struct *p = task_of(se); | |
809 | if (p->sched_class != &fair_sched_class) { | |
810 | /* | |
811 | * For !fair tasks do: | |
812 | * | |
813 | update_cfs_rq_load_avg(now, cfs_rq, false); | |
814 | attach_entity_load_avg(cfs_rq, se); | |
815 | switched_from_fair(rq, p); | |
816 | * | |
817 | * such that the next switched_to_fair() has the | |
818 | * expected state. | |
819 | */ | |
df217913 | 820 | se->avg.last_update_time = cfs_rq_clock_task(cfs_rq); |
7dc603c9 PZ |
821 | return; |
822 | } | |
823 | } | |
824 | ||
df217913 | 825 | attach_entity_cfs_rq(se); |
2b8c41da YD |
826 | } |
827 | ||
7dc603c9 | 828 | #else /* !CONFIG_SMP */ |
540247fb | 829 | void init_entity_runnable_average(struct sched_entity *se) |
a75cdaa9 AS |
830 | { |
831 | } | |
2b8c41da YD |
832 | void post_init_entity_util_avg(struct sched_entity *se) |
833 | { | |
834 | } | |
3d30544f PZ |
835 | static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
836 | { | |
837 | } | |
7dc603c9 | 838 | #endif /* CONFIG_SMP */ |
a75cdaa9 | 839 | |
bf0f6f24 | 840 | /* |
9dbdb155 | 841 | * Update the current task's runtime statistics. |
bf0f6f24 | 842 | */ |
b7cc0896 | 843 | static void update_curr(struct cfs_rq *cfs_rq) |
bf0f6f24 | 844 | { |
429d43bc | 845 | struct sched_entity *curr = cfs_rq->curr; |
78becc27 | 846 | u64 now = rq_clock_task(rq_of(cfs_rq)); |
9dbdb155 | 847 | u64 delta_exec; |
bf0f6f24 IM |
848 | |
849 | if (unlikely(!curr)) | |
850 | return; | |
851 | ||
9dbdb155 PZ |
852 | delta_exec = now - curr->exec_start; |
853 | if (unlikely((s64)delta_exec <= 0)) | |
34f28ecd | 854 | return; |
bf0f6f24 | 855 | |
8ebc91d9 | 856 | curr->exec_start = now; |
d842de87 | 857 | |
9dbdb155 PZ |
858 | schedstat_set(curr->statistics.exec_max, |
859 | max(delta_exec, curr->statistics.exec_max)); | |
860 | ||
861 | curr->sum_exec_runtime += delta_exec; | |
ae92882e | 862 | schedstat_add(cfs_rq->exec_clock, delta_exec); |
9dbdb155 PZ |
863 | |
864 | curr->vruntime += calc_delta_fair(delta_exec, curr); | |
865 | update_min_vruntime(cfs_rq); | |
866 | ||
d842de87 SV |
867 | if (entity_is_task(curr)) { |
868 | struct task_struct *curtask = task_of(curr); | |
869 | ||
f977bb49 | 870 | trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); |
d842de87 | 871 | cpuacct_charge(curtask, delta_exec); |
f06febc9 | 872 | account_group_exec_runtime(curtask, delta_exec); |
d842de87 | 873 | } |
ec12cb7f PT |
874 | |
875 | account_cfs_rq_runtime(cfs_rq, delta_exec); | |
bf0f6f24 IM |
876 | } |
877 | ||
6e998916 SG |
878 | static void update_curr_fair(struct rq *rq) |
879 | { | |
880 | update_curr(cfs_rq_of(&rq->curr->se)); | |
881 | } | |
882 | ||
bf0f6f24 | 883 | static inline void |
5870db5b | 884 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 885 | { |
4fa8d299 JP |
886 | u64 wait_start, prev_wait_start; |
887 | ||
888 | if (!schedstat_enabled()) | |
889 | return; | |
890 | ||
891 | wait_start = rq_clock(rq_of(cfs_rq)); | |
892 | prev_wait_start = schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
893 | |
894 | if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && | |
4fa8d299 JP |
895 | likely(wait_start > prev_wait_start)) |
896 | wait_start -= prev_wait_start; | |
3ea94de1 | 897 | |
4fa8d299 | 898 | schedstat_set(se->statistics.wait_start, wait_start); |
bf0f6f24 IM |
899 | } |
900 | ||
4fa8d299 | 901 | static inline void |
3ea94de1 JP |
902 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
903 | { | |
904 | struct task_struct *p; | |
cb251765 MG |
905 | u64 delta; |
906 | ||
4fa8d299 JP |
907 | if (!schedstat_enabled()) |
908 | return; | |
909 | ||
910 | delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); | |
3ea94de1 JP |
911 | |
912 | if (entity_is_task(se)) { | |
913 | p = task_of(se); | |
914 | if (task_on_rq_migrating(p)) { | |
915 | /* | |
916 | * Preserve migrating task's wait time so wait_start | |
917 | * time stamp can be adjusted to accumulate wait time | |
918 | * prior to migration. | |
919 | */ | |
4fa8d299 | 920 | schedstat_set(se->statistics.wait_start, delta); |
3ea94de1 JP |
921 | return; |
922 | } | |
923 | trace_sched_stat_wait(p, delta); | |
924 | } | |
925 | ||
4fa8d299 JP |
926 | schedstat_set(se->statistics.wait_max, |
927 | max(schedstat_val(se->statistics.wait_max), delta)); | |
928 | schedstat_inc(se->statistics.wait_count); | |
929 | schedstat_add(se->statistics.wait_sum, delta); | |
930 | schedstat_set(se->statistics.wait_start, 0); | |
3ea94de1 | 931 | } |
3ea94de1 | 932 | |
4fa8d299 | 933 | static inline void |
1a3d027c JP |
934 | update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
935 | { | |
936 | struct task_struct *tsk = NULL; | |
4fa8d299 JP |
937 | u64 sleep_start, block_start; |
938 | ||
939 | if (!schedstat_enabled()) | |
940 | return; | |
941 | ||
942 | sleep_start = schedstat_val(se->statistics.sleep_start); | |
943 | block_start = schedstat_val(se->statistics.block_start); | |
1a3d027c JP |
944 | |
945 | if (entity_is_task(se)) | |
946 | tsk = task_of(se); | |
947 | ||
4fa8d299 JP |
948 | if (sleep_start) { |
949 | u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; | |
1a3d027c JP |
950 | |
951 | if ((s64)delta < 0) | |
952 | delta = 0; | |
953 | ||
4fa8d299 JP |
954 | if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) |
955 | schedstat_set(se->statistics.sleep_max, delta); | |
1a3d027c | 956 | |
4fa8d299 JP |
957 | schedstat_set(se->statistics.sleep_start, 0); |
958 | schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
959 | |
960 | if (tsk) { | |
961 | account_scheduler_latency(tsk, delta >> 10, 1); | |
962 | trace_sched_stat_sleep(tsk, delta); | |
963 | } | |
964 | } | |
4fa8d299 JP |
965 | if (block_start) { |
966 | u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; | |
1a3d027c JP |
967 | |
968 | if ((s64)delta < 0) | |
969 | delta = 0; | |
970 | ||
4fa8d299 JP |
971 | if (unlikely(delta > schedstat_val(se->statistics.block_max))) |
972 | schedstat_set(se->statistics.block_max, delta); | |
1a3d027c | 973 | |
4fa8d299 JP |
974 | schedstat_set(se->statistics.block_start, 0); |
975 | schedstat_add(se->statistics.sum_sleep_runtime, delta); | |
1a3d027c JP |
976 | |
977 | if (tsk) { | |
978 | if (tsk->in_iowait) { | |
4fa8d299 JP |
979 | schedstat_add(se->statistics.iowait_sum, delta); |
980 | schedstat_inc(se->statistics.iowait_count); | |
1a3d027c JP |
981 | trace_sched_stat_iowait(tsk, delta); |
982 | } | |
983 | ||
984 | trace_sched_stat_blocked(tsk, delta); | |
985 | ||
986 | /* | |
987 | * Blocking time is in units of nanosecs, so shift by | |
988 | * 20 to get a milliseconds-range estimation of the | |
989 | * amount of time that the task spent sleeping: | |
990 | */ | |
991 | if (unlikely(prof_on == SLEEP_PROFILING)) { | |
992 | profile_hits(SLEEP_PROFILING, | |
993 | (void *)get_wchan(tsk), | |
994 | delta >> 20); | |
995 | } | |
996 | account_scheduler_latency(tsk, delta >> 10, 0); | |
997 | } | |
998 | } | |
3ea94de1 | 999 | } |
3ea94de1 | 1000 | |
bf0f6f24 IM |
1001 | /* |
1002 | * Task is being enqueued - update stats: | |
1003 | */ | |
cb251765 | 1004 | static inline void |
1a3d027c | 1005 | update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1006 | { |
4fa8d299 JP |
1007 | if (!schedstat_enabled()) |
1008 | return; | |
1009 | ||
bf0f6f24 IM |
1010 | /* |
1011 | * Are we enqueueing a waiting task? (for current tasks | |
1012 | * a dequeue/enqueue event is a NOP) | |
1013 | */ | |
429d43bc | 1014 | if (se != cfs_rq->curr) |
5870db5b | 1015 | update_stats_wait_start(cfs_rq, se); |
1a3d027c JP |
1016 | |
1017 | if (flags & ENQUEUE_WAKEUP) | |
1018 | update_stats_enqueue_sleeper(cfs_rq, se); | |
bf0f6f24 IM |
1019 | } |
1020 | ||
bf0f6f24 | 1021 | static inline void |
cb251765 | 1022 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 1023 | { |
4fa8d299 JP |
1024 | |
1025 | if (!schedstat_enabled()) | |
1026 | return; | |
1027 | ||
bf0f6f24 IM |
1028 | /* |
1029 | * Mark the end of the wait period if dequeueing a | |
1030 | * waiting task: | |
1031 | */ | |
429d43bc | 1032 | if (se != cfs_rq->curr) |
9ef0a961 | 1033 | update_stats_wait_end(cfs_rq, se); |
cb251765 | 1034 | |
4fa8d299 JP |
1035 | if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { |
1036 | struct task_struct *tsk = task_of(se); | |
cb251765 | 1037 | |
4fa8d299 JP |
1038 | if (tsk->state & TASK_INTERRUPTIBLE) |
1039 | schedstat_set(se->statistics.sleep_start, | |
1040 | rq_clock(rq_of(cfs_rq))); | |
1041 | if (tsk->state & TASK_UNINTERRUPTIBLE) | |
1042 | schedstat_set(se->statistics.block_start, | |
1043 | rq_clock(rq_of(cfs_rq))); | |
cb251765 | 1044 | } |
cb251765 MG |
1045 | } |
1046 | ||
bf0f6f24 IM |
1047 | /* |
1048 | * We are picking a new current task - update its stats: | |
1049 | */ | |
1050 | static inline void | |
79303e9e | 1051 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 IM |
1052 | { |
1053 | /* | |
1054 | * We are starting a new run period: | |
1055 | */ | |
78becc27 | 1056 | se->exec_start = rq_clock_task(rq_of(cfs_rq)); |
bf0f6f24 IM |
1057 | } |
1058 | ||
bf0f6f24 IM |
1059 | /************************************************** |
1060 | * Scheduling class queueing methods: | |
1061 | */ | |
1062 | ||
cbee9f88 PZ |
1063 | #ifdef CONFIG_NUMA_BALANCING |
1064 | /* | |
598f0ec0 MG |
1065 | * Approximate time to scan a full NUMA task in ms. The task scan period is |
1066 | * calculated based on the tasks virtual memory size and | |
1067 | * numa_balancing_scan_size. | |
cbee9f88 | 1068 | */ |
598f0ec0 MG |
1069 | unsigned int sysctl_numa_balancing_scan_period_min = 1000; |
1070 | unsigned int sysctl_numa_balancing_scan_period_max = 60000; | |
6e5fb223 PZ |
1071 | |
1072 | /* Portion of address space to scan in MB */ | |
1073 | unsigned int sysctl_numa_balancing_scan_size = 256; | |
cbee9f88 | 1074 | |
4b96a29b PZ |
1075 | /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ |
1076 | unsigned int sysctl_numa_balancing_scan_delay = 1000; | |
1077 | ||
598f0ec0 MG |
1078 | static unsigned int task_nr_scan_windows(struct task_struct *p) |
1079 | { | |
1080 | unsigned long rss = 0; | |
1081 | unsigned long nr_scan_pages; | |
1082 | ||
1083 | /* | |
1084 | * Calculations based on RSS as non-present and empty pages are skipped | |
1085 | * by the PTE scanner and NUMA hinting faults should be trapped based | |
1086 | * on resident pages | |
1087 | */ | |
1088 | nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); | |
1089 | rss = get_mm_rss(p->mm); | |
1090 | if (!rss) | |
1091 | rss = nr_scan_pages; | |
1092 | ||
1093 | rss = round_up(rss, nr_scan_pages); | |
1094 | return rss / nr_scan_pages; | |
1095 | } | |
1096 | ||
1097 | /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ | |
1098 | #define MAX_SCAN_WINDOW 2560 | |
1099 | ||
1100 | static unsigned int task_scan_min(struct task_struct *p) | |
1101 | { | |
316c1608 | 1102 | unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); |
598f0ec0 MG |
1103 | unsigned int scan, floor; |
1104 | unsigned int windows = 1; | |
1105 | ||
64192658 KT |
1106 | if (scan_size < MAX_SCAN_WINDOW) |
1107 | windows = MAX_SCAN_WINDOW / scan_size; | |
598f0ec0 MG |
1108 | floor = 1000 / windows; |
1109 | ||
1110 | scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); | |
1111 | return max_t(unsigned int, floor, scan); | |
1112 | } | |
1113 | ||
1114 | static unsigned int task_scan_max(struct task_struct *p) | |
1115 | { | |
1116 | unsigned int smin = task_scan_min(p); | |
1117 | unsigned int smax; | |
1118 | ||
1119 | /* Watch for min being lower than max due to floor calculations */ | |
1120 | smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); | |
1121 | return max(smin, smax); | |
1122 | } | |
1123 | ||
0ec8aa00 PZ |
1124 | static void account_numa_enqueue(struct rq *rq, struct task_struct *p) |
1125 | { | |
1126 | rq->nr_numa_running += (p->numa_preferred_nid != -1); | |
1127 | rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); | |
1128 | } | |
1129 | ||
1130 | static void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
1131 | { | |
1132 | rq->nr_numa_running -= (p->numa_preferred_nid != -1); | |
1133 | rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); | |
1134 | } | |
1135 | ||
8c8a743c PZ |
1136 | struct numa_group { |
1137 | atomic_t refcount; | |
1138 | ||
1139 | spinlock_t lock; /* nr_tasks, tasks */ | |
1140 | int nr_tasks; | |
e29cf08b | 1141 | pid_t gid; |
4142c3eb | 1142 | int active_nodes; |
8c8a743c PZ |
1143 | |
1144 | struct rcu_head rcu; | |
989348b5 | 1145 | unsigned long total_faults; |
4142c3eb | 1146 | unsigned long max_faults_cpu; |
7e2703e6 RR |
1147 | /* |
1148 | * Faults_cpu is used to decide whether memory should move | |
1149 | * towards the CPU. As a consequence, these stats are weighted | |
1150 | * more by CPU use than by memory faults. | |
1151 | */ | |
50ec8a40 | 1152 | unsigned long *faults_cpu; |
989348b5 | 1153 | unsigned long faults[0]; |
8c8a743c PZ |
1154 | }; |
1155 | ||
be1e4e76 RR |
1156 | /* Shared or private faults. */ |
1157 | #define NR_NUMA_HINT_FAULT_TYPES 2 | |
1158 | ||
1159 | /* Memory and CPU locality */ | |
1160 | #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) | |
1161 | ||
1162 | /* Averaged statistics, and temporary buffers. */ | |
1163 | #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) | |
1164 | ||
e29cf08b MG |
1165 | pid_t task_numa_group_id(struct task_struct *p) |
1166 | { | |
1167 | return p->numa_group ? p->numa_group->gid : 0; | |
1168 | } | |
1169 | ||
44dba3d5 IM |
1170 | /* |
1171 | * The averaged statistics, shared & private, memory & cpu, | |
1172 | * occupy the first half of the array. The second half of the | |
1173 | * array is for current counters, which are averaged into the | |
1174 | * first set by task_numa_placement. | |
1175 | */ | |
1176 | static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) | |
ac8e895b | 1177 | { |
44dba3d5 | 1178 | return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; |
ac8e895b MG |
1179 | } |
1180 | ||
1181 | static inline unsigned long task_faults(struct task_struct *p, int nid) | |
1182 | { | |
44dba3d5 | 1183 | if (!p->numa_faults) |
ac8e895b MG |
1184 | return 0; |
1185 | ||
44dba3d5 IM |
1186 | return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1187 | p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
ac8e895b MG |
1188 | } |
1189 | ||
83e1d2cd MG |
1190 | static inline unsigned long group_faults(struct task_struct *p, int nid) |
1191 | { | |
1192 | if (!p->numa_group) | |
1193 | return 0; | |
1194 | ||
44dba3d5 IM |
1195 | return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] + |
1196 | p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)]; | |
83e1d2cd MG |
1197 | } |
1198 | ||
20e07dea RR |
1199 | static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) |
1200 | { | |
44dba3d5 IM |
1201 | return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + |
1202 | group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; | |
20e07dea RR |
1203 | } |
1204 | ||
4142c3eb RR |
1205 | /* |
1206 | * A node triggering more than 1/3 as many NUMA faults as the maximum is | |
1207 | * considered part of a numa group's pseudo-interleaving set. Migrations | |
1208 | * between these nodes are slowed down, to allow things to settle down. | |
1209 | */ | |
1210 | #define ACTIVE_NODE_FRACTION 3 | |
1211 | ||
1212 | static bool numa_is_active_node(int nid, struct numa_group *ng) | |
1213 | { | |
1214 | return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; | |
1215 | } | |
1216 | ||
6c6b1193 RR |
1217 | /* Handle placement on systems where not all nodes are directly connected. */ |
1218 | static unsigned long score_nearby_nodes(struct task_struct *p, int nid, | |
1219 | int maxdist, bool task) | |
1220 | { | |
1221 | unsigned long score = 0; | |
1222 | int node; | |
1223 | ||
1224 | /* | |
1225 | * All nodes are directly connected, and the same distance | |
1226 | * from each other. No need for fancy placement algorithms. | |
1227 | */ | |
1228 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1229 | return 0; | |
1230 | ||
1231 | /* | |
1232 | * This code is called for each node, introducing N^2 complexity, | |
1233 | * which should be ok given the number of nodes rarely exceeds 8. | |
1234 | */ | |
1235 | for_each_online_node(node) { | |
1236 | unsigned long faults; | |
1237 | int dist = node_distance(nid, node); | |
1238 | ||
1239 | /* | |
1240 | * The furthest away nodes in the system are not interesting | |
1241 | * for placement; nid was already counted. | |
1242 | */ | |
1243 | if (dist == sched_max_numa_distance || node == nid) | |
1244 | continue; | |
1245 | ||
1246 | /* | |
1247 | * On systems with a backplane NUMA topology, compare groups | |
1248 | * of nodes, and move tasks towards the group with the most | |
1249 | * memory accesses. When comparing two nodes at distance | |
1250 | * "hoplimit", only nodes closer by than "hoplimit" are part | |
1251 | * of each group. Skip other nodes. | |
1252 | */ | |
1253 | if (sched_numa_topology_type == NUMA_BACKPLANE && | |
1254 | dist > maxdist) | |
1255 | continue; | |
1256 | ||
1257 | /* Add up the faults from nearby nodes. */ | |
1258 | if (task) | |
1259 | faults = task_faults(p, node); | |
1260 | else | |
1261 | faults = group_faults(p, node); | |
1262 | ||
1263 | /* | |
1264 | * On systems with a glueless mesh NUMA topology, there are | |
1265 | * no fixed "groups of nodes". Instead, nodes that are not | |
1266 | * directly connected bounce traffic through intermediate | |
1267 | * nodes; a numa_group can occupy any set of nodes. | |
1268 | * The further away a node is, the less the faults count. | |
1269 | * This seems to result in good task placement. | |
1270 | */ | |
1271 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
1272 | faults *= (sched_max_numa_distance - dist); | |
1273 | faults /= (sched_max_numa_distance - LOCAL_DISTANCE); | |
1274 | } | |
1275 | ||
1276 | score += faults; | |
1277 | } | |
1278 | ||
1279 | return score; | |
1280 | } | |
1281 | ||
83e1d2cd MG |
1282 | /* |
1283 | * These return the fraction of accesses done by a particular task, or | |
1284 | * task group, on a particular numa node. The group weight is given a | |
1285 | * larger multiplier, in order to group tasks together that are almost | |
1286 | * evenly spread out between numa nodes. | |
1287 | */ | |
7bd95320 RR |
1288 | static inline unsigned long task_weight(struct task_struct *p, int nid, |
1289 | int dist) | |
83e1d2cd | 1290 | { |
7bd95320 | 1291 | unsigned long faults, total_faults; |
83e1d2cd | 1292 | |
44dba3d5 | 1293 | if (!p->numa_faults) |
83e1d2cd MG |
1294 | return 0; |
1295 | ||
1296 | total_faults = p->total_numa_faults; | |
1297 | ||
1298 | if (!total_faults) | |
1299 | return 0; | |
1300 | ||
7bd95320 | 1301 | faults = task_faults(p, nid); |
6c6b1193 RR |
1302 | faults += score_nearby_nodes(p, nid, dist, true); |
1303 | ||
7bd95320 | 1304 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1305 | } |
1306 | ||
7bd95320 RR |
1307 | static inline unsigned long group_weight(struct task_struct *p, int nid, |
1308 | int dist) | |
83e1d2cd | 1309 | { |
7bd95320 RR |
1310 | unsigned long faults, total_faults; |
1311 | ||
1312 | if (!p->numa_group) | |
1313 | return 0; | |
1314 | ||
1315 | total_faults = p->numa_group->total_faults; | |
1316 | ||
1317 | if (!total_faults) | |
83e1d2cd MG |
1318 | return 0; |
1319 | ||
7bd95320 | 1320 | faults = group_faults(p, nid); |
6c6b1193 RR |
1321 | faults += score_nearby_nodes(p, nid, dist, false); |
1322 | ||
7bd95320 | 1323 | return 1000 * faults / total_faults; |
83e1d2cd MG |
1324 | } |
1325 | ||
10f39042 RR |
1326 | bool should_numa_migrate_memory(struct task_struct *p, struct page * page, |
1327 | int src_nid, int dst_cpu) | |
1328 | { | |
1329 | struct numa_group *ng = p->numa_group; | |
1330 | int dst_nid = cpu_to_node(dst_cpu); | |
1331 | int last_cpupid, this_cpupid; | |
1332 | ||
1333 | this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); | |
1334 | ||
1335 | /* | |
1336 | * Multi-stage node selection is used in conjunction with a periodic | |
1337 | * migration fault to build a temporal task<->page relation. By using | |
1338 | * a two-stage filter we remove short/unlikely relations. | |
1339 | * | |
1340 | * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate | |
1341 | * a task's usage of a particular page (n_p) per total usage of this | |
1342 | * page (n_t) (in a given time-span) to a probability. | |
1343 | * | |
1344 | * Our periodic faults will sample this probability and getting the | |
1345 | * same result twice in a row, given these samples are fully | |
1346 | * independent, is then given by P(n)^2, provided our sample period | |
1347 | * is sufficiently short compared to the usage pattern. | |
1348 | * | |
1349 | * This quadric squishes small probabilities, making it less likely we | |
1350 | * act on an unlikely task<->page relation. | |
1351 | */ | |
1352 | last_cpupid = page_cpupid_xchg_last(page, this_cpupid); | |
1353 | if (!cpupid_pid_unset(last_cpupid) && | |
1354 | cpupid_to_nid(last_cpupid) != dst_nid) | |
1355 | return false; | |
1356 | ||
1357 | /* Always allow migrate on private faults */ | |
1358 | if (cpupid_match_pid(p, last_cpupid)) | |
1359 | return true; | |
1360 | ||
1361 | /* A shared fault, but p->numa_group has not been set up yet. */ | |
1362 | if (!ng) | |
1363 | return true; | |
1364 | ||
1365 | /* | |
4142c3eb RR |
1366 | * Destination node is much more heavily used than the source |
1367 | * node? Allow migration. | |
10f39042 | 1368 | */ |
4142c3eb RR |
1369 | if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * |
1370 | ACTIVE_NODE_FRACTION) | |
10f39042 RR |
1371 | return true; |
1372 | ||
1373 | /* | |
4142c3eb RR |
1374 | * Distribute memory according to CPU & memory use on each node, |
1375 | * with 3/4 hysteresis to avoid unnecessary memory migrations: | |
1376 | * | |
1377 | * faults_cpu(dst) 3 faults_cpu(src) | |
1378 | * --------------- * - > --------------- | |
1379 | * faults_mem(dst) 4 faults_mem(src) | |
10f39042 | 1380 | */ |
4142c3eb RR |
1381 | return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > |
1382 | group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; | |
10f39042 RR |
1383 | } |
1384 | ||
e6628d5b | 1385 | static unsigned long weighted_cpuload(const int cpu); |
58d081b5 MG |
1386 | static unsigned long source_load(int cpu, int type); |
1387 | static unsigned long target_load(int cpu, int type); | |
ced549fa | 1388 | static unsigned long capacity_of(int cpu); |
58d081b5 MG |
1389 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg); |
1390 | ||
fb13c7ee | 1391 | /* Cached statistics for all CPUs within a node */ |
58d081b5 | 1392 | struct numa_stats { |
fb13c7ee | 1393 | unsigned long nr_running; |
58d081b5 | 1394 | unsigned long load; |
fb13c7ee MG |
1395 | |
1396 | /* Total compute capacity of CPUs on a node */ | |
5ef20ca1 | 1397 | unsigned long compute_capacity; |
fb13c7ee MG |
1398 | |
1399 | /* Approximate capacity in terms of runnable tasks on a node */ | |
5ef20ca1 | 1400 | unsigned long task_capacity; |
1b6a7495 | 1401 | int has_free_capacity; |
58d081b5 | 1402 | }; |
e6628d5b | 1403 | |
fb13c7ee MG |
1404 | /* |
1405 | * XXX borrowed from update_sg_lb_stats | |
1406 | */ | |
1407 | static void update_numa_stats(struct numa_stats *ns, int nid) | |
1408 | { | |
83d7f242 RR |
1409 | int smt, cpu, cpus = 0; |
1410 | unsigned long capacity; | |
fb13c7ee MG |
1411 | |
1412 | memset(ns, 0, sizeof(*ns)); | |
1413 | for_each_cpu(cpu, cpumask_of_node(nid)) { | |
1414 | struct rq *rq = cpu_rq(cpu); | |
1415 | ||
1416 | ns->nr_running += rq->nr_running; | |
1417 | ns->load += weighted_cpuload(cpu); | |
ced549fa | 1418 | ns->compute_capacity += capacity_of(cpu); |
5eca82a9 PZ |
1419 | |
1420 | cpus++; | |
fb13c7ee MG |
1421 | } |
1422 | ||
5eca82a9 PZ |
1423 | /* |
1424 | * If we raced with hotplug and there are no CPUs left in our mask | |
1425 | * the @ns structure is NULL'ed and task_numa_compare() will | |
1426 | * not find this node attractive. | |
1427 | * | |
1b6a7495 NP |
1428 | * We'll either bail at !has_free_capacity, or we'll detect a huge |
1429 | * imbalance and bail there. | |
5eca82a9 PZ |
1430 | */ |
1431 | if (!cpus) | |
1432 | return; | |
1433 | ||
83d7f242 RR |
1434 | /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */ |
1435 | smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity); | |
1436 | capacity = cpus / smt; /* cores */ | |
1437 | ||
1438 | ns->task_capacity = min_t(unsigned, capacity, | |
1439 | DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE)); | |
1b6a7495 | 1440 | ns->has_free_capacity = (ns->nr_running < ns->task_capacity); |
fb13c7ee MG |
1441 | } |
1442 | ||
58d081b5 MG |
1443 | struct task_numa_env { |
1444 | struct task_struct *p; | |
e6628d5b | 1445 | |
58d081b5 MG |
1446 | int src_cpu, src_nid; |
1447 | int dst_cpu, dst_nid; | |
e6628d5b | 1448 | |
58d081b5 | 1449 | struct numa_stats src_stats, dst_stats; |
e6628d5b | 1450 | |
40ea2b42 | 1451 | int imbalance_pct; |
7bd95320 | 1452 | int dist; |
fb13c7ee MG |
1453 | |
1454 | struct task_struct *best_task; | |
1455 | long best_imp; | |
58d081b5 MG |
1456 | int best_cpu; |
1457 | }; | |
1458 | ||
fb13c7ee MG |
1459 | static void task_numa_assign(struct task_numa_env *env, |
1460 | struct task_struct *p, long imp) | |
1461 | { | |
1462 | if (env->best_task) | |
1463 | put_task_struct(env->best_task); | |
bac78573 ON |
1464 | if (p) |
1465 | get_task_struct(p); | |
fb13c7ee MG |
1466 | |
1467 | env->best_task = p; | |
1468 | env->best_imp = imp; | |
1469 | env->best_cpu = env->dst_cpu; | |
1470 | } | |
1471 | ||
28a21745 | 1472 | static bool load_too_imbalanced(long src_load, long dst_load, |
e63da036 RR |
1473 | struct task_numa_env *env) |
1474 | { | |
e4991b24 RR |
1475 | long imb, old_imb; |
1476 | long orig_src_load, orig_dst_load; | |
28a21745 RR |
1477 | long src_capacity, dst_capacity; |
1478 | ||
1479 | /* | |
1480 | * The load is corrected for the CPU capacity available on each node. | |
1481 | * | |
1482 | * src_load dst_load | |
1483 | * ------------ vs --------- | |
1484 | * src_capacity dst_capacity | |
1485 | */ | |
1486 | src_capacity = env->src_stats.compute_capacity; | |
1487 | dst_capacity = env->dst_stats.compute_capacity; | |
e63da036 RR |
1488 | |
1489 | /* We care about the slope of the imbalance, not the direction. */ | |
e4991b24 RR |
1490 | if (dst_load < src_load) |
1491 | swap(dst_load, src_load); | |
e63da036 RR |
1492 | |
1493 | /* Is the difference below the threshold? */ | |
e4991b24 RR |
1494 | imb = dst_load * src_capacity * 100 - |
1495 | src_load * dst_capacity * env->imbalance_pct; | |
e63da036 RR |
1496 | if (imb <= 0) |
1497 | return false; | |
1498 | ||
1499 | /* | |
1500 | * The imbalance is above the allowed threshold. | |
e4991b24 | 1501 | * Compare it with the old imbalance. |
e63da036 | 1502 | */ |
28a21745 | 1503 | orig_src_load = env->src_stats.load; |
e4991b24 | 1504 | orig_dst_load = env->dst_stats.load; |
28a21745 | 1505 | |
e4991b24 RR |
1506 | if (orig_dst_load < orig_src_load) |
1507 | swap(orig_dst_load, orig_src_load); | |
e63da036 | 1508 | |
e4991b24 RR |
1509 | old_imb = orig_dst_load * src_capacity * 100 - |
1510 | orig_src_load * dst_capacity * env->imbalance_pct; | |
1511 | ||
1512 | /* Would this change make things worse? */ | |
1513 | return (imb > old_imb); | |
e63da036 RR |
1514 | } |
1515 | ||
fb13c7ee MG |
1516 | /* |
1517 | * This checks if the overall compute and NUMA accesses of the system would | |
1518 | * be improved if the source tasks was migrated to the target dst_cpu taking | |
1519 | * into account that it might be best if task running on the dst_cpu should | |
1520 | * be exchanged with the source task | |
1521 | */ | |
887c290e RR |
1522 | static void task_numa_compare(struct task_numa_env *env, |
1523 | long taskimp, long groupimp) | |
fb13c7ee MG |
1524 | { |
1525 | struct rq *src_rq = cpu_rq(env->src_cpu); | |
1526 | struct rq *dst_rq = cpu_rq(env->dst_cpu); | |
1527 | struct task_struct *cur; | |
28a21745 | 1528 | long src_load, dst_load; |
fb13c7ee | 1529 | long load; |
1c5d3eb3 | 1530 | long imp = env->p->numa_group ? groupimp : taskimp; |
0132c3e1 | 1531 | long moveimp = imp; |
7bd95320 | 1532 | int dist = env->dist; |
fb13c7ee MG |
1533 | |
1534 | rcu_read_lock(); | |
bac78573 ON |
1535 | cur = task_rcu_dereference(&dst_rq->curr); |
1536 | if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) | |
fb13c7ee MG |
1537 | cur = NULL; |
1538 | ||
7af68335 PZ |
1539 | /* |
1540 | * Because we have preemption enabled we can get migrated around and | |
1541 | * end try selecting ourselves (current == env->p) as a swap candidate. | |
1542 | */ | |
1543 | if (cur == env->p) | |
1544 | goto unlock; | |
1545 | ||
fb13c7ee MG |
1546 | /* |
1547 | * "imp" is the fault differential for the source task between the | |
1548 | * source and destination node. Calculate the total differential for | |
1549 | * the source task and potential destination task. The more negative | |
1550 | * the value is, the more rmeote accesses that would be expected to | |
1551 | * be incurred if the tasks were swapped. | |
1552 | */ | |
1553 | if (cur) { | |
1554 | /* Skip this swap candidate if cannot move to the source cpu */ | |
0c98d344 | 1555 | if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed)) |
fb13c7ee MG |
1556 | goto unlock; |
1557 | ||
887c290e RR |
1558 | /* |
1559 | * If dst and source tasks are in the same NUMA group, or not | |
ca28aa53 | 1560 | * in any group then look only at task weights. |
887c290e | 1561 | */ |
ca28aa53 | 1562 | if (cur->numa_group == env->p->numa_group) { |
7bd95320 RR |
1563 | imp = taskimp + task_weight(cur, env->src_nid, dist) - |
1564 | task_weight(cur, env->dst_nid, dist); | |
ca28aa53 RR |
1565 | /* |
1566 | * Add some hysteresis to prevent swapping the | |
1567 | * tasks within a group over tiny differences. | |
1568 | */ | |
1569 | if (cur->numa_group) | |
1570 | imp -= imp/16; | |
887c290e | 1571 | } else { |
ca28aa53 RR |
1572 | /* |
1573 | * Compare the group weights. If a task is all by | |
1574 | * itself (not part of a group), use the task weight | |
1575 | * instead. | |
1576 | */ | |
ca28aa53 | 1577 | if (cur->numa_group) |
7bd95320 RR |
1578 | imp += group_weight(cur, env->src_nid, dist) - |
1579 | group_weight(cur, env->dst_nid, dist); | |
ca28aa53 | 1580 | else |
7bd95320 RR |
1581 | imp += task_weight(cur, env->src_nid, dist) - |
1582 | task_weight(cur, env->dst_nid, dist); | |
887c290e | 1583 | } |
fb13c7ee MG |
1584 | } |
1585 | ||
0132c3e1 | 1586 | if (imp <= env->best_imp && moveimp <= env->best_imp) |
fb13c7ee MG |
1587 | goto unlock; |
1588 | ||
1589 | if (!cur) { | |
1590 | /* Is there capacity at our destination? */ | |
b932c03c | 1591 | if (env->src_stats.nr_running <= env->src_stats.task_capacity && |
1b6a7495 | 1592 | !env->dst_stats.has_free_capacity) |
fb13c7ee MG |
1593 | goto unlock; |
1594 | ||
1595 | goto balance; | |
1596 | } | |
1597 | ||
1598 | /* Balance doesn't matter much if we're running a task per cpu */ | |
0132c3e1 RR |
1599 | if (imp > env->best_imp && src_rq->nr_running == 1 && |
1600 | dst_rq->nr_running == 1) | |
fb13c7ee MG |
1601 | goto assign; |
1602 | ||
1603 | /* | |
1604 | * In the overloaded case, try and keep the load balanced. | |
1605 | */ | |
1606 | balance: | |
e720fff6 PZ |
1607 | load = task_h_load(env->p); |
1608 | dst_load = env->dst_stats.load + load; | |
1609 | src_load = env->src_stats.load - load; | |
fb13c7ee | 1610 | |
0132c3e1 RR |
1611 | if (moveimp > imp && moveimp > env->best_imp) { |
1612 | /* | |
1613 | * If the improvement from just moving env->p direction is | |
1614 | * better than swapping tasks around, check if a move is | |
1615 | * possible. Store a slightly smaller score than moveimp, | |
1616 | * so an actually idle CPU will win. | |
1617 | */ | |
1618 | if (!load_too_imbalanced(src_load, dst_load, env)) { | |
1619 | imp = moveimp - 1; | |
1620 | cur = NULL; | |
1621 | goto assign; | |
1622 | } | |
1623 | } | |
1624 | ||
1625 | if (imp <= env->best_imp) | |
1626 | goto unlock; | |
1627 | ||
fb13c7ee | 1628 | if (cur) { |
e720fff6 PZ |
1629 | load = task_h_load(cur); |
1630 | dst_load -= load; | |
1631 | src_load += load; | |
fb13c7ee MG |
1632 | } |
1633 | ||
28a21745 | 1634 | if (load_too_imbalanced(src_load, dst_load, env)) |
fb13c7ee MG |
1635 | goto unlock; |
1636 | ||
ba7e5a27 RR |
1637 | /* |
1638 | * One idle CPU per node is evaluated for a task numa move. | |
1639 | * Call select_idle_sibling to maybe find a better one. | |
1640 | */ | |
10e2f1ac PZ |
1641 | if (!cur) { |
1642 | /* | |
1643 | * select_idle_siblings() uses an per-cpu cpumask that | |
1644 | * can be used from IRQ context. | |
1645 | */ | |
1646 | local_irq_disable(); | |
772bd008 MR |
1647 | env->dst_cpu = select_idle_sibling(env->p, env->src_cpu, |
1648 | env->dst_cpu); | |
10e2f1ac PZ |
1649 | local_irq_enable(); |
1650 | } | |
ba7e5a27 | 1651 | |
fb13c7ee MG |
1652 | assign: |
1653 | task_numa_assign(env, cur, imp); | |
1654 | unlock: | |
1655 | rcu_read_unlock(); | |
1656 | } | |
1657 | ||
887c290e RR |
1658 | static void task_numa_find_cpu(struct task_numa_env *env, |
1659 | long taskimp, long groupimp) | |
2c8a50aa MG |
1660 | { |
1661 | int cpu; | |
1662 | ||
1663 | for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { | |
1664 | /* Skip this CPU if the source task cannot migrate */ | |
0c98d344 | 1665 | if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed)) |
2c8a50aa MG |
1666 | continue; |
1667 | ||
1668 | env->dst_cpu = cpu; | |
887c290e | 1669 | task_numa_compare(env, taskimp, groupimp); |
2c8a50aa MG |
1670 | } |
1671 | } | |
1672 | ||
6f9aad0b RR |
1673 | /* Only move tasks to a NUMA node less busy than the current node. */ |
1674 | static bool numa_has_capacity(struct task_numa_env *env) | |
1675 | { | |
1676 | struct numa_stats *src = &env->src_stats; | |
1677 | struct numa_stats *dst = &env->dst_stats; | |
1678 | ||
1679 | if (src->has_free_capacity && !dst->has_free_capacity) | |
1680 | return false; | |
1681 | ||
1682 | /* | |
1683 | * Only consider a task move if the source has a higher load | |
1684 | * than the destination, corrected for CPU capacity on each node. | |
1685 | * | |
1686 | * src->load dst->load | |
1687 | * --------------------- vs --------------------- | |
1688 | * src->compute_capacity dst->compute_capacity | |
1689 | */ | |
44dcb04f SD |
1690 | if (src->load * dst->compute_capacity * env->imbalance_pct > |
1691 | ||
1692 | dst->load * src->compute_capacity * 100) | |
6f9aad0b RR |
1693 | return true; |
1694 | ||
1695 | return false; | |
1696 | } | |
1697 | ||
58d081b5 MG |
1698 | static int task_numa_migrate(struct task_struct *p) |
1699 | { | |
58d081b5 MG |
1700 | struct task_numa_env env = { |
1701 | .p = p, | |
fb13c7ee | 1702 | |
58d081b5 | 1703 | .src_cpu = task_cpu(p), |
b32e86b4 | 1704 | .src_nid = task_node(p), |
fb13c7ee MG |
1705 | |
1706 | .imbalance_pct = 112, | |
1707 | ||
1708 | .best_task = NULL, | |
1709 | .best_imp = 0, | |
4142c3eb | 1710 | .best_cpu = -1, |
58d081b5 MG |
1711 | }; |
1712 | struct sched_domain *sd; | |
887c290e | 1713 | unsigned long taskweight, groupweight; |
7bd95320 | 1714 | int nid, ret, dist; |
887c290e | 1715 | long taskimp, groupimp; |
e6628d5b | 1716 | |
58d081b5 | 1717 | /* |
fb13c7ee MG |
1718 | * Pick the lowest SD_NUMA domain, as that would have the smallest |
1719 | * imbalance and would be the first to start moving tasks about. | |
1720 | * | |
1721 | * And we want to avoid any moving of tasks about, as that would create | |
1722 | * random movement of tasks -- counter the numa conditions we're trying | |
1723 | * to satisfy here. | |
58d081b5 MG |
1724 | */ |
1725 | rcu_read_lock(); | |
fb13c7ee | 1726 | sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); |
46a73e8a RR |
1727 | if (sd) |
1728 | env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; | |
e6628d5b MG |
1729 | rcu_read_unlock(); |
1730 | ||
46a73e8a RR |
1731 | /* |
1732 | * Cpusets can break the scheduler domain tree into smaller | |
1733 | * balance domains, some of which do not cross NUMA boundaries. | |
1734 | * Tasks that are "trapped" in such domains cannot be migrated | |
1735 | * elsewhere, so there is no point in (re)trying. | |
1736 | */ | |
1737 | if (unlikely(!sd)) { | |
de1b301a | 1738 | p->numa_preferred_nid = task_node(p); |
46a73e8a RR |
1739 | return -EINVAL; |
1740 | } | |
1741 | ||
2c8a50aa | 1742 | env.dst_nid = p->numa_preferred_nid; |
7bd95320 RR |
1743 | dist = env.dist = node_distance(env.src_nid, env.dst_nid); |
1744 | taskweight = task_weight(p, env.src_nid, dist); | |
1745 | groupweight = group_weight(p, env.src_nid, dist); | |
1746 | update_numa_stats(&env.src_stats, env.src_nid); | |
1747 | taskimp = task_weight(p, env.dst_nid, dist) - taskweight; | |
1748 | groupimp = group_weight(p, env.dst_nid, dist) - groupweight; | |
2c8a50aa | 1749 | update_numa_stats(&env.dst_stats, env.dst_nid); |
58d081b5 | 1750 | |
a43455a1 | 1751 | /* Try to find a spot on the preferred nid. */ |
6f9aad0b RR |
1752 | if (numa_has_capacity(&env)) |
1753 | task_numa_find_cpu(&env, taskimp, groupimp); | |
e1dda8a7 | 1754 | |
9de05d48 RR |
1755 | /* |
1756 | * Look at other nodes in these cases: | |
1757 | * - there is no space available on the preferred_nid | |
1758 | * - the task is part of a numa_group that is interleaved across | |
1759 | * multiple NUMA nodes; in order to better consolidate the group, | |
1760 | * we need to check other locations. | |
1761 | */ | |
4142c3eb | 1762 | if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) { |
2c8a50aa MG |
1763 | for_each_online_node(nid) { |
1764 | if (nid == env.src_nid || nid == p->numa_preferred_nid) | |
1765 | continue; | |
58d081b5 | 1766 | |
7bd95320 | 1767 | dist = node_distance(env.src_nid, env.dst_nid); |
6c6b1193 RR |
1768 | if (sched_numa_topology_type == NUMA_BACKPLANE && |
1769 | dist != env.dist) { | |
1770 | taskweight = task_weight(p, env.src_nid, dist); | |
1771 | groupweight = group_weight(p, env.src_nid, dist); | |
1772 | } | |
7bd95320 | 1773 | |
83e1d2cd | 1774 | /* Only consider nodes where both task and groups benefit */ |
7bd95320 RR |
1775 | taskimp = task_weight(p, nid, dist) - taskweight; |
1776 | groupimp = group_weight(p, nid, dist) - groupweight; | |
887c290e | 1777 | if (taskimp < 0 && groupimp < 0) |
fb13c7ee MG |
1778 | continue; |
1779 | ||
7bd95320 | 1780 | env.dist = dist; |
2c8a50aa MG |
1781 | env.dst_nid = nid; |
1782 | update_numa_stats(&env.dst_stats, env.dst_nid); | |
6f9aad0b RR |
1783 | if (numa_has_capacity(&env)) |
1784 | task_numa_find_cpu(&env, taskimp, groupimp); | |
58d081b5 MG |
1785 | } |
1786 | } | |
1787 | ||
68d1b02a RR |
1788 | /* |
1789 | * If the task is part of a workload that spans multiple NUMA nodes, | |
1790 | * and is migrating into one of the workload's active nodes, remember | |
1791 | * this node as the task's preferred numa node, so the workload can | |
1792 | * settle down. | |
1793 | * A task that migrated to a second choice node will be better off | |
1794 | * trying for a better one later. Do not set the preferred node here. | |
1795 | */ | |
db015dae | 1796 | if (p->numa_group) { |
4142c3eb RR |
1797 | struct numa_group *ng = p->numa_group; |
1798 | ||
db015dae RR |
1799 | if (env.best_cpu == -1) |
1800 | nid = env.src_nid; | |
1801 | else | |
1802 | nid = env.dst_nid; | |
1803 | ||
4142c3eb | 1804 | if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng)) |
db015dae RR |
1805 | sched_setnuma(p, env.dst_nid); |
1806 | } | |
1807 | ||
1808 | /* No better CPU than the current one was found. */ | |
1809 | if (env.best_cpu == -1) | |
1810 | return -EAGAIN; | |
0ec8aa00 | 1811 | |
04bb2f94 RR |
1812 | /* |
1813 | * Reset the scan period if the task is being rescheduled on an | |
1814 | * alternative node to recheck if the tasks is now properly placed. | |
1815 | */ | |
1816 | p->numa_scan_period = task_scan_min(p); | |
1817 | ||
fb13c7ee | 1818 | if (env.best_task == NULL) { |
286549dc MG |
1819 | ret = migrate_task_to(p, env.best_cpu); |
1820 | if (ret != 0) | |
1821 | trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); | |
fb13c7ee MG |
1822 | return ret; |
1823 | } | |
1824 | ||
1825 | ret = migrate_swap(p, env.best_task); | |
286549dc MG |
1826 | if (ret != 0) |
1827 | trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); | |
fb13c7ee MG |
1828 | put_task_struct(env.best_task); |
1829 | return ret; | |
e6628d5b MG |
1830 | } |
1831 | ||
6b9a7460 MG |
1832 | /* Attempt to migrate a task to a CPU on the preferred node. */ |
1833 | static void numa_migrate_preferred(struct task_struct *p) | |
1834 | { | |
5085e2a3 RR |
1835 | unsigned long interval = HZ; |
1836 | ||
2739d3ee | 1837 | /* This task has no NUMA fault statistics yet */ |
44dba3d5 | 1838 | if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults)) |
6b9a7460 MG |
1839 | return; |
1840 | ||
2739d3ee | 1841 | /* Periodically retry migrating the task to the preferred node */ |
5085e2a3 RR |
1842 | interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); |
1843 | p->numa_migrate_retry = jiffies + interval; | |
2739d3ee RR |
1844 | |
1845 | /* Success if task is already running on preferred CPU */ | |
de1b301a | 1846 | if (task_node(p) == p->numa_preferred_nid) |
6b9a7460 MG |
1847 | return; |
1848 | ||
1849 | /* Otherwise, try migrate to a CPU on the preferred node */ | |
2739d3ee | 1850 | task_numa_migrate(p); |
6b9a7460 MG |
1851 | } |
1852 | ||
20e07dea | 1853 | /* |
4142c3eb | 1854 | * Find out how many nodes on the workload is actively running on. Do this by |
20e07dea RR |
1855 | * tracking the nodes from which NUMA hinting faults are triggered. This can |
1856 | * be different from the set of nodes where the workload's memory is currently | |
1857 | * located. | |
20e07dea | 1858 | */ |
4142c3eb | 1859 | static void numa_group_count_active_nodes(struct numa_group *numa_group) |
20e07dea RR |
1860 | { |
1861 | unsigned long faults, max_faults = 0; | |
4142c3eb | 1862 | int nid, active_nodes = 0; |
20e07dea RR |
1863 | |
1864 | for_each_online_node(nid) { | |
1865 | faults = group_faults_cpu(numa_group, nid); | |
1866 | if (faults > max_faults) | |
1867 | max_faults = faults; | |
1868 | } | |
1869 | ||
1870 | for_each_online_node(nid) { | |
1871 | faults = group_faults_cpu(numa_group, nid); | |
4142c3eb RR |
1872 | if (faults * ACTIVE_NODE_FRACTION > max_faults) |
1873 | active_nodes++; | |
20e07dea | 1874 | } |
4142c3eb RR |
1875 | |
1876 | numa_group->max_faults_cpu = max_faults; | |
1877 | numa_group->active_nodes = active_nodes; | |
20e07dea RR |
1878 | } |
1879 | ||
04bb2f94 RR |
1880 | /* |
1881 | * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS | |
1882 | * increments. The more local the fault statistics are, the higher the scan | |
a22b4b01 RR |
1883 | * period will be for the next scan window. If local/(local+remote) ratio is |
1884 | * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) | |
1885 | * the scan period will decrease. Aim for 70% local accesses. | |
04bb2f94 RR |
1886 | */ |
1887 | #define NUMA_PERIOD_SLOTS 10 | |
a22b4b01 | 1888 | #define NUMA_PERIOD_THRESHOLD 7 |
04bb2f94 RR |
1889 | |
1890 | /* | |
1891 | * Increase the scan period (slow down scanning) if the majority of | |
1892 | * our memory is already on our local node, or if the majority of | |
1893 | * the page accesses are shared with other processes. | |
1894 | * Otherwise, decrease the scan period. | |
1895 | */ | |
1896 | static void update_task_scan_period(struct task_struct *p, | |
1897 | unsigned long shared, unsigned long private) | |
1898 | { | |
1899 | unsigned int period_slot; | |
1900 | int ratio; | |
1901 | int diff; | |
1902 | ||
1903 | unsigned long remote = p->numa_faults_locality[0]; | |
1904 | unsigned long local = p->numa_faults_locality[1]; | |
1905 | ||
1906 | /* | |
1907 | * If there were no record hinting faults then either the task is | |
1908 | * completely idle or all activity is areas that are not of interest | |
074c2381 MG |
1909 | * to automatic numa balancing. Related to that, if there were failed |
1910 | * migration then it implies we are migrating too quickly or the local | |
1911 | * node is overloaded. In either case, scan slower | |
04bb2f94 | 1912 | */ |
074c2381 | 1913 | if (local + shared == 0 || p->numa_faults_locality[2]) { |
04bb2f94 RR |
1914 | p->numa_scan_period = min(p->numa_scan_period_max, |
1915 | p->numa_scan_period << 1); | |
1916 | ||
1917 | p->mm->numa_next_scan = jiffies + | |
1918 | msecs_to_jiffies(p->numa_scan_period); | |
1919 | ||
1920 | return; | |
1921 | } | |
1922 | ||
1923 | /* | |
1924 | * Prepare to scale scan period relative to the current period. | |
1925 | * == NUMA_PERIOD_THRESHOLD scan period stays the same | |
1926 | * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) | |
1927 | * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) | |
1928 | */ | |
1929 | period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); | |
1930 | ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); | |
1931 | if (ratio >= NUMA_PERIOD_THRESHOLD) { | |
1932 | int slot = ratio - NUMA_PERIOD_THRESHOLD; | |
1933 | if (!slot) | |
1934 | slot = 1; | |
1935 | diff = slot * period_slot; | |
1936 | } else { | |
1937 | diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; | |
1938 | ||
1939 | /* | |
1940 | * Scale scan rate increases based on sharing. There is an | |
1941 | * inverse relationship between the degree of sharing and | |
1942 | * the adjustment made to the scanning period. Broadly | |
1943 | * speaking the intent is that there is little point | |
1944 | * scanning faster if shared accesses dominate as it may | |
1945 | * simply bounce migrations uselessly | |
1946 | */ | |
2847c90e | 1947 | ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1)); |
04bb2f94 RR |
1948 | diff = (diff * ratio) / NUMA_PERIOD_SLOTS; |
1949 | } | |
1950 | ||
1951 | p->numa_scan_period = clamp(p->numa_scan_period + diff, | |
1952 | task_scan_min(p), task_scan_max(p)); | |
1953 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); | |
1954 | } | |
1955 | ||
7e2703e6 RR |
1956 | /* |
1957 | * Get the fraction of time the task has been running since the last | |
1958 | * NUMA placement cycle. The scheduler keeps similar statistics, but | |
1959 | * decays those on a 32ms period, which is orders of magnitude off | |
1960 | * from the dozens-of-seconds NUMA balancing period. Use the scheduler | |
1961 | * stats only if the task is so new there are no NUMA statistics yet. | |
1962 | */ | |
1963 | static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) | |
1964 | { | |
1965 | u64 runtime, delta, now; | |
1966 | /* Use the start of this time slice to avoid calculations. */ | |
1967 | now = p->se.exec_start; | |
1968 | runtime = p->se.sum_exec_runtime; | |
1969 | ||
1970 | if (p->last_task_numa_placement) { | |
1971 | delta = runtime - p->last_sum_exec_runtime; | |
1972 | *period = now - p->last_task_numa_placement; | |
1973 | } else { | |
9d89c257 YD |
1974 | delta = p->se.avg.load_sum / p->se.load.weight; |
1975 | *period = LOAD_AVG_MAX; | |
7e2703e6 RR |
1976 | } |
1977 | ||
1978 | p->last_sum_exec_runtime = runtime; | |
1979 | p->last_task_numa_placement = now; | |
1980 | ||
1981 | return delta; | |
1982 | } | |
1983 | ||
54009416 RR |
1984 | /* |
1985 | * Determine the preferred nid for a task in a numa_group. This needs to | |
1986 | * be done in a way that produces consistent results with group_weight, | |
1987 | * otherwise workloads might not converge. | |
1988 | */ | |
1989 | static int preferred_group_nid(struct task_struct *p, int nid) | |
1990 | { | |
1991 | nodemask_t nodes; | |
1992 | int dist; | |
1993 | ||
1994 | /* Direct connections between all NUMA nodes. */ | |
1995 | if (sched_numa_topology_type == NUMA_DIRECT) | |
1996 | return nid; | |
1997 | ||
1998 | /* | |
1999 | * On a system with glueless mesh NUMA topology, group_weight | |
2000 | * scores nodes according to the number of NUMA hinting faults on | |
2001 | * both the node itself, and on nearby nodes. | |
2002 | */ | |
2003 | if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { | |
2004 | unsigned long score, max_score = 0; | |
2005 | int node, max_node = nid; | |
2006 | ||
2007 | dist = sched_max_numa_distance; | |
2008 | ||
2009 | for_each_online_node(node) { | |
2010 | score = group_weight(p, node, dist); | |
2011 | if (score > max_score) { | |
2012 | max_score = score; | |
2013 | max_node = node; | |
2014 | } | |
2015 | } | |
2016 | return max_node; | |
2017 | } | |
2018 | ||
2019 | /* | |
2020 | * Finding the preferred nid in a system with NUMA backplane | |
2021 | * interconnect topology is more involved. The goal is to locate | |
2022 | * tasks from numa_groups near each other in the system, and | |
2023 | * untangle workloads from different sides of the system. This requires | |
2024 | * searching down the hierarchy of node groups, recursively searching | |
2025 | * inside the highest scoring group of nodes. The nodemask tricks | |
2026 | * keep the complexity of the search down. | |
2027 | */ | |
2028 | nodes = node_online_map; | |
2029 | for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { | |
2030 | unsigned long max_faults = 0; | |
81907478 | 2031 | nodemask_t max_group = NODE_MASK_NONE; |
54009416 RR |
2032 | int a, b; |
2033 | ||
2034 | /* Are there nodes at this distance from each other? */ | |
2035 | if (!find_numa_distance(dist)) | |
2036 | continue; | |
2037 | ||
2038 | for_each_node_mask(a, nodes) { | |
2039 | unsigned long faults = 0; | |
2040 | nodemask_t this_group; | |
2041 | nodes_clear(this_group); | |
2042 | ||
2043 | /* Sum group's NUMA faults; includes a==b case. */ | |
2044 | for_each_node_mask(b, nodes) { | |
2045 | if (node_distance(a, b) < dist) { | |
2046 | faults += group_faults(p, b); | |
2047 | node_set(b, this_group); | |
2048 | node_clear(b, nodes); | |
2049 | } | |
2050 | } | |
2051 | ||
2052 | /* Remember the top group. */ | |
2053 | if (faults > max_faults) { | |
2054 | max_faults = faults; | |
2055 | max_group = this_group; | |
2056 | /* | |
2057 | * subtle: at the smallest distance there is | |
2058 | * just one node left in each "group", the | |
2059 | * winner is the preferred nid. | |
2060 | */ | |
2061 | nid = a; | |
2062 | } | |
2063 | } | |
2064 | /* Next round, evaluate the nodes within max_group. */ | |
890a5409 JB |
2065 | if (!max_faults) |
2066 | break; | |
54009416 RR |
2067 | nodes = max_group; |
2068 | } | |
2069 | return nid; | |
2070 | } | |
2071 | ||
cbee9f88 PZ |
2072 | static void task_numa_placement(struct task_struct *p) |
2073 | { | |
83e1d2cd MG |
2074 | int seq, nid, max_nid = -1, max_group_nid = -1; |
2075 | unsigned long max_faults = 0, max_group_faults = 0; | |
04bb2f94 | 2076 | unsigned long fault_types[2] = { 0, 0 }; |
7e2703e6 RR |
2077 | unsigned long total_faults; |
2078 | u64 runtime, period; | |
7dbd13ed | 2079 | spinlock_t *group_lock = NULL; |
cbee9f88 | 2080 | |
7e5a2c17 JL |
2081 | /* |
2082 | * The p->mm->numa_scan_seq field gets updated without | |
2083 | * exclusive access. Use READ_ONCE() here to ensure | |
2084 | * that the field is read in a single access: | |
2085 | */ | |
316c1608 | 2086 | seq = READ_ONCE(p->mm->numa_scan_seq); |
cbee9f88 PZ |
2087 | if (p->numa_scan_seq == seq) |
2088 | return; | |
2089 | p->numa_scan_seq = seq; | |
598f0ec0 | 2090 | p->numa_scan_period_max = task_scan_max(p); |
cbee9f88 | 2091 | |
7e2703e6 RR |
2092 | total_faults = p->numa_faults_locality[0] + |
2093 | p->numa_faults_locality[1]; | |
2094 | runtime = numa_get_avg_runtime(p, &period); | |
2095 | ||
7dbd13ed MG |
2096 | /* If the task is part of a group prevent parallel updates to group stats */ |
2097 | if (p->numa_group) { | |
2098 | group_lock = &p->numa_group->lock; | |
60e69eed | 2099 | spin_lock_irq(group_lock); |
7dbd13ed MG |
2100 | } |
2101 | ||
688b7585 MG |
2102 | /* Find the node with the highest number of faults */ |
2103 | for_each_online_node(nid) { | |
44dba3d5 IM |
2104 | /* Keep track of the offsets in numa_faults array */ |
2105 | int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; | |
83e1d2cd | 2106 | unsigned long faults = 0, group_faults = 0; |
44dba3d5 | 2107 | int priv; |
745d6147 | 2108 | |
be1e4e76 | 2109 | for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { |
7e2703e6 | 2110 | long diff, f_diff, f_weight; |
8c8a743c | 2111 | |
44dba3d5 IM |
2112 | mem_idx = task_faults_idx(NUMA_MEM, nid, priv); |
2113 | membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); | |
2114 | cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); | |
2115 | cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); | |
745d6147 | 2116 | |
ac8e895b | 2117 | /* Decay existing window, copy faults since last scan */ |
44dba3d5 IM |
2118 | diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; |
2119 | fault_types[priv] += p->numa_faults[membuf_idx]; | |
2120 | p->numa_faults[membuf_idx] = 0; | |
fb13c7ee | 2121 | |
7e2703e6 RR |
2122 | /* |
2123 | * Normalize the faults_from, so all tasks in a group | |
2124 | * count according to CPU use, instead of by the raw | |
2125 | * number of faults. Tasks with little runtime have | |
2126 | * little over-all impact on throughput, and thus their | |
2127 | * faults are less important. | |
2128 | */ | |
2129 | f_weight = div64_u64(runtime << 16, period + 1); | |
44dba3d5 | 2130 | f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / |
7e2703e6 | 2131 | (total_faults + 1); |
44dba3d5 IM |
2132 | f_diff = f_weight - p->numa_faults[cpu_idx] / 2; |
2133 | p->numa_faults[cpubuf_idx] = 0; | |
50ec8a40 | 2134 | |
44dba3d5 IM |
2135 | p->numa_faults[mem_idx] += diff; |
2136 | p->numa_faults[cpu_idx] += f_diff; | |
2137 | faults += p->numa_faults[mem_idx]; | |
83e1d2cd | 2138 | p->total_numa_faults += diff; |
8c8a743c | 2139 | if (p->numa_group) { |
44dba3d5 IM |
2140 | /* |
2141 | * safe because we can only change our own group | |
2142 | * | |
2143 | * mem_idx represents the offset for a given | |
2144 | * nid and priv in a specific region because it | |
2145 | * is at the beginning of the numa_faults array. | |
2146 | */ | |
2147 | p->numa_group->faults[mem_idx] += diff; | |
2148 | p->numa_group->faults_cpu[mem_idx] += f_diff; | |
989348b5 | 2149 | p->numa_group->total_faults += diff; |
44dba3d5 | 2150 | group_faults += p->numa_group->faults[mem_idx]; |
8c8a743c | 2151 | } |
ac8e895b MG |
2152 | } |
2153 | ||
688b7585 MG |
2154 | if (faults > max_faults) { |
2155 | max_faults = faults; | |
2156 | max_nid = nid; | |
2157 | } | |
83e1d2cd MG |
2158 | |
2159 | if (group_faults > max_group_faults) { | |
2160 | max_group_faults = group_faults; | |
2161 | max_group_nid = nid; | |
2162 | } | |
2163 | } | |
2164 | ||
04bb2f94 RR |
2165 | update_task_scan_period(p, fault_types[0], fault_types[1]); |
2166 | ||
7dbd13ed | 2167 | if (p->numa_group) { |
4142c3eb | 2168 | numa_group_count_active_nodes(p->numa_group); |
60e69eed | 2169 | spin_unlock_irq(group_lock); |
54009416 | 2170 | max_nid = preferred_group_nid(p, max_group_nid); |
688b7585 MG |
2171 | } |
2172 | ||
bb97fc31 RR |
2173 | if (max_faults) { |
2174 | /* Set the new preferred node */ | |
2175 | if (max_nid != p->numa_preferred_nid) | |
2176 | sched_setnuma(p, max_nid); | |
2177 | ||
2178 | if (task_node(p) != p->numa_preferred_nid) | |
2179 | numa_migrate_preferred(p); | |
3a7053b3 | 2180 | } |
cbee9f88 PZ |
2181 | } |
2182 | ||
8c8a743c PZ |
2183 | static inline int get_numa_group(struct numa_group *grp) |
2184 | { | |
2185 | return atomic_inc_not_zero(&grp->refcount); | |
2186 | } | |
2187 | ||
2188 | static inline void put_numa_group(struct numa_group *grp) | |
2189 | { | |
2190 | if (atomic_dec_and_test(&grp->refcount)) | |
2191 | kfree_rcu(grp, rcu); | |
2192 | } | |
2193 | ||
3e6a9418 MG |
2194 | static void task_numa_group(struct task_struct *p, int cpupid, int flags, |
2195 | int *priv) | |
8c8a743c PZ |
2196 | { |
2197 | struct numa_group *grp, *my_grp; | |
2198 | struct task_struct *tsk; | |
2199 | bool join = false; | |
2200 | int cpu = cpupid_to_cpu(cpupid); | |
2201 | int i; | |
2202 | ||
2203 | if (unlikely(!p->numa_group)) { | |
2204 | unsigned int size = sizeof(struct numa_group) + | |
50ec8a40 | 2205 | 4*nr_node_ids*sizeof(unsigned long); |
8c8a743c PZ |
2206 | |
2207 | grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); | |
2208 | if (!grp) | |
2209 | return; | |
2210 | ||
2211 | atomic_set(&grp->refcount, 1); | |
4142c3eb RR |
2212 | grp->active_nodes = 1; |
2213 | grp->max_faults_cpu = 0; | |
8c8a743c | 2214 | spin_lock_init(&grp->lock); |
e29cf08b | 2215 | grp->gid = p->pid; |
50ec8a40 | 2216 | /* Second half of the array tracks nids where faults happen */ |
be1e4e76 RR |
2217 | grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * |
2218 | nr_node_ids; | |
8c8a743c | 2219 | |
be1e4e76 | 2220 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2221 | grp->faults[i] = p->numa_faults[i]; |
8c8a743c | 2222 | |
989348b5 | 2223 | grp->total_faults = p->total_numa_faults; |
83e1d2cd | 2224 | |
8c8a743c PZ |
2225 | grp->nr_tasks++; |
2226 | rcu_assign_pointer(p->numa_group, grp); | |
2227 | } | |
2228 | ||
2229 | rcu_read_lock(); | |
316c1608 | 2230 | tsk = READ_ONCE(cpu_rq(cpu)->curr); |
8c8a743c PZ |
2231 | |
2232 | if (!cpupid_match_pid(tsk, cpupid)) | |
3354781a | 2233 | goto no_join; |
8c8a743c PZ |
2234 | |
2235 | grp = rcu_dereference(tsk->numa_group); | |
2236 | if (!grp) | |
3354781a | 2237 | goto no_join; |
8c8a743c PZ |
2238 | |
2239 | my_grp = p->numa_group; | |
2240 | if (grp == my_grp) | |
3354781a | 2241 | goto no_join; |
8c8a743c PZ |
2242 | |
2243 | /* | |
2244 | * Only join the other group if its bigger; if we're the bigger group, | |
2245 | * the other task will join us. | |
2246 | */ | |
2247 | if (my_grp->nr_tasks > grp->nr_tasks) | |
3354781a | 2248 | goto no_join; |
8c8a743c PZ |
2249 | |
2250 | /* | |
2251 | * Tie-break on the grp address. | |
2252 | */ | |
2253 | if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) | |
3354781a | 2254 | goto no_join; |
8c8a743c | 2255 | |
dabe1d99 RR |
2256 | /* Always join threads in the same process. */ |
2257 | if (tsk->mm == current->mm) | |
2258 | join = true; | |
2259 | ||
2260 | /* Simple filter to avoid false positives due to PID collisions */ | |
2261 | if (flags & TNF_SHARED) | |
2262 | join = true; | |
8c8a743c | 2263 | |
3e6a9418 MG |
2264 | /* Update priv based on whether false sharing was detected */ |
2265 | *priv = !join; | |
2266 | ||
dabe1d99 | 2267 | if (join && !get_numa_group(grp)) |
3354781a | 2268 | goto no_join; |
8c8a743c | 2269 | |
8c8a743c PZ |
2270 | rcu_read_unlock(); |
2271 | ||
2272 | if (!join) | |
2273 | return; | |
2274 | ||
60e69eed MG |
2275 | BUG_ON(irqs_disabled()); |
2276 | double_lock_irq(&my_grp->lock, &grp->lock); | |
989348b5 | 2277 | |
be1e4e76 | 2278 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { |
44dba3d5 IM |
2279 | my_grp->faults[i] -= p->numa_faults[i]; |
2280 | grp->faults[i] += p->numa_faults[i]; | |
8c8a743c | 2281 | } |
989348b5 MG |
2282 | my_grp->total_faults -= p->total_numa_faults; |
2283 | grp->total_faults += p->total_numa_faults; | |
8c8a743c | 2284 | |
8c8a743c PZ |
2285 | my_grp->nr_tasks--; |
2286 | grp->nr_tasks++; | |
2287 | ||
2288 | spin_unlock(&my_grp->lock); | |
60e69eed | 2289 | spin_unlock_irq(&grp->lock); |
8c8a743c PZ |
2290 | |
2291 | rcu_assign_pointer(p->numa_group, grp); | |
2292 | ||
2293 | put_numa_group(my_grp); | |
3354781a PZ |
2294 | return; |
2295 | ||
2296 | no_join: | |
2297 | rcu_read_unlock(); | |
2298 | return; | |
8c8a743c PZ |
2299 | } |
2300 | ||
2301 | void task_numa_free(struct task_struct *p) | |
2302 | { | |
2303 | struct numa_group *grp = p->numa_group; | |
44dba3d5 | 2304 | void *numa_faults = p->numa_faults; |
e9dd685c SR |
2305 | unsigned long flags; |
2306 | int i; | |
8c8a743c PZ |
2307 | |
2308 | if (grp) { | |
e9dd685c | 2309 | spin_lock_irqsave(&grp->lock, flags); |
be1e4e76 | 2310 | for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) |
44dba3d5 | 2311 | grp->faults[i] -= p->numa_faults[i]; |
989348b5 | 2312 | grp->total_faults -= p->total_numa_faults; |
83e1d2cd | 2313 | |
8c8a743c | 2314 | grp->nr_tasks--; |
e9dd685c | 2315 | spin_unlock_irqrestore(&grp->lock, flags); |
35b123e2 | 2316 | RCU_INIT_POINTER(p->numa_group, NULL); |
8c8a743c PZ |
2317 | put_numa_group(grp); |
2318 | } | |
2319 | ||
44dba3d5 | 2320 | p->numa_faults = NULL; |
82727018 | 2321 | kfree(numa_faults); |
8c8a743c PZ |
2322 | } |
2323 | ||
cbee9f88 PZ |
2324 | /* |
2325 | * Got a PROT_NONE fault for a page on @node. | |
2326 | */ | |
58b46da3 | 2327 | void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) |
cbee9f88 PZ |
2328 | { |
2329 | struct task_struct *p = current; | |
6688cc05 | 2330 | bool migrated = flags & TNF_MIGRATED; |
58b46da3 | 2331 | int cpu_node = task_node(current); |
792568ec | 2332 | int local = !!(flags & TNF_FAULT_LOCAL); |
4142c3eb | 2333 | struct numa_group *ng; |
ac8e895b | 2334 | int priv; |
cbee9f88 | 2335 | |
2a595721 | 2336 | if (!static_branch_likely(&sched_numa_balancing)) |
1a687c2e MG |
2337 | return; |
2338 | ||
9ff1d9ff MG |
2339 | /* for example, ksmd faulting in a user's mm */ |
2340 | if (!p->mm) | |
2341 | return; | |
2342 | ||
f809ca9a | 2343 | /* Allocate buffer to track faults on a per-node basis */ |
44dba3d5 IM |
2344 | if (unlikely(!p->numa_faults)) { |
2345 | int size = sizeof(*p->numa_faults) * | |
be1e4e76 | 2346 | NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; |
f809ca9a | 2347 | |
44dba3d5 IM |
2348 | p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); |
2349 | if (!p->numa_faults) | |
f809ca9a | 2350 | return; |
745d6147 | 2351 | |
83e1d2cd | 2352 | p->total_numa_faults = 0; |
04bb2f94 | 2353 | memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); |
f809ca9a | 2354 | } |
cbee9f88 | 2355 | |
8c8a743c PZ |
2356 | /* |
2357 | * First accesses are treated as private, otherwise consider accesses | |
2358 | * to be private if the accessing pid has not changed | |
2359 | */ | |
2360 | if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { | |
2361 | priv = 1; | |
2362 | } else { | |
2363 | priv = cpupid_match_pid(p, last_cpupid); | |
6688cc05 | 2364 | if (!priv && !(flags & TNF_NO_GROUP)) |
3e6a9418 | 2365 | task_numa_group(p, last_cpupid, flags, &priv); |
8c8a743c PZ |
2366 | } |
2367 | ||
792568ec RR |
2368 | /* |
2369 | * If a workload spans multiple NUMA nodes, a shared fault that | |
2370 | * occurs wholly within the set of nodes that the workload is | |
2371 | * actively using should be counted as local. This allows the | |
2372 | * scan rate to slow down when a workload has settled down. | |
2373 | */ | |
4142c3eb RR |
2374 | ng = p->numa_group; |
2375 | if (!priv && !local && ng && ng->active_nodes > 1 && | |
2376 | numa_is_active_node(cpu_node, ng) && | |
2377 | numa_is_active_node(mem_node, ng)) | |
792568ec RR |
2378 | local = 1; |
2379 | ||
cbee9f88 | 2380 | task_numa_placement(p); |
f809ca9a | 2381 | |
2739d3ee RR |
2382 | /* |
2383 | * Retry task to preferred node migration periodically, in case it | |
2384 | * case it previously failed, or the scheduler moved us. | |
2385 | */ | |
2386 | if (time_after(jiffies, p->numa_migrate_retry)) | |
6b9a7460 MG |
2387 | numa_migrate_preferred(p); |
2388 | ||
b32e86b4 IM |
2389 | if (migrated) |
2390 | p->numa_pages_migrated += pages; | |
074c2381 MG |
2391 | if (flags & TNF_MIGRATE_FAIL) |
2392 | p->numa_faults_locality[2] += pages; | |
b32e86b4 | 2393 | |
44dba3d5 IM |
2394 | p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; |
2395 | p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; | |
792568ec | 2396 | p->numa_faults_locality[local] += pages; |
cbee9f88 PZ |
2397 | } |
2398 | ||
6e5fb223 PZ |
2399 | static void reset_ptenuma_scan(struct task_struct *p) |
2400 | { | |
7e5a2c17 JL |
2401 | /* |
2402 | * We only did a read acquisition of the mmap sem, so | |
2403 | * p->mm->numa_scan_seq is written to without exclusive access | |
2404 | * and the update is not guaranteed to be atomic. That's not | |
2405 | * much of an issue though, since this is just used for | |
2406 | * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not | |
2407 | * expensive, to avoid any form of compiler optimizations: | |
2408 | */ | |
316c1608 | 2409 | WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); |
6e5fb223 PZ |
2410 | p->mm->numa_scan_offset = 0; |
2411 | } | |
2412 | ||
cbee9f88 PZ |
2413 | /* |
2414 | * The expensive part of numa migration is done from task_work context. | |
2415 | * Triggered from task_tick_numa(). | |
2416 | */ | |
2417 | void task_numa_work(struct callback_head *work) | |
2418 | { | |
2419 | unsigned long migrate, next_scan, now = jiffies; | |
2420 | struct task_struct *p = current; | |
2421 | struct mm_struct *mm = p->mm; | |
51170840 | 2422 | u64 runtime = p->se.sum_exec_runtime; |
6e5fb223 | 2423 | struct vm_area_struct *vma; |
9f40604c | 2424 | unsigned long start, end; |
598f0ec0 | 2425 | unsigned long nr_pte_updates = 0; |
4620f8c1 | 2426 | long pages, virtpages; |
cbee9f88 | 2427 | |
9148a3a1 | 2428 | SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); |
cbee9f88 PZ |
2429 | |
2430 | work->next = work; /* protect against double add */ | |
2431 | /* | |
2432 | * Who cares about NUMA placement when they're dying. | |
2433 | * | |
2434 | * NOTE: make sure not to dereference p->mm before this check, | |
2435 | * exit_task_work() happens _after_ exit_mm() so we could be called | |
2436 | * without p->mm even though we still had it when we enqueued this | |
2437 | * work. | |
2438 | */ | |
2439 | if (p->flags & PF_EXITING) | |
2440 | return; | |
2441 | ||
930aa174 | 2442 | if (!mm->numa_next_scan) { |
7e8d16b6 MG |
2443 | mm->numa_next_scan = now + |
2444 | msecs_to_jiffies(sysctl_numa_balancing_scan_delay); | |
b8593bfd MG |
2445 | } |
2446 | ||
cbee9f88 PZ |
2447 | /* |
2448 | * Enforce maximal scan/migration frequency.. | |
2449 | */ | |
2450 | migrate = mm->numa_next_scan; | |
2451 | if (time_before(now, migrate)) | |
2452 | return; | |
2453 | ||
598f0ec0 MG |
2454 | if (p->numa_scan_period == 0) { |
2455 | p->numa_scan_period_max = task_scan_max(p); | |
2456 | p->numa_scan_period = task_scan_min(p); | |
2457 | } | |
cbee9f88 | 2458 | |
fb003b80 | 2459 | next_scan = now + msecs_to_jiffies(p->numa_scan_period); |
cbee9f88 PZ |
2460 | if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) |
2461 | return; | |
2462 | ||
19a78d11 PZ |
2463 | /* |
2464 | * Delay this task enough that another task of this mm will likely win | |
2465 | * the next time around. | |
2466 | */ | |
2467 | p->node_stamp += 2 * TICK_NSEC; | |
2468 | ||
9f40604c MG |
2469 | start = mm->numa_scan_offset; |
2470 | pages = sysctl_numa_balancing_scan_size; | |
2471 | pages <<= 20 - PAGE_SHIFT; /* MB in pages */ | |
4620f8c1 | 2472 | virtpages = pages * 8; /* Scan up to this much virtual space */ |
9f40604c MG |
2473 | if (!pages) |
2474 | return; | |
cbee9f88 | 2475 | |
4620f8c1 | 2476 | |
6e5fb223 | 2477 | down_read(&mm->mmap_sem); |
9f40604c | 2478 | vma = find_vma(mm, start); |
6e5fb223 PZ |
2479 | if (!vma) { |
2480 | reset_ptenuma_scan(p); | |
9f40604c | 2481 | start = 0; |
6e5fb223 PZ |
2482 | vma = mm->mmap; |
2483 | } | |
9f40604c | 2484 | for (; vma; vma = vma->vm_next) { |
6b79c57b | 2485 | if (!vma_migratable(vma) || !vma_policy_mof(vma) || |
8e76d4ee | 2486 | is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { |
6e5fb223 | 2487 | continue; |
6b79c57b | 2488 | } |
6e5fb223 | 2489 | |
4591ce4f MG |
2490 | /* |
2491 | * Shared library pages mapped by multiple processes are not | |
2492 | * migrated as it is expected they are cache replicated. Avoid | |
2493 | * hinting faults in read-only file-backed mappings or the vdso | |
2494 | * as migrating the pages will be of marginal benefit. | |
2495 | */ | |
2496 | if (!vma->vm_mm || | |
2497 | (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) | |
2498 | continue; | |
2499 | ||
3c67f474 MG |
2500 | /* |
2501 | * Skip inaccessible VMAs to avoid any confusion between | |
2502 | * PROT_NONE and NUMA hinting ptes | |
2503 | */ | |
2504 | if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) | |
2505 | continue; | |
4591ce4f | 2506 | |
9f40604c MG |
2507 | do { |
2508 | start = max(start, vma->vm_start); | |
2509 | end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); | |
2510 | end = min(end, vma->vm_end); | |
4620f8c1 | 2511 | nr_pte_updates = change_prot_numa(vma, start, end); |
598f0ec0 MG |
2512 | |
2513 | /* | |
4620f8c1 RR |
2514 | * Try to scan sysctl_numa_balancing_size worth of |
2515 | * hpages that have at least one present PTE that | |
2516 | * is not already pte-numa. If the VMA contains | |
2517 | * areas that are unused or already full of prot_numa | |
2518 | * PTEs, scan up to virtpages, to skip through those | |
2519 | * areas faster. | |
598f0ec0 MG |
2520 | */ |
2521 | if (nr_pte_updates) | |
2522 | pages -= (end - start) >> PAGE_SHIFT; | |
4620f8c1 | 2523 | virtpages -= (end - start) >> PAGE_SHIFT; |
6e5fb223 | 2524 | |
9f40604c | 2525 | start = end; |
4620f8c1 | 2526 | if (pages <= 0 || virtpages <= 0) |
9f40604c | 2527 | goto out; |
3cf1962c RR |
2528 | |
2529 | cond_resched(); | |
9f40604c | 2530 | } while (end != vma->vm_end); |
cbee9f88 | 2531 | } |
6e5fb223 | 2532 | |
9f40604c | 2533 | out: |
6e5fb223 | 2534 | /* |
c69307d5 PZ |
2535 | * It is possible to reach the end of the VMA list but the last few |
2536 | * VMAs are not guaranteed to the vma_migratable. If they are not, we | |
2537 | * would find the !migratable VMA on the next scan but not reset the | |
2538 | * scanner to the start so check it now. | |
6e5fb223 PZ |
2539 | */ |
2540 | if (vma) | |
9f40604c | 2541 | mm->numa_scan_offset = start; |
6e5fb223 PZ |
2542 | else |
2543 | reset_ptenuma_scan(p); | |
2544 | up_read(&mm->mmap_sem); | |
51170840 RR |
2545 | |
2546 | /* | |
2547 | * Make sure tasks use at least 32x as much time to run other code | |
2548 | * than they used here, to limit NUMA PTE scanning overhead to 3% max. | |
2549 | * Usually update_task_scan_period slows down scanning enough; on an | |
2550 | * overloaded system we need to limit overhead on a per task basis. | |
2551 | */ | |
2552 | if (unlikely(p->se.sum_exec_runtime != runtime)) { | |
2553 | u64 diff = p->se.sum_exec_runtime - runtime; | |
2554 | p->node_stamp += 32 * diff; | |
2555 | } | |
cbee9f88 PZ |
2556 | } |
2557 | ||
2558 | /* | |
2559 | * Drive the periodic memory faults.. | |
2560 | */ | |
2561 | void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2562 | { | |
2563 | struct callback_head *work = &curr->numa_work; | |
2564 | u64 period, now; | |
2565 | ||
2566 | /* | |
2567 | * We don't care about NUMA placement if we don't have memory. | |
2568 | */ | |
2569 | if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) | |
2570 | return; | |
2571 | ||
2572 | /* | |
2573 | * Using runtime rather than walltime has the dual advantage that | |
2574 | * we (mostly) drive the selection from busy threads and that the | |
2575 | * task needs to have done some actual work before we bother with | |
2576 | * NUMA placement. | |
2577 | */ | |
2578 | now = curr->se.sum_exec_runtime; | |
2579 | period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; | |
2580 | ||
25b3e5a3 | 2581 | if (now > curr->node_stamp + period) { |
4b96a29b | 2582 | if (!curr->node_stamp) |
598f0ec0 | 2583 | curr->numa_scan_period = task_scan_min(curr); |
19a78d11 | 2584 | curr->node_stamp += period; |
cbee9f88 PZ |
2585 | |
2586 | if (!time_before(jiffies, curr->mm->numa_next_scan)) { | |
2587 | init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ | |
2588 | task_work_add(curr, work, true); | |
2589 | } | |
2590 | } | |
2591 | } | |
2592 | #else | |
2593 | static void task_tick_numa(struct rq *rq, struct task_struct *curr) | |
2594 | { | |
2595 | } | |
0ec8aa00 PZ |
2596 | |
2597 | static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) | |
2598 | { | |
2599 | } | |
2600 | ||
2601 | static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) | |
2602 | { | |
2603 | } | |
cbee9f88 PZ |
2604 | #endif /* CONFIG_NUMA_BALANCING */ |
2605 | ||
30cfdcfc DA |
2606 | static void |
2607 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2608 | { | |
2609 | update_load_add(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2610 | if (!parent_entity(se)) |
029632fb | 2611 | update_load_add(&rq_of(cfs_rq)->load, se->load.weight); |
367456c7 | 2612 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2613 | if (entity_is_task(se)) { |
2614 | struct rq *rq = rq_of(cfs_rq); | |
2615 | ||
2616 | account_numa_enqueue(rq, task_of(se)); | |
2617 | list_add(&se->group_node, &rq->cfs_tasks); | |
2618 | } | |
367456c7 | 2619 | #endif |
30cfdcfc | 2620 | cfs_rq->nr_running++; |
30cfdcfc DA |
2621 | } |
2622 | ||
2623 | static void | |
2624 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2625 | { | |
2626 | update_load_sub(&cfs_rq->load, se->load.weight); | |
c09595f6 | 2627 | if (!parent_entity(se)) |
029632fb | 2628 | update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); |
bfdb198c | 2629 | #ifdef CONFIG_SMP |
0ec8aa00 PZ |
2630 | if (entity_is_task(se)) { |
2631 | account_numa_dequeue(rq_of(cfs_rq), task_of(se)); | |
b87f1724 | 2632 | list_del_init(&se->group_node); |
0ec8aa00 | 2633 | } |
bfdb198c | 2634 | #endif |
30cfdcfc | 2635 | cfs_rq->nr_running--; |
30cfdcfc DA |
2636 | } |
2637 | ||
3ff6dcac YZ |
2638 | #ifdef CONFIG_FAIR_GROUP_SCHED |
2639 | # ifdef CONFIG_SMP | |
ea1dc6fc | 2640 | static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
cf5f0acf | 2641 | { |
ea1dc6fc | 2642 | long tg_weight, load, shares; |
cf5f0acf PZ |
2643 | |
2644 | /* | |
ea1dc6fc PZ |
2645 | * This really should be: cfs_rq->avg.load_avg, but instead we use |
2646 | * cfs_rq->load.weight, which is its upper bound. This helps ramp up | |
2647 | * the shares for small weight interactive tasks. | |
cf5f0acf | 2648 | */ |
ea1dc6fc | 2649 | load = scale_load_down(cfs_rq->load.weight); |
cf5f0acf | 2650 | |
ea1dc6fc | 2651 | tg_weight = atomic_long_read(&tg->load_avg); |
3ff6dcac | 2652 | |
ea1dc6fc PZ |
2653 | /* Ensure tg_weight >= load */ |
2654 | tg_weight -= cfs_rq->tg_load_avg_contrib; | |
2655 | tg_weight += load; | |
3ff6dcac | 2656 | |
3ff6dcac | 2657 | shares = (tg->shares * load); |
cf5f0acf PZ |
2658 | if (tg_weight) |
2659 | shares /= tg_weight; | |
3ff6dcac | 2660 | |
b8fd8423 DE |
2661 | /* |
2662 | * MIN_SHARES has to be unscaled here to support per-CPU partitioning | |
2663 | * of a group with small tg->shares value. It is a floor value which is | |
2664 | * assigned as a minimum load.weight to the sched_entity representing | |
2665 | * the group on a CPU. | |
2666 | * | |
2667 | * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 | |
2668 | * on an 8-core system with 8 tasks each runnable on one CPU shares has | |
2669 | * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In | |
2670 | * case no task is runnable on a CPU MIN_SHARES=2 should be returned | |
2671 | * instead of 0. | |
2672 | */ | |
3ff6dcac YZ |
2673 | if (shares < MIN_SHARES) |
2674 | shares = MIN_SHARES; | |
2675 | if (shares > tg->shares) | |
2676 | shares = tg->shares; | |
2677 | ||
2678 | return shares; | |
2679 | } | |
3ff6dcac | 2680 | # else /* CONFIG_SMP */ |
6d5ab293 | 2681 | static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) |
3ff6dcac YZ |
2682 | { |
2683 | return tg->shares; | |
2684 | } | |
3ff6dcac | 2685 | # endif /* CONFIG_SMP */ |
ea1dc6fc | 2686 | |
2069dd75 PZ |
2687 | static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, |
2688 | unsigned long weight) | |
2689 | { | |
19e5eebb PT |
2690 | if (se->on_rq) { |
2691 | /* commit outstanding execution time */ | |
2692 | if (cfs_rq->curr == se) | |
2693 | update_curr(cfs_rq); | |
2069dd75 | 2694 | account_entity_dequeue(cfs_rq, se); |
19e5eebb | 2695 | } |
2069dd75 PZ |
2696 | |
2697 | update_load_set(&se->load, weight); | |
2698 | ||
2699 | if (se->on_rq) | |
2700 | account_entity_enqueue(cfs_rq, se); | |
2701 | } | |
2702 | ||
82958366 PT |
2703 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); |
2704 | ||
89ee048f | 2705 | static void update_cfs_shares(struct sched_entity *se) |
2069dd75 | 2706 | { |
89ee048f | 2707 | struct cfs_rq *cfs_rq = group_cfs_rq(se); |
2069dd75 | 2708 | struct task_group *tg; |
3ff6dcac | 2709 | long shares; |
2069dd75 | 2710 | |
89ee048f VG |
2711 | if (!cfs_rq) |
2712 | return; | |
2713 | ||
2714 | if (throttled_hierarchy(cfs_rq)) | |
2069dd75 | 2715 | return; |
89ee048f VG |
2716 | |
2717 | tg = cfs_rq->tg; | |
2718 | ||
3ff6dcac YZ |
2719 | #ifndef CONFIG_SMP |
2720 | if (likely(se->load.weight == tg->shares)) | |
2721 | return; | |
2722 | #endif | |
6d5ab293 | 2723 | shares = calc_cfs_shares(cfs_rq, tg); |
2069dd75 PZ |
2724 | |
2725 | reweight_entity(cfs_rq_of(se), se, shares); | |
2726 | } | |
89ee048f | 2727 | |
2069dd75 | 2728 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
89ee048f | 2729 | static inline void update_cfs_shares(struct sched_entity *se) |
2069dd75 PZ |
2730 | { |
2731 | } | |
2732 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
2733 | ||
141965c7 | 2734 | #ifdef CONFIG_SMP |
5b51f2f8 PT |
2735 | /* Precomputed fixed inverse multiplies for multiplication by y^n */ |
2736 | static const u32 runnable_avg_yN_inv[] = { | |
2737 | 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, | |
2738 | 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, | |
2739 | 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, | |
2740 | 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, | |
2741 | 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, | |
2742 | 0x85aac367, 0x82cd8698, | |
2743 | }; | |
2744 | ||
9d85f21c PT |
2745 | /* |
2746 | * Approximate: | |
2747 | * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) | |
2748 | */ | |
a481db34 | 2749 | static u64 decay_load(u64 val, u64 n) |
9d85f21c | 2750 | { |
5b51f2f8 PT |
2751 | unsigned int local_n; |
2752 | ||
05296e75 | 2753 | if (unlikely(n > LOAD_AVG_PERIOD * 63)) |
5b51f2f8 PT |
2754 | return 0; |
2755 | ||
2756 | /* after bounds checking we can collapse to 32-bit */ | |
2757 | local_n = n; | |
2758 | ||
2759 | /* | |
2760 | * As y^PERIOD = 1/2, we can combine | |
9c58c79a ZZ |
2761 | * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD) |
2762 | * With a look-up table which covers y^n (n<PERIOD) | |
5b51f2f8 PT |
2763 | * |
2764 | * To achieve constant time decay_load. | |
2765 | */ | |
2766 | if (unlikely(local_n >= LOAD_AVG_PERIOD)) { | |
2767 | val >>= local_n / LOAD_AVG_PERIOD; | |
2768 | local_n %= LOAD_AVG_PERIOD; | |
9d85f21c PT |
2769 | } |
2770 | ||
9d89c257 YD |
2771 | val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32); |
2772 | return val; | |
5b51f2f8 PT |
2773 | } |
2774 | ||
05296e75 | 2775 | static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3) |
5b51f2f8 | 2776 | { |
05296e75 | 2777 | u32 c1, c2, c3 = d3; /* y^0 == 1 */ |
5b51f2f8 | 2778 | |
a481db34 YD |
2779 | /* |
2780 | * c1 = d1 y^(p+1) | |
2781 | */ | |
05296e75 | 2782 | c1 = decay_load((u64)d1, periods); |
a481db34 | 2783 | |
a481db34 | 2784 | /* |
05296e75 PZ |
2785 | * p |
2786 | * c2 = 1024 \Sum y^n | |
2787 | * n=1 | |
a481db34 | 2788 | * |
05296e75 PZ |
2789 | * inf inf |
2790 | * = 1024 ( \Sum y^n - \Sum y^n - y^0 ) | |
2791 | * n=0 n=p+1 | |
a481db34 | 2792 | */ |
05296e75 | 2793 | c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024; |
a481db34 YD |
2794 | |
2795 | return c1 + c2 + c3; | |
9d85f21c PT |
2796 | } |
2797 | ||
54a21385 | 2798 | #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT) |
e0f5f3af | 2799 | |
a481db34 YD |
2800 | /* |
2801 | * Accumulate the three separate parts of the sum; d1 the remainder | |
2802 | * of the last (incomplete) period, d2 the span of full periods and d3 | |
2803 | * the remainder of the (incomplete) current period. | |
2804 | * | |
2805 | * d1 d2 d3 | |
2806 | * ^ ^ ^ | |
2807 | * | | | | |
2808 | * |<->|<----------------->|<--->| | |
2809 | * ... |---x---|------| ... |------|-----x (now) | |
2810 | * | |
2811 | * p | |
2812 | * u' = (u + d1) y^(p+1) + 1024 \Sum y^n + d3 y^0 | |
2813 | * n=1 | |
2814 | * | |
2815 | * = u y^(p+1) + (Step 1) | |
2816 | * | |
2817 | * p | |
2818 | * d1 y^(p+1) + 1024 \Sum y^n + d3 y^0 (Step 2) | |
2819 | * n=1 | |
2820 | */ | |
2821 | static __always_inline u32 | |
2822 | accumulate_sum(u64 delta, int cpu, struct sched_avg *sa, | |
2823 | unsigned long weight, int running, struct cfs_rq *cfs_rq) | |
2824 | { | |
2825 | unsigned long scale_freq, scale_cpu; | |
05296e75 | 2826 | u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */ |
a481db34 | 2827 | u64 periods; |
a481db34 YD |
2828 | |
2829 | scale_freq = arch_scale_freq_capacity(NULL, cpu); | |
2830 | scale_cpu = arch_scale_cpu_capacity(NULL, cpu); | |
2831 | ||
2832 | delta += sa->period_contrib; | |
2833 | periods = delta / 1024; /* A period is 1024us (~1ms) */ | |
2834 | ||
2835 | /* | |
2836 | * Step 1: decay old *_sum if we crossed period boundaries. | |
2837 | */ | |
2838 | if (periods) { | |
2839 | sa->load_sum = decay_load(sa->load_sum, periods); | |
2840 | if (cfs_rq) { | |
2841 | cfs_rq->runnable_load_sum = | |
2842 | decay_load(cfs_rq->runnable_load_sum, periods); | |
2843 | } | |
2844 | sa->util_sum = decay_load((u64)(sa->util_sum), periods); | |
a481db34 | 2845 | |
05296e75 PZ |
2846 | /* |
2847 | * Step 2 | |
2848 | */ | |
2849 | delta %= 1024; | |
2850 | contrib = __accumulate_pelt_segments(periods, | |
2851 | 1024 - sa->period_contrib, delta); | |
2852 | } | |
a481db34 YD |
2853 | sa->period_contrib = delta; |
2854 | ||
2855 | contrib = cap_scale(contrib, scale_freq); | |
2856 | if (weight) { | |
2857 | sa->load_sum += weight * contrib; | |
2858 | if (cfs_rq) | |
2859 | cfs_rq->runnable_load_sum += weight * contrib; | |
2860 | } | |
2861 | if (running) | |
2862 | sa->util_sum += contrib * scale_cpu; | |
2863 | ||
2864 | return periods; | |
2865 | } | |
2866 | ||
9d85f21c PT |
2867 | /* |
2868 | * We can represent the historical contribution to runnable average as the | |
2869 | * coefficients of a geometric series. To do this we sub-divide our runnable | |
2870 | * history into segments of approximately 1ms (1024us); label the segment that | |
2871 | * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. | |
2872 | * | |
2873 | * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... | |
2874 | * p0 p1 p2 | |
2875 | * (now) (~1ms ago) (~2ms ago) | |
2876 | * | |
2877 | * Let u_i denote the fraction of p_i that the entity was runnable. | |
2878 | * | |
2879 | * We then designate the fractions u_i as our co-efficients, yielding the | |
2880 | * following representation of historical load: | |
2881 | * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... | |
2882 | * | |
2883 | * We choose y based on the with of a reasonably scheduling period, fixing: | |
2884 | * y^32 = 0.5 | |
2885 | * | |
2886 | * This means that the contribution to load ~32ms ago (u_32) will be weighted | |
2887 | * approximately half as much as the contribution to load within the last ms | |
2888 | * (u_0). | |
2889 | * | |
2890 | * When a period "rolls over" and we have new u_0`, multiplying the previous | |
2891 | * sum again by y is sufficient to update: | |
2892 | * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) | |
2893 | * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] | |
2894 | */ | |
9d89c257 | 2895 | static __always_inline int |
0ccb977f | 2896 | ___update_load_avg(u64 now, int cpu, struct sched_avg *sa, |
13962234 | 2897 | unsigned long weight, int running, struct cfs_rq *cfs_rq) |
9d85f21c | 2898 | { |
a481db34 | 2899 | u64 delta; |
9d85f21c | 2900 | |
9d89c257 | 2901 | delta = now - sa->last_update_time; |
9d85f21c PT |
2902 | /* |
2903 | * This should only happen when time goes backwards, which it | |
2904 | * unfortunately does during sched clock init when we swap over to TSC. | |
2905 | */ | |
2906 | if ((s64)delta < 0) { | |
9d89c257 | 2907 | sa->last_update_time = now; |
9d85f21c PT |
2908 | return 0; |
2909 | } | |
2910 | ||
2911 | /* | |
2912 | * Use 1024ns as the unit of measurement since it's a reasonable | |
2913 | * approximation of 1us and fast to compute. | |
2914 | */ | |
2915 | delta >>= 10; | |
2916 | if (!delta) | |
2917 | return 0; | |
9d89c257 | 2918 | sa->last_update_time = now; |
9d85f21c | 2919 | |
a481db34 YD |
2920 | /* |
2921 | * Now we know we crossed measurement unit boundaries. The *_avg | |
2922 | * accrues by two steps: | |
2923 | * | |
2924 | * Step 1: accumulate *_sum since last_update_time. If we haven't | |
2925 | * crossed period boundaries, finish. | |
2926 | */ | |
2927 | if (!accumulate_sum(delta, cpu, sa, weight, running, cfs_rq)) | |
2928 | return 0; | |
9ee474f5 | 2929 | |
a481db34 YD |
2930 | /* |
2931 | * Step 2: update *_avg. | |
2932 | */ | |
2933 | sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX); | |
2934 | if (cfs_rq) { | |
2935 | cfs_rq->runnable_load_avg = | |
2936 | div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX); | |
9d89c257 | 2937 | } |
a481db34 | 2938 | sa->util_avg = sa->util_sum / LOAD_AVG_MAX; |
aff3e498 | 2939 | |
a481db34 | 2940 | return 1; |
9ee474f5 PT |
2941 | } |
2942 | ||
0ccb977f PZ |
2943 | static int |
2944 | __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se) | |
2945 | { | |
2946 | return ___update_load_avg(now, cpu, &se->avg, 0, 0, NULL); | |
2947 | } | |
2948 | ||
2949 | static int | |
2950 | __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se) | |
2951 | { | |
2952 | return ___update_load_avg(now, cpu, &se->avg, | |
2953 | se->on_rq * scale_load_down(se->load.weight), | |
2954 | cfs_rq->curr == se, NULL); | |
2955 | } | |
2956 | ||
2957 | static int | |
2958 | __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq) | |
2959 | { | |
2960 | return ___update_load_avg(now, cpu, &cfs_rq->avg, | |
2961 | scale_load_down(cfs_rq->load.weight), | |
2962 | cfs_rq->curr != NULL, cfs_rq); | |
2963 | } | |
2964 | ||
09a43ace VG |
2965 | /* |
2966 | * Signed add and clamp on underflow. | |
2967 | * | |
2968 | * Explicitly do a load-store to ensure the intermediate value never hits | |
2969 | * memory. This allows lockless observations without ever seeing the negative | |
2970 | * values. | |
2971 | */ | |
2972 | #define add_positive(_ptr, _val) do { \ | |
2973 | typeof(_ptr) ptr = (_ptr); \ | |
2974 | typeof(_val) val = (_val); \ | |
2975 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
2976 | \ | |
2977 | res = var + val; \ | |
2978 | \ | |
2979 | if (val < 0 && res > var) \ | |
2980 | res = 0; \ | |
2981 | \ | |
2982 | WRITE_ONCE(*ptr, res); \ | |
2983 | } while (0) | |
2984 | ||
c566e8e9 | 2985 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7c3edd2c PZ |
2986 | /** |
2987 | * update_tg_load_avg - update the tg's load avg | |
2988 | * @cfs_rq: the cfs_rq whose avg changed | |
2989 | * @force: update regardless of how small the difference | |
2990 | * | |
2991 | * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. | |
2992 | * However, because tg->load_avg is a global value there are performance | |
2993 | * considerations. | |
2994 | * | |
2995 | * In order to avoid having to look at the other cfs_rq's, we use a | |
2996 | * differential update where we store the last value we propagated. This in | |
2997 | * turn allows skipping updates if the differential is 'small'. | |
2998 | * | |
2999 | * Updating tg's load_avg is necessary before update_cfs_share() (which is | |
3000 | * done) and effective_load() (which is not done because it is too costly). | |
bb17f655 | 3001 | */ |
9d89c257 | 3002 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) |
bb17f655 | 3003 | { |
9d89c257 | 3004 | long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; |
bb17f655 | 3005 | |
aa0b7ae0 WL |
3006 | /* |
3007 | * No need to update load_avg for root_task_group as it is not used. | |
3008 | */ | |
3009 | if (cfs_rq->tg == &root_task_group) | |
3010 | return; | |
3011 | ||
9d89c257 YD |
3012 | if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { |
3013 | atomic_long_add(delta, &cfs_rq->tg->load_avg); | |
3014 | cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; | |
bb17f655 | 3015 | } |
8165e145 | 3016 | } |
f5f9739d | 3017 | |
ad936d86 BP |
3018 | /* |
3019 | * Called within set_task_rq() right before setting a task's cpu. The | |
3020 | * caller only guarantees p->pi_lock is held; no other assumptions, | |
3021 | * including the state of rq->lock, should be made. | |
3022 | */ | |
3023 | void set_task_rq_fair(struct sched_entity *se, | |
3024 | struct cfs_rq *prev, struct cfs_rq *next) | |
3025 | { | |
0ccb977f PZ |
3026 | u64 p_last_update_time; |
3027 | u64 n_last_update_time; | |
3028 | ||
ad936d86 BP |
3029 | if (!sched_feat(ATTACH_AGE_LOAD)) |
3030 | return; | |
3031 | ||
3032 | /* | |
3033 | * We are supposed to update the task to "current" time, then its up to | |
3034 | * date and ready to go to new CPU/cfs_rq. But we have difficulty in | |
3035 | * getting what current time is, so simply throw away the out-of-date | |
3036 | * time. This will result in the wakee task is less decayed, but giving | |
3037 | * the wakee more load sounds not bad. | |
3038 | */ | |
0ccb977f PZ |
3039 | if (!(se->avg.last_update_time && prev)) |
3040 | return; | |
ad936d86 BP |
3041 | |
3042 | #ifndef CONFIG_64BIT | |
0ccb977f | 3043 | { |
ad936d86 BP |
3044 | u64 p_last_update_time_copy; |
3045 | u64 n_last_update_time_copy; | |
3046 | ||
3047 | do { | |
3048 | p_last_update_time_copy = prev->load_last_update_time_copy; | |
3049 | n_last_update_time_copy = next->load_last_update_time_copy; | |
3050 | ||
3051 | smp_rmb(); | |
3052 | ||
3053 | p_last_update_time = prev->avg.last_update_time; | |
3054 | n_last_update_time = next->avg.last_update_time; | |
3055 | ||
3056 | } while (p_last_update_time != p_last_update_time_copy || | |
3057 | n_last_update_time != n_last_update_time_copy); | |
0ccb977f | 3058 | } |
ad936d86 | 3059 | #else |
0ccb977f PZ |
3060 | p_last_update_time = prev->avg.last_update_time; |
3061 | n_last_update_time = next->avg.last_update_time; | |
ad936d86 | 3062 | #endif |
0ccb977f PZ |
3063 | __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se); |
3064 | se->avg.last_update_time = n_last_update_time; | |
ad936d86 | 3065 | } |
09a43ace VG |
3066 | |
3067 | /* Take into account change of utilization of a child task group */ | |
3068 | static inline void | |
3069 | update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3070 | { | |
3071 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3072 | long delta = gcfs_rq->avg.util_avg - se->avg.util_avg; | |
3073 | ||
3074 | /* Nothing to update */ | |
3075 | if (!delta) | |
3076 | return; | |
3077 | ||
3078 | /* Set new sched_entity's utilization */ | |
3079 | se->avg.util_avg = gcfs_rq->avg.util_avg; | |
3080 | se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX; | |
3081 | ||
3082 | /* Update parent cfs_rq utilization */ | |
3083 | add_positive(&cfs_rq->avg.util_avg, delta); | |
3084 | cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX; | |
3085 | } | |
3086 | ||
3087 | /* Take into account change of load of a child task group */ | |
3088 | static inline void | |
3089 | update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3090 | { | |
3091 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3092 | long delta, load = gcfs_rq->avg.load_avg; | |
3093 | ||
3094 | /* | |
3095 | * If the load of group cfs_rq is null, the load of the | |
3096 | * sched_entity will also be null so we can skip the formula | |
3097 | */ | |
3098 | if (load) { | |
3099 | long tg_load; | |
3100 | ||
3101 | /* Get tg's load and ensure tg_load > 0 */ | |
3102 | tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1; | |
3103 | ||
3104 | /* Ensure tg_load >= load and updated with current load*/ | |
3105 | tg_load -= gcfs_rq->tg_load_avg_contrib; | |
3106 | tg_load += load; | |
3107 | ||
3108 | /* | |
3109 | * We need to compute a correction term in the case that the | |
3110 | * task group is consuming more CPU than a task of equal | |
3111 | * weight. A task with a weight equals to tg->shares will have | |
3112 | * a load less or equal to scale_load_down(tg->shares). | |
3113 | * Similarly, the sched_entities that represent the task group | |
3114 | * at parent level, can't have a load higher than | |
3115 | * scale_load_down(tg->shares). And the Sum of sched_entities' | |
3116 | * load must be <= scale_load_down(tg->shares). | |
3117 | */ | |
3118 | if (tg_load > scale_load_down(gcfs_rq->tg->shares)) { | |
3119 | /* scale gcfs_rq's load into tg's shares*/ | |
3120 | load *= scale_load_down(gcfs_rq->tg->shares); | |
3121 | load /= tg_load; | |
3122 | } | |
3123 | } | |
3124 | ||
3125 | delta = load - se->avg.load_avg; | |
3126 | ||
3127 | /* Nothing to update */ | |
3128 | if (!delta) | |
3129 | return; | |
3130 | ||
3131 | /* Set new sched_entity's load */ | |
3132 | se->avg.load_avg = load; | |
3133 | se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX; | |
3134 | ||
3135 | /* Update parent cfs_rq load */ | |
3136 | add_positive(&cfs_rq->avg.load_avg, delta); | |
3137 | cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX; | |
3138 | ||
3139 | /* | |
3140 | * If the sched_entity is already enqueued, we also have to update the | |
3141 | * runnable load avg. | |
3142 | */ | |
3143 | if (se->on_rq) { | |
3144 | /* Update parent cfs_rq runnable_load_avg */ | |
3145 | add_positive(&cfs_rq->runnable_load_avg, delta); | |
3146 | cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX; | |
3147 | } | |
3148 | } | |
3149 | ||
3150 | static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) | |
3151 | { | |
3152 | cfs_rq->propagate_avg = 1; | |
3153 | } | |
3154 | ||
3155 | static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se) | |
3156 | { | |
3157 | struct cfs_rq *cfs_rq = group_cfs_rq(se); | |
3158 | ||
3159 | if (!cfs_rq->propagate_avg) | |
3160 | return 0; | |
3161 | ||
3162 | cfs_rq->propagate_avg = 0; | |
3163 | return 1; | |
3164 | } | |
3165 | ||
3166 | /* Update task and its cfs_rq load average */ | |
3167 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3168 | { | |
3169 | struct cfs_rq *cfs_rq; | |
3170 | ||
3171 | if (entity_is_task(se)) | |
3172 | return 0; | |
3173 | ||
3174 | if (!test_and_clear_tg_cfs_propagate(se)) | |
3175 | return 0; | |
3176 | ||
3177 | cfs_rq = cfs_rq_of(se); | |
3178 | ||
3179 | set_tg_cfs_propagate(cfs_rq); | |
3180 | ||
3181 | update_tg_cfs_util(cfs_rq, se); | |
3182 | update_tg_cfs_load(cfs_rq, se); | |
3183 | ||
3184 | return 1; | |
3185 | } | |
3186 | ||
bc427898 VG |
3187 | /* |
3188 | * Check if we need to update the load and the utilization of a blocked | |
3189 | * group_entity: | |
3190 | */ | |
3191 | static inline bool skip_blocked_update(struct sched_entity *se) | |
3192 | { | |
3193 | struct cfs_rq *gcfs_rq = group_cfs_rq(se); | |
3194 | ||
3195 | /* | |
3196 | * If sched_entity still have not zero load or utilization, we have to | |
3197 | * decay it: | |
3198 | */ | |
3199 | if (se->avg.load_avg || se->avg.util_avg) | |
3200 | return false; | |
3201 | ||
3202 | /* | |
3203 | * If there is a pending propagation, we have to update the load and | |
3204 | * the utilization of the sched_entity: | |
3205 | */ | |
3206 | if (gcfs_rq->propagate_avg) | |
3207 | return false; | |
3208 | ||
3209 | /* | |
3210 | * Otherwise, the load and the utilization of the sched_entity is | |
3211 | * already zero and there is no pending propagation, so it will be a | |
3212 | * waste of time to try to decay it: | |
3213 | */ | |
3214 | return true; | |
3215 | } | |
3216 | ||
6e83125c | 3217 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
09a43ace | 3218 | |
9d89c257 | 3219 | static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {} |
09a43ace VG |
3220 | |
3221 | static inline int propagate_entity_load_avg(struct sched_entity *se) | |
3222 | { | |
3223 | return 0; | |
3224 | } | |
3225 | ||
3226 | static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {} | |
3227 | ||
6e83125c | 3228 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
c566e8e9 | 3229 | |
a2c6c91f SM |
3230 | static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq) |
3231 | { | |
58919e83 | 3232 | if (&this_rq()->cfs == cfs_rq) { |
a2c6c91f SM |
3233 | /* |
3234 | * There are a few boundary cases this might miss but it should | |
3235 | * get called often enough that that should (hopefully) not be | |
3236 | * a real problem -- added to that it only calls on the local | |
3237 | * CPU, so if we enqueue remotely we'll miss an update, but | |
3238 | * the next tick/schedule should update. | |
3239 | * | |
3240 | * It will not get called when we go idle, because the idle | |
3241 | * thread is a different class (!fair), nor will the utilization | |
3242 | * number include things like RT tasks. | |
3243 | * | |
3244 | * As is, the util number is not freq-invariant (we'd have to | |
3245 | * implement arch_scale_freq_capacity() for that). | |
3246 | * | |
3247 | * See cpu_util(). | |
3248 | */ | |
12bde33d | 3249 | cpufreq_update_util(rq_of(cfs_rq), 0); |
a2c6c91f SM |
3250 | } |
3251 | } | |
3252 | ||
89741892 PZ |
3253 | /* |
3254 | * Unsigned subtract and clamp on underflow. | |
3255 | * | |
3256 | * Explicitly do a load-store to ensure the intermediate value never hits | |
3257 | * memory. This allows lockless observations without ever seeing the negative | |
3258 | * values. | |
3259 | */ | |
3260 | #define sub_positive(_ptr, _val) do { \ | |
3261 | typeof(_ptr) ptr = (_ptr); \ | |
3262 | typeof(*ptr) val = (_val); \ | |
3263 | typeof(*ptr) res, var = READ_ONCE(*ptr); \ | |
3264 | res = var - val; \ | |
3265 | if (res > var) \ | |
3266 | res = 0; \ | |
3267 | WRITE_ONCE(*ptr, res); \ | |
3268 | } while (0) | |
3269 | ||
3d30544f PZ |
3270 | /** |
3271 | * update_cfs_rq_load_avg - update the cfs_rq's load/util averages | |
3272 | * @now: current time, as per cfs_rq_clock_task() | |
3273 | * @cfs_rq: cfs_rq to update | |
3274 | * @update_freq: should we call cfs_rq_util_change() or will the call do so | |
3275 | * | |
3276 | * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) | |
3277 | * avg. The immediate corollary is that all (fair) tasks must be attached, see | |
3278 | * post_init_entity_util_avg(). | |
3279 | * | |
3280 | * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. | |
3281 | * | |
7c3edd2c PZ |
3282 | * Returns true if the load decayed or we removed load. |
3283 | * | |
3284 | * Since both these conditions indicate a changed cfs_rq->avg.load we should | |
3285 | * call update_tg_load_avg() when this function returns true. | |
3d30544f | 3286 | */ |
a2c6c91f SM |
3287 | static inline int |
3288 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq) | |
2dac754e | 3289 | { |
9d89c257 | 3290 | struct sched_avg *sa = &cfs_rq->avg; |
41e0d37f | 3291 | int decayed, removed_load = 0, removed_util = 0; |
2dac754e | 3292 | |
9d89c257 | 3293 | if (atomic_long_read(&cfs_rq->removed_load_avg)) { |
9e0e83a1 | 3294 | s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0); |
89741892 PZ |
3295 | sub_positive(&sa->load_avg, r); |
3296 | sub_positive(&sa->load_sum, r * LOAD_AVG_MAX); | |
41e0d37f | 3297 | removed_load = 1; |
4e516076 | 3298 | set_tg_cfs_propagate(cfs_rq); |
8165e145 | 3299 | } |
2dac754e | 3300 | |
9d89c257 YD |
3301 | if (atomic_long_read(&cfs_rq->removed_util_avg)) { |
3302 | long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0); | |
89741892 PZ |
3303 | sub_positive(&sa->util_avg, r); |
3304 | sub_positive(&sa->util_sum, r * LOAD_AVG_MAX); | |
41e0d37f | 3305 | removed_util = 1; |
4e516076 | 3306 | set_tg_cfs_propagate(cfs_rq); |
9d89c257 | 3307 | } |
36ee28e4 | 3308 | |
0ccb977f | 3309 | decayed = __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq); |
36ee28e4 | 3310 | |
9d89c257 YD |
3311 | #ifndef CONFIG_64BIT |
3312 | smp_wmb(); | |
3313 | cfs_rq->load_last_update_time_copy = sa->last_update_time; | |
3314 | #endif | |
36ee28e4 | 3315 | |
a2c6c91f SM |
3316 | if (update_freq && (decayed || removed_util)) |
3317 | cfs_rq_util_change(cfs_rq); | |
21e96f88 | 3318 | |
41e0d37f | 3319 | return decayed || removed_load; |
21e96f88 SM |
3320 | } |
3321 | ||
d31b1a66 VG |
3322 | /* |
3323 | * Optional action to be done while updating the load average | |
3324 | */ | |
3325 | #define UPDATE_TG 0x1 | |
3326 | #define SKIP_AGE_LOAD 0x2 | |
3327 | ||
21e96f88 | 3328 | /* Update task and its cfs_rq load average */ |
d31b1a66 | 3329 | static inline void update_load_avg(struct sched_entity *se, int flags) |
21e96f88 SM |
3330 | { |
3331 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3332 | u64 now = cfs_rq_clock_task(cfs_rq); | |
3333 | struct rq *rq = rq_of(cfs_rq); | |
3334 | int cpu = cpu_of(rq); | |
09a43ace | 3335 | int decayed; |
21e96f88 SM |
3336 | |
3337 | /* | |
3338 | * Track task load average for carrying it to new CPU after migrated, and | |
3339 | * track group sched_entity load average for task_h_load calc in migration | |
3340 | */ | |
0ccb977f PZ |
3341 | if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) |
3342 | __update_load_avg_se(now, cpu, cfs_rq, se); | |
21e96f88 | 3343 | |
09a43ace VG |
3344 | decayed = update_cfs_rq_load_avg(now, cfs_rq, true); |
3345 | decayed |= propagate_entity_load_avg(se); | |
3346 | ||
3347 | if (decayed && (flags & UPDATE_TG)) | |
21e96f88 | 3348 | update_tg_load_avg(cfs_rq, 0); |
9ee474f5 PT |
3349 | } |
3350 | ||
3d30544f PZ |
3351 | /** |
3352 | * attach_entity_load_avg - attach this entity to its cfs_rq load avg | |
3353 | * @cfs_rq: cfs_rq to attach to | |
3354 | * @se: sched_entity to attach | |
3355 | * | |
3356 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3357 | * cfs_rq->avg.last_update_time being current. | |
3358 | */ | |
a05e8c51 BP |
3359 | static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3360 | { | |
3361 | se->avg.last_update_time = cfs_rq->avg.last_update_time; | |
3362 | cfs_rq->avg.load_avg += se->avg.load_avg; | |
3363 | cfs_rq->avg.load_sum += se->avg.load_sum; | |
3364 | cfs_rq->avg.util_avg += se->avg.util_avg; | |
3365 | cfs_rq->avg.util_sum += se->avg.util_sum; | |
09a43ace | 3366 | set_tg_cfs_propagate(cfs_rq); |
a2c6c91f SM |
3367 | |
3368 | cfs_rq_util_change(cfs_rq); | |
a05e8c51 BP |
3369 | } |
3370 | ||
3d30544f PZ |
3371 | /** |
3372 | * detach_entity_load_avg - detach this entity from its cfs_rq load avg | |
3373 | * @cfs_rq: cfs_rq to detach from | |
3374 | * @se: sched_entity to detach | |
3375 | * | |
3376 | * Must call update_cfs_rq_load_avg() before this, since we rely on | |
3377 | * cfs_rq->avg.last_update_time being current. | |
3378 | */ | |
a05e8c51 BP |
3379 | static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3380 | { | |
a05e8c51 | 3381 | |
89741892 PZ |
3382 | sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); |
3383 | sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum); | |
3384 | sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); | |
3385 | sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); | |
09a43ace | 3386 | set_tg_cfs_propagate(cfs_rq); |
a2c6c91f SM |
3387 | |
3388 | cfs_rq_util_change(cfs_rq); | |
a05e8c51 BP |
3389 | } |
3390 | ||
9d89c257 YD |
3391 | /* Add the load generated by se into cfs_rq's load average */ |
3392 | static inline void | |
3393 | enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
9ee474f5 | 3394 | { |
9d89c257 | 3395 | struct sched_avg *sa = &se->avg; |
18bf2805 | 3396 | |
13962234 YD |
3397 | cfs_rq->runnable_load_avg += sa->load_avg; |
3398 | cfs_rq->runnable_load_sum += sa->load_sum; | |
3399 | ||
d31b1a66 | 3400 | if (!sa->last_update_time) { |
a05e8c51 | 3401 | attach_entity_load_avg(cfs_rq, se); |
9d89c257 | 3402 | update_tg_load_avg(cfs_rq, 0); |
d31b1a66 | 3403 | } |
2dac754e PT |
3404 | } |
3405 | ||
13962234 YD |
3406 | /* Remove the runnable load generated by se from cfs_rq's runnable load average */ |
3407 | static inline void | |
3408 | dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) | |
3409 | { | |
13962234 YD |
3410 | cfs_rq->runnable_load_avg = |
3411 | max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0); | |
3412 | cfs_rq->runnable_load_sum = | |
a05e8c51 | 3413 | max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0); |
13962234 YD |
3414 | } |
3415 | ||
9d89c257 | 3416 | #ifndef CONFIG_64BIT |
0905f04e YD |
3417 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3418 | { | |
9d89c257 | 3419 | u64 last_update_time_copy; |
0905f04e | 3420 | u64 last_update_time; |
9ee474f5 | 3421 | |
9d89c257 YD |
3422 | do { |
3423 | last_update_time_copy = cfs_rq->load_last_update_time_copy; | |
3424 | smp_rmb(); | |
3425 | last_update_time = cfs_rq->avg.last_update_time; | |
3426 | } while (last_update_time != last_update_time_copy); | |
0905f04e YD |
3427 | |
3428 | return last_update_time; | |
3429 | } | |
9d89c257 | 3430 | #else |
0905f04e YD |
3431 | static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) |
3432 | { | |
3433 | return cfs_rq->avg.last_update_time; | |
3434 | } | |
9d89c257 YD |
3435 | #endif |
3436 | ||
104cb16d MR |
3437 | /* |
3438 | * Synchronize entity load avg of dequeued entity without locking | |
3439 | * the previous rq. | |
3440 | */ | |
3441 | void sync_entity_load_avg(struct sched_entity *se) | |
3442 | { | |
3443 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
3444 | u64 last_update_time; | |
3445 | ||
3446 | last_update_time = cfs_rq_last_update_time(cfs_rq); | |
0ccb977f | 3447 | __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se); |
104cb16d MR |
3448 | } |
3449 | ||
0905f04e YD |
3450 | /* |
3451 | * Task first catches up with cfs_rq, and then subtract | |
3452 | * itself from the cfs_rq (task must be off the queue now). | |
3453 | */ | |
3454 | void remove_entity_load_avg(struct sched_entity *se) | |
3455 | { | |
3456 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
0905f04e YD |
3457 | |
3458 | /* | |
7dc603c9 PZ |
3459 | * tasks cannot exit without having gone through wake_up_new_task() -> |
3460 | * post_init_entity_util_avg() which will have added things to the | |
3461 | * cfs_rq, so we can remove unconditionally. | |
3462 | * | |
3463 | * Similarly for groups, they will have passed through | |
3464 | * post_init_entity_util_avg() before unregister_sched_fair_group() | |
3465 | * calls this. | |
0905f04e | 3466 | */ |
0905f04e | 3467 | |
104cb16d | 3468 | sync_entity_load_avg(se); |
9d89c257 YD |
3469 | atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg); |
3470 | atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg); | |
2dac754e | 3471 | } |
642dbc39 | 3472 | |
7ea241af YD |
3473 | static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq) |
3474 | { | |
3475 | return cfs_rq->runnable_load_avg; | |
3476 | } | |
3477 | ||
3478 | static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) | |
3479 | { | |
3480 | return cfs_rq->avg.load_avg; | |
3481 | } | |
3482 | ||
46f69fa3 | 3483 | static int idle_balance(struct rq *this_rq, struct rq_flags *rf); |
6e83125c | 3484 | |
38033c37 PZ |
3485 | #else /* CONFIG_SMP */ |
3486 | ||
01011473 PZ |
3487 | static inline int |
3488 | update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq) | |
3489 | { | |
3490 | return 0; | |
3491 | } | |
3492 | ||
d31b1a66 VG |
3493 | #define UPDATE_TG 0x0 |
3494 | #define SKIP_AGE_LOAD 0x0 | |
3495 | ||
3496 | static inline void update_load_avg(struct sched_entity *se, int not_used1) | |
536bd00c | 3497 | { |
12bde33d | 3498 | cpufreq_update_util(rq_of(cfs_rq_of(se)), 0); |
536bd00c RW |
3499 | } |
3500 | ||
9d89c257 YD |
3501 | static inline void |
3502 | enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
13962234 YD |
3503 | static inline void |
3504 | dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
9d89c257 | 3505 | static inline void remove_entity_load_avg(struct sched_entity *se) {} |
6e83125c | 3506 | |
a05e8c51 BP |
3507 | static inline void |
3508 | attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
3509 | static inline void | |
3510 | detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} | |
3511 | ||
46f69fa3 | 3512 | static inline int idle_balance(struct rq *rq, struct rq_flags *rf) |
6e83125c PZ |
3513 | { |
3514 | return 0; | |
3515 | } | |
3516 | ||
38033c37 | 3517 | #endif /* CONFIG_SMP */ |
9d85f21c | 3518 | |
ddc97297 PZ |
3519 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3520 | { | |
3521 | #ifdef CONFIG_SCHED_DEBUG | |
3522 | s64 d = se->vruntime - cfs_rq->min_vruntime; | |
3523 | ||
3524 | if (d < 0) | |
3525 | d = -d; | |
3526 | ||
3527 | if (d > 3*sysctl_sched_latency) | |
ae92882e | 3528 | schedstat_inc(cfs_rq->nr_spread_over); |
ddc97297 PZ |
3529 | #endif |
3530 | } | |
3531 | ||
aeb73b04 PZ |
3532 | static void |
3533 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) | |
3534 | { | |
1af5f730 | 3535 | u64 vruntime = cfs_rq->min_vruntime; |
94dfb5e7 | 3536 | |
2cb8600e PZ |
3537 | /* |
3538 | * The 'current' period is already promised to the current tasks, | |
3539 | * however the extra weight of the new task will slow them down a | |
3540 | * little, place the new task so that it fits in the slot that | |
3541 | * stays open at the end. | |
3542 | */ | |
94dfb5e7 | 3543 | if (initial && sched_feat(START_DEBIT)) |
f9c0b095 | 3544 | vruntime += sched_vslice(cfs_rq, se); |
aeb73b04 | 3545 | |
a2e7a7eb | 3546 | /* sleeps up to a single latency don't count. */ |
5ca9880c | 3547 | if (!initial) { |
a2e7a7eb | 3548 | unsigned long thresh = sysctl_sched_latency; |
a7be37ac | 3549 | |
a2e7a7eb MG |
3550 | /* |
3551 | * Halve their sleep time's effect, to allow | |
3552 | * for a gentler effect of sleepers: | |
3553 | */ | |
3554 | if (sched_feat(GENTLE_FAIR_SLEEPERS)) | |
3555 | thresh >>= 1; | |
51e0304c | 3556 | |
a2e7a7eb | 3557 | vruntime -= thresh; |
aeb73b04 PZ |
3558 | } |
3559 | ||
b5d9d734 | 3560 | /* ensure we never gain time by being placed backwards. */ |
16c8f1c7 | 3561 | se->vruntime = max_vruntime(se->vruntime, vruntime); |
aeb73b04 PZ |
3562 | } |
3563 | ||
d3d9dc33 PT |
3564 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq); |
3565 | ||
cb251765 MG |
3566 | static inline void check_schedstat_required(void) |
3567 | { | |
3568 | #ifdef CONFIG_SCHEDSTATS | |
3569 | if (schedstat_enabled()) | |
3570 | return; | |
3571 | ||
3572 | /* Force schedstat enabled if a dependent tracepoint is active */ | |
3573 | if (trace_sched_stat_wait_enabled() || | |
3574 | trace_sched_stat_sleep_enabled() || | |
3575 | trace_sched_stat_iowait_enabled() || | |
3576 | trace_sched_stat_blocked_enabled() || | |
3577 | trace_sched_stat_runtime_enabled()) { | |
eda8dca5 | 3578 | printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " |
cb251765 MG |
3579 | "stat_blocked and stat_runtime require the " |
3580 | "kernel parameter schedstats=enabled or " | |
3581 | "kernel.sched_schedstats=1\n"); | |
3582 | } | |
3583 | #endif | |
3584 | } | |
3585 | ||
b5179ac7 PZ |
3586 | |
3587 | /* | |
3588 | * MIGRATION | |
3589 | * | |
3590 | * dequeue | |
3591 | * update_curr() | |
3592 | * update_min_vruntime() | |
3593 | * vruntime -= min_vruntime | |
3594 | * | |
3595 | * enqueue | |
3596 | * update_curr() | |
3597 | * update_min_vruntime() | |
3598 | * vruntime += min_vruntime | |
3599 | * | |
3600 | * this way the vruntime transition between RQs is done when both | |
3601 | * min_vruntime are up-to-date. | |
3602 | * | |
3603 | * WAKEUP (remote) | |
3604 | * | |
59efa0ba | 3605 | * ->migrate_task_rq_fair() (p->state == TASK_WAKING) |
b5179ac7 PZ |
3606 | * vruntime -= min_vruntime |
3607 | * | |
3608 | * enqueue | |
3609 | * update_curr() | |
3610 | * update_min_vruntime() | |
3611 | * vruntime += min_vruntime | |
3612 | * | |
3613 | * this way we don't have the most up-to-date min_vruntime on the originating | |
3614 | * CPU and an up-to-date min_vruntime on the destination CPU. | |
3615 | */ | |
3616 | ||
bf0f6f24 | 3617 | static void |
88ec22d3 | 3618 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3619 | { |
2f950354 PZ |
3620 | bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); |
3621 | bool curr = cfs_rq->curr == se; | |
3622 | ||
88ec22d3 | 3623 | /* |
2f950354 PZ |
3624 | * If we're the current task, we must renormalise before calling |
3625 | * update_curr(). | |
88ec22d3 | 3626 | */ |
2f950354 | 3627 | if (renorm && curr) |
88ec22d3 PZ |
3628 | se->vruntime += cfs_rq->min_vruntime; |
3629 | ||
2f950354 PZ |
3630 | update_curr(cfs_rq); |
3631 | ||
bf0f6f24 | 3632 | /* |
2f950354 PZ |
3633 | * Otherwise, renormalise after, such that we're placed at the current |
3634 | * moment in time, instead of some random moment in the past. Being | |
3635 | * placed in the past could significantly boost this task to the | |
3636 | * fairness detriment of existing tasks. | |
bf0f6f24 | 3637 | */ |
2f950354 PZ |
3638 | if (renorm && !curr) |
3639 | se->vruntime += cfs_rq->min_vruntime; | |
3640 | ||
89ee048f VG |
3641 | /* |
3642 | * When enqueuing a sched_entity, we must: | |
3643 | * - Update loads to have both entity and cfs_rq synced with now. | |
3644 | * - Add its load to cfs_rq->runnable_avg | |
3645 | * - For group_entity, update its weight to reflect the new share of | |
3646 | * its group cfs_rq | |
3647 | * - Add its new weight to cfs_rq->load.weight | |
3648 | */ | |
d31b1a66 | 3649 | update_load_avg(se, UPDATE_TG); |
9d89c257 | 3650 | enqueue_entity_load_avg(cfs_rq, se); |
89ee048f | 3651 | update_cfs_shares(se); |
17bc14b7 | 3652 | account_entity_enqueue(cfs_rq, se); |
bf0f6f24 | 3653 | |
1a3d027c | 3654 | if (flags & ENQUEUE_WAKEUP) |
aeb73b04 | 3655 | place_entity(cfs_rq, se, 0); |
bf0f6f24 | 3656 | |
cb251765 | 3657 | check_schedstat_required(); |
4fa8d299 JP |
3658 | update_stats_enqueue(cfs_rq, se, flags); |
3659 | check_spread(cfs_rq, se); | |
2f950354 | 3660 | if (!curr) |
83b699ed | 3661 | __enqueue_entity(cfs_rq, se); |
2069dd75 | 3662 | se->on_rq = 1; |
3d4b47b4 | 3663 | |
d3d9dc33 | 3664 | if (cfs_rq->nr_running == 1) { |
3d4b47b4 | 3665 | list_add_leaf_cfs_rq(cfs_rq); |
d3d9dc33 PT |
3666 | check_enqueue_throttle(cfs_rq); |
3667 | } | |
bf0f6f24 IM |
3668 | } |
3669 | ||
2c13c919 | 3670 | static void __clear_buddies_last(struct sched_entity *se) |
2002c695 | 3671 | { |
2c13c919 RR |
3672 | for_each_sched_entity(se) { |
3673 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3674 | if (cfs_rq->last != se) |
2c13c919 | 3675 | break; |
f1044799 PZ |
3676 | |
3677 | cfs_rq->last = NULL; | |
2c13c919 RR |
3678 | } |
3679 | } | |
2002c695 | 3680 | |
2c13c919 RR |
3681 | static void __clear_buddies_next(struct sched_entity *se) |
3682 | { | |
3683 | for_each_sched_entity(se) { | |
3684 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3685 | if (cfs_rq->next != se) |
2c13c919 | 3686 | break; |
f1044799 PZ |
3687 | |
3688 | cfs_rq->next = NULL; | |
2c13c919 | 3689 | } |
2002c695 PZ |
3690 | } |
3691 | ||
ac53db59 RR |
3692 | static void __clear_buddies_skip(struct sched_entity *se) |
3693 | { | |
3694 | for_each_sched_entity(se) { | |
3695 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
f1044799 | 3696 | if (cfs_rq->skip != se) |
ac53db59 | 3697 | break; |
f1044799 PZ |
3698 | |
3699 | cfs_rq->skip = NULL; | |
ac53db59 RR |
3700 | } |
3701 | } | |
3702 | ||
a571bbea PZ |
3703 | static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) |
3704 | { | |
2c13c919 RR |
3705 | if (cfs_rq->last == se) |
3706 | __clear_buddies_last(se); | |
3707 | ||
3708 | if (cfs_rq->next == se) | |
3709 | __clear_buddies_next(se); | |
ac53db59 RR |
3710 | |
3711 | if (cfs_rq->skip == se) | |
3712 | __clear_buddies_skip(se); | |
a571bbea PZ |
3713 | } |
3714 | ||
6c16a6dc | 3715 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d8b4986d | 3716 | |
bf0f6f24 | 3717 | static void |
371fd7e7 | 3718 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) |
bf0f6f24 | 3719 | { |
a2a2d680 DA |
3720 | /* |
3721 | * Update run-time statistics of the 'current'. | |
3722 | */ | |
3723 | update_curr(cfs_rq); | |
89ee048f VG |
3724 | |
3725 | /* | |
3726 | * When dequeuing a sched_entity, we must: | |
3727 | * - Update loads to have both entity and cfs_rq synced with now. | |
3728 | * - Substract its load from the cfs_rq->runnable_avg. | |
3729 | * - Substract its previous weight from cfs_rq->load.weight. | |
3730 | * - For group entity, update its weight to reflect the new share | |
3731 | * of its group cfs_rq. | |
3732 | */ | |
d31b1a66 | 3733 | update_load_avg(se, UPDATE_TG); |
13962234 | 3734 | dequeue_entity_load_avg(cfs_rq, se); |
a2a2d680 | 3735 | |
4fa8d299 | 3736 | update_stats_dequeue(cfs_rq, se, flags); |
67e9fb2a | 3737 | |
2002c695 | 3738 | clear_buddies(cfs_rq, se); |
4793241b | 3739 | |
83b699ed | 3740 | if (se != cfs_rq->curr) |
30cfdcfc | 3741 | __dequeue_entity(cfs_rq, se); |
17bc14b7 | 3742 | se->on_rq = 0; |
30cfdcfc | 3743 | account_entity_dequeue(cfs_rq, se); |
88ec22d3 PZ |
3744 | |
3745 | /* | |
b60205c7 PZ |
3746 | * Normalize after update_curr(); which will also have moved |
3747 | * min_vruntime if @se is the one holding it back. But before doing | |
3748 | * update_min_vruntime() again, which will discount @se's position and | |
3749 | * can move min_vruntime forward still more. | |
88ec22d3 | 3750 | */ |
371fd7e7 | 3751 | if (!(flags & DEQUEUE_SLEEP)) |
88ec22d3 | 3752 | se->vruntime -= cfs_rq->min_vruntime; |
1e876231 | 3753 | |
d8b4986d PT |
3754 | /* return excess runtime on last dequeue */ |
3755 | return_cfs_rq_runtime(cfs_rq); | |
3756 | ||
89ee048f | 3757 | update_cfs_shares(se); |
b60205c7 PZ |
3758 | |
3759 | /* | |
3760 | * Now advance min_vruntime if @se was the entity holding it back, | |
3761 | * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be | |
3762 | * put back on, and if we advance min_vruntime, we'll be placed back | |
3763 | * further than we started -- ie. we'll be penalized. | |
3764 | */ | |
3765 | if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE) | |
3766 | update_min_vruntime(cfs_rq); | |
bf0f6f24 IM |
3767 | } |
3768 | ||
3769 | /* | |
3770 | * Preempt the current task with a newly woken task if needed: | |
3771 | */ | |
7c92e54f | 3772 | static void |
2e09bf55 | 3773 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
bf0f6f24 | 3774 | { |
11697830 | 3775 | unsigned long ideal_runtime, delta_exec; |
f4cfb33e WX |
3776 | struct sched_entity *se; |
3777 | s64 delta; | |
11697830 | 3778 | |
6d0f0ebd | 3779 | ideal_runtime = sched_slice(cfs_rq, curr); |
11697830 | 3780 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
a9f3e2b5 | 3781 | if (delta_exec > ideal_runtime) { |
8875125e | 3782 | resched_curr(rq_of(cfs_rq)); |
a9f3e2b5 MG |
3783 | /* |
3784 | * The current task ran long enough, ensure it doesn't get | |
3785 | * re-elected due to buddy favours. | |
3786 | */ | |
3787 | clear_buddies(cfs_rq, curr); | |
f685ceac MG |
3788 | return; |
3789 | } | |
3790 | ||
3791 | /* | |
3792 | * Ensure that a task that missed wakeup preemption by a | |
3793 | * narrow margin doesn't have to wait for a full slice. | |
3794 | * This also mitigates buddy induced latencies under load. | |
3795 | */ | |
f685ceac MG |
3796 | if (delta_exec < sysctl_sched_min_granularity) |
3797 | return; | |
3798 | ||
f4cfb33e WX |
3799 | se = __pick_first_entity(cfs_rq); |
3800 | delta = curr->vruntime - se->vruntime; | |
f685ceac | 3801 | |
f4cfb33e WX |
3802 | if (delta < 0) |
3803 | return; | |
d7d82944 | 3804 | |
f4cfb33e | 3805 | if (delta > ideal_runtime) |
8875125e | 3806 | resched_curr(rq_of(cfs_rq)); |
bf0f6f24 IM |
3807 | } |
3808 | ||
83b699ed | 3809 | static void |
8494f412 | 3810 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
bf0f6f24 | 3811 | { |
83b699ed SV |
3812 | /* 'current' is not kept within the tree. */ |
3813 | if (se->on_rq) { | |
3814 | /* | |
3815 | * Any task has to be enqueued before it get to execute on | |
3816 | * a CPU. So account for the time it spent waiting on the | |
3817 | * runqueue. | |
3818 | */ | |
4fa8d299 | 3819 | update_stats_wait_end(cfs_rq, se); |
83b699ed | 3820 | __dequeue_entity(cfs_rq, se); |
d31b1a66 | 3821 | update_load_avg(se, UPDATE_TG); |
83b699ed SV |
3822 | } |
3823 | ||
79303e9e | 3824 | update_stats_curr_start(cfs_rq, se); |
429d43bc | 3825 | cfs_rq->curr = se; |
4fa8d299 | 3826 | |
eba1ed4b IM |
3827 | /* |
3828 | * Track our maximum slice length, if the CPU's load is at | |
3829 | * least twice that of our own weight (i.e. dont track it | |
3830 | * when there are only lesser-weight tasks around): | |
3831 | */ | |
cb251765 | 3832 | if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
4fa8d299 JP |
3833 | schedstat_set(se->statistics.slice_max, |
3834 | max((u64)schedstat_val(se->statistics.slice_max), | |
3835 | se->sum_exec_runtime - se->prev_sum_exec_runtime)); | |
eba1ed4b | 3836 | } |
4fa8d299 | 3837 | |
4a55b450 | 3838 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
bf0f6f24 IM |
3839 | } |
3840 | ||
3f3a4904 PZ |
3841 | static int |
3842 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); | |
3843 | ||
ac53db59 RR |
3844 | /* |
3845 | * Pick the next process, keeping these things in mind, in this order: | |
3846 | * 1) keep things fair between processes/task groups | |
3847 | * 2) pick the "next" process, since someone really wants that to run | |
3848 | * 3) pick the "last" process, for cache locality | |
3849 | * 4) do not run the "skip" process, if something else is available | |
3850 | */ | |
678d5718 PZ |
3851 | static struct sched_entity * |
3852 | pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) | |
aa2ac252 | 3853 | { |
678d5718 PZ |
3854 | struct sched_entity *left = __pick_first_entity(cfs_rq); |
3855 | struct sched_entity *se; | |
3856 | ||
3857 | /* | |
3858 | * If curr is set we have to see if its left of the leftmost entity | |
3859 | * still in the tree, provided there was anything in the tree at all. | |
3860 | */ | |
3861 | if (!left || (curr && entity_before(curr, left))) | |
3862 | left = curr; | |
3863 | ||
3864 | se = left; /* ideally we run the leftmost entity */ | |
f4b6755f | 3865 | |
ac53db59 RR |
3866 | /* |
3867 | * Avoid running the skip buddy, if running something else can | |
3868 | * be done without getting too unfair. | |
3869 | */ | |
3870 | if (cfs_rq->skip == se) { | |
678d5718 PZ |
3871 | struct sched_entity *second; |
3872 | ||
3873 | if (se == curr) { | |
3874 | second = __pick_first_entity(cfs_rq); | |
3875 | } else { | |
3876 | second = __pick_next_entity(se); | |
3877 | if (!second || (curr && entity_before(curr, second))) | |
3878 | second = curr; | |
3879 | } | |
3880 | ||
ac53db59 RR |
3881 | if (second && wakeup_preempt_entity(second, left) < 1) |
3882 | se = second; | |
3883 | } | |
aa2ac252 | 3884 | |
f685ceac MG |
3885 | /* |
3886 | * Prefer last buddy, try to return the CPU to a preempted task. | |
3887 | */ | |
3888 | if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) | |
3889 | se = cfs_rq->last; | |
3890 | ||
ac53db59 RR |
3891 | /* |
3892 | * Someone really wants this to run. If it's not unfair, run it. | |
3893 | */ | |
3894 | if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) | |
3895 | se = cfs_rq->next; | |
3896 | ||
f685ceac | 3897 | clear_buddies(cfs_rq, se); |
4793241b PZ |
3898 | |
3899 | return se; | |
aa2ac252 PZ |
3900 | } |
3901 | ||
678d5718 | 3902 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); |
d3d9dc33 | 3903 | |
ab6cde26 | 3904 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
bf0f6f24 IM |
3905 | { |
3906 | /* | |
3907 | * If still on the runqueue then deactivate_task() | |
3908 | * was not called and update_curr() has to be done: | |
3909 | */ | |
3910 | if (prev->on_rq) | |
b7cc0896 | 3911 | update_curr(cfs_rq); |
bf0f6f24 | 3912 | |
d3d9dc33 PT |
3913 | /* throttle cfs_rqs exceeding runtime */ |
3914 | check_cfs_rq_runtime(cfs_rq); | |
3915 | ||
4fa8d299 | 3916 | check_spread(cfs_rq, prev); |
cb251765 | 3917 | |
30cfdcfc | 3918 | if (prev->on_rq) { |
4fa8d299 | 3919 | update_stats_wait_start(cfs_rq, prev); |
30cfdcfc DA |
3920 | /* Put 'current' back into the tree. */ |
3921 | __enqueue_entity(cfs_rq, prev); | |
9d85f21c | 3922 | /* in !on_rq case, update occurred at dequeue */ |
9d89c257 | 3923 | update_load_avg(prev, 0); |
30cfdcfc | 3924 | } |
429d43bc | 3925 | cfs_rq->curr = NULL; |
bf0f6f24 IM |
3926 | } |
3927 | ||
8f4d37ec PZ |
3928 | static void |
3929 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) | |
bf0f6f24 | 3930 | { |
bf0f6f24 | 3931 | /* |
30cfdcfc | 3932 | * Update run-time statistics of the 'current'. |
bf0f6f24 | 3933 | */ |
30cfdcfc | 3934 | update_curr(cfs_rq); |
bf0f6f24 | 3935 | |
9d85f21c PT |
3936 | /* |
3937 | * Ensure that runnable average is periodically updated. | |
3938 | */ | |
d31b1a66 | 3939 | update_load_avg(curr, UPDATE_TG); |
89ee048f | 3940 | update_cfs_shares(curr); |
9d85f21c | 3941 | |
8f4d37ec PZ |
3942 | #ifdef CONFIG_SCHED_HRTICK |
3943 | /* | |
3944 | * queued ticks are scheduled to match the slice, so don't bother | |
3945 | * validating it and just reschedule. | |
3946 | */ | |
983ed7a6 | 3947 | if (queued) { |
8875125e | 3948 | resched_curr(rq_of(cfs_rq)); |
983ed7a6 HH |
3949 | return; |
3950 | } | |
8f4d37ec PZ |
3951 | /* |
3952 | * don't let the period tick interfere with the hrtick preemption | |
3953 | */ | |
3954 | if (!sched_feat(DOUBLE_TICK) && | |
3955 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) | |
3956 | return; | |
3957 | #endif | |
3958 | ||
2c2efaed | 3959 | if (cfs_rq->nr_running > 1) |
2e09bf55 | 3960 | check_preempt_tick(cfs_rq, curr); |
bf0f6f24 IM |
3961 | } |
3962 | ||
ab84d31e PT |
3963 | |
3964 | /************************************************** | |
3965 | * CFS bandwidth control machinery | |
3966 | */ | |
3967 | ||
3968 | #ifdef CONFIG_CFS_BANDWIDTH | |
029632fb PZ |
3969 | |
3970 | #ifdef HAVE_JUMP_LABEL | |
c5905afb | 3971 | static struct static_key __cfs_bandwidth_used; |
029632fb PZ |
3972 | |
3973 | static inline bool cfs_bandwidth_used(void) | |
3974 | { | |
c5905afb | 3975 | return static_key_false(&__cfs_bandwidth_used); |
029632fb PZ |
3976 | } |
3977 | ||
1ee14e6c | 3978 | void cfs_bandwidth_usage_inc(void) |
029632fb | 3979 | { |
1ee14e6c BS |
3980 | static_key_slow_inc(&__cfs_bandwidth_used); |
3981 | } | |
3982 | ||
3983 | void cfs_bandwidth_usage_dec(void) | |
3984 | { | |
3985 | static_key_slow_dec(&__cfs_bandwidth_used); | |
029632fb PZ |
3986 | } |
3987 | #else /* HAVE_JUMP_LABEL */ | |
3988 | static bool cfs_bandwidth_used(void) | |
3989 | { | |
3990 | return true; | |
3991 | } | |
3992 | ||
1ee14e6c BS |
3993 | void cfs_bandwidth_usage_inc(void) {} |
3994 | void cfs_bandwidth_usage_dec(void) {} | |
029632fb PZ |
3995 | #endif /* HAVE_JUMP_LABEL */ |
3996 | ||
ab84d31e PT |
3997 | /* |
3998 | * default period for cfs group bandwidth. | |
3999 | * default: 0.1s, units: nanoseconds | |
4000 | */ | |
4001 | static inline u64 default_cfs_period(void) | |
4002 | { | |
4003 | return 100000000ULL; | |
4004 | } | |
ec12cb7f PT |
4005 | |
4006 | static inline u64 sched_cfs_bandwidth_slice(void) | |
4007 | { | |
4008 | return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; | |
4009 | } | |
4010 | ||
a9cf55b2 PT |
4011 | /* |
4012 | * Replenish runtime according to assigned quota and update expiration time. | |
4013 | * We use sched_clock_cpu directly instead of rq->clock to avoid adding | |
4014 | * additional synchronization around rq->lock. | |
4015 | * | |
4016 | * requires cfs_b->lock | |
4017 | */ | |
029632fb | 4018 | void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) |
a9cf55b2 PT |
4019 | { |
4020 | u64 now; | |
4021 | ||
4022 | if (cfs_b->quota == RUNTIME_INF) | |
4023 | return; | |
4024 | ||
4025 | now = sched_clock_cpu(smp_processor_id()); | |
4026 | cfs_b->runtime = cfs_b->quota; | |
4027 | cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); | |
4028 | } | |
4029 | ||
029632fb PZ |
4030 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4031 | { | |
4032 | return &tg->cfs_bandwidth; | |
4033 | } | |
4034 | ||
f1b17280 PT |
4035 | /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ |
4036 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) | |
4037 | { | |
4038 | if (unlikely(cfs_rq->throttle_count)) | |
1a99ae3f | 4039 | return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time; |
f1b17280 | 4040 | |
78becc27 | 4041 | return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; |
f1b17280 PT |
4042 | } |
4043 | ||
85dac906 PT |
4044 | /* returns 0 on failure to allocate runtime */ |
4045 | static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f PT |
4046 | { |
4047 | struct task_group *tg = cfs_rq->tg; | |
4048 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); | |
a9cf55b2 | 4049 | u64 amount = 0, min_amount, expires; |
ec12cb7f PT |
4050 | |
4051 | /* note: this is a positive sum as runtime_remaining <= 0 */ | |
4052 | min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; | |
4053 | ||
4054 | raw_spin_lock(&cfs_b->lock); | |
4055 | if (cfs_b->quota == RUNTIME_INF) | |
4056 | amount = min_amount; | |
58088ad0 | 4057 | else { |
77a4d1a1 | 4058 | start_cfs_bandwidth(cfs_b); |
58088ad0 PT |
4059 | |
4060 | if (cfs_b->runtime > 0) { | |
4061 | amount = min(cfs_b->runtime, min_amount); | |
4062 | cfs_b->runtime -= amount; | |
4063 | cfs_b->idle = 0; | |
4064 | } | |
ec12cb7f | 4065 | } |
a9cf55b2 | 4066 | expires = cfs_b->runtime_expires; |
ec12cb7f PT |
4067 | raw_spin_unlock(&cfs_b->lock); |
4068 | ||
4069 | cfs_rq->runtime_remaining += amount; | |
a9cf55b2 PT |
4070 | /* |
4071 | * we may have advanced our local expiration to account for allowed | |
4072 | * spread between our sched_clock and the one on which runtime was | |
4073 | * issued. | |
4074 | */ | |
4075 | if ((s64)(expires - cfs_rq->runtime_expires) > 0) | |
4076 | cfs_rq->runtime_expires = expires; | |
85dac906 PT |
4077 | |
4078 | return cfs_rq->runtime_remaining > 0; | |
ec12cb7f PT |
4079 | } |
4080 | ||
a9cf55b2 PT |
4081 | /* |
4082 | * Note: This depends on the synchronization provided by sched_clock and the | |
4083 | * fact that rq->clock snapshots this value. | |
4084 | */ | |
4085 | static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
ec12cb7f | 4086 | { |
a9cf55b2 | 4087 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); |
a9cf55b2 PT |
4088 | |
4089 | /* if the deadline is ahead of our clock, nothing to do */ | |
78becc27 | 4090 | if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) |
ec12cb7f PT |
4091 | return; |
4092 | ||
a9cf55b2 PT |
4093 | if (cfs_rq->runtime_remaining < 0) |
4094 | return; | |
4095 | ||
4096 | /* | |
4097 | * If the local deadline has passed we have to consider the | |
4098 | * possibility that our sched_clock is 'fast' and the global deadline | |
4099 | * has not truly expired. | |
4100 | * | |
4101 | * Fortunately we can check determine whether this the case by checking | |
51f2176d BS |
4102 | * whether the global deadline has advanced. It is valid to compare |
4103 | * cfs_b->runtime_expires without any locks since we only care about | |
4104 | * exact equality, so a partial write will still work. | |
a9cf55b2 PT |
4105 | */ |
4106 | ||
51f2176d | 4107 | if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { |
a9cf55b2 PT |
4108 | /* extend local deadline, drift is bounded above by 2 ticks */ |
4109 | cfs_rq->runtime_expires += TICK_NSEC; | |
4110 | } else { | |
4111 | /* global deadline is ahead, expiration has passed */ | |
4112 | cfs_rq->runtime_remaining = 0; | |
4113 | } | |
4114 | } | |
4115 | ||
9dbdb155 | 4116 | static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
a9cf55b2 PT |
4117 | { |
4118 | /* dock delta_exec before expiring quota (as it could span periods) */ | |
ec12cb7f | 4119 | cfs_rq->runtime_remaining -= delta_exec; |
a9cf55b2 PT |
4120 | expire_cfs_rq_runtime(cfs_rq); |
4121 | ||
4122 | if (likely(cfs_rq->runtime_remaining > 0)) | |
ec12cb7f PT |
4123 | return; |
4124 | ||
85dac906 PT |
4125 | /* |
4126 | * if we're unable to extend our runtime we resched so that the active | |
4127 | * hierarchy can be throttled | |
4128 | */ | |
4129 | if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) | |
8875125e | 4130 | resched_curr(rq_of(cfs_rq)); |
ec12cb7f PT |
4131 | } |
4132 | ||
6c16a6dc | 4133 | static __always_inline |
9dbdb155 | 4134 | void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) |
ec12cb7f | 4135 | { |
56f570e5 | 4136 | if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) |
ec12cb7f PT |
4137 | return; |
4138 | ||
4139 | __account_cfs_rq_runtime(cfs_rq, delta_exec); | |
4140 | } | |
4141 | ||
85dac906 PT |
4142 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) |
4143 | { | |
56f570e5 | 4144 | return cfs_bandwidth_used() && cfs_rq->throttled; |
85dac906 PT |
4145 | } |
4146 | ||
64660c86 PT |
4147 | /* check whether cfs_rq, or any parent, is throttled */ |
4148 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4149 | { | |
56f570e5 | 4150 | return cfs_bandwidth_used() && cfs_rq->throttle_count; |
64660c86 PT |
4151 | } |
4152 | ||
4153 | /* | |
4154 | * Ensure that neither of the group entities corresponding to src_cpu or | |
4155 | * dest_cpu are members of a throttled hierarchy when performing group | |
4156 | * load-balance operations. | |
4157 | */ | |
4158 | static inline int throttled_lb_pair(struct task_group *tg, | |
4159 | int src_cpu, int dest_cpu) | |
4160 | { | |
4161 | struct cfs_rq *src_cfs_rq, *dest_cfs_rq; | |
4162 | ||
4163 | src_cfs_rq = tg->cfs_rq[src_cpu]; | |
4164 | dest_cfs_rq = tg->cfs_rq[dest_cpu]; | |
4165 | ||
4166 | return throttled_hierarchy(src_cfs_rq) || | |
4167 | throttled_hierarchy(dest_cfs_rq); | |
4168 | } | |
4169 | ||
4170 | /* updated child weight may affect parent so we have to do this bottom up */ | |
4171 | static int tg_unthrottle_up(struct task_group *tg, void *data) | |
4172 | { | |
4173 | struct rq *rq = data; | |
4174 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4175 | ||
4176 | cfs_rq->throttle_count--; | |
64660c86 | 4177 | if (!cfs_rq->throttle_count) { |
f1b17280 | 4178 | /* adjust cfs_rq_clock_task() */ |
78becc27 | 4179 | cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - |
f1b17280 | 4180 | cfs_rq->throttled_clock_task; |
64660c86 | 4181 | } |
64660c86 PT |
4182 | |
4183 | return 0; | |
4184 | } | |
4185 | ||
4186 | static int tg_throttle_down(struct task_group *tg, void *data) | |
4187 | { | |
4188 | struct rq *rq = data; | |
4189 | struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; | |
4190 | ||
82958366 PT |
4191 | /* group is entering throttled state, stop time */ |
4192 | if (!cfs_rq->throttle_count) | |
78becc27 | 4193 | cfs_rq->throttled_clock_task = rq_clock_task(rq); |
64660c86 PT |
4194 | cfs_rq->throttle_count++; |
4195 | ||
4196 | return 0; | |
4197 | } | |
4198 | ||
d3d9dc33 | 4199 | static void throttle_cfs_rq(struct cfs_rq *cfs_rq) |
85dac906 PT |
4200 | { |
4201 | struct rq *rq = rq_of(cfs_rq); | |
4202 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4203 | struct sched_entity *se; | |
4204 | long task_delta, dequeue = 1; | |
77a4d1a1 | 4205 | bool empty; |
85dac906 PT |
4206 | |
4207 | se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; | |
4208 | ||
f1b17280 | 4209 | /* freeze hierarchy runnable averages while throttled */ |
64660c86 PT |
4210 | rcu_read_lock(); |
4211 | walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); | |
4212 | rcu_read_unlock(); | |
85dac906 PT |
4213 | |
4214 | task_delta = cfs_rq->h_nr_running; | |
4215 | for_each_sched_entity(se) { | |
4216 | struct cfs_rq *qcfs_rq = cfs_rq_of(se); | |
4217 | /* throttled entity or throttle-on-deactivate */ | |
4218 | if (!se->on_rq) | |
4219 | break; | |
4220 | ||
4221 | if (dequeue) | |
4222 | dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); | |
4223 | qcfs_rq->h_nr_running -= task_delta; | |
4224 | ||
4225 | if (qcfs_rq->load.weight) | |
4226 | dequeue = 0; | |
4227 | } | |
4228 | ||
4229 | if (!se) | |
72465447 | 4230 | sub_nr_running(rq, task_delta); |
85dac906 PT |
4231 | |
4232 | cfs_rq->throttled = 1; | |
78becc27 | 4233 | cfs_rq->throttled_clock = rq_clock(rq); |
85dac906 | 4234 | raw_spin_lock(&cfs_b->lock); |
d49db342 | 4235 | empty = list_empty(&cfs_b->throttled_cfs_rq); |
77a4d1a1 | 4236 | |
c06f04c7 BS |
4237 | /* |
4238 | * Add to the _head_ of the list, so that an already-started | |
4239 | * distribute_cfs_runtime will not see us | |
4240 | */ | |
4241 | list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); | |
77a4d1a1 PZ |
4242 | |
4243 | /* | |
4244 | * If we're the first throttled task, make sure the bandwidth | |
4245 | * timer is running. | |
4246 | */ | |
4247 | if (empty) | |
4248 | start_cfs_bandwidth(cfs_b); | |
4249 | ||
85dac906 PT |
4250 | raw_spin_unlock(&cfs_b->lock); |
4251 | } | |
4252 | ||
029632fb | 4253 | void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) |
671fd9da PT |
4254 | { |
4255 | struct rq *rq = rq_of(cfs_rq); | |
4256 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4257 | struct sched_entity *se; | |
4258 | int enqueue = 1; | |
4259 | long task_delta; | |
4260 | ||
22b958d8 | 4261 | se = cfs_rq->tg->se[cpu_of(rq)]; |
671fd9da PT |
4262 | |
4263 | cfs_rq->throttled = 0; | |
1a55af2e FW |
4264 | |
4265 | update_rq_clock(rq); | |
4266 | ||
671fd9da | 4267 | raw_spin_lock(&cfs_b->lock); |
78becc27 | 4268 | cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; |
671fd9da PT |
4269 | list_del_rcu(&cfs_rq->throttled_list); |
4270 | raw_spin_unlock(&cfs_b->lock); | |
4271 | ||
64660c86 PT |
4272 | /* update hierarchical throttle state */ |
4273 | walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); | |
4274 | ||
671fd9da PT |
4275 | if (!cfs_rq->load.weight) |
4276 | return; | |
4277 | ||
4278 | task_delta = cfs_rq->h_nr_running; | |
4279 | for_each_sched_entity(se) { | |
4280 | if (se->on_rq) | |
4281 | enqueue = 0; | |
4282 | ||
4283 | cfs_rq = cfs_rq_of(se); | |
4284 | if (enqueue) | |
4285 | enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); | |
4286 | cfs_rq->h_nr_running += task_delta; | |
4287 | ||
4288 | if (cfs_rq_throttled(cfs_rq)) | |
4289 | break; | |
4290 | } | |
4291 | ||
4292 | if (!se) | |
72465447 | 4293 | add_nr_running(rq, task_delta); |
671fd9da PT |
4294 | |
4295 | /* determine whether we need to wake up potentially idle cpu */ | |
4296 | if (rq->curr == rq->idle && rq->cfs.nr_running) | |
8875125e | 4297 | resched_curr(rq); |
671fd9da PT |
4298 | } |
4299 | ||
4300 | static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, | |
4301 | u64 remaining, u64 expires) | |
4302 | { | |
4303 | struct cfs_rq *cfs_rq; | |
c06f04c7 BS |
4304 | u64 runtime; |
4305 | u64 starting_runtime = remaining; | |
671fd9da PT |
4306 | |
4307 | rcu_read_lock(); | |
4308 | list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, | |
4309 | throttled_list) { | |
4310 | struct rq *rq = rq_of(cfs_rq); | |
8a8c69c3 | 4311 | struct rq_flags rf; |
671fd9da | 4312 | |
8a8c69c3 | 4313 | rq_lock(rq, &rf); |
671fd9da PT |
4314 | if (!cfs_rq_throttled(cfs_rq)) |
4315 | goto next; | |
4316 | ||
4317 | runtime = -cfs_rq->runtime_remaining + 1; | |
4318 | if (runtime > remaining) | |
4319 | runtime = remaining; | |
4320 | remaining -= runtime; | |
4321 | ||
4322 | cfs_rq->runtime_remaining += runtime; | |
4323 | cfs_rq->runtime_expires = expires; | |
4324 | ||
4325 | /* we check whether we're throttled above */ | |
4326 | if (cfs_rq->runtime_remaining > 0) | |
4327 | unthrottle_cfs_rq(cfs_rq); | |
4328 | ||
4329 | next: | |
8a8c69c3 | 4330 | rq_unlock(rq, &rf); |
671fd9da PT |
4331 | |
4332 | if (!remaining) | |
4333 | break; | |
4334 | } | |
4335 | rcu_read_unlock(); | |
4336 | ||
c06f04c7 | 4337 | return starting_runtime - remaining; |
671fd9da PT |
4338 | } |
4339 | ||
58088ad0 PT |
4340 | /* |
4341 | * Responsible for refilling a task_group's bandwidth and unthrottling its | |
4342 | * cfs_rqs as appropriate. If there has been no activity within the last | |
4343 | * period the timer is deactivated until scheduling resumes; cfs_b->idle is | |
4344 | * used to track this state. | |
4345 | */ | |
4346 | static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) | |
4347 | { | |
671fd9da | 4348 | u64 runtime, runtime_expires; |
51f2176d | 4349 | int throttled; |
58088ad0 | 4350 | |
58088ad0 PT |
4351 | /* no need to continue the timer with no bandwidth constraint */ |
4352 | if (cfs_b->quota == RUNTIME_INF) | |
51f2176d | 4353 | goto out_deactivate; |
58088ad0 | 4354 | |
671fd9da | 4355 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); |
e8da1b18 | 4356 | cfs_b->nr_periods += overrun; |
671fd9da | 4357 | |
51f2176d BS |
4358 | /* |
4359 | * idle depends on !throttled (for the case of a large deficit), and if | |
4360 | * we're going inactive then everything else can be deferred | |
4361 | */ | |
4362 | if (cfs_b->idle && !throttled) | |
4363 | goto out_deactivate; | |
a9cf55b2 PT |
4364 | |
4365 | __refill_cfs_bandwidth_runtime(cfs_b); | |
4366 | ||
671fd9da PT |
4367 | if (!throttled) { |
4368 | /* mark as potentially idle for the upcoming period */ | |
4369 | cfs_b->idle = 1; | |
51f2176d | 4370 | return 0; |
671fd9da PT |
4371 | } |
4372 | ||
e8da1b18 NR |
4373 | /* account preceding periods in which throttling occurred */ |
4374 | cfs_b->nr_throttled += overrun; | |
4375 | ||
671fd9da | 4376 | runtime_expires = cfs_b->runtime_expires; |
671fd9da PT |
4377 | |
4378 | /* | |
c06f04c7 BS |
4379 | * This check is repeated as we are holding onto the new bandwidth while |
4380 | * we unthrottle. This can potentially race with an unthrottled group | |
4381 | * trying to acquire new bandwidth from the global pool. This can result | |
4382 | * in us over-using our runtime if it is all used during this loop, but | |
4383 | * only by limited amounts in that extreme case. | |
671fd9da | 4384 | */ |
c06f04c7 BS |
4385 | while (throttled && cfs_b->runtime > 0) { |
4386 | runtime = cfs_b->runtime; | |
671fd9da PT |
4387 | raw_spin_unlock(&cfs_b->lock); |
4388 | /* we can't nest cfs_b->lock while distributing bandwidth */ | |
4389 | runtime = distribute_cfs_runtime(cfs_b, runtime, | |
4390 | runtime_expires); | |
4391 | raw_spin_lock(&cfs_b->lock); | |
4392 | ||
4393 | throttled = !list_empty(&cfs_b->throttled_cfs_rq); | |
c06f04c7 BS |
4394 | |
4395 | cfs_b->runtime -= min(runtime, cfs_b->runtime); | |
671fd9da | 4396 | } |
58088ad0 | 4397 | |
671fd9da PT |
4398 | /* |
4399 | * While we are ensured activity in the period following an | |
4400 | * unthrottle, this also covers the case in which the new bandwidth is | |
4401 | * insufficient to cover the existing bandwidth deficit. (Forcing the | |
4402 | * timer to remain active while there are any throttled entities.) | |
4403 | */ | |
4404 | cfs_b->idle = 0; | |
58088ad0 | 4405 | |
51f2176d BS |
4406 | return 0; |
4407 | ||
4408 | out_deactivate: | |
51f2176d | 4409 | return 1; |
58088ad0 | 4410 | } |
d3d9dc33 | 4411 | |
d8b4986d PT |
4412 | /* a cfs_rq won't donate quota below this amount */ |
4413 | static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; | |
4414 | /* minimum remaining period time to redistribute slack quota */ | |
4415 | static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; | |
4416 | /* how long we wait to gather additional slack before distributing */ | |
4417 | static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; | |
4418 | ||
db06e78c BS |
4419 | /* |
4420 | * Are we near the end of the current quota period? | |
4421 | * | |
4422 | * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the | |
4961b6e1 | 4423 | * hrtimer base being cleared by hrtimer_start. In the case of |
db06e78c BS |
4424 | * migrate_hrtimers, base is never cleared, so we are fine. |
4425 | */ | |
d8b4986d PT |
4426 | static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) |
4427 | { | |
4428 | struct hrtimer *refresh_timer = &cfs_b->period_timer; | |
4429 | u64 remaining; | |
4430 | ||
4431 | /* if the call-back is running a quota refresh is already occurring */ | |
4432 | if (hrtimer_callback_running(refresh_timer)) | |
4433 | return 1; | |
4434 | ||
4435 | /* is a quota refresh about to occur? */ | |
4436 | remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); | |
4437 | if (remaining < min_expire) | |
4438 | return 1; | |
4439 | ||
4440 | return 0; | |
4441 | } | |
4442 | ||
4443 | static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) | |
4444 | { | |
4445 | u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; | |
4446 | ||
4447 | /* if there's a quota refresh soon don't bother with slack */ | |
4448 | if (runtime_refresh_within(cfs_b, min_left)) | |
4449 | return; | |
4450 | ||
4cfafd30 PZ |
4451 | hrtimer_start(&cfs_b->slack_timer, |
4452 | ns_to_ktime(cfs_bandwidth_slack_period), | |
4453 | HRTIMER_MODE_REL); | |
d8b4986d PT |
4454 | } |
4455 | ||
4456 | /* we know any runtime found here is valid as update_curr() precedes return */ | |
4457 | static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4458 | { | |
4459 | struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); | |
4460 | s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; | |
4461 | ||
4462 | if (slack_runtime <= 0) | |
4463 | return; | |
4464 | ||
4465 | raw_spin_lock(&cfs_b->lock); | |
4466 | if (cfs_b->quota != RUNTIME_INF && | |
4467 | cfs_rq->runtime_expires == cfs_b->runtime_expires) { | |
4468 | cfs_b->runtime += slack_runtime; | |
4469 | ||
4470 | /* we are under rq->lock, defer unthrottling using a timer */ | |
4471 | if (cfs_b->runtime > sched_cfs_bandwidth_slice() && | |
4472 | !list_empty(&cfs_b->throttled_cfs_rq)) | |
4473 | start_cfs_slack_bandwidth(cfs_b); | |
4474 | } | |
4475 | raw_spin_unlock(&cfs_b->lock); | |
4476 | ||
4477 | /* even if it's not valid for return we don't want to try again */ | |
4478 | cfs_rq->runtime_remaining -= slack_runtime; | |
4479 | } | |
4480 | ||
4481 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4482 | { | |
56f570e5 PT |
4483 | if (!cfs_bandwidth_used()) |
4484 | return; | |
4485 | ||
fccfdc6f | 4486 | if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) |
d8b4986d PT |
4487 | return; |
4488 | ||
4489 | __return_cfs_rq_runtime(cfs_rq); | |
4490 | } | |
4491 | ||
4492 | /* | |
4493 | * This is done with a timer (instead of inline with bandwidth return) since | |
4494 | * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. | |
4495 | */ | |
4496 | static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) | |
4497 | { | |
4498 | u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); | |
4499 | u64 expires; | |
4500 | ||
4501 | /* confirm we're still not at a refresh boundary */ | |
db06e78c BS |
4502 | raw_spin_lock(&cfs_b->lock); |
4503 | if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { | |
4504 | raw_spin_unlock(&cfs_b->lock); | |
d8b4986d | 4505 | return; |
db06e78c | 4506 | } |
d8b4986d | 4507 | |
c06f04c7 | 4508 | if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) |
d8b4986d | 4509 | runtime = cfs_b->runtime; |
c06f04c7 | 4510 | |
d8b4986d PT |
4511 | expires = cfs_b->runtime_expires; |
4512 | raw_spin_unlock(&cfs_b->lock); | |
4513 | ||
4514 | if (!runtime) | |
4515 | return; | |
4516 | ||
4517 | runtime = distribute_cfs_runtime(cfs_b, runtime, expires); | |
4518 | ||
4519 | raw_spin_lock(&cfs_b->lock); | |
4520 | if (expires == cfs_b->runtime_expires) | |
c06f04c7 | 4521 | cfs_b->runtime -= min(runtime, cfs_b->runtime); |
d8b4986d PT |
4522 | raw_spin_unlock(&cfs_b->lock); |
4523 | } | |
4524 | ||
d3d9dc33 PT |
4525 | /* |
4526 | * When a group wakes up we want to make sure that its quota is not already | |
4527 | * expired/exceeded, otherwise it may be allowed to steal additional ticks of | |
4528 | * runtime as update_curr() throttling can not not trigger until it's on-rq. | |
4529 | */ | |
4530 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) | |
4531 | { | |
56f570e5 PT |
4532 | if (!cfs_bandwidth_used()) |
4533 | return; | |
4534 | ||
d3d9dc33 PT |
4535 | /* an active group must be handled by the update_curr()->put() path */ |
4536 | if (!cfs_rq->runtime_enabled || cfs_rq->curr) | |
4537 | return; | |
4538 | ||
4539 | /* ensure the group is not already throttled */ | |
4540 | if (cfs_rq_throttled(cfs_rq)) | |
4541 | return; | |
4542 | ||
4543 | /* update runtime allocation */ | |
4544 | account_cfs_rq_runtime(cfs_rq, 0); | |
4545 | if (cfs_rq->runtime_remaining <= 0) | |
4546 | throttle_cfs_rq(cfs_rq); | |
4547 | } | |
4548 | ||
55e16d30 PZ |
4549 | static void sync_throttle(struct task_group *tg, int cpu) |
4550 | { | |
4551 | struct cfs_rq *pcfs_rq, *cfs_rq; | |
4552 | ||
4553 | if (!cfs_bandwidth_used()) | |
4554 | return; | |
4555 | ||
4556 | if (!tg->parent) | |
4557 | return; | |
4558 | ||
4559 | cfs_rq = tg->cfs_rq[cpu]; | |
4560 | pcfs_rq = tg->parent->cfs_rq[cpu]; | |
4561 | ||
4562 | cfs_rq->throttle_count = pcfs_rq->throttle_count; | |
b8922125 | 4563 | cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); |
55e16d30 PZ |
4564 | } |
4565 | ||
d3d9dc33 | 4566 | /* conditionally throttle active cfs_rq's from put_prev_entity() */ |
678d5718 | 4567 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) |
d3d9dc33 | 4568 | { |
56f570e5 | 4569 | if (!cfs_bandwidth_used()) |
678d5718 | 4570 | return false; |
56f570e5 | 4571 | |
d3d9dc33 | 4572 | if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) |
678d5718 | 4573 | return false; |
d3d9dc33 PT |
4574 | |
4575 | /* | |
4576 | * it's possible for a throttled entity to be forced into a running | |
4577 | * state (e.g. set_curr_task), in this case we're finished. | |
4578 | */ | |
4579 | if (cfs_rq_throttled(cfs_rq)) | |
678d5718 | 4580 | return true; |
d3d9dc33 PT |
4581 | |
4582 | throttle_cfs_rq(cfs_rq); | |
678d5718 | 4583 | return true; |
d3d9dc33 | 4584 | } |
029632fb | 4585 | |
029632fb PZ |
4586 | static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) |
4587 | { | |
4588 | struct cfs_bandwidth *cfs_b = | |
4589 | container_of(timer, struct cfs_bandwidth, slack_timer); | |
77a4d1a1 | 4590 | |
029632fb PZ |
4591 | do_sched_cfs_slack_timer(cfs_b); |
4592 | ||
4593 | return HRTIMER_NORESTART; | |
4594 | } | |
4595 | ||
4596 | static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) | |
4597 | { | |
4598 | struct cfs_bandwidth *cfs_b = | |
4599 | container_of(timer, struct cfs_bandwidth, period_timer); | |
029632fb PZ |
4600 | int overrun; |
4601 | int idle = 0; | |
4602 | ||
51f2176d | 4603 | raw_spin_lock(&cfs_b->lock); |
029632fb | 4604 | for (;;) { |
77a4d1a1 | 4605 | overrun = hrtimer_forward_now(timer, cfs_b->period); |
029632fb PZ |
4606 | if (!overrun) |
4607 | break; | |
4608 | ||
4609 | idle = do_sched_cfs_period_timer(cfs_b, overrun); | |
4610 | } | |
4cfafd30 PZ |
4611 | if (idle) |
4612 | cfs_b->period_active = 0; | |
51f2176d | 4613 | raw_spin_unlock(&cfs_b->lock); |
029632fb PZ |
4614 | |
4615 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
4616 | } | |
4617 | ||
4618 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4619 | { | |
4620 | raw_spin_lock_init(&cfs_b->lock); | |
4621 | cfs_b->runtime = 0; | |
4622 | cfs_b->quota = RUNTIME_INF; | |
4623 | cfs_b->period = ns_to_ktime(default_cfs_period()); | |
4624 | ||
4625 | INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); | |
4cfafd30 | 4626 | hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); |
029632fb PZ |
4627 | cfs_b->period_timer.function = sched_cfs_period_timer; |
4628 | hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
4629 | cfs_b->slack_timer.function = sched_cfs_slack_timer; | |
4630 | } | |
4631 | ||
4632 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) | |
4633 | { | |
4634 | cfs_rq->runtime_enabled = 0; | |
4635 | INIT_LIST_HEAD(&cfs_rq->throttled_list); | |
4636 | } | |
4637 | ||
77a4d1a1 | 4638 | void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) |
029632fb | 4639 | { |
4cfafd30 | 4640 | lockdep_assert_held(&cfs_b->lock); |
029632fb | 4641 | |
4cfafd30 PZ |
4642 | if (!cfs_b->period_active) { |
4643 | cfs_b->period_active = 1; | |
4644 | hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); | |
4645 | hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); | |
4646 | } | |
029632fb PZ |
4647 | } |
4648 | ||
4649 | static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) | |
4650 | { | |
7f1a169b TH |
4651 | /* init_cfs_bandwidth() was not called */ |
4652 | if (!cfs_b->throttled_cfs_rq.next) | |
4653 | return; | |
4654 | ||
029632fb PZ |
4655 | hrtimer_cancel(&cfs_b->period_timer); |
4656 | hrtimer_cancel(&cfs_b->slack_timer); | |
4657 | } | |
4658 | ||
0e59bdae KT |
4659 | static void __maybe_unused update_runtime_enabled(struct rq *rq) |
4660 | { | |
4661 | struct cfs_rq *cfs_rq; | |
4662 | ||
4663 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
4664 | struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth; | |
4665 | ||
4666 | raw_spin_lock(&cfs_b->lock); | |
4667 | cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; | |
4668 | raw_spin_unlock(&cfs_b->lock); | |
4669 | } | |
4670 | } | |
4671 | ||
38dc3348 | 4672 | static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) |
029632fb PZ |
4673 | { |
4674 | struct cfs_rq *cfs_rq; | |
4675 | ||
4676 | for_each_leaf_cfs_rq(rq, cfs_rq) { | |
029632fb PZ |
4677 | if (!cfs_rq->runtime_enabled) |
4678 | continue; | |
4679 | ||
4680 | /* | |
4681 | * clock_task is not advancing so we just need to make sure | |
4682 | * there's some valid quota amount | |
4683 | */ | |
51f2176d | 4684 | cfs_rq->runtime_remaining = 1; |
0e59bdae KT |
4685 | /* |
4686 | * Offline rq is schedulable till cpu is completely disabled | |
4687 | * in take_cpu_down(), so we prevent new cfs throttling here. | |
4688 | */ | |
4689 | cfs_rq->runtime_enabled = 0; | |
4690 | ||
029632fb PZ |
4691 | if (cfs_rq_throttled(cfs_rq)) |
4692 | unthrottle_cfs_rq(cfs_rq); | |
4693 | } | |
4694 | } | |
4695 | ||
4696 | #else /* CONFIG_CFS_BANDWIDTH */ | |
f1b17280 PT |
4697 | static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) |
4698 | { | |
78becc27 | 4699 | return rq_clock_task(rq_of(cfs_rq)); |
f1b17280 PT |
4700 | } |
4701 | ||
9dbdb155 | 4702 | static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} |
678d5718 | 4703 | static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } |
d3d9dc33 | 4704 | static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} |
55e16d30 | 4705 | static inline void sync_throttle(struct task_group *tg, int cpu) {} |
6c16a6dc | 4706 | static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} |
85dac906 PT |
4707 | |
4708 | static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) | |
4709 | { | |
4710 | return 0; | |
4711 | } | |
64660c86 PT |
4712 | |
4713 | static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) | |
4714 | { | |
4715 | return 0; | |
4716 | } | |
4717 | ||
4718 | static inline int throttled_lb_pair(struct task_group *tg, | |
4719 | int src_cpu, int dest_cpu) | |
4720 | { | |
4721 | return 0; | |
4722 | } | |
029632fb PZ |
4723 | |
4724 | void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
4725 | ||
4726 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
4727 | static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} | |
ab84d31e PT |
4728 | #endif |
4729 | ||
029632fb PZ |
4730 | static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) |
4731 | { | |
4732 | return NULL; | |
4733 | } | |
4734 | static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} | |
0e59bdae | 4735 | static inline void update_runtime_enabled(struct rq *rq) {} |
a4c96ae3 | 4736 | static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} |
029632fb PZ |
4737 | |
4738 | #endif /* CONFIG_CFS_BANDWIDTH */ | |
4739 | ||
bf0f6f24 IM |
4740 | /************************************************** |
4741 | * CFS operations on tasks: | |
4742 | */ | |
4743 | ||
8f4d37ec PZ |
4744 | #ifdef CONFIG_SCHED_HRTICK |
4745 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
4746 | { | |
8f4d37ec PZ |
4747 | struct sched_entity *se = &p->se; |
4748 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
4749 | ||
9148a3a1 | 4750 | SCHED_WARN_ON(task_rq(p) != rq); |
8f4d37ec | 4751 | |
8bf46a39 | 4752 | if (rq->cfs.h_nr_running > 1) { |
8f4d37ec PZ |
4753 | u64 slice = sched_slice(cfs_rq, se); |
4754 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; | |
4755 | s64 delta = slice - ran; | |
4756 | ||
4757 | if (delta < 0) { | |
4758 | if (rq->curr == p) | |
8875125e | 4759 | resched_curr(rq); |
8f4d37ec PZ |
4760 | return; |
4761 | } | |
31656519 | 4762 | hrtick_start(rq, delta); |
8f4d37ec PZ |
4763 | } |
4764 | } | |
a4c2f00f PZ |
4765 | |
4766 | /* | |
4767 | * called from enqueue/dequeue and updates the hrtick when the | |
4768 | * current task is from our class and nr_running is low enough | |
4769 | * to matter. | |
4770 | */ | |
4771 | static void hrtick_update(struct rq *rq) | |
4772 | { | |
4773 | struct task_struct *curr = rq->curr; | |
4774 | ||
b39e66ea | 4775 | if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) |
a4c2f00f PZ |
4776 | return; |
4777 | ||
4778 | if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) | |
4779 | hrtick_start_fair(rq, curr); | |
4780 | } | |
55e12e5e | 4781 | #else /* !CONFIG_SCHED_HRTICK */ |
8f4d37ec PZ |
4782 | static inline void |
4783 | hrtick_start_fair(struct rq *rq, struct task_struct *p) | |
4784 | { | |
4785 | } | |
a4c2f00f PZ |
4786 | |
4787 | static inline void hrtick_update(struct rq *rq) | |
4788 | { | |
4789 | } | |
8f4d37ec PZ |
4790 | #endif |
4791 | ||
bf0f6f24 IM |
4792 | /* |
4793 | * The enqueue_task method is called before nr_running is | |
4794 | * increased. Here we update the fair scheduling stats and | |
4795 | * then put the task into the rbtree: | |
4796 | */ | |
ea87bb78 | 4797 | static void |
371fd7e7 | 4798 | enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
4799 | { |
4800 | struct cfs_rq *cfs_rq; | |
62fb1851 | 4801 | struct sched_entity *se = &p->se; |
bf0f6f24 | 4802 | |
8c34ab19 RW |
4803 | /* |
4804 | * If in_iowait is set, the code below may not trigger any cpufreq | |
4805 | * utilization updates, so do it here explicitly with the IOWAIT flag | |
4806 | * passed. | |
4807 | */ | |
4808 | if (p->in_iowait) | |
4809 | cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT); | |
4810 | ||
bf0f6f24 | 4811 | for_each_sched_entity(se) { |
62fb1851 | 4812 | if (se->on_rq) |
bf0f6f24 IM |
4813 | break; |
4814 | cfs_rq = cfs_rq_of(se); | |
88ec22d3 | 4815 | enqueue_entity(cfs_rq, se, flags); |
85dac906 PT |
4816 | |
4817 | /* | |
4818 | * end evaluation on encountering a throttled cfs_rq | |
4819 | * | |
4820 | * note: in the case of encountering a throttled cfs_rq we will | |
4821 | * post the final h_nr_running increment below. | |
e210bffd | 4822 | */ |
85dac906 PT |
4823 | if (cfs_rq_throttled(cfs_rq)) |
4824 | break; | |
953bfcd1 | 4825 | cfs_rq->h_nr_running++; |
85dac906 | 4826 | |
88ec22d3 | 4827 | flags = ENQUEUE_WAKEUP; |
bf0f6f24 | 4828 | } |
8f4d37ec | 4829 | |
2069dd75 | 4830 | for_each_sched_entity(se) { |
0f317143 | 4831 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 4832 | cfs_rq->h_nr_running++; |
2069dd75 | 4833 | |
85dac906 PT |
4834 | if (cfs_rq_throttled(cfs_rq)) |
4835 | break; | |
4836 | ||
d31b1a66 | 4837 | update_load_avg(se, UPDATE_TG); |
89ee048f | 4838 | update_cfs_shares(se); |
2069dd75 PZ |
4839 | } |
4840 | ||
cd126afe | 4841 | if (!se) |
72465447 | 4842 | add_nr_running(rq, 1); |
cd126afe | 4843 | |
a4c2f00f | 4844 | hrtick_update(rq); |
bf0f6f24 IM |
4845 | } |
4846 | ||
2f36825b VP |
4847 | static void set_next_buddy(struct sched_entity *se); |
4848 | ||
bf0f6f24 IM |
4849 | /* |
4850 | * The dequeue_task method is called before nr_running is | |
4851 | * decreased. We remove the task from the rbtree and | |
4852 | * update the fair scheduling stats: | |
4853 | */ | |
371fd7e7 | 4854 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) |
bf0f6f24 IM |
4855 | { |
4856 | struct cfs_rq *cfs_rq; | |
62fb1851 | 4857 | struct sched_entity *se = &p->se; |
2f36825b | 4858 | int task_sleep = flags & DEQUEUE_SLEEP; |
bf0f6f24 IM |
4859 | |
4860 | for_each_sched_entity(se) { | |
4861 | cfs_rq = cfs_rq_of(se); | |
371fd7e7 | 4862 | dequeue_entity(cfs_rq, se, flags); |
85dac906 PT |
4863 | |
4864 | /* | |
4865 | * end evaluation on encountering a throttled cfs_rq | |
4866 | * | |
4867 | * note: in the case of encountering a throttled cfs_rq we will | |
4868 | * post the final h_nr_running decrement below. | |
4869 | */ | |
4870 | if (cfs_rq_throttled(cfs_rq)) | |
4871 | break; | |
953bfcd1 | 4872 | cfs_rq->h_nr_running--; |
2069dd75 | 4873 | |
bf0f6f24 | 4874 | /* Don't dequeue parent if it has other entities besides us */ |
2f36825b | 4875 | if (cfs_rq->load.weight) { |
754bd598 KK |
4876 | /* Avoid re-evaluating load for this entity: */ |
4877 | se = parent_entity(se); | |
2f36825b VP |
4878 | /* |
4879 | * Bias pick_next to pick a task from this cfs_rq, as | |
4880 | * p is sleeping when it is within its sched_slice. | |
4881 | */ | |
754bd598 KK |
4882 | if (task_sleep && se && !throttled_hierarchy(cfs_rq)) |
4883 | set_next_buddy(se); | |
bf0f6f24 | 4884 | break; |
2f36825b | 4885 | } |
371fd7e7 | 4886 | flags |= DEQUEUE_SLEEP; |
bf0f6f24 | 4887 | } |
8f4d37ec | 4888 | |
2069dd75 | 4889 | for_each_sched_entity(se) { |
0f317143 | 4890 | cfs_rq = cfs_rq_of(se); |
953bfcd1 | 4891 | cfs_rq->h_nr_running--; |
2069dd75 | 4892 | |
85dac906 PT |
4893 | if (cfs_rq_throttled(cfs_rq)) |
4894 | break; | |
4895 | ||
d31b1a66 | 4896 | update_load_avg(se, UPDATE_TG); |
89ee048f | 4897 | update_cfs_shares(se); |
2069dd75 PZ |
4898 | } |
4899 | ||
cd126afe | 4900 | if (!se) |
72465447 | 4901 | sub_nr_running(rq, 1); |
cd126afe | 4902 | |
a4c2f00f | 4903 | hrtick_update(rq); |
bf0f6f24 IM |
4904 | } |
4905 | ||
e7693a36 | 4906 | #ifdef CONFIG_SMP |
10e2f1ac PZ |
4907 | |
4908 | /* Working cpumask for: load_balance, load_balance_newidle. */ | |
4909 | DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); | |
4910 | DEFINE_PER_CPU(cpumask_var_t, select_idle_mask); | |
4911 | ||
9fd81dd5 | 4912 | #ifdef CONFIG_NO_HZ_COMMON |
3289bdb4 PZ |
4913 | /* |
4914 | * per rq 'load' arrray crap; XXX kill this. | |
4915 | */ | |
4916 | ||
4917 | /* | |
d937cdc5 | 4918 | * The exact cpuload calculated at every tick would be: |
3289bdb4 | 4919 | * |
d937cdc5 PZ |
4920 | * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load |
4921 | * | |
4922 | * If a cpu misses updates for n ticks (as it was idle) and update gets | |
4923 | * called on the n+1-th tick when cpu may be busy, then we have: | |
4924 | * | |
4925 | * load_n = (1 - 1/2^i)^n * load_0 | |
4926 | * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load | |
3289bdb4 PZ |
4927 | * |
4928 | * decay_load_missed() below does efficient calculation of | |
3289bdb4 | 4929 | * |
d937cdc5 PZ |
4930 | * load' = (1 - 1/2^i)^n * load |
4931 | * | |
4932 | * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors. | |
4933 | * This allows us to precompute the above in said factors, thereby allowing the | |
4934 | * reduction of an arbitrary n in O(log_2 n) steps. (See also | |
4935 | * fixed_power_int()) | |
3289bdb4 | 4936 | * |
d937cdc5 | 4937 | * The calculation is approximated on a 128 point scale. |
3289bdb4 PZ |
4938 | */ |
4939 | #define DEGRADE_SHIFT 7 | |
d937cdc5 PZ |
4940 | |
4941 | static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; | |
4942 | static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { | |
4943 | { 0, 0, 0, 0, 0, 0, 0, 0 }, | |
4944 | { 64, 32, 8, 0, 0, 0, 0, 0 }, | |
4945 | { 96, 72, 40, 12, 1, 0, 0, 0 }, | |
4946 | { 112, 98, 75, 43, 15, 1, 0, 0 }, | |
4947 | { 120, 112, 98, 76, 45, 16, 2, 0 } | |
4948 | }; | |
3289bdb4 PZ |
4949 | |
4950 | /* | |
4951 | * Update cpu_load for any missed ticks, due to tickless idle. The backlog | |
4952 | * would be when CPU is idle and so we just decay the old load without | |
4953 | * adding any new load. | |
4954 | */ | |
4955 | static unsigned long | |
4956 | decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) | |
4957 | { | |
4958 | int j = 0; | |
4959 | ||
4960 | if (!missed_updates) | |
4961 | return load; | |
4962 | ||
4963 | if (missed_updates >= degrade_zero_ticks[idx]) | |
4964 | return 0; | |
4965 | ||
4966 | if (idx == 1) | |
4967 | return load >> missed_updates; | |
4968 | ||
4969 | while (missed_updates) { | |
4970 | if (missed_updates % 2) | |
4971 | load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; | |
4972 | ||
4973 | missed_updates >>= 1; | |
4974 | j++; | |
4975 | } | |
4976 | return load; | |
4977 | } | |
9fd81dd5 | 4978 | #endif /* CONFIG_NO_HZ_COMMON */ |
3289bdb4 | 4979 | |
59543275 | 4980 | /** |
cee1afce | 4981 | * __cpu_load_update - update the rq->cpu_load[] statistics |
59543275 BP |
4982 | * @this_rq: The rq to update statistics for |
4983 | * @this_load: The current load | |
4984 | * @pending_updates: The number of missed updates | |
59543275 | 4985 | * |
3289bdb4 | 4986 | * Update rq->cpu_load[] statistics. This function is usually called every |
59543275 BP |
4987 | * scheduler tick (TICK_NSEC). |
4988 | * | |
4989 | * This function computes a decaying average: | |
4990 | * | |
4991 | * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load | |
4992 | * | |
4993 | * Because of NOHZ it might not get called on every tick which gives need for | |
4994 | * the @pending_updates argument. | |
4995 | * | |
4996 | * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1 | |
4997 | * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load | |
4998 | * = A * (A * load[i]_n-2 + B) + B | |
4999 | * = A * (A * (A * load[i]_n-3 + B) + B) + B | |
5000 | * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B | |
5001 | * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B | |
5002 | * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B | |
5003 | * = (1 - 1/2^i)^n * (load[i]_0 - load) + load | |
5004 | * | |
5005 | * In the above we've assumed load_n := load, which is true for NOHZ_FULL as | |
5006 | * any change in load would have resulted in the tick being turned back on. | |
5007 | * | |
5008 | * For regular NOHZ, this reduces to: | |
5009 | * | |
5010 | * load[i]_n = (1 - 1/2^i)^n * load[i]_0 | |
5011 | * | |
5012 | * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra | |
1f41906a | 5013 | * term. |
3289bdb4 | 5014 | */ |
1f41906a FW |
5015 | static void cpu_load_update(struct rq *this_rq, unsigned long this_load, |
5016 | unsigned long pending_updates) | |
3289bdb4 | 5017 | { |
9fd81dd5 | 5018 | unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0]; |
3289bdb4 PZ |
5019 | int i, scale; |
5020 | ||
5021 | this_rq->nr_load_updates++; | |
5022 | ||
5023 | /* Update our load: */ | |
5024 | this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ | |
5025 | for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
5026 | unsigned long old_load, new_load; | |
5027 | ||
5028 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
5029 | ||
7400d3bb | 5030 | old_load = this_rq->cpu_load[i]; |
9fd81dd5 | 5031 | #ifdef CONFIG_NO_HZ_COMMON |
3289bdb4 | 5032 | old_load = decay_load_missed(old_load, pending_updates - 1, i); |
7400d3bb BP |
5033 | if (tickless_load) { |
5034 | old_load -= decay_load_missed(tickless_load, pending_updates - 1, i); | |
5035 | /* | |
5036 | * old_load can never be a negative value because a | |
5037 | * decayed tickless_load cannot be greater than the | |
5038 | * original tickless_load. | |
5039 | */ | |
5040 | old_load += tickless_load; | |
5041 | } | |
9fd81dd5 | 5042 | #endif |
3289bdb4 PZ |
5043 | new_load = this_load; |
5044 | /* | |
5045 | * Round up the averaging division if load is increasing. This | |
5046 | * prevents us from getting stuck on 9 if the load is 10, for | |
5047 | * example. | |
5048 | */ | |
5049 | if (new_load > old_load) | |
5050 | new_load += scale - 1; | |
5051 | ||
5052 | this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; | |
5053 | } | |
5054 | ||
5055 | sched_avg_update(this_rq); | |
5056 | } | |
5057 | ||
7ea241af YD |
5058 | /* Used instead of source_load when we know the type == 0 */ |
5059 | static unsigned long weighted_cpuload(const int cpu) | |
5060 | { | |
5061 | return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs); | |
5062 | } | |
5063 | ||
3289bdb4 | 5064 | #ifdef CONFIG_NO_HZ_COMMON |
1f41906a FW |
5065 | /* |
5066 | * There is no sane way to deal with nohz on smp when using jiffies because the | |
5067 | * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading | |
5068 | * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. | |
5069 | * | |
5070 | * Therefore we need to avoid the delta approach from the regular tick when | |
5071 | * possible since that would seriously skew the load calculation. This is why we | |
5072 | * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on | |
5073 | * jiffies deltas for updates happening while in nohz mode (idle ticks, idle | |
5074 | * loop exit, nohz_idle_balance, nohz full exit...) | |
5075 | * | |
5076 | * This means we might still be one tick off for nohz periods. | |
5077 | */ | |
5078 | ||
5079 | static void cpu_load_update_nohz(struct rq *this_rq, | |
5080 | unsigned long curr_jiffies, | |
5081 | unsigned long load) | |
be68a682 FW |
5082 | { |
5083 | unsigned long pending_updates; | |
5084 | ||
5085 | pending_updates = curr_jiffies - this_rq->last_load_update_tick; | |
5086 | if (pending_updates) { | |
5087 | this_rq->last_load_update_tick = curr_jiffies; | |
5088 | /* | |
5089 | * In the regular NOHZ case, we were idle, this means load 0. | |
5090 | * In the NOHZ_FULL case, we were non-idle, we should consider | |
5091 | * its weighted load. | |
5092 | */ | |
1f41906a | 5093 | cpu_load_update(this_rq, load, pending_updates); |
be68a682 FW |
5094 | } |
5095 | } | |
5096 | ||
3289bdb4 PZ |
5097 | /* |
5098 | * Called from nohz_idle_balance() to update the load ratings before doing the | |
5099 | * idle balance. | |
5100 | */ | |
cee1afce | 5101 | static void cpu_load_update_idle(struct rq *this_rq) |
3289bdb4 | 5102 | { |
3289bdb4 PZ |
5103 | /* |
5104 | * bail if there's load or we're actually up-to-date. | |
5105 | */ | |
be68a682 | 5106 | if (weighted_cpuload(cpu_of(this_rq))) |
3289bdb4 PZ |
5107 | return; |
5108 | ||
1f41906a | 5109 | cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0); |
3289bdb4 PZ |
5110 | } |
5111 | ||
5112 | /* | |
1f41906a FW |
5113 | * Record CPU load on nohz entry so we know the tickless load to account |
5114 | * on nohz exit. cpu_load[0] happens then to be updated more frequently | |
5115 | * than other cpu_load[idx] but it should be fine as cpu_load readers | |
5116 | * shouldn't rely into synchronized cpu_load[*] updates. | |
3289bdb4 | 5117 | */ |
1f41906a | 5118 | void cpu_load_update_nohz_start(void) |
3289bdb4 PZ |
5119 | { |
5120 | struct rq *this_rq = this_rq(); | |
1f41906a FW |
5121 | |
5122 | /* | |
5123 | * This is all lockless but should be fine. If weighted_cpuload changes | |
5124 | * concurrently we'll exit nohz. And cpu_load write can race with | |
5125 | * cpu_load_update_idle() but both updater would be writing the same. | |
5126 | */ | |
5127 | this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq)); | |
5128 | } | |
5129 | ||
5130 | /* | |
5131 | * Account the tickless load in the end of a nohz frame. | |
5132 | */ | |
5133 | void cpu_load_update_nohz_stop(void) | |
5134 | { | |
316c1608 | 5135 | unsigned long curr_jiffies = READ_ONCE(jiffies); |
1f41906a FW |
5136 | struct rq *this_rq = this_rq(); |
5137 | unsigned long load; | |
8a8c69c3 | 5138 | struct rq_flags rf; |
3289bdb4 PZ |
5139 | |
5140 | if (curr_jiffies == this_rq->last_load_update_tick) | |
5141 | return; | |
5142 | ||
1f41906a | 5143 | load = weighted_cpuload(cpu_of(this_rq)); |
8a8c69c3 | 5144 | rq_lock(this_rq, &rf); |
b52fad2d | 5145 | update_rq_clock(this_rq); |
1f41906a | 5146 | cpu_load_update_nohz(this_rq, curr_jiffies, load); |
8a8c69c3 | 5147 | rq_unlock(this_rq, &rf); |
3289bdb4 | 5148 | } |
1f41906a FW |
5149 | #else /* !CONFIG_NO_HZ_COMMON */ |
5150 | static inline void cpu_load_update_nohz(struct rq *this_rq, | |
5151 | unsigned long curr_jiffies, | |
5152 | unsigned long load) { } | |
5153 | #endif /* CONFIG_NO_HZ_COMMON */ | |
5154 | ||
5155 | static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load) | |
5156 | { | |
9fd81dd5 | 5157 | #ifdef CONFIG_NO_HZ_COMMON |
1f41906a FW |
5158 | /* See the mess around cpu_load_update_nohz(). */ |
5159 | this_rq->last_load_update_tick = READ_ONCE(jiffies); | |
9fd81dd5 | 5160 | #endif |
1f41906a FW |
5161 | cpu_load_update(this_rq, load, 1); |
5162 | } | |
3289bdb4 PZ |
5163 | |
5164 | /* | |
5165 | * Called from scheduler_tick() | |
5166 | */ | |
cee1afce | 5167 | void cpu_load_update_active(struct rq *this_rq) |
3289bdb4 | 5168 | { |
7ea241af | 5169 | unsigned long load = weighted_cpuload(cpu_of(this_rq)); |
1f41906a FW |
5170 | |
5171 | if (tick_nohz_tick_stopped()) | |
5172 | cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load); | |
5173 | else | |
5174 | cpu_load_update_periodic(this_rq, load); | |
3289bdb4 PZ |
5175 | } |
5176 | ||
029632fb PZ |
5177 | /* |
5178 | * Return a low guess at the load of a migration-source cpu weighted | |
5179 | * according to the scheduling class and "nice" value. | |
5180 | * | |
5181 | * We want to under-estimate the load of migration sources, to | |
5182 | * balance conservatively. | |
5183 | */ | |
5184 | static unsigned long source_load(int cpu, int type) | |
5185 | { | |
5186 | struct rq *rq = cpu_rq(cpu); | |
5187 | unsigned long total = weighted_cpuload(cpu); | |
5188 | ||
5189 | if (type == 0 || !sched_feat(LB_BIAS)) | |
5190 | return total; | |
5191 | ||
5192 | return min(rq->cpu_load[type-1], total); | |
5193 | } | |
5194 | ||
5195 | /* | |
5196 | * Return a high guess at the load of a migration-target cpu weighted | |
5197 | * according to the scheduling class and "nice" value. | |
5198 | */ | |
5199 | static unsigned long target_load(int cpu, int type) | |
5200 | { | |
5201 | struct rq *rq = cpu_rq(cpu); | |
5202 | unsigned long total = weighted_cpuload(cpu); | |
5203 | ||
5204 | if (type == 0 || !sched_feat(LB_BIAS)) | |
5205 | return total; | |
5206 | ||
5207 | return max(rq->cpu_load[type-1], total); | |
5208 | } | |
5209 | ||
ced549fa | 5210 | static unsigned long capacity_of(int cpu) |
029632fb | 5211 | { |
ced549fa | 5212 | return cpu_rq(cpu)->cpu_capacity; |
029632fb PZ |
5213 | } |
5214 | ||
ca6d75e6 VG |
5215 | static unsigned long capacity_orig_of(int cpu) |
5216 | { | |
5217 | return cpu_rq(cpu)->cpu_capacity_orig; | |
5218 | } | |
5219 | ||
029632fb PZ |
5220 | static unsigned long cpu_avg_load_per_task(int cpu) |
5221 | { | |
5222 | struct rq *rq = cpu_rq(cpu); | |
316c1608 | 5223 | unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running); |
7ea241af | 5224 | unsigned long load_avg = weighted_cpuload(cpu); |
029632fb PZ |
5225 | |
5226 | if (nr_running) | |
b92486cb | 5227 | return load_avg / nr_running; |
029632fb PZ |
5228 | |
5229 | return 0; | |
5230 | } | |
5231 | ||
bb3469ac | 5232 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f5bfb7d9 PZ |
5233 | /* |
5234 | * effective_load() calculates the load change as seen from the root_task_group | |
5235 | * | |
5236 | * Adding load to a group doesn't make a group heavier, but can cause movement | |
5237 | * of group shares between cpus. Assuming the shares were perfectly aligned one | |
5238 | * can calculate the shift in shares. | |
cf5f0acf PZ |
5239 | * |
5240 | * Calculate the effective load difference if @wl is added (subtracted) to @tg | |
5241 | * on this @cpu and results in a total addition (subtraction) of @wg to the | |
5242 | * total group weight. | |
5243 | * | |
5244 | * Given a runqueue weight distribution (rw_i) we can compute a shares | |
5245 | * distribution (s_i) using: | |
5246 | * | |
5247 | * s_i = rw_i / \Sum rw_j (1) | |
5248 | * | |
5249 | * Suppose we have 4 CPUs and our @tg is a direct child of the root group and | |
5250 | * has 7 equal weight tasks, distributed as below (rw_i), with the resulting | |
5251 | * shares distribution (s_i): | |
5252 | * | |
5253 | * rw_i = { 2, 4, 1, 0 } | |
5254 | * s_i = { 2/7, 4/7, 1/7, 0 } | |
5255 | * | |
5256 | * As per wake_affine() we're interested in the load of two CPUs (the CPU the | |
5257 | * task used to run on and the CPU the waker is running on), we need to | |
5258 | * compute the effect of waking a task on either CPU and, in case of a sync | |
5259 | * wakeup, compute the effect of the current task going to sleep. | |
5260 | * | |
5261 | * So for a change of @wl to the local @cpu with an overall group weight change | |
5262 | * of @wl we can compute the new shares distribution (s'_i) using: | |
5263 | * | |
5264 | * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) | |
5265 | * | |
5266 | * Suppose we're interested in CPUs 0 and 1, and want to compute the load | |
5267 | * differences in waking a task to CPU 0. The additional task changes the | |
5268 | * weight and shares distributions like: | |
5269 | * | |
5270 | * rw'_i = { 3, 4, 1, 0 } | |
5271 | * s'_i = { 3/8, 4/8, 1/8, 0 } | |
5272 | * | |
5273 | * We can then compute the difference in effective weight by using: | |
5274 | * | |
5275 | * dw_i = S * (s'_i - s_i) (3) | |
5276 | * | |
5277 | * Where 'S' is the group weight as seen by its parent. | |
5278 | * | |
5279 | * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) | |
5280 | * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - | |
5281 | * 4/7) times the weight of the group. | |
f5bfb7d9 | 5282 | */ |
2069dd75 | 5283 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
bb3469ac | 5284 | { |
4be9daaa | 5285 | struct sched_entity *se = tg->se[cpu]; |
f1d239f7 | 5286 | |
9722c2da | 5287 | if (!tg->parent) /* the trivial, non-cgroup case */ |
f1d239f7 PZ |
5288 | return wl; |
5289 | ||
4be9daaa | 5290 | for_each_sched_entity(se) { |
7dd49125 PZ |
5291 | struct cfs_rq *cfs_rq = se->my_q; |
5292 | long W, w = cfs_rq_load_avg(cfs_rq); | |
4be9daaa | 5293 | |
7dd49125 | 5294 | tg = cfs_rq->tg; |
bb3469ac | 5295 | |
cf5f0acf PZ |
5296 | /* |
5297 | * W = @wg + \Sum rw_j | |
5298 | */ | |
7dd49125 PZ |
5299 | W = wg + atomic_long_read(&tg->load_avg); |
5300 | ||
5301 | /* Ensure \Sum rw_j >= rw_i */ | |
5302 | W -= cfs_rq->tg_load_avg_contrib; | |
5303 | W += w; | |
4be9daaa | 5304 | |
cf5f0acf PZ |
5305 | /* |
5306 | * w = rw_i + @wl | |
5307 | */ | |
7dd49125 | 5308 | w += wl; |
940959e9 | 5309 | |
cf5f0acf PZ |
5310 | /* |
5311 | * wl = S * s'_i; see (2) | |
5312 | */ | |
5313 | if (W > 0 && w < W) | |
ab522e33 | 5314 | wl = (w * (long)scale_load_down(tg->shares)) / W; |
977dda7c | 5315 | else |
ab522e33 | 5316 | wl = scale_load_down(tg->shares); |
940959e9 | 5317 | |
cf5f0acf PZ |
5318 | /* |
5319 | * Per the above, wl is the new se->load.weight value; since | |
5320 | * those are clipped to [MIN_SHARES, ...) do so now. See | |
5321 | * calc_cfs_shares(). | |
5322 | */ | |
977dda7c PT |
5323 | if (wl < MIN_SHARES) |
5324 | wl = MIN_SHARES; | |
cf5f0acf PZ |
5325 | |
5326 | /* | |
5327 | * wl = dw_i = S * (s'_i - s_i); see (3) | |
5328 | */ | |
9d89c257 | 5329 | wl -= se->avg.load_avg; |
cf5f0acf PZ |
5330 | |
5331 | /* | |
5332 | * Recursively apply this logic to all parent groups to compute | |
5333 | * the final effective load change on the root group. Since | |
5334 | * only the @tg group gets extra weight, all parent groups can | |
5335 | * only redistribute existing shares. @wl is the shift in shares | |
5336 | * resulting from this level per the above. | |
5337 | */ | |
4be9daaa | 5338 | wg = 0; |
4be9daaa | 5339 | } |
bb3469ac | 5340 | |
4be9daaa | 5341 | return wl; |
bb3469ac PZ |
5342 | } |
5343 | #else | |
4be9daaa | 5344 | |
58d081b5 | 5345 | static long effective_load(struct task_group *tg, int cpu, long wl, long wg) |
4be9daaa | 5346 | { |
83378269 | 5347 | return wl; |
bb3469ac | 5348 | } |
4be9daaa | 5349 | |
bb3469ac PZ |
5350 | #endif |
5351 | ||
c58d25f3 PZ |
5352 | static void record_wakee(struct task_struct *p) |
5353 | { | |
5354 | /* | |
5355 | * Only decay a single time; tasks that have less then 1 wakeup per | |
5356 | * jiffy will not have built up many flips. | |
5357 | */ | |
5358 | if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { | |
5359 | current->wakee_flips >>= 1; | |
5360 | current->wakee_flip_decay_ts = jiffies; | |
5361 | } | |
5362 | ||
5363 | if (current->last_wakee != p) { | |
5364 | current->last_wakee = p; | |
5365 | current->wakee_flips++; | |
5366 | } | |
5367 | } | |
5368 | ||
63b0e9ed MG |
5369 | /* |
5370 | * Detect M:N waker/wakee relationships via a switching-frequency heuristic. | |
c58d25f3 | 5371 | * |
63b0e9ed | 5372 | * A waker of many should wake a different task than the one last awakened |
c58d25f3 PZ |
5373 | * at a frequency roughly N times higher than one of its wakees. |
5374 | * | |
5375 | * In order to determine whether we should let the load spread vs consolidating | |
5376 | * to shared cache, we look for a minimum 'flip' frequency of llc_size in one | |
5377 | * partner, and a factor of lls_size higher frequency in the other. | |
5378 | * | |
5379 | * With both conditions met, we can be relatively sure that the relationship is | |
5380 | * non-monogamous, with partner count exceeding socket size. | |
5381 | * | |
5382 | * Waker/wakee being client/server, worker/dispatcher, interrupt source or | |
5383 | * whatever is irrelevant, spread criteria is apparent partner count exceeds | |
5384 | * socket size. | |
63b0e9ed | 5385 | */ |
62470419 MW |
5386 | static int wake_wide(struct task_struct *p) |
5387 | { | |
63b0e9ed MG |
5388 | unsigned int master = current->wakee_flips; |
5389 | unsigned int slave = p->wakee_flips; | |
7d9ffa89 | 5390 | int factor = this_cpu_read(sd_llc_size); |
62470419 | 5391 | |
63b0e9ed MG |
5392 | if (master < slave) |
5393 | swap(master, slave); | |
5394 | if (slave < factor || master < slave * factor) | |
5395 | return 0; | |
5396 | return 1; | |
62470419 MW |
5397 | } |
5398 | ||
772bd008 MR |
5399 | static int wake_affine(struct sched_domain *sd, struct task_struct *p, |
5400 | int prev_cpu, int sync) | |
098fb9db | 5401 | { |
e37b6a7b | 5402 | s64 this_load, load; |
bd61c98f | 5403 | s64 this_eff_load, prev_eff_load; |
772bd008 | 5404 | int idx, this_cpu; |
c88d5910 | 5405 | struct task_group *tg; |
83378269 | 5406 | unsigned long weight; |
b3137bc8 | 5407 | int balanced; |
098fb9db | 5408 | |
c88d5910 PZ |
5409 | idx = sd->wake_idx; |
5410 | this_cpu = smp_processor_id(); | |
c88d5910 PZ |
5411 | load = source_load(prev_cpu, idx); |
5412 | this_load = target_load(this_cpu, idx); | |
098fb9db | 5413 | |
b3137bc8 MG |
5414 | /* |
5415 | * If sync wakeup then subtract the (maximum possible) | |
5416 | * effect of the currently running task from the load | |
5417 | * of the current CPU: | |
5418 | */ | |
83378269 PZ |
5419 | if (sync) { |
5420 | tg = task_group(current); | |
9d89c257 | 5421 | weight = current->se.avg.load_avg; |
83378269 | 5422 | |
c88d5910 | 5423 | this_load += effective_load(tg, this_cpu, -weight, -weight); |
83378269 PZ |
5424 | load += effective_load(tg, prev_cpu, 0, -weight); |
5425 | } | |
b3137bc8 | 5426 | |
83378269 | 5427 | tg = task_group(p); |
9d89c257 | 5428 | weight = p->se.avg.load_avg; |
b3137bc8 | 5429 | |
71a29aa7 PZ |
5430 | /* |
5431 | * In low-load situations, where prev_cpu is idle and this_cpu is idle | |
c88d5910 PZ |
5432 | * due to the sync cause above having dropped this_load to 0, we'll |
5433 | * always have an imbalance, but there's really nothing you can do | |
5434 | * about that, so that's good too. | |
71a29aa7 PZ |
5435 | * |
5436 | * Otherwise check if either cpus are near enough in load to allow this | |
5437 | * task to be woken on this_cpu. | |
5438 | */ | |
bd61c98f VG |
5439 | this_eff_load = 100; |
5440 | this_eff_load *= capacity_of(prev_cpu); | |
e51fd5e2 | 5441 | |
bd61c98f VG |
5442 | prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; |
5443 | prev_eff_load *= capacity_of(this_cpu); | |
e51fd5e2 | 5444 | |
bd61c98f | 5445 | if (this_load > 0) { |
e51fd5e2 PZ |
5446 | this_eff_load *= this_load + |
5447 | effective_load(tg, this_cpu, weight, weight); | |
5448 | ||
e51fd5e2 | 5449 | prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); |
bd61c98f | 5450 | } |
e51fd5e2 | 5451 | |
bd61c98f | 5452 | balanced = this_eff_load <= prev_eff_load; |
098fb9db | 5453 | |
ae92882e | 5454 | schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); |
098fb9db | 5455 | |
05bfb65f VG |
5456 | if (!balanced) |
5457 | return 0; | |
098fb9db | 5458 | |
ae92882e JP |
5459 | schedstat_inc(sd->ttwu_move_affine); |
5460 | schedstat_inc(p->se.statistics.nr_wakeups_affine); | |
05bfb65f VG |
5461 | |
5462 | return 1; | |
098fb9db IM |
5463 | } |
5464 | ||
6a0b19c0 MR |
5465 | static inline int task_util(struct task_struct *p); |
5466 | static int cpu_util_wake(int cpu, struct task_struct *p); | |
5467 | ||
5468 | static unsigned long capacity_spare_wake(int cpu, struct task_struct *p) | |
5469 | { | |
5470 | return capacity_orig_of(cpu) - cpu_util_wake(cpu, p); | |
5471 | } | |
5472 | ||
aaee1203 PZ |
5473 | /* |
5474 | * find_idlest_group finds and returns the least busy CPU group within the | |
5475 | * domain. | |
5476 | */ | |
5477 | static struct sched_group * | |
78e7ed53 | 5478 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, |
c44f2a02 | 5479 | int this_cpu, int sd_flag) |
e7693a36 | 5480 | { |
b3bd3de6 | 5481 | struct sched_group *idlest = NULL, *group = sd->groups; |
6a0b19c0 | 5482 | struct sched_group *most_spare_sg = NULL; |
6b94780e VG |
5483 | unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0; |
5484 | unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0; | |
6a0b19c0 | 5485 | unsigned long most_spare = 0, this_spare = 0; |
c44f2a02 | 5486 | int load_idx = sd->forkexec_idx; |
6b94780e VG |
5487 | int imbalance_scale = 100 + (sd->imbalance_pct-100)/2; |
5488 | unsigned long imbalance = scale_load_down(NICE_0_LOAD) * | |
5489 | (sd->imbalance_pct-100) / 100; | |
e7693a36 | 5490 | |
c44f2a02 VG |
5491 | if (sd_flag & SD_BALANCE_WAKE) |
5492 | load_idx = sd->wake_idx; | |
5493 | ||
aaee1203 | 5494 | do { |
6b94780e VG |
5495 | unsigned long load, avg_load, runnable_load; |
5496 | unsigned long spare_cap, max_spare_cap; | |
aaee1203 PZ |
5497 | int local_group; |
5498 | int i; | |
e7693a36 | 5499 | |
aaee1203 PZ |
5500 | /* Skip over this group if it has no CPUs allowed */ |
5501 | if (!cpumask_intersects(sched_group_cpus(group), | |
0c98d344 | 5502 | &p->cpus_allowed)) |
aaee1203 PZ |
5503 | continue; |
5504 | ||
5505 | local_group = cpumask_test_cpu(this_cpu, | |
5506 | sched_group_cpus(group)); | |
5507 | ||
6a0b19c0 MR |
5508 | /* |
5509 | * Tally up the load of all CPUs in the group and find | |
5510 | * the group containing the CPU with most spare capacity. | |
5511 | */ | |
aaee1203 | 5512 | avg_load = 0; |
6b94780e | 5513 | runnable_load = 0; |
6a0b19c0 | 5514 | max_spare_cap = 0; |
aaee1203 PZ |
5515 | |
5516 | for_each_cpu(i, sched_group_cpus(group)) { | |
5517 | /* Bias balancing toward cpus of our domain */ | |
5518 | if (local_group) | |
5519 | load = source_load(i, load_idx); | |
5520 | else | |
5521 | load = target_load(i, load_idx); | |
5522 | ||
6b94780e VG |
5523 | runnable_load += load; |
5524 | ||
5525 | avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs); | |
6a0b19c0 MR |
5526 | |
5527 | spare_cap = capacity_spare_wake(i, p); | |
5528 | ||
5529 | if (spare_cap > max_spare_cap) | |
5530 | max_spare_cap = spare_cap; | |
aaee1203 PZ |
5531 | } |
5532 | ||
63b2ca30 | 5533 | /* Adjust by relative CPU capacity of the group */ |
6b94780e VG |
5534 | avg_load = (avg_load * SCHED_CAPACITY_SCALE) / |
5535 | group->sgc->capacity; | |
5536 | runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) / | |
5537 | group->sgc->capacity; | |
aaee1203 PZ |
5538 | |
5539 | if (local_group) { | |
6b94780e VG |
5540 | this_runnable_load = runnable_load; |
5541 | this_avg_load = avg_load; | |
6a0b19c0 MR |
5542 | this_spare = max_spare_cap; |
5543 | } else { | |
6b94780e VG |
5544 | if (min_runnable_load > (runnable_load + imbalance)) { |
5545 | /* | |
5546 | * The runnable load is significantly smaller | |
5547 | * so we can pick this new cpu | |
5548 | */ | |
5549 | min_runnable_load = runnable_load; | |
5550 | min_avg_load = avg_load; | |
5551 | idlest = group; | |
5552 | } else if ((runnable_load < (min_runnable_load + imbalance)) && | |
5553 | (100*min_avg_load > imbalance_scale*avg_load)) { | |
5554 | /* | |
5555 | * The runnable loads are close so take the | |
5556 | * blocked load into account through avg_load. | |
5557 | */ | |
5558 | min_avg_load = avg_load; | |
6a0b19c0 MR |
5559 | idlest = group; |
5560 | } | |
5561 | ||
5562 | if (most_spare < max_spare_cap) { | |
5563 | most_spare = max_spare_cap; | |
5564 | most_spare_sg = group; | |
5565 | } | |
aaee1203 PZ |
5566 | } |
5567 | } while (group = group->next, group != sd->groups); | |
5568 | ||
6a0b19c0 MR |
5569 | /* |
5570 | * The cross-over point between using spare capacity or least load | |
5571 | * is too conservative for high utilization tasks on partially | |
5572 | * utilized systems if we require spare_capacity > task_util(p), | |
5573 | * so we allow for some task stuffing by using | |
5574 | * spare_capacity > task_util(p)/2. | |
f519a3f1 VG |
5575 | * |
5576 | * Spare capacity can't be used for fork because the utilization has | |
5577 | * not been set yet, we must first select a rq to compute the initial | |
5578 | * utilization. | |
6a0b19c0 | 5579 | */ |
f519a3f1 VG |
5580 | if (sd_flag & SD_BALANCE_FORK) |
5581 | goto skip_spare; | |
5582 | ||
6a0b19c0 | 5583 | if (this_spare > task_util(p) / 2 && |
6b94780e | 5584 | imbalance_scale*this_spare > 100*most_spare) |
6a0b19c0 | 5585 | return NULL; |
6b94780e VG |
5586 | |
5587 | if (most_spare > task_util(p) / 2) | |
6a0b19c0 MR |
5588 | return most_spare_sg; |
5589 | ||
f519a3f1 | 5590 | skip_spare: |
6b94780e VG |
5591 | if (!idlest) |
5592 | return NULL; | |
5593 | ||
5594 | if (min_runnable_load > (this_runnable_load + imbalance)) | |
aaee1203 | 5595 | return NULL; |
6b94780e VG |
5596 | |
5597 | if ((this_runnable_load < (min_runnable_load + imbalance)) && | |
5598 | (100*this_avg_load < imbalance_scale*min_avg_load)) | |
5599 | return NULL; | |
5600 | ||
aaee1203 PZ |
5601 | return idlest; |
5602 | } | |
5603 | ||
5604 | /* | |
5605 | * find_idlest_cpu - find the idlest cpu among the cpus in group. | |
5606 | */ | |
5607 | static int | |
5608 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) | |
5609 | { | |
5610 | unsigned long load, min_load = ULONG_MAX; | |
83a0a96a NP |
5611 | unsigned int min_exit_latency = UINT_MAX; |
5612 | u64 latest_idle_timestamp = 0; | |
5613 | int least_loaded_cpu = this_cpu; | |
5614 | int shallowest_idle_cpu = -1; | |
aaee1203 PZ |
5615 | int i; |
5616 | ||
eaecf41f MR |
5617 | /* Check if we have any choice: */ |
5618 | if (group->group_weight == 1) | |
5619 | return cpumask_first(sched_group_cpus(group)); | |
5620 | ||
aaee1203 | 5621 | /* Traverse only the allowed CPUs */ |
0c98d344 | 5622 | for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) { |
83a0a96a NP |
5623 | if (idle_cpu(i)) { |
5624 | struct rq *rq = cpu_rq(i); | |
5625 | struct cpuidle_state *idle = idle_get_state(rq); | |
5626 | if (idle && idle->exit_latency < min_exit_latency) { | |
5627 | /* | |
5628 | * We give priority to a CPU whose idle state | |
5629 | * has the smallest exit latency irrespective | |
5630 | * of any idle timestamp. | |
5631 | */ | |
5632 | min_exit_latency = idle->exit_latency; | |
5633 | latest_idle_timestamp = rq->idle_stamp; | |
5634 | shallowest_idle_cpu = i; | |
5635 | } else if ((!idle || idle->exit_latency == min_exit_latency) && | |
5636 | rq->idle_stamp > latest_idle_timestamp) { | |
5637 | /* | |
5638 | * If equal or no active idle state, then | |
5639 | * the most recently idled CPU might have | |
5640 | * a warmer cache. | |
5641 | */ | |
5642 | latest_idle_timestamp = rq->idle_stamp; | |
5643 | shallowest_idle_cpu = i; | |
5644 | } | |
9f96742a | 5645 | } else if (shallowest_idle_cpu == -1) { |
83a0a96a NP |
5646 | load = weighted_cpuload(i); |
5647 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
5648 | min_load = load; | |
5649 | least_loaded_cpu = i; | |
5650 | } | |
e7693a36 GH |
5651 | } |
5652 | } | |
5653 | ||
83a0a96a | 5654 | return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; |
aaee1203 | 5655 | } |
e7693a36 | 5656 | |
a50bde51 | 5657 | /* |
10e2f1ac PZ |
5658 | * Implement a for_each_cpu() variant that starts the scan at a given cpu |
5659 | * (@start), and wraps around. | |
5660 | * | |
5661 | * This is used to scan for idle CPUs; such that not all CPUs looking for an | |
5662 | * idle CPU find the same CPU. The down-side is that tasks tend to cycle | |
5663 | * through the LLC domain. | |
5664 | * | |
5665 | * Especially tbench is found sensitive to this. | |
5666 | */ | |
5667 | ||
5668 | static int cpumask_next_wrap(int n, const struct cpumask *mask, int start, int *wrapped) | |
5669 | { | |
5670 | int next; | |
5671 | ||
5672 | again: | |
5673 | next = find_next_bit(cpumask_bits(mask), nr_cpumask_bits, n+1); | |
5674 | ||
5675 | if (*wrapped) { | |
5676 | if (next >= start) | |
5677 | return nr_cpumask_bits; | |
5678 | } else { | |
5679 | if (next >= nr_cpumask_bits) { | |
5680 | *wrapped = 1; | |
5681 | n = -1; | |
5682 | goto again; | |
5683 | } | |
5684 | } | |
5685 | ||
5686 | return next; | |
5687 | } | |
5688 | ||
5689 | #define for_each_cpu_wrap(cpu, mask, start, wrap) \ | |
5690 | for ((wrap) = 0, (cpu) = (start)-1; \ | |
5691 | (cpu) = cpumask_next_wrap((cpu), (mask), (start), &(wrap)), \ | |
5692 | (cpu) < nr_cpumask_bits; ) | |
5693 | ||
5694 | #ifdef CONFIG_SCHED_SMT | |
5695 | ||
5696 | static inline void set_idle_cores(int cpu, int val) | |
5697 | { | |
5698 | struct sched_domain_shared *sds; | |
5699 | ||
5700 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
5701 | if (sds) | |
5702 | WRITE_ONCE(sds->has_idle_cores, val); | |
5703 | } | |
5704 | ||
5705 | static inline bool test_idle_cores(int cpu, bool def) | |
5706 | { | |
5707 | struct sched_domain_shared *sds; | |
5708 | ||
5709 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); | |
5710 | if (sds) | |
5711 | return READ_ONCE(sds->has_idle_cores); | |
5712 | ||
5713 | return def; | |
5714 | } | |
5715 | ||
5716 | /* | |
5717 | * Scans the local SMT mask to see if the entire core is idle, and records this | |
5718 | * information in sd_llc_shared->has_idle_cores. | |
5719 | * | |
5720 | * Since SMT siblings share all cache levels, inspecting this limited remote | |
5721 | * state should be fairly cheap. | |
5722 | */ | |
1b568f0a | 5723 | void __update_idle_core(struct rq *rq) |
10e2f1ac PZ |
5724 | { |
5725 | int core = cpu_of(rq); | |
5726 | int cpu; | |
5727 | ||
5728 | rcu_read_lock(); | |
5729 | if (test_idle_cores(core, true)) | |
5730 | goto unlock; | |
5731 | ||
5732 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
5733 | if (cpu == core) | |
5734 | continue; | |
5735 | ||
5736 | if (!idle_cpu(cpu)) | |
5737 | goto unlock; | |
5738 | } | |
5739 | ||
5740 | set_idle_cores(core, 1); | |
5741 | unlock: | |
5742 | rcu_read_unlock(); | |
5743 | } | |
5744 | ||
5745 | /* | |
5746 | * Scan the entire LLC domain for idle cores; this dynamically switches off if | |
5747 | * there are no idle cores left in the system; tracked through | |
5748 | * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. | |
5749 | */ | |
5750 | static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) | |
5751 | { | |
5752 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); | |
5753 | int core, cpu, wrap; | |
5754 | ||
1b568f0a PZ |
5755 | if (!static_branch_likely(&sched_smt_present)) |
5756 | return -1; | |
5757 | ||
10e2f1ac PZ |
5758 | if (!test_idle_cores(target, false)) |
5759 | return -1; | |
5760 | ||
0c98d344 | 5761 | cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed); |
10e2f1ac PZ |
5762 | |
5763 | for_each_cpu_wrap(core, cpus, target, wrap) { | |
5764 | bool idle = true; | |
5765 | ||
5766 | for_each_cpu(cpu, cpu_smt_mask(core)) { | |
5767 | cpumask_clear_cpu(cpu, cpus); | |
5768 | if (!idle_cpu(cpu)) | |
5769 | idle = false; | |
5770 | } | |
5771 | ||
5772 | if (idle) | |
5773 | return core; | |
5774 | } | |
5775 | ||
5776 | /* | |
5777 | * Failed to find an idle core; stop looking for one. | |
5778 | */ | |
5779 | set_idle_cores(target, 0); | |
5780 | ||
5781 | return -1; | |
5782 | } | |
5783 | ||
5784 | /* | |
5785 | * Scan the local SMT mask for idle CPUs. | |
5786 | */ | |
5787 | static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) | |
5788 | { | |
5789 | int cpu; | |
5790 | ||
1b568f0a PZ |
5791 | if (!static_branch_likely(&sched_smt_present)) |
5792 | return -1; | |
5793 | ||
10e2f1ac | 5794 | for_each_cpu(cpu, cpu_smt_mask(target)) { |
0c98d344 | 5795 | if (!cpumask_test_cpu(cpu, &p->cpus_allowed)) |
10e2f1ac PZ |
5796 | continue; |
5797 | if (idle_cpu(cpu)) | |
5798 | return cpu; | |
5799 | } | |
5800 | ||
5801 | return -1; | |
5802 | } | |
5803 | ||
5804 | #else /* CONFIG_SCHED_SMT */ | |
5805 | ||
5806 | static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) | |
5807 | { | |
5808 | return -1; | |
5809 | } | |
5810 | ||
5811 | static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) | |
5812 | { | |
5813 | return -1; | |
5814 | } | |
5815 | ||
5816 | #endif /* CONFIG_SCHED_SMT */ | |
5817 | ||
5818 | /* | |
5819 | * Scan the LLC domain for idle CPUs; this is dynamically regulated by | |
5820 | * comparing the average scan cost (tracked in sd->avg_scan_cost) against the | |
5821 | * average idle time for this rq (as found in rq->avg_idle). | |
a50bde51 | 5822 | */ |
10e2f1ac PZ |
5823 | static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target) |
5824 | { | |
9cfb38a7 WL |
5825 | struct sched_domain *this_sd; |
5826 | u64 avg_cost, avg_idle = this_rq()->avg_idle; | |
10e2f1ac PZ |
5827 | u64 time, cost; |
5828 | s64 delta; | |
5829 | int cpu, wrap; | |
5830 | ||
9cfb38a7 WL |
5831 | this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); |
5832 | if (!this_sd) | |
5833 | return -1; | |
5834 | ||
5835 | avg_cost = this_sd->avg_scan_cost; | |
5836 | ||
10e2f1ac PZ |
5837 | /* |
5838 | * Due to large variance we need a large fuzz factor; hackbench in | |
5839 | * particularly is sensitive here. | |
5840 | */ | |
4c77b18c | 5841 | if (sched_feat(SIS_AVG_CPU) && (avg_idle / 512) < avg_cost) |
10e2f1ac PZ |
5842 | return -1; |
5843 | ||
5844 | time = local_clock(); | |
5845 | ||
5846 | for_each_cpu_wrap(cpu, sched_domain_span(sd), target, wrap) { | |
0c98d344 | 5847 | if (!cpumask_test_cpu(cpu, &p->cpus_allowed)) |
10e2f1ac PZ |
5848 | continue; |
5849 | if (idle_cpu(cpu)) | |
5850 | break; | |
5851 | } | |
5852 | ||
5853 | time = local_clock() - time; | |
5854 | cost = this_sd->avg_scan_cost; | |
5855 | delta = (s64)(time - cost) / 8; | |
5856 | this_sd->avg_scan_cost += delta; | |
5857 | ||
5858 | return cpu; | |
5859 | } | |
5860 | ||
5861 | /* | |
5862 | * Try and locate an idle core/thread in the LLC cache domain. | |
a50bde51 | 5863 | */ |
772bd008 | 5864 | static int select_idle_sibling(struct task_struct *p, int prev, int target) |
a50bde51 | 5865 | { |
99bd5e2f | 5866 | struct sched_domain *sd; |
10e2f1ac | 5867 | int i; |
a50bde51 | 5868 | |
e0a79f52 MG |
5869 | if (idle_cpu(target)) |
5870 | return target; | |
99bd5e2f SS |
5871 | |
5872 | /* | |
10e2f1ac | 5873 | * If the previous cpu is cache affine and idle, don't be stupid. |
99bd5e2f | 5874 | */ |
772bd008 MR |
5875 | if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev)) |
5876 | return prev; | |
a50bde51 | 5877 | |
518cd623 | 5878 | sd = rcu_dereference(per_cpu(sd_llc, target)); |
10e2f1ac PZ |
5879 | if (!sd) |
5880 | return target; | |
772bd008 | 5881 | |
10e2f1ac PZ |
5882 | i = select_idle_core(p, sd, target); |
5883 | if ((unsigned)i < nr_cpumask_bits) | |
5884 | return i; | |
37407ea7 | 5885 | |
10e2f1ac PZ |
5886 | i = select_idle_cpu(p, sd, target); |
5887 | if ((unsigned)i < nr_cpumask_bits) | |
5888 | return i; | |
5889 | ||
5890 | i = select_idle_smt(p, sd, target); | |
5891 | if ((unsigned)i < nr_cpumask_bits) | |
5892 | return i; | |
970e1789 | 5893 | |
a50bde51 PZ |
5894 | return target; |
5895 | } | |
231678b7 | 5896 | |
8bb5b00c | 5897 | /* |
9e91d61d | 5898 | * cpu_util returns the amount of capacity of a CPU that is used by CFS |
8bb5b00c | 5899 | * tasks. The unit of the return value must be the one of capacity so we can |
9e91d61d DE |
5900 | * compare the utilization with the capacity of the CPU that is available for |
5901 | * CFS task (ie cpu_capacity). | |
231678b7 DE |
5902 | * |
5903 | * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the | |
5904 | * recent utilization of currently non-runnable tasks on a CPU. It represents | |
5905 | * the amount of utilization of a CPU in the range [0..capacity_orig] where | |
5906 | * capacity_orig is the cpu_capacity available at the highest frequency | |
5907 | * (arch_scale_freq_capacity()). | |
5908 | * The utilization of a CPU converges towards a sum equal to or less than the | |
5909 | * current capacity (capacity_curr <= capacity_orig) of the CPU because it is | |
5910 | * the running time on this CPU scaled by capacity_curr. | |
5911 | * | |
5912 | * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even | |
5913 | * higher than capacity_orig because of unfortunate rounding in | |
5914 | * cfs.avg.util_avg or just after migrating tasks and new task wakeups until | |
5915 | * the average stabilizes with the new running time. We need to check that the | |
5916 | * utilization stays within the range of [0..capacity_orig] and cap it if | |
5917 | * necessary. Without utilization capping, a group could be seen as overloaded | |
5918 | * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of | |
5919 | * available capacity. We allow utilization to overshoot capacity_curr (but not | |
5920 | * capacity_orig) as it useful for predicting the capacity required after task | |
5921 | * migrations (scheduler-driven DVFS). | |
8bb5b00c | 5922 | */ |
9e91d61d | 5923 | static int cpu_util(int cpu) |
8bb5b00c | 5924 | { |
9e91d61d | 5925 | unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg; |
8bb5b00c VG |
5926 | unsigned long capacity = capacity_orig_of(cpu); |
5927 | ||
231678b7 | 5928 | return (util >= capacity) ? capacity : util; |
8bb5b00c | 5929 | } |
a50bde51 | 5930 | |
3273163c MR |
5931 | static inline int task_util(struct task_struct *p) |
5932 | { | |
5933 | return p->se.avg.util_avg; | |
5934 | } | |
5935 | ||
104cb16d MR |
5936 | /* |
5937 | * cpu_util_wake: Compute cpu utilization with any contributions from | |
5938 | * the waking task p removed. | |
5939 | */ | |
5940 | static int cpu_util_wake(int cpu, struct task_struct *p) | |
5941 | { | |
5942 | unsigned long util, capacity; | |
5943 | ||
5944 | /* Task has no contribution or is new */ | |
5945 | if (cpu != task_cpu(p) || !p->se.avg.last_update_time) | |
5946 | return cpu_util(cpu); | |
5947 | ||
5948 | capacity = capacity_orig_of(cpu); | |
5949 | util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0); | |
5950 | ||
5951 | return (util >= capacity) ? capacity : util; | |
5952 | } | |
5953 | ||
3273163c MR |
5954 | /* |
5955 | * Disable WAKE_AFFINE in the case where task @p doesn't fit in the | |
5956 | * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu. | |
5957 | * | |
5958 | * In that case WAKE_AFFINE doesn't make sense and we'll let | |
5959 | * BALANCE_WAKE sort things out. | |
5960 | */ | |
5961 | static int wake_cap(struct task_struct *p, int cpu, int prev_cpu) | |
5962 | { | |
5963 | long min_cap, max_cap; | |
5964 | ||
5965 | min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu)); | |
5966 | max_cap = cpu_rq(cpu)->rd->max_cpu_capacity; | |
5967 | ||
5968 | /* Minimum capacity is close to max, no need to abort wake_affine */ | |
5969 | if (max_cap - min_cap < max_cap >> 3) | |
5970 | return 0; | |
5971 | ||
104cb16d MR |
5972 | /* Bring task utilization in sync with prev_cpu */ |
5973 | sync_entity_load_avg(&p->se); | |
5974 | ||
3273163c MR |
5975 | return min_cap * 1024 < task_util(p) * capacity_margin; |
5976 | } | |
5977 | ||
aaee1203 | 5978 | /* |
de91b9cb MR |
5979 | * select_task_rq_fair: Select target runqueue for the waking task in domains |
5980 | * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, | |
5981 | * SD_BALANCE_FORK, or SD_BALANCE_EXEC. | |
aaee1203 | 5982 | * |
de91b9cb MR |
5983 | * Balances load by selecting the idlest cpu in the idlest group, or under |
5984 | * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. | |
aaee1203 | 5985 | * |
de91b9cb | 5986 | * Returns the target cpu number. |
aaee1203 PZ |
5987 | * |
5988 | * preempt must be disabled. | |
5989 | */ | |
0017d735 | 5990 | static int |
ac66f547 | 5991 | select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) |
aaee1203 | 5992 | { |
29cd8bae | 5993 | struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; |
c88d5910 | 5994 | int cpu = smp_processor_id(); |
63b0e9ed | 5995 | int new_cpu = prev_cpu; |
99bd5e2f | 5996 | int want_affine = 0; |
5158f4e4 | 5997 | int sync = wake_flags & WF_SYNC; |
c88d5910 | 5998 | |
c58d25f3 PZ |
5999 | if (sd_flag & SD_BALANCE_WAKE) { |
6000 | record_wakee(p); | |
3273163c | 6001 | want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) |
0c98d344 | 6002 | && cpumask_test_cpu(cpu, &p->cpus_allowed); |
c58d25f3 | 6003 | } |
aaee1203 | 6004 | |
dce840a0 | 6005 | rcu_read_lock(); |
aaee1203 | 6006 | for_each_domain(cpu, tmp) { |
e4f42888 | 6007 | if (!(tmp->flags & SD_LOAD_BALANCE)) |
63b0e9ed | 6008 | break; |
e4f42888 | 6009 | |
fe3bcfe1 | 6010 | /* |
99bd5e2f SS |
6011 | * If both cpu and prev_cpu are part of this domain, |
6012 | * cpu is a valid SD_WAKE_AFFINE target. | |
fe3bcfe1 | 6013 | */ |
99bd5e2f SS |
6014 | if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && |
6015 | cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { | |
6016 | affine_sd = tmp; | |
29cd8bae | 6017 | break; |
f03542a7 | 6018 | } |
29cd8bae | 6019 | |
f03542a7 | 6020 | if (tmp->flags & sd_flag) |
29cd8bae | 6021 | sd = tmp; |
63b0e9ed MG |
6022 | else if (!want_affine) |
6023 | break; | |
29cd8bae PZ |
6024 | } |
6025 | ||
63b0e9ed MG |
6026 | if (affine_sd) { |
6027 | sd = NULL; /* Prefer wake_affine over balance flags */ | |
772bd008 | 6028 | if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync)) |
63b0e9ed | 6029 | new_cpu = cpu; |
8b911acd | 6030 | } |
e7693a36 | 6031 | |
63b0e9ed MG |
6032 | if (!sd) { |
6033 | if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */ | |
772bd008 | 6034 | new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); |
63b0e9ed MG |
6035 | |
6036 | } else while (sd) { | |
aaee1203 | 6037 | struct sched_group *group; |
c88d5910 | 6038 | int weight; |
098fb9db | 6039 | |
0763a660 | 6040 | if (!(sd->flags & sd_flag)) { |
aaee1203 PZ |
6041 | sd = sd->child; |
6042 | continue; | |
6043 | } | |
098fb9db | 6044 | |
c44f2a02 | 6045 | group = find_idlest_group(sd, p, cpu, sd_flag); |
aaee1203 PZ |
6046 | if (!group) { |
6047 | sd = sd->child; | |
6048 | continue; | |
6049 | } | |
4ae7d5ce | 6050 | |
d7c33c49 | 6051 | new_cpu = find_idlest_cpu(group, p, cpu); |
aaee1203 PZ |
6052 | if (new_cpu == -1 || new_cpu == cpu) { |
6053 | /* Now try balancing at a lower domain level of cpu */ | |
6054 | sd = sd->child; | |
6055 | continue; | |
e7693a36 | 6056 | } |
aaee1203 PZ |
6057 | |
6058 | /* Now try balancing at a lower domain level of new_cpu */ | |
6059 | cpu = new_cpu; | |
669c55e9 | 6060 | weight = sd->span_weight; |
aaee1203 PZ |
6061 | sd = NULL; |
6062 | for_each_domain(cpu, tmp) { | |
669c55e9 | 6063 | if (weight <= tmp->span_weight) |
aaee1203 | 6064 | break; |
0763a660 | 6065 | if (tmp->flags & sd_flag) |
aaee1203 PZ |
6066 | sd = tmp; |
6067 | } | |
6068 | /* while loop will break here if sd == NULL */ | |
e7693a36 | 6069 | } |
dce840a0 | 6070 | rcu_read_unlock(); |
e7693a36 | 6071 | |
c88d5910 | 6072 | return new_cpu; |
e7693a36 | 6073 | } |
0a74bef8 PT |
6074 | |
6075 | /* | |
6076 | * Called immediately before a task is migrated to a new cpu; task_cpu(p) and | |
6077 | * cfs_rq_of(p) references at time of call are still valid and identify the | |
525628c7 | 6078 | * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held. |
0a74bef8 | 6079 | */ |
5a4fd036 | 6080 | static void migrate_task_rq_fair(struct task_struct *p) |
0a74bef8 | 6081 | { |
59efa0ba PZ |
6082 | /* |
6083 | * As blocked tasks retain absolute vruntime the migration needs to | |
6084 | * deal with this by subtracting the old and adding the new | |
6085 | * min_vruntime -- the latter is done by enqueue_entity() when placing | |
6086 | * the task on the new runqueue. | |
6087 | */ | |
6088 | if (p->state == TASK_WAKING) { | |
6089 | struct sched_entity *se = &p->se; | |
6090 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
6091 | u64 min_vruntime; | |
6092 | ||
6093 | #ifndef CONFIG_64BIT | |
6094 | u64 min_vruntime_copy; | |
6095 | ||
6096 | do { | |
6097 | min_vruntime_copy = cfs_rq->min_vruntime_copy; | |
6098 | smp_rmb(); | |
6099 | min_vruntime = cfs_rq->min_vruntime; | |
6100 | } while (min_vruntime != min_vruntime_copy); | |
6101 | #else | |
6102 | min_vruntime = cfs_rq->min_vruntime; | |
6103 | #endif | |
6104 | ||
6105 | se->vruntime -= min_vruntime; | |
6106 | } | |
6107 | ||
aff3e498 | 6108 | /* |
9d89c257 YD |
6109 | * We are supposed to update the task to "current" time, then its up to date |
6110 | * and ready to go to new CPU/cfs_rq. But we have difficulty in getting | |
6111 | * what current time is, so simply throw away the out-of-date time. This | |
6112 | * will result in the wakee task is less decayed, but giving the wakee more | |
6113 | * load sounds not bad. | |
aff3e498 | 6114 | */ |
9d89c257 YD |
6115 | remove_entity_load_avg(&p->se); |
6116 | ||
6117 | /* Tell new CPU we are migrated */ | |
6118 | p->se.avg.last_update_time = 0; | |
3944a927 BS |
6119 | |
6120 | /* We have migrated, no longer consider this task hot */ | |
9d89c257 | 6121 | p->se.exec_start = 0; |
0a74bef8 | 6122 | } |
12695578 YD |
6123 | |
6124 | static void task_dead_fair(struct task_struct *p) | |
6125 | { | |
6126 | remove_entity_load_avg(&p->se); | |
6127 | } | |
e7693a36 GH |
6128 | #endif /* CONFIG_SMP */ |
6129 | ||
e52fb7c0 PZ |
6130 | static unsigned long |
6131 | wakeup_gran(struct sched_entity *curr, struct sched_entity *se) | |
0bbd3336 PZ |
6132 | { |
6133 | unsigned long gran = sysctl_sched_wakeup_granularity; | |
6134 | ||
6135 | /* | |
e52fb7c0 PZ |
6136 | * Since its curr running now, convert the gran from real-time |
6137 | * to virtual-time in his units. | |
13814d42 MG |
6138 | * |
6139 | * By using 'se' instead of 'curr' we penalize light tasks, so | |
6140 | * they get preempted easier. That is, if 'se' < 'curr' then | |
6141 | * the resulting gran will be larger, therefore penalizing the | |
6142 | * lighter, if otoh 'se' > 'curr' then the resulting gran will | |
6143 | * be smaller, again penalizing the lighter task. | |
6144 | * | |
6145 | * This is especially important for buddies when the leftmost | |
6146 | * task is higher priority than the buddy. | |
0bbd3336 | 6147 | */ |
f4ad9bd2 | 6148 | return calc_delta_fair(gran, se); |
0bbd3336 PZ |
6149 | } |
6150 | ||
464b7527 PZ |
6151 | /* |
6152 | * Should 'se' preempt 'curr'. | |
6153 | * | |
6154 | * |s1 | |
6155 | * |s2 | |
6156 | * |s3 | |
6157 | * g | |
6158 | * |<--->|c | |
6159 | * | |
6160 | * w(c, s1) = -1 | |
6161 | * w(c, s2) = 0 | |
6162 | * w(c, s3) = 1 | |
6163 | * | |
6164 | */ | |
6165 | static int | |
6166 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) | |
6167 | { | |
6168 | s64 gran, vdiff = curr->vruntime - se->vruntime; | |
6169 | ||
6170 | if (vdiff <= 0) | |
6171 | return -1; | |
6172 | ||
e52fb7c0 | 6173 | gran = wakeup_gran(curr, se); |
464b7527 PZ |
6174 | if (vdiff > gran) |
6175 | return 1; | |
6176 | ||
6177 | return 0; | |
6178 | } | |
6179 | ||
02479099 PZ |
6180 | static void set_last_buddy(struct sched_entity *se) |
6181 | { | |
69c80f3e VP |
6182 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
6183 | return; | |
6184 | ||
6185 | for_each_sched_entity(se) | |
6186 | cfs_rq_of(se)->last = se; | |
02479099 PZ |
6187 | } |
6188 | ||
6189 | static void set_next_buddy(struct sched_entity *se) | |
6190 | { | |
69c80f3e VP |
6191 | if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) |
6192 | return; | |
6193 | ||
6194 | for_each_sched_entity(se) | |
6195 | cfs_rq_of(se)->next = se; | |
02479099 PZ |
6196 | } |
6197 | ||
ac53db59 RR |
6198 | static void set_skip_buddy(struct sched_entity *se) |
6199 | { | |
69c80f3e VP |
6200 | for_each_sched_entity(se) |
6201 | cfs_rq_of(se)->skip = se; | |
ac53db59 RR |
6202 | } |
6203 | ||
bf0f6f24 IM |
6204 | /* |
6205 | * Preempt the current task with a newly woken task if needed: | |
6206 | */ | |
5a9b86f6 | 6207 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) |
bf0f6f24 IM |
6208 | { |
6209 | struct task_struct *curr = rq->curr; | |
8651a86c | 6210 | struct sched_entity *se = &curr->se, *pse = &p->se; |
03e89e45 | 6211 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
f685ceac | 6212 | int scale = cfs_rq->nr_running >= sched_nr_latency; |
2f36825b | 6213 | int next_buddy_marked = 0; |
bf0f6f24 | 6214 | |
4ae7d5ce IM |
6215 | if (unlikely(se == pse)) |
6216 | return; | |
6217 | ||
5238cdd3 | 6218 | /* |
163122b7 | 6219 | * This is possible from callers such as attach_tasks(), in which we |
5238cdd3 PT |
6220 | * unconditionally check_prempt_curr() after an enqueue (which may have |
6221 | * lead to a throttle). This both saves work and prevents false | |
6222 | * next-buddy nomination below. | |
6223 | */ | |
6224 | if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) | |
6225 | return; | |
6226 | ||
2f36825b | 6227 | if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { |
3cb63d52 | 6228 | set_next_buddy(pse); |
2f36825b VP |
6229 | next_buddy_marked = 1; |
6230 | } | |
57fdc26d | 6231 | |
aec0a514 BR |
6232 | /* |
6233 | * We can come here with TIF_NEED_RESCHED already set from new task | |
6234 | * wake up path. | |
5238cdd3 PT |
6235 | * |
6236 | * Note: this also catches the edge-case of curr being in a throttled | |
6237 | * group (e.g. via set_curr_task), since update_curr() (in the | |
6238 | * enqueue of curr) will have resulted in resched being set. This | |
6239 | * prevents us from potentially nominating it as a false LAST_BUDDY | |
6240 | * below. | |
aec0a514 BR |
6241 | */ |
6242 | if (test_tsk_need_resched(curr)) | |
6243 | return; | |
6244 | ||
a2f5c9ab DH |
6245 | /* Idle tasks are by definition preempted by non-idle tasks. */ |
6246 | if (unlikely(curr->policy == SCHED_IDLE) && | |
6247 | likely(p->policy != SCHED_IDLE)) | |
6248 | goto preempt; | |
6249 | ||
91c234b4 | 6250 | /* |
a2f5c9ab DH |
6251 | * Batch and idle tasks do not preempt non-idle tasks (their preemption |
6252 | * is driven by the tick): | |
91c234b4 | 6253 | */ |
8ed92e51 | 6254 | if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) |
91c234b4 | 6255 | return; |
bf0f6f24 | 6256 | |
464b7527 | 6257 | find_matching_se(&se, &pse); |
9bbd7374 | 6258 | update_curr(cfs_rq_of(se)); |
002f128b | 6259 | BUG_ON(!pse); |
2f36825b VP |
6260 | if (wakeup_preempt_entity(se, pse) == 1) { |
6261 | /* | |
6262 | * Bias pick_next to pick the sched entity that is | |
6263 | * triggering this preemption. | |
6264 | */ | |
6265 | if (!next_buddy_marked) | |
6266 | set_next_buddy(pse); | |
3a7e73a2 | 6267 | goto preempt; |
2f36825b | 6268 | } |
464b7527 | 6269 | |
3a7e73a2 | 6270 | return; |
a65ac745 | 6271 | |
3a7e73a2 | 6272 | preempt: |
8875125e | 6273 | resched_curr(rq); |
3a7e73a2 PZ |
6274 | /* |
6275 | * Only set the backward buddy when the current task is still | |
6276 | * on the rq. This can happen when a wakeup gets interleaved | |
6277 | * with schedule on the ->pre_schedule() or idle_balance() | |
6278 | * point, either of which can * drop the rq lock. | |
6279 | * | |
6280 | * Also, during early boot the idle thread is in the fair class, | |
6281 | * for obvious reasons its a bad idea to schedule back to it. | |
6282 | */ | |
6283 | if (unlikely(!se->on_rq || curr == rq->idle)) | |
6284 | return; | |
6285 | ||
6286 | if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) | |
6287 | set_last_buddy(se); | |
bf0f6f24 IM |
6288 | } |
6289 | ||
606dba2e | 6290 | static struct task_struct * |
d8ac8971 | 6291 | pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) |
bf0f6f24 IM |
6292 | { |
6293 | struct cfs_rq *cfs_rq = &rq->cfs; | |
6294 | struct sched_entity *se; | |
678d5718 | 6295 | struct task_struct *p; |
37e117c0 | 6296 | int new_tasks; |
678d5718 | 6297 | |
6e83125c | 6298 | again: |
678d5718 PZ |
6299 | #ifdef CONFIG_FAIR_GROUP_SCHED |
6300 | if (!cfs_rq->nr_running) | |
38033c37 | 6301 | goto idle; |
678d5718 | 6302 | |
3f1d2a31 | 6303 | if (prev->sched_class != &fair_sched_class) |
678d5718 PZ |
6304 | goto simple; |
6305 | ||
6306 | /* | |
6307 | * Because of the set_next_buddy() in dequeue_task_fair() it is rather | |
6308 | * likely that a next task is from the same cgroup as the current. | |
6309 | * | |
6310 | * Therefore attempt to avoid putting and setting the entire cgroup | |
6311 | * hierarchy, only change the part that actually changes. | |
6312 | */ | |
6313 | ||
6314 | do { | |
6315 | struct sched_entity *curr = cfs_rq->curr; | |
6316 | ||
6317 | /* | |
6318 | * Since we got here without doing put_prev_entity() we also | |
6319 | * have to consider cfs_rq->curr. If it is still a runnable | |
6320 | * entity, update_curr() will update its vruntime, otherwise | |
6321 | * forget we've ever seen it. | |
6322 | */ | |
54d27365 BS |
6323 | if (curr) { |
6324 | if (curr->on_rq) | |
6325 | update_curr(cfs_rq); | |
6326 | else | |
6327 | curr = NULL; | |
678d5718 | 6328 | |
54d27365 BS |
6329 | /* |
6330 | * This call to check_cfs_rq_runtime() will do the | |
6331 | * throttle and dequeue its entity in the parent(s). | |
6332 | * Therefore the 'simple' nr_running test will indeed | |
6333 | * be correct. | |
6334 | */ | |
6335 | if (unlikely(check_cfs_rq_runtime(cfs_rq))) | |
6336 | goto simple; | |
6337 | } | |
678d5718 PZ |
6338 | |
6339 | se = pick_next_entity(cfs_rq, curr); | |
6340 | cfs_rq = group_cfs_rq(se); | |
6341 | } while (cfs_rq); | |
6342 | ||
6343 | p = task_of(se); | |
6344 | ||
6345 | /* | |
6346 | * Since we haven't yet done put_prev_entity and if the selected task | |
6347 | * is a different task than we started out with, try and touch the | |
6348 | * least amount of cfs_rqs. | |
6349 | */ | |
6350 | if (prev != p) { | |
6351 | struct sched_entity *pse = &prev->se; | |
6352 | ||
6353 | while (!(cfs_rq = is_same_group(se, pse))) { | |
6354 | int se_depth = se->depth; | |
6355 | int pse_depth = pse->depth; | |
6356 | ||
6357 | if (se_depth <= pse_depth) { | |
6358 | put_prev_entity(cfs_rq_of(pse), pse); | |
6359 | pse = parent_entity(pse); | |
6360 | } | |
6361 | if (se_depth >= pse_depth) { | |
6362 | set_next_entity(cfs_rq_of(se), se); | |
6363 | se = parent_entity(se); | |
6364 | } | |
6365 | } | |
6366 | ||
6367 | put_prev_entity(cfs_rq, pse); | |
6368 | set_next_entity(cfs_rq, se); | |
6369 | } | |
6370 | ||
6371 | if (hrtick_enabled(rq)) | |
6372 | hrtick_start_fair(rq, p); | |
6373 | ||
6374 | return p; | |
6375 | simple: | |
6376 | cfs_rq = &rq->cfs; | |
6377 | #endif | |
bf0f6f24 | 6378 | |
36ace27e | 6379 | if (!cfs_rq->nr_running) |
38033c37 | 6380 | goto idle; |
bf0f6f24 | 6381 | |
3f1d2a31 | 6382 | put_prev_task(rq, prev); |
606dba2e | 6383 | |
bf0f6f24 | 6384 | do { |
678d5718 | 6385 | se = pick_next_entity(cfs_rq, NULL); |
f4b6755f | 6386 | set_next_entity(cfs_rq, se); |
bf0f6f24 IM |
6387 | cfs_rq = group_cfs_rq(se); |
6388 | } while (cfs_rq); | |
6389 | ||
8f4d37ec | 6390 | p = task_of(se); |
678d5718 | 6391 | |
b39e66ea MG |
6392 | if (hrtick_enabled(rq)) |
6393 | hrtick_start_fair(rq, p); | |
8f4d37ec PZ |
6394 | |
6395 | return p; | |
38033c37 PZ |
6396 | |
6397 | idle: | |
46f69fa3 MF |
6398 | new_tasks = idle_balance(rq, rf); |
6399 | ||
37e117c0 PZ |
6400 | /* |
6401 | * Because idle_balance() releases (and re-acquires) rq->lock, it is | |
6402 | * possible for any higher priority task to appear. In that case we | |
6403 | * must re-start the pick_next_entity() loop. | |
6404 | */ | |
e4aa358b | 6405 | if (new_tasks < 0) |
37e117c0 PZ |
6406 | return RETRY_TASK; |
6407 | ||
e4aa358b | 6408 | if (new_tasks > 0) |
38033c37 | 6409 | goto again; |
38033c37 PZ |
6410 | |
6411 | return NULL; | |
bf0f6f24 IM |
6412 | } |
6413 | ||
6414 | /* | |
6415 | * Account for a descheduled task: | |
6416 | */ | |
31ee529c | 6417 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
bf0f6f24 IM |
6418 | { |
6419 | struct sched_entity *se = &prev->se; | |
6420 | struct cfs_rq *cfs_rq; | |
6421 | ||
6422 | for_each_sched_entity(se) { | |
6423 | cfs_rq = cfs_rq_of(se); | |
ab6cde26 | 6424 | put_prev_entity(cfs_rq, se); |
bf0f6f24 IM |
6425 | } |
6426 | } | |
6427 | ||
ac53db59 RR |
6428 | /* |
6429 | * sched_yield() is very simple | |
6430 | * | |
6431 | * The magic of dealing with the ->skip buddy is in pick_next_entity. | |
6432 | */ | |
6433 | static void yield_task_fair(struct rq *rq) | |
6434 | { | |
6435 | struct task_struct *curr = rq->curr; | |
6436 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); | |
6437 | struct sched_entity *se = &curr->se; | |
6438 | ||
6439 | /* | |
6440 | * Are we the only task in the tree? | |
6441 | */ | |
6442 | if (unlikely(rq->nr_running == 1)) | |
6443 | return; | |
6444 | ||
6445 | clear_buddies(cfs_rq, se); | |
6446 | ||
6447 | if (curr->policy != SCHED_BATCH) { | |
6448 | update_rq_clock(rq); | |
6449 | /* | |
6450 | * Update run-time statistics of the 'current'. | |
6451 | */ | |
6452 | update_curr(cfs_rq); | |
916671c0 MG |
6453 | /* |
6454 | * Tell update_rq_clock() that we've just updated, | |
6455 | * so we don't do microscopic update in schedule() | |
6456 | * and double the fastpath cost. | |
6457 | */ | |
9edfbfed | 6458 | rq_clock_skip_update(rq, true); |
ac53db59 RR |
6459 | } |
6460 | ||
6461 | set_skip_buddy(se); | |
6462 | } | |
6463 | ||
d95f4122 MG |
6464 | static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) |
6465 | { | |
6466 | struct sched_entity *se = &p->se; | |
6467 | ||
5238cdd3 PT |
6468 | /* throttled hierarchies are not runnable */ |
6469 | if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) | |
d95f4122 MG |
6470 | return false; |
6471 | ||
6472 | /* Tell the scheduler that we'd really like pse to run next. */ | |
6473 | set_next_buddy(se); | |
6474 | ||
d95f4122 MG |
6475 | yield_task_fair(rq); |
6476 | ||
6477 | return true; | |
6478 | } | |
6479 | ||
681f3e68 | 6480 | #ifdef CONFIG_SMP |
bf0f6f24 | 6481 | /************************************************** |
e9c84cb8 PZ |
6482 | * Fair scheduling class load-balancing methods. |
6483 | * | |
6484 | * BASICS | |
6485 | * | |
6486 | * The purpose of load-balancing is to achieve the same basic fairness the | |
6487 | * per-cpu scheduler provides, namely provide a proportional amount of compute | |
6488 | * time to each task. This is expressed in the following equation: | |
6489 | * | |
6490 | * W_i,n/P_i == W_j,n/P_j for all i,j (1) | |
6491 | * | |
6492 | * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight | |
6493 | * W_i,0 is defined as: | |
6494 | * | |
6495 | * W_i,0 = \Sum_j w_i,j (2) | |
6496 | * | |
6497 | * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight | |
1c3de5e1 | 6498 | * is derived from the nice value as per sched_prio_to_weight[]. |
e9c84cb8 PZ |
6499 | * |
6500 | * The weight average is an exponential decay average of the instantaneous | |
6501 | * weight: | |
6502 | * | |
6503 | * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) | |
6504 | * | |
ced549fa | 6505 | * C_i is the compute capacity of cpu i, typically it is the |
e9c84cb8 PZ |
6506 | * fraction of 'recent' time available for SCHED_OTHER task execution. But it |
6507 | * can also include other factors [XXX]. | |
6508 | * | |
6509 | * To achieve this balance we define a measure of imbalance which follows | |
6510 | * directly from (1): | |
6511 | * | |
ced549fa | 6512 | * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) |
e9c84cb8 PZ |
6513 | * |
6514 | * We them move tasks around to minimize the imbalance. In the continuous | |
6515 | * function space it is obvious this converges, in the discrete case we get | |
6516 | * a few fun cases generally called infeasible weight scenarios. | |
6517 | * | |
6518 | * [XXX expand on: | |
6519 | * - infeasible weights; | |
6520 | * - local vs global optima in the discrete case. ] | |
6521 | * | |
6522 | * | |
6523 | * SCHED DOMAINS | |
6524 | * | |
6525 | * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) | |
6526 | * for all i,j solution, we create a tree of cpus that follows the hardware | |
6527 | * topology where each level pairs two lower groups (or better). This results | |
6528 | * in O(log n) layers. Furthermore we reduce the number of cpus going up the | |
6529 | * tree to only the first of the previous level and we decrease the frequency | |
6530 | * of load-balance at each level inv. proportional to the number of cpus in | |
6531 | * the groups. | |
6532 | * | |
6533 | * This yields: | |
6534 | * | |
6535 | * log_2 n 1 n | |
6536 | * \Sum { --- * --- * 2^i } = O(n) (5) | |
6537 | * i = 0 2^i 2^i | |
6538 | * `- size of each group | |
6539 | * | | `- number of cpus doing load-balance | |
6540 | * | `- freq | |
6541 | * `- sum over all levels | |
6542 | * | |
6543 | * Coupled with a limit on how many tasks we can migrate every balance pass, | |
6544 | * this makes (5) the runtime complexity of the balancer. | |
6545 | * | |
6546 | * An important property here is that each CPU is still (indirectly) connected | |
6547 | * to every other cpu in at most O(log n) steps: | |
6548 | * | |
6549 | * The adjacency matrix of the resulting graph is given by: | |
6550 | * | |
97a7142f | 6551 | * log_2 n |
e9c84cb8 PZ |
6552 | * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) |
6553 | * k = 0 | |
6554 | * | |
6555 | * And you'll find that: | |
6556 | * | |
6557 | * A^(log_2 n)_i,j != 0 for all i,j (7) | |
6558 | * | |
6559 | * Showing there's indeed a path between every cpu in at most O(log n) steps. | |
6560 | * The task movement gives a factor of O(m), giving a convergence complexity | |
6561 | * of: | |
6562 | * | |
6563 | * O(nm log n), n := nr_cpus, m := nr_tasks (8) | |
6564 | * | |
6565 | * | |
6566 | * WORK CONSERVING | |
6567 | * | |
6568 | * In order to avoid CPUs going idle while there's still work to do, new idle | |
6569 | * balancing is more aggressive and has the newly idle cpu iterate up the domain | |
6570 | * tree itself instead of relying on other CPUs to bring it work. | |
6571 | * | |
6572 | * This adds some complexity to both (5) and (8) but it reduces the total idle | |
6573 | * time. | |
6574 | * | |
6575 | * [XXX more?] | |
6576 | * | |
6577 | * | |
6578 | * CGROUPS | |
6579 | * | |
6580 | * Cgroups make a horror show out of (2), instead of a simple sum we get: | |
6581 | * | |
6582 | * s_k,i | |
6583 | * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) | |
6584 | * S_k | |
6585 | * | |
6586 | * Where | |
6587 | * | |
6588 | * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) | |
6589 | * | |
6590 | * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. | |
6591 | * | |
6592 | * The big problem is S_k, its a global sum needed to compute a local (W_i) | |
6593 | * property. | |
6594 | * | |
6595 | * [XXX write more on how we solve this.. _after_ merging pjt's patches that | |
6596 | * rewrite all of this once again.] | |
97a7142f | 6597 | */ |
bf0f6f24 | 6598 | |
ed387b78 HS |
6599 | static unsigned long __read_mostly max_load_balance_interval = HZ/10; |
6600 | ||
0ec8aa00 PZ |
6601 | enum fbq_type { regular, remote, all }; |
6602 | ||
ddcdf6e7 | 6603 | #define LBF_ALL_PINNED 0x01 |
367456c7 | 6604 | #define LBF_NEED_BREAK 0x02 |
6263322c PZ |
6605 | #define LBF_DST_PINNED 0x04 |
6606 | #define LBF_SOME_PINNED 0x08 | |
ddcdf6e7 PZ |
6607 | |
6608 | struct lb_env { | |
6609 | struct sched_domain *sd; | |
6610 | ||
ddcdf6e7 | 6611 | struct rq *src_rq; |
85c1e7da | 6612 | int src_cpu; |
ddcdf6e7 PZ |
6613 | |
6614 | int dst_cpu; | |
6615 | struct rq *dst_rq; | |
6616 | ||
88b8dac0 SV |
6617 | struct cpumask *dst_grpmask; |
6618 | int new_dst_cpu; | |
ddcdf6e7 | 6619 | enum cpu_idle_type idle; |
bd939f45 | 6620 | long imbalance; |
b9403130 MW |
6621 | /* The set of CPUs under consideration for load-balancing */ |
6622 | struct cpumask *cpus; | |
6623 | ||
ddcdf6e7 | 6624 | unsigned int flags; |
367456c7 PZ |
6625 | |
6626 | unsigned int loop; | |
6627 | unsigned int loop_break; | |
6628 | unsigned int loop_max; | |
0ec8aa00 PZ |
6629 | |
6630 | enum fbq_type fbq_type; | |
163122b7 | 6631 | struct list_head tasks; |
ddcdf6e7 PZ |
6632 | }; |
6633 | ||
029632fb PZ |
6634 | /* |
6635 | * Is this task likely cache-hot: | |
6636 | */ | |
5d5e2b1b | 6637 | static int task_hot(struct task_struct *p, struct lb_env *env) |
029632fb PZ |
6638 | { |
6639 | s64 delta; | |
6640 | ||
e5673f28 KT |
6641 | lockdep_assert_held(&env->src_rq->lock); |
6642 | ||
029632fb PZ |
6643 | if (p->sched_class != &fair_sched_class) |
6644 | return 0; | |
6645 | ||
6646 | if (unlikely(p->policy == SCHED_IDLE)) | |
6647 | return 0; | |
6648 | ||
6649 | /* | |
6650 | * Buddy candidates are cache hot: | |
6651 | */ | |
5d5e2b1b | 6652 | if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && |
029632fb PZ |
6653 | (&p->se == cfs_rq_of(&p->se)->next || |
6654 | &p->se == cfs_rq_of(&p->se)->last)) | |
6655 | return 1; | |
6656 | ||
6657 | if (sysctl_sched_migration_cost == -1) | |
6658 | return 1; | |
6659 | if (sysctl_sched_migration_cost == 0) | |
6660 | return 0; | |
6661 | ||
5d5e2b1b | 6662 | delta = rq_clock_task(env->src_rq) - p->se.exec_start; |
029632fb PZ |
6663 | |
6664 | return delta < (s64)sysctl_sched_migration_cost; | |
6665 | } | |
6666 | ||
3a7053b3 | 6667 | #ifdef CONFIG_NUMA_BALANCING |
c1ceac62 | 6668 | /* |
2a1ed24c SD |
6669 | * Returns 1, if task migration degrades locality |
6670 | * Returns 0, if task migration improves locality i.e migration preferred. | |
6671 | * Returns -1, if task migration is not affected by locality. | |
c1ceac62 | 6672 | */ |
2a1ed24c | 6673 | static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) |
3a7053b3 | 6674 | { |
b1ad065e | 6675 | struct numa_group *numa_group = rcu_dereference(p->numa_group); |
c1ceac62 | 6676 | unsigned long src_faults, dst_faults; |
3a7053b3 MG |
6677 | int src_nid, dst_nid; |
6678 | ||
2a595721 | 6679 | if (!static_branch_likely(&sched_numa_balancing)) |
2a1ed24c SD |
6680 | return -1; |
6681 | ||
c3b9bc5b | 6682 | if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) |
2a1ed24c | 6683 | return -1; |
7a0f3083 MG |
6684 | |
6685 | src_nid = cpu_to_node(env->src_cpu); | |
6686 | dst_nid = cpu_to_node(env->dst_cpu); | |
6687 | ||
83e1d2cd | 6688 | if (src_nid == dst_nid) |
2a1ed24c | 6689 | return -1; |
7a0f3083 | 6690 | |
2a1ed24c SD |
6691 | /* Migrating away from the preferred node is always bad. */ |
6692 | if (src_nid == p->numa_preferred_nid) { | |
6693 | if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) | |
6694 | return 1; | |
6695 | else | |
6696 | return -1; | |
6697 | } | |
b1ad065e | 6698 | |
c1ceac62 RR |
6699 | /* Encourage migration to the preferred node. */ |
6700 | if (dst_nid == p->numa_preferred_nid) | |
2a1ed24c | 6701 | return 0; |
b1ad065e | 6702 | |
c1ceac62 RR |
6703 | if (numa_group) { |
6704 | src_faults = group_faults(p, src_nid); | |
6705 | dst_faults = group_faults(p, dst_nid); | |
6706 | } else { | |
6707 | src_faults = task_faults(p, src_nid); | |
6708 | dst_faults = task_faults(p, dst_nid); | |
b1ad065e RR |
6709 | } |
6710 | ||
c1ceac62 | 6711 | return dst_faults < src_faults; |
7a0f3083 MG |
6712 | } |
6713 | ||
3a7053b3 | 6714 | #else |
2a1ed24c | 6715 | static inline int migrate_degrades_locality(struct task_struct *p, |
3a7053b3 MG |
6716 | struct lb_env *env) |
6717 | { | |
2a1ed24c | 6718 | return -1; |
7a0f3083 | 6719 | } |
3a7053b3 MG |
6720 | #endif |
6721 | ||
1e3c88bd PZ |
6722 | /* |
6723 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
6724 | */ | |
6725 | static | |
8e45cb54 | 6726 | int can_migrate_task(struct task_struct *p, struct lb_env *env) |
1e3c88bd | 6727 | { |
2a1ed24c | 6728 | int tsk_cache_hot; |
e5673f28 KT |
6729 | |
6730 | lockdep_assert_held(&env->src_rq->lock); | |
6731 | ||
1e3c88bd PZ |
6732 | /* |
6733 | * We do not migrate tasks that are: | |
d3198084 | 6734 | * 1) throttled_lb_pair, or |
1e3c88bd | 6735 | * 2) cannot be migrated to this CPU due to cpus_allowed, or |
d3198084 JK |
6736 | * 3) running (obviously), or |
6737 | * 4) are cache-hot on their current CPU. | |
1e3c88bd | 6738 | */ |
d3198084 JK |
6739 | if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) |
6740 | return 0; | |
6741 | ||
0c98d344 | 6742 | if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) { |
e02e60c1 | 6743 | int cpu; |
88b8dac0 | 6744 | |
ae92882e | 6745 | schedstat_inc(p->se.statistics.nr_failed_migrations_affine); |
88b8dac0 | 6746 | |
6263322c PZ |
6747 | env->flags |= LBF_SOME_PINNED; |
6748 | ||
88b8dac0 SV |
6749 | /* |
6750 | * Remember if this task can be migrated to any other cpu in | |
6751 | * our sched_group. We may want to revisit it if we couldn't | |
6752 | * meet load balance goals by pulling other tasks on src_cpu. | |
6753 | * | |
6754 | * Also avoid computing new_dst_cpu if we have already computed | |
6755 | * one in current iteration. | |
6756 | */ | |
6263322c | 6757 | if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) |
88b8dac0 SV |
6758 | return 0; |
6759 | ||
e02e60c1 JK |
6760 | /* Prevent to re-select dst_cpu via env's cpus */ |
6761 | for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { | |
0c98d344 | 6762 | if (cpumask_test_cpu(cpu, &p->cpus_allowed)) { |
6263322c | 6763 | env->flags |= LBF_DST_PINNED; |
e02e60c1 JK |
6764 | env->new_dst_cpu = cpu; |
6765 | break; | |
6766 | } | |
88b8dac0 | 6767 | } |
e02e60c1 | 6768 | |
1e3c88bd PZ |
6769 | return 0; |
6770 | } | |
88b8dac0 SV |
6771 | |
6772 | /* Record that we found atleast one task that could run on dst_cpu */ | |
8e45cb54 | 6773 | env->flags &= ~LBF_ALL_PINNED; |
1e3c88bd | 6774 | |
ddcdf6e7 | 6775 | if (task_running(env->src_rq, p)) { |
ae92882e | 6776 | schedstat_inc(p->se.statistics.nr_failed_migrations_running); |
1e3c88bd PZ |
6777 | return 0; |
6778 | } | |
6779 | ||
6780 | /* | |
6781 | * Aggressive migration if: | |
3a7053b3 MG |
6782 | * 1) destination numa is preferred |
6783 | * 2) task is cache cold, or | |
6784 | * 3) too many balance attempts have failed. | |
1e3c88bd | 6785 | */ |
2a1ed24c SD |
6786 | tsk_cache_hot = migrate_degrades_locality(p, env); |
6787 | if (tsk_cache_hot == -1) | |
6788 | tsk_cache_hot = task_hot(p, env); | |
3a7053b3 | 6789 | |
2a1ed24c | 6790 | if (tsk_cache_hot <= 0 || |
7a96c231 | 6791 | env->sd->nr_balance_failed > env->sd->cache_nice_tries) { |
2a1ed24c | 6792 | if (tsk_cache_hot == 1) { |
ae92882e JP |
6793 | schedstat_inc(env->sd->lb_hot_gained[env->idle]); |
6794 | schedstat_inc(p->se.statistics.nr_forced_migrations); | |
3a7053b3 | 6795 | } |
1e3c88bd PZ |
6796 | return 1; |
6797 | } | |
6798 | ||
ae92882e | 6799 | schedstat_inc(p->se.statistics.nr_failed_migrations_hot); |
4e2dcb73 | 6800 | return 0; |
1e3c88bd PZ |
6801 | } |
6802 | ||
897c395f | 6803 | /* |
163122b7 KT |
6804 | * detach_task() -- detach the task for the migration specified in env |
6805 | */ | |
6806 | static void detach_task(struct task_struct *p, struct lb_env *env) | |
6807 | { | |
6808 | lockdep_assert_held(&env->src_rq->lock); | |
6809 | ||
163122b7 | 6810 | p->on_rq = TASK_ON_RQ_MIGRATING; |
5704ac0a | 6811 | deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); |
163122b7 KT |
6812 | set_task_cpu(p, env->dst_cpu); |
6813 | } | |
6814 | ||
897c395f | 6815 | /* |
e5673f28 | 6816 | * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as |
897c395f | 6817 | * part of active balancing operations within "domain". |
897c395f | 6818 | * |
e5673f28 | 6819 | * Returns a task if successful and NULL otherwise. |
897c395f | 6820 | */ |
e5673f28 | 6821 | static struct task_struct *detach_one_task(struct lb_env *env) |
897c395f PZ |
6822 | { |
6823 | struct task_struct *p, *n; | |
897c395f | 6824 | |
e5673f28 KT |
6825 | lockdep_assert_held(&env->src_rq->lock); |
6826 | ||
367456c7 | 6827 | list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { |
367456c7 PZ |
6828 | if (!can_migrate_task(p, env)) |
6829 | continue; | |
897c395f | 6830 | |
163122b7 | 6831 | detach_task(p, env); |
e5673f28 | 6832 | |
367456c7 | 6833 | /* |
e5673f28 | 6834 | * Right now, this is only the second place where |
163122b7 | 6835 | * lb_gained[env->idle] is updated (other is detach_tasks) |
e5673f28 | 6836 | * so we can safely collect stats here rather than |
163122b7 | 6837 | * inside detach_tasks(). |
367456c7 | 6838 | */ |
ae92882e | 6839 | schedstat_inc(env->sd->lb_gained[env->idle]); |
e5673f28 | 6840 | return p; |
897c395f | 6841 | } |
e5673f28 | 6842 | return NULL; |
897c395f PZ |
6843 | } |
6844 | ||
eb95308e PZ |
6845 | static const unsigned int sched_nr_migrate_break = 32; |
6846 | ||
5d6523eb | 6847 | /* |
163122b7 KT |
6848 | * detach_tasks() -- tries to detach up to imbalance weighted load from |
6849 | * busiest_rq, as part of a balancing operation within domain "sd". | |
5d6523eb | 6850 | * |
163122b7 | 6851 | * Returns number of detached tasks if successful and 0 otherwise. |
5d6523eb | 6852 | */ |
163122b7 | 6853 | static int detach_tasks(struct lb_env *env) |
1e3c88bd | 6854 | { |
5d6523eb PZ |
6855 | struct list_head *tasks = &env->src_rq->cfs_tasks; |
6856 | struct task_struct *p; | |
367456c7 | 6857 | unsigned long load; |
163122b7 KT |
6858 | int detached = 0; |
6859 | ||
6860 | lockdep_assert_held(&env->src_rq->lock); | |
1e3c88bd | 6861 | |
bd939f45 | 6862 | if (env->imbalance <= 0) |
5d6523eb | 6863 | return 0; |
1e3c88bd | 6864 | |
5d6523eb | 6865 | while (!list_empty(tasks)) { |
985d3a4c YD |
6866 | /* |
6867 | * We don't want to steal all, otherwise we may be treated likewise, | |
6868 | * which could at worst lead to a livelock crash. | |
6869 | */ | |
6870 | if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) | |
6871 | break; | |
6872 | ||
5d6523eb | 6873 | p = list_first_entry(tasks, struct task_struct, se.group_node); |
1e3c88bd | 6874 | |
367456c7 PZ |
6875 | env->loop++; |
6876 | /* We've more or less seen every task there is, call it quits */ | |
5d6523eb | 6877 | if (env->loop > env->loop_max) |
367456c7 | 6878 | break; |
5d6523eb PZ |
6879 | |
6880 | /* take a breather every nr_migrate tasks */ | |
367456c7 | 6881 | if (env->loop > env->loop_break) { |
eb95308e | 6882 | env->loop_break += sched_nr_migrate_break; |
8e45cb54 | 6883 | env->flags |= LBF_NEED_BREAK; |
ee00e66f | 6884 | break; |
a195f004 | 6885 | } |
1e3c88bd | 6886 | |
d3198084 | 6887 | if (!can_migrate_task(p, env)) |
367456c7 PZ |
6888 | goto next; |
6889 | ||
6890 | load = task_h_load(p); | |
5d6523eb | 6891 | |
eb95308e | 6892 | if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) |
367456c7 PZ |
6893 | goto next; |
6894 | ||
bd939f45 | 6895 | if ((load / 2) > env->imbalance) |
367456c7 | 6896 | goto next; |
1e3c88bd | 6897 | |
163122b7 KT |
6898 | detach_task(p, env); |
6899 | list_add(&p->se.group_node, &env->tasks); | |
6900 | ||
6901 | detached++; | |
bd939f45 | 6902 | env->imbalance -= load; |
1e3c88bd PZ |
6903 | |
6904 | #ifdef CONFIG_PREEMPT | |
ee00e66f PZ |
6905 | /* |
6906 | * NEWIDLE balancing is a source of latency, so preemptible | |
163122b7 | 6907 | * kernels will stop after the first task is detached to minimize |
ee00e66f PZ |
6908 | * the critical section. |
6909 | */ | |
5d6523eb | 6910 | if (env->idle == CPU_NEWLY_IDLE) |
ee00e66f | 6911 | break; |
1e3c88bd PZ |
6912 | #endif |
6913 | ||
ee00e66f PZ |
6914 | /* |
6915 | * We only want to steal up to the prescribed amount of | |
6916 | * weighted load. | |
6917 | */ | |
bd939f45 | 6918 | if (env->imbalance <= 0) |
ee00e66f | 6919 | break; |
367456c7 PZ |
6920 | |
6921 | continue; | |
6922 | next: | |
5d6523eb | 6923 | list_move_tail(&p->se.group_node, tasks); |
1e3c88bd | 6924 | } |
5d6523eb | 6925 | |
1e3c88bd | 6926 | /* |
163122b7 KT |
6927 | * Right now, this is one of only two places we collect this stat |
6928 | * so we can safely collect detach_one_task() stats here rather | |
6929 | * than inside detach_one_task(). | |
1e3c88bd | 6930 | */ |
ae92882e | 6931 | schedstat_add(env->sd->lb_gained[env->idle], detached); |
1e3c88bd | 6932 | |
163122b7 KT |
6933 | return detached; |
6934 | } | |
6935 | ||
6936 | /* | |
6937 | * attach_task() -- attach the task detached by detach_task() to its new rq. | |
6938 | */ | |
6939 | static void attach_task(struct rq *rq, struct task_struct *p) | |
6940 | { | |
6941 | lockdep_assert_held(&rq->lock); | |
6942 | ||
6943 | BUG_ON(task_rq(p) != rq); | |
5704ac0a | 6944 | activate_task(rq, p, ENQUEUE_NOCLOCK); |
3ea94de1 | 6945 | p->on_rq = TASK_ON_RQ_QUEUED; |
163122b7 KT |
6946 | check_preempt_curr(rq, p, 0); |
6947 | } | |
6948 | ||
6949 | /* | |
6950 | * attach_one_task() -- attaches the task returned from detach_one_task() to | |
6951 | * its new rq. | |
6952 | */ | |
6953 | static void attach_one_task(struct rq *rq, struct task_struct *p) | |
6954 | { | |
8a8c69c3 PZ |
6955 | struct rq_flags rf; |
6956 | ||
6957 | rq_lock(rq, &rf); | |
5704ac0a | 6958 | update_rq_clock(rq); |
163122b7 | 6959 | attach_task(rq, p); |
8a8c69c3 | 6960 | rq_unlock(rq, &rf); |
163122b7 KT |
6961 | } |
6962 | ||
6963 | /* | |
6964 | * attach_tasks() -- attaches all tasks detached by detach_tasks() to their | |
6965 | * new rq. | |
6966 | */ | |
6967 | static void attach_tasks(struct lb_env *env) | |
6968 | { | |
6969 | struct list_head *tasks = &env->tasks; | |
6970 | struct task_struct *p; | |
8a8c69c3 | 6971 | struct rq_flags rf; |
163122b7 | 6972 | |
8a8c69c3 | 6973 | rq_lock(env->dst_rq, &rf); |
5704ac0a | 6974 | update_rq_clock(env->dst_rq); |
163122b7 KT |
6975 | |
6976 | while (!list_empty(tasks)) { | |
6977 | p = list_first_entry(tasks, struct task_struct, se.group_node); | |
6978 | list_del_init(&p->se.group_node); | |
1e3c88bd | 6979 | |
163122b7 KT |
6980 | attach_task(env->dst_rq, p); |
6981 | } | |
6982 | ||
8a8c69c3 | 6983 | rq_unlock(env->dst_rq, &rf); |
1e3c88bd PZ |
6984 | } |
6985 | ||
230059de | 6986 | #ifdef CONFIG_FAIR_GROUP_SCHED |
48a16753 | 6987 | static void update_blocked_averages(int cpu) |
9e3081ca | 6988 | { |
9e3081ca | 6989 | struct rq *rq = cpu_rq(cpu); |
48a16753 | 6990 | struct cfs_rq *cfs_rq; |
8a8c69c3 | 6991 | struct rq_flags rf; |
9e3081ca | 6992 | |
8a8c69c3 | 6993 | rq_lock_irqsave(rq, &rf); |
48a16753 | 6994 | update_rq_clock(rq); |
9d89c257 | 6995 | |
9763b67f PZ |
6996 | /* |
6997 | * Iterates the task_group tree in a bottom up fashion, see | |
6998 | * list_add_leaf_cfs_rq() for details. | |
6999 | */ | |
64660c86 | 7000 | for_each_leaf_cfs_rq(rq, cfs_rq) { |
bc427898 VG |
7001 | struct sched_entity *se; |
7002 | ||
9d89c257 YD |
7003 | /* throttled entities do not contribute to load */ |
7004 | if (throttled_hierarchy(cfs_rq)) | |
7005 | continue; | |
48a16753 | 7006 | |
a2c6c91f | 7007 | if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true)) |
9d89c257 | 7008 | update_tg_load_avg(cfs_rq, 0); |
4e516076 | 7009 | |
bc427898 VG |
7010 | /* Propagate pending load changes to the parent, if any: */ |
7011 | se = cfs_rq->tg->se[cpu]; | |
7012 | if (se && !skip_blocked_update(se)) | |
7013 | update_load_avg(se, 0); | |
9d89c257 | 7014 | } |
8a8c69c3 | 7015 | rq_unlock_irqrestore(rq, &rf); |
9e3081ca PZ |
7016 | } |
7017 | ||
9763b67f | 7018 | /* |
68520796 | 7019 | * Compute the hierarchical load factor for cfs_rq and all its ascendants. |
9763b67f PZ |
7020 | * This needs to be done in a top-down fashion because the load of a child |
7021 | * group is a fraction of its parents load. | |
7022 | */ | |
68520796 | 7023 | static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) |
9763b67f | 7024 | { |
68520796 VD |
7025 | struct rq *rq = rq_of(cfs_rq); |
7026 | struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; | |
a35b6466 | 7027 | unsigned long now = jiffies; |
68520796 | 7028 | unsigned long load; |
a35b6466 | 7029 | |
68520796 | 7030 | if (cfs_rq->last_h_load_update == now) |
a35b6466 PZ |
7031 | return; |
7032 | ||
68520796 VD |
7033 | cfs_rq->h_load_next = NULL; |
7034 | for_each_sched_entity(se) { | |
7035 | cfs_rq = cfs_rq_of(se); | |
7036 | cfs_rq->h_load_next = se; | |
7037 | if (cfs_rq->last_h_load_update == now) | |
7038 | break; | |
7039 | } | |
a35b6466 | 7040 | |
68520796 | 7041 | if (!se) { |
7ea241af | 7042 | cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); |
68520796 VD |
7043 | cfs_rq->last_h_load_update = now; |
7044 | } | |
7045 | ||
7046 | while ((se = cfs_rq->h_load_next) != NULL) { | |
7047 | load = cfs_rq->h_load; | |
7ea241af YD |
7048 | load = div64_ul(load * se->avg.load_avg, |
7049 | cfs_rq_load_avg(cfs_rq) + 1); | |
68520796 VD |
7050 | cfs_rq = group_cfs_rq(se); |
7051 | cfs_rq->h_load = load; | |
7052 | cfs_rq->last_h_load_update = now; | |
7053 | } | |
9763b67f PZ |
7054 | } |
7055 | ||
367456c7 | 7056 | static unsigned long task_h_load(struct task_struct *p) |
230059de | 7057 | { |
367456c7 | 7058 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
230059de | 7059 | |
68520796 | 7060 | update_cfs_rq_h_load(cfs_rq); |
9d89c257 | 7061 | return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, |
7ea241af | 7062 | cfs_rq_load_avg(cfs_rq) + 1); |
230059de PZ |
7063 | } |
7064 | #else | |
48a16753 | 7065 | static inline void update_blocked_averages(int cpu) |
9e3081ca | 7066 | { |
6c1d47c0 VG |
7067 | struct rq *rq = cpu_rq(cpu); |
7068 | struct cfs_rq *cfs_rq = &rq->cfs; | |
8a8c69c3 | 7069 | struct rq_flags rf; |
6c1d47c0 | 7070 | |
8a8c69c3 | 7071 | rq_lock_irqsave(rq, &rf); |
6c1d47c0 | 7072 | update_rq_clock(rq); |
a2c6c91f | 7073 | update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true); |
8a8c69c3 | 7074 | rq_unlock_irqrestore(rq, &rf); |
9e3081ca PZ |
7075 | } |
7076 | ||
367456c7 | 7077 | static unsigned long task_h_load(struct task_struct *p) |
1e3c88bd | 7078 | { |
9d89c257 | 7079 | return p->se.avg.load_avg; |
1e3c88bd | 7080 | } |
230059de | 7081 | #endif |
1e3c88bd | 7082 | |
1e3c88bd | 7083 | /********** Helpers for find_busiest_group ************************/ |
caeb178c RR |
7084 | |
7085 | enum group_type { | |
7086 | group_other = 0, | |
7087 | group_imbalanced, | |
7088 | group_overloaded, | |
7089 | }; | |
7090 | ||
1e3c88bd PZ |
7091 | /* |
7092 | * sg_lb_stats - stats of a sched_group required for load_balancing | |
7093 | */ | |
7094 | struct sg_lb_stats { | |
7095 | unsigned long avg_load; /*Avg load across the CPUs of the group */ | |
7096 | unsigned long group_load; /* Total load over the CPUs of the group */ | |
1e3c88bd | 7097 | unsigned long sum_weighted_load; /* Weighted load of group's tasks */ |
56cf515b | 7098 | unsigned long load_per_task; |
63b2ca30 | 7099 | unsigned long group_capacity; |
9e91d61d | 7100 | unsigned long group_util; /* Total utilization of the group */ |
147c5fc2 | 7101 | unsigned int sum_nr_running; /* Nr tasks running in the group */ |
147c5fc2 PZ |
7102 | unsigned int idle_cpus; |
7103 | unsigned int group_weight; | |
caeb178c | 7104 | enum group_type group_type; |
ea67821b | 7105 | int group_no_capacity; |
0ec8aa00 PZ |
7106 | #ifdef CONFIG_NUMA_BALANCING |
7107 | unsigned int nr_numa_running; | |
7108 | unsigned int nr_preferred_running; | |
7109 | #endif | |
1e3c88bd PZ |
7110 | }; |
7111 | ||
56cf515b JK |
7112 | /* |
7113 | * sd_lb_stats - Structure to store the statistics of a sched_domain | |
7114 | * during load balancing. | |
7115 | */ | |
7116 | struct sd_lb_stats { | |
7117 | struct sched_group *busiest; /* Busiest group in this sd */ | |
7118 | struct sched_group *local; /* Local group in this sd */ | |
7119 | unsigned long total_load; /* Total load of all groups in sd */ | |
63b2ca30 | 7120 | unsigned long total_capacity; /* Total capacity of all groups in sd */ |
56cf515b JK |
7121 | unsigned long avg_load; /* Average load across all groups in sd */ |
7122 | ||
56cf515b | 7123 | struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ |
147c5fc2 | 7124 | struct sg_lb_stats local_stat; /* Statistics of the local group */ |
56cf515b JK |
7125 | }; |
7126 | ||
147c5fc2 PZ |
7127 | static inline void init_sd_lb_stats(struct sd_lb_stats *sds) |
7128 | { | |
7129 | /* | |
7130 | * Skimp on the clearing to avoid duplicate work. We can avoid clearing | |
7131 | * local_stat because update_sg_lb_stats() does a full clear/assignment. | |
7132 | * We must however clear busiest_stat::avg_load because | |
7133 | * update_sd_pick_busiest() reads this before assignment. | |
7134 | */ | |
7135 | *sds = (struct sd_lb_stats){ | |
7136 | .busiest = NULL, | |
7137 | .local = NULL, | |
7138 | .total_load = 0UL, | |
63b2ca30 | 7139 | .total_capacity = 0UL, |
147c5fc2 PZ |
7140 | .busiest_stat = { |
7141 | .avg_load = 0UL, | |
caeb178c RR |
7142 | .sum_nr_running = 0, |
7143 | .group_type = group_other, | |
147c5fc2 PZ |
7144 | }, |
7145 | }; | |
7146 | } | |
7147 | ||
1e3c88bd PZ |
7148 | /** |
7149 | * get_sd_load_idx - Obtain the load index for a given sched domain. | |
7150 | * @sd: The sched_domain whose load_idx is to be obtained. | |
ed1b7732 | 7151 | * @idle: The idle status of the CPU for whose sd load_idx is obtained. |
e69f6186 YB |
7152 | * |
7153 | * Return: The load index. | |
1e3c88bd PZ |
7154 | */ |
7155 | static inline int get_sd_load_idx(struct sched_domain *sd, | |
7156 | enum cpu_idle_type idle) | |
7157 | { | |
7158 | int load_idx; | |
7159 | ||
7160 | switch (idle) { | |
7161 | case CPU_NOT_IDLE: | |
7162 | load_idx = sd->busy_idx; | |
7163 | break; | |
7164 | ||
7165 | case CPU_NEWLY_IDLE: | |
7166 | load_idx = sd->newidle_idx; | |
7167 | break; | |
7168 | default: | |
7169 | load_idx = sd->idle_idx; | |
7170 | break; | |
7171 | } | |
7172 | ||
7173 | return load_idx; | |
7174 | } | |
7175 | ||
ced549fa | 7176 | static unsigned long scale_rt_capacity(int cpu) |
1e3c88bd PZ |
7177 | { |
7178 | struct rq *rq = cpu_rq(cpu); | |
b5b4860d | 7179 | u64 total, used, age_stamp, avg; |
cadefd3d | 7180 | s64 delta; |
1e3c88bd | 7181 | |
b654f7de PZ |
7182 | /* |
7183 | * Since we're reading these variables without serialization make sure | |
7184 | * we read them once before doing sanity checks on them. | |
7185 | */ | |
316c1608 JL |
7186 | age_stamp = READ_ONCE(rq->age_stamp); |
7187 | avg = READ_ONCE(rq->rt_avg); | |
cebde6d6 | 7188 | delta = __rq_clock_broken(rq) - age_stamp; |
b654f7de | 7189 | |
cadefd3d PZ |
7190 | if (unlikely(delta < 0)) |
7191 | delta = 0; | |
7192 | ||
7193 | total = sched_avg_period() + delta; | |
aa483808 | 7194 | |
b5b4860d | 7195 | used = div_u64(avg, total); |
1e3c88bd | 7196 | |
b5b4860d VG |
7197 | if (likely(used < SCHED_CAPACITY_SCALE)) |
7198 | return SCHED_CAPACITY_SCALE - used; | |
1e3c88bd | 7199 | |
b5b4860d | 7200 | return 1; |
1e3c88bd PZ |
7201 | } |
7202 | ||
ced549fa | 7203 | static void update_cpu_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd | 7204 | { |
8cd5601c | 7205 | unsigned long capacity = arch_scale_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
7206 | struct sched_group *sdg = sd->groups; |
7207 | ||
ca6d75e6 | 7208 | cpu_rq(cpu)->cpu_capacity_orig = capacity; |
9d5efe05 | 7209 | |
ced549fa | 7210 | capacity *= scale_rt_capacity(cpu); |
ca8ce3d0 | 7211 | capacity >>= SCHED_CAPACITY_SHIFT; |
1e3c88bd | 7212 | |
ced549fa NP |
7213 | if (!capacity) |
7214 | capacity = 1; | |
1e3c88bd | 7215 | |
ced549fa NP |
7216 | cpu_rq(cpu)->cpu_capacity = capacity; |
7217 | sdg->sgc->capacity = capacity; | |
bf475ce0 | 7218 | sdg->sgc->min_capacity = capacity; |
1e3c88bd PZ |
7219 | } |
7220 | ||
63b2ca30 | 7221 | void update_group_capacity(struct sched_domain *sd, int cpu) |
1e3c88bd PZ |
7222 | { |
7223 | struct sched_domain *child = sd->child; | |
7224 | struct sched_group *group, *sdg = sd->groups; | |
bf475ce0 | 7225 | unsigned long capacity, min_capacity; |
4ec4412e VG |
7226 | unsigned long interval; |
7227 | ||
7228 | interval = msecs_to_jiffies(sd->balance_interval); | |
7229 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
63b2ca30 | 7230 | sdg->sgc->next_update = jiffies + interval; |
1e3c88bd PZ |
7231 | |
7232 | if (!child) { | |
ced549fa | 7233 | update_cpu_capacity(sd, cpu); |
1e3c88bd PZ |
7234 | return; |
7235 | } | |
7236 | ||
dc7ff76e | 7237 | capacity = 0; |
bf475ce0 | 7238 | min_capacity = ULONG_MAX; |
1e3c88bd | 7239 | |
74a5ce20 PZ |
7240 | if (child->flags & SD_OVERLAP) { |
7241 | /* | |
7242 | * SD_OVERLAP domains cannot assume that child groups | |
7243 | * span the current group. | |
7244 | */ | |
7245 | ||
863bffc8 | 7246 | for_each_cpu(cpu, sched_group_cpus(sdg)) { |
63b2ca30 | 7247 | struct sched_group_capacity *sgc; |
9abf24d4 | 7248 | struct rq *rq = cpu_rq(cpu); |
863bffc8 | 7249 | |
9abf24d4 | 7250 | /* |
63b2ca30 | 7251 | * build_sched_domains() -> init_sched_groups_capacity() |
9abf24d4 SD |
7252 | * gets here before we've attached the domains to the |
7253 | * runqueues. | |
7254 | * | |
ced549fa NP |
7255 | * Use capacity_of(), which is set irrespective of domains |
7256 | * in update_cpu_capacity(). | |
9abf24d4 | 7257 | * |
dc7ff76e | 7258 | * This avoids capacity from being 0 and |
9abf24d4 | 7259 | * causing divide-by-zero issues on boot. |
9abf24d4 SD |
7260 | */ |
7261 | if (unlikely(!rq->sd)) { | |
ced549fa | 7262 | capacity += capacity_of(cpu); |
bf475ce0 MR |
7263 | } else { |
7264 | sgc = rq->sd->groups->sgc; | |
7265 | capacity += sgc->capacity; | |
9abf24d4 | 7266 | } |
863bffc8 | 7267 | |
bf475ce0 | 7268 | min_capacity = min(capacity, min_capacity); |
863bffc8 | 7269 | } |
74a5ce20 PZ |
7270 | } else { |
7271 | /* | |
7272 | * !SD_OVERLAP domains can assume that child groups | |
7273 | * span the current group. | |
97a7142f | 7274 | */ |
74a5ce20 PZ |
7275 | |
7276 | group = child->groups; | |
7277 | do { | |
bf475ce0 MR |
7278 | struct sched_group_capacity *sgc = group->sgc; |
7279 | ||
7280 | capacity += sgc->capacity; | |
7281 | min_capacity = min(sgc->min_capacity, min_capacity); | |
74a5ce20 PZ |
7282 | group = group->next; |
7283 | } while (group != child->groups); | |
7284 | } | |
1e3c88bd | 7285 | |
63b2ca30 | 7286 | sdg->sgc->capacity = capacity; |
bf475ce0 | 7287 | sdg->sgc->min_capacity = min_capacity; |
1e3c88bd PZ |
7288 | } |
7289 | ||
9d5efe05 | 7290 | /* |
ea67821b VG |
7291 | * Check whether the capacity of the rq has been noticeably reduced by side |
7292 | * activity. The imbalance_pct is used for the threshold. | |
7293 | * Return true is the capacity is reduced | |
9d5efe05 SV |
7294 | */ |
7295 | static inline int | |
ea67821b | 7296 | check_cpu_capacity(struct rq *rq, struct sched_domain *sd) |
9d5efe05 | 7297 | { |
ea67821b VG |
7298 | return ((rq->cpu_capacity * sd->imbalance_pct) < |
7299 | (rq->cpu_capacity_orig * 100)); | |
9d5efe05 SV |
7300 | } |
7301 | ||
30ce5dab PZ |
7302 | /* |
7303 | * Group imbalance indicates (and tries to solve) the problem where balancing | |
0c98d344 | 7304 | * groups is inadequate due to ->cpus_allowed constraints. |
30ce5dab PZ |
7305 | * |
7306 | * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a | |
7307 | * cpumask covering 1 cpu of the first group and 3 cpus of the second group. | |
7308 | * Something like: | |
7309 | * | |
2b4d5b25 IM |
7310 | * { 0 1 2 3 } { 4 5 6 7 } |
7311 | * * * * * | |
30ce5dab PZ |
7312 | * |
7313 | * If we were to balance group-wise we'd place two tasks in the first group and | |
7314 | * two tasks in the second group. Clearly this is undesired as it will overload | |
7315 | * cpu 3 and leave one of the cpus in the second group unused. | |
7316 | * | |
7317 | * The current solution to this issue is detecting the skew in the first group | |
6263322c PZ |
7318 | * by noticing the lower domain failed to reach balance and had difficulty |
7319 | * moving tasks due to affinity constraints. | |
30ce5dab PZ |
7320 | * |
7321 | * When this is so detected; this group becomes a candidate for busiest; see | |
ed1b7732 | 7322 | * update_sd_pick_busiest(). And calculate_imbalance() and |
6263322c | 7323 | * find_busiest_group() avoid some of the usual balance conditions to allow it |
30ce5dab PZ |
7324 | * to create an effective group imbalance. |
7325 | * | |
7326 | * This is a somewhat tricky proposition since the next run might not find the | |
7327 | * group imbalance and decide the groups need to be balanced again. A most | |
7328 | * subtle and fragile situation. | |
7329 | */ | |
7330 | ||
6263322c | 7331 | static inline int sg_imbalanced(struct sched_group *group) |
30ce5dab | 7332 | { |
63b2ca30 | 7333 | return group->sgc->imbalance; |
30ce5dab PZ |
7334 | } |
7335 | ||
b37d9316 | 7336 | /* |
ea67821b VG |
7337 | * group_has_capacity returns true if the group has spare capacity that could |
7338 | * be used by some tasks. | |
7339 | * We consider that a group has spare capacity if the * number of task is | |
9e91d61d DE |
7340 | * smaller than the number of CPUs or if the utilization is lower than the |
7341 | * available capacity for CFS tasks. | |
ea67821b VG |
7342 | * For the latter, we use a threshold to stabilize the state, to take into |
7343 | * account the variance of the tasks' load and to return true if the available | |
7344 | * capacity in meaningful for the load balancer. | |
7345 | * As an example, an available capacity of 1% can appear but it doesn't make | |
7346 | * any benefit for the load balance. | |
b37d9316 | 7347 | */ |
ea67821b VG |
7348 | static inline bool |
7349 | group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs) | |
b37d9316 | 7350 | { |
ea67821b VG |
7351 | if (sgs->sum_nr_running < sgs->group_weight) |
7352 | return true; | |
c61037e9 | 7353 | |
ea67821b | 7354 | if ((sgs->group_capacity * 100) > |
9e91d61d | 7355 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 7356 | return true; |
b37d9316 | 7357 | |
ea67821b VG |
7358 | return false; |
7359 | } | |
7360 | ||
7361 | /* | |
7362 | * group_is_overloaded returns true if the group has more tasks than it can | |
7363 | * handle. | |
7364 | * group_is_overloaded is not equals to !group_has_capacity because a group | |
7365 | * with the exact right number of tasks, has no more spare capacity but is not | |
7366 | * overloaded so both group_has_capacity and group_is_overloaded return | |
7367 | * false. | |
7368 | */ | |
7369 | static inline bool | |
7370 | group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs) | |
7371 | { | |
7372 | if (sgs->sum_nr_running <= sgs->group_weight) | |
7373 | return false; | |
b37d9316 | 7374 | |
ea67821b | 7375 | if ((sgs->group_capacity * 100) < |
9e91d61d | 7376 | (sgs->group_util * env->sd->imbalance_pct)) |
ea67821b | 7377 | return true; |
b37d9316 | 7378 | |
ea67821b | 7379 | return false; |
b37d9316 PZ |
7380 | } |
7381 | ||
9e0994c0 MR |
7382 | /* |
7383 | * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller | |
7384 | * per-CPU capacity than sched_group ref. | |
7385 | */ | |
7386 | static inline bool | |
7387 | group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref) | |
7388 | { | |
7389 | return sg->sgc->min_capacity * capacity_margin < | |
7390 | ref->sgc->min_capacity * 1024; | |
7391 | } | |
7392 | ||
79a89f92 LY |
7393 | static inline enum |
7394 | group_type group_classify(struct sched_group *group, | |
7395 | struct sg_lb_stats *sgs) | |
caeb178c | 7396 | { |
ea67821b | 7397 | if (sgs->group_no_capacity) |
caeb178c RR |
7398 | return group_overloaded; |
7399 | ||
7400 | if (sg_imbalanced(group)) | |
7401 | return group_imbalanced; | |
7402 | ||
7403 | return group_other; | |
7404 | } | |
7405 | ||
1e3c88bd PZ |
7406 | /** |
7407 | * update_sg_lb_stats - Update sched_group's statistics for load balancing. | |
cd96891d | 7408 | * @env: The load balancing environment. |
1e3c88bd | 7409 | * @group: sched_group whose statistics are to be updated. |
1e3c88bd | 7410 | * @load_idx: Load index of sched_domain of this_cpu for load calc. |
1e3c88bd | 7411 | * @local_group: Does group contain this_cpu. |
1e3c88bd | 7412 | * @sgs: variable to hold the statistics for this group. |
cd3bd4e6 | 7413 | * @overload: Indicate more than one runnable task for any CPU. |
1e3c88bd | 7414 | */ |
bd939f45 PZ |
7415 | static inline void update_sg_lb_stats(struct lb_env *env, |
7416 | struct sched_group *group, int load_idx, | |
4486edd1 TC |
7417 | int local_group, struct sg_lb_stats *sgs, |
7418 | bool *overload) | |
1e3c88bd | 7419 | { |
30ce5dab | 7420 | unsigned long load; |
a426f99c | 7421 | int i, nr_running; |
1e3c88bd | 7422 | |
b72ff13c PZ |
7423 | memset(sgs, 0, sizeof(*sgs)); |
7424 | ||
b9403130 | 7425 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
1e3c88bd PZ |
7426 | struct rq *rq = cpu_rq(i); |
7427 | ||
1e3c88bd | 7428 | /* Bias balancing toward cpus of our domain */ |
6263322c | 7429 | if (local_group) |
04f733b4 | 7430 | load = target_load(i, load_idx); |
6263322c | 7431 | else |
1e3c88bd | 7432 | load = source_load(i, load_idx); |
1e3c88bd PZ |
7433 | |
7434 | sgs->group_load += load; | |
9e91d61d | 7435 | sgs->group_util += cpu_util(i); |
65fdac08 | 7436 | sgs->sum_nr_running += rq->cfs.h_nr_running; |
4486edd1 | 7437 | |
a426f99c WL |
7438 | nr_running = rq->nr_running; |
7439 | if (nr_running > 1) | |
4486edd1 TC |
7440 | *overload = true; |
7441 | ||
0ec8aa00 PZ |
7442 | #ifdef CONFIG_NUMA_BALANCING |
7443 | sgs->nr_numa_running += rq->nr_numa_running; | |
7444 | sgs->nr_preferred_running += rq->nr_preferred_running; | |
7445 | #endif | |
1e3c88bd | 7446 | sgs->sum_weighted_load += weighted_cpuload(i); |
a426f99c WL |
7447 | /* |
7448 | * No need to call idle_cpu() if nr_running is not 0 | |
7449 | */ | |
7450 | if (!nr_running && idle_cpu(i)) | |
aae6d3dd | 7451 | sgs->idle_cpus++; |
1e3c88bd PZ |
7452 | } |
7453 | ||
63b2ca30 NP |
7454 | /* Adjust by relative CPU capacity of the group */ |
7455 | sgs->group_capacity = group->sgc->capacity; | |
ca8ce3d0 | 7456 | sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; |
1e3c88bd | 7457 | |
dd5feea1 | 7458 | if (sgs->sum_nr_running) |
38d0f770 | 7459 | sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; |
1e3c88bd | 7460 | |
aae6d3dd | 7461 | sgs->group_weight = group->group_weight; |
b37d9316 | 7462 | |
ea67821b | 7463 | sgs->group_no_capacity = group_is_overloaded(env, sgs); |
79a89f92 | 7464 | sgs->group_type = group_classify(group, sgs); |
1e3c88bd PZ |
7465 | } |
7466 | ||
532cb4c4 MN |
7467 | /** |
7468 | * update_sd_pick_busiest - return 1 on busiest group | |
cd96891d | 7469 | * @env: The load balancing environment. |
532cb4c4 MN |
7470 | * @sds: sched_domain statistics |
7471 | * @sg: sched_group candidate to be checked for being the busiest | |
b6b12294 | 7472 | * @sgs: sched_group statistics |
532cb4c4 MN |
7473 | * |
7474 | * Determine if @sg is a busier group than the previously selected | |
7475 | * busiest group. | |
e69f6186 YB |
7476 | * |
7477 | * Return: %true if @sg is a busier group than the previously selected | |
7478 | * busiest group. %false otherwise. | |
532cb4c4 | 7479 | */ |
bd939f45 | 7480 | static bool update_sd_pick_busiest(struct lb_env *env, |
532cb4c4 MN |
7481 | struct sd_lb_stats *sds, |
7482 | struct sched_group *sg, | |
bd939f45 | 7483 | struct sg_lb_stats *sgs) |
532cb4c4 | 7484 | { |
caeb178c | 7485 | struct sg_lb_stats *busiest = &sds->busiest_stat; |
532cb4c4 | 7486 | |
caeb178c | 7487 | if (sgs->group_type > busiest->group_type) |
532cb4c4 MN |
7488 | return true; |
7489 | ||
caeb178c RR |
7490 | if (sgs->group_type < busiest->group_type) |
7491 | return false; | |
7492 | ||
7493 | if (sgs->avg_load <= busiest->avg_load) | |
7494 | return false; | |
7495 | ||
9e0994c0 MR |
7496 | if (!(env->sd->flags & SD_ASYM_CPUCAPACITY)) |
7497 | goto asym_packing; | |
7498 | ||
7499 | /* | |
7500 | * Candidate sg has no more than one task per CPU and | |
7501 | * has higher per-CPU capacity. Migrating tasks to less | |
7502 | * capable CPUs may harm throughput. Maximize throughput, | |
7503 | * power/energy consequences are not considered. | |
7504 | */ | |
7505 | if (sgs->sum_nr_running <= sgs->group_weight && | |
7506 | group_smaller_cpu_capacity(sds->local, sg)) | |
7507 | return false; | |
7508 | ||
7509 | asym_packing: | |
caeb178c RR |
7510 | /* This is the busiest node in its class. */ |
7511 | if (!(env->sd->flags & SD_ASYM_PACKING)) | |
532cb4c4 MN |
7512 | return true; |
7513 | ||
1f621e02 SD |
7514 | /* No ASYM_PACKING if target cpu is already busy */ |
7515 | if (env->idle == CPU_NOT_IDLE) | |
7516 | return true; | |
532cb4c4 | 7517 | /* |
afe06efd TC |
7518 | * ASYM_PACKING needs to move all the work to the highest |
7519 | * prority CPUs in the group, therefore mark all groups | |
7520 | * of lower priority than ourself as busy. | |
532cb4c4 | 7521 | */ |
afe06efd TC |
7522 | if (sgs->sum_nr_running && |
7523 | sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) { | |
532cb4c4 MN |
7524 | if (!sds->busiest) |
7525 | return true; | |
7526 | ||
afe06efd TC |
7527 | /* Prefer to move from lowest priority cpu's work */ |
7528 | if (sched_asym_prefer(sds->busiest->asym_prefer_cpu, | |
7529 | sg->asym_prefer_cpu)) | |
532cb4c4 MN |
7530 | return true; |
7531 | } | |
7532 | ||
7533 | return false; | |
7534 | } | |
7535 | ||
0ec8aa00 PZ |
7536 | #ifdef CONFIG_NUMA_BALANCING |
7537 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
7538 | { | |
7539 | if (sgs->sum_nr_running > sgs->nr_numa_running) | |
7540 | return regular; | |
7541 | if (sgs->sum_nr_running > sgs->nr_preferred_running) | |
7542 | return remote; | |
7543 | return all; | |
7544 | } | |
7545 | ||
7546 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
7547 | { | |
7548 | if (rq->nr_running > rq->nr_numa_running) | |
7549 | return regular; | |
7550 | if (rq->nr_running > rq->nr_preferred_running) | |
7551 | return remote; | |
7552 | return all; | |
7553 | } | |
7554 | #else | |
7555 | static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) | |
7556 | { | |
7557 | return all; | |
7558 | } | |
7559 | ||
7560 | static inline enum fbq_type fbq_classify_rq(struct rq *rq) | |
7561 | { | |
7562 | return regular; | |
7563 | } | |
7564 | #endif /* CONFIG_NUMA_BALANCING */ | |
7565 | ||
1e3c88bd | 7566 | /** |
461819ac | 7567 | * update_sd_lb_stats - Update sched_domain's statistics for load balancing. |
cd96891d | 7568 | * @env: The load balancing environment. |
1e3c88bd PZ |
7569 | * @sds: variable to hold the statistics for this sched_domain. |
7570 | */ | |
0ec8aa00 | 7571 | static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 7572 | { |
bd939f45 PZ |
7573 | struct sched_domain *child = env->sd->child; |
7574 | struct sched_group *sg = env->sd->groups; | |
05b40e05 | 7575 | struct sg_lb_stats *local = &sds->local_stat; |
56cf515b | 7576 | struct sg_lb_stats tmp_sgs; |
1e3c88bd | 7577 | int load_idx, prefer_sibling = 0; |
4486edd1 | 7578 | bool overload = false; |
1e3c88bd PZ |
7579 | |
7580 | if (child && child->flags & SD_PREFER_SIBLING) | |
7581 | prefer_sibling = 1; | |
7582 | ||
bd939f45 | 7583 | load_idx = get_sd_load_idx(env->sd, env->idle); |
1e3c88bd PZ |
7584 | |
7585 | do { | |
56cf515b | 7586 | struct sg_lb_stats *sgs = &tmp_sgs; |
1e3c88bd PZ |
7587 | int local_group; |
7588 | ||
bd939f45 | 7589 | local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); |
56cf515b JK |
7590 | if (local_group) { |
7591 | sds->local = sg; | |
05b40e05 | 7592 | sgs = local; |
b72ff13c PZ |
7593 | |
7594 | if (env->idle != CPU_NEWLY_IDLE || | |
63b2ca30 NP |
7595 | time_after_eq(jiffies, sg->sgc->next_update)) |
7596 | update_group_capacity(env->sd, env->dst_cpu); | |
56cf515b | 7597 | } |
1e3c88bd | 7598 | |
4486edd1 TC |
7599 | update_sg_lb_stats(env, sg, load_idx, local_group, sgs, |
7600 | &overload); | |
1e3c88bd | 7601 | |
b72ff13c PZ |
7602 | if (local_group) |
7603 | goto next_group; | |
7604 | ||
1e3c88bd PZ |
7605 | /* |
7606 | * In case the child domain prefers tasks go to siblings | |
ea67821b | 7607 | * first, lower the sg capacity so that we'll try |
75dd321d NR |
7608 | * and move all the excess tasks away. We lower the capacity |
7609 | * of a group only if the local group has the capacity to fit | |
ea67821b VG |
7610 | * these excess tasks. The extra check prevents the case where |
7611 | * you always pull from the heaviest group when it is already | |
7612 | * under-utilized (possible with a large weight task outweighs | |
7613 | * the tasks on the system). | |
1e3c88bd | 7614 | */ |
b72ff13c | 7615 | if (prefer_sibling && sds->local && |
05b40e05 SD |
7616 | group_has_capacity(env, local) && |
7617 | (sgs->sum_nr_running > local->sum_nr_running + 1)) { | |
ea67821b | 7618 | sgs->group_no_capacity = 1; |
79a89f92 | 7619 | sgs->group_type = group_classify(sg, sgs); |
cb0b9f24 | 7620 | } |
1e3c88bd | 7621 | |
b72ff13c | 7622 | if (update_sd_pick_busiest(env, sds, sg, sgs)) { |
532cb4c4 | 7623 | sds->busiest = sg; |
56cf515b | 7624 | sds->busiest_stat = *sgs; |
1e3c88bd PZ |
7625 | } |
7626 | ||
b72ff13c PZ |
7627 | next_group: |
7628 | /* Now, start updating sd_lb_stats */ | |
7629 | sds->total_load += sgs->group_load; | |
63b2ca30 | 7630 | sds->total_capacity += sgs->group_capacity; |
b72ff13c | 7631 | |
532cb4c4 | 7632 | sg = sg->next; |
bd939f45 | 7633 | } while (sg != env->sd->groups); |
0ec8aa00 PZ |
7634 | |
7635 | if (env->sd->flags & SD_NUMA) | |
7636 | env->fbq_type = fbq_classify_group(&sds->busiest_stat); | |
4486edd1 TC |
7637 | |
7638 | if (!env->sd->parent) { | |
7639 | /* update overload indicator if we are at root domain */ | |
7640 | if (env->dst_rq->rd->overload != overload) | |
7641 | env->dst_rq->rd->overload = overload; | |
7642 | } | |
7643 | ||
532cb4c4 MN |
7644 | } |
7645 | ||
532cb4c4 MN |
7646 | /** |
7647 | * check_asym_packing - Check to see if the group is packed into the | |
7648 | * sched doman. | |
7649 | * | |
7650 | * This is primarily intended to used at the sibling level. Some | |
7651 | * cores like POWER7 prefer to use lower numbered SMT threads. In the | |
7652 | * case of POWER7, it can move to lower SMT modes only when higher | |
7653 | * threads are idle. When in lower SMT modes, the threads will | |
7654 | * perform better since they share less core resources. Hence when we | |
7655 | * have idle threads, we want them to be the higher ones. | |
7656 | * | |
7657 | * This packing function is run on idle threads. It checks to see if | |
7658 | * the busiest CPU in this domain (core in the P7 case) has a higher | |
7659 | * CPU number than the packing function is being run on. Here we are | |
7660 | * assuming lower CPU number will be equivalent to lower a SMT thread | |
7661 | * number. | |
7662 | * | |
e69f6186 | 7663 | * Return: 1 when packing is required and a task should be moved to |
b6b12294 MN |
7664 | * this CPU. The amount of the imbalance is returned in *imbalance. |
7665 | * | |
cd96891d | 7666 | * @env: The load balancing environment. |
532cb4c4 | 7667 | * @sds: Statistics of the sched_domain which is to be packed |
532cb4c4 | 7668 | */ |
bd939f45 | 7669 | static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) |
532cb4c4 MN |
7670 | { |
7671 | int busiest_cpu; | |
7672 | ||
bd939f45 | 7673 | if (!(env->sd->flags & SD_ASYM_PACKING)) |
532cb4c4 MN |
7674 | return 0; |
7675 | ||
1f621e02 SD |
7676 | if (env->idle == CPU_NOT_IDLE) |
7677 | return 0; | |
7678 | ||
532cb4c4 MN |
7679 | if (!sds->busiest) |
7680 | return 0; | |
7681 | ||
afe06efd TC |
7682 | busiest_cpu = sds->busiest->asym_prefer_cpu; |
7683 | if (sched_asym_prefer(busiest_cpu, env->dst_cpu)) | |
532cb4c4 MN |
7684 | return 0; |
7685 | ||
bd939f45 | 7686 | env->imbalance = DIV_ROUND_CLOSEST( |
63b2ca30 | 7687 | sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity, |
ca8ce3d0 | 7688 | SCHED_CAPACITY_SCALE); |
bd939f45 | 7689 | |
532cb4c4 | 7690 | return 1; |
1e3c88bd PZ |
7691 | } |
7692 | ||
7693 | /** | |
7694 | * fix_small_imbalance - Calculate the minor imbalance that exists | |
7695 | * amongst the groups of a sched_domain, during | |
7696 | * load balancing. | |
cd96891d | 7697 | * @env: The load balancing environment. |
1e3c88bd | 7698 | * @sds: Statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 7699 | */ |
bd939f45 PZ |
7700 | static inline |
7701 | void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) | |
1e3c88bd | 7702 | { |
63b2ca30 | 7703 | unsigned long tmp, capa_now = 0, capa_move = 0; |
1e3c88bd | 7704 | unsigned int imbn = 2; |
dd5feea1 | 7705 | unsigned long scaled_busy_load_per_task; |
56cf515b | 7706 | struct sg_lb_stats *local, *busiest; |
1e3c88bd | 7707 | |
56cf515b JK |
7708 | local = &sds->local_stat; |
7709 | busiest = &sds->busiest_stat; | |
1e3c88bd | 7710 | |
56cf515b JK |
7711 | if (!local->sum_nr_running) |
7712 | local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); | |
7713 | else if (busiest->load_per_task > local->load_per_task) | |
7714 | imbn = 1; | |
dd5feea1 | 7715 | |
56cf515b | 7716 | scaled_busy_load_per_task = |
ca8ce3d0 | 7717 | (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 7718 | busiest->group_capacity; |
56cf515b | 7719 | |
3029ede3 VD |
7720 | if (busiest->avg_load + scaled_busy_load_per_task >= |
7721 | local->avg_load + (scaled_busy_load_per_task * imbn)) { | |
56cf515b | 7722 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
7723 | return; |
7724 | } | |
7725 | ||
7726 | /* | |
7727 | * OK, we don't have enough imbalance to justify moving tasks, | |
ced549fa | 7728 | * however we may be able to increase total CPU capacity used by |
1e3c88bd PZ |
7729 | * moving them. |
7730 | */ | |
7731 | ||
63b2ca30 | 7732 | capa_now += busiest->group_capacity * |
56cf515b | 7733 | min(busiest->load_per_task, busiest->avg_load); |
63b2ca30 | 7734 | capa_now += local->group_capacity * |
56cf515b | 7735 | min(local->load_per_task, local->avg_load); |
ca8ce3d0 | 7736 | capa_now /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
7737 | |
7738 | /* Amount of load we'd subtract */ | |
a2cd4260 | 7739 | if (busiest->avg_load > scaled_busy_load_per_task) { |
63b2ca30 | 7740 | capa_move += busiest->group_capacity * |
56cf515b | 7741 | min(busiest->load_per_task, |
a2cd4260 | 7742 | busiest->avg_load - scaled_busy_load_per_task); |
56cf515b | 7743 | } |
1e3c88bd PZ |
7744 | |
7745 | /* Amount of load we'd add */ | |
63b2ca30 | 7746 | if (busiest->avg_load * busiest->group_capacity < |
ca8ce3d0 | 7747 | busiest->load_per_task * SCHED_CAPACITY_SCALE) { |
63b2ca30 NP |
7748 | tmp = (busiest->avg_load * busiest->group_capacity) / |
7749 | local->group_capacity; | |
56cf515b | 7750 | } else { |
ca8ce3d0 | 7751 | tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / |
63b2ca30 | 7752 | local->group_capacity; |
56cf515b | 7753 | } |
63b2ca30 | 7754 | capa_move += local->group_capacity * |
3ae11c90 | 7755 | min(local->load_per_task, local->avg_load + tmp); |
ca8ce3d0 | 7756 | capa_move /= SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
7757 | |
7758 | /* Move if we gain throughput */ | |
63b2ca30 | 7759 | if (capa_move > capa_now) |
56cf515b | 7760 | env->imbalance = busiest->load_per_task; |
1e3c88bd PZ |
7761 | } |
7762 | ||
7763 | /** | |
7764 | * calculate_imbalance - Calculate the amount of imbalance present within the | |
7765 | * groups of a given sched_domain during load balance. | |
bd939f45 | 7766 | * @env: load balance environment |
1e3c88bd | 7767 | * @sds: statistics of the sched_domain whose imbalance is to be calculated. |
1e3c88bd | 7768 | */ |
bd939f45 | 7769 | static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) |
1e3c88bd | 7770 | { |
dd5feea1 | 7771 | unsigned long max_pull, load_above_capacity = ~0UL; |
56cf515b JK |
7772 | struct sg_lb_stats *local, *busiest; |
7773 | ||
7774 | local = &sds->local_stat; | |
56cf515b | 7775 | busiest = &sds->busiest_stat; |
dd5feea1 | 7776 | |
caeb178c | 7777 | if (busiest->group_type == group_imbalanced) { |
30ce5dab PZ |
7778 | /* |
7779 | * In the group_imb case we cannot rely on group-wide averages | |
7780 | * to ensure cpu-load equilibrium, look at wider averages. XXX | |
7781 | */ | |
56cf515b JK |
7782 | busiest->load_per_task = |
7783 | min(busiest->load_per_task, sds->avg_load); | |
dd5feea1 SS |
7784 | } |
7785 | ||
1e3c88bd | 7786 | /* |
885e542c DE |
7787 | * Avg load of busiest sg can be less and avg load of local sg can |
7788 | * be greater than avg load across all sgs of sd because avg load | |
7789 | * factors in sg capacity and sgs with smaller group_type are | |
7790 | * skipped when updating the busiest sg: | |
1e3c88bd | 7791 | */ |
b1885550 VD |
7792 | if (busiest->avg_load <= sds->avg_load || |
7793 | local->avg_load >= sds->avg_load) { | |
bd939f45 PZ |
7794 | env->imbalance = 0; |
7795 | return fix_small_imbalance(env, sds); | |
1e3c88bd PZ |
7796 | } |
7797 | ||
9a5d9ba6 PZ |
7798 | /* |
7799 | * If there aren't any idle cpus, avoid creating some. | |
7800 | */ | |
7801 | if (busiest->group_type == group_overloaded && | |
7802 | local->group_type == group_overloaded) { | |
1be0eb2a | 7803 | load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE; |
cfa10334 | 7804 | if (load_above_capacity > busiest->group_capacity) { |
ea67821b | 7805 | load_above_capacity -= busiest->group_capacity; |
26656215 | 7806 | load_above_capacity *= scale_load_down(NICE_0_LOAD); |
cfa10334 MR |
7807 | load_above_capacity /= busiest->group_capacity; |
7808 | } else | |
ea67821b | 7809 | load_above_capacity = ~0UL; |
dd5feea1 SS |
7810 | } |
7811 | ||
7812 | /* | |
7813 | * We're trying to get all the cpus to the average_load, so we don't | |
7814 | * want to push ourselves above the average load, nor do we wish to | |
7815 | * reduce the max loaded cpu below the average load. At the same time, | |
0a9b23ce DE |
7816 | * we also don't want to reduce the group load below the group |
7817 | * capacity. Thus we look for the minimum possible imbalance. | |
dd5feea1 | 7818 | */ |
30ce5dab | 7819 | max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); |
1e3c88bd PZ |
7820 | |
7821 | /* How much load to actually move to equalise the imbalance */ | |
56cf515b | 7822 | env->imbalance = min( |
63b2ca30 NP |
7823 | max_pull * busiest->group_capacity, |
7824 | (sds->avg_load - local->avg_load) * local->group_capacity | |
ca8ce3d0 | 7825 | ) / SCHED_CAPACITY_SCALE; |
1e3c88bd PZ |
7826 | |
7827 | /* | |
7828 | * if *imbalance is less than the average load per runnable task | |
25985edc | 7829 | * there is no guarantee that any tasks will be moved so we'll have |
1e3c88bd PZ |
7830 | * a think about bumping its value to force at least one task to be |
7831 | * moved | |
7832 | */ | |
56cf515b | 7833 | if (env->imbalance < busiest->load_per_task) |
bd939f45 | 7834 | return fix_small_imbalance(env, sds); |
1e3c88bd | 7835 | } |
fab47622 | 7836 | |
1e3c88bd PZ |
7837 | /******* find_busiest_group() helpers end here *********************/ |
7838 | ||
7839 | /** | |
7840 | * find_busiest_group - Returns the busiest group within the sched_domain | |
0a9b23ce | 7841 | * if there is an imbalance. |
1e3c88bd PZ |
7842 | * |
7843 | * Also calculates the amount of weighted load which should be moved | |
7844 | * to restore balance. | |
7845 | * | |
cd96891d | 7846 | * @env: The load balancing environment. |
1e3c88bd | 7847 | * |
e69f6186 | 7848 | * Return: - The busiest group if imbalance exists. |
1e3c88bd | 7849 | */ |
56cf515b | 7850 | static struct sched_group *find_busiest_group(struct lb_env *env) |
1e3c88bd | 7851 | { |
56cf515b | 7852 | struct sg_lb_stats *local, *busiest; |
1e3c88bd PZ |
7853 | struct sd_lb_stats sds; |
7854 | ||
147c5fc2 | 7855 | init_sd_lb_stats(&sds); |
1e3c88bd PZ |
7856 | |
7857 | /* | |
7858 | * Compute the various statistics relavent for load balancing at | |
7859 | * this level. | |
7860 | */ | |
23f0d209 | 7861 | update_sd_lb_stats(env, &sds); |
56cf515b JK |
7862 | local = &sds.local_stat; |
7863 | busiest = &sds.busiest_stat; | |
1e3c88bd | 7864 | |
ea67821b | 7865 | /* ASYM feature bypasses nice load balance check */ |
1f621e02 | 7866 | if (check_asym_packing(env, &sds)) |
532cb4c4 MN |
7867 | return sds.busiest; |
7868 | ||
cc57aa8f | 7869 | /* There is no busy sibling group to pull tasks from */ |
56cf515b | 7870 | if (!sds.busiest || busiest->sum_nr_running == 0) |
1e3c88bd PZ |
7871 | goto out_balanced; |
7872 | ||
ca8ce3d0 NP |
7873 | sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) |
7874 | / sds.total_capacity; | |
b0432d8f | 7875 | |
866ab43e PZ |
7876 | /* |
7877 | * If the busiest group is imbalanced the below checks don't | |
30ce5dab | 7878 | * work because they assume all things are equal, which typically |
866ab43e PZ |
7879 | * isn't true due to cpus_allowed constraints and the like. |
7880 | */ | |
caeb178c | 7881 | if (busiest->group_type == group_imbalanced) |
866ab43e PZ |
7882 | goto force_balance; |
7883 | ||
cc57aa8f | 7884 | /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ |
ea67821b VG |
7885 | if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) && |
7886 | busiest->group_no_capacity) | |
fab47622 NR |
7887 | goto force_balance; |
7888 | ||
cc57aa8f | 7889 | /* |
9c58c79a | 7890 | * If the local group is busier than the selected busiest group |
cc57aa8f PZ |
7891 | * don't try and pull any tasks. |
7892 | */ | |
56cf515b | 7893 | if (local->avg_load >= busiest->avg_load) |
1e3c88bd PZ |
7894 | goto out_balanced; |
7895 | ||
cc57aa8f PZ |
7896 | /* |
7897 | * Don't pull any tasks if this group is already above the domain | |
7898 | * average load. | |
7899 | */ | |
56cf515b | 7900 | if (local->avg_load >= sds.avg_load) |
1e3c88bd PZ |
7901 | goto out_balanced; |
7902 | ||
bd939f45 | 7903 | if (env->idle == CPU_IDLE) { |
aae6d3dd | 7904 | /* |
43f4d666 VG |
7905 | * This cpu is idle. If the busiest group is not overloaded |
7906 | * and there is no imbalance between this and busiest group | |
7907 | * wrt idle cpus, it is balanced. The imbalance becomes | |
7908 | * significant if the diff is greater than 1 otherwise we | |
7909 | * might end up to just move the imbalance on another group | |
aae6d3dd | 7910 | */ |
43f4d666 VG |
7911 | if ((busiest->group_type != group_overloaded) && |
7912 | (local->idle_cpus <= (busiest->idle_cpus + 1))) | |
aae6d3dd | 7913 | goto out_balanced; |
c186fafe PZ |
7914 | } else { |
7915 | /* | |
7916 | * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use | |
7917 | * imbalance_pct to be conservative. | |
7918 | */ | |
56cf515b JK |
7919 | if (100 * busiest->avg_load <= |
7920 | env->sd->imbalance_pct * local->avg_load) | |
c186fafe | 7921 | goto out_balanced; |
aae6d3dd | 7922 | } |
1e3c88bd | 7923 | |
fab47622 | 7924 | force_balance: |
1e3c88bd | 7925 | /* Looks like there is an imbalance. Compute it */ |
bd939f45 | 7926 | calculate_imbalance(env, &sds); |
1e3c88bd PZ |
7927 | return sds.busiest; |
7928 | ||
7929 | out_balanced: | |
bd939f45 | 7930 | env->imbalance = 0; |
1e3c88bd PZ |
7931 | return NULL; |
7932 | } | |
7933 | ||
7934 | /* | |
7935 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
7936 | */ | |
bd939f45 | 7937 | static struct rq *find_busiest_queue(struct lb_env *env, |
b9403130 | 7938 | struct sched_group *group) |
1e3c88bd PZ |
7939 | { |
7940 | struct rq *busiest = NULL, *rq; | |
ced549fa | 7941 | unsigned long busiest_load = 0, busiest_capacity = 1; |
1e3c88bd PZ |
7942 | int i; |
7943 | ||
6906a408 | 7944 | for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { |
ea67821b | 7945 | unsigned long capacity, wl; |
0ec8aa00 PZ |
7946 | enum fbq_type rt; |
7947 | ||
7948 | rq = cpu_rq(i); | |
7949 | rt = fbq_classify_rq(rq); | |
1e3c88bd | 7950 | |
0ec8aa00 PZ |
7951 | /* |
7952 | * We classify groups/runqueues into three groups: | |
7953 | * - regular: there are !numa tasks | |
7954 | * - remote: there are numa tasks that run on the 'wrong' node | |
7955 | * - all: there is no distinction | |
7956 | * | |
7957 | * In order to avoid migrating ideally placed numa tasks, | |
7958 | * ignore those when there's better options. | |
7959 | * | |
7960 | * If we ignore the actual busiest queue to migrate another | |
7961 | * task, the next balance pass can still reduce the busiest | |
7962 | * queue by moving tasks around inside the node. | |
7963 | * | |
7964 | * If we cannot move enough load due to this classification | |
7965 | * the next pass will adjust the group classification and | |
7966 | * allow migration of more tasks. | |
7967 | * | |
7968 | * Both cases only affect the total convergence complexity. | |
7969 | */ | |
7970 | if (rt > env->fbq_type) | |
7971 | continue; | |
7972 | ||
ced549fa | 7973 | capacity = capacity_of(i); |
9d5efe05 | 7974 | |
6e40f5bb | 7975 | wl = weighted_cpuload(i); |
1e3c88bd | 7976 | |
6e40f5bb TG |
7977 | /* |
7978 | * When comparing with imbalance, use weighted_cpuload() | |
ced549fa | 7979 | * which is not scaled with the cpu capacity. |
6e40f5bb | 7980 | */ |
ea67821b VG |
7981 | |
7982 | if (rq->nr_running == 1 && wl > env->imbalance && | |
7983 | !check_cpu_capacity(rq, env->sd)) | |
1e3c88bd PZ |
7984 | continue; |
7985 | ||
6e40f5bb TG |
7986 | /* |
7987 | * For the load comparisons with the other cpu's, consider | |
ced549fa NP |
7988 | * the weighted_cpuload() scaled with the cpu capacity, so |
7989 | * that the load can be moved away from the cpu that is | |
7990 | * potentially running at a lower capacity. | |
95a79b80 | 7991 | * |
ced549fa | 7992 | * Thus we're looking for max(wl_i / capacity_i), crosswise |
95a79b80 | 7993 | * multiplication to rid ourselves of the division works out |
ced549fa NP |
7994 | * to: wl_i * capacity_j > wl_j * capacity_i; where j is |
7995 | * our previous maximum. | |
6e40f5bb | 7996 | */ |
ced549fa | 7997 | if (wl * busiest_capacity > busiest_load * capacity) { |
95a79b80 | 7998 | busiest_load = wl; |
ced549fa | 7999 | busiest_capacity = capacity; |
1e3c88bd PZ |
8000 | busiest = rq; |
8001 | } | |
8002 | } | |
8003 | ||
8004 | return busiest; | |
8005 | } | |
8006 | ||
8007 | /* | |
8008 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
8009 | * so long as it is large enough. | |
8010 | */ | |
8011 | #define MAX_PINNED_INTERVAL 512 | |
8012 | ||
bd939f45 | 8013 | static int need_active_balance(struct lb_env *env) |
1af3ed3d | 8014 | { |
bd939f45 PZ |
8015 | struct sched_domain *sd = env->sd; |
8016 | ||
8017 | if (env->idle == CPU_NEWLY_IDLE) { | |
532cb4c4 MN |
8018 | |
8019 | /* | |
8020 | * ASYM_PACKING needs to force migrate tasks from busy but | |
afe06efd TC |
8021 | * lower priority CPUs in order to pack all tasks in the |
8022 | * highest priority CPUs. | |
532cb4c4 | 8023 | */ |
afe06efd TC |
8024 | if ((sd->flags & SD_ASYM_PACKING) && |
8025 | sched_asym_prefer(env->dst_cpu, env->src_cpu)) | |
532cb4c4 | 8026 | return 1; |
1af3ed3d PZ |
8027 | } |
8028 | ||
1aaf90a4 VG |
8029 | /* |
8030 | * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. | |
8031 | * It's worth migrating the task if the src_cpu's capacity is reduced | |
8032 | * because of other sched_class or IRQs if more capacity stays | |
8033 | * available on dst_cpu. | |
8034 | */ | |
8035 | if ((env->idle != CPU_NOT_IDLE) && | |
8036 | (env->src_rq->cfs.h_nr_running == 1)) { | |
8037 | if ((check_cpu_capacity(env->src_rq, sd)) && | |
8038 | (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) | |
8039 | return 1; | |
8040 | } | |
8041 | ||
1af3ed3d PZ |
8042 | return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); |
8043 | } | |
8044 | ||
969c7921 TH |
8045 | static int active_load_balance_cpu_stop(void *data); |
8046 | ||
23f0d209 JK |
8047 | static int should_we_balance(struct lb_env *env) |
8048 | { | |
8049 | struct sched_group *sg = env->sd->groups; | |
8050 | struct cpumask *sg_cpus, *sg_mask; | |
8051 | int cpu, balance_cpu = -1; | |
8052 | ||
8053 | /* | |
8054 | * In the newly idle case, we will allow all the cpu's | |
8055 | * to do the newly idle load balance. | |
8056 | */ | |
8057 | if (env->idle == CPU_NEWLY_IDLE) | |
8058 | return 1; | |
8059 | ||
8060 | sg_cpus = sched_group_cpus(sg); | |
8061 | sg_mask = sched_group_mask(sg); | |
8062 | /* Try to find first idle cpu */ | |
8063 | for_each_cpu_and(cpu, sg_cpus, env->cpus) { | |
8064 | if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) | |
8065 | continue; | |
8066 | ||
8067 | balance_cpu = cpu; | |
8068 | break; | |
8069 | } | |
8070 | ||
8071 | if (balance_cpu == -1) | |
8072 | balance_cpu = group_balance_cpu(sg); | |
8073 | ||
8074 | /* | |
8075 | * First idle cpu or the first cpu(busiest) in this sched group | |
8076 | * is eligible for doing load balancing at this and above domains. | |
8077 | */ | |
b0cff9d8 | 8078 | return balance_cpu == env->dst_cpu; |
23f0d209 JK |
8079 | } |
8080 | ||
1e3c88bd PZ |
8081 | /* |
8082 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
8083 | * tasks if there is an imbalance. | |
8084 | */ | |
8085 | static int load_balance(int this_cpu, struct rq *this_rq, | |
8086 | struct sched_domain *sd, enum cpu_idle_type idle, | |
23f0d209 | 8087 | int *continue_balancing) |
1e3c88bd | 8088 | { |
88b8dac0 | 8089 | int ld_moved, cur_ld_moved, active_balance = 0; |
6263322c | 8090 | struct sched_domain *sd_parent = sd->parent; |
1e3c88bd | 8091 | struct sched_group *group; |
1e3c88bd | 8092 | struct rq *busiest; |
8a8c69c3 | 8093 | struct rq_flags rf; |
4ba29684 | 8094 | struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); |
1e3c88bd | 8095 | |
8e45cb54 PZ |
8096 | struct lb_env env = { |
8097 | .sd = sd, | |
ddcdf6e7 PZ |
8098 | .dst_cpu = this_cpu, |
8099 | .dst_rq = this_rq, | |
88b8dac0 | 8100 | .dst_grpmask = sched_group_cpus(sd->groups), |
8e45cb54 | 8101 | .idle = idle, |
eb95308e | 8102 | .loop_break = sched_nr_migrate_break, |
b9403130 | 8103 | .cpus = cpus, |
0ec8aa00 | 8104 | .fbq_type = all, |
163122b7 | 8105 | .tasks = LIST_HEAD_INIT(env.tasks), |
8e45cb54 PZ |
8106 | }; |
8107 | ||
cfc03118 JK |
8108 | /* |
8109 | * For NEWLY_IDLE load_balancing, we don't need to consider | |
8110 | * other cpus in our group | |
8111 | */ | |
e02e60c1 | 8112 | if (idle == CPU_NEWLY_IDLE) |
cfc03118 | 8113 | env.dst_grpmask = NULL; |
cfc03118 | 8114 | |
1e3c88bd PZ |
8115 | cpumask_copy(cpus, cpu_active_mask); |
8116 | ||
ae92882e | 8117 | schedstat_inc(sd->lb_count[idle]); |
1e3c88bd PZ |
8118 | |
8119 | redo: | |
23f0d209 JK |
8120 | if (!should_we_balance(&env)) { |
8121 | *continue_balancing = 0; | |
1e3c88bd | 8122 | goto out_balanced; |
23f0d209 | 8123 | } |
1e3c88bd | 8124 | |
23f0d209 | 8125 | group = find_busiest_group(&env); |
1e3c88bd | 8126 | if (!group) { |
ae92882e | 8127 | schedstat_inc(sd->lb_nobusyg[idle]); |
1e3c88bd PZ |
8128 | goto out_balanced; |
8129 | } | |
8130 | ||
b9403130 | 8131 | busiest = find_busiest_queue(&env, group); |
1e3c88bd | 8132 | if (!busiest) { |
ae92882e | 8133 | schedstat_inc(sd->lb_nobusyq[idle]); |
1e3c88bd PZ |
8134 | goto out_balanced; |
8135 | } | |
8136 | ||
78feefc5 | 8137 | BUG_ON(busiest == env.dst_rq); |
1e3c88bd | 8138 | |
ae92882e | 8139 | schedstat_add(sd->lb_imbalance[idle], env.imbalance); |
1e3c88bd | 8140 | |
1aaf90a4 VG |
8141 | env.src_cpu = busiest->cpu; |
8142 | env.src_rq = busiest; | |
8143 | ||
1e3c88bd PZ |
8144 | ld_moved = 0; |
8145 | if (busiest->nr_running > 1) { | |
8146 | /* | |
8147 | * Attempt to move tasks. If find_busiest_group has found | |
8148 | * an imbalance but busiest->nr_running <= 1, the group is | |
8149 | * still unbalanced. ld_moved simply stays zero, so it is | |
8150 | * correctly treated as an imbalance. | |
8151 | */ | |
8e45cb54 | 8152 | env.flags |= LBF_ALL_PINNED; |
c82513e5 | 8153 | env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); |
8e45cb54 | 8154 | |
5d6523eb | 8155 | more_balance: |
8a8c69c3 | 8156 | rq_lock_irqsave(busiest, &rf); |
3bed5e21 | 8157 | update_rq_clock(busiest); |
88b8dac0 SV |
8158 | |
8159 | /* | |
8160 | * cur_ld_moved - load moved in current iteration | |
8161 | * ld_moved - cumulative load moved across iterations | |
8162 | */ | |
163122b7 | 8163 | cur_ld_moved = detach_tasks(&env); |
1e3c88bd PZ |
8164 | |
8165 | /* | |
163122b7 KT |
8166 | * We've detached some tasks from busiest_rq. Every |
8167 | * task is masked "TASK_ON_RQ_MIGRATING", so we can safely | |
8168 | * unlock busiest->lock, and we are able to be sure | |
8169 | * that nobody can manipulate the tasks in parallel. | |
8170 | * See task_rq_lock() family for the details. | |
1e3c88bd | 8171 | */ |
163122b7 | 8172 | |
8a8c69c3 | 8173 | rq_unlock(busiest, &rf); |
163122b7 KT |
8174 | |
8175 | if (cur_ld_moved) { | |
8176 | attach_tasks(&env); | |
8177 | ld_moved += cur_ld_moved; | |
8178 | } | |
8179 | ||
8a8c69c3 | 8180 | local_irq_restore(rf.flags); |
88b8dac0 | 8181 | |
f1cd0858 JK |
8182 | if (env.flags & LBF_NEED_BREAK) { |
8183 | env.flags &= ~LBF_NEED_BREAK; | |
8184 | goto more_balance; | |
8185 | } | |
8186 | ||
88b8dac0 SV |
8187 | /* |
8188 | * Revisit (affine) tasks on src_cpu that couldn't be moved to | |
8189 | * us and move them to an alternate dst_cpu in our sched_group | |
8190 | * where they can run. The upper limit on how many times we | |
8191 | * iterate on same src_cpu is dependent on number of cpus in our | |
8192 | * sched_group. | |
8193 | * | |
8194 | * This changes load balance semantics a bit on who can move | |
8195 | * load to a given_cpu. In addition to the given_cpu itself | |
8196 | * (or a ilb_cpu acting on its behalf where given_cpu is | |
8197 | * nohz-idle), we now have balance_cpu in a position to move | |
8198 | * load to given_cpu. In rare situations, this may cause | |
8199 | * conflicts (balance_cpu and given_cpu/ilb_cpu deciding | |
8200 | * _independently_ and at _same_ time to move some load to | |
8201 | * given_cpu) causing exceess load to be moved to given_cpu. | |
8202 | * This however should not happen so much in practice and | |
8203 | * moreover subsequent load balance cycles should correct the | |
8204 | * excess load moved. | |
8205 | */ | |
6263322c | 8206 | if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { |
88b8dac0 | 8207 | |
7aff2e3a VD |
8208 | /* Prevent to re-select dst_cpu via env's cpus */ |
8209 | cpumask_clear_cpu(env.dst_cpu, env.cpus); | |
8210 | ||
78feefc5 | 8211 | env.dst_rq = cpu_rq(env.new_dst_cpu); |
88b8dac0 | 8212 | env.dst_cpu = env.new_dst_cpu; |
6263322c | 8213 | env.flags &= ~LBF_DST_PINNED; |
88b8dac0 SV |
8214 | env.loop = 0; |
8215 | env.loop_break = sched_nr_migrate_break; | |
e02e60c1 | 8216 | |
88b8dac0 SV |
8217 | /* |
8218 | * Go back to "more_balance" rather than "redo" since we | |
8219 | * need to continue with same src_cpu. | |
8220 | */ | |
8221 | goto more_balance; | |
8222 | } | |
1e3c88bd | 8223 | |
6263322c PZ |
8224 | /* |
8225 | * We failed to reach balance because of affinity. | |
8226 | */ | |
8227 | if (sd_parent) { | |
63b2ca30 | 8228 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; |
6263322c | 8229 | |
afdeee05 | 8230 | if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) |
6263322c | 8231 | *group_imbalance = 1; |
6263322c PZ |
8232 | } |
8233 | ||
1e3c88bd | 8234 | /* All tasks on this runqueue were pinned by CPU affinity */ |
8e45cb54 | 8235 | if (unlikely(env.flags & LBF_ALL_PINNED)) { |
1e3c88bd | 8236 | cpumask_clear_cpu(cpu_of(busiest), cpus); |
bbf18b19 PN |
8237 | if (!cpumask_empty(cpus)) { |
8238 | env.loop = 0; | |
8239 | env.loop_break = sched_nr_migrate_break; | |
1e3c88bd | 8240 | goto redo; |
bbf18b19 | 8241 | } |
afdeee05 | 8242 | goto out_all_pinned; |
1e3c88bd PZ |
8243 | } |
8244 | } | |
8245 | ||
8246 | if (!ld_moved) { | |
ae92882e | 8247 | schedstat_inc(sd->lb_failed[idle]); |
58b26c4c VP |
8248 | /* |
8249 | * Increment the failure counter only on periodic balance. | |
8250 | * We do not want newidle balance, which can be very | |
8251 | * frequent, pollute the failure counter causing | |
8252 | * excessive cache_hot migrations and active balances. | |
8253 | */ | |
8254 | if (idle != CPU_NEWLY_IDLE) | |
8255 | sd->nr_balance_failed++; | |
1e3c88bd | 8256 | |
bd939f45 | 8257 | if (need_active_balance(&env)) { |
8a8c69c3 PZ |
8258 | unsigned long flags; |
8259 | ||
1e3c88bd PZ |
8260 | raw_spin_lock_irqsave(&busiest->lock, flags); |
8261 | ||
969c7921 TH |
8262 | /* don't kick the active_load_balance_cpu_stop, |
8263 | * if the curr task on busiest cpu can't be | |
8264 | * moved to this_cpu | |
1e3c88bd | 8265 | */ |
0c98d344 | 8266 | if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) { |
1e3c88bd PZ |
8267 | raw_spin_unlock_irqrestore(&busiest->lock, |
8268 | flags); | |
8e45cb54 | 8269 | env.flags |= LBF_ALL_PINNED; |
1e3c88bd PZ |
8270 | goto out_one_pinned; |
8271 | } | |
8272 | ||
969c7921 TH |
8273 | /* |
8274 | * ->active_balance synchronizes accesses to | |
8275 | * ->active_balance_work. Once set, it's cleared | |
8276 | * only after active load balance is finished. | |
8277 | */ | |
1e3c88bd PZ |
8278 | if (!busiest->active_balance) { |
8279 | busiest->active_balance = 1; | |
8280 | busiest->push_cpu = this_cpu; | |
8281 | active_balance = 1; | |
8282 | } | |
8283 | raw_spin_unlock_irqrestore(&busiest->lock, flags); | |
969c7921 | 8284 | |
bd939f45 | 8285 | if (active_balance) { |
969c7921 TH |
8286 | stop_one_cpu_nowait(cpu_of(busiest), |
8287 | active_load_balance_cpu_stop, busiest, | |
8288 | &busiest->active_balance_work); | |
bd939f45 | 8289 | } |
1e3c88bd | 8290 | |
d02c0711 | 8291 | /* We've kicked active balancing, force task migration. */ |
1e3c88bd PZ |
8292 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
8293 | } | |
8294 | } else | |
8295 | sd->nr_balance_failed = 0; | |
8296 | ||
8297 | if (likely(!active_balance)) { | |
8298 | /* We were unbalanced, so reset the balancing interval */ | |
8299 | sd->balance_interval = sd->min_interval; | |
8300 | } else { | |
8301 | /* | |
8302 | * If we've begun active balancing, start to back off. This | |
8303 | * case may not be covered by the all_pinned logic if there | |
8304 | * is only 1 task on the busy runqueue (because we don't call | |
163122b7 | 8305 | * detach_tasks). |
1e3c88bd PZ |
8306 | */ |
8307 | if (sd->balance_interval < sd->max_interval) | |
8308 | sd->balance_interval *= 2; | |
8309 | } | |
8310 | ||
1e3c88bd PZ |
8311 | goto out; |
8312 | ||
8313 | out_balanced: | |
afdeee05 VG |
8314 | /* |
8315 | * We reach balance although we may have faced some affinity | |
8316 | * constraints. Clear the imbalance flag if it was set. | |
8317 | */ | |
8318 | if (sd_parent) { | |
8319 | int *group_imbalance = &sd_parent->groups->sgc->imbalance; | |
8320 | ||
8321 | if (*group_imbalance) | |
8322 | *group_imbalance = 0; | |
8323 | } | |
8324 | ||
8325 | out_all_pinned: | |
8326 | /* | |
8327 | * We reach balance because all tasks are pinned at this level so | |
8328 | * we can't migrate them. Let the imbalance flag set so parent level | |
8329 | * can try to migrate them. | |
8330 | */ | |
ae92882e | 8331 | schedstat_inc(sd->lb_balanced[idle]); |
1e3c88bd PZ |
8332 | |
8333 | sd->nr_balance_failed = 0; | |
8334 | ||
8335 | out_one_pinned: | |
8336 | /* tune up the balancing interval */ | |
8e45cb54 | 8337 | if (((env.flags & LBF_ALL_PINNED) && |
5b54b56b | 8338 | sd->balance_interval < MAX_PINNED_INTERVAL) || |
1e3c88bd PZ |
8339 | (sd->balance_interval < sd->max_interval)) |
8340 | sd->balance_interval *= 2; | |
8341 | ||
46e49b38 | 8342 | ld_moved = 0; |
1e3c88bd | 8343 | out: |
1e3c88bd PZ |
8344 | return ld_moved; |
8345 | } | |
8346 | ||
52a08ef1 JL |
8347 | static inline unsigned long |
8348 | get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) | |
8349 | { | |
8350 | unsigned long interval = sd->balance_interval; | |
8351 | ||
8352 | if (cpu_busy) | |
8353 | interval *= sd->busy_factor; | |
8354 | ||
8355 | /* scale ms to jiffies */ | |
8356 | interval = msecs_to_jiffies(interval); | |
8357 | interval = clamp(interval, 1UL, max_load_balance_interval); | |
8358 | ||
8359 | return interval; | |
8360 | } | |
8361 | ||
8362 | static inline void | |
31851a98 | 8363 | update_next_balance(struct sched_domain *sd, unsigned long *next_balance) |
52a08ef1 JL |
8364 | { |
8365 | unsigned long interval, next; | |
8366 | ||
31851a98 LY |
8367 | /* used by idle balance, so cpu_busy = 0 */ |
8368 | interval = get_sd_balance_interval(sd, 0); | |
52a08ef1 JL |
8369 | next = sd->last_balance + interval; |
8370 | ||
8371 | if (time_after(*next_balance, next)) | |
8372 | *next_balance = next; | |
8373 | } | |
8374 | ||
1e3c88bd PZ |
8375 | /* |
8376 | * idle_balance is called by schedule() if this_cpu is about to become | |
8377 | * idle. Attempts to pull tasks from other CPUs. | |
8378 | */ | |
46f69fa3 | 8379 | static int idle_balance(struct rq *this_rq, struct rq_flags *rf) |
1e3c88bd | 8380 | { |
52a08ef1 JL |
8381 | unsigned long next_balance = jiffies + HZ; |
8382 | int this_cpu = this_rq->cpu; | |
1e3c88bd PZ |
8383 | struct sched_domain *sd; |
8384 | int pulled_task = 0; | |
9bd721c5 | 8385 | u64 curr_cost = 0; |
1e3c88bd | 8386 | |
6e83125c PZ |
8387 | /* |
8388 | * We must set idle_stamp _before_ calling idle_balance(), such that we | |
8389 | * measure the duration of idle_balance() as idle time. | |
8390 | */ | |
8391 | this_rq->idle_stamp = rq_clock(this_rq); | |
8392 | ||
46f69fa3 MF |
8393 | /* |
8394 | * This is OK, because current is on_cpu, which avoids it being picked | |
8395 | * for load-balance and preemption/IRQs are still disabled avoiding | |
8396 | * further scheduler activity on it and we're being very careful to | |
8397 | * re-start the picking loop. | |
8398 | */ | |
8399 | rq_unpin_lock(this_rq, rf); | |
8400 | ||
4486edd1 TC |
8401 | if (this_rq->avg_idle < sysctl_sched_migration_cost || |
8402 | !this_rq->rd->overload) { | |
52a08ef1 JL |
8403 | rcu_read_lock(); |
8404 | sd = rcu_dereference_check_sched_domain(this_rq->sd); | |
8405 | if (sd) | |
31851a98 | 8406 | update_next_balance(sd, &next_balance); |
52a08ef1 JL |
8407 | rcu_read_unlock(); |
8408 | ||
6e83125c | 8409 | goto out; |
52a08ef1 | 8410 | } |
1e3c88bd | 8411 | |
f492e12e PZ |
8412 | raw_spin_unlock(&this_rq->lock); |
8413 | ||
48a16753 | 8414 | update_blocked_averages(this_cpu); |
dce840a0 | 8415 | rcu_read_lock(); |
1e3c88bd | 8416 | for_each_domain(this_cpu, sd) { |
23f0d209 | 8417 | int continue_balancing = 1; |
9bd721c5 | 8418 | u64 t0, domain_cost; |
1e3c88bd PZ |
8419 | |
8420 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
8421 | continue; | |
8422 | ||
52a08ef1 | 8423 | if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { |
31851a98 | 8424 | update_next_balance(sd, &next_balance); |
9bd721c5 | 8425 | break; |
52a08ef1 | 8426 | } |
9bd721c5 | 8427 | |
f492e12e | 8428 | if (sd->flags & SD_BALANCE_NEWIDLE) { |
9bd721c5 JL |
8429 | t0 = sched_clock_cpu(this_cpu); |
8430 | ||
f492e12e | 8431 | pulled_task = load_balance(this_cpu, this_rq, |
23f0d209 JK |
8432 | sd, CPU_NEWLY_IDLE, |
8433 | &continue_balancing); | |
9bd721c5 JL |
8434 | |
8435 | domain_cost = sched_clock_cpu(this_cpu) - t0; | |
8436 | if (domain_cost > sd->max_newidle_lb_cost) | |
8437 | sd->max_newidle_lb_cost = domain_cost; | |
8438 | ||
8439 | curr_cost += domain_cost; | |
f492e12e | 8440 | } |
1e3c88bd | 8441 | |
31851a98 | 8442 | update_next_balance(sd, &next_balance); |
39a4d9ca JL |
8443 | |
8444 | /* | |
8445 | * Stop searching for tasks to pull if there are | |
8446 | * now runnable tasks on this rq. | |
8447 | */ | |
8448 | if (pulled_task || this_rq->nr_running > 0) | |
1e3c88bd | 8449 | break; |
1e3c88bd | 8450 | } |
dce840a0 | 8451 | rcu_read_unlock(); |
f492e12e PZ |
8452 | |
8453 | raw_spin_lock(&this_rq->lock); | |
8454 | ||
0e5b5337 JL |
8455 | if (curr_cost > this_rq->max_idle_balance_cost) |
8456 | this_rq->max_idle_balance_cost = curr_cost; | |
8457 | ||
e5fc6611 | 8458 | /* |
0e5b5337 JL |
8459 | * While browsing the domains, we released the rq lock, a task could |
8460 | * have been enqueued in the meantime. Since we're not going idle, | |
8461 | * pretend we pulled a task. | |
e5fc6611 | 8462 | */ |
0e5b5337 | 8463 | if (this_rq->cfs.h_nr_running && !pulled_task) |
6e83125c | 8464 | pulled_task = 1; |
e5fc6611 | 8465 | |
52a08ef1 JL |
8466 | out: |
8467 | /* Move the next balance forward */ | |
8468 | if (time_after(this_rq->next_balance, next_balance)) | |
1e3c88bd | 8469 | this_rq->next_balance = next_balance; |
9bd721c5 | 8470 | |
e4aa358b | 8471 | /* Is there a task of a high priority class? */ |
46383648 | 8472 | if (this_rq->nr_running != this_rq->cfs.h_nr_running) |
e4aa358b KT |
8473 | pulled_task = -1; |
8474 | ||
38c6ade2 | 8475 | if (pulled_task) |
6e83125c PZ |
8476 | this_rq->idle_stamp = 0; |
8477 | ||
46f69fa3 MF |
8478 | rq_repin_lock(this_rq, rf); |
8479 | ||
3c4017c1 | 8480 | return pulled_task; |
1e3c88bd PZ |
8481 | } |
8482 | ||
8483 | /* | |
969c7921 TH |
8484 | * active_load_balance_cpu_stop is run by cpu stopper. It pushes |
8485 | * running tasks off the busiest CPU onto idle CPUs. It requires at | |
8486 | * least 1 task to be running on each physical CPU where possible, and | |
8487 | * avoids physical / logical imbalances. | |
1e3c88bd | 8488 | */ |
969c7921 | 8489 | static int active_load_balance_cpu_stop(void *data) |
1e3c88bd | 8490 | { |
969c7921 TH |
8491 | struct rq *busiest_rq = data; |
8492 | int busiest_cpu = cpu_of(busiest_rq); | |
1e3c88bd | 8493 | int target_cpu = busiest_rq->push_cpu; |
969c7921 | 8494 | struct rq *target_rq = cpu_rq(target_cpu); |
1e3c88bd | 8495 | struct sched_domain *sd; |
e5673f28 | 8496 | struct task_struct *p = NULL; |
8a8c69c3 | 8497 | struct rq_flags rf; |
969c7921 | 8498 | |
8a8c69c3 | 8499 | rq_lock_irq(busiest_rq, &rf); |
969c7921 TH |
8500 | |
8501 | /* make sure the requested cpu hasn't gone down in the meantime */ | |
8502 | if (unlikely(busiest_cpu != smp_processor_id() || | |
8503 | !busiest_rq->active_balance)) | |
8504 | goto out_unlock; | |
1e3c88bd PZ |
8505 | |
8506 | /* Is there any task to move? */ | |
8507 | if (busiest_rq->nr_running <= 1) | |
969c7921 | 8508 | goto out_unlock; |
1e3c88bd PZ |
8509 | |
8510 | /* | |
8511 | * This condition is "impossible", if it occurs | |
8512 | * we need to fix it. Originally reported by | |
8513 | * Bjorn Helgaas on a 128-cpu setup. | |
8514 | */ | |
8515 | BUG_ON(busiest_rq == target_rq); | |
8516 | ||
1e3c88bd | 8517 | /* Search for an sd spanning us and the target CPU. */ |
dce840a0 | 8518 | rcu_read_lock(); |
1e3c88bd PZ |
8519 | for_each_domain(target_cpu, sd) { |
8520 | if ((sd->flags & SD_LOAD_BALANCE) && | |
8521 | cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) | |
8522 | break; | |
8523 | } | |
8524 | ||
8525 | if (likely(sd)) { | |
8e45cb54 PZ |
8526 | struct lb_env env = { |
8527 | .sd = sd, | |
ddcdf6e7 PZ |
8528 | .dst_cpu = target_cpu, |
8529 | .dst_rq = target_rq, | |
8530 | .src_cpu = busiest_rq->cpu, | |
8531 | .src_rq = busiest_rq, | |
8e45cb54 PZ |
8532 | .idle = CPU_IDLE, |
8533 | }; | |
8534 | ||
ae92882e | 8535 | schedstat_inc(sd->alb_count); |
3bed5e21 | 8536 | update_rq_clock(busiest_rq); |
1e3c88bd | 8537 | |
e5673f28 | 8538 | p = detach_one_task(&env); |
d02c0711 | 8539 | if (p) { |
ae92882e | 8540 | schedstat_inc(sd->alb_pushed); |
d02c0711 SD |
8541 | /* Active balancing done, reset the failure counter. */ |
8542 | sd->nr_balance_failed = 0; | |
8543 | } else { | |
ae92882e | 8544 | schedstat_inc(sd->alb_failed); |
d02c0711 | 8545 | } |
1e3c88bd | 8546 | } |
dce840a0 | 8547 | rcu_read_unlock(); |
969c7921 TH |
8548 | out_unlock: |
8549 | busiest_rq->active_balance = 0; | |
8a8c69c3 | 8550 | rq_unlock(busiest_rq, &rf); |
e5673f28 KT |
8551 | |
8552 | if (p) | |
8553 | attach_one_task(target_rq, p); | |
8554 | ||
8555 | local_irq_enable(); | |
8556 | ||
969c7921 | 8557 | return 0; |
1e3c88bd PZ |
8558 | } |
8559 | ||
d987fc7f MG |
8560 | static inline int on_null_domain(struct rq *rq) |
8561 | { | |
8562 | return unlikely(!rcu_dereference_sched(rq->sd)); | |
8563 | } | |
8564 | ||
3451d024 | 8565 | #ifdef CONFIG_NO_HZ_COMMON |
83cd4fe2 VP |
8566 | /* |
8567 | * idle load balancing details | |
83cd4fe2 VP |
8568 | * - When one of the busy CPUs notice that there may be an idle rebalancing |
8569 | * needed, they will kick the idle load balancer, which then does idle | |
8570 | * load balancing for all the idle CPUs. | |
8571 | */ | |
1e3c88bd | 8572 | static struct { |
83cd4fe2 | 8573 | cpumask_var_t idle_cpus_mask; |
0b005cf5 | 8574 | atomic_t nr_cpus; |
83cd4fe2 VP |
8575 | unsigned long next_balance; /* in jiffy units */ |
8576 | } nohz ____cacheline_aligned; | |
1e3c88bd | 8577 | |
3dd0337d | 8578 | static inline int find_new_ilb(void) |
1e3c88bd | 8579 | { |
0b005cf5 | 8580 | int ilb = cpumask_first(nohz.idle_cpus_mask); |
1e3c88bd | 8581 | |
786d6dc7 SS |
8582 | if (ilb < nr_cpu_ids && idle_cpu(ilb)) |
8583 | return ilb; | |
8584 | ||
8585 | return nr_cpu_ids; | |
1e3c88bd | 8586 | } |
1e3c88bd | 8587 | |
83cd4fe2 VP |
8588 | /* |
8589 | * Kick a CPU to do the nohz balancing, if it is time for it. We pick the | |
8590 | * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle | |
8591 | * CPU (if there is one). | |
8592 | */ | |
0aeeeeba | 8593 | static void nohz_balancer_kick(void) |
83cd4fe2 VP |
8594 | { |
8595 | int ilb_cpu; | |
8596 | ||
8597 | nohz.next_balance++; | |
8598 | ||
3dd0337d | 8599 | ilb_cpu = find_new_ilb(); |
83cd4fe2 | 8600 | |
0b005cf5 SS |
8601 | if (ilb_cpu >= nr_cpu_ids) |
8602 | return; | |
83cd4fe2 | 8603 | |
cd490c5b | 8604 | if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) |
1c792db7 SS |
8605 | return; |
8606 | /* | |
8607 | * Use smp_send_reschedule() instead of resched_cpu(). | |
8608 | * This way we generate a sched IPI on the target cpu which | |
8609 | * is idle. And the softirq performing nohz idle load balance | |
8610 | * will be run before returning from the IPI. | |
8611 | */ | |
8612 | smp_send_reschedule(ilb_cpu); | |
83cd4fe2 VP |
8613 | return; |
8614 | } | |
8615 | ||
20a5c8cc | 8616 | void nohz_balance_exit_idle(unsigned int cpu) |
71325960 SS |
8617 | { |
8618 | if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { | |
d987fc7f MG |
8619 | /* |
8620 | * Completely isolated CPUs don't ever set, so we must test. | |
8621 | */ | |
8622 | if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) { | |
8623 | cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); | |
8624 | atomic_dec(&nohz.nr_cpus); | |
8625 | } | |
71325960 SS |
8626 | clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); |
8627 | } | |
8628 | } | |
8629 | ||
69e1e811 SS |
8630 | static inline void set_cpu_sd_state_busy(void) |
8631 | { | |
8632 | struct sched_domain *sd; | |
37dc6b50 | 8633 | int cpu = smp_processor_id(); |
69e1e811 | 8634 | |
69e1e811 | 8635 | rcu_read_lock(); |
0e369d75 | 8636 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
8637 | |
8638 | if (!sd || !sd->nohz_idle) | |
8639 | goto unlock; | |
8640 | sd->nohz_idle = 0; | |
8641 | ||
0e369d75 | 8642 | atomic_inc(&sd->shared->nr_busy_cpus); |
25f55d9d | 8643 | unlock: |
69e1e811 SS |
8644 | rcu_read_unlock(); |
8645 | } | |
8646 | ||
8647 | void set_cpu_sd_state_idle(void) | |
8648 | { | |
8649 | struct sched_domain *sd; | |
37dc6b50 | 8650 | int cpu = smp_processor_id(); |
69e1e811 | 8651 | |
69e1e811 | 8652 | rcu_read_lock(); |
0e369d75 | 8653 | sd = rcu_dereference(per_cpu(sd_llc, cpu)); |
25f55d9d VG |
8654 | |
8655 | if (!sd || sd->nohz_idle) | |
8656 | goto unlock; | |
8657 | sd->nohz_idle = 1; | |
8658 | ||
0e369d75 | 8659 | atomic_dec(&sd->shared->nr_busy_cpus); |
25f55d9d | 8660 | unlock: |
69e1e811 SS |
8661 | rcu_read_unlock(); |
8662 | } | |
8663 | ||
1e3c88bd | 8664 | /* |
c1cc017c | 8665 | * This routine will record that the cpu is going idle with tick stopped. |
0b005cf5 | 8666 | * This info will be used in performing idle load balancing in the future. |
1e3c88bd | 8667 | */ |
c1cc017c | 8668 | void nohz_balance_enter_idle(int cpu) |
1e3c88bd | 8669 | { |
71325960 SS |
8670 | /* |
8671 | * If this cpu is going down, then nothing needs to be done. | |
8672 | */ | |
8673 | if (!cpu_active(cpu)) | |
8674 | return; | |
8675 | ||
c1cc017c AS |
8676 | if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) |
8677 | return; | |
1e3c88bd | 8678 | |
d987fc7f MG |
8679 | /* |
8680 | * If we're a completely isolated CPU, we don't play. | |
8681 | */ | |
8682 | if (on_null_domain(cpu_rq(cpu))) | |
8683 | return; | |
8684 | ||
c1cc017c AS |
8685 | cpumask_set_cpu(cpu, nohz.idle_cpus_mask); |
8686 | atomic_inc(&nohz.nr_cpus); | |
8687 | set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); | |
1e3c88bd PZ |
8688 | } |
8689 | #endif | |
8690 | ||
8691 | static DEFINE_SPINLOCK(balancing); | |
8692 | ||
49c022e6 PZ |
8693 | /* |
8694 | * Scale the max load_balance interval with the number of CPUs in the system. | |
8695 | * This trades load-balance latency on larger machines for less cross talk. | |
8696 | */ | |
029632fb | 8697 | void update_max_interval(void) |
49c022e6 PZ |
8698 | { |
8699 | max_load_balance_interval = HZ*num_online_cpus()/10; | |
8700 | } | |
8701 | ||
1e3c88bd PZ |
8702 | /* |
8703 | * It checks each scheduling domain to see if it is due to be balanced, | |
8704 | * and initiates a balancing operation if so. | |
8705 | * | |
b9b0853a | 8706 | * Balancing parameters are set up in init_sched_domains. |
1e3c88bd | 8707 | */ |
f7ed0a89 | 8708 | static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) |
1e3c88bd | 8709 | { |
23f0d209 | 8710 | int continue_balancing = 1; |
f7ed0a89 | 8711 | int cpu = rq->cpu; |
1e3c88bd | 8712 | unsigned long interval; |
04f733b4 | 8713 | struct sched_domain *sd; |
1e3c88bd PZ |
8714 | /* Earliest time when we have to do rebalance again */ |
8715 | unsigned long next_balance = jiffies + 60*HZ; | |
8716 | int update_next_balance = 0; | |
f48627e6 JL |
8717 | int need_serialize, need_decay = 0; |
8718 | u64 max_cost = 0; | |
1e3c88bd | 8719 | |
48a16753 | 8720 | update_blocked_averages(cpu); |
2069dd75 | 8721 | |
dce840a0 | 8722 | rcu_read_lock(); |
1e3c88bd | 8723 | for_each_domain(cpu, sd) { |
f48627e6 JL |
8724 | /* |
8725 | * Decay the newidle max times here because this is a regular | |
8726 | * visit to all the domains. Decay ~1% per second. | |
8727 | */ | |
8728 | if (time_after(jiffies, sd->next_decay_max_lb_cost)) { | |
8729 | sd->max_newidle_lb_cost = | |
8730 | (sd->max_newidle_lb_cost * 253) / 256; | |
8731 | sd->next_decay_max_lb_cost = jiffies + HZ; | |
8732 | need_decay = 1; | |
8733 | } | |
8734 | max_cost += sd->max_newidle_lb_cost; | |
8735 | ||
1e3c88bd PZ |
8736 | if (!(sd->flags & SD_LOAD_BALANCE)) |
8737 | continue; | |
8738 | ||
f48627e6 JL |
8739 | /* |
8740 | * Stop the load balance at this level. There is another | |
8741 | * CPU in our sched group which is doing load balancing more | |
8742 | * actively. | |
8743 | */ | |
8744 | if (!continue_balancing) { | |
8745 | if (need_decay) | |
8746 | continue; | |
8747 | break; | |
8748 | } | |
8749 | ||
52a08ef1 | 8750 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
8751 | |
8752 | need_serialize = sd->flags & SD_SERIALIZE; | |
1e3c88bd PZ |
8753 | if (need_serialize) { |
8754 | if (!spin_trylock(&balancing)) | |
8755 | goto out; | |
8756 | } | |
8757 | ||
8758 | if (time_after_eq(jiffies, sd->last_balance + interval)) { | |
23f0d209 | 8759 | if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { |
1e3c88bd | 8760 | /* |
6263322c | 8761 | * The LBF_DST_PINNED logic could have changed |
de5eb2dd JK |
8762 | * env->dst_cpu, so we can't know our idle |
8763 | * state even if we migrated tasks. Update it. | |
1e3c88bd | 8764 | */ |
de5eb2dd | 8765 | idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; |
1e3c88bd PZ |
8766 | } |
8767 | sd->last_balance = jiffies; | |
52a08ef1 | 8768 | interval = get_sd_balance_interval(sd, idle != CPU_IDLE); |
1e3c88bd PZ |
8769 | } |
8770 | if (need_serialize) | |
8771 | spin_unlock(&balancing); | |
8772 | out: | |
8773 | if (time_after(next_balance, sd->last_balance + interval)) { | |
8774 | next_balance = sd->last_balance + interval; | |
8775 | update_next_balance = 1; | |
8776 | } | |
f48627e6 JL |
8777 | } |
8778 | if (need_decay) { | |
1e3c88bd | 8779 | /* |
f48627e6 JL |
8780 | * Ensure the rq-wide value also decays but keep it at a |
8781 | * reasonable floor to avoid funnies with rq->avg_idle. | |
1e3c88bd | 8782 | */ |
f48627e6 JL |
8783 | rq->max_idle_balance_cost = |
8784 | max((u64)sysctl_sched_migration_cost, max_cost); | |
1e3c88bd | 8785 | } |
dce840a0 | 8786 | rcu_read_unlock(); |
1e3c88bd PZ |
8787 | |
8788 | /* | |
8789 | * next_balance will be updated only when there is a need. | |
8790 | * When the cpu is attached to null domain for ex, it will not be | |
8791 | * updated. | |
8792 | */ | |
c5afb6a8 | 8793 | if (likely(update_next_balance)) { |
1e3c88bd | 8794 | rq->next_balance = next_balance; |
c5afb6a8 VG |
8795 | |
8796 | #ifdef CONFIG_NO_HZ_COMMON | |
8797 | /* | |
8798 | * If this CPU has been elected to perform the nohz idle | |
8799 | * balance. Other idle CPUs have already rebalanced with | |
8800 | * nohz_idle_balance() and nohz.next_balance has been | |
8801 | * updated accordingly. This CPU is now running the idle load | |
8802 | * balance for itself and we need to update the | |
8803 | * nohz.next_balance accordingly. | |
8804 | */ | |
8805 | if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance)) | |
8806 | nohz.next_balance = rq->next_balance; | |
8807 | #endif | |
8808 | } | |
1e3c88bd PZ |
8809 | } |
8810 | ||
3451d024 | 8811 | #ifdef CONFIG_NO_HZ_COMMON |
1e3c88bd | 8812 | /* |
3451d024 | 8813 | * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the |
1e3c88bd PZ |
8814 | * rebalancing for all the cpus for whom scheduler ticks are stopped. |
8815 | */ | |
208cb16b | 8816 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) |
83cd4fe2 | 8817 | { |
208cb16b | 8818 | int this_cpu = this_rq->cpu; |
83cd4fe2 VP |
8819 | struct rq *rq; |
8820 | int balance_cpu; | |
c5afb6a8 VG |
8821 | /* Earliest time when we have to do rebalance again */ |
8822 | unsigned long next_balance = jiffies + 60*HZ; | |
8823 | int update_next_balance = 0; | |
83cd4fe2 | 8824 | |
1c792db7 SS |
8825 | if (idle != CPU_IDLE || |
8826 | !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) | |
8827 | goto end; | |
83cd4fe2 VP |
8828 | |
8829 | for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { | |
8a6d42d1 | 8830 | if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) |
83cd4fe2 VP |
8831 | continue; |
8832 | ||
8833 | /* | |
8834 | * If this cpu gets work to do, stop the load balancing | |
8835 | * work being done for other cpus. Next load | |
8836 | * balancing owner will pick it up. | |
8837 | */ | |
1c792db7 | 8838 | if (need_resched()) |
83cd4fe2 | 8839 | break; |
83cd4fe2 | 8840 | |
5ed4f1d9 VG |
8841 | rq = cpu_rq(balance_cpu); |
8842 | ||
ed61bbc6 TC |
8843 | /* |
8844 | * If time for next balance is due, | |
8845 | * do the balance. | |
8846 | */ | |
8847 | if (time_after_eq(jiffies, rq->next_balance)) { | |
8a8c69c3 PZ |
8848 | struct rq_flags rf; |
8849 | ||
8850 | rq_lock_irq(rq, &rf); | |
ed61bbc6 | 8851 | update_rq_clock(rq); |
cee1afce | 8852 | cpu_load_update_idle(rq); |
8a8c69c3 PZ |
8853 | rq_unlock_irq(rq, &rf); |
8854 | ||
ed61bbc6 TC |
8855 | rebalance_domains(rq, CPU_IDLE); |
8856 | } | |
83cd4fe2 | 8857 | |
c5afb6a8 VG |
8858 | if (time_after(next_balance, rq->next_balance)) { |
8859 | next_balance = rq->next_balance; | |
8860 | update_next_balance = 1; | |
8861 | } | |
83cd4fe2 | 8862 | } |
c5afb6a8 VG |
8863 | |
8864 | /* | |
8865 | * next_balance will be updated only when there is a need. | |
8866 | * When the CPU is attached to null domain for ex, it will not be | |
8867 | * updated. | |
8868 | */ | |
8869 | if (likely(update_next_balance)) | |
8870 | nohz.next_balance = next_balance; | |
1c792db7 SS |
8871 | end: |
8872 | clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); | |
83cd4fe2 VP |
8873 | } |
8874 | ||
8875 | /* | |
0b005cf5 | 8876 | * Current heuristic for kicking the idle load balancer in the presence |
1aaf90a4 | 8877 | * of an idle cpu in the system. |
0b005cf5 | 8878 | * - This rq has more than one task. |
1aaf90a4 VG |
8879 | * - This rq has at least one CFS task and the capacity of the CPU is |
8880 | * significantly reduced because of RT tasks or IRQs. | |
8881 | * - At parent of LLC scheduler domain level, this cpu's scheduler group has | |
8882 | * multiple busy cpu. | |
0b005cf5 SS |
8883 | * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler |
8884 | * domain span are idle. | |
83cd4fe2 | 8885 | */ |
1aaf90a4 | 8886 | static inline bool nohz_kick_needed(struct rq *rq) |
83cd4fe2 VP |
8887 | { |
8888 | unsigned long now = jiffies; | |
0e369d75 | 8889 | struct sched_domain_shared *sds; |
0b005cf5 | 8890 | struct sched_domain *sd; |
afe06efd | 8891 | int nr_busy, i, cpu = rq->cpu; |
1aaf90a4 | 8892 | bool kick = false; |
83cd4fe2 | 8893 | |
4a725627 | 8894 | if (unlikely(rq->idle_balance)) |
1aaf90a4 | 8895 | return false; |
83cd4fe2 | 8896 | |
1c792db7 SS |
8897 | /* |
8898 | * We may be recently in ticked or tickless idle mode. At the first | |
8899 | * busy tick after returning from idle, we will update the busy stats. | |
8900 | */ | |
69e1e811 | 8901 | set_cpu_sd_state_busy(); |
c1cc017c | 8902 | nohz_balance_exit_idle(cpu); |
0b005cf5 SS |
8903 | |
8904 | /* | |
8905 | * None are in tickless mode and hence no need for NOHZ idle load | |
8906 | * balancing. | |
8907 | */ | |
8908 | if (likely(!atomic_read(&nohz.nr_cpus))) | |
1aaf90a4 | 8909 | return false; |
1c792db7 SS |
8910 | |
8911 | if (time_before(now, nohz.next_balance)) | |
1aaf90a4 | 8912 | return false; |
83cd4fe2 | 8913 | |
0b005cf5 | 8914 | if (rq->nr_running >= 2) |
1aaf90a4 | 8915 | return true; |
83cd4fe2 | 8916 | |
067491b7 | 8917 | rcu_read_lock(); |
0e369d75 PZ |
8918 | sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); |
8919 | if (sds) { | |
8920 | /* | |
8921 | * XXX: write a coherent comment on why we do this. | |
8922 | * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com | |
8923 | */ | |
8924 | nr_busy = atomic_read(&sds->nr_busy_cpus); | |
1aaf90a4 VG |
8925 | if (nr_busy > 1) { |
8926 | kick = true; | |
8927 | goto unlock; | |
8928 | } | |
8929 | ||
83cd4fe2 | 8930 | } |
37dc6b50 | 8931 | |
1aaf90a4 VG |
8932 | sd = rcu_dereference(rq->sd); |
8933 | if (sd) { | |
8934 | if ((rq->cfs.h_nr_running >= 1) && | |
8935 | check_cpu_capacity(rq, sd)) { | |
8936 | kick = true; | |
8937 | goto unlock; | |
8938 | } | |
8939 | } | |
37dc6b50 | 8940 | |
1aaf90a4 | 8941 | sd = rcu_dereference(per_cpu(sd_asym, cpu)); |
afe06efd TC |
8942 | if (sd) { |
8943 | for_each_cpu(i, sched_domain_span(sd)) { | |
8944 | if (i == cpu || | |
8945 | !cpumask_test_cpu(i, nohz.idle_cpus_mask)) | |
8946 | continue; | |
067491b7 | 8947 | |
afe06efd TC |
8948 | if (sched_asym_prefer(i, cpu)) { |
8949 | kick = true; | |
8950 | goto unlock; | |
8951 | } | |
8952 | } | |
8953 | } | |
1aaf90a4 | 8954 | unlock: |
067491b7 | 8955 | rcu_read_unlock(); |
1aaf90a4 | 8956 | return kick; |
83cd4fe2 VP |
8957 | } |
8958 | #else | |
208cb16b | 8959 | static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } |
83cd4fe2 VP |
8960 | #endif |
8961 | ||
8962 | /* | |
8963 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
8964 | * Also triggered for nohz idle balancing (with nohz_balancing_kick set). | |
8965 | */ | |
0766f788 | 8966 | static __latent_entropy void run_rebalance_domains(struct softirq_action *h) |
1e3c88bd | 8967 | { |
208cb16b | 8968 | struct rq *this_rq = this_rq(); |
6eb57e0d | 8969 | enum cpu_idle_type idle = this_rq->idle_balance ? |
1e3c88bd PZ |
8970 | CPU_IDLE : CPU_NOT_IDLE; |
8971 | ||
1e3c88bd | 8972 | /* |
83cd4fe2 | 8973 | * If this cpu has a pending nohz_balance_kick, then do the |
1e3c88bd | 8974 | * balancing on behalf of the other idle cpus whose ticks are |
d4573c3e PM |
8975 | * stopped. Do nohz_idle_balance *before* rebalance_domains to |
8976 | * give the idle cpus a chance to load balance. Else we may | |
8977 | * load balance only within the local sched_domain hierarchy | |
8978 | * and abort nohz_idle_balance altogether if we pull some load. | |
1e3c88bd | 8979 | */ |
208cb16b | 8980 | nohz_idle_balance(this_rq, idle); |
d4573c3e | 8981 | rebalance_domains(this_rq, idle); |
1e3c88bd PZ |
8982 | } |
8983 | ||
1e3c88bd PZ |
8984 | /* |
8985 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
1e3c88bd | 8986 | */ |
7caff66f | 8987 | void trigger_load_balance(struct rq *rq) |
1e3c88bd | 8988 | { |
1e3c88bd | 8989 | /* Don't need to rebalance while attached to NULL domain */ |
c726099e DL |
8990 | if (unlikely(on_null_domain(rq))) |
8991 | return; | |
8992 | ||
8993 | if (time_after_eq(jiffies, rq->next_balance)) | |
1e3c88bd | 8994 | raise_softirq(SCHED_SOFTIRQ); |
3451d024 | 8995 | #ifdef CONFIG_NO_HZ_COMMON |
c726099e | 8996 | if (nohz_kick_needed(rq)) |
0aeeeeba | 8997 | nohz_balancer_kick(); |
83cd4fe2 | 8998 | #endif |
1e3c88bd PZ |
8999 | } |
9000 | ||
0bcdcf28 CE |
9001 | static void rq_online_fair(struct rq *rq) |
9002 | { | |
9003 | update_sysctl(); | |
0e59bdae KT |
9004 | |
9005 | update_runtime_enabled(rq); | |
0bcdcf28 CE |
9006 | } |
9007 | ||
9008 | static void rq_offline_fair(struct rq *rq) | |
9009 | { | |
9010 | update_sysctl(); | |
a4c96ae3 PB |
9011 | |
9012 | /* Ensure any throttled groups are reachable by pick_next_task */ | |
9013 | unthrottle_offline_cfs_rqs(rq); | |
0bcdcf28 CE |
9014 | } |
9015 | ||
55e12e5e | 9016 | #endif /* CONFIG_SMP */ |
e1d1484f | 9017 | |
bf0f6f24 IM |
9018 | /* |
9019 | * scheduler tick hitting a task of our scheduling class: | |
9020 | */ | |
8f4d37ec | 9021 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
bf0f6f24 IM |
9022 | { |
9023 | struct cfs_rq *cfs_rq; | |
9024 | struct sched_entity *se = &curr->se; | |
9025 | ||
9026 | for_each_sched_entity(se) { | |
9027 | cfs_rq = cfs_rq_of(se); | |
8f4d37ec | 9028 | entity_tick(cfs_rq, se, queued); |
bf0f6f24 | 9029 | } |
18bf2805 | 9030 | |
b52da86e | 9031 | if (static_branch_unlikely(&sched_numa_balancing)) |
cbee9f88 | 9032 | task_tick_numa(rq, curr); |
bf0f6f24 IM |
9033 | } |
9034 | ||
9035 | /* | |
cd29fe6f PZ |
9036 | * called on fork with the child task as argument from the parent's context |
9037 | * - child not yet on the tasklist | |
9038 | * - preemption disabled | |
bf0f6f24 | 9039 | */ |
cd29fe6f | 9040 | static void task_fork_fair(struct task_struct *p) |
bf0f6f24 | 9041 | { |
4fc420c9 DN |
9042 | struct cfs_rq *cfs_rq; |
9043 | struct sched_entity *se = &p->se, *curr; | |
cd29fe6f | 9044 | struct rq *rq = this_rq(); |
8a8c69c3 | 9045 | struct rq_flags rf; |
bf0f6f24 | 9046 | |
8a8c69c3 | 9047 | rq_lock(rq, &rf); |
861d034e PZ |
9048 | update_rq_clock(rq); |
9049 | ||
4fc420c9 DN |
9050 | cfs_rq = task_cfs_rq(current); |
9051 | curr = cfs_rq->curr; | |
e210bffd PZ |
9052 | if (curr) { |
9053 | update_curr(cfs_rq); | |
b5d9d734 | 9054 | se->vruntime = curr->vruntime; |
e210bffd | 9055 | } |
aeb73b04 | 9056 | place_entity(cfs_rq, se, 1); |
4d78e7b6 | 9057 | |
cd29fe6f | 9058 | if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { |
87fefa38 | 9059 | /* |
edcb60a3 IM |
9060 | * Upon rescheduling, sched_class::put_prev_task() will place |
9061 | * 'current' within the tree based on its new key value. | |
9062 | */ | |
4d78e7b6 | 9063 | swap(curr->vruntime, se->vruntime); |
8875125e | 9064 | resched_curr(rq); |
4d78e7b6 | 9065 | } |
bf0f6f24 | 9066 | |
88ec22d3 | 9067 | se->vruntime -= cfs_rq->min_vruntime; |
8a8c69c3 | 9068 | rq_unlock(rq, &rf); |
bf0f6f24 IM |
9069 | } |
9070 | ||
cb469845 SR |
9071 | /* |
9072 | * Priority of the task has changed. Check to see if we preempt | |
9073 | * the current task. | |
9074 | */ | |
da7a735e PZ |
9075 | static void |
9076 | prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) | |
cb469845 | 9077 | { |
da0c1e65 | 9078 | if (!task_on_rq_queued(p)) |
da7a735e PZ |
9079 | return; |
9080 | ||
cb469845 SR |
9081 | /* |
9082 | * Reschedule if we are currently running on this runqueue and | |
9083 | * our priority decreased, or if we are not currently running on | |
9084 | * this runqueue and our priority is higher than the current's | |
9085 | */ | |
da7a735e | 9086 | if (rq->curr == p) { |
cb469845 | 9087 | if (p->prio > oldprio) |
8875125e | 9088 | resched_curr(rq); |
cb469845 | 9089 | } else |
15afe09b | 9090 | check_preempt_curr(rq, p, 0); |
cb469845 SR |
9091 | } |
9092 | ||
daa59407 | 9093 | static inline bool vruntime_normalized(struct task_struct *p) |
da7a735e PZ |
9094 | { |
9095 | struct sched_entity *se = &p->se; | |
da7a735e PZ |
9096 | |
9097 | /* | |
daa59407 BP |
9098 | * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, |
9099 | * the dequeue_entity(.flags=0) will already have normalized the | |
9100 | * vruntime. | |
9101 | */ | |
9102 | if (p->on_rq) | |
9103 | return true; | |
9104 | ||
9105 | /* | |
9106 | * When !on_rq, vruntime of the task has usually NOT been normalized. | |
9107 | * But there are some cases where it has already been normalized: | |
da7a735e | 9108 | * |
daa59407 BP |
9109 | * - A forked child which is waiting for being woken up by |
9110 | * wake_up_new_task(). | |
9111 | * - A task which has been woken up by try_to_wake_up() and | |
9112 | * waiting for actually being woken up by sched_ttwu_pending(). | |
da7a735e | 9113 | */ |
daa59407 BP |
9114 | if (!se->sum_exec_runtime || p->state == TASK_WAKING) |
9115 | return true; | |
9116 | ||
9117 | return false; | |
9118 | } | |
9119 | ||
09a43ace VG |
9120 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9121 | /* | |
9122 | * Propagate the changes of the sched_entity across the tg tree to make it | |
9123 | * visible to the root | |
9124 | */ | |
9125 | static void propagate_entity_cfs_rq(struct sched_entity *se) | |
9126 | { | |
9127 | struct cfs_rq *cfs_rq; | |
9128 | ||
9129 | /* Start to propagate at parent */ | |
9130 | se = se->parent; | |
9131 | ||
9132 | for_each_sched_entity(se) { | |
9133 | cfs_rq = cfs_rq_of(se); | |
9134 | ||
9135 | if (cfs_rq_throttled(cfs_rq)) | |
9136 | break; | |
9137 | ||
9138 | update_load_avg(se, UPDATE_TG); | |
9139 | } | |
9140 | } | |
9141 | #else | |
9142 | static void propagate_entity_cfs_rq(struct sched_entity *se) { } | |
9143 | #endif | |
9144 | ||
df217913 | 9145 | static void detach_entity_cfs_rq(struct sched_entity *se) |
daa59407 | 9146 | { |
daa59407 BP |
9147 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
9148 | ||
9d89c257 | 9149 | /* Catch up with the cfs_rq and remove our load when we leave */ |
d31b1a66 | 9150 | update_load_avg(se, 0); |
a05e8c51 | 9151 | detach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 9152 | update_tg_load_avg(cfs_rq, false); |
09a43ace | 9153 | propagate_entity_cfs_rq(se); |
da7a735e PZ |
9154 | } |
9155 | ||
df217913 | 9156 | static void attach_entity_cfs_rq(struct sched_entity *se) |
cb469845 | 9157 | { |
daa59407 | 9158 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
7855a35a BP |
9159 | |
9160 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
eb7a59b2 M |
9161 | /* |
9162 | * Since the real-depth could have been changed (only FAIR | |
9163 | * class maintain depth value), reset depth properly. | |
9164 | */ | |
9165 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
9166 | #endif | |
7855a35a | 9167 | |
df217913 | 9168 | /* Synchronize entity with its cfs_rq */ |
d31b1a66 | 9169 | update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); |
daa59407 | 9170 | attach_entity_load_avg(cfs_rq, se); |
7c3edd2c | 9171 | update_tg_load_avg(cfs_rq, false); |
09a43ace | 9172 | propagate_entity_cfs_rq(se); |
df217913 VG |
9173 | } |
9174 | ||
9175 | static void detach_task_cfs_rq(struct task_struct *p) | |
9176 | { | |
9177 | struct sched_entity *se = &p->se; | |
9178 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
9179 | ||
9180 | if (!vruntime_normalized(p)) { | |
9181 | /* | |
9182 | * Fix up our vruntime so that the current sleep doesn't | |
9183 | * cause 'unlimited' sleep bonus. | |
9184 | */ | |
9185 | place_entity(cfs_rq, se, 0); | |
9186 | se->vruntime -= cfs_rq->min_vruntime; | |
9187 | } | |
9188 | ||
9189 | detach_entity_cfs_rq(se); | |
9190 | } | |
9191 | ||
9192 | static void attach_task_cfs_rq(struct task_struct *p) | |
9193 | { | |
9194 | struct sched_entity *se = &p->se; | |
9195 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
9196 | ||
9197 | attach_entity_cfs_rq(se); | |
daa59407 BP |
9198 | |
9199 | if (!vruntime_normalized(p)) | |
9200 | se->vruntime += cfs_rq->min_vruntime; | |
9201 | } | |
6efdb105 | 9202 | |
daa59407 BP |
9203 | static void switched_from_fair(struct rq *rq, struct task_struct *p) |
9204 | { | |
9205 | detach_task_cfs_rq(p); | |
9206 | } | |
9207 | ||
9208 | static void switched_to_fair(struct rq *rq, struct task_struct *p) | |
9209 | { | |
9210 | attach_task_cfs_rq(p); | |
7855a35a | 9211 | |
daa59407 | 9212 | if (task_on_rq_queued(p)) { |
7855a35a | 9213 | /* |
daa59407 BP |
9214 | * We were most likely switched from sched_rt, so |
9215 | * kick off the schedule if running, otherwise just see | |
9216 | * if we can still preempt the current task. | |
7855a35a | 9217 | */ |
daa59407 BP |
9218 | if (rq->curr == p) |
9219 | resched_curr(rq); | |
9220 | else | |
9221 | check_preempt_curr(rq, p, 0); | |
7855a35a | 9222 | } |
cb469845 SR |
9223 | } |
9224 | ||
83b699ed SV |
9225 | /* Account for a task changing its policy or group. |
9226 | * | |
9227 | * This routine is mostly called to set cfs_rq->curr field when a task | |
9228 | * migrates between groups/classes. | |
9229 | */ | |
9230 | static void set_curr_task_fair(struct rq *rq) | |
9231 | { | |
9232 | struct sched_entity *se = &rq->curr->se; | |
9233 | ||
ec12cb7f PT |
9234 | for_each_sched_entity(se) { |
9235 | struct cfs_rq *cfs_rq = cfs_rq_of(se); | |
9236 | ||
9237 | set_next_entity(cfs_rq, se); | |
9238 | /* ensure bandwidth has been allocated on our new cfs_rq */ | |
9239 | account_cfs_rq_runtime(cfs_rq, 0); | |
9240 | } | |
83b699ed SV |
9241 | } |
9242 | ||
029632fb PZ |
9243 | void init_cfs_rq(struct cfs_rq *cfs_rq) |
9244 | { | |
9245 | cfs_rq->tasks_timeline = RB_ROOT; | |
029632fb PZ |
9246 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
9247 | #ifndef CONFIG_64BIT | |
9248 | cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; | |
9249 | #endif | |
141965c7 | 9250 | #ifdef CONFIG_SMP |
09a43ace VG |
9251 | #ifdef CONFIG_FAIR_GROUP_SCHED |
9252 | cfs_rq->propagate_avg = 0; | |
9253 | #endif | |
9d89c257 YD |
9254 | atomic_long_set(&cfs_rq->removed_load_avg, 0); |
9255 | atomic_long_set(&cfs_rq->removed_util_avg, 0); | |
9ee474f5 | 9256 | #endif |
029632fb PZ |
9257 | } |
9258 | ||
810b3817 | 9259 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b VG |
9260 | static void task_set_group_fair(struct task_struct *p) |
9261 | { | |
9262 | struct sched_entity *se = &p->se; | |
9263 | ||
9264 | set_task_rq(p, task_cpu(p)); | |
9265 | se->depth = se->parent ? se->parent->depth + 1 : 0; | |
9266 | } | |
9267 | ||
bc54da21 | 9268 | static void task_move_group_fair(struct task_struct *p) |
810b3817 | 9269 | { |
daa59407 | 9270 | detach_task_cfs_rq(p); |
b2b5ce02 | 9271 | set_task_rq(p, task_cpu(p)); |
6efdb105 BP |
9272 | |
9273 | #ifdef CONFIG_SMP | |
9274 | /* Tell se's cfs_rq has been changed -- migrated */ | |
9275 | p->se.avg.last_update_time = 0; | |
9276 | #endif | |
daa59407 | 9277 | attach_task_cfs_rq(p); |
810b3817 | 9278 | } |
029632fb | 9279 | |
ea86cb4b VG |
9280 | static void task_change_group_fair(struct task_struct *p, int type) |
9281 | { | |
9282 | switch (type) { | |
9283 | case TASK_SET_GROUP: | |
9284 | task_set_group_fair(p); | |
9285 | break; | |
9286 | ||
9287 | case TASK_MOVE_GROUP: | |
9288 | task_move_group_fair(p); | |
9289 | break; | |
9290 | } | |
9291 | } | |
9292 | ||
029632fb PZ |
9293 | void free_fair_sched_group(struct task_group *tg) |
9294 | { | |
9295 | int i; | |
9296 | ||
9297 | destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
9298 | ||
9299 | for_each_possible_cpu(i) { | |
9300 | if (tg->cfs_rq) | |
9301 | kfree(tg->cfs_rq[i]); | |
6fe1f348 | 9302 | if (tg->se) |
029632fb PZ |
9303 | kfree(tg->se[i]); |
9304 | } | |
9305 | ||
9306 | kfree(tg->cfs_rq); | |
9307 | kfree(tg->se); | |
9308 | } | |
9309 | ||
9310 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
9311 | { | |
029632fb | 9312 | struct sched_entity *se; |
b7fa30c9 | 9313 | struct cfs_rq *cfs_rq; |
029632fb PZ |
9314 | int i; |
9315 | ||
9316 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); | |
9317 | if (!tg->cfs_rq) | |
9318 | goto err; | |
9319 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); | |
9320 | if (!tg->se) | |
9321 | goto err; | |
9322 | ||
9323 | tg->shares = NICE_0_LOAD; | |
9324 | ||
9325 | init_cfs_bandwidth(tg_cfs_bandwidth(tg)); | |
9326 | ||
9327 | for_each_possible_cpu(i) { | |
9328 | cfs_rq = kzalloc_node(sizeof(struct cfs_rq), | |
9329 | GFP_KERNEL, cpu_to_node(i)); | |
9330 | if (!cfs_rq) | |
9331 | goto err; | |
9332 | ||
9333 | se = kzalloc_node(sizeof(struct sched_entity), | |
9334 | GFP_KERNEL, cpu_to_node(i)); | |
9335 | if (!se) | |
9336 | goto err_free_rq; | |
9337 | ||
9338 | init_cfs_rq(cfs_rq); | |
9339 | init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); | |
540247fb | 9340 | init_entity_runnable_average(se); |
029632fb PZ |
9341 | } |
9342 | ||
9343 | return 1; | |
9344 | ||
9345 | err_free_rq: | |
9346 | kfree(cfs_rq); | |
9347 | err: | |
9348 | return 0; | |
9349 | } | |
9350 | ||
8663e24d PZ |
9351 | void online_fair_sched_group(struct task_group *tg) |
9352 | { | |
9353 | struct sched_entity *se; | |
9354 | struct rq *rq; | |
9355 | int i; | |
9356 | ||
9357 | for_each_possible_cpu(i) { | |
9358 | rq = cpu_rq(i); | |
9359 | se = tg->se[i]; | |
9360 | ||
9361 | raw_spin_lock_irq(&rq->lock); | |
4126bad6 | 9362 | update_rq_clock(rq); |
d0326691 | 9363 | attach_entity_cfs_rq(se); |
55e16d30 | 9364 | sync_throttle(tg, i); |
8663e24d PZ |
9365 | raw_spin_unlock_irq(&rq->lock); |
9366 | } | |
9367 | } | |
9368 | ||
6fe1f348 | 9369 | void unregister_fair_sched_group(struct task_group *tg) |
029632fb | 9370 | { |
029632fb | 9371 | unsigned long flags; |
6fe1f348 PZ |
9372 | struct rq *rq; |
9373 | int cpu; | |
029632fb | 9374 | |
6fe1f348 PZ |
9375 | for_each_possible_cpu(cpu) { |
9376 | if (tg->se[cpu]) | |
9377 | remove_entity_load_avg(tg->se[cpu]); | |
029632fb | 9378 | |
6fe1f348 PZ |
9379 | /* |
9380 | * Only empty task groups can be destroyed; so we can speculatively | |
9381 | * check on_list without danger of it being re-added. | |
9382 | */ | |
9383 | if (!tg->cfs_rq[cpu]->on_list) | |
9384 | continue; | |
9385 | ||
9386 | rq = cpu_rq(cpu); | |
9387 | ||
9388 | raw_spin_lock_irqsave(&rq->lock, flags); | |
9389 | list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); | |
9390 | raw_spin_unlock_irqrestore(&rq->lock, flags); | |
9391 | } | |
029632fb PZ |
9392 | } |
9393 | ||
9394 | void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, | |
9395 | struct sched_entity *se, int cpu, | |
9396 | struct sched_entity *parent) | |
9397 | { | |
9398 | struct rq *rq = cpu_rq(cpu); | |
9399 | ||
9400 | cfs_rq->tg = tg; | |
9401 | cfs_rq->rq = rq; | |
029632fb PZ |
9402 | init_cfs_rq_runtime(cfs_rq); |
9403 | ||
9404 | tg->cfs_rq[cpu] = cfs_rq; | |
9405 | tg->se[cpu] = se; | |
9406 | ||
9407 | /* se could be NULL for root_task_group */ | |
9408 | if (!se) | |
9409 | return; | |
9410 | ||
fed14d45 | 9411 | if (!parent) { |
029632fb | 9412 | se->cfs_rq = &rq->cfs; |
fed14d45 PZ |
9413 | se->depth = 0; |
9414 | } else { | |
029632fb | 9415 | se->cfs_rq = parent->my_q; |
fed14d45 PZ |
9416 | se->depth = parent->depth + 1; |
9417 | } | |
029632fb PZ |
9418 | |
9419 | se->my_q = cfs_rq; | |
0ac9b1c2 PT |
9420 | /* guarantee group entities always have weight */ |
9421 | update_load_set(&se->load, NICE_0_LOAD); | |
029632fb PZ |
9422 | se->parent = parent; |
9423 | } | |
9424 | ||
9425 | static DEFINE_MUTEX(shares_mutex); | |
9426 | ||
9427 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) | |
9428 | { | |
9429 | int i; | |
029632fb PZ |
9430 | |
9431 | /* | |
9432 | * We can't change the weight of the root cgroup. | |
9433 | */ | |
9434 | if (!tg->se[0]) | |
9435 | return -EINVAL; | |
9436 | ||
9437 | shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); | |
9438 | ||
9439 | mutex_lock(&shares_mutex); | |
9440 | if (tg->shares == shares) | |
9441 | goto done; | |
9442 | ||
9443 | tg->shares = shares; | |
9444 | for_each_possible_cpu(i) { | |
9445 | struct rq *rq = cpu_rq(i); | |
8a8c69c3 PZ |
9446 | struct sched_entity *se = tg->se[i]; |
9447 | struct rq_flags rf; | |
029632fb | 9448 | |
029632fb | 9449 | /* Propagate contribution to hierarchy */ |
8a8c69c3 | 9450 | rq_lock_irqsave(rq, &rf); |
71b1da46 | 9451 | update_rq_clock(rq); |
89ee048f VG |
9452 | for_each_sched_entity(se) { |
9453 | update_load_avg(se, UPDATE_TG); | |
9454 | update_cfs_shares(se); | |
9455 | } | |
8a8c69c3 | 9456 | rq_unlock_irqrestore(rq, &rf); |
029632fb PZ |
9457 | } |
9458 | ||
9459 | done: | |
9460 | mutex_unlock(&shares_mutex); | |
9461 | return 0; | |
9462 | } | |
9463 | #else /* CONFIG_FAIR_GROUP_SCHED */ | |
9464 | ||
9465 | void free_fair_sched_group(struct task_group *tg) { } | |
9466 | ||
9467 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
9468 | { | |
9469 | return 1; | |
9470 | } | |
9471 | ||
8663e24d PZ |
9472 | void online_fair_sched_group(struct task_group *tg) { } |
9473 | ||
6fe1f348 | 9474 | void unregister_fair_sched_group(struct task_group *tg) { } |
029632fb PZ |
9475 | |
9476 | #endif /* CONFIG_FAIR_GROUP_SCHED */ | |
9477 | ||
810b3817 | 9478 | |
6d686f45 | 9479 | static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) |
0d721cea PW |
9480 | { |
9481 | struct sched_entity *se = &task->se; | |
0d721cea PW |
9482 | unsigned int rr_interval = 0; |
9483 | ||
9484 | /* | |
9485 | * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise | |
9486 | * idle runqueue: | |
9487 | */ | |
0d721cea | 9488 | if (rq->cfs.load.weight) |
a59f4e07 | 9489 | rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); |
0d721cea PW |
9490 | |
9491 | return rr_interval; | |
9492 | } | |
9493 | ||
bf0f6f24 IM |
9494 | /* |
9495 | * All the scheduling class methods: | |
9496 | */ | |
029632fb | 9497 | const struct sched_class fair_sched_class = { |
5522d5d5 | 9498 | .next = &idle_sched_class, |
bf0f6f24 IM |
9499 | .enqueue_task = enqueue_task_fair, |
9500 | .dequeue_task = dequeue_task_fair, | |
9501 | .yield_task = yield_task_fair, | |
d95f4122 | 9502 | .yield_to_task = yield_to_task_fair, |
bf0f6f24 | 9503 | |
2e09bf55 | 9504 | .check_preempt_curr = check_preempt_wakeup, |
bf0f6f24 IM |
9505 | |
9506 | .pick_next_task = pick_next_task_fair, | |
9507 | .put_prev_task = put_prev_task_fair, | |
9508 | ||
681f3e68 | 9509 | #ifdef CONFIG_SMP |
4ce72a2c | 9510 | .select_task_rq = select_task_rq_fair, |
0a74bef8 | 9511 | .migrate_task_rq = migrate_task_rq_fair, |
141965c7 | 9512 | |
0bcdcf28 CE |
9513 | .rq_online = rq_online_fair, |
9514 | .rq_offline = rq_offline_fair, | |
88ec22d3 | 9515 | |
12695578 | 9516 | .task_dead = task_dead_fair, |
c5b28038 | 9517 | .set_cpus_allowed = set_cpus_allowed_common, |
681f3e68 | 9518 | #endif |
bf0f6f24 | 9519 | |
83b699ed | 9520 | .set_curr_task = set_curr_task_fair, |
bf0f6f24 | 9521 | .task_tick = task_tick_fair, |
cd29fe6f | 9522 | .task_fork = task_fork_fair, |
cb469845 SR |
9523 | |
9524 | .prio_changed = prio_changed_fair, | |
da7a735e | 9525 | .switched_from = switched_from_fair, |
cb469845 | 9526 | .switched_to = switched_to_fair, |
810b3817 | 9527 | |
0d721cea PW |
9528 | .get_rr_interval = get_rr_interval_fair, |
9529 | ||
6e998916 SG |
9530 | .update_curr = update_curr_fair, |
9531 | ||
810b3817 | 9532 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ea86cb4b | 9533 | .task_change_group = task_change_group_fair, |
810b3817 | 9534 | #endif |
bf0f6f24 IM |
9535 | }; |
9536 | ||
9537 | #ifdef CONFIG_SCHED_DEBUG | |
029632fb | 9538 | void print_cfs_stats(struct seq_file *m, int cpu) |
bf0f6f24 | 9539 | { |
bf0f6f24 IM |
9540 | struct cfs_rq *cfs_rq; |
9541 | ||
5973e5b9 | 9542 | rcu_read_lock(); |
c3b64f1e | 9543 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
5cef9eca | 9544 | print_cfs_rq(m, cpu, cfs_rq); |
5973e5b9 | 9545 | rcu_read_unlock(); |
bf0f6f24 | 9546 | } |
397f2378 SD |
9547 | |
9548 | #ifdef CONFIG_NUMA_BALANCING | |
9549 | void show_numa_stats(struct task_struct *p, struct seq_file *m) | |
9550 | { | |
9551 | int node; | |
9552 | unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; | |
9553 | ||
9554 | for_each_online_node(node) { | |
9555 | if (p->numa_faults) { | |
9556 | tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; | |
9557 | tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
9558 | } | |
9559 | if (p->numa_group) { | |
9560 | gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)], | |
9561 | gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)]; | |
9562 | } | |
9563 | print_numa_stats(m, node, tsf, tpf, gsf, gpf); | |
9564 | } | |
9565 | } | |
9566 | #endif /* CONFIG_NUMA_BALANCING */ | |
9567 | #endif /* CONFIG_SCHED_DEBUG */ | |
029632fb PZ |
9568 | |
9569 | __init void init_sched_fair_class(void) | |
9570 | { | |
9571 | #ifdef CONFIG_SMP | |
9572 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); | |
9573 | ||
3451d024 | 9574 | #ifdef CONFIG_NO_HZ_COMMON |
554cecaf | 9575 | nohz.next_balance = jiffies; |
029632fb | 9576 | zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); |
029632fb PZ |
9577 | #endif |
9578 | #endif /* SMP */ | |
9579 | ||
9580 | } |