blob: 8d5868771cb307c8dbfe5f2b08e6e16b1018e18b [file] [log] [blame]
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001/*
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>
Peter Zijlstra21805082007-08-25 18:41:53 +020018 *
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
Peter Zijlstra90eec102015-11-16 11:08:45 +010020 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020021 */
22
Ingo Molnar589ee622017-02-04 00:16:44 +010023#include <linux/sched/mm.h>
Ingo Molnar105ab3d2017-02-01 16:36:40 +010024#include <linux/sched/topology.h>
25
Mel Gormancb251762016-02-05 09:08:36 +000026#include <linux/latencytop.h>
Sisir Koppaka3436ae12011-03-26 18:22:55 +053027#include <linux/cpumask.h>
Nicolas Pitre83a0a962014-09-04 11:32:10 -040028#include <linux/cpuidle.h>
Peter Zijlstra029632f2011-10-25 10:00:11 +020029#include <linux/slab.h>
30#include <linux/profile.h>
31#include <linux/interrupt.h>
Peter Zijlstracbee9f82012-10-25 14:16:43 +020032#include <linux/mempolicy.h>
Mel Gormane14808b2012-11-19 10:59:15 +000033#include <linux/migrate.h>
Peter Zijlstracbee9f82012-10-25 14:16:43 +020034#include <linux/task_work.h>
Peter Zijlstra029632f2011-10-25 10:00:11 +020035
36#include <trace/events/sched.h>
37
38#include "sched.h"
Arjan van de Ven97455122008-01-25 21:08:34 +010039
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020040/*
Peter Zijlstra21805082007-08-25 18:41:53 +020041 * Targeted preemption latency for CPU-bound tasks:
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020042 *
Peter Zijlstra21805082007-08-25 18:41:53 +020043 * NOTE: this latency value is not the same as the concept of
Ingo Molnard274a4c2007-10-15 17:00:14 +020044 * 'timeslice length' - timeslices in CFS are of variable length
45 * and have no persistent notion like in traditional, time-slice
46 * based scheduling concepts.
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020047 *
Ingo Molnard274a4c2007-10-15 17:00:14 +020048 * (to see the precise effective timeslice length of your workload,
49 * run vmstat and monitor the context-switches (cs) field)
Ingo Molnar2b4d5b22016-11-23 07:37:00 +010050 *
51 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020052 */
Ingo Molnar2b4d5b22016-11-23 07:37:00 +010053unsigned int sysctl_sched_latency = 6000000ULL;
54unsigned int normalized_sysctl_sched_latency = 6000000ULL;
Ingo Molnar2bd8e6d2007-10-15 17:00:02 +020055
56/*
Christian Ehrhardt1983a922009-11-30 12:16:47 +010057 * The initial- and re-scaling of tunables is configurable
Christian Ehrhardt1983a922009-11-30 12:16:47 +010058 *
59 * Options are:
Ingo Molnar2b4d5b22016-11-23 07:37:00 +010060 *
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))
Christian Ehrhardt1983a922009-11-30 12:16:47 +010066 */
Ingo Molnar2b4d5b22016-11-23 07:37:00 +010067enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
Christian Ehrhardt1983a922009-11-30 12:16:47 +010068
69/*
Peter Zijlstrab2be5e92007-11-09 22:39:37 +010070 * Minimal preemption granularity for CPU-bound tasks:
Ingo Molnar2b4d5b22016-11-23 07:37:00 +010071 *
Takuya Yoshikawa864616e2010-10-14 16:09:13 +090072 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
Peter Zijlstrab2be5e92007-11-09 22:39:37 +010073 */
Ingo Molnar2b4d5b22016-11-23 07:37:00 +010074unsigned int sysctl_sched_min_granularity = 750000ULL;
75unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
Peter Zijlstrab2be5e92007-11-09 22:39:37 +010076
77/*
Ingo Molnar2b4d5b22016-11-23 07:37:00 +010078 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
Peter Zijlstrab2be5e92007-11-09 22:39:37 +010079 */
Ingo Molnar0bf377b2010-09-12 08:14:52 +020080static unsigned int sched_nr_latency = 8;
Peter Zijlstrab2be5e92007-11-09 22:39:37 +010081
82/*
Mike Galbraith2bba22c2009-09-09 15:41:37 +020083 * After fork, child runs first. If set to 0 (default) then
Ingo Molnar2bd8e6d2007-10-15 17:00:02 +020084 * parent will (try to) run first.
85 */
Mike Galbraith2bba22c2009-09-09 15:41:37 +020086unsigned int sysctl_sched_child_runs_first __read_mostly;
Peter Zijlstra21805082007-08-25 18:41:53 +020087
88/*
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020089 * SCHED_OTHER wake-up granularity.
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020090 *
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.
Ingo Molnar2b4d5b22016-11-23 07:37:00 +010094 *
95 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020096 */
Ingo Molnar2b4d5b22016-11-23 07:37:00 +010097unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
98unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +020099
Ingo Molnar2b4d5b22016-11-23 07:37:00 +0100100const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
Ingo Molnarda84d962007-10-15 17:00:18 +0200101
Tim Chenafe06ef2016-11-22 12:23:53 -0800102#ifdef CONFIG_SMP
103/*
104 * For asym packing, by default the lower numbered cpu has higher priority.
105 */
106int __weak arch_asym_cpu_priority(int cpu)
107{
108 return -cpu;
109}
110#endif
111
Paul Turnerec12cb72011-07-21 09:43:30 -0700112#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 *
Ingo Molnar2b4d5b22016-11-23 07:37:00 +0100121 * (default: 5 msec, units: microseconds)
122 */
123unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
Paul Turnerec12cb72011-07-21 09:43:30 -0700124#endif
125
Morten Rasmussen32731632016-07-25 14:34:26 +0100126/*
127 * The margin used when comparing utilization with CPU capacity:
Morten Rasmussen893c5d22016-10-14 14:41:12 +0100128 * util * margin < capacity * 1024
Ingo Molnar2b4d5b22016-11-23 07:37:00 +0100129 *
130 * (default: ~20%)
Morten Rasmussen32731632016-07-25 14:34:26 +0100131 */
Ingo Molnar2b4d5b22016-11-23 07:37:00 +0100132unsigned int capacity_margin = 1280;
Morten Rasmussen32731632016-07-25 14:34:26 +0100133
Paul Gortmaker85276322013-04-19 15:10:50 -0400134static inline void update_load_add(struct load_weight *lw, unsigned long inc)
135{
136 lw->weight += inc;
137 lw->inv_weight = 0;
138}
139
140static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
141{
142 lw->weight -= dec;
143 lw->inv_weight = 0;
144}
145
146static inline void update_load_set(struct load_weight *lw, unsigned long w)
147{
148 lw->weight = w;
149 lw->inv_weight = 0;
150}
151
Peter Zijlstra029632f2011-10-25 10:00:11 +0200152/*
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 */
Nicholas Mc Guire58ac93e2015-05-15 21:05:42 +0200161static unsigned int get_update_sysctl_factor(void)
Peter Zijlstra029632f2011-10-25 10:00:11 +0200162{
Nicholas Mc Guire58ac93e2015-05-15 21:05:42 +0200163 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
Peter Zijlstra029632f2011-10-25 10:00:11 +0200164 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
182static 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
194void sched_init_granularity(void)
195{
196 update_sysctl();
197}
198
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100199#define WMULT_CONST (~0U)
Peter Zijlstra029632f2011-10-25 10:00:11 +0200200#define WMULT_SHIFT 32
201
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100202static void __update_inv_weight(struct load_weight *lw)
Peter Zijlstra029632f2011-10-25 10:00:11 +0200203{
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100204 unsigned long w;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200205
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100206 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;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200215 else
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100216 lw->inv_weight = WMULT_CONST / w;
217}
Peter Zijlstra029632f2011-10-25 10:00:11 +0200218
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100219/*
220 * delta_exec * weight / lw.weight
221 * OR
222 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
223 *
Yuyang Du1c3de5e2016-03-30 07:07:51 +0800224 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100225 * 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.
230 */
231static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
232{
233 u64 fact = scale_load_down(weight);
234 int shift = WMULT_SHIFT;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200235
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100236 __update_inv_weight(lw);
237
238 if (unlikely(fact >> 32)) {
239 while (fact >> 32) {
240 fact >>= 1;
241 shift--;
242 }
Peter Zijlstra029632f2011-10-25 10:00:11 +0200243 }
244
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100245 /* hint to use a 32x32->64 mul */
246 fact = (u64)(u32)fact * lw->inv_weight;
Peter Zijlstra029632f2011-10-25 10:00:11 +0200247
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100248 while (fact >> 32) {
249 fact >>= 1;
250 shift--;
251 }
252
253 return mul_u64_u32_shr(delta_exec, fact, shift);
Peter Zijlstra029632f2011-10-25 10:00:11 +0200254}
255
256
257const struct sched_class fair_sched_class;
Peter Zijlstraa4c2f002008-10-17 19:27:03 +0200258
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200259/**************************************************************
260 * CFS operations on generic schedulable entities:
261 */
262
263#ifdef CONFIG_FAIR_GROUP_SCHED
264
265/* cpu runqueue to which this cfs_rq is attached */
266static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
267{
268 return cfs_rq->rq;
269}
270
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200271/* An entity is a task if it doesn't "own" a runqueue */
272#define entity_is_task(se) (!se->my_q)
273
Peter Zijlstra8f488942009-07-24 12:25:30 +0200274static inline struct task_struct *task_of(struct sched_entity *se)
275{
Peter Zijlstra9148a3a2016-09-20 22:34:51 +0200276 SCHED_WARN_ON(!entity_is_task(se));
Peter Zijlstra8f488942009-07-24 12:25:30 +0200277 return container_of(se, struct task_struct, se);
278}
279
Peter Zijlstrab7581492008-04-19 19:45:00 +0200280/* Walk up scheduling entities hierarchy */
281#define for_each_sched_entity(se) \
282 for (; se; se = se->parent)
283
284static 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 */
290static 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 */
296static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
297{
298 return grp->my_q;
299}
300
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -0800301static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
302{
303 if (!cfs_rq->on_list) {
Vincent Guittot9c2791f2016-11-08 10:53:43 +0100304 struct rq *rq = rq_of(cfs_rq);
305 int cpu = cpu_of(rq);
Paul Turner67e86252010-11-15 15:47:05 -0800306 /*
307 * Ensure we either appear before our parent (if already
308 * enqueued) or force our parent to appear after us when it is
Vincent Guittot9c2791f2016-11-08 10:53:43 +0100309 * 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
Paul Turner67e86252010-11-15 15:47:05 -0800314 */
315 if (cfs_rq->tg->parent &&
Vincent Guittot9c2791f2016-11-08 10:53:43 +0100316 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 */
Paul Turner67e86252010-11-15 15:47:05 -0800323 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
Vincent Guittot9c2791f2016-11-08 10:53:43 +0100324 &(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 */
336 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
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;
Paul Turner67e86252010-11-15 15:47:05 -0800357 }
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -0800358
359 cfs_rq->on_list = 1;
360 }
361}
362
363static 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
Peter Zijlstrab7581492008-04-19 19:45:00 +0200371/* Iterate thr' all leaf cfs_rq's on a runqueue */
Tejun Heoa9e7f652017-04-25 17:43:50 -0700372#define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
373 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
374 leaf_cfs_rq_list)
Peter Zijlstrab7581492008-04-19 19:45:00 +0200375
376/* Do the two (enqueued) entities belong to the same group ? */
Peter Zijlstrafed14d42012-02-11 06:05:00 +0100377static inline struct cfs_rq *
Peter Zijlstrab7581492008-04-19 19:45:00 +0200378is_same_group(struct sched_entity *se, struct sched_entity *pse)
379{
380 if (se->cfs_rq == pse->cfs_rq)
Peter Zijlstrafed14d42012-02-11 06:05:00 +0100381 return se->cfs_rq;
Peter Zijlstrab7581492008-04-19 19:45:00 +0200382
Peter Zijlstrafed14d42012-02-11 06:05:00 +0100383 return NULL;
Peter Zijlstrab7581492008-04-19 19:45:00 +0200384}
385
386static inline struct sched_entity *parent_entity(struct sched_entity *se)
387{
388 return se->parent;
389}
390
Peter Zijlstra464b7522008-10-24 11:06:15 +0200391static void
392find_matching_se(struct sched_entity **se, struct sched_entity **pse)
393{
394 int se_depth, pse_depth;
395
396 /*
397 * preemption test can be made between sibling entities who are in the
398 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
399 * both tasks until we find their ancestors who are siblings of common
400 * parent.
401 */
402
403 /* First walk up until both entities are at same depth */
Peter Zijlstrafed14d42012-02-11 06:05:00 +0100404 se_depth = (*se)->depth;
405 pse_depth = (*pse)->depth;
Peter Zijlstra464b7522008-10-24 11:06:15 +0200406
407 while (se_depth > pse_depth) {
408 se_depth--;
409 *se = parent_entity(*se);
410 }
411
412 while (pse_depth > se_depth) {
413 pse_depth--;
414 *pse = parent_entity(*pse);
415 }
416
417 while (!is_same_group(*se, *pse)) {
418 *se = parent_entity(*se);
419 *pse = parent_entity(*pse);
420 }
421}
422
Peter Zijlstra8f488942009-07-24 12:25:30 +0200423#else /* !CONFIG_FAIR_GROUP_SCHED */
424
425static inline struct task_struct *task_of(struct sched_entity *se)
426{
427 return container_of(se, struct task_struct, se);
428}
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200429
430static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
431{
432 return container_of(cfs_rq, struct rq, cfs);
433}
434
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200435#define entity_is_task(se) 1
436
Peter Zijlstrab7581492008-04-19 19:45:00 +0200437#define for_each_sched_entity(se) \
438 for (; se; se = NULL)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200439
Peter Zijlstrab7581492008-04-19 19:45:00 +0200440static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200441{
Peter Zijlstrab7581492008-04-19 19:45:00 +0200442 return &task_rq(p)->cfs;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200443}
444
Peter Zijlstrab7581492008-04-19 19:45:00 +0200445static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
446{
447 struct task_struct *p = task_of(se);
448 struct rq *rq = task_rq(p);
449
450 return &rq->cfs;
451}
452
453/* runqueue "owned" by this group */
454static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
455{
456 return NULL;
457}
458
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -0800459static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
460{
461}
462
463static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
464{
465}
466
Tejun Heoa9e7f652017-04-25 17:43:50 -0700467#define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
468 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
Peter Zijlstrab7581492008-04-19 19:45:00 +0200469
Peter Zijlstrab7581492008-04-19 19:45:00 +0200470static inline struct sched_entity *parent_entity(struct sched_entity *se)
471{
472 return NULL;
473}
474
Peter Zijlstra464b7522008-10-24 11:06:15 +0200475static inline void
476find_matching_se(struct sched_entity **se, struct sched_entity **pse)
477{
478}
479
Peter Zijlstrab7581492008-04-19 19:45:00 +0200480#endif /* CONFIG_FAIR_GROUP_SCHED */
481
Peter Zijlstra6c16a6d2012-03-21 13:07:16 -0700482static __always_inline
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100483void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200484
485/**************************************************************
486 * Scheduling class tree data structure manipulation methods:
487 */
488
Andrei Epure1bf08232013-03-12 21:12:24 +0200489static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
Peter Zijlstra02e04312007-10-15 17:00:07 +0200490{
Andrei Epure1bf08232013-03-12 21:12:24 +0200491 s64 delta = (s64)(vruntime - max_vruntime);
Peter Zijlstra368059a2007-10-15 17:00:11 +0200492 if (delta > 0)
Andrei Epure1bf08232013-03-12 21:12:24 +0200493 max_vruntime = vruntime;
Peter Zijlstra02e04312007-10-15 17:00:07 +0200494
Andrei Epure1bf08232013-03-12 21:12:24 +0200495 return max_vruntime;
Peter Zijlstra02e04312007-10-15 17:00:07 +0200496}
497
Ingo Molnar0702e3e2007-10-15 17:00:14 +0200498static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
Peter Zijlstrab0ffd242007-10-15 17:00:12 +0200499{
500 s64 delta = (s64)(vruntime - min_vruntime);
501 if (delta < 0)
502 min_vruntime = vruntime;
503
504 return min_vruntime;
505}
506
Fabio Checconi54fdc582009-07-16 12:32:27 +0200507static inline int entity_before(struct sched_entity *a,
508 struct sched_entity *b)
509{
510 return (s64)(a->vruntime - b->vruntime) < 0;
511}
512
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200513static void update_min_vruntime(struct cfs_rq *cfs_rq)
514{
Peter Zijlstrab60205c2016-09-20 21:58:12 +0200515 struct sched_entity *curr = cfs_rq->curr;
516
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200517 u64 vruntime = cfs_rq->min_vruntime;
518
Peter Zijlstrab60205c2016-09-20 21:58:12 +0200519 if (curr) {
520 if (curr->on_rq)
521 vruntime = curr->vruntime;
522 else
523 curr = NULL;
524 }
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200525
526 if (cfs_rq->rb_leftmost) {
527 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
528 struct sched_entity,
529 run_node);
530
Peter Zijlstrab60205c2016-09-20 21:58:12 +0200531 if (!curr)
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200532 vruntime = se->vruntime;
533 else
534 vruntime = min_vruntime(vruntime, se->vruntime);
535 }
536
Andrei Epure1bf08232013-03-12 21:12:24 +0200537 /* ensure we never gain time by being placed backwards. */
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200538 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
Peter Zijlstra3fe16982011-04-05 17:23:48 +0200539#ifndef CONFIG_64BIT
540 smp_wmb();
541 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
542#endif
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200543}
544
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200545/*
546 * Enqueue an entity into the rb-tree:
547 */
Ingo Molnar0702e3e2007-10-15 17:00:14 +0200548static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200549{
550 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
551 struct rb_node *parent = NULL;
552 struct sched_entity *entry;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200553 int leftmost = 1;
554
555 /*
556 * Find the right place in the rbtree:
557 */
558 while (*link) {
559 parent = *link;
560 entry = rb_entry(parent, struct sched_entity, run_node);
561 /*
562 * We dont care about collisions. Nodes with
563 * the same key stay together.
564 */
Stephan Baerwolf2bd2d6f2011-07-20 14:46:59 +0200565 if (entity_before(se, entry)) {
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200566 link = &parent->rb_left;
567 } else {
568 link = &parent->rb_right;
569 leftmost = 0;
570 }
571 }
572
573 /*
574 * Maintain a cache of leftmost tree entries (it is frequently
575 * used):
576 */
Peter Zijlstra1af5f732008-10-24 11:06:13 +0200577 if (leftmost)
Ingo Molnar57cb4992007-10-15 17:00:11 +0200578 cfs_rq->rb_leftmost = &se->run_node;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200579
580 rb_link_node(&se->run_node, parent, link);
581 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200582}
583
Ingo Molnar0702e3e2007-10-15 17:00:14 +0200584static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200585{
Peter Zijlstra3fe69742008-03-14 20:55:51 +0100586 if (cfs_rq->rb_leftmost == &se->run_node) {
587 struct rb_node *next_node;
Peter Zijlstra3fe69742008-03-14 20:55:51 +0100588
589 next_node = rb_next(&se->run_node);
590 cfs_rq->rb_leftmost = next_node;
Peter Zijlstra3fe69742008-03-14 20:55:51 +0100591 }
Ingo Molnare9acbff2007-10-15 17:00:04 +0200592
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200593 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200594}
595
Peter Zijlstra029632f2011-10-25 10:00:11 +0200596struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200597{
Peter Zijlstraf4b67552008-11-04 21:25:07 +0100598 struct rb_node *left = cfs_rq->rb_leftmost;
599
600 if (!left)
601 return NULL;
602
603 return rb_entry(left, struct sched_entity, run_node);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200604}
605
Rik van Rielac53db52011-02-01 09:51:03 -0500606static struct sched_entity *__pick_next_entity(struct sched_entity *se)
607{
608 struct rb_node *next = rb_next(&se->run_node);
609
610 if (!next)
611 return NULL;
612
613 return rb_entry(next, struct sched_entity, run_node);
614}
615
616#ifdef CONFIG_SCHED_DEBUG
Peter Zijlstra029632f2011-10-25 10:00:11 +0200617struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200618{
Ingo Molnar7eee3e62008-02-22 10:32:21 +0100619 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200620
Balbir Singh70eee742008-02-22 13:25:53 +0530621 if (!last)
622 return NULL;
Ingo Molnar7eee3e62008-02-22 10:32:21 +0100623
624 return rb_entry(last, struct sched_entity, run_node);
Peter Zijlstraaeb73b02007-10-15 17:00:05 +0200625}
626
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200627/**************************************************************
628 * Scheduling class statistics methods:
629 */
630
Christian Ehrhardtacb4a842009-11-30 12:16:48 +0100631int sched_proc_update_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700632 void __user *buffer, size_t *lenp,
Peter Zijlstrab2be5e92007-11-09 22:39:37 +0100633 loff_t *ppos)
634{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700635 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Nicholas Mc Guire58ac93e2015-05-15 21:05:42 +0200636 unsigned int factor = get_update_sysctl_factor();
Peter Zijlstrab2be5e92007-11-09 22:39:37 +0100637
638 if (ret || !write)
639 return ret;
640
641 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
642 sysctl_sched_min_granularity);
643
Christian Ehrhardtacb4a842009-11-30 12:16:48 +0100644#define WRT_SYSCTL(name) \
645 (normalized_sysctl_##name = sysctl_##name / (factor))
646 WRT_SYSCTL(sched_min_granularity);
647 WRT_SYSCTL(sched_latency);
648 WRT_SYSCTL(sched_wakeup_granularity);
Christian Ehrhardtacb4a842009-11-30 12:16:48 +0100649#undef WRT_SYSCTL
650
Peter Zijlstrab2be5e92007-11-09 22:39:37 +0100651 return 0;
652}
653#endif
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200654
655/*
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200656 * delta /= w
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200657 */
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100658static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200659{
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200660 if (unlikely(se->load.weight != NICE_0_LOAD))
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100661 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200662
663 return delta;
664}
665
666/*
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200667 * The idea is to set a period in which each task runs once.
668 *
Borislav Petkov532b1852012-08-08 16:16:04 +0200669 * When there are too many tasks (sched_nr_latency) we have to stretch
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200670 * this period because otherwise the slices get too small.
671 *
672 * p = (nr <= nl) ? l : l*nr/nl
673 */
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +0200674static u64 __sched_period(unsigned long nr_running)
675{
Boqun Feng8e2b0bf2015-07-02 22:25:52 +0800676 if (unlikely(nr_running > sched_nr_latency))
677 return nr_running * sysctl_sched_min_granularity;
678 else
679 return sysctl_sched_latency;
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +0200680}
681
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200682/*
683 * We calculate the wall-time slice from the period by taking a part
684 * proportional to the weight.
685 *
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200686 * s = p*P[w/rw]
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200687 */
Peter Zijlstra6d0f0eb2007-10-15 17:00:05 +0200688static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
Peter Zijlstra21805082007-08-25 18:41:53 +0200689{
Mike Galbraith0a582442009-01-02 12:16:42 +0100690 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200691
Mike Galbraith0a582442009-01-02 12:16:42 +0100692 for_each_sched_entity(se) {
Lin Ming6272d682009-01-15 17:17:15 +0100693 struct load_weight *load;
Christian Engelmayer3104bf02009-06-16 10:35:12 +0200694 struct load_weight lw;
Lin Ming6272d682009-01-15 17:17:15 +0100695
696 cfs_rq = cfs_rq_of(se);
697 load = &cfs_rq->load;
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200698
Mike Galbraith0a582442009-01-02 12:16:42 +0100699 if (unlikely(!se->on_rq)) {
Christian Engelmayer3104bf02009-06-16 10:35:12 +0200700 lw = cfs_rq->load;
Mike Galbraith0a582442009-01-02 12:16:42 +0100701
702 update_load_add(&lw, se->load.weight);
703 load = &lw;
704 }
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100705 slice = __calc_delta(slice, se->load.weight, load);
Mike Galbraith0a582442009-01-02 12:16:42 +0100706 }
707 return slice;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200708}
709
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200710/*
Andrei Epure660cc002013-03-11 12:03:20 +0200711 * We calculate the vruntime slice of a to-be-inserted task.
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200712 *
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200713 * vs = s/w
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200714 */
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200715static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnar647e7ca2007-10-15 17:00:13 +0200716{
Peter Zijlstraf9c0b092008-10-17 19:27:04 +0200717 return calc_delta_fair(sched_slice(cfs_rq, se), se);
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200718}
719
Alex Shia75cdaa2013-06-20 10:18:47 +0800720#ifdef CONFIG_SMP
Peter Zijlstra283e2ed2017-04-11 11:08:42 +0200721
722#include "sched-pelt.h"
723
Morten Rasmussen772bd008c2016-06-22 18:03:13 +0100724static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
Mel Gormanfb13c7e2013-10-07 11:29:17 +0100725static unsigned long task_h_load(struct task_struct *p);
726
Yuyang Du540247f2015-07-15 08:04:39 +0800727/* Give new sched_entity start runnable values to heavy its load in infant time */
728void init_entity_runnable_average(struct sched_entity *se)
Alex Shia75cdaa2013-06-20 10:18:47 +0800729{
Yuyang Du540247f2015-07-15 08:04:39 +0800730 struct sched_avg *sa = &se->avg;
Alex Shia75cdaa2013-06-20 10:18:47 +0800731
Yuyang Du9d89c252015-07-15 08:04:37 +0800732 sa->last_update_time = 0;
733 /*
734 * sched_avg's period_contrib should be strictly less then 1024, so
735 * we give it 1023 to make sure it is almost a period (1024us), and
736 * will definitely be update (after enqueue).
737 */
738 sa->period_contrib = 1023;
Vincent Guittotb5a9b342016-10-19 14:45:23 +0200739 /*
740 * Tasks are intialized with full load to be seen as heavy tasks until
741 * they get a chance to stabilize to their real load level.
742 * Group entities are intialized with zero load to reflect the fact that
743 * nothing has been attached to the task group yet.
744 */
745 if (entity_is_task(se))
746 sa->load_avg = scale_load_down(se->load.weight);
Yuyang Du9d89c252015-07-15 08:04:37 +0800747 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
Yuyang Du2b8c41d2016-03-30 04:30:56 +0800748 /*
749 * At this point, util_avg won't be used in select_task_rq_fair anyway
750 */
751 sa->util_avg = 0;
752 sa->util_sum = 0;
Yuyang Du9d89c252015-07-15 08:04:37 +0800753 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
Alex Shia75cdaa2013-06-20 10:18:47 +0800754}
Yuyang Du7ea241a2015-07-15 08:04:42 +0800755
Peter Zijlstra7dc603c2016-06-16 13:29:28 +0200756static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
Vincent Guittotdf217912016-11-08 10:53:42 +0100757static void attach_entity_cfs_rq(struct sched_entity *se);
Peter Zijlstra7dc603c2016-06-16 13:29:28 +0200758
Yuyang Du2b8c41d2016-03-30 04:30:56 +0800759/*
760 * With new tasks being created, their initial util_avgs are extrapolated
761 * based on the cfs_rq's current util_avg:
762 *
763 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
764 *
765 * However, in many cases, the above util_avg does not give a desired
766 * value. Moreover, the sum of the util_avgs may be divergent, such
767 * as when the series is a harmonic series.
768 *
769 * To solve this problem, we also cap the util_avg of successive tasks to
770 * only 1/2 of the left utilization budget:
771 *
772 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
773 *
774 * where n denotes the nth task.
775 *
776 * For example, a simplest series from the beginning would be like:
777 *
778 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
779 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
780 *
781 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
782 * if util_avg > util_avg_cap.
783 */
784void post_init_entity_util_avg(struct sched_entity *se)
785{
786 struct cfs_rq *cfs_rq = cfs_rq_of(se);
787 struct sched_avg *sa = &se->avg;
Yuyang Du172895e2016-04-05 12:12:27 +0800788 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
Yuyang Du2b8c41d2016-03-30 04:30:56 +0800789
790 if (cap > 0) {
791 if (cfs_rq->avg.util_avg != 0) {
792 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
793 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
794
795 if (sa->util_avg > cap)
796 sa->util_avg = cap;
797 } else {
798 sa->util_avg = cap;
799 }
800 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
801 }
Peter Zijlstra7dc603c2016-06-16 13:29:28 +0200802
803 if (entity_is_task(se)) {
804 struct task_struct *p = task_of(se);
805 if (p->sched_class != &fair_sched_class) {
806 /*
807 * For !fair tasks do:
808 *
Viresh Kumar3a123bb2017-05-24 10:59:56 +0530809 update_cfs_rq_load_avg(now, cfs_rq);
Peter Zijlstra7dc603c2016-06-16 13:29:28 +0200810 attach_entity_load_avg(cfs_rq, se);
811 switched_from_fair(rq, p);
812 *
813 * such that the next switched_to_fair() has the
814 * expected state.
815 */
Vincent Guittotdf217912016-11-08 10:53:42 +0100816 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
Peter Zijlstra7dc603c2016-06-16 13:29:28 +0200817 return;
818 }
819 }
820
Vincent Guittotdf217912016-11-08 10:53:42 +0100821 attach_entity_cfs_rq(se);
Yuyang Du2b8c41d2016-03-30 04:30:56 +0800822}
823
Peter Zijlstra7dc603c2016-06-16 13:29:28 +0200824#else /* !CONFIG_SMP */
Yuyang Du540247f2015-07-15 08:04:39 +0800825void init_entity_runnable_average(struct sched_entity *se)
Alex Shia75cdaa2013-06-20 10:18:47 +0800826{
827}
Yuyang Du2b8c41d2016-03-30 04:30:56 +0800828void post_init_entity_util_avg(struct sched_entity *se)
829{
830}
Peter Zijlstra3d30544f2016-06-21 14:27:50 +0200831static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
832{
833}
Peter Zijlstra7dc603c2016-06-16 13:29:28 +0200834#endif /* CONFIG_SMP */
Alex Shia75cdaa2013-06-20 10:18:47 +0800835
Peter Zijlstraa7be37a2008-06-27 13:41:11 +0200836/*
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100837 * Update the current task's runtime statistics.
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200838 */
Ingo Molnarb7cc0892007-08-09 11:16:47 +0200839static void update_curr(struct cfs_rq *cfs_rq)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200840{
Ingo Molnar429d43b2007-10-15 17:00:03 +0200841 struct sched_entity *curr = cfs_rq->curr;
Frederic Weisbecker78becc22013-04-12 01:51:02 +0200842 u64 now = rq_clock_task(rq_of(cfs_rq));
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100843 u64 delta_exec;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200844
845 if (unlikely(!curr))
846 return;
847
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100848 delta_exec = now - curr->exec_start;
849 if (unlikely((s64)delta_exec <= 0))
Peter Zijlstra34f28ec2008-12-16 08:45:31 +0100850 return;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200851
Ingo Molnar8ebc91d2007-10-15 17:00:03 +0200852 curr->exec_start = now;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100853
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100854 schedstat_set(curr->statistics.exec_max,
855 max(delta_exec, curr->statistics.exec_max));
856
857 curr->sum_exec_runtime += delta_exec;
Josh Poimboeufae928822016-06-17 12:43:24 -0500858 schedstat_add(cfs_rq->exec_clock, delta_exec);
Peter Zijlstra9dbdb152013-11-18 18:27:06 +0100859
860 curr->vruntime += calc_delta_fair(delta_exec, curr);
861 update_min_vruntime(cfs_rq);
862
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100863 if (entity_is_task(curr)) {
864 struct task_struct *curtask = task_of(curr);
865
Ingo Molnarf977bb42009-09-13 18:15:54 +0200866 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100867 cpuacct_charge(curtask, delta_exec);
Frank Mayharf06febc2008-09-12 09:54:39 -0700868 account_group_exec_runtime(curtask, delta_exec);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +0100869 }
Paul Turnerec12cb72011-07-21 09:43:30 -0700870
871 account_cfs_rq_runtime(cfs_rq, delta_exec);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200872}
873
Stanislaw Gruszka6e998912014-11-12 16:58:44 +0100874static void update_curr_fair(struct rq *rq)
875{
876 update_curr(cfs_rq_of(&rq->curr->se));
877}
878
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200879static inline void
Ingo Molnar5870db52007-08-09 11:16:47 +0200880update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200881{
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500882 u64 wait_start, prev_wait_start;
883
884 if (!schedstat_enabled())
885 return;
886
887 wait_start = rq_clock(rq_of(cfs_rq));
888 prev_wait_start = schedstat_val(se->statistics.wait_start);
Joonwoo Park3ea94de2015-11-12 19:38:54 -0800889
890 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500891 likely(wait_start > prev_wait_start))
892 wait_start -= prev_wait_start;
Joonwoo Park3ea94de2015-11-12 19:38:54 -0800893
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500894 schedstat_set(se->statistics.wait_start, wait_start);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200895}
896
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500897static inline void
Joonwoo Park3ea94de2015-11-12 19:38:54 -0800898update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
899{
900 struct task_struct *p;
Mel Gormancb251762016-02-05 09:08:36 +0000901 u64 delta;
902
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500903 if (!schedstat_enabled())
904 return;
905
906 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
Joonwoo Park3ea94de2015-11-12 19:38:54 -0800907
908 if (entity_is_task(se)) {
909 p = task_of(se);
910 if (task_on_rq_migrating(p)) {
911 /*
912 * Preserve migrating task's wait time so wait_start
913 * time stamp can be adjusted to accumulate wait time
914 * prior to migration.
915 */
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500916 schedstat_set(se->statistics.wait_start, delta);
Joonwoo Park3ea94de2015-11-12 19:38:54 -0800917 return;
918 }
919 trace_sched_stat_wait(p, delta);
920 }
921
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500922 schedstat_set(se->statistics.wait_max,
923 max(schedstat_val(se->statistics.wait_max), delta));
924 schedstat_inc(se->statistics.wait_count);
925 schedstat_add(se->statistics.wait_sum, delta);
926 schedstat_set(se->statistics.wait_start, 0);
Joonwoo Park3ea94de2015-11-12 19:38:54 -0800927}
Joonwoo Park3ea94de2015-11-12 19:38:54 -0800928
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500929static inline void
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -0500930update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
931{
932 struct task_struct *tsk = NULL;
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500933 u64 sleep_start, block_start;
934
935 if (!schedstat_enabled())
936 return;
937
938 sleep_start = schedstat_val(se->statistics.sleep_start);
939 block_start = schedstat_val(se->statistics.block_start);
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -0500940
941 if (entity_is_task(se))
942 tsk = task_of(se);
943
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500944 if (sleep_start) {
945 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -0500946
947 if ((s64)delta < 0)
948 delta = 0;
949
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500950 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
951 schedstat_set(se->statistics.sleep_max, delta);
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -0500952
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500953 schedstat_set(se->statistics.sleep_start, 0);
954 schedstat_add(se->statistics.sum_sleep_runtime, delta);
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -0500955
956 if (tsk) {
957 account_scheduler_latency(tsk, delta >> 10, 1);
958 trace_sched_stat_sleep(tsk, delta);
959 }
960 }
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500961 if (block_start) {
962 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -0500963
964 if ((s64)delta < 0)
965 delta = 0;
966
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500967 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
968 schedstat_set(se->statistics.block_max, delta);
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -0500969
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500970 schedstat_set(se->statistics.block_start, 0);
971 schedstat_add(se->statistics.sum_sleep_runtime, delta);
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -0500972
973 if (tsk) {
974 if (tsk->in_iowait) {
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -0500975 schedstat_add(se->statistics.iowait_sum, delta);
976 schedstat_inc(se->statistics.iowait_count);
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -0500977 trace_sched_stat_iowait(tsk, delta);
978 }
979
980 trace_sched_stat_blocked(tsk, delta);
981
982 /*
983 * Blocking time is in units of nanosecs, so shift by
984 * 20 to get a milliseconds-range estimation of the
985 * amount of time that the task spent sleeping:
986 */
987 if (unlikely(prof_on == SLEEP_PROFILING)) {
988 profile_hits(SLEEP_PROFILING,
989 (void *)get_wchan(tsk),
990 delta >> 20);
991 }
992 account_scheduler_latency(tsk, delta >> 10, 0);
993 }
994 }
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200995}
996
Ingo Molnarbf0f6f22007-07-09 18:51:58 +0200997/*
998 * Task is being enqueued - update stats:
999 */
Mel Gormancb251762016-02-05 09:08:36 +00001000static inline void
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -05001001update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001002{
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05001003 if (!schedstat_enabled())
1004 return;
1005
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001006 /*
1007 * Are we enqueueing a waiting task? (for current tasks
1008 * a dequeue/enqueue event is a NOP)
1009 */
Ingo Molnar429d43b2007-10-15 17:00:03 +02001010 if (se != cfs_rq->curr)
Ingo Molnar5870db52007-08-09 11:16:47 +02001011 update_stats_wait_start(cfs_rq, se);
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -05001012
1013 if (flags & ENQUEUE_WAKEUP)
1014 update_stats_enqueue_sleeper(cfs_rq, se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001015}
1016
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001017static inline void
Mel Gormancb251762016-02-05 09:08:36 +00001018update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001019{
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05001020
1021 if (!schedstat_enabled())
1022 return;
1023
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001024 /*
1025 * Mark the end of the wait period if dequeueing a
1026 * waiting task:
1027 */
Ingo Molnar429d43b2007-10-15 17:00:03 +02001028 if (se != cfs_rq->curr)
Ingo Molnar9ef0a962007-08-09 11:16:47 +02001029 update_stats_wait_end(cfs_rq, se);
Mel Gormancb251762016-02-05 09:08:36 +00001030
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05001031 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1032 struct task_struct *tsk = task_of(se);
Mel Gormancb251762016-02-05 09:08:36 +00001033
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05001034 if (tsk->state & TASK_INTERRUPTIBLE)
1035 schedstat_set(se->statistics.sleep_start,
1036 rq_clock(rq_of(cfs_rq)));
1037 if (tsk->state & TASK_UNINTERRUPTIBLE)
1038 schedstat_set(se->statistics.block_start,
1039 rq_clock(rq_of(cfs_rq)));
Mel Gormancb251762016-02-05 09:08:36 +00001040 }
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001041}
1042
1043/*
1044 * We are picking a new current task - update its stats:
1045 */
1046static inline void
Ingo Molnar79303e92007-08-09 11:16:47 +02001047update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001048{
1049 /*
1050 * We are starting a new run period:
1051 */
Frederic Weisbecker78becc22013-04-12 01:51:02 +02001052 se->exec_start = rq_clock_task(rq_of(cfs_rq));
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001053}
1054
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02001055/**************************************************
1056 * Scheduling class queueing methods:
1057 */
1058
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001059#ifdef CONFIG_NUMA_BALANCING
1060/*
Mel Gorman598f0ec2013-10-07 11:28:55 +01001061 * Approximate time to scan a full NUMA task in ms. The task scan period is
1062 * calculated based on the tasks virtual memory size and
1063 * numa_balancing_scan_size.
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001064 */
Mel Gorman598f0ec2013-10-07 11:28:55 +01001065unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1066unsigned int sysctl_numa_balancing_scan_period_max = 60000;
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02001067
1068/* Portion of address space to scan in MB */
1069unsigned int sysctl_numa_balancing_scan_size = 256;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02001070
Peter Zijlstra4b96a292012-10-25 14:16:47 +02001071/* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1072unsigned int sysctl_numa_balancing_scan_delay = 1000;
1073
Rik van Rielb5dd77c2017-07-31 15:28:47 -04001074struct numa_group {
1075 atomic_t refcount;
1076
1077 spinlock_t lock; /* nr_tasks, tasks */
1078 int nr_tasks;
1079 pid_t gid;
1080 int active_nodes;
1081
1082 struct rcu_head rcu;
1083 unsigned long total_faults;
1084 unsigned long max_faults_cpu;
1085 /*
1086 * Faults_cpu is used to decide whether memory should move
1087 * towards the CPU. As a consequence, these stats are weighted
1088 * more by CPU use than by memory faults.
1089 */
1090 unsigned long *faults_cpu;
1091 unsigned long faults[0];
1092};
1093
1094static inline unsigned long group_faults_priv(struct numa_group *ng);
1095static inline unsigned long group_faults_shared(struct numa_group *ng);
1096
Mel Gorman598f0ec2013-10-07 11:28:55 +01001097static unsigned int task_nr_scan_windows(struct task_struct *p)
1098{
1099 unsigned long rss = 0;
1100 unsigned long nr_scan_pages;
1101
1102 /*
1103 * Calculations based on RSS as non-present and empty pages are skipped
1104 * by the PTE scanner and NUMA hinting faults should be trapped based
1105 * on resident pages
1106 */
1107 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1108 rss = get_mm_rss(p->mm);
1109 if (!rss)
1110 rss = nr_scan_pages;
1111
1112 rss = round_up(rss, nr_scan_pages);
1113 return rss / nr_scan_pages;
1114}
1115
1116/* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1117#define MAX_SCAN_WINDOW 2560
1118
1119static unsigned int task_scan_min(struct task_struct *p)
1120{
Jason Low316c1608d2015-04-28 13:00:20 -07001121 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
Mel Gorman598f0ec2013-10-07 11:28:55 +01001122 unsigned int scan, floor;
1123 unsigned int windows = 1;
1124
Kirill Tkhai64192652014-10-16 14:39:37 +04001125 if (scan_size < MAX_SCAN_WINDOW)
1126 windows = MAX_SCAN_WINDOW / scan_size;
Mel Gorman598f0ec2013-10-07 11:28:55 +01001127 floor = 1000 / windows;
1128
1129 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1130 return max_t(unsigned int, floor, scan);
1131}
1132
Rik van Rielb5dd77c2017-07-31 15:28:47 -04001133static unsigned int task_scan_start(struct task_struct *p)
1134{
1135 unsigned long smin = task_scan_min(p);
1136 unsigned long period = smin;
1137
1138 /* Scale the maximum scan period with the amount of shared memory. */
1139 if (p->numa_group) {
1140 struct numa_group *ng = p->numa_group;
1141 unsigned long shared = group_faults_shared(ng);
1142 unsigned long private = group_faults_priv(ng);
1143
1144 period *= atomic_read(&ng->refcount);
1145 period *= shared + 1;
1146 period /= private + shared + 1;
1147 }
1148
1149 return max(smin, period);
1150}
1151
Mel Gorman598f0ec2013-10-07 11:28:55 +01001152static unsigned int task_scan_max(struct task_struct *p)
1153{
Rik van Rielb5dd77c2017-07-31 15:28:47 -04001154 unsigned long smin = task_scan_min(p);
1155 unsigned long smax;
Mel Gorman598f0ec2013-10-07 11:28:55 +01001156
1157 /* Watch for min being lower than max due to floor calculations */
1158 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
Rik van Rielb5dd77c2017-07-31 15:28:47 -04001159
1160 /* Scale the maximum scan period with the amount of shared memory. */
1161 if (p->numa_group) {
1162 struct numa_group *ng = p->numa_group;
1163 unsigned long shared = group_faults_shared(ng);
1164 unsigned long private = group_faults_priv(ng);
1165 unsigned long period = smax;
1166
1167 period *= atomic_read(&ng->refcount);
1168 period *= shared + 1;
1169 period /= private + shared + 1;
1170
1171 smax = max(smax, period);
1172 }
1173
Mel Gorman598f0ec2013-10-07 11:28:55 +01001174 return max(smin, smax);
1175}
1176
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01001177static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1178{
1179 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1180 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1181}
1182
1183static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1184{
1185 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1186 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1187}
1188
Rik van Rielbe1e4e72014-01-27 17:03:48 -05001189/* Shared or private faults. */
1190#define NR_NUMA_HINT_FAULT_TYPES 2
1191
1192/* Memory and CPU locality */
1193#define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1194
1195/* Averaged statistics, and temporary buffers. */
1196#define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1197
Mel Gormane29cf082013-10-07 11:29:22 +01001198pid_t task_numa_group_id(struct task_struct *p)
1199{
1200 return p->numa_group ? p->numa_group->gid : 0;
1201}
1202
Iulia Manda44dba3d2014-10-31 02:13:31 +02001203/*
1204 * The averaged statistics, shared & private, memory & cpu,
1205 * occupy the first half of the array. The second half of the
1206 * array is for current counters, which are averaged into the
1207 * first set by task_numa_placement.
1208 */
1209static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
Mel Gormanac8e8952013-10-07 11:29:03 +01001210{
Iulia Manda44dba3d2014-10-31 02:13:31 +02001211 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
Mel Gormanac8e8952013-10-07 11:29:03 +01001212}
1213
1214static inline unsigned long task_faults(struct task_struct *p, int nid)
1215{
Iulia Manda44dba3d2014-10-31 02:13:31 +02001216 if (!p->numa_faults)
Mel Gormanac8e8952013-10-07 11:29:03 +01001217 return 0;
1218
Iulia Manda44dba3d2014-10-31 02:13:31 +02001219 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1220 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
Mel Gormanac8e8952013-10-07 11:29:03 +01001221}
1222
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001223static inline unsigned long group_faults(struct task_struct *p, int nid)
1224{
1225 if (!p->numa_group)
1226 return 0;
1227
Iulia Manda44dba3d2014-10-31 02:13:31 +02001228 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1229 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001230}
1231
Rik van Riel20e07de2014-01-27 17:03:43 -05001232static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1233{
Iulia Manda44dba3d2014-10-31 02:13:31 +02001234 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1235 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
Rik van Riel20e07de2014-01-27 17:03:43 -05001236}
1237
Rik van Rielb5dd77c2017-07-31 15:28:47 -04001238static inline unsigned long group_faults_priv(struct numa_group *ng)
1239{
1240 unsigned long faults = 0;
1241 int node;
1242
1243 for_each_online_node(node) {
1244 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1245 }
1246
1247 return faults;
1248}
1249
1250static inline unsigned long group_faults_shared(struct numa_group *ng)
1251{
1252 unsigned long faults = 0;
1253 int node;
1254
1255 for_each_online_node(node) {
1256 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1257 }
1258
1259 return faults;
1260}
1261
Rik van Riel4142c3e2016-01-25 17:07:39 -05001262/*
1263 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1264 * considered part of a numa group's pseudo-interleaving set. Migrations
1265 * between these nodes are slowed down, to allow things to settle down.
1266 */
1267#define ACTIVE_NODE_FRACTION 3
1268
1269static bool numa_is_active_node(int nid, struct numa_group *ng)
1270{
1271 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1272}
1273
Rik van Riel6c6b1192014-10-17 03:29:52 -04001274/* Handle placement on systems where not all nodes are directly connected. */
1275static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1276 int maxdist, bool task)
1277{
1278 unsigned long score = 0;
1279 int node;
1280
1281 /*
1282 * All nodes are directly connected, and the same distance
1283 * from each other. No need for fancy placement algorithms.
1284 */
1285 if (sched_numa_topology_type == NUMA_DIRECT)
1286 return 0;
1287
1288 /*
1289 * This code is called for each node, introducing N^2 complexity,
1290 * which should be ok given the number of nodes rarely exceeds 8.
1291 */
1292 for_each_online_node(node) {
1293 unsigned long faults;
1294 int dist = node_distance(nid, node);
1295
1296 /*
1297 * The furthest away nodes in the system are not interesting
1298 * for placement; nid was already counted.
1299 */
1300 if (dist == sched_max_numa_distance || node == nid)
1301 continue;
1302
1303 /*
1304 * On systems with a backplane NUMA topology, compare groups
1305 * of nodes, and move tasks towards the group with the most
1306 * memory accesses. When comparing two nodes at distance
1307 * "hoplimit", only nodes closer by than "hoplimit" are part
1308 * of each group. Skip other nodes.
1309 */
1310 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1311 dist > maxdist)
1312 continue;
1313
1314 /* Add up the faults from nearby nodes. */
1315 if (task)
1316 faults = task_faults(p, node);
1317 else
1318 faults = group_faults(p, node);
1319
1320 /*
1321 * On systems with a glueless mesh NUMA topology, there are
1322 * no fixed "groups of nodes". Instead, nodes that are not
1323 * directly connected bounce traffic through intermediate
1324 * nodes; a numa_group can occupy any set of nodes.
1325 * The further away a node is, the less the faults count.
1326 * This seems to result in good task placement.
1327 */
1328 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1329 faults *= (sched_max_numa_distance - dist);
1330 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1331 }
1332
1333 score += faults;
1334 }
1335
1336 return score;
1337}
1338
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001339/*
1340 * These return the fraction of accesses done by a particular task, or
1341 * task group, on a particular numa node. The group weight is given a
1342 * larger multiplier, in order to group tasks together that are almost
1343 * evenly spread out between numa nodes.
1344 */
Rik van Riel7bd95322014-10-17 03:29:51 -04001345static inline unsigned long task_weight(struct task_struct *p, int nid,
1346 int dist)
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001347{
Rik van Riel7bd95322014-10-17 03:29:51 -04001348 unsigned long faults, total_faults;
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001349
Iulia Manda44dba3d2014-10-31 02:13:31 +02001350 if (!p->numa_faults)
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001351 return 0;
1352
1353 total_faults = p->total_numa_faults;
1354
1355 if (!total_faults)
1356 return 0;
1357
Rik van Riel7bd95322014-10-17 03:29:51 -04001358 faults = task_faults(p, nid);
Rik van Riel6c6b1192014-10-17 03:29:52 -04001359 faults += score_nearby_nodes(p, nid, dist, true);
1360
Rik van Riel7bd95322014-10-17 03:29:51 -04001361 return 1000 * faults / total_faults;
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001362}
1363
Rik van Riel7bd95322014-10-17 03:29:51 -04001364static inline unsigned long group_weight(struct task_struct *p, int nid,
1365 int dist)
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001366{
Rik van Riel7bd95322014-10-17 03:29:51 -04001367 unsigned long faults, total_faults;
1368
1369 if (!p->numa_group)
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001370 return 0;
1371
Rik van Riel7bd95322014-10-17 03:29:51 -04001372 total_faults = p->numa_group->total_faults;
1373
1374 if (!total_faults)
1375 return 0;
1376
1377 faults = group_faults(p, nid);
Rik van Riel6c6b1192014-10-17 03:29:52 -04001378 faults += score_nearby_nodes(p, nid, dist, false);
1379
Rik van Riel7bd95322014-10-17 03:29:51 -04001380 return 1000 * faults / total_faults;
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001381}
1382
Rik van Riel10f39042014-01-27 17:03:44 -05001383bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1384 int src_nid, int dst_cpu)
1385{
1386 struct numa_group *ng = p->numa_group;
1387 int dst_nid = cpu_to_node(dst_cpu);
1388 int last_cpupid, this_cpupid;
1389
1390 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1391
1392 /*
1393 * Multi-stage node selection is used in conjunction with a periodic
1394 * migration fault to build a temporal task<->page relation. By using
1395 * a two-stage filter we remove short/unlikely relations.
1396 *
1397 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1398 * a task's usage of a particular page (n_p) per total usage of this
1399 * page (n_t) (in a given time-span) to a probability.
1400 *
1401 * Our periodic faults will sample this probability and getting the
1402 * same result twice in a row, given these samples are fully
1403 * independent, is then given by P(n)^2, provided our sample period
1404 * is sufficiently short compared to the usage pattern.
1405 *
1406 * This quadric squishes small probabilities, making it less likely we
1407 * act on an unlikely task<->page relation.
1408 */
1409 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1410 if (!cpupid_pid_unset(last_cpupid) &&
1411 cpupid_to_nid(last_cpupid) != dst_nid)
1412 return false;
1413
1414 /* Always allow migrate on private faults */
1415 if (cpupid_match_pid(p, last_cpupid))
1416 return true;
1417
1418 /* A shared fault, but p->numa_group has not been set up yet. */
1419 if (!ng)
1420 return true;
1421
1422 /*
Rik van Riel4142c3e2016-01-25 17:07:39 -05001423 * Destination node is much more heavily used than the source
1424 * node? Allow migration.
Rik van Riel10f39042014-01-27 17:03:44 -05001425 */
Rik van Riel4142c3e2016-01-25 17:07:39 -05001426 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1427 ACTIVE_NODE_FRACTION)
Rik van Riel10f39042014-01-27 17:03:44 -05001428 return true;
1429
1430 /*
Rik van Riel4142c3e2016-01-25 17:07:39 -05001431 * Distribute memory according to CPU & memory use on each node,
1432 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1433 *
1434 * faults_cpu(dst) 3 faults_cpu(src)
1435 * --------------- * - > ---------------
1436 * faults_mem(dst) 4 faults_mem(src)
Rik van Riel10f39042014-01-27 17:03:44 -05001437 */
Rik van Riel4142c3e2016-01-25 17:07:39 -05001438 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1439 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
Rik van Riel10f39042014-01-27 17:03:44 -05001440}
1441
Viresh Kumarc7132dd2017-05-24 10:59:54 +05301442static unsigned long weighted_cpuload(struct rq *rq);
Mel Gorman58d081b2013-10-07 11:29:10 +01001443static unsigned long source_load(int cpu, int type);
1444static unsigned long target_load(int cpu, int type);
Nicolas Pitreced549f2014-05-26 18:19:38 -04001445static unsigned long capacity_of(int cpu);
Mel Gormane6628d52013-10-07 11:29:02 +01001446
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001447/* Cached statistics for all CPUs within a node */
Mel Gorman58d081b2013-10-07 11:29:10 +01001448struct numa_stats {
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001449 unsigned long nr_running;
Mel Gorman58d081b2013-10-07 11:29:10 +01001450 unsigned long load;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001451
1452 /* Total compute capacity of CPUs on a node */
Nicolas Pitre5ef20ca2014-05-26 18:19:34 -04001453 unsigned long compute_capacity;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001454
1455 /* Approximate capacity in terms of runnable tasks on a node */
Nicolas Pitre5ef20ca2014-05-26 18:19:34 -04001456 unsigned long task_capacity;
Nicolas Pitre1b6a7492014-05-26 18:19:35 -04001457 int has_free_capacity;
Mel Gorman58d081b2013-10-07 11:29:10 +01001458};
Mel Gormane6628d52013-10-07 11:29:02 +01001459
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001460/*
1461 * XXX borrowed from update_sg_lb_stats
1462 */
1463static void update_numa_stats(struct numa_stats *ns, int nid)
1464{
Rik van Riel83d7f242014-08-04 13:23:28 -04001465 int smt, cpu, cpus = 0;
1466 unsigned long capacity;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001467
1468 memset(ns, 0, sizeof(*ns));
1469 for_each_cpu(cpu, cpumask_of_node(nid)) {
1470 struct rq *rq = cpu_rq(cpu);
1471
1472 ns->nr_running += rq->nr_running;
Viresh Kumarc7132dd2017-05-24 10:59:54 +05301473 ns->load += weighted_cpuload(rq);
Nicolas Pitreced549f2014-05-26 18:19:38 -04001474 ns->compute_capacity += capacity_of(cpu);
Peter Zijlstra5eca82a2013-11-06 18:47:57 +01001475
1476 cpus++;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001477 }
1478
Peter Zijlstra5eca82a2013-11-06 18:47:57 +01001479 /*
1480 * If we raced with hotplug and there are no CPUs left in our mask
1481 * the @ns structure is NULL'ed and task_numa_compare() will
1482 * not find this node attractive.
1483 *
Nicolas Pitre1b6a7492014-05-26 18:19:35 -04001484 * We'll either bail at !has_free_capacity, or we'll detect a huge
1485 * imbalance and bail there.
Peter Zijlstra5eca82a2013-11-06 18:47:57 +01001486 */
1487 if (!cpus)
1488 return;
1489
Rik van Riel83d7f242014-08-04 13:23:28 -04001490 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1491 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1492 capacity = cpus / smt; /* cores */
1493
1494 ns->task_capacity = min_t(unsigned, capacity,
1495 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
Nicolas Pitre1b6a7492014-05-26 18:19:35 -04001496 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001497}
1498
Mel Gorman58d081b2013-10-07 11:29:10 +01001499struct task_numa_env {
1500 struct task_struct *p;
1501
1502 int src_cpu, src_nid;
1503 int dst_cpu, dst_nid;
1504
1505 struct numa_stats src_stats, dst_stats;
1506
Wanpeng Li40ea2b42013-12-05 19:10:17 +08001507 int imbalance_pct;
Rik van Riel7bd95322014-10-17 03:29:51 -04001508 int dist;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001509
1510 struct task_struct *best_task;
1511 long best_imp;
Mel Gorman58d081b2013-10-07 11:29:10 +01001512 int best_cpu;
1513};
1514
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001515static void task_numa_assign(struct task_numa_env *env,
1516 struct task_struct *p, long imp)
1517{
1518 if (env->best_task)
1519 put_task_struct(env->best_task);
Oleg Nesterovbac78572016-05-18 21:57:33 +02001520 if (p)
1521 get_task_struct(p);
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001522
1523 env->best_task = p;
1524 env->best_imp = imp;
1525 env->best_cpu = env->dst_cpu;
1526}
1527
Rik van Riel28a21742014-06-23 11:46:13 -04001528static bool load_too_imbalanced(long src_load, long dst_load,
Rik van Riele63da032014-05-14 13:22:21 -04001529 struct task_numa_env *env)
1530{
Rik van Riele4991b22015-05-27 15:04:27 -04001531 long imb, old_imb;
1532 long orig_src_load, orig_dst_load;
Rik van Riel28a21742014-06-23 11:46:13 -04001533 long src_capacity, dst_capacity;
1534
1535 /*
1536 * The load is corrected for the CPU capacity available on each node.
1537 *
1538 * src_load dst_load
1539 * ------------ vs ---------
1540 * src_capacity dst_capacity
1541 */
1542 src_capacity = env->src_stats.compute_capacity;
1543 dst_capacity = env->dst_stats.compute_capacity;
Rik van Riele63da032014-05-14 13:22:21 -04001544
1545 /* We care about the slope of the imbalance, not the direction. */
Rik van Riele4991b22015-05-27 15:04:27 -04001546 if (dst_load < src_load)
1547 swap(dst_load, src_load);
Rik van Riele63da032014-05-14 13:22:21 -04001548
1549 /* Is the difference below the threshold? */
Rik van Riele4991b22015-05-27 15:04:27 -04001550 imb = dst_load * src_capacity * 100 -
1551 src_load * dst_capacity * env->imbalance_pct;
Rik van Riele63da032014-05-14 13:22:21 -04001552 if (imb <= 0)
1553 return false;
1554
1555 /*
1556 * The imbalance is above the allowed threshold.
Rik van Riele4991b22015-05-27 15:04:27 -04001557 * Compare it with the old imbalance.
Rik van Riele63da032014-05-14 13:22:21 -04001558 */
Rik van Riel28a21742014-06-23 11:46:13 -04001559 orig_src_load = env->src_stats.load;
Rik van Riele4991b22015-05-27 15:04:27 -04001560 orig_dst_load = env->dst_stats.load;
Rik van Riel28a21742014-06-23 11:46:13 -04001561
Rik van Riele4991b22015-05-27 15:04:27 -04001562 if (orig_dst_load < orig_src_load)
1563 swap(orig_dst_load, orig_src_load);
Rik van Riele63da032014-05-14 13:22:21 -04001564
Rik van Riele4991b22015-05-27 15:04:27 -04001565 old_imb = orig_dst_load * src_capacity * 100 -
1566 orig_src_load * dst_capacity * env->imbalance_pct;
1567
1568 /* Would this change make things worse? */
1569 return (imb > old_imb);
Rik van Riele63da032014-05-14 13:22:21 -04001570}
1571
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001572/*
1573 * This checks if the overall compute and NUMA accesses of the system would
1574 * be improved if the source tasks was migrated to the target dst_cpu taking
1575 * into account that it might be best if task running on the dst_cpu should
1576 * be exchanged with the source task
1577 */
Rik van Riel887c2902013-10-07 11:29:31 +01001578static void task_numa_compare(struct task_numa_env *env,
1579 long taskimp, long groupimp)
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001580{
1581 struct rq *src_rq = cpu_rq(env->src_cpu);
1582 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1583 struct task_struct *cur;
Rik van Riel28a21742014-06-23 11:46:13 -04001584 long src_load, dst_load;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001585 long load;
Rik van Riel1c5d3eb2014-06-23 11:46:15 -04001586 long imp = env->p->numa_group ? groupimp : taskimp;
Rik van Riel0132c3e2014-06-23 11:46:16 -04001587 long moveimp = imp;
Rik van Riel7bd95322014-10-17 03:29:51 -04001588 int dist = env->dist;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001589
1590 rcu_read_lock();
Oleg Nesterovbac78572016-05-18 21:57:33 +02001591 cur = task_rcu_dereference(&dst_rq->curr);
1592 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001593 cur = NULL;
1594
1595 /*
Peter Zijlstra7af68332014-11-10 10:54:35 +01001596 * Because we have preemption enabled we can get migrated around and
1597 * end try selecting ourselves (current == env->p) as a swap candidate.
1598 */
1599 if (cur == env->p)
1600 goto unlock;
1601
1602 /*
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001603 * "imp" is the fault differential for the source task between the
1604 * source and destination node. Calculate the total differential for
1605 * the source task and potential destination task. The more negative
1606 * the value is, the more rmeote accesses that would be expected to
1607 * be incurred if the tasks were swapped.
1608 */
1609 if (cur) {
1610 /* Skip this swap candidate if cannot move to the source cpu */
Ingo Molnar0c98d342017-02-05 15:38:10 +01001611 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001612 goto unlock;
1613
Rik van Riel887c2902013-10-07 11:29:31 +01001614 /*
1615 * If dst and source tasks are in the same NUMA group, or not
Rik van Rielca28aa532013-10-07 11:29:32 +01001616 * in any group then look only at task weights.
Rik van Riel887c2902013-10-07 11:29:31 +01001617 */
Rik van Rielca28aa532013-10-07 11:29:32 +01001618 if (cur->numa_group == env->p->numa_group) {
Rik van Riel7bd95322014-10-17 03:29:51 -04001619 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1620 task_weight(cur, env->dst_nid, dist);
Rik van Rielca28aa532013-10-07 11:29:32 +01001621 /*
1622 * Add some hysteresis to prevent swapping the
1623 * tasks within a group over tiny differences.
1624 */
1625 if (cur->numa_group)
1626 imp -= imp/16;
Rik van Riel887c2902013-10-07 11:29:31 +01001627 } else {
Rik van Rielca28aa532013-10-07 11:29:32 +01001628 /*
1629 * Compare the group weights. If a task is all by
1630 * itself (not part of a group), use the task weight
1631 * instead.
1632 */
Rik van Rielca28aa532013-10-07 11:29:32 +01001633 if (cur->numa_group)
Rik van Riel7bd95322014-10-17 03:29:51 -04001634 imp += group_weight(cur, env->src_nid, dist) -
1635 group_weight(cur, env->dst_nid, dist);
Rik van Rielca28aa532013-10-07 11:29:32 +01001636 else
Rik van Riel7bd95322014-10-17 03:29:51 -04001637 imp += task_weight(cur, env->src_nid, dist) -
1638 task_weight(cur, env->dst_nid, dist);
Rik van Riel887c2902013-10-07 11:29:31 +01001639 }
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001640 }
1641
Rik van Riel0132c3e2014-06-23 11:46:16 -04001642 if (imp <= env->best_imp && moveimp <= env->best_imp)
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001643 goto unlock;
1644
1645 if (!cur) {
1646 /* Is there capacity at our destination? */
Rik van Rielb932c032014-08-04 13:23:27 -04001647 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
Nicolas Pitre1b6a7492014-05-26 18:19:35 -04001648 !env->dst_stats.has_free_capacity)
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001649 goto unlock;
1650
1651 goto balance;
1652 }
1653
1654 /* Balance doesn't matter much if we're running a task per cpu */
Rik van Riel0132c3e2014-06-23 11:46:16 -04001655 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1656 dst_rq->nr_running == 1)
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001657 goto assign;
1658
1659 /*
1660 * In the overloaded case, try and keep the load balanced.
1661 */
1662balance:
Peter Zijlstrae720fff2014-07-11 16:01:53 +02001663 load = task_h_load(env->p);
1664 dst_load = env->dst_stats.load + load;
1665 src_load = env->src_stats.load - load;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001666
Rik van Riel0132c3e2014-06-23 11:46:16 -04001667 if (moveimp > imp && moveimp > env->best_imp) {
1668 /*
1669 * If the improvement from just moving env->p direction is
1670 * better than swapping tasks around, check if a move is
1671 * possible. Store a slightly smaller score than moveimp,
1672 * so an actually idle CPU will win.
1673 */
1674 if (!load_too_imbalanced(src_load, dst_load, env)) {
1675 imp = moveimp - 1;
1676 cur = NULL;
1677 goto assign;
1678 }
1679 }
1680
1681 if (imp <= env->best_imp)
1682 goto unlock;
1683
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001684 if (cur) {
Peter Zijlstrae720fff2014-07-11 16:01:53 +02001685 load = task_h_load(cur);
1686 dst_load -= load;
1687 src_load += load;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001688 }
1689
Rik van Riel28a21742014-06-23 11:46:13 -04001690 if (load_too_imbalanced(src_load, dst_load, env))
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001691 goto unlock;
1692
Rik van Rielba7e5a22014-09-04 16:35:30 -04001693 /*
1694 * One idle CPU per node is evaluated for a task numa move.
1695 * Call select_idle_sibling to maybe find a better one.
1696 */
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02001697 if (!cur) {
1698 /*
1699 * select_idle_siblings() uses an per-cpu cpumask that
1700 * can be used from IRQ context.
1701 */
1702 local_irq_disable();
Morten Rasmussen772bd008c2016-06-22 18:03:13 +01001703 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1704 env->dst_cpu);
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02001705 local_irq_enable();
1706 }
Rik van Rielba7e5a22014-09-04 16:35:30 -04001707
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001708assign:
1709 task_numa_assign(env, cur, imp);
1710unlock:
1711 rcu_read_unlock();
1712}
1713
Rik van Riel887c2902013-10-07 11:29:31 +01001714static void task_numa_find_cpu(struct task_numa_env *env,
1715 long taskimp, long groupimp)
Mel Gorman2c8a50a2013-10-07 11:29:18 +01001716{
1717 int cpu;
1718
1719 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1720 /* Skip this CPU if the source task cannot migrate */
Ingo Molnar0c98d342017-02-05 15:38:10 +01001721 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
Mel Gorman2c8a50a2013-10-07 11:29:18 +01001722 continue;
1723
1724 env->dst_cpu = cpu;
Rik van Riel887c2902013-10-07 11:29:31 +01001725 task_numa_compare(env, taskimp, groupimp);
Mel Gorman2c8a50a2013-10-07 11:29:18 +01001726 }
1727}
1728
Rik van Riel6f9aad02015-05-28 09:52:49 -04001729/* Only move tasks to a NUMA node less busy than the current node. */
1730static bool numa_has_capacity(struct task_numa_env *env)
1731{
1732 struct numa_stats *src = &env->src_stats;
1733 struct numa_stats *dst = &env->dst_stats;
1734
1735 if (src->has_free_capacity && !dst->has_free_capacity)
1736 return false;
1737
1738 /*
1739 * Only consider a task move if the source has a higher load
1740 * than the destination, corrected for CPU capacity on each node.
1741 *
1742 * src->load dst->load
1743 * --------------------- vs ---------------------
1744 * src->compute_capacity dst->compute_capacity
1745 */
Srikar Dronamraju44dcb042015-06-16 17:26:00 +05301746 if (src->load * dst->compute_capacity * env->imbalance_pct >
1747
1748 dst->load * src->compute_capacity * 100)
Rik van Riel6f9aad02015-05-28 09:52:49 -04001749 return true;
1750
1751 return false;
1752}
1753
Mel Gorman58d081b2013-10-07 11:29:10 +01001754static int task_numa_migrate(struct task_struct *p)
Mel Gormane6628d52013-10-07 11:29:02 +01001755{
Mel Gorman58d081b2013-10-07 11:29:10 +01001756 struct task_numa_env env = {
1757 .p = p,
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001758
Mel Gorman58d081b2013-10-07 11:29:10 +01001759 .src_cpu = task_cpu(p),
Ingo Molnarb32e86b2013-10-07 11:29:30 +01001760 .src_nid = task_node(p),
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001761
1762 .imbalance_pct = 112,
1763
1764 .best_task = NULL,
1765 .best_imp = 0,
Rik van Riel4142c3e2016-01-25 17:07:39 -05001766 .best_cpu = -1,
Mel Gorman58d081b2013-10-07 11:29:10 +01001767 };
1768 struct sched_domain *sd;
Rik van Riel887c2902013-10-07 11:29:31 +01001769 unsigned long taskweight, groupweight;
Rik van Riel7bd95322014-10-17 03:29:51 -04001770 int nid, ret, dist;
Rik van Riel887c2902013-10-07 11:29:31 +01001771 long taskimp, groupimp;
Mel Gormane6628d52013-10-07 11:29:02 +01001772
Mel Gorman58d081b2013-10-07 11:29:10 +01001773 /*
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001774 * Pick the lowest SD_NUMA domain, as that would have the smallest
1775 * imbalance and would be the first to start moving tasks about.
1776 *
1777 * And we want to avoid any moving of tasks about, as that would create
1778 * random movement of tasks -- counter the numa conditions we're trying
1779 * to satisfy here.
Mel Gorman58d081b2013-10-07 11:29:10 +01001780 */
Mel Gormane6628d52013-10-07 11:29:02 +01001781 rcu_read_lock();
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001782 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
Rik van Riel46a73e82013-11-11 19:29:25 -05001783 if (sd)
1784 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
Mel Gormane6628d52013-10-07 11:29:02 +01001785 rcu_read_unlock();
1786
Rik van Riel46a73e82013-11-11 19:29:25 -05001787 /*
1788 * Cpusets can break the scheduler domain tree into smaller
1789 * balance domains, some of which do not cross NUMA boundaries.
1790 * Tasks that are "trapped" in such domains cannot be migrated
1791 * elsewhere, so there is no point in (re)trying.
1792 */
1793 if (unlikely(!sd)) {
Wanpeng Lide1b3012013-12-12 15:23:24 +08001794 p->numa_preferred_nid = task_node(p);
Rik van Riel46a73e82013-11-11 19:29:25 -05001795 return -EINVAL;
1796 }
1797
Mel Gorman2c8a50a2013-10-07 11:29:18 +01001798 env.dst_nid = p->numa_preferred_nid;
Rik van Riel7bd95322014-10-17 03:29:51 -04001799 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1800 taskweight = task_weight(p, env.src_nid, dist);
1801 groupweight = group_weight(p, env.src_nid, dist);
1802 update_numa_stats(&env.src_stats, env.src_nid);
1803 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1804 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
Mel Gorman2c8a50a2013-10-07 11:29:18 +01001805 update_numa_stats(&env.dst_stats, env.dst_nid);
Mel Gorman58d081b2013-10-07 11:29:10 +01001806
Rik van Riela43455a2014-06-04 16:09:42 -04001807 /* Try to find a spot on the preferred nid. */
Rik van Riel6f9aad02015-05-28 09:52:49 -04001808 if (numa_has_capacity(&env))
1809 task_numa_find_cpu(&env, taskimp, groupimp);
Rik van Riele1dda8a2013-10-07 11:29:19 +01001810
Rik van Riel9de05d42014-10-09 17:27:47 -04001811 /*
1812 * Look at other nodes in these cases:
1813 * - there is no space available on the preferred_nid
1814 * - the task is part of a numa_group that is interleaved across
1815 * multiple NUMA nodes; in order to better consolidate the group,
1816 * we need to check other locations.
1817 */
Rik van Riel4142c3e2016-01-25 17:07:39 -05001818 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
Mel Gorman2c8a50a2013-10-07 11:29:18 +01001819 for_each_online_node(nid) {
1820 if (nid == env.src_nid || nid == p->numa_preferred_nid)
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001821 continue;
1822
Rik van Riel7bd95322014-10-17 03:29:51 -04001823 dist = node_distance(env.src_nid, env.dst_nid);
Rik van Riel6c6b1192014-10-17 03:29:52 -04001824 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1825 dist != env.dist) {
1826 taskweight = task_weight(p, env.src_nid, dist);
1827 groupweight = group_weight(p, env.src_nid, dist);
1828 }
Rik van Riel7bd95322014-10-17 03:29:51 -04001829
Mel Gorman83e1d2c2013-10-07 11:29:27 +01001830 /* Only consider nodes where both task and groups benefit */
Rik van Riel7bd95322014-10-17 03:29:51 -04001831 taskimp = task_weight(p, nid, dist) - taskweight;
1832 groupimp = group_weight(p, nid, dist) - groupweight;
Rik van Riel887c2902013-10-07 11:29:31 +01001833 if (taskimp < 0 && groupimp < 0)
Mel Gorman2c8a50a2013-10-07 11:29:18 +01001834 continue;
1835
Rik van Riel7bd95322014-10-17 03:29:51 -04001836 env.dist = dist;
Mel Gorman2c8a50a2013-10-07 11:29:18 +01001837 env.dst_nid = nid;
1838 update_numa_stats(&env.dst_stats, env.dst_nid);
Rik van Riel6f9aad02015-05-28 09:52:49 -04001839 if (numa_has_capacity(&env))
1840 task_numa_find_cpu(&env, taskimp, groupimp);
Mel Gorman58d081b2013-10-07 11:29:10 +01001841 }
1842 }
1843
Rik van Riel68d1b022014-04-11 13:00:29 -04001844 /*
1845 * If the task is part of a workload that spans multiple NUMA nodes,
1846 * and is migrating into one of the workload's active nodes, remember
1847 * this node as the task's preferred numa node, so the workload can
1848 * settle down.
1849 * A task that migrated to a second choice node will be better off
1850 * trying for a better one later. Do not set the preferred node here.
1851 */
Rik van Rieldb015da2014-06-23 11:41:34 -04001852 if (p->numa_group) {
Rik van Riel4142c3e2016-01-25 17:07:39 -05001853 struct numa_group *ng = p->numa_group;
1854
Rik van Rieldb015da2014-06-23 11:41:34 -04001855 if (env.best_cpu == -1)
1856 nid = env.src_nid;
1857 else
1858 nid = env.dst_nid;
1859
Rik van Riel4142c3e2016-01-25 17:07:39 -05001860 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
Rik van Rieldb015da2014-06-23 11:41:34 -04001861 sched_setnuma(p, env.dst_nid);
1862 }
1863
1864 /* No better CPU than the current one was found. */
1865 if (env.best_cpu == -1)
1866 return -EAGAIN;
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01001867
Rik van Riel04bb2f92013-10-07 11:29:36 +01001868 /*
1869 * Reset the scan period if the task is being rescheduled on an
1870 * alternative node to recheck if the tasks is now properly placed.
1871 */
Rik van Rielb5dd77c2017-07-31 15:28:47 -04001872 p->numa_scan_period = task_scan_start(p);
Rik van Riel04bb2f92013-10-07 11:29:36 +01001873
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001874 if (env.best_task == NULL) {
Mel Gorman286549d2014-01-21 15:51:03 -08001875 ret = migrate_task_to(p, env.best_cpu);
1876 if (ret != 0)
1877 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001878 return ret;
1879 }
1880
1881 ret = migrate_swap(p, env.best_task);
Mel Gorman286549d2014-01-21 15:51:03 -08001882 if (ret != 0)
1883 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
Mel Gormanfb13c7e2013-10-07 11:29:17 +01001884 put_task_struct(env.best_task);
1885 return ret;
Mel Gormane6628d52013-10-07 11:29:02 +01001886}
1887
Mel Gorman6b9a7462013-10-07 11:29:11 +01001888/* Attempt to migrate a task to a CPU on the preferred node. */
1889static void numa_migrate_preferred(struct task_struct *p)
1890{
Rik van Riel5085e2a2014-04-11 13:00:28 -04001891 unsigned long interval = HZ;
1892
Rik van Riel2739d3e2013-10-07 11:29:41 +01001893 /* This task has no NUMA fault statistics yet */
Iulia Manda44dba3d2014-10-31 02:13:31 +02001894 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
Rik van Riel2739d3e2013-10-07 11:29:41 +01001895 return;
1896
1897 /* Periodically retry migrating the task to the preferred node */
Rik van Riel5085e2a2014-04-11 13:00:28 -04001898 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1899 p->numa_migrate_retry = jiffies + interval;
Rik van Riel2739d3e2013-10-07 11:29:41 +01001900
Mel Gorman6b9a7462013-10-07 11:29:11 +01001901 /* Success if task is already running on preferred CPU */
Wanpeng Lide1b3012013-12-12 15:23:24 +08001902 if (task_node(p) == p->numa_preferred_nid)
Mel Gorman6b9a7462013-10-07 11:29:11 +01001903 return;
1904
Mel Gorman6b9a7462013-10-07 11:29:11 +01001905 /* Otherwise, try migrate to a CPU on the preferred node */
Rik van Riel2739d3e2013-10-07 11:29:41 +01001906 task_numa_migrate(p);
Mel Gorman6b9a7462013-10-07 11:29:11 +01001907}
1908
Rik van Riel04bb2f92013-10-07 11:29:36 +01001909/*
Rik van Riel4142c3e2016-01-25 17:07:39 -05001910 * Find out how many nodes on the workload is actively running on. Do this by
Rik van Riel20e07de2014-01-27 17:03:43 -05001911 * tracking the nodes from which NUMA hinting faults are triggered. This can
1912 * be different from the set of nodes where the workload's memory is currently
1913 * located.
Rik van Riel20e07de2014-01-27 17:03:43 -05001914 */
Rik van Riel4142c3e2016-01-25 17:07:39 -05001915static void numa_group_count_active_nodes(struct numa_group *numa_group)
Rik van Riel20e07de2014-01-27 17:03:43 -05001916{
1917 unsigned long faults, max_faults = 0;
Rik van Riel4142c3e2016-01-25 17:07:39 -05001918 int nid, active_nodes = 0;
Rik van Riel20e07de2014-01-27 17:03:43 -05001919
1920 for_each_online_node(nid) {
1921 faults = group_faults_cpu(numa_group, nid);
1922 if (faults > max_faults)
1923 max_faults = faults;
1924 }
1925
1926 for_each_online_node(nid) {
1927 faults = group_faults_cpu(numa_group, nid);
Rik van Riel4142c3e2016-01-25 17:07:39 -05001928 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1929 active_nodes++;
Rik van Riel20e07de2014-01-27 17:03:43 -05001930 }
Rik van Riel4142c3e2016-01-25 17:07:39 -05001931
1932 numa_group->max_faults_cpu = max_faults;
1933 numa_group->active_nodes = active_nodes;
Rik van Riel20e07de2014-01-27 17:03:43 -05001934}
1935
1936/*
Rik van Riel04bb2f92013-10-07 11:29:36 +01001937 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1938 * increments. The more local the fault statistics are, the higher the scan
Rik van Riela22b4b02014-06-23 11:41:35 -04001939 * period will be for the next scan window. If local/(local+remote) ratio is
1940 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1941 * the scan period will decrease. Aim for 70% local accesses.
Rik van Riel04bb2f92013-10-07 11:29:36 +01001942 */
1943#define NUMA_PERIOD_SLOTS 10
Rik van Riela22b4b02014-06-23 11:41:35 -04001944#define NUMA_PERIOD_THRESHOLD 7
Rik van Riel04bb2f92013-10-07 11:29:36 +01001945
1946/*
1947 * Increase the scan period (slow down scanning) if the majority of
1948 * our memory is already on our local node, or if the majority of
1949 * the page accesses are shared with other processes.
1950 * Otherwise, decrease the scan period.
1951 */
1952static void update_task_scan_period(struct task_struct *p,
1953 unsigned long shared, unsigned long private)
1954{
1955 unsigned int period_slot;
Rik van Riel37ec97de2017-07-31 15:28:46 -04001956 int lr_ratio, ps_ratio;
Rik van Riel04bb2f92013-10-07 11:29:36 +01001957 int diff;
1958
1959 unsigned long remote = p->numa_faults_locality[0];
1960 unsigned long local = p->numa_faults_locality[1];
1961
1962 /*
1963 * If there were no record hinting faults then either the task is
1964 * completely idle or all activity is areas that are not of interest
Mel Gorman074c2382015-03-25 15:55:42 -07001965 * to automatic numa balancing. Related to that, if there were failed
1966 * migration then it implies we are migrating too quickly or the local
1967 * node is overloaded. In either case, scan slower
Rik van Riel04bb2f92013-10-07 11:29:36 +01001968 */
Mel Gorman074c2382015-03-25 15:55:42 -07001969 if (local + shared == 0 || p->numa_faults_locality[2]) {
Rik van Riel04bb2f92013-10-07 11:29:36 +01001970 p->numa_scan_period = min(p->numa_scan_period_max,
1971 p->numa_scan_period << 1);
1972
1973 p->mm->numa_next_scan = jiffies +
1974 msecs_to_jiffies(p->numa_scan_period);
1975
1976 return;
1977 }
1978
1979 /*
1980 * Prepare to scale scan period relative to the current period.
1981 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1982 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1983 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1984 */
1985 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
Rik van Riel37ec97de2017-07-31 15:28:46 -04001986 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1987 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1988
1989 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1990 /*
1991 * Most memory accesses are local. There is no need to
1992 * do fast NUMA scanning, since memory is already local.
1993 */
1994 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1995 if (!slot)
1996 slot = 1;
1997 diff = slot * period_slot;
1998 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1999 /*
2000 * Most memory accesses are shared with other tasks.
2001 * There is no point in continuing fast NUMA scanning,
2002 * since other tasks may just move the memory elsewhere.
2003 */
2004 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
Rik van Riel04bb2f92013-10-07 11:29:36 +01002005 if (!slot)
2006 slot = 1;
2007 diff = slot * period_slot;
2008 } else {
Rik van Riel04bb2f92013-10-07 11:29:36 +01002009 /*
Rik van Riel37ec97de2017-07-31 15:28:46 -04002010 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2011 * yet they are not on the local NUMA node. Speed up
2012 * NUMA scanning to get the memory moved over.
Rik van Riel04bb2f92013-10-07 11:29:36 +01002013 */
Rik van Riel37ec97de2017-07-31 15:28:46 -04002014 int ratio = max(lr_ratio, ps_ratio);
2015 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
Rik van Riel04bb2f92013-10-07 11:29:36 +01002016 }
2017
2018 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2019 task_scan_min(p), task_scan_max(p));
2020 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2021}
2022
Rik van Riel7e2703e2014-01-27 17:03:45 -05002023/*
2024 * Get the fraction of time the task has been running since the last
2025 * NUMA placement cycle. The scheduler keeps similar statistics, but
2026 * decays those on a 32ms period, which is orders of magnitude off
2027 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2028 * stats only if the task is so new there are no NUMA statistics yet.
2029 */
2030static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2031{
2032 u64 runtime, delta, now;
2033 /* Use the start of this time slice to avoid calculations. */
2034 now = p->se.exec_start;
2035 runtime = p->se.sum_exec_runtime;
2036
2037 if (p->last_task_numa_placement) {
2038 delta = runtime - p->last_sum_exec_runtime;
2039 *period = now - p->last_task_numa_placement;
2040 } else {
Yuyang Du9d89c252015-07-15 08:04:37 +08002041 delta = p->se.avg.load_sum / p->se.load.weight;
2042 *period = LOAD_AVG_MAX;
Rik van Riel7e2703e2014-01-27 17:03:45 -05002043 }
2044
2045 p->last_sum_exec_runtime = runtime;
2046 p->last_task_numa_placement = now;
2047
2048 return delta;
2049}
2050
Rik van Riel54009412014-10-17 03:29:53 -04002051/*
2052 * Determine the preferred nid for a task in a numa_group. This needs to
2053 * be done in a way that produces consistent results with group_weight,
2054 * otherwise workloads might not converge.
2055 */
2056static int preferred_group_nid(struct task_struct *p, int nid)
2057{
2058 nodemask_t nodes;
2059 int dist;
2060
2061 /* Direct connections between all NUMA nodes. */
2062 if (sched_numa_topology_type == NUMA_DIRECT)
2063 return nid;
2064
2065 /*
2066 * On a system with glueless mesh NUMA topology, group_weight
2067 * scores nodes according to the number of NUMA hinting faults on
2068 * both the node itself, and on nearby nodes.
2069 */
2070 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2071 unsigned long score, max_score = 0;
2072 int node, max_node = nid;
2073
2074 dist = sched_max_numa_distance;
2075
2076 for_each_online_node(node) {
2077 score = group_weight(p, node, dist);
2078 if (score > max_score) {
2079 max_score = score;
2080 max_node = node;
2081 }
2082 }
2083 return max_node;
2084 }
2085
2086 /*
2087 * Finding the preferred nid in a system with NUMA backplane
2088 * interconnect topology is more involved. The goal is to locate
2089 * tasks from numa_groups near each other in the system, and
2090 * untangle workloads from different sides of the system. This requires
2091 * searching down the hierarchy of node groups, recursively searching
2092 * inside the highest scoring group of nodes. The nodemask tricks
2093 * keep the complexity of the search down.
2094 */
2095 nodes = node_online_map;
2096 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2097 unsigned long max_faults = 0;
Jan Beulich81907472015-01-23 08:25:38 +00002098 nodemask_t max_group = NODE_MASK_NONE;
Rik van Riel54009412014-10-17 03:29:53 -04002099 int a, b;
2100
2101 /* Are there nodes at this distance from each other? */
2102 if (!find_numa_distance(dist))
2103 continue;
2104
2105 for_each_node_mask(a, nodes) {
2106 unsigned long faults = 0;
2107 nodemask_t this_group;
2108 nodes_clear(this_group);
2109
2110 /* Sum group's NUMA faults; includes a==b case. */
2111 for_each_node_mask(b, nodes) {
2112 if (node_distance(a, b) < dist) {
2113 faults += group_faults(p, b);
2114 node_set(b, this_group);
2115 node_clear(b, nodes);
2116 }
2117 }
2118
2119 /* Remember the top group. */
2120 if (faults > max_faults) {
2121 max_faults = faults;
2122 max_group = this_group;
2123 /*
2124 * subtle: at the smallest distance there is
2125 * just one node left in each "group", the
2126 * winner is the preferred nid.
2127 */
2128 nid = a;
2129 }
2130 }
2131 /* Next round, evaluate the nodes within max_group. */
Jan Beulich890a5402015-02-09 12:30:00 +01002132 if (!max_faults)
2133 break;
Rik van Riel54009412014-10-17 03:29:53 -04002134 nodes = max_group;
2135 }
2136 return nid;
2137}
2138
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002139static void task_numa_placement(struct task_struct *p)
2140{
Mel Gorman83e1d2c2013-10-07 11:29:27 +01002141 int seq, nid, max_nid = -1, max_group_nid = -1;
2142 unsigned long max_faults = 0, max_group_faults = 0;
Rik van Riel04bb2f92013-10-07 11:29:36 +01002143 unsigned long fault_types[2] = { 0, 0 };
Rik van Riel7e2703e2014-01-27 17:03:45 -05002144 unsigned long total_faults;
2145 u64 runtime, period;
Mel Gorman7dbd13e2013-10-07 11:29:29 +01002146 spinlock_t *group_lock = NULL;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002147
Jason Low7e5a2c12015-04-30 17:28:14 -07002148 /*
2149 * The p->mm->numa_scan_seq field gets updated without
2150 * exclusive access. Use READ_ONCE() here to ensure
2151 * that the field is read in a single access:
2152 */
Jason Low316c1608d2015-04-28 13:00:20 -07002153 seq = READ_ONCE(p->mm->numa_scan_seq);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002154 if (p->numa_scan_seq == seq)
2155 return;
2156 p->numa_scan_seq = seq;
Mel Gorman598f0ec2013-10-07 11:28:55 +01002157 p->numa_scan_period_max = task_scan_max(p);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002158
Rik van Riel7e2703e2014-01-27 17:03:45 -05002159 total_faults = p->numa_faults_locality[0] +
2160 p->numa_faults_locality[1];
2161 runtime = numa_get_avg_runtime(p, &period);
2162
Mel Gorman7dbd13e2013-10-07 11:29:29 +01002163 /* If the task is part of a group prevent parallel updates to group stats */
2164 if (p->numa_group) {
2165 group_lock = &p->numa_group->lock;
Mike Galbraith60e69ee2014-04-07 10:55:15 +02002166 spin_lock_irq(group_lock);
Mel Gorman7dbd13e2013-10-07 11:29:29 +01002167 }
2168
Mel Gorman688b7582013-10-07 11:28:58 +01002169 /* Find the node with the highest number of faults */
2170 for_each_online_node(nid) {
Iulia Manda44dba3d2014-10-31 02:13:31 +02002171 /* Keep track of the offsets in numa_faults array */
2172 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
Mel Gorman83e1d2c2013-10-07 11:29:27 +01002173 unsigned long faults = 0, group_faults = 0;
Iulia Manda44dba3d2014-10-31 02:13:31 +02002174 int priv;
Mel Gorman745d6142013-10-07 11:28:59 +01002175
Rik van Rielbe1e4e72014-01-27 17:03:48 -05002176 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
Rik van Riel7e2703e2014-01-27 17:03:45 -05002177 long diff, f_diff, f_weight;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002178
Iulia Manda44dba3d2014-10-31 02:13:31 +02002179 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2180 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2181 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2182 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
Mel Gorman745d6142013-10-07 11:28:59 +01002183
Mel Gormanac8e8952013-10-07 11:29:03 +01002184 /* Decay existing window, copy faults since last scan */
Iulia Manda44dba3d2014-10-31 02:13:31 +02002185 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2186 fault_types[priv] += p->numa_faults[membuf_idx];
2187 p->numa_faults[membuf_idx] = 0;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01002188
Rik van Riel7e2703e2014-01-27 17:03:45 -05002189 /*
2190 * Normalize the faults_from, so all tasks in a group
2191 * count according to CPU use, instead of by the raw
2192 * number of faults. Tasks with little runtime have
2193 * little over-all impact on throughput, and thus their
2194 * faults are less important.
2195 */
2196 f_weight = div64_u64(runtime << 16, period + 1);
Iulia Manda44dba3d2014-10-31 02:13:31 +02002197 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
Rik van Riel7e2703e2014-01-27 17:03:45 -05002198 (total_faults + 1);
Iulia Manda44dba3d2014-10-31 02:13:31 +02002199 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2200 p->numa_faults[cpubuf_idx] = 0;
Rik van Riel50ec8a42014-01-27 17:03:42 -05002201
Iulia Manda44dba3d2014-10-31 02:13:31 +02002202 p->numa_faults[mem_idx] += diff;
2203 p->numa_faults[cpu_idx] += f_diff;
2204 faults += p->numa_faults[mem_idx];
Mel Gorman83e1d2c2013-10-07 11:29:27 +01002205 p->total_numa_faults += diff;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002206 if (p->numa_group) {
Iulia Manda44dba3d2014-10-31 02:13:31 +02002207 /*
2208 * safe because we can only change our own group
2209 *
2210 * mem_idx represents the offset for a given
2211 * nid and priv in a specific region because it
2212 * is at the beginning of the numa_faults array.
2213 */
2214 p->numa_group->faults[mem_idx] += diff;
2215 p->numa_group->faults_cpu[mem_idx] += f_diff;
Mel Gorman989348b2013-10-07 11:29:40 +01002216 p->numa_group->total_faults += diff;
Iulia Manda44dba3d2014-10-31 02:13:31 +02002217 group_faults += p->numa_group->faults[mem_idx];
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002218 }
Mel Gormanac8e8952013-10-07 11:29:03 +01002219 }
2220
Mel Gorman688b7582013-10-07 11:28:58 +01002221 if (faults > max_faults) {
2222 max_faults = faults;
2223 max_nid = nid;
2224 }
Mel Gorman83e1d2c2013-10-07 11:29:27 +01002225
2226 if (group_faults > max_group_faults) {
2227 max_group_faults = group_faults;
2228 max_group_nid = nid;
2229 }
2230 }
2231
Rik van Riel04bb2f92013-10-07 11:29:36 +01002232 update_task_scan_period(p, fault_types[0], fault_types[1]);
2233
Mel Gorman7dbd13e2013-10-07 11:29:29 +01002234 if (p->numa_group) {
Rik van Riel4142c3e2016-01-25 17:07:39 -05002235 numa_group_count_active_nodes(p->numa_group);
Mike Galbraith60e69ee2014-04-07 10:55:15 +02002236 spin_unlock_irq(group_lock);
Rik van Riel54009412014-10-17 03:29:53 -04002237 max_nid = preferred_group_nid(p, max_group_nid);
Mel Gorman688b7582013-10-07 11:28:58 +01002238 }
2239
Rik van Rielbb97fc32014-06-04 16:33:15 -04002240 if (max_faults) {
2241 /* Set the new preferred node */
2242 if (max_nid != p->numa_preferred_nid)
2243 sched_setnuma(p, max_nid);
2244
2245 if (task_node(p) != p->numa_preferred_nid)
2246 numa_migrate_preferred(p);
Mel Gorman3a7053b2013-10-07 11:29:00 +01002247 }
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002248}
2249
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002250static inline int get_numa_group(struct numa_group *grp)
2251{
2252 return atomic_inc_not_zero(&grp->refcount);
2253}
2254
2255static inline void put_numa_group(struct numa_group *grp)
2256{
2257 if (atomic_dec_and_test(&grp->refcount))
2258 kfree_rcu(grp, rcu);
2259}
2260
Mel Gorman3e6a9412013-10-07 11:29:35 +01002261static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2262 int *priv)
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002263{
2264 struct numa_group *grp, *my_grp;
2265 struct task_struct *tsk;
2266 bool join = false;
2267 int cpu = cpupid_to_cpu(cpupid);
2268 int i;
2269
2270 if (unlikely(!p->numa_group)) {
2271 unsigned int size = sizeof(struct numa_group) +
Rik van Riel50ec8a42014-01-27 17:03:42 -05002272 4*nr_node_ids*sizeof(unsigned long);
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002273
2274 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2275 if (!grp)
2276 return;
2277
2278 atomic_set(&grp->refcount, 1);
Rik van Riel4142c3e2016-01-25 17:07:39 -05002279 grp->active_nodes = 1;
2280 grp->max_faults_cpu = 0;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002281 spin_lock_init(&grp->lock);
Mel Gormane29cf082013-10-07 11:29:22 +01002282 grp->gid = p->pid;
Rik van Riel50ec8a42014-01-27 17:03:42 -05002283 /* Second half of the array tracks nids where faults happen */
Rik van Rielbe1e4e72014-01-27 17:03:48 -05002284 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2285 nr_node_ids;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002286
Rik van Rielbe1e4e72014-01-27 17:03:48 -05002287 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
Iulia Manda44dba3d2014-10-31 02:13:31 +02002288 grp->faults[i] = p->numa_faults[i];
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002289
Mel Gorman989348b2013-10-07 11:29:40 +01002290 grp->total_faults = p->total_numa_faults;
Mel Gorman83e1d2c2013-10-07 11:29:27 +01002291
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002292 grp->nr_tasks++;
2293 rcu_assign_pointer(p->numa_group, grp);
2294 }
2295
2296 rcu_read_lock();
Jason Low316c1608d2015-04-28 13:00:20 -07002297 tsk = READ_ONCE(cpu_rq(cpu)->curr);
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002298
2299 if (!cpupid_match_pid(tsk, cpupid))
Peter Zijlstra33547812013-10-09 10:24:48 +02002300 goto no_join;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002301
2302 grp = rcu_dereference(tsk->numa_group);
2303 if (!grp)
Peter Zijlstra33547812013-10-09 10:24:48 +02002304 goto no_join;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002305
2306 my_grp = p->numa_group;
2307 if (grp == my_grp)
Peter Zijlstra33547812013-10-09 10:24:48 +02002308 goto no_join;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002309
2310 /*
2311 * Only join the other group if its bigger; if we're the bigger group,
2312 * the other task will join us.
2313 */
2314 if (my_grp->nr_tasks > grp->nr_tasks)
Peter Zijlstra33547812013-10-09 10:24:48 +02002315 goto no_join;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002316
2317 /*
2318 * Tie-break on the grp address.
2319 */
2320 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
Peter Zijlstra33547812013-10-09 10:24:48 +02002321 goto no_join;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002322
Rik van Rieldabe1d92013-10-07 11:29:34 +01002323 /* Always join threads in the same process. */
2324 if (tsk->mm == current->mm)
2325 join = true;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002326
Rik van Rieldabe1d92013-10-07 11:29:34 +01002327 /* Simple filter to avoid false positives due to PID collisions */
2328 if (flags & TNF_SHARED)
2329 join = true;
2330
Mel Gorman3e6a9412013-10-07 11:29:35 +01002331 /* Update priv based on whether false sharing was detected */
2332 *priv = !join;
2333
Rik van Rieldabe1d92013-10-07 11:29:34 +01002334 if (join && !get_numa_group(grp))
Peter Zijlstra33547812013-10-09 10:24:48 +02002335 goto no_join;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002336
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002337 rcu_read_unlock();
2338
2339 if (!join)
2340 return;
2341
Mike Galbraith60e69ee2014-04-07 10:55:15 +02002342 BUG_ON(irqs_disabled());
2343 double_lock_irq(&my_grp->lock, &grp->lock);
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002344
Rik van Rielbe1e4e72014-01-27 17:03:48 -05002345 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
Iulia Manda44dba3d2014-10-31 02:13:31 +02002346 my_grp->faults[i] -= p->numa_faults[i];
2347 grp->faults[i] += p->numa_faults[i];
Mel Gorman989348b2013-10-07 11:29:40 +01002348 }
2349 my_grp->total_faults -= p->total_numa_faults;
2350 grp->total_faults += p->total_numa_faults;
2351
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002352 my_grp->nr_tasks--;
2353 grp->nr_tasks++;
2354
2355 spin_unlock(&my_grp->lock);
Mike Galbraith60e69ee2014-04-07 10:55:15 +02002356 spin_unlock_irq(&grp->lock);
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002357
2358 rcu_assign_pointer(p->numa_group, grp);
2359
2360 put_numa_group(my_grp);
Peter Zijlstra33547812013-10-09 10:24:48 +02002361 return;
2362
2363no_join:
2364 rcu_read_unlock();
2365 return;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002366}
2367
2368void task_numa_free(struct task_struct *p)
2369{
2370 struct numa_group *grp = p->numa_group;
Iulia Manda44dba3d2014-10-31 02:13:31 +02002371 void *numa_faults = p->numa_faults;
Steven Rostedte9dd6852014-05-27 17:02:04 -04002372 unsigned long flags;
2373 int i;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002374
2375 if (grp) {
Steven Rostedte9dd6852014-05-27 17:02:04 -04002376 spin_lock_irqsave(&grp->lock, flags);
Rik van Rielbe1e4e72014-01-27 17:03:48 -05002377 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
Iulia Manda44dba3d2014-10-31 02:13:31 +02002378 grp->faults[i] -= p->numa_faults[i];
Mel Gorman989348b2013-10-07 11:29:40 +01002379 grp->total_faults -= p->total_numa_faults;
2380
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002381 grp->nr_tasks--;
Steven Rostedte9dd6852014-05-27 17:02:04 -04002382 spin_unlock_irqrestore(&grp->lock, flags);
Andreea-Cristina Bernat35b123e2014-08-22 17:50:43 +03002383 RCU_INIT_POINTER(p->numa_group, NULL);
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002384 put_numa_group(grp);
2385 }
2386
Iulia Manda44dba3d2014-10-31 02:13:31 +02002387 p->numa_faults = NULL;
Rik van Riel82727012013-10-07 11:29:28 +01002388 kfree(numa_faults);
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002389}
2390
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002391/*
2392 * Got a PROT_NONE fault for a page on @node.
2393 */
Rik van Riel58b46da2014-01-27 17:03:47 -05002394void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002395{
2396 struct task_struct *p = current;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01002397 bool migrated = flags & TNF_MIGRATED;
Rik van Riel58b46da2014-01-27 17:03:47 -05002398 int cpu_node = task_node(current);
Rik van Riel792568e2014-04-11 13:00:27 -04002399 int local = !!(flags & TNF_FAULT_LOCAL);
Rik van Riel4142c3e2016-01-25 17:07:39 -05002400 struct numa_group *ng;
Mel Gormanac8e8952013-10-07 11:29:03 +01002401 int priv;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002402
Srikar Dronamraju2a595722015-08-11 21:54:21 +05302403 if (!static_branch_likely(&sched_numa_balancing))
Mel Gorman1a687c22012-11-22 11:16:36 +00002404 return;
2405
Mel Gorman9ff1d9f2013-10-07 11:29:04 +01002406 /* for example, ksmd faulting in a user's mm */
2407 if (!p->mm)
2408 return;
2409
Mel Gormanf809ca92013-10-07 11:28:57 +01002410 /* Allocate buffer to track faults on a per-node basis */
Iulia Manda44dba3d2014-10-31 02:13:31 +02002411 if (unlikely(!p->numa_faults)) {
2412 int size = sizeof(*p->numa_faults) *
Rik van Rielbe1e4e72014-01-27 17:03:48 -05002413 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
Mel Gormanf809ca92013-10-07 11:28:57 +01002414
Iulia Manda44dba3d2014-10-31 02:13:31 +02002415 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2416 if (!p->numa_faults)
Mel Gormanf809ca92013-10-07 11:28:57 +01002417 return;
Mel Gorman745d6142013-10-07 11:28:59 +01002418
Mel Gorman83e1d2c2013-10-07 11:29:27 +01002419 p->total_numa_faults = 0;
Rik van Riel04bb2f92013-10-07 11:29:36 +01002420 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
Mel Gormanf809ca92013-10-07 11:28:57 +01002421 }
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002422
Mel Gormanfb003b82012-11-15 09:01:14 +00002423 /*
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002424 * First accesses are treated as private, otherwise consider accesses
2425 * to be private if the accessing pid has not changed
2426 */
2427 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2428 priv = 1;
2429 } else {
2430 priv = cpupid_match_pid(p, last_cpupid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01002431 if (!priv && !(flags & TNF_NO_GROUP))
Mel Gorman3e6a9412013-10-07 11:29:35 +01002432 task_numa_group(p, last_cpupid, flags, &priv);
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002433 }
2434
Rik van Riel792568e2014-04-11 13:00:27 -04002435 /*
2436 * If a workload spans multiple NUMA nodes, a shared fault that
2437 * occurs wholly within the set of nodes that the workload is
2438 * actively using should be counted as local. This allows the
2439 * scan rate to slow down when a workload has settled down.
2440 */
Rik van Riel4142c3e2016-01-25 17:07:39 -05002441 ng = p->numa_group;
2442 if (!priv && !local && ng && ng->active_nodes > 1 &&
2443 numa_is_active_node(cpu_node, ng) &&
2444 numa_is_active_node(mem_node, ng))
Rik van Riel792568e2014-04-11 13:00:27 -04002445 local = 1;
2446
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002447 task_numa_placement(p);
Mel Gormanf809ca92013-10-07 11:28:57 +01002448
Rik van Riel2739d3e2013-10-07 11:29:41 +01002449 /*
2450 * Retry task to preferred node migration periodically, in case it
2451 * case it previously failed, or the scheduler moved us.
2452 */
2453 if (time_after(jiffies, p->numa_migrate_retry))
Mel Gorman6b9a7462013-10-07 11:29:11 +01002454 numa_migrate_preferred(p);
2455
Ingo Molnarb32e86b2013-10-07 11:29:30 +01002456 if (migrated)
2457 p->numa_pages_migrated += pages;
Mel Gorman074c2382015-03-25 15:55:42 -07002458 if (flags & TNF_MIGRATE_FAIL)
2459 p->numa_faults_locality[2] += pages;
Ingo Molnarb32e86b2013-10-07 11:29:30 +01002460
Iulia Manda44dba3d2014-10-31 02:13:31 +02002461 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2462 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
Rik van Riel792568e2014-04-11 13:00:27 -04002463 p->numa_faults_locality[local] += pages;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002464}
2465
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002466static void reset_ptenuma_scan(struct task_struct *p)
2467{
Jason Low7e5a2c12015-04-30 17:28:14 -07002468 /*
2469 * We only did a read acquisition of the mmap sem, so
2470 * p->mm->numa_scan_seq is written to without exclusive access
2471 * and the update is not guaranteed to be atomic. That's not
2472 * much of an issue though, since this is just used for
2473 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2474 * expensive, to avoid any form of compiler optimizations:
2475 */
Jason Low316c1608d2015-04-28 13:00:20 -07002476 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002477 p->mm->numa_scan_offset = 0;
2478}
2479
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002480/*
2481 * The expensive part of numa migration is done from task_work context.
2482 * Triggered from task_tick_numa().
2483 */
2484void task_numa_work(struct callback_head *work)
2485{
2486 unsigned long migrate, next_scan, now = jiffies;
2487 struct task_struct *p = current;
2488 struct mm_struct *mm = p->mm;
Rik van Riel51170842015-11-05 15:56:23 -05002489 u64 runtime = p->se.sum_exec_runtime;
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002490 struct vm_area_struct *vma;
Mel Gorman9f406042012-11-14 18:34:32 +00002491 unsigned long start, end;
Mel Gorman598f0ec2013-10-07 11:28:55 +01002492 unsigned long nr_pte_updates = 0;
Rik van Riel4620f8c2015-09-11 09:00:27 -04002493 long pages, virtpages;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002494
Peter Zijlstra9148a3a2016-09-20 22:34:51 +02002495 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002496
2497 work->next = work; /* protect against double add */
2498 /*
2499 * Who cares about NUMA placement when they're dying.
2500 *
2501 * NOTE: make sure not to dereference p->mm before this check,
2502 * exit_task_work() happens _after_ exit_mm() so we could be called
2503 * without p->mm even though we still had it when we enqueued this
2504 * work.
2505 */
2506 if (p->flags & PF_EXITING)
2507 return;
2508
Mel Gorman930aa172013-10-07 11:29:37 +01002509 if (!mm->numa_next_scan) {
Mel Gorman7e8d16b2013-10-07 11:28:54 +01002510 mm->numa_next_scan = now +
2511 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
Mel Gormanb8593bf2012-11-21 01:18:23 +00002512 }
2513
2514 /*
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002515 * Enforce maximal scan/migration frequency..
2516 */
2517 migrate = mm->numa_next_scan;
2518 if (time_before(now, migrate))
2519 return;
2520
Mel Gorman598f0ec2013-10-07 11:28:55 +01002521 if (p->numa_scan_period == 0) {
2522 p->numa_scan_period_max = task_scan_max(p);
Rik van Rielb5dd77c2017-07-31 15:28:47 -04002523 p->numa_scan_period = task_scan_start(p);
Mel Gorman598f0ec2013-10-07 11:28:55 +01002524 }
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002525
Mel Gormanfb003b82012-11-15 09:01:14 +00002526 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002527 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2528 return;
2529
Mel Gormane14808b2012-11-19 10:59:15 +00002530 /*
Peter Zijlstra19a78d12013-10-07 11:28:51 +01002531 * Delay this task enough that another task of this mm will likely win
2532 * the next time around.
2533 */
2534 p->node_stamp += 2 * TICK_NSEC;
2535
Mel Gorman9f406042012-11-14 18:34:32 +00002536 start = mm->numa_scan_offset;
2537 pages = sysctl_numa_balancing_scan_size;
2538 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
Rik van Riel4620f8c2015-09-11 09:00:27 -04002539 virtpages = pages * 8; /* Scan up to this much virtual space */
Mel Gorman9f406042012-11-14 18:34:32 +00002540 if (!pages)
2541 return;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002542
Rik van Riel4620f8c2015-09-11 09:00:27 -04002543
Vlastimil Babka8655d542017-05-15 15:13:16 +02002544 if (!down_read_trylock(&mm->mmap_sem))
2545 return;
Mel Gorman9f406042012-11-14 18:34:32 +00002546 vma = find_vma(mm, start);
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002547 if (!vma) {
2548 reset_ptenuma_scan(p);
Mel Gorman9f406042012-11-14 18:34:32 +00002549 start = 0;
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002550 vma = mm->mmap;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002551 }
Mel Gorman9f406042012-11-14 18:34:32 +00002552 for (; vma; vma = vma->vm_next) {
Naoya Horiguchi6b79c572015-04-07 14:26:47 -07002553 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
Mel Gorman8e76d4e2015-06-10 11:15:00 -07002554 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002555 continue;
Naoya Horiguchi6b79c572015-04-07 14:26:47 -07002556 }
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002557
Mel Gorman4591ce4f2013-10-07 11:29:13 +01002558 /*
2559 * Shared library pages mapped by multiple processes are not
2560 * migrated as it is expected they are cache replicated. Avoid
2561 * hinting faults in read-only file-backed mappings or the vdso
2562 * as migrating the pages will be of marginal benefit.
2563 */
2564 if (!vma->vm_mm ||
2565 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2566 continue;
2567
Mel Gorman3c67f472013-12-18 17:08:40 -08002568 /*
2569 * Skip inaccessible VMAs to avoid any confusion between
2570 * PROT_NONE and NUMA hinting ptes
2571 */
2572 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2573 continue;
2574
Mel Gorman9f406042012-11-14 18:34:32 +00002575 do {
2576 start = max(start, vma->vm_start);
2577 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2578 end = min(end, vma->vm_end);
Rik van Riel4620f8c2015-09-11 09:00:27 -04002579 nr_pte_updates = change_prot_numa(vma, start, end);
Mel Gorman598f0ec2013-10-07 11:28:55 +01002580
2581 /*
Rik van Riel4620f8c2015-09-11 09:00:27 -04002582 * Try to scan sysctl_numa_balancing_size worth of
2583 * hpages that have at least one present PTE that
2584 * is not already pte-numa. If the VMA contains
2585 * areas that are unused or already full of prot_numa
2586 * PTEs, scan up to virtpages, to skip through those
2587 * areas faster.
Mel Gorman598f0ec2013-10-07 11:28:55 +01002588 */
2589 if (nr_pte_updates)
2590 pages -= (end - start) >> PAGE_SHIFT;
Rik van Riel4620f8c2015-09-11 09:00:27 -04002591 virtpages -= (end - start) >> PAGE_SHIFT;
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002592
Mel Gorman9f406042012-11-14 18:34:32 +00002593 start = end;
Rik van Riel4620f8c2015-09-11 09:00:27 -04002594 if (pages <= 0 || virtpages <= 0)
Mel Gorman9f406042012-11-14 18:34:32 +00002595 goto out;
Rik van Riel3cf19622014-02-18 17:12:44 -05002596
2597 cond_resched();
Mel Gorman9f406042012-11-14 18:34:32 +00002598 } while (end != vma->vm_end);
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002599 }
2600
Mel Gorman9f406042012-11-14 18:34:32 +00002601out:
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002602 /*
Peter Zijlstrac69307d2013-10-07 11:28:41 +01002603 * It is possible to reach the end of the VMA list but the last few
2604 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2605 * would find the !migratable VMA on the next scan but not reset the
2606 * scanner to the start so check it now.
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002607 */
2608 if (vma)
Mel Gorman9f406042012-11-14 18:34:32 +00002609 mm->numa_scan_offset = start;
Peter Zijlstra6e5fb222012-10-25 14:16:45 +02002610 else
2611 reset_ptenuma_scan(p);
2612 up_read(&mm->mmap_sem);
Rik van Riel51170842015-11-05 15:56:23 -05002613
2614 /*
2615 * Make sure tasks use at least 32x as much time to run other code
2616 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2617 * Usually update_task_scan_period slows down scanning enough; on an
2618 * overloaded system we need to limit overhead on a per task basis.
2619 */
2620 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2621 u64 diff = p->se.sum_exec_runtime - runtime;
2622 p->node_stamp += 32 * diff;
2623 }
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002624}
2625
2626/*
2627 * Drive the periodic memory faults..
2628 */
2629void task_tick_numa(struct rq *rq, struct task_struct *curr)
2630{
2631 struct callback_head *work = &curr->numa_work;
2632 u64 period, now;
2633
2634 /*
2635 * We don't care about NUMA placement if we don't have memory.
2636 */
2637 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2638 return;
2639
2640 /*
2641 * Using runtime rather than walltime has the dual advantage that
2642 * we (mostly) drive the selection from busy threads and that the
2643 * task needs to have done some actual work before we bother with
2644 * NUMA placement.
2645 */
2646 now = curr->se.sum_exec_runtime;
2647 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2648
Rik van Riel25b3e5a2015-11-05 15:56:22 -05002649 if (now > curr->node_stamp + period) {
Peter Zijlstra4b96a292012-10-25 14:16:47 +02002650 if (!curr->node_stamp)
Rik van Rielb5dd77c2017-07-31 15:28:47 -04002651 curr->numa_scan_period = task_scan_start(curr);
Peter Zijlstra19a78d12013-10-07 11:28:51 +01002652 curr->node_stamp += period;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002653
2654 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2655 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2656 task_work_add(curr, work, true);
2657 }
2658 }
2659}
Rik van Riel3fed3822017-06-23 12:55:29 -04002660
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002661#else
2662static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2663{
2664}
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01002665
2666static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2667{
2668}
2669
2670static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2671{
2672}
Rik van Riel3fed3822017-06-23 12:55:29 -04002673
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002674#endif /* CONFIG_NUMA_BALANCING */
2675
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02002676static void
2677account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2678{
2679 update_load_add(&cfs_rq->load, se->load.weight);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02002680 if (!parent_entity(se))
Peter Zijlstra029632f2011-10-25 10:00:11 +02002681 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
Peter Zijlstra367456c2012-02-20 21:49:09 +01002682#ifdef CONFIG_SMP
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01002683 if (entity_is_task(se)) {
2684 struct rq *rq = rq_of(cfs_rq);
2685
2686 account_numa_enqueue(rq, task_of(se));
2687 list_add(&se->group_node, &rq->cfs_tasks);
2688 }
Peter Zijlstra367456c2012-02-20 21:49:09 +01002689#endif
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02002690 cfs_rq->nr_running++;
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02002691}
2692
2693static void
2694account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2695{
2696 update_load_sub(&cfs_rq->load, se->load.weight);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02002697 if (!parent_entity(se))
Peter Zijlstra029632f2011-10-25 10:00:11 +02002698 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
Tim Chenbfdb1982016-02-01 14:47:59 -08002699#ifdef CONFIG_SMP
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01002700 if (entity_is_task(se)) {
2701 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
Bharata B Raob87f1722008-09-25 09:53:54 +05302702 list_del_init(&se->group_node);
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01002703 }
Tim Chenbfdb1982016-02-01 14:47:59 -08002704#endif
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02002705 cfs_rq->nr_running--;
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02002706}
2707
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002708#ifdef CONFIG_FAIR_GROUP_SCHED
2709# ifdef CONFIG_SMP
Paul Turner6d5ab292011-01-21 20:45:01 -08002710static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002711{
Peter Zijlstracf5f0ac2011-10-13 16:52:28 +02002712 long tg_weight, load, shares;
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002713
Peter Zijlstraea1dc6f2016-06-24 16:11:02 +02002714 /*
2715 * This really should be: cfs_rq->avg.load_avg, but instead we use
2716 * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2717 * the shares for small weight interactive tasks.
2718 */
2719 load = scale_load_down(cfs_rq->load.weight);
2720
2721 tg_weight = atomic_long_read(&tg->load_avg);
2722
2723 /* Ensure tg_weight >= load */
2724 tg_weight -= cfs_rq->tg_load_avg_contrib;
2725 tg_weight += load;
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002726
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002727 shares = (tg->shares * load);
Peter Zijlstracf5f0ac2011-10-13 16:52:28 +02002728 if (tg_weight)
2729 shares /= tg_weight;
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002730
Dietmar Eggemannb8fd8422017-01-11 11:29:47 +00002731 /*
2732 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2733 * of a group with small tg->shares value. It is a floor value which is
2734 * assigned as a minimum load.weight to the sched_entity representing
2735 * the group on a CPU.
2736 *
2737 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2738 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2739 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2740 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2741 * instead of 0.
2742 */
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002743 if (shares < MIN_SHARES)
2744 shares = MIN_SHARES;
2745 if (shares > tg->shares)
2746 shares = tg->shares;
2747
2748 return shares;
2749}
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002750# else /* CONFIG_SMP */
Paul Turner6d5ab292011-01-21 20:45:01 -08002751static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002752{
2753 return tg->shares;
2754}
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002755# endif /* CONFIG_SMP */
Peter Zijlstraea1dc6f2016-06-24 16:11:02 +02002756
Peter Zijlstra2069dd72010-11-15 15:47:00 -08002757static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2758 unsigned long weight)
2759{
Paul Turner19e5eeb2010-12-15 19:10:18 -08002760 if (se->on_rq) {
2761 /* commit outstanding execution time */
2762 if (cfs_rq->curr == se)
2763 update_curr(cfs_rq);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08002764 account_entity_dequeue(cfs_rq, se);
Paul Turner19e5eeb2010-12-15 19:10:18 -08002765 }
Peter Zijlstra2069dd72010-11-15 15:47:00 -08002766
2767 update_load_set(&se->load, weight);
2768
2769 if (se->on_rq)
2770 account_entity_enqueue(cfs_rq, se);
2771}
2772
Paul Turner82958362012-10-04 13:18:31 +02002773static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2774
Vincent Guittot89ee0482016-12-21 16:50:26 +01002775static void update_cfs_shares(struct sched_entity *se)
Peter Zijlstra2069dd72010-11-15 15:47:00 -08002776{
Vincent Guittot89ee0482016-12-21 16:50:26 +01002777 struct cfs_rq *cfs_rq = group_cfs_rq(se);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08002778 struct task_group *tg;
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002779 long shares;
Peter Zijlstra2069dd72010-11-15 15:47:00 -08002780
Vincent Guittot89ee0482016-12-21 16:50:26 +01002781 if (!cfs_rq)
Peter Zijlstra2069dd72010-11-15 15:47:00 -08002782 return;
Vincent Guittot89ee0482016-12-21 16:50:26 +01002783
2784 if (throttled_hierarchy(cfs_rq))
2785 return;
2786
2787 tg = cfs_rq->tg;
2788
Yong Zhang3ff6dca2011-01-24 15:33:52 +08002789#ifndef CONFIG_SMP
2790 if (likely(se->load.weight == tg->shares))
2791 return;
2792#endif
Paul Turner6d5ab292011-01-21 20:45:01 -08002793 shares = calc_cfs_shares(cfs_rq, tg);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08002794
2795 reweight_entity(cfs_rq_of(se), se, shares);
2796}
Vincent Guittot89ee0482016-12-21 16:50:26 +01002797
Peter Zijlstra2069dd72010-11-15 15:47:00 -08002798#else /* CONFIG_FAIR_GROUP_SCHED */
Vincent Guittot89ee0482016-12-21 16:50:26 +01002799static inline void update_cfs_shares(struct sched_entity *se)
Peter Zijlstra2069dd72010-11-15 15:47:00 -08002800{
2801}
2802#endif /* CONFIG_FAIR_GROUP_SCHED */
2803
Viresh Kumara030d732017-05-24 10:59:52 +05302804static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2805{
2806 if (&this_rq()->cfs == cfs_rq) {
2807 /*
2808 * There are a few boundary cases this might miss but it should
2809 * get called often enough that that should (hopefully) not be
2810 * a real problem -- added to that it only calls on the local
2811 * CPU, so if we enqueue remotely we'll miss an update, but
2812 * the next tick/schedule should update.
2813 *
2814 * It will not get called when we go idle, because the idle
2815 * thread is a different class (!fair), nor will the utilization
2816 * number include things like RT tasks.
2817 *
2818 * As is, the util number is not freq-invariant (we'd have to
2819 * implement arch_scale_freq_capacity() for that).
2820 *
2821 * See cpu_util().
2822 */
2823 cpufreq_update_util(rq_of(cfs_rq), 0);
2824 }
2825}
2826
Alex Shi141965c2013-06-26 13:05:39 +08002827#ifdef CONFIG_SMP
Paul Turner5b51f2f2012-10-04 13:18:32 +02002828/*
Paul Turner9d85f212012-10-04 13:18:29 +02002829 * Approximate:
2830 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2831 */
Yuyang Dua481db32017-02-13 05:44:23 +08002832static u64 decay_load(u64 val, u64 n)
Paul Turner9d85f212012-10-04 13:18:29 +02002833{
Paul Turner5b51f2f2012-10-04 13:18:32 +02002834 unsigned int local_n;
2835
Peter Zijlstra05296e72017-03-31 10:51:41 +02002836 if (unlikely(n > LOAD_AVG_PERIOD * 63))
Paul Turner5b51f2f2012-10-04 13:18:32 +02002837 return 0;
2838
2839 /* after bounds checking we can collapse to 32-bit */
2840 local_n = n;
2841
2842 /*
2843 * As y^PERIOD = 1/2, we can combine
Zhihui Zhang9c58c792014-09-20 21:24:36 -04002844 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2845 * With a look-up table which covers y^n (n<PERIOD)
Paul Turner5b51f2f2012-10-04 13:18:32 +02002846 *
2847 * To achieve constant time decay_load.
2848 */
2849 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2850 val >>= local_n / LOAD_AVG_PERIOD;
2851 local_n %= LOAD_AVG_PERIOD;
Paul Turner9d85f212012-10-04 13:18:29 +02002852 }
2853
Yuyang Du9d89c252015-07-15 08:04:37 +08002854 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2855 return val;
Paul Turner5b51f2f2012-10-04 13:18:32 +02002856}
2857
Peter Zijlstra05296e72017-03-31 10:51:41 +02002858static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
Paul Turner5b51f2f2012-10-04 13:18:32 +02002859{
Peter Zijlstra05296e72017-03-31 10:51:41 +02002860 u32 c1, c2, c3 = d3; /* y^0 == 1 */
Paul Turner5b51f2f2012-10-04 13:18:32 +02002861
Yuyang Dua481db32017-02-13 05:44:23 +08002862 /*
Peter Zijlstra3841cdc2017-04-10 12:47:33 +02002863 * c1 = d1 y^p
Yuyang Dua481db32017-02-13 05:44:23 +08002864 */
Peter Zijlstra05296e72017-03-31 10:51:41 +02002865 c1 = decay_load((u64)d1, periods);
Yuyang Dua481db32017-02-13 05:44:23 +08002866
Yuyang Dua481db32017-02-13 05:44:23 +08002867 /*
Peter Zijlstra3841cdc2017-04-10 12:47:33 +02002868 * p-1
Peter Zijlstra05296e72017-03-31 10:51:41 +02002869 * c2 = 1024 \Sum y^n
2870 * n=1
Yuyang Dua481db32017-02-13 05:44:23 +08002871 *
Peter Zijlstra05296e72017-03-31 10:51:41 +02002872 * inf inf
2873 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
Peter Zijlstra3841cdc2017-04-10 12:47:33 +02002874 * n=0 n=p
Yuyang Dua481db32017-02-13 05:44:23 +08002875 */
Peter Zijlstra05296e72017-03-31 10:51:41 +02002876 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
Yuyang Dua481db32017-02-13 05:44:23 +08002877
2878 return c1 + c2 + c3;
Paul Turner9d85f212012-10-04 13:18:29 +02002879}
2880
Peter Zijlstra54a21382015-09-07 15:05:42 +02002881#define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
Dietmar Eggemanne0f5f3a2015-08-14 17:23:09 +01002882
Paul Turner9d85f212012-10-04 13:18:29 +02002883/*
Yuyang Dua481db32017-02-13 05:44:23 +08002884 * Accumulate the three separate parts of the sum; d1 the remainder
2885 * of the last (incomplete) period, d2 the span of full periods and d3
2886 * the remainder of the (incomplete) current period.
2887 *
2888 * d1 d2 d3
2889 * ^ ^ ^
2890 * | | |
2891 * |<->|<----------------->|<--->|
2892 * ... |---x---|------| ... |------|-----x (now)
2893 *
Peter Zijlstra3841cdc2017-04-10 12:47:33 +02002894 * p-1
2895 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
2896 * n=1
Yuyang Dua481db32017-02-13 05:44:23 +08002897 *
Peter Zijlstra3841cdc2017-04-10 12:47:33 +02002898 * = u y^p + (Step 1)
Yuyang Dua481db32017-02-13 05:44:23 +08002899 *
Peter Zijlstra3841cdc2017-04-10 12:47:33 +02002900 * p-1
2901 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
2902 * n=1
Yuyang Dua481db32017-02-13 05:44:23 +08002903 */
2904static __always_inline u32
2905accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
2906 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2907{
2908 unsigned long scale_freq, scale_cpu;
Peter Zijlstra05296e72017-03-31 10:51:41 +02002909 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
Yuyang Dua481db32017-02-13 05:44:23 +08002910 u64 periods;
Yuyang Dua481db32017-02-13 05:44:23 +08002911
2912 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2913 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2914
2915 delta += sa->period_contrib;
2916 periods = delta / 1024; /* A period is 1024us (~1ms) */
2917
2918 /*
2919 * Step 1: decay old *_sum if we crossed period boundaries.
2920 */
2921 if (periods) {
2922 sa->load_sum = decay_load(sa->load_sum, periods);
2923 if (cfs_rq) {
2924 cfs_rq->runnable_load_sum =
2925 decay_load(cfs_rq->runnable_load_sum, periods);
2926 }
2927 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
Yuyang Dua481db32017-02-13 05:44:23 +08002928
Peter Zijlstra05296e72017-03-31 10:51:41 +02002929 /*
2930 * Step 2
2931 */
2932 delta %= 1024;
2933 contrib = __accumulate_pelt_segments(periods,
2934 1024 - sa->period_contrib, delta);
2935 }
Yuyang Dua481db32017-02-13 05:44:23 +08002936 sa->period_contrib = delta;
2937
2938 contrib = cap_scale(contrib, scale_freq);
2939 if (weight) {
2940 sa->load_sum += weight * contrib;
2941 if (cfs_rq)
2942 cfs_rq->runnable_load_sum += weight * contrib;
2943 }
2944 if (running)
2945 sa->util_sum += contrib * scale_cpu;
2946
2947 return periods;
2948}
2949
2950/*
Paul Turner9d85f212012-10-04 13:18:29 +02002951 * We can represent the historical contribution to runnable average as the
2952 * coefficients of a geometric series. To do this we sub-divide our runnable
2953 * history into segments of approximately 1ms (1024us); label the segment that
2954 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2955 *
2956 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2957 * p0 p1 p2
2958 * (now) (~1ms ago) (~2ms ago)
2959 *
2960 * Let u_i denote the fraction of p_i that the entity was runnable.
2961 *
2962 * We then designate the fractions u_i as our co-efficients, yielding the
2963 * following representation of historical load:
2964 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2965 *
2966 * We choose y based on the with of a reasonably scheduling period, fixing:
2967 * y^32 = 0.5
2968 *
2969 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2970 * approximately half as much as the contribution to load within the last ms
2971 * (u_0).
2972 *
2973 * When a period "rolls over" and we have new u_0`, multiplying the previous
2974 * sum again by y is sufficient to update:
2975 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2976 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2977 */
Yuyang Du9d89c252015-07-15 08:04:37 +08002978static __always_inline int
Peter Zijlstra0ccb9772017-03-28 11:08:20 +02002979___update_load_avg(u64 now, int cpu, struct sched_avg *sa,
Yuyang Du13962232015-07-15 08:04:41 +08002980 unsigned long weight, int running, struct cfs_rq *cfs_rq)
Paul Turner9d85f212012-10-04 13:18:29 +02002981{
Yuyang Dua481db32017-02-13 05:44:23 +08002982 u64 delta;
Paul Turner9d85f212012-10-04 13:18:29 +02002983
Yuyang Du9d89c252015-07-15 08:04:37 +08002984 delta = now - sa->last_update_time;
Paul Turner9d85f212012-10-04 13:18:29 +02002985 /*
2986 * This should only happen when time goes backwards, which it
2987 * unfortunately does during sched clock init when we swap over to TSC.
2988 */
2989 if ((s64)delta < 0) {
Yuyang Du9d89c252015-07-15 08:04:37 +08002990 sa->last_update_time = now;
Paul Turner9d85f212012-10-04 13:18:29 +02002991 return 0;
2992 }
2993
2994 /*
2995 * Use 1024ns as the unit of measurement since it's a reasonable
2996 * approximation of 1us and fast to compute.
2997 */
2998 delta >>= 10;
2999 if (!delta)
3000 return 0;
Peter Zijlstrabb0bd042017-04-10 13:20:45 +02003001
3002 sa->last_update_time += delta << 10;
Paul Turner9d85f212012-10-04 13:18:29 +02003003
Yuyang Dua481db32017-02-13 05:44:23 +08003004 /*
Vincent Guittotf235a542017-07-01 07:06:13 +02003005 * running is a subset of runnable (weight) so running can't be set if
3006 * runnable is clear. But there are some corner cases where the current
3007 * se has been already dequeued but cfs_rq->curr still points to it.
3008 * This means that weight will be 0 but not running for a sched_entity
3009 * but also for a cfs_rq if the latter becomes idle. As an example,
3010 * this happens during idle_balance() which calls
3011 * update_blocked_averages()
3012 */
3013 if (!weight)
3014 running = 0;
3015
3016 /*
Yuyang Dua481db32017-02-13 05:44:23 +08003017 * Now we know we crossed measurement unit boundaries. The *_avg
3018 * accrues by two steps:
3019 *
3020 * Step 1: accumulate *_sum since last_update_time. If we haven't
3021 * crossed period boundaries, finish.
3022 */
3023 if (!accumulate_sum(delta, cpu, sa, weight, running, cfs_rq))
3024 return 0;
Dietmar Eggemann6f2b0452015-09-07 14:57:22 +01003025
Yuyang Dua481db32017-02-13 05:44:23 +08003026 /*
3027 * Step 2: update *_avg.
3028 */
Vincent Guittot625ed2b2017-04-26 08:27:56 +02003029 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
Yuyang Dua481db32017-02-13 05:44:23 +08003030 if (cfs_rq) {
3031 cfs_rq->runnable_load_avg =
Vincent Guittot625ed2b2017-04-26 08:27:56 +02003032 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
Paul Turner9d85f212012-10-04 13:18:29 +02003033 }
Vincent Guittot625ed2b2017-04-26 08:27:56 +02003034 sa->util_avg = sa->util_sum / (LOAD_AVG_MAX - 1024 + sa->period_contrib);
Paul Turner9d85f212012-10-04 13:18:29 +02003035
Yuyang Dua481db32017-02-13 05:44:23 +08003036 return 1;
Paul Turner9d85f212012-10-04 13:18:29 +02003037}
3038
Peter Zijlstra0ccb9772017-03-28 11:08:20 +02003039static int
3040__update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3041{
3042 return ___update_load_avg(now, cpu, &se->avg, 0, 0, NULL);
3043}
3044
3045static int
3046__update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3047{
3048 return ___update_load_avg(now, cpu, &se->avg,
3049 se->on_rq * scale_load_down(se->load.weight),
3050 cfs_rq->curr == se, NULL);
3051}
3052
3053static int
3054__update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3055{
3056 return ___update_load_avg(now, cpu, &cfs_rq->avg,
3057 scale_load_down(cfs_rq->load.weight),
3058 cfs_rq->curr != NULL, cfs_rq);
3059}
3060
Vincent Guittot09a43ac2016-11-08 10:53:45 +01003061/*
3062 * Signed add and clamp on underflow.
3063 *
3064 * Explicitly do a load-store to ensure the intermediate value never hits
3065 * memory. This allows lockless observations without ever seeing the negative
3066 * values.
3067 */
3068#define add_positive(_ptr, _val) do { \
3069 typeof(_ptr) ptr = (_ptr); \
3070 typeof(_val) val = (_val); \
3071 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3072 \
3073 res = var + val; \
3074 \
3075 if (val < 0 && res > var) \
3076 res = 0; \
3077 \
3078 WRITE_ONCE(*ptr, res); \
3079} while (0)
3080
Paul Turnerc566e8e2012-10-04 13:18:30 +02003081#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra7c3edd22016-07-13 10:56:25 +02003082/**
3083 * update_tg_load_avg - update the tg's load avg
3084 * @cfs_rq: the cfs_rq whose avg changed
3085 * @force: update regardless of how small the difference
3086 *
3087 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3088 * However, because tg->load_avg is a global value there are performance
3089 * considerations.
3090 *
3091 * In order to avoid having to look at the other cfs_rq's, we use a
3092 * differential update where we store the last value we propagated. This in
3093 * turn allows skipping updates if the differential is 'small'.
3094 *
Rik van Riel815abf52017-06-23 12:55:30 -04003095 * Updating tg's load_avg is necessary before update_cfs_share().
Paul Turnerbb17f652012-10-04 13:18:31 +02003096 */
Yuyang Du9d89c252015-07-15 08:04:37 +08003097static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
Paul Turnerbb17f652012-10-04 13:18:31 +02003098{
Yuyang Du9d89c252015-07-15 08:04:37 +08003099 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
Paul Turnerbb17f652012-10-04 13:18:31 +02003100
Waiman Longaa0b7ae2015-12-02 13:41:50 -05003101 /*
3102 * No need to update load_avg for root_task_group as it is not used.
3103 */
3104 if (cfs_rq->tg == &root_task_group)
3105 return;
3106
Yuyang Du9d89c252015-07-15 08:04:37 +08003107 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3108 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3109 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
Paul Turnerbb17f652012-10-04 13:18:31 +02003110 }
Paul Turner8165e142012-10-04 13:18:31 +02003111}
Dietmar Eggemannf5f97392014-02-26 11:19:33 +00003112
Byungchul Parkad936d82015-10-24 01:16:19 +09003113/*
3114 * Called within set_task_rq() right before setting a task's cpu. The
3115 * caller only guarantees p->pi_lock is held; no other assumptions,
3116 * including the state of rq->lock, should be made.
3117 */
3118void set_task_rq_fair(struct sched_entity *se,
3119 struct cfs_rq *prev, struct cfs_rq *next)
3120{
Peter Zijlstra0ccb9772017-03-28 11:08:20 +02003121 u64 p_last_update_time;
3122 u64 n_last_update_time;
3123
Byungchul Parkad936d82015-10-24 01:16:19 +09003124 if (!sched_feat(ATTACH_AGE_LOAD))
3125 return;
3126
3127 /*
3128 * We are supposed to update the task to "current" time, then its up to
3129 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3130 * getting what current time is, so simply throw away the out-of-date
3131 * time. This will result in the wakee task is less decayed, but giving
3132 * the wakee more load sounds not bad.
3133 */
Peter Zijlstra0ccb9772017-03-28 11:08:20 +02003134 if (!(se->avg.last_update_time && prev))
3135 return;
Byungchul Parkad936d82015-10-24 01:16:19 +09003136
3137#ifndef CONFIG_64BIT
Peter Zijlstra0ccb9772017-03-28 11:08:20 +02003138 {
Byungchul Parkad936d82015-10-24 01:16:19 +09003139 u64 p_last_update_time_copy;
3140 u64 n_last_update_time_copy;
3141
3142 do {
3143 p_last_update_time_copy = prev->load_last_update_time_copy;
3144 n_last_update_time_copy = next->load_last_update_time_copy;
3145
3146 smp_rmb();
3147
3148 p_last_update_time = prev->avg.last_update_time;
3149 n_last_update_time = next->avg.last_update_time;
3150
3151 } while (p_last_update_time != p_last_update_time_copy ||
3152 n_last_update_time != n_last_update_time_copy);
Byungchul Parkad936d82015-10-24 01:16:19 +09003153 }
Peter Zijlstra0ccb9772017-03-28 11:08:20 +02003154#else
3155 p_last_update_time = prev->avg.last_update_time;
3156 n_last_update_time = next->avg.last_update_time;
3157#endif
3158 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3159 se->avg.last_update_time = n_last_update_time;
Byungchul Parkad936d82015-10-24 01:16:19 +09003160}
Vincent Guittot09a43ac2016-11-08 10:53:45 +01003161
3162/* Take into account change of utilization of a child task group */
3163static inline void
3164update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
3165{
3166 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3167 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3168
3169 /* Nothing to update */
3170 if (!delta)
3171 return;
3172
3173 /* Set new sched_entity's utilization */
3174 se->avg.util_avg = gcfs_rq->avg.util_avg;
3175 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3176
3177 /* Update parent cfs_rq utilization */
3178 add_positive(&cfs_rq->avg.util_avg, delta);
3179 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3180}
3181
3182/* Take into account change of load of a child task group */
3183static inline void
3184update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
3185{
3186 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3187 long delta, load = gcfs_rq->avg.load_avg;
3188
3189 /*
3190 * If the load of group cfs_rq is null, the load of the
3191 * sched_entity will also be null so we can skip the formula
3192 */
3193 if (load) {
3194 long tg_load;
3195
3196 /* Get tg's load and ensure tg_load > 0 */
3197 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
3198
3199 /* Ensure tg_load >= load and updated with current load*/
3200 tg_load -= gcfs_rq->tg_load_avg_contrib;
3201 tg_load += load;
3202
3203 /*
3204 * We need to compute a correction term in the case that the
3205 * task group is consuming more CPU than a task of equal
3206 * weight. A task with a weight equals to tg->shares will have
3207 * a load less or equal to scale_load_down(tg->shares).
3208 * Similarly, the sched_entities that represent the task group
3209 * at parent level, can't have a load higher than
3210 * scale_load_down(tg->shares). And the Sum of sched_entities'
3211 * load must be <= scale_load_down(tg->shares).
3212 */
3213 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3214 /* scale gcfs_rq's load into tg's shares*/
3215 load *= scale_load_down(gcfs_rq->tg->shares);
3216 load /= tg_load;
3217 }
3218 }
3219
3220 delta = load - se->avg.load_avg;
3221
3222 /* Nothing to update */
3223 if (!delta)
3224 return;
3225
3226 /* Set new sched_entity's load */
3227 se->avg.load_avg = load;
3228 se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3229
3230 /* Update parent cfs_rq load */
3231 add_positive(&cfs_rq->avg.load_avg, delta);
3232 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3233
3234 /*
3235 * If the sched_entity is already enqueued, we also have to update the
3236 * runnable load avg.
3237 */
3238 if (se->on_rq) {
3239 /* Update parent cfs_rq runnable_load_avg */
3240 add_positive(&cfs_rq->runnable_load_avg, delta);
3241 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3242 }
3243}
3244
3245static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3246{
3247 cfs_rq->propagate_avg = 1;
3248}
3249
3250static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3251{
3252 struct cfs_rq *cfs_rq = group_cfs_rq(se);
3253
3254 if (!cfs_rq->propagate_avg)
3255 return 0;
3256
3257 cfs_rq->propagate_avg = 0;
3258 return 1;
3259}
3260
3261/* Update task and its cfs_rq load average */
3262static inline int propagate_entity_load_avg(struct sched_entity *se)
3263{
3264 struct cfs_rq *cfs_rq;
3265
3266 if (entity_is_task(se))
3267 return 0;
3268
3269 if (!test_and_clear_tg_cfs_propagate(se))
3270 return 0;
3271
3272 cfs_rq = cfs_rq_of(se);
3273
3274 set_tg_cfs_propagate(cfs_rq);
3275
3276 update_tg_cfs_util(cfs_rq, se);
3277 update_tg_cfs_load(cfs_rq, se);
3278
3279 return 1;
3280}
3281
Vincent Guittotbc427892017-03-17 14:47:22 +01003282/*
3283 * Check if we need to update the load and the utilization of a blocked
3284 * group_entity:
3285 */
3286static inline bool skip_blocked_update(struct sched_entity *se)
3287{
3288 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3289
3290 /*
3291 * If sched_entity still have not zero load or utilization, we have to
3292 * decay it:
3293 */
3294 if (se->avg.load_avg || se->avg.util_avg)
3295 return false;
3296
3297 /*
3298 * If there is a pending propagation, we have to update the load and
3299 * the utilization of the sched_entity:
3300 */
3301 if (gcfs_rq->propagate_avg)
3302 return false;
3303
3304 /*
3305 * Otherwise, the load and the utilization of the sched_entity is
3306 * already zero and there is no pending propagation, so it will be a
3307 * waste of time to try to decay it:
3308 */
3309 return true;
3310}
3311
Peter Zijlstra6e831252014-02-11 16:11:48 +01003312#else /* CONFIG_FAIR_GROUP_SCHED */
Vincent Guittot09a43ac2016-11-08 10:53:45 +01003313
Yuyang Du9d89c252015-07-15 08:04:37 +08003314static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
Vincent Guittot09a43ac2016-11-08 10:53:45 +01003315
3316static inline int propagate_entity_load_avg(struct sched_entity *se)
3317{
3318 return 0;
3319}
3320
3321static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3322
Peter Zijlstra6e831252014-02-11 16:11:48 +01003323#endif /* CONFIG_FAIR_GROUP_SCHED */
Paul Turnerc566e8e2012-10-04 13:18:30 +02003324
Peter Zijlstra89741892016-06-16 10:50:40 +02003325/*
3326 * Unsigned subtract and clamp on underflow.
3327 *
3328 * Explicitly do a load-store to ensure the intermediate value never hits
3329 * memory. This allows lockless observations without ever seeing the negative
3330 * values.
3331 */
3332#define sub_positive(_ptr, _val) do { \
3333 typeof(_ptr) ptr = (_ptr); \
3334 typeof(*ptr) val = (_val); \
3335 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3336 res = var - val; \
3337 if (res > var) \
3338 res = 0; \
3339 WRITE_ONCE(*ptr, res); \
3340} while (0)
3341
Peter Zijlstra3d30544f2016-06-21 14:27:50 +02003342/**
3343 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3344 * @now: current time, as per cfs_rq_clock_task()
3345 * @cfs_rq: cfs_rq to update
Peter Zijlstra3d30544f2016-06-21 14:27:50 +02003346 *
3347 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3348 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3349 * post_init_entity_util_avg().
3350 *
3351 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3352 *
Peter Zijlstra7c3edd22016-07-13 10:56:25 +02003353 * Returns true if the load decayed or we removed load.
3354 *
3355 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3356 * call update_tg_load_avg() when this function returns true.
Peter Zijlstra3d30544f2016-06-21 14:27:50 +02003357 */
Steve Mucklea2c6c912016-03-24 15:26:07 -07003358static inline int
Viresh Kumar3a123bb2017-05-24 10:59:56 +05303359update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
Steve Mucklea2c6c912016-03-24 15:26:07 -07003360{
3361 struct sched_avg *sa = &cfs_rq->avg;
3362 int decayed, removed_load = 0, removed_util = 0;
3363
3364 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3365 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
Peter Zijlstra89741892016-06-16 10:50:40 +02003366 sub_positive(&sa->load_avg, r);
3367 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
Steve Mucklea2c6c912016-03-24 15:26:07 -07003368 removed_load = 1;
Vincent Guittot4e516072016-11-08 10:53:46 +01003369 set_tg_cfs_propagate(cfs_rq);
Steve Mucklea2c6c912016-03-24 15:26:07 -07003370 }
3371
3372 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3373 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
Peter Zijlstra89741892016-06-16 10:50:40 +02003374 sub_positive(&sa->util_avg, r);
3375 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
Steve Mucklea2c6c912016-03-24 15:26:07 -07003376 removed_util = 1;
Vincent Guittot4e516072016-11-08 10:53:46 +01003377 set_tg_cfs_propagate(cfs_rq);
Steve Mucklea2c6c912016-03-24 15:26:07 -07003378 }
3379
Peter Zijlstra0ccb9772017-03-28 11:08:20 +02003380 decayed = __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
Steve Mucklea2c6c912016-03-24 15:26:07 -07003381
3382#ifndef CONFIG_64BIT
3383 smp_wmb();
3384 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3385#endif
3386
Viresh Kumar3a123bb2017-05-24 10:59:56 +05303387 if (decayed || removed_util)
Steve Mucklea2c6c912016-03-24 15:26:07 -07003388 cfs_rq_util_change(cfs_rq);
Steve Muckle21e96f82016-03-21 17:21:07 -07003389
Steve Muckle41e0d372016-03-21 17:21:08 -07003390 return decayed || removed_load;
Yuyang Du9d89c252015-07-15 08:04:37 +08003391}
3392
Vincent Guittotd31b1a62016-11-08 10:53:44 +01003393/*
3394 * Optional action to be done while updating the load average
3395 */
3396#define UPDATE_TG 0x1
3397#define SKIP_AGE_LOAD 0x2
3398
Yuyang Du9d89c252015-07-15 08:04:37 +08003399/* Update task and its cfs_rq load average */
Vincent Guittotd31b1a62016-11-08 10:53:44 +01003400static inline void update_load_avg(struct sched_entity *se, int flags)
Paul Turner9d85f212012-10-04 13:18:29 +02003401{
Paul Turner2dac7542012-10-04 13:18:30 +02003402 struct cfs_rq *cfs_rq = cfs_rq_of(se);
Yuyang Du9d89c252015-07-15 08:04:37 +08003403 u64 now = cfs_rq_clock_task(cfs_rq);
Rafael J. Wysocki34e2c552016-02-15 20:20:42 +01003404 struct rq *rq = rq_of(cfs_rq);
3405 int cpu = cpu_of(rq);
Vincent Guittot09a43ac2016-11-08 10:53:45 +01003406 int decayed;
Paul Turner2dac7542012-10-04 13:18:30 +02003407
Paul Turnerf1b17282012-10-04 13:18:31 +02003408 /*
Yuyang Du9d89c252015-07-15 08:04:37 +08003409 * Track task load average for carrying it to new CPU after migrated, and
3410 * track group sched_entity load average for task_h_load calc in migration
Paul Turnerf1b17282012-10-04 13:18:31 +02003411 */
Peter Zijlstra0ccb9772017-03-28 11:08:20 +02003412 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3413 __update_load_avg_se(now, cpu, cfs_rq, se);
Paul Turnerf1b17282012-10-04 13:18:31 +02003414
Viresh Kumar3a123bb2017-05-24 10:59:56 +05303415 decayed = update_cfs_rq_load_avg(now, cfs_rq);
Vincent Guittot09a43ac2016-11-08 10:53:45 +01003416 decayed |= propagate_entity_load_avg(se);
3417
3418 if (decayed && (flags & UPDATE_TG))
Yuyang Du9d89c252015-07-15 08:04:37 +08003419 update_tg_load_avg(cfs_rq, 0);
3420}
Paul Turner2dac7542012-10-04 13:18:30 +02003421
Peter Zijlstra3d30544f2016-06-21 14:27:50 +02003422/**
3423 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3424 * @cfs_rq: cfs_rq to attach to
3425 * @se: sched_entity to attach
3426 *
3427 * Must call update_cfs_rq_load_avg() before this, since we rely on
3428 * cfs_rq->avg.last_update_time being current.
3429 */
Byungchul Parka05e8c52015-08-20 20:21:56 +09003430static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3431{
3432 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3433 cfs_rq->avg.load_avg += se->avg.load_avg;
3434 cfs_rq->avg.load_sum += se->avg.load_sum;
3435 cfs_rq->avg.util_avg += se->avg.util_avg;
3436 cfs_rq->avg.util_sum += se->avg.util_sum;
Vincent Guittot09a43ac2016-11-08 10:53:45 +01003437 set_tg_cfs_propagate(cfs_rq);
Steve Mucklea2c6c912016-03-24 15:26:07 -07003438
3439 cfs_rq_util_change(cfs_rq);
Byungchul Parka05e8c52015-08-20 20:21:56 +09003440}
3441
Peter Zijlstra3d30544f2016-06-21 14:27:50 +02003442/**
3443 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3444 * @cfs_rq: cfs_rq to detach from
3445 * @se: sched_entity to detach
3446 *
3447 * Must call update_cfs_rq_load_avg() before this, since we rely on
3448 * cfs_rq->avg.last_update_time being current.
3449 */
Byungchul Parka05e8c52015-08-20 20:21:56 +09003450static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3451{
Byungchul Parka05e8c52015-08-20 20:21:56 +09003452
Peter Zijlstra89741892016-06-16 10:50:40 +02003453 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3454 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3455 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3456 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
Vincent Guittot09a43ac2016-11-08 10:53:45 +01003457 set_tg_cfs_propagate(cfs_rq);
Steve Mucklea2c6c912016-03-24 15:26:07 -07003458
3459 cfs_rq_util_change(cfs_rq);
Byungchul Parka05e8c52015-08-20 20:21:56 +09003460}
3461
Yuyang Du9d89c252015-07-15 08:04:37 +08003462/* Add the load generated by se into cfs_rq's load average */
3463static inline void
3464enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3465{
3466 struct sched_avg *sa = &se->avg;
Yuyang Du9d89c252015-07-15 08:04:37 +08003467
Yuyang Du13962232015-07-15 08:04:41 +08003468 cfs_rq->runnable_load_avg += sa->load_avg;
3469 cfs_rq->runnable_load_sum += sa->load_sum;
3470
Vincent Guittotd31b1a62016-11-08 10:53:44 +01003471 if (!sa->last_update_time) {
Byungchul Parka05e8c52015-08-20 20:21:56 +09003472 attach_entity_load_avg(cfs_rq, se);
Yuyang Du9d89c252015-07-15 08:04:37 +08003473 update_tg_load_avg(cfs_rq, 0);
Vincent Guittotd31b1a62016-11-08 10:53:44 +01003474 }
Paul Turner9ee474f2012-10-04 13:18:30 +02003475}
3476
Yuyang Du13962232015-07-15 08:04:41 +08003477/* Remove the runnable load generated by se from cfs_rq's runnable load average */
3478static inline void
3479dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3480{
Yuyang Du13962232015-07-15 08:04:41 +08003481 cfs_rq->runnable_load_avg =
3482 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3483 cfs_rq->runnable_load_sum =
Byungchul Parka05e8c52015-08-20 20:21:56 +09003484 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
Yuyang Du13962232015-07-15 08:04:41 +08003485}
3486
Yuyang Du0905f042015-12-17 07:34:27 +08003487#ifndef CONFIG_64BIT
3488static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3489{
3490 u64 last_update_time_copy;
3491 u64 last_update_time;
3492
3493 do {
3494 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3495 smp_rmb();
3496 last_update_time = cfs_rq->avg.last_update_time;
3497 } while (last_update_time != last_update_time_copy);
3498
3499 return last_update_time;
3500}
3501#else
3502static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3503{
3504 return cfs_rq->avg.last_update_time;
3505}
3506#endif
3507
Paul Turner9ee474f2012-10-04 13:18:30 +02003508/*
Morten Rasmussen104cb162016-10-14 14:41:07 +01003509 * Synchronize entity load avg of dequeued entity without locking
3510 * the previous rq.
3511 */
3512void sync_entity_load_avg(struct sched_entity *se)
3513{
3514 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3515 u64 last_update_time;
3516
3517 last_update_time = cfs_rq_last_update_time(cfs_rq);
Peter Zijlstra0ccb9772017-03-28 11:08:20 +02003518 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
Morten Rasmussen104cb162016-10-14 14:41:07 +01003519}
3520
3521/*
Yuyang Du9d89c252015-07-15 08:04:37 +08003522 * Task first catches up with cfs_rq, and then subtract
3523 * itself from the cfs_rq (task must be off the queue now).
Paul Turner9ee474f2012-10-04 13:18:30 +02003524 */
Yuyang Du9d89c252015-07-15 08:04:37 +08003525void remove_entity_load_avg(struct sched_entity *se)
Paul Turner9ee474f2012-10-04 13:18:30 +02003526{
Yuyang Du9d89c252015-07-15 08:04:37 +08003527 struct cfs_rq *cfs_rq = cfs_rq_of(se);
Paul Turner9ee474f2012-10-04 13:18:30 +02003528
Yuyang Du0905f042015-12-17 07:34:27 +08003529 /*
Peter Zijlstra7dc603c2016-06-16 13:29:28 +02003530 * tasks cannot exit without having gone through wake_up_new_task() ->
3531 * post_init_entity_util_avg() which will have added things to the
3532 * cfs_rq, so we can remove unconditionally.
3533 *
3534 * Similarly for groups, they will have passed through
3535 * post_init_entity_util_avg() before unregister_sched_fair_group()
3536 * calls this.
Yuyang Du0905f042015-12-17 07:34:27 +08003537 */
Paul Turner9ee474f2012-10-04 13:18:30 +02003538
Morten Rasmussen104cb162016-10-14 14:41:07 +01003539 sync_entity_load_avg(se);
Yuyang Du9d89c252015-07-15 08:04:37 +08003540 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3541 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
Paul Turner2dac7542012-10-04 13:18:30 +02003542}
Vincent Guittot642dbc32013-04-18 18:34:26 +02003543
Yuyang Du7ea241a2015-07-15 08:04:42 +08003544static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3545{
3546 return cfs_rq->runnable_load_avg;
3547}
3548
3549static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3550{
3551 return cfs_rq->avg.load_avg;
3552}
3553
Matt Fleming46f69fa2016-09-21 14:38:12 +01003554static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
Peter Zijlstra6e831252014-02-11 16:11:48 +01003555
Peter Zijlstra38033c32014-01-23 20:32:21 +01003556#else /* CONFIG_SMP */
3557
Peter Zijlstra01011472016-06-17 11:20:46 +02003558static inline int
Viresh Kumar3a123bb2017-05-24 10:59:56 +05303559update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
Peter Zijlstra01011472016-06-17 11:20:46 +02003560{
3561 return 0;
3562}
3563
Vincent Guittotd31b1a62016-11-08 10:53:44 +01003564#define UPDATE_TG 0x0
3565#define SKIP_AGE_LOAD 0x0
3566
3567static inline void update_load_avg(struct sched_entity *se, int not_used1)
Rafael J. Wysocki536bd002016-05-06 14:58:43 +02003568{
Viresh Kumara030d732017-05-24 10:59:52 +05303569 cfs_rq_util_change(cfs_rq_of(se));
Rafael J. Wysocki536bd002016-05-06 14:58:43 +02003570}
3571
Yuyang Du9d89c252015-07-15 08:04:37 +08003572static inline void
3573enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
Yuyang Du13962232015-07-15 08:04:41 +08003574static inline void
3575dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
Yuyang Du9d89c252015-07-15 08:04:37 +08003576static inline void remove_entity_load_avg(struct sched_entity *se) {}
Peter Zijlstra6e831252014-02-11 16:11:48 +01003577
Byungchul Parka05e8c52015-08-20 20:21:56 +09003578static inline void
3579attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3580static inline void
3581detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3582
Matt Fleming46f69fa2016-09-21 14:38:12 +01003583static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
Peter Zijlstra6e831252014-02-11 16:11:48 +01003584{
3585 return 0;
3586}
3587
Peter Zijlstra38033c32014-01-23 20:32:21 +01003588#endif /* CONFIG_SMP */
Paul Turner9d85f212012-10-04 13:18:29 +02003589
Peter Zijlstraddc97292007-10-15 17:00:10 +02003590static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3591{
3592#ifdef CONFIG_SCHED_DEBUG
3593 s64 d = se->vruntime - cfs_rq->min_vruntime;
3594
3595 if (d < 0)
3596 d = -d;
3597
3598 if (d > 3*sysctl_sched_latency)
Josh Poimboeufae928822016-06-17 12:43:24 -05003599 schedstat_inc(cfs_rq->nr_spread_over);
Peter Zijlstraddc97292007-10-15 17:00:10 +02003600#endif
3601}
3602
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003603static void
Peter Zijlstraaeb73b02007-10-15 17:00:05 +02003604place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3605{
Peter Zijlstra1af5f732008-10-24 11:06:13 +02003606 u64 vruntime = cfs_rq->min_vruntime;
Peter Zijlstra94dfb5e2007-10-15 17:00:05 +02003607
Peter Zijlstra2cb86002007-11-09 22:39:37 +01003608 /*
3609 * The 'current' period is already promised to the current tasks,
3610 * however the extra weight of the new task will slow them down a
3611 * little, place the new task so that it fits in the slot that
3612 * stays open at the end.
3613 */
Peter Zijlstra94dfb5e2007-10-15 17:00:05 +02003614 if (initial && sched_feat(START_DEBIT))
Peter Zijlstraf9c0b092008-10-17 19:27:04 +02003615 vruntime += sched_vslice(cfs_rq, se);
Peter Zijlstraaeb73b02007-10-15 17:00:05 +02003616
Mike Galbraitha2e7a7e2009-09-18 09:19:25 +02003617 /* sleeps up to a single latency don't count. */
Mike Galbraith5ca98802010-03-11 17:17:17 +01003618 if (!initial) {
Mike Galbraitha2e7a7e2009-09-18 09:19:25 +02003619 unsigned long thresh = sysctl_sched_latency;
Peter Zijlstraa7be37a2008-06-27 13:41:11 +02003620
Mike Galbraitha2e7a7e2009-09-18 09:19:25 +02003621 /*
Mike Galbraitha2e7a7e2009-09-18 09:19:25 +02003622 * Halve their sleep time's effect, to allow
3623 * for a gentler effect of sleepers:
3624 */
3625 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3626 thresh >>= 1;
Ingo Molnar51e03042009-09-16 08:54:45 +02003627
Mike Galbraitha2e7a7e2009-09-18 09:19:25 +02003628 vruntime -= thresh;
Peter Zijlstraaeb73b02007-10-15 17:00:05 +02003629 }
3630
Mike Galbraithb5d9d732009-09-08 11:12:28 +02003631 /* ensure we never gain time by being placed backwards. */
Viresh Kumar16c8f1c2012-11-08 13:33:46 +05303632 se->vruntime = max_vruntime(se->vruntime, vruntime);
Peter Zijlstraaeb73b02007-10-15 17:00:05 +02003633}
3634
Paul Turnerd3d9dc32011-07-21 09:43:39 -07003635static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3636
Mel Gormancb251762016-02-05 09:08:36 +00003637static inline void check_schedstat_required(void)
3638{
3639#ifdef CONFIG_SCHEDSTATS
3640 if (schedstat_enabled())
3641 return;
3642
3643 /* Force schedstat enabled if a dependent tracepoint is active */
3644 if (trace_sched_stat_wait_enabled() ||
3645 trace_sched_stat_sleep_enabled() ||
3646 trace_sched_stat_iowait_enabled() ||
3647 trace_sched_stat_blocked_enabled() ||
3648 trace_sched_stat_runtime_enabled()) {
Josh Poimboeufeda8dca2016-06-13 02:32:09 -05003649 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
Mel Gormancb251762016-02-05 09:08:36 +00003650 "stat_blocked and stat_runtime require the "
Marcin Nowakowskif67abed2017-06-09 10:00:29 +02003651 "kernel parameter schedstats=enable or "
Mel Gormancb251762016-02-05 09:08:36 +00003652 "kernel.sched_schedstats=1\n");
3653 }
3654#endif
3655}
3656
Peter Zijlstrab5179ac2016-05-11 16:10:34 +02003657
3658/*
3659 * MIGRATION
3660 *
3661 * dequeue
3662 * update_curr()
3663 * update_min_vruntime()
3664 * vruntime -= min_vruntime
3665 *
3666 * enqueue
3667 * update_curr()
3668 * update_min_vruntime()
3669 * vruntime += min_vruntime
3670 *
3671 * this way the vruntime transition between RQs is done when both
3672 * min_vruntime are up-to-date.
3673 *
3674 * WAKEUP (remote)
3675 *
Peter Zijlstra59efa0b2016-05-10 18:24:37 +02003676 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
Peter Zijlstrab5179ac2016-05-11 16:10:34 +02003677 * vruntime -= min_vruntime
3678 *
3679 * enqueue
3680 * update_curr()
3681 * update_min_vruntime()
3682 * vruntime += min_vruntime
3683 *
3684 * this way we don't have the most up-to-date min_vruntime on the originating
3685 * CPU and an up-to-date min_vruntime on the destination CPU.
3686 */
3687
Peter Zijlstraaeb73b02007-10-15 17:00:05 +02003688static void
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01003689enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003690{
Peter Zijlstra2f950352016-05-11 19:27:56 +02003691 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3692 bool curr = cfs_rq->curr == se;
Peter Zijlstra3a47d512016-03-09 13:04:03 +01003693
Ingo Molnar53d3bc72016-05-11 08:25:53 +02003694 /*
Peter Zijlstra2f950352016-05-11 19:27:56 +02003695 * If we're the current task, we must renormalise before calling
3696 * update_curr().
Ingo Molnar53d3bc72016-05-11 08:25:53 +02003697 */
Peter Zijlstra2f950352016-05-11 19:27:56 +02003698 if (renorm && curr)
3699 se->vruntime += cfs_rq->min_vruntime;
3700
Ingo Molnarb7cc0892007-08-09 11:16:47 +02003701 update_curr(cfs_rq);
Peter Zijlstra2f950352016-05-11 19:27:56 +02003702
3703 /*
3704 * Otherwise, renormalise after, such that we're placed at the current
3705 * moment in time, instead of some random moment in the past. Being
3706 * placed in the past could significantly boost this task to the
3707 * fairness detriment of existing tasks.
3708 */
3709 if (renorm && !curr)
3710 se->vruntime += cfs_rq->min_vruntime;
3711
Vincent Guittot89ee0482016-12-21 16:50:26 +01003712 /*
3713 * When enqueuing a sched_entity, we must:
3714 * - Update loads to have both entity and cfs_rq synced with now.
3715 * - Add its load to cfs_rq->runnable_avg
3716 * - For group_entity, update its weight to reflect the new share of
3717 * its group cfs_rq
3718 * - Add its new weight to cfs_rq->load.weight
3719 */
Vincent Guittotd31b1a62016-11-08 10:53:44 +01003720 update_load_avg(se, UPDATE_TG);
Yuyang Du9d89c252015-07-15 08:04:37 +08003721 enqueue_entity_load_avg(cfs_rq, se);
Vincent Guittot89ee0482016-12-21 16:50:26 +01003722 update_cfs_shares(se);
Linus Torvalds17bc14b2012-12-14 07:20:43 -08003723 account_entity_enqueue(cfs_rq, se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003724
Josh Poimboeuf1a3d0272016-06-17 12:43:23 -05003725 if (flags & ENQUEUE_WAKEUP)
Peter Zijlstraaeb73b02007-10-15 17:00:05 +02003726 place_entity(cfs_rq, se, 0);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003727
Mel Gormancb251762016-02-05 09:08:36 +00003728 check_schedstat_required();
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003729 update_stats_enqueue(cfs_rq, se, flags);
3730 check_spread(cfs_rq, se);
Peter Zijlstra2f950352016-05-11 19:27:56 +02003731 if (!curr)
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02003732 __enqueue_entity(cfs_rq, se);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08003733 se->on_rq = 1;
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -08003734
Paul Turnerd3d9dc32011-07-21 09:43:39 -07003735 if (cfs_rq->nr_running == 1) {
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -08003736 list_add_leaf_cfs_rq(cfs_rq);
Paul Turnerd3d9dc32011-07-21 09:43:39 -07003737 check_enqueue_throttle(cfs_rq);
3738 }
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003739}
3740
Rik van Riel2c13c9192011-02-01 09:48:37 -05003741static void __clear_buddies_last(struct sched_entity *se)
Peter Zijlstra2002c692008-11-11 11:52:33 +01003742{
Rik van Riel2c13c9192011-02-01 09:48:37 -05003743 for_each_sched_entity(se) {
3744 struct cfs_rq *cfs_rq = cfs_rq_of(se);
Peter Zijlstraf1044792012-02-11 06:05:00 +01003745 if (cfs_rq->last != se)
Rik van Riel2c13c9192011-02-01 09:48:37 -05003746 break;
Peter Zijlstraf1044792012-02-11 06:05:00 +01003747
3748 cfs_rq->last = NULL;
Rik van Riel2c13c9192011-02-01 09:48:37 -05003749 }
3750}
Peter Zijlstra2002c692008-11-11 11:52:33 +01003751
Rik van Riel2c13c9192011-02-01 09:48:37 -05003752static void __clear_buddies_next(struct sched_entity *se)
3753{
3754 for_each_sched_entity(se) {
3755 struct cfs_rq *cfs_rq = cfs_rq_of(se);
Peter Zijlstraf1044792012-02-11 06:05:00 +01003756 if (cfs_rq->next != se)
Rik van Riel2c13c9192011-02-01 09:48:37 -05003757 break;
Peter Zijlstraf1044792012-02-11 06:05:00 +01003758
3759 cfs_rq->next = NULL;
Rik van Riel2c13c9192011-02-01 09:48:37 -05003760 }
Peter Zijlstra2002c692008-11-11 11:52:33 +01003761}
3762
Rik van Rielac53db52011-02-01 09:51:03 -05003763static void __clear_buddies_skip(struct sched_entity *se)
3764{
3765 for_each_sched_entity(se) {
3766 struct cfs_rq *cfs_rq = cfs_rq_of(se);
Peter Zijlstraf1044792012-02-11 06:05:00 +01003767 if (cfs_rq->skip != se)
Rik van Rielac53db52011-02-01 09:51:03 -05003768 break;
Peter Zijlstraf1044792012-02-11 06:05:00 +01003769
3770 cfs_rq->skip = NULL;
Rik van Rielac53db52011-02-01 09:51:03 -05003771 }
3772}
3773
Peter Zijlstraa571bbe2009-01-28 14:51:40 +01003774static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3775{
Rik van Riel2c13c9192011-02-01 09:48:37 -05003776 if (cfs_rq->last == se)
3777 __clear_buddies_last(se);
3778
3779 if (cfs_rq->next == se)
3780 __clear_buddies_next(se);
Rik van Rielac53db52011-02-01 09:51:03 -05003781
3782 if (cfs_rq->skip == se)
3783 __clear_buddies_skip(se);
Peter Zijlstraa571bbe2009-01-28 14:51:40 +01003784}
3785
Peter Zijlstra6c16a6d2012-03-21 13:07:16 -07003786static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
Paul Turnerd8b49862011-07-21 09:43:41 -07003787
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003788static void
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01003789dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003790{
Dmitry Adamushkoa2a2d682007-10-15 17:00:13 +02003791 /*
3792 * Update run-time statistics of the 'current'.
3793 */
3794 update_curr(cfs_rq);
Vincent Guittot89ee0482016-12-21 16:50:26 +01003795
3796 /*
3797 * When dequeuing a sched_entity, we must:
3798 * - Update loads to have both entity and cfs_rq synced with now.
3799 * - Substract its load from the cfs_rq->runnable_avg.
3800 * - Substract its previous weight from cfs_rq->load.weight.
3801 * - For group entity, update its weight to reflect the new share
3802 * of its group cfs_rq.
3803 */
Vincent Guittotd31b1a62016-11-08 10:53:44 +01003804 update_load_avg(se, UPDATE_TG);
Yuyang Du13962232015-07-15 08:04:41 +08003805 dequeue_entity_load_avg(cfs_rq, se);
Dmitry Adamushkoa2a2d682007-10-15 17:00:13 +02003806
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003807 update_stats_dequeue(cfs_rq, se, flags);
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02003808
Peter Zijlstra2002c692008-11-11 11:52:33 +01003809 clear_buddies(cfs_rq, se);
Peter Zijlstra47932412008-11-04 21:25:09 +01003810
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02003811 if (se != cfs_rq->curr)
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02003812 __dequeue_entity(cfs_rq, se);
Linus Torvalds17bc14b2012-12-14 07:20:43 -08003813 se->on_rq = 0;
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02003814 account_entity_dequeue(cfs_rq, se);
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01003815
3816 /*
Peter Zijlstrab60205c2016-09-20 21:58:12 +02003817 * Normalize after update_curr(); which will also have moved
3818 * min_vruntime if @se is the one holding it back. But before doing
3819 * update_min_vruntime() again, which will discount @se's position and
3820 * can move min_vruntime forward still more.
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01003821 */
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01003822 if (!(flags & DEQUEUE_SLEEP))
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01003823 se->vruntime -= cfs_rq->min_vruntime;
Peter Zijlstra1e876232011-05-17 16:21:10 -07003824
Paul Turnerd8b49862011-07-21 09:43:41 -07003825 /* return excess runtime on last dequeue */
3826 return_cfs_rq_runtime(cfs_rq);
3827
Vincent Guittot89ee0482016-12-21 16:50:26 +01003828 update_cfs_shares(se);
Peter Zijlstrab60205c2016-09-20 21:58:12 +02003829
3830 /*
3831 * Now advance min_vruntime if @se was the entity holding it back,
3832 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3833 * put back on, and if we advance min_vruntime, we'll be placed back
3834 * further than we started -- ie. we'll be penalized.
3835 */
3836 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3837 update_min_vruntime(cfs_rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003838}
3839
3840/*
3841 * Preempt the current task with a newly woken task if needed:
3842 */
Peter Zijlstra7c92e542007-09-05 14:32:49 +02003843static void
Ingo Molnar2e09bf52007-10-15 17:00:05 +02003844check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003845{
Peter Zijlstra11697832007-09-05 14:32:49 +02003846 unsigned long ideal_runtime, delta_exec;
Wang Xingchaof4cfb332011-09-16 13:35:52 -04003847 struct sched_entity *se;
3848 s64 delta;
Peter Zijlstra11697832007-09-05 14:32:49 +02003849
Peter Zijlstra6d0f0eb2007-10-15 17:00:05 +02003850 ideal_runtime = sched_slice(cfs_rq, curr);
Peter Zijlstra11697832007-09-05 14:32:49 +02003851 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
Mike Galbraitha9f3e2b2009-01-28 14:51:39 +01003852 if (delta_exec > ideal_runtime) {
Kirill Tkhai88751252014-06-29 00:03:57 +04003853 resched_curr(rq_of(cfs_rq));
Mike Galbraitha9f3e2b2009-01-28 14:51:39 +01003854 /*
3855 * The current task ran long enough, ensure it doesn't get
3856 * re-elected due to buddy favours.
3857 */
3858 clear_buddies(cfs_rq, curr);
Mike Galbraithf685cea2009-10-23 23:09:22 +02003859 return;
3860 }
3861
3862 /*
3863 * Ensure that a task that missed wakeup preemption by a
3864 * narrow margin doesn't have to wait for a full slice.
3865 * This also mitigates buddy induced latencies under load.
3866 */
Mike Galbraithf685cea2009-10-23 23:09:22 +02003867 if (delta_exec < sysctl_sched_min_granularity)
3868 return;
3869
Wang Xingchaof4cfb332011-09-16 13:35:52 -04003870 se = __pick_first_entity(cfs_rq);
3871 delta = curr->vruntime - se->vruntime;
Mike Galbraithf685cea2009-10-23 23:09:22 +02003872
Wang Xingchaof4cfb332011-09-16 13:35:52 -04003873 if (delta < 0)
3874 return;
Mike Galbraithd7d82942011-01-05 05:41:17 +01003875
Wang Xingchaof4cfb332011-09-16 13:35:52 -04003876 if (delta > ideal_runtime)
Kirill Tkhai88751252014-06-29 00:03:57 +04003877 resched_curr(rq_of(cfs_rq));
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003878}
3879
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02003880static void
Ingo Molnar8494f412007-08-09 11:16:48 +02003881set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003882{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02003883 /* 'current' is not kept within the tree. */
3884 if (se->on_rq) {
3885 /*
3886 * Any task has to be enqueued before it get to execute on
3887 * a CPU. So account for the time it spent waiting on the
3888 * runqueue.
3889 */
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003890 update_stats_wait_end(cfs_rq, se);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02003891 __dequeue_entity(cfs_rq, se);
Vincent Guittotd31b1a62016-11-08 10:53:44 +01003892 update_load_avg(se, UPDATE_TG);
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02003893 }
3894
Ingo Molnar79303e92007-08-09 11:16:47 +02003895 update_stats_curr_start(cfs_rq, se);
Ingo Molnar429d43b2007-10-15 17:00:03 +02003896 cfs_rq->curr = se;
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003897
Ingo Molnareba1ed42007-10-15 17:00:02 +02003898 /*
3899 * Track our maximum slice length, if the CPU's load is at
3900 * least twice that of our own weight (i.e. dont track it
3901 * when there are only lesser-weight tasks around):
3902 */
Mel Gormancb251762016-02-05 09:08:36 +00003903 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003904 schedstat_set(se->statistics.slice_max,
3905 max((u64)schedstat_val(se->statistics.slice_max),
3906 se->sum_exec_runtime - se->prev_sum_exec_runtime));
Ingo Molnareba1ed42007-10-15 17:00:02 +02003907 }
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003908
Peter Zijlstra4a55b452007-09-05 14:32:49 +02003909 se->prev_sum_exec_runtime = se->sum_exec_runtime;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003910}
3911
Peter Zijlstra3f3a4902008-10-24 11:06:16 +02003912static int
3913wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3914
Rik van Rielac53db52011-02-01 09:51:03 -05003915/*
3916 * Pick the next process, keeping these things in mind, in this order:
3917 * 1) keep things fair between processes/task groups
3918 * 2) pick the "next" process, since someone really wants that to run
3919 * 3) pick the "last" process, for cache locality
3920 * 4) do not run the "skip" process, if something else is available
3921 */
Peter Zijlstra678d5712012-02-11 06:05:00 +01003922static struct sched_entity *
3923pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
Peter Zijlstraaa2ac252008-03-14 21:12:12 +01003924{
Peter Zijlstra678d5712012-02-11 06:05:00 +01003925 struct sched_entity *left = __pick_first_entity(cfs_rq);
3926 struct sched_entity *se;
3927
3928 /*
3929 * If curr is set we have to see if its left of the leftmost entity
3930 * still in the tree, provided there was anything in the tree at all.
3931 */
3932 if (!left || (curr && entity_before(curr, left)))
3933 left = curr;
3934
3935 se = left; /* ideally we run the leftmost entity */
Peter Zijlstraf4b67552008-11-04 21:25:07 +01003936
Rik van Rielac53db52011-02-01 09:51:03 -05003937 /*
3938 * Avoid running the skip buddy, if running something else can
3939 * be done without getting too unfair.
3940 */
3941 if (cfs_rq->skip == se) {
Peter Zijlstra678d5712012-02-11 06:05:00 +01003942 struct sched_entity *second;
3943
3944 if (se == curr) {
3945 second = __pick_first_entity(cfs_rq);
3946 } else {
3947 second = __pick_next_entity(se);
3948 if (!second || (curr && entity_before(curr, second)))
3949 second = curr;
3950 }
3951
Rik van Rielac53db52011-02-01 09:51:03 -05003952 if (second && wakeup_preempt_entity(second, left) < 1)
3953 se = second;
3954 }
Peter Zijlstraaa2ac252008-03-14 21:12:12 +01003955
Mike Galbraithf685cea2009-10-23 23:09:22 +02003956 /*
3957 * Prefer last buddy, try to return the CPU to a preempted task.
3958 */
3959 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3960 se = cfs_rq->last;
3961
Rik van Rielac53db52011-02-01 09:51:03 -05003962 /*
3963 * Someone really wants this to run. If it's not unfair, run it.
3964 */
3965 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3966 se = cfs_rq->next;
3967
Mike Galbraithf685cea2009-10-23 23:09:22 +02003968 clear_buddies(cfs_rq, se);
Peter Zijlstra47932412008-11-04 21:25:09 +01003969
3970 return se;
Peter Zijlstraaa2ac252008-03-14 21:12:12 +01003971}
3972
Peter Zijlstra678d5712012-02-11 06:05:00 +01003973static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
Paul Turnerd3d9dc32011-07-21 09:43:39 -07003974
Ingo Molnarab6cde22007-08-09 11:16:48 +02003975static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003976{
3977 /*
3978 * If still on the runqueue then deactivate_task()
3979 * was not called and update_curr() has to be done:
3980 */
3981 if (prev->on_rq)
Ingo Molnarb7cc0892007-08-09 11:16:47 +02003982 update_curr(cfs_rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003983
Paul Turnerd3d9dc32011-07-21 09:43:39 -07003984 /* throttle cfs_rqs exceeding runtime */
3985 check_cfs_rq_runtime(cfs_rq);
3986
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003987 check_spread(cfs_rq, prev);
Mel Gormancb251762016-02-05 09:08:36 +00003988
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02003989 if (prev->on_rq) {
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003990 update_stats_wait_start(cfs_rq, prev);
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02003991 /* Put 'current' back into the tree. */
3992 __enqueue_entity(cfs_rq, prev);
Paul Turner9d85f212012-10-04 13:18:29 +02003993 /* in !on_rq case, update occurred at dequeue */
Yuyang Du9d89c252015-07-15 08:04:37 +08003994 update_load_avg(prev, 0);
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02003995 }
Ingo Molnar429d43b2007-10-15 17:00:03 +02003996 cfs_rq->curr = NULL;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02003997}
3998
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01003999static void
4000entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004001{
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004002 /*
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02004003 * Update run-time statistics of the 'current'.
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004004 */
Dmitry Adamushko30cfdcf2007-10-15 17:00:07 +02004005 update_curr(cfs_rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004006
Paul Turner43365bd2010-12-15 19:10:17 -08004007 /*
Paul Turner9d85f212012-10-04 13:18:29 +02004008 * Ensure that runnable average is periodically updated.
4009 */
Vincent Guittotd31b1a62016-11-08 10:53:44 +01004010 update_load_avg(curr, UPDATE_TG);
Vincent Guittot89ee0482016-12-21 16:50:26 +01004011 update_cfs_shares(curr);
Paul Turner9d85f212012-10-04 13:18:29 +02004012
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004013#ifdef CONFIG_SCHED_HRTICK
4014 /*
4015 * queued ticks are scheduled to match the slice, so don't bother
4016 * validating it and just reschedule.
4017 */
Harvey Harrison983ed7a2008-04-24 18:17:55 -07004018 if (queued) {
Kirill Tkhai88751252014-06-29 00:03:57 +04004019 resched_curr(rq_of(cfs_rq));
Harvey Harrison983ed7a2008-04-24 18:17:55 -07004020 return;
4021 }
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004022 /*
4023 * don't let the period tick interfere with the hrtick preemption
4024 */
4025 if (!sched_feat(DOUBLE_TICK) &&
4026 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4027 return;
4028#endif
4029
Yong Zhang2c2efae2011-07-29 16:20:33 +08004030 if (cfs_rq->nr_running > 1)
Ingo Molnar2e09bf52007-10-15 17:00:05 +02004031 check_preempt_tick(cfs_rq, curr);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004032}
4033
Paul Turnerab84d312011-07-21 09:43:28 -07004034
4035/**************************************************
4036 * CFS bandwidth control machinery
4037 */
4038
4039#ifdef CONFIG_CFS_BANDWIDTH
Peter Zijlstra029632f2011-10-25 10:00:11 +02004040
4041#ifdef HAVE_JUMP_LABEL
Ingo Molnarc5905af2012-02-24 08:31:31 +01004042static struct static_key __cfs_bandwidth_used;
Peter Zijlstra029632f2011-10-25 10:00:11 +02004043
4044static inline bool cfs_bandwidth_used(void)
4045{
Ingo Molnarc5905af2012-02-24 08:31:31 +01004046 return static_key_false(&__cfs_bandwidth_used);
Peter Zijlstra029632f2011-10-25 10:00:11 +02004047}
4048
Ben Segall1ee14e62013-10-16 11:16:12 -07004049void cfs_bandwidth_usage_inc(void)
Peter Zijlstra029632f2011-10-25 10:00:11 +02004050{
Ben Segall1ee14e62013-10-16 11:16:12 -07004051 static_key_slow_inc(&__cfs_bandwidth_used);
4052}
4053
4054void cfs_bandwidth_usage_dec(void)
4055{
4056 static_key_slow_dec(&__cfs_bandwidth_used);
Peter Zijlstra029632f2011-10-25 10:00:11 +02004057}
4058#else /* HAVE_JUMP_LABEL */
4059static bool cfs_bandwidth_used(void)
4060{
4061 return true;
4062}
4063
Ben Segall1ee14e62013-10-16 11:16:12 -07004064void cfs_bandwidth_usage_inc(void) {}
4065void cfs_bandwidth_usage_dec(void) {}
Peter Zijlstra029632f2011-10-25 10:00:11 +02004066#endif /* HAVE_JUMP_LABEL */
4067
Paul Turnerab84d312011-07-21 09:43:28 -07004068/*
4069 * default period for cfs group bandwidth.
4070 * default: 0.1s, units: nanoseconds
4071 */
4072static inline u64 default_cfs_period(void)
4073{
4074 return 100000000ULL;
4075}
Paul Turnerec12cb72011-07-21 09:43:30 -07004076
4077static inline u64 sched_cfs_bandwidth_slice(void)
4078{
4079 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4080}
4081
Paul Turnera9cf55b2011-07-21 09:43:32 -07004082/*
4083 * Replenish runtime according to assigned quota and update expiration time.
4084 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4085 * additional synchronization around rq->lock.
4086 *
4087 * requires cfs_b->lock
4088 */
Peter Zijlstra029632f2011-10-25 10:00:11 +02004089void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
Paul Turnera9cf55b2011-07-21 09:43:32 -07004090{
4091 u64 now;
4092
4093 if (cfs_b->quota == RUNTIME_INF)
4094 return;
4095
4096 now = sched_clock_cpu(smp_processor_id());
4097 cfs_b->runtime = cfs_b->quota;
4098 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4099}
4100
Peter Zijlstra029632f2011-10-25 10:00:11 +02004101static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4102{
4103 return &tg->cfs_bandwidth;
4104}
4105
Paul Turnerf1b17282012-10-04 13:18:31 +02004106/* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4107static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4108{
4109 if (unlikely(cfs_rq->throttle_count))
Xunlei Pang1a99ae32016-05-10 21:03:18 +08004110 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
Paul Turnerf1b17282012-10-04 13:18:31 +02004111
Frederic Weisbecker78becc22013-04-12 01:51:02 +02004112 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
Paul Turnerf1b17282012-10-04 13:18:31 +02004113}
4114
Paul Turner85dac902011-07-21 09:43:33 -07004115/* returns 0 on failure to allocate runtime */
4116static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
Paul Turnerec12cb72011-07-21 09:43:30 -07004117{
4118 struct task_group *tg = cfs_rq->tg;
4119 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
Paul Turnera9cf55b2011-07-21 09:43:32 -07004120 u64 amount = 0, min_amount, expires;
Paul Turnerec12cb72011-07-21 09:43:30 -07004121
4122 /* note: this is a positive sum as runtime_remaining <= 0 */
4123 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4124
4125 raw_spin_lock(&cfs_b->lock);
4126 if (cfs_b->quota == RUNTIME_INF)
4127 amount = min_amount;
Paul Turner58088ad2011-07-21 09:43:31 -07004128 else {
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +02004129 start_cfs_bandwidth(cfs_b);
Paul Turner58088ad2011-07-21 09:43:31 -07004130
4131 if (cfs_b->runtime > 0) {
4132 amount = min(cfs_b->runtime, min_amount);
4133 cfs_b->runtime -= amount;
4134 cfs_b->idle = 0;
4135 }
Paul Turnerec12cb72011-07-21 09:43:30 -07004136 }
Paul Turnera9cf55b2011-07-21 09:43:32 -07004137 expires = cfs_b->runtime_expires;
Paul Turnerec12cb72011-07-21 09:43:30 -07004138 raw_spin_unlock(&cfs_b->lock);
4139
4140 cfs_rq->runtime_remaining += amount;
Paul Turnera9cf55b2011-07-21 09:43:32 -07004141 /*
4142 * we may have advanced our local expiration to account for allowed
4143 * spread between our sched_clock and the one on which runtime was
4144 * issued.
4145 */
4146 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4147 cfs_rq->runtime_expires = expires;
Paul Turner85dac902011-07-21 09:43:33 -07004148
4149 return cfs_rq->runtime_remaining > 0;
Paul Turnera9cf55b2011-07-21 09:43:32 -07004150}
4151
4152/*
4153 * Note: This depends on the synchronization provided by sched_clock and the
4154 * fact that rq->clock snapshots this value.
4155 */
4156static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4157{
4158 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
Paul Turnera9cf55b2011-07-21 09:43:32 -07004159
4160 /* if the deadline is ahead of our clock, nothing to do */
Frederic Weisbecker78becc22013-04-12 01:51:02 +02004161 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
Paul Turnera9cf55b2011-07-21 09:43:32 -07004162 return;
4163
4164 if (cfs_rq->runtime_remaining < 0)
4165 return;
4166
4167 /*
4168 * If the local deadline has passed we have to consider the
4169 * possibility that our sched_clock is 'fast' and the global deadline
4170 * has not truly expired.
4171 *
4172 * Fortunately we can check determine whether this the case by checking
Ben Segall51f21762014-05-19 15:49:45 -07004173 * whether the global deadline has advanced. It is valid to compare
4174 * cfs_b->runtime_expires without any locks since we only care about
4175 * exact equality, so a partial write will still work.
Paul Turnera9cf55b2011-07-21 09:43:32 -07004176 */
4177
Ben Segall51f21762014-05-19 15:49:45 -07004178 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
Paul Turnera9cf55b2011-07-21 09:43:32 -07004179 /* extend local deadline, drift is bounded above by 2 ticks */
4180 cfs_rq->runtime_expires += TICK_NSEC;
4181 } else {
4182 /* global deadline is ahead, expiration has passed */
4183 cfs_rq->runtime_remaining = 0;
4184 }
Paul Turnerec12cb72011-07-21 09:43:30 -07004185}
4186
Peter Zijlstra9dbdb152013-11-18 18:27:06 +01004187static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
Paul Turnerec12cb72011-07-21 09:43:30 -07004188{
Paul Turnera9cf55b2011-07-21 09:43:32 -07004189 /* dock delta_exec before expiring quota (as it could span periods) */
Paul Turnerec12cb72011-07-21 09:43:30 -07004190 cfs_rq->runtime_remaining -= delta_exec;
Paul Turnera9cf55b2011-07-21 09:43:32 -07004191 expire_cfs_rq_runtime(cfs_rq);
4192
4193 if (likely(cfs_rq->runtime_remaining > 0))
Paul Turnerec12cb72011-07-21 09:43:30 -07004194 return;
4195
Paul Turner85dac902011-07-21 09:43:33 -07004196 /*
4197 * if we're unable to extend our runtime we resched so that the active
4198 * hierarchy can be throttled
4199 */
4200 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
Kirill Tkhai88751252014-06-29 00:03:57 +04004201 resched_curr(rq_of(cfs_rq));
Paul Turnerec12cb72011-07-21 09:43:30 -07004202}
4203
Peter Zijlstra6c16a6d2012-03-21 13:07:16 -07004204static __always_inline
Peter Zijlstra9dbdb152013-11-18 18:27:06 +01004205void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
Paul Turnerec12cb72011-07-21 09:43:30 -07004206{
Paul Turner56f570e2011-11-07 20:26:33 -08004207 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
Paul Turnerec12cb72011-07-21 09:43:30 -07004208 return;
4209
4210 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4211}
4212
Paul Turner85dac902011-07-21 09:43:33 -07004213static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4214{
Paul Turner56f570e2011-11-07 20:26:33 -08004215 return cfs_bandwidth_used() && cfs_rq->throttled;
Paul Turner85dac902011-07-21 09:43:33 -07004216}
4217
Paul Turner64660c82011-07-21 09:43:36 -07004218/* check whether cfs_rq, or any parent, is throttled */
4219static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4220{
Paul Turner56f570e2011-11-07 20:26:33 -08004221 return cfs_bandwidth_used() && cfs_rq->throttle_count;
Paul Turner64660c82011-07-21 09:43:36 -07004222}
4223
4224/*
4225 * Ensure that neither of the group entities corresponding to src_cpu or
4226 * dest_cpu are members of a throttled hierarchy when performing group
4227 * load-balance operations.
4228 */
4229static inline int throttled_lb_pair(struct task_group *tg,
4230 int src_cpu, int dest_cpu)
4231{
4232 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4233
4234 src_cfs_rq = tg->cfs_rq[src_cpu];
4235 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4236
4237 return throttled_hierarchy(src_cfs_rq) ||
4238 throttled_hierarchy(dest_cfs_rq);
4239}
4240
4241/* updated child weight may affect parent so we have to do this bottom up */
4242static int tg_unthrottle_up(struct task_group *tg, void *data)
4243{
4244 struct rq *rq = data;
4245 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4246
4247 cfs_rq->throttle_count--;
Paul Turner64660c82011-07-21 09:43:36 -07004248 if (!cfs_rq->throttle_count) {
Paul Turnerf1b17282012-10-04 13:18:31 +02004249 /* adjust cfs_rq_clock_task() */
Frederic Weisbecker78becc22013-04-12 01:51:02 +02004250 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
Paul Turnerf1b17282012-10-04 13:18:31 +02004251 cfs_rq->throttled_clock_task;
Paul Turner64660c82011-07-21 09:43:36 -07004252 }
Paul Turner64660c82011-07-21 09:43:36 -07004253
4254 return 0;
4255}
4256
4257static int tg_throttle_down(struct task_group *tg, void *data)
4258{
4259 struct rq *rq = data;
4260 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4261
Paul Turner82958362012-10-04 13:18:31 +02004262 /* group is entering throttled state, stop time */
4263 if (!cfs_rq->throttle_count)
Frederic Weisbecker78becc22013-04-12 01:51:02 +02004264 cfs_rq->throttled_clock_task = rq_clock_task(rq);
Paul Turner64660c82011-07-21 09:43:36 -07004265 cfs_rq->throttle_count++;
4266
4267 return 0;
4268}
4269
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004270static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
Paul Turner85dac902011-07-21 09:43:33 -07004271{
4272 struct rq *rq = rq_of(cfs_rq);
4273 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4274 struct sched_entity *se;
4275 long task_delta, dequeue = 1;
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +02004276 bool empty;
Paul Turner85dac902011-07-21 09:43:33 -07004277
4278 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4279
Paul Turnerf1b17282012-10-04 13:18:31 +02004280 /* freeze hierarchy runnable averages while throttled */
Paul Turner64660c82011-07-21 09:43:36 -07004281 rcu_read_lock();
4282 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4283 rcu_read_unlock();
Paul Turner85dac902011-07-21 09:43:33 -07004284
4285 task_delta = cfs_rq->h_nr_running;
4286 for_each_sched_entity(se) {
4287 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4288 /* throttled entity or throttle-on-deactivate */
4289 if (!se->on_rq)
4290 break;
4291
4292 if (dequeue)
4293 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4294 qcfs_rq->h_nr_running -= task_delta;
4295
4296 if (qcfs_rq->load.weight)
4297 dequeue = 0;
4298 }
4299
4300 if (!se)
Kirill Tkhai72465442014-05-09 03:00:14 +04004301 sub_nr_running(rq, task_delta);
Paul Turner85dac902011-07-21 09:43:33 -07004302
4303 cfs_rq->throttled = 1;
Frederic Weisbecker78becc22013-04-12 01:51:02 +02004304 cfs_rq->throttled_clock = rq_clock(rq);
Paul Turner85dac902011-07-21 09:43:33 -07004305 raw_spin_lock(&cfs_b->lock);
Cong Wangd49db342015-06-24 12:41:47 -07004306 empty = list_empty(&cfs_b->throttled_cfs_rq);
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +02004307
Ben Segallc06f04c2014-06-20 15:21:20 -07004308 /*
4309 * Add to the _head_ of the list, so that an already-started
4310 * distribute_cfs_runtime will not see us
4311 */
4312 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +02004313
4314 /*
4315 * If we're the first throttled task, make sure the bandwidth
4316 * timer is running.
4317 */
4318 if (empty)
4319 start_cfs_bandwidth(cfs_b);
4320
Paul Turner85dac902011-07-21 09:43:33 -07004321 raw_spin_unlock(&cfs_b->lock);
4322}
4323
Peter Zijlstra029632f2011-10-25 10:00:11 +02004324void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
Paul Turner671fd9d2011-07-21 09:43:34 -07004325{
4326 struct rq *rq = rq_of(cfs_rq);
4327 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4328 struct sched_entity *se;
4329 int enqueue = 1;
4330 long task_delta;
4331
Michael Wang22b958d2013-06-04 14:23:39 +08004332 se = cfs_rq->tg->se[cpu_of(rq)];
Paul Turner671fd9d2011-07-21 09:43:34 -07004333
4334 cfs_rq->throttled = 0;
Frederic Weisbecker1a55af22013-04-12 01:51:01 +02004335
4336 update_rq_clock(rq);
4337
Paul Turner671fd9d2011-07-21 09:43:34 -07004338 raw_spin_lock(&cfs_b->lock);
Frederic Weisbecker78becc22013-04-12 01:51:02 +02004339 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
Paul Turner671fd9d2011-07-21 09:43:34 -07004340 list_del_rcu(&cfs_rq->throttled_list);
4341 raw_spin_unlock(&cfs_b->lock);
4342
Paul Turner64660c82011-07-21 09:43:36 -07004343 /* update hierarchical throttle state */
4344 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4345
Paul Turner671fd9d2011-07-21 09:43:34 -07004346 if (!cfs_rq->load.weight)
4347 return;
4348
4349 task_delta = cfs_rq->h_nr_running;
4350 for_each_sched_entity(se) {
4351 if (se->on_rq)
4352 enqueue = 0;
4353
4354 cfs_rq = cfs_rq_of(se);
4355 if (enqueue)
4356 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4357 cfs_rq->h_nr_running += task_delta;
4358
4359 if (cfs_rq_throttled(cfs_rq))
4360 break;
4361 }
4362
4363 if (!se)
Kirill Tkhai72465442014-05-09 03:00:14 +04004364 add_nr_running(rq, task_delta);
Paul Turner671fd9d2011-07-21 09:43:34 -07004365
4366 /* determine whether we need to wake up potentially idle cpu */
4367 if (rq->curr == rq->idle && rq->cfs.nr_running)
Kirill Tkhai88751252014-06-29 00:03:57 +04004368 resched_curr(rq);
Paul Turner671fd9d2011-07-21 09:43:34 -07004369}
4370
4371static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4372 u64 remaining, u64 expires)
4373{
4374 struct cfs_rq *cfs_rq;
Ben Segallc06f04c2014-06-20 15:21:20 -07004375 u64 runtime;
4376 u64 starting_runtime = remaining;
Paul Turner671fd9d2011-07-21 09:43:34 -07004377
4378 rcu_read_lock();
4379 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4380 throttled_list) {
4381 struct rq *rq = rq_of(cfs_rq);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02004382 struct rq_flags rf;
Paul Turner671fd9d2011-07-21 09:43:34 -07004383
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02004384 rq_lock(rq, &rf);
Paul Turner671fd9d2011-07-21 09:43:34 -07004385 if (!cfs_rq_throttled(cfs_rq))
4386 goto next;
4387
4388 runtime = -cfs_rq->runtime_remaining + 1;
4389 if (runtime > remaining)
4390 runtime = remaining;
4391 remaining -= runtime;
4392
4393 cfs_rq->runtime_remaining += runtime;
4394 cfs_rq->runtime_expires = expires;
4395
4396 /* we check whether we're throttled above */
4397 if (cfs_rq->runtime_remaining > 0)
4398 unthrottle_cfs_rq(cfs_rq);
4399
4400next:
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02004401 rq_unlock(rq, &rf);
Paul Turner671fd9d2011-07-21 09:43:34 -07004402
4403 if (!remaining)
4404 break;
4405 }
4406 rcu_read_unlock();
4407
Ben Segallc06f04c2014-06-20 15:21:20 -07004408 return starting_runtime - remaining;
Paul Turner671fd9d2011-07-21 09:43:34 -07004409}
4410
Paul Turner58088ad2011-07-21 09:43:31 -07004411/*
4412 * Responsible for refilling a task_group's bandwidth and unthrottling its
4413 * cfs_rqs as appropriate. If there has been no activity within the last
4414 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4415 * used to track this state.
4416 */
4417static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4418{
Paul Turner671fd9d2011-07-21 09:43:34 -07004419 u64 runtime, runtime_expires;
Ben Segall51f21762014-05-19 15:49:45 -07004420 int throttled;
Paul Turner58088ad2011-07-21 09:43:31 -07004421
Paul Turner58088ad2011-07-21 09:43:31 -07004422 /* no need to continue the timer with no bandwidth constraint */
4423 if (cfs_b->quota == RUNTIME_INF)
Ben Segall51f21762014-05-19 15:49:45 -07004424 goto out_deactivate;
Paul Turner58088ad2011-07-21 09:43:31 -07004425
Paul Turner671fd9d2011-07-21 09:43:34 -07004426 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
Nikhil Raoe8da1b12011-07-21 09:43:40 -07004427 cfs_b->nr_periods += overrun;
Paul Turner671fd9d2011-07-21 09:43:34 -07004428
Ben Segall51f21762014-05-19 15:49:45 -07004429 /*
4430 * idle depends on !throttled (for the case of a large deficit), and if
4431 * we're going inactive then everything else can be deferred
4432 */
4433 if (cfs_b->idle && !throttled)
4434 goto out_deactivate;
Paul Turnera9cf55b2011-07-21 09:43:32 -07004435
4436 __refill_cfs_bandwidth_runtime(cfs_b);
4437
Paul Turner671fd9d2011-07-21 09:43:34 -07004438 if (!throttled) {
4439 /* mark as potentially idle for the upcoming period */
4440 cfs_b->idle = 1;
Ben Segall51f21762014-05-19 15:49:45 -07004441 return 0;
Paul Turner671fd9d2011-07-21 09:43:34 -07004442 }
Paul Turner58088ad2011-07-21 09:43:31 -07004443
Nikhil Raoe8da1b12011-07-21 09:43:40 -07004444 /* account preceding periods in which throttling occurred */
4445 cfs_b->nr_throttled += overrun;
4446
Paul Turner671fd9d2011-07-21 09:43:34 -07004447 runtime_expires = cfs_b->runtime_expires;
Paul Turner671fd9d2011-07-21 09:43:34 -07004448
4449 /*
Ben Segallc06f04c2014-06-20 15:21:20 -07004450 * This check is repeated as we are holding onto the new bandwidth while
4451 * we unthrottle. This can potentially race with an unthrottled group
4452 * trying to acquire new bandwidth from the global pool. This can result
4453 * in us over-using our runtime if it is all used during this loop, but
4454 * only by limited amounts in that extreme case.
Paul Turner671fd9d2011-07-21 09:43:34 -07004455 */
Ben Segallc06f04c2014-06-20 15:21:20 -07004456 while (throttled && cfs_b->runtime > 0) {
4457 runtime = cfs_b->runtime;
Paul Turner671fd9d2011-07-21 09:43:34 -07004458 raw_spin_unlock(&cfs_b->lock);
4459 /* we can't nest cfs_b->lock while distributing bandwidth */
4460 runtime = distribute_cfs_runtime(cfs_b, runtime,
4461 runtime_expires);
4462 raw_spin_lock(&cfs_b->lock);
4463
4464 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
Ben Segallc06f04c2014-06-20 15:21:20 -07004465
4466 cfs_b->runtime -= min(runtime, cfs_b->runtime);
Paul Turner671fd9d2011-07-21 09:43:34 -07004467 }
4468
Paul Turner671fd9d2011-07-21 09:43:34 -07004469 /*
4470 * While we are ensured activity in the period following an
4471 * unthrottle, this also covers the case in which the new bandwidth is
4472 * insufficient to cover the existing bandwidth deficit. (Forcing the
4473 * timer to remain active while there are any throttled entities.)
4474 */
4475 cfs_b->idle = 0;
Paul Turner58088ad2011-07-21 09:43:31 -07004476
Ben Segall51f21762014-05-19 15:49:45 -07004477 return 0;
4478
4479out_deactivate:
Ben Segall51f21762014-05-19 15:49:45 -07004480 return 1;
Paul Turner58088ad2011-07-21 09:43:31 -07004481}
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004482
Paul Turnerd8b49862011-07-21 09:43:41 -07004483/* a cfs_rq won't donate quota below this amount */
4484static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4485/* minimum remaining period time to redistribute slack quota */
4486static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4487/* how long we wait to gather additional slack before distributing */
4488static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4489
Ben Segalldb06e782013-10-16 11:16:17 -07004490/*
4491 * Are we near the end of the current quota period?
4492 *
4493 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
Thomas Gleixner4961b6e2015-04-14 21:09:05 +00004494 * hrtimer base being cleared by hrtimer_start. In the case of
Ben Segalldb06e782013-10-16 11:16:17 -07004495 * migrate_hrtimers, base is never cleared, so we are fine.
4496 */
Paul Turnerd8b49862011-07-21 09:43:41 -07004497static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4498{
4499 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4500 u64 remaining;
4501
4502 /* if the call-back is running a quota refresh is already occurring */
4503 if (hrtimer_callback_running(refresh_timer))
4504 return 1;
4505
4506 /* is a quota refresh about to occur? */
4507 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4508 if (remaining < min_expire)
4509 return 1;
4510
4511 return 0;
4512}
4513
4514static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4515{
4516 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4517
4518 /* if there's a quota refresh soon don't bother with slack */
4519 if (runtime_refresh_within(cfs_b, min_left))
4520 return;
4521
Peter Zijlstra4cfafd32015-05-14 12:23:11 +02004522 hrtimer_start(&cfs_b->slack_timer,
4523 ns_to_ktime(cfs_bandwidth_slack_period),
4524 HRTIMER_MODE_REL);
Paul Turnerd8b49862011-07-21 09:43:41 -07004525}
4526
4527/* we know any runtime found here is valid as update_curr() precedes return */
4528static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4529{
4530 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4531 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4532
4533 if (slack_runtime <= 0)
4534 return;
4535
4536 raw_spin_lock(&cfs_b->lock);
4537 if (cfs_b->quota != RUNTIME_INF &&
4538 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4539 cfs_b->runtime += slack_runtime;
4540
4541 /* we are under rq->lock, defer unthrottling using a timer */
4542 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4543 !list_empty(&cfs_b->throttled_cfs_rq))
4544 start_cfs_slack_bandwidth(cfs_b);
4545 }
4546 raw_spin_unlock(&cfs_b->lock);
4547
4548 /* even if it's not valid for return we don't want to try again */
4549 cfs_rq->runtime_remaining -= slack_runtime;
4550}
4551
4552static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4553{
Paul Turner56f570e2011-11-07 20:26:33 -08004554 if (!cfs_bandwidth_used())
4555 return;
4556
Paul Turnerfccfdc62011-11-07 20:26:34 -08004557 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
Paul Turnerd8b49862011-07-21 09:43:41 -07004558 return;
4559
4560 __return_cfs_rq_runtime(cfs_rq);
4561}
4562
4563/*
4564 * This is done with a timer (instead of inline with bandwidth return) since
4565 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4566 */
4567static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4568{
4569 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4570 u64 expires;
4571
4572 /* confirm we're still not at a refresh boundary */
Paul Turnerd8b49862011-07-21 09:43:41 -07004573 raw_spin_lock(&cfs_b->lock);
Ben Segalldb06e782013-10-16 11:16:17 -07004574 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4575 raw_spin_unlock(&cfs_b->lock);
4576 return;
4577 }
4578
Ben Segallc06f04c2014-06-20 15:21:20 -07004579 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
Paul Turnerd8b49862011-07-21 09:43:41 -07004580 runtime = cfs_b->runtime;
Ben Segallc06f04c2014-06-20 15:21:20 -07004581
Paul Turnerd8b49862011-07-21 09:43:41 -07004582 expires = cfs_b->runtime_expires;
4583 raw_spin_unlock(&cfs_b->lock);
4584
4585 if (!runtime)
4586 return;
4587
4588 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4589
4590 raw_spin_lock(&cfs_b->lock);
4591 if (expires == cfs_b->runtime_expires)
Ben Segallc06f04c2014-06-20 15:21:20 -07004592 cfs_b->runtime -= min(runtime, cfs_b->runtime);
Paul Turnerd8b49862011-07-21 09:43:41 -07004593 raw_spin_unlock(&cfs_b->lock);
4594}
4595
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004596/*
4597 * When a group wakes up we want to make sure that its quota is not already
4598 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4599 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4600 */
4601static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4602{
Paul Turner56f570e2011-11-07 20:26:33 -08004603 if (!cfs_bandwidth_used())
4604 return;
4605
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004606 /* an active group must be handled by the update_curr()->put() path */
4607 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4608 return;
4609
4610 /* ensure the group is not already throttled */
4611 if (cfs_rq_throttled(cfs_rq))
4612 return;
4613
4614 /* update runtime allocation */
4615 account_cfs_rq_runtime(cfs_rq, 0);
4616 if (cfs_rq->runtime_remaining <= 0)
4617 throttle_cfs_rq(cfs_rq);
4618}
4619
Peter Zijlstra55e16d32016-06-22 15:14:26 +02004620static void sync_throttle(struct task_group *tg, int cpu)
4621{
4622 struct cfs_rq *pcfs_rq, *cfs_rq;
4623
4624 if (!cfs_bandwidth_used())
4625 return;
4626
4627 if (!tg->parent)
4628 return;
4629
4630 cfs_rq = tg->cfs_rq[cpu];
4631 pcfs_rq = tg->parent->cfs_rq[cpu];
4632
4633 cfs_rq->throttle_count = pcfs_rq->throttle_count;
Xunlei Pangb8922122016-07-09 15:54:22 +08004634 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
Peter Zijlstra55e16d32016-06-22 15:14:26 +02004635}
4636
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004637/* conditionally throttle active cfs_rq's from put_prev_entity() */
Peter Zijlstra678d5712012-02-11 06:05:00 +01004638static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004639{
Paul Turner56f570e2011-11-07 20:26:33 -08004640 if (!cfs_bandwidth_used())
Peter Zijlstra678d5712012-02-11 06:05:00 +01004641 return false;
Paul Turner56f570e2011-11-07 20:26:33 -08004642
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004643 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
Peter Zijlstra678d5712012-02-11 06:05:00 +01004644 return false;
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004645
4646 /*
4647 * it's possible for a throttled entity to be forced into a running
4648 * state (e.g. set_curr_task), in this case we're finished.
4649 */
4650 if (cfs_rq_throttled(cfs_rq))
Peter Zijlstra678d5712012-02-11 06:05:00 +01004651 return true;
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004652
4653 throttle_cfs_rq(cfs_rq);
Peter Zijlstra678d5712012-02-11 06:05:00 +01004654 return true;
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004655}
Peter Zijlstra029632f2011-10-25 10:00:11 +02004656
Peter Zijlstra029632f2011-10-25 10:00:11 +02004657static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4658{
4659 struct cfs_bandwidth *cfs_b =
4660 container_of(timer, struct cfs_bandwidth, slack_timer);
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +02004661
Peter Zijlstra029632f2011-10-25 10:00:11 +02004662 do_sched_cfs_slack_timer(cfs_b);
4663
4664 return HRTIMER_NORESTART;
4665}
4666
4667static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4668{
4669 struct cfs_bandwidth *cfs_b =
4670 container_of(timer, struct cfs_bandwidth, period_timer);
Peter Zijlstra029632f2011-10-25 10:00:11 +02004671 int overrun;
4672 int idle = 0;
4673
Ben Segall51f21762014-05-19 15:49:45 -07004674 raw_spin_lock(&cfs_b->lock);
Peter Zijlstra029632f2011-10-25 10:00:11 +02004675 for (;;) {
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +02004676 overrun = hrtimer_forward_now(timer, cfs_b->period);
Peter Zijlstra029632f2011-10-25 10:00:11 +02004677 if (!overrun)
4678 break;
4679
4680 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4681 }
Peter Zijlstra4cfafd32015-05-14 12:23:11 +02004682 if (idle)
4683 cfs_b->period_active = 0;
Ben Segall51f21762014-05-19 15:49:45 -07004684 raw_spin_unlock(&cfs_b->lock);
Peter Zijlstra029632f2011-10-25 10:00:11 +02004685
4686 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4687}
4688
4689void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4690{
4691 raw_spin_lock_init(&cfs_b->lock);
4692 cfs_b->runtime = 0;
4693 cfs_b->quota = RUNTIME_INF;
4694 cfs_b->period = ns_to_ktime(default_cfs_period());
4695
4696 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
Peter Zijlstra4cfafd32015-05-14 12:23:11 +02004697 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
Peter Zijlstra029632f2011-10-25 10:00:11 +02004698 cfs_b->period_timer.function = sched_cfs_period_timer;
4699 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4700 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4701}
4702
4703static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4704{
4705 cfs_rq->runtime_enabled = 0;
4706 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4707}
4708
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +02004709void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
Peter Zijlstra029632f2011-10-25 10:00:11 +02004710{
Peter Zijlstra4cfafd32015-05-14 12:23:11 +02004711 lockdep_assert_held(&cfs_b->lock);
Peter Zijlstra029632f2011-10-25 10:00:11 +02004712
Peter Zijlstra4cfafd32015-05-14 12:23:11 +02004713 if (!cfs_b->period_active) {
4714 cfs_b->period_active = 1;
4715 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4716 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4717 }
Peter Zijlstra029632f2011-10-25 10:00:11 +02004718}
4719
4720static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4721{
Tetsuo Handa7f1a1692014-12-25 15:51:21 +09004722 /* init_cfs_bandwidth() was not called */
4723 if (!cfs_b->throttled_cfs_rq.next)
4724 return;
4725
Peter Zijlstra029632f2011-10-25 10:00:11 +02004726 hrtimer_cancel(&cfs_b->period_timer);
4727 hrtimer_cancel(&cfs_b->slack_timer);
4728}
4729
Peter Zijlstra502ce002017-05-04 15:31:22 +02004730/*
4731 * Both these cpu hotplug callbacks race against unregister_fair_sched_group()
4732 *
4733 * The race is harmless, since modifying bandwidth settings of unhooked group
4734 * bits doesn't do much.
4735 */
4736
4737/* cpu online calback */
Kirill Tkhai0e59bda2014-06-25 12:19:42 +04004738static void __maybe_unused update_runtime_enabled(struct rq *rq)
4739{
Peter Zijlstra502ce002017-05-04 15:31:22 +02004740 struct task_group *tg;
Kirill Tkhai0e59bda2014-06-25 12:19:42 +04004741
Peter Zijlstra502ce002017-05-04 15:31:22 +02004742 lockdep_assert_held(&rq->lock);
4743
4744 rcu_read_lock();
4745 list_for_each_entry_rcu(tg, &task_groups, list) {
4746 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4747 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
Kirill Tkhai0e59bda2014-06-25 12:19:42 +04004748
4749 raw_spin_lock(&cfs_b->lock);
4750 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4751 raw_spin_unlock(&cfs_b->lock);
4752 }
Peter Zijlstra502ce002017-05-04 15:31:22 +02004753 rcu_read_unlock();
Kirill Tkhai0e59bda2014-06-25 12:19:42 +04004754}
4755
Peter Zijlstra502ce002017-05-04 15:31:22 +02004756/* cpu offline callback */
Arnd Bergmann38dc3342013-01-25 14:14:22 +00004757static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
Peter Zijlstra029632f2011-10-25 10:00:11 +02004758{
Peter Zijlstra502ce002017-05-04 15:31:22 +02004759 struct task_group *tg;
Peter Zijlstra029632f2011-10-25 10:00:11 +02004760
Peter Zijlstra502ce002017-05-04 15:31:22 +02004761 lockdep_assert_held(&rq->lock);
4762
4763 rcu_read_lock();
4764 list_for_each_entry_rcu(tg, &task_groups, list) {
4765 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4766
Peter Zijlstra029632f2011-10-25 10:00:11 +02004767 if (!cfs_rq->runtime_enabled)
4768 continue;
4769
4770 /*
4771 * clock_task is not advancing so we just need to make sure
4772 * there's some valid quota amount
4773 */
Ben Segall51f21762014-05-19 15:49:45 -07004774 cfs_rq->runtime_remaining = 1;
Kirill Tkhai0e59bda2014-06-25 12:19:42 +04004775 /*
4776 * Offline rq is schedulable till cpu is completely disabled
4777 * in take_cpu_down(), so we prevent new cfs throttling here.
4778 */
4779 cfs_rq->runtime_enabled = 0;
4780
Peter Zijlstra029632f2011-10-25 10:00:11 +02004781 if (cfs_rq_throttled(cfs_rq))
4782 unthrottle_cfs_rq(cfs_rq);
4783 }
Peter Zijlstra502ce002017-05-04 15:31:22 +02004784 rcu_read_unlock();
Peter Zijlstra029632f2011-10-25 10:00:11 +02004785}
4786
4787#else /* CONFIG_CFS_BANDWIDTH */
Paul Turnerf1b17282012-10-04 13:18:31 +02004788static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4789{
Frederic Weisbecker78becc22013-04-12 01:51:02 +02004790 return rq_clock_task(rq_of(cfs_rq));
Paul Turnerf1b17282012-10-04 13:18:31 +02004791}
4792
Peter Zijlstra9dbdb152013-11-18 18:27:06 +01004793static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
Peter Zijlstra678d5712012-02-11 06:05:00 +01004794static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
Paul Turnerd3d9dc32011-07-21 09:43:39 -07004795static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
Peter Zijlstra55e16d32016-06-22 15:14:26 +02004796static inline void sync_throttle(struct task_group *tg, int cpu) {}
Peter Zijlstra6c16a6d2012-03-21 13:07:16 -07004797static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
Paul Turner85dac902011-07-21 09:43:33 -07004798
4799static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4800{
4801 return 0;
4802}
Paul Turner64660c82011-07-21 09:43:36 -07004803
4804static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4805{
4806 return 0;
4807}
4808
4809static inline int throttled_lb_pair(struct task_group *tg,
4810 int src_cpu, int dest_cpu)
4811{
4812 return 0;
4813}
Peter Zijlstra029632f2011-10-25 10:00:11 +02004814
4815void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4816
4817#ifdef CONFIG_FAIR_GROUP_SCHED
4818static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
Paul Turnerab84d312011-07-21 09:43:28 -07004819#endif
4820
Peter Zijlstra029632f2011-10-25 10:00:11 +02004821static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4822{
4823 return NULL;
4824}
4825static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
Kirill Tkhai0e59bda2014-06-25 12:19:42 +04004826static inline void update_runtime_enabled(struct rq *rq) {}
Peter Boonstoppela4c96ae2012-08-09 15:34:47 -07004827static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
Peter Zijlstra029632f2011-10-25 10:00:11 +02004828
4829#endif /* CONFIG_CFS_BANDWIDTH */
4830
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004831/**************************************************
4832 * CFS operations on tasks:
4833 */
4834
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004835#ifdef CONFIG_SCHED_HRTICK
4836static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4837{
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004838 struct sched_entity *se = &p->se;
4839 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4840
Peter Zijlstra9148a3a2016-09-20 22:34:51 +02004841 SCHED_WARN_ON(task_rq(p) != rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004842
Srivatsa Vaddagiri8bf46a32016-09-16 18:28:51 -07004843 if (rq->cfs.h_nr_running > 1) {
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004844 u64 slice = sched_slice(cfs_rq, se);
4845 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4846 s64 delta = slice - ran;
4847
4848 if (delta < 0) {
4849 if (rq->curr == p)
Kirill Tkhai88751252014-06-29 00:03:57 +04004850 resched_curr(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004851 return;
4852 }
Peter Zijlstra31656512008-07-18 18:01:23 +02004853 hrtick_start(rq, delta);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004854 }
4855}
Peter Zijlstraa4c2f002008-10-17 19:27:03 +02004856
4857/*
4858 * called from enqueue/dequeue and updates the hrtick when the
4859 * current task is from our class and nr_running is low enough
4860 * to matter.
4861 */
4862static void hrtick_update(struct rq *rq)
4863{
4864 struct task_struct *curr = rq->curr;
4865
Mike Galbraithb39e66e2011-11-22 15:20:07 +01004866 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
Peter Zijlstraa4c2f002008-10-17 19:27:03 +02004867 return;
4868
4869 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4870 hrtick_start_fair(rq, curr);
4871}
Dhaval Giani55e12e52008-06-24 23:39:43 +05304872#else /* !CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004873static inline void
4874hrtick_start_fair(struct rq *rq, struct task_struct *p)
4875{
4876}
Peter Zijlstraa4c2f002008-10-17 19:27:03 +02004877
4878static inline void hrtick_update(struct rq *rq)
4879{
4880}
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004881#endif
4882
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004883/*
4884 * The enqueue_task method is called before nr_running is
4885 * increased. Here we update the fair scheduling stats and
4886 * then put the task into the rbtree:
4887 */
Thomas Gleixnerea87bb72010-01-20 20:58:57 +00004888static void
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01004889enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004890{
4891 struct cfs_rq *cfs_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01004892 struct sched_entity *se = &p->se;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004893
Rafael J. Wysocki8c34ab12016-09-09 23:59:33 +02004894 /*
4895 * If in_iowait is set, the code below may not trigger any cpufreq
4896 * utilization updates, so do it here explicitly with the IOWAIT flag
4897 * passed.
4898 */
4899 if (p->in_iowait)
4900 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4901
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004902 for_each_sched_entity(se) {
Peter Zijlstra62fb1852008-02-25 17:34:02 +01004903 if (se->on_rq)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004904 break;
4905 cfs_rq = cfs_rq_of(se);
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01004906 enqueue_entity(cfs_rq, se, flags);
Paul Turner85dac902011-07-21 09:43:33 -07004907
4908 /*
4909 * end evaluation on encountering a throttled cfs_rq
4910 *
4911 * note: in the case of encountering a throttled cfs_rq we will
4912 * post the final h_nr_running increment below.
Peter Zijlstrae210bff2016-06-16 18:51:48 +02004913 */
Paul Turner85dac902011-07-21 09:43:33 -07004914 if (cfs_rq_throttled(cfs_rq))
4915 break;
Paul Turner953bfcd2011-07-21 09:43:27 -07004916 cfs_rq->h_nr_running++;
Paul Turner85dac902011-07-21 09:43:33 -07004917
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01004918 flags = ENQUEUE_WAKEUP;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004919 }
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004920
Peter Zijlstra2069dd72010-11-15 15:47:00 -08004921 for_each_sched_entity(se) {
Lin Ming0f317142011-07-22 09:14:31 +08004922 cfs_rq = cfs_rq_of(se);
Paul Turner953bfcd2011-07-21 09:43:27 -07004923 cfs_rq->h_nr_running++;
Peter Zijlstra2069dd72010-11-15 15:47:00 -08004924
Paul Turner85dac902011-07-21 09:43:33 -07004925 if (cfs_rq_throttled(cfs_rq))
4926 break;
4927
Vincent Guittotd31b1a62016-11-08 10:53:44 +01004928 update_load_avg(se, UPDATE_TG);
Vincent Guittot89ee0482016-12-21 16:50:26 +01004929 update_cfs_shares(se);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08004930 }
4931
Yuyang Ducd126af2015-07-15 08:04:36 +08004932 if (!se)
Kirill Tkhai72465442014-05-09 03:00:14 +04004933 add_nr_running(rq, 1);
Yuyang Ducd126af2015-07-15 08:04:36 +08004934
Peter Zijlstraa4c2f002008-10-17 19:27:03 +02004935 hrtick_update(rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004936}
4937
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07004938static void set_next_buddy(struct sched_entity *se);
4939
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004940/*
4941 * The dequeue_task method is called before nr_running is
4942 * decreased. We remove the task from the rbtree and
4943 * update the fair scheduling stats:
4944 */
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01004945static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004946{
4947 struct cfs_rq *cfs_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01004948 struct sched_entity *se = &p->se;
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07004949 int task_sleep = flags & DEQUEUE_SLEEP;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004950
4951 for_each_sched_entity(se) {
4952 cfs_rq = cfs_rq_of(se);
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01004953 dequeue_entity(cfs_rq, se, flags);
Paul Turner85dac902011-07-21 09:43:33 -07004954
4955 /*
4956 * end evaluation on encountering a throttled cfs_rq
4957 *
4958 * note: in the case of encountering a throttled cfs_rq we will
4959 * post the final h_nr_running decrement below.
4960 */
4961 if (cfs_rq_throttled(cfs_rq))
4962 break;
Paul Turner953bfcd2011-07-21 09:43:27 -07004963 cfs_rq->h_nr_running--;
Peter Zijlstra2069dd72010-11-15 15:47:00 -08004964
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004965 /* Don't dequeue parent if it has other entities besides us */
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07004966 if (cfs_rq->load.weight) {
Konstantin Khlebnikov754bd592016-06-16 15:57:15 +03004967 /* Avoid re-evaluating load for this entity: */
4968 se = parent_entity(se);
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07004969 /*
4970 * Bias pick_next to pick a task from this cfs_rq, as
4971 * p is sleeping when it is within its sched_slice.
4972 */
Konstantin Khlebnikov754bd592016-06-16 15:57:15 +03004973 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4974 set_next_buddy(se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004975 break;
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07004976 }
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01004977 flags |= DEQUEUE_SLEEP;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004978 }
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004979
Peter Zijlstra2069dd72010-11-15 15:47:00 -08004980 for_each_sched_entity(se) {
Lin Ming0f317142011-07-22 09:14:31 +08004981 cfs_rq = cfs_rq_of(se);
Paul Turner953bfcd2011-07-21 09:43:27 -07004982 cfs_rq->h_nr_running--;
Peter Zijlstra2069dd72010-11-15 15:47:00 -08004983
Paul Turner85dac902011-07-21 09:43:33 -07004984 if (cfs_rq_throttled(cfs_rq))
4985 break;
4986
Vincent Guittotd31b1a62016-11-08 10:53:44 +01004987 update_load_avg(se, UPDATE_TG);
Vincent Guittot89ee0482016-12-21 16:50:26 +01004988 update_cfs_shares(se);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08004989 }
4990
Yuyang Ducd126af2015-07-15 08:04:36 +08004991 if (!se)
Kirill Tkhai72465442014-05-09 03:00:14 +04004992 sub_nr_running(rq, 1);
Yuyang Ducd126af2015-07-15 08:04:36 +08004993
Peter Zijlstraa4c2f002008-10-17 19:27:03 +02004994 hrtick_update(rq);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02004995}
4996
Gregory Haskinse7693a32008-01-25 21:08:09 +01004997#ifdef CONFIG_SMP
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02004998
4999/* Working cpumask for: load_balance, load_balance_newidle. */
5000DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5001DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5002
Frederic Weisbecker9fd81dd2016-04-19 17:36:51 +02005003#ifdef CONFIG_NO_HZ_COMMON
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005004/*
5005 * per rq 'load' arrray crap; XXX kill this.
5006 */
5007
5008/*
Peter Zijlstrad937cdc2015-10-19 13:49:30 +02005009 * The exact cpuload calculated at every tick would be:
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005010 *
Peter Zijlstrad937cdc2015-10-19 13:49:30 +02005011 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5012 *
5013 * If a cpu misses updates for n ticks (as it was idle) and update gets
5014 * called on the n+1-th tick when cpu may be busy, then we have:
5015 *
5016 * load_n = (1 - 1/2^i)^n * load_0
5017 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005018 *
5019 * decay_load_missed() below does efficient calculation of
Peter Zijlstrad937cdc2015-10-19 13:49:30 +02005020 *
5021 * load' = (1 - 1/2^i)^n * load
5022 *
5023 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5024 * This allows us to precompute the above in said factors, thereby allowing the
5025 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5026 * fixed_power_int())
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005027 *
5028 * The calculation is approximated on a 128 point scale.
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005029 */
5030#define DEGRADE_SHIFT 7
Peter Zijlstrad937cdc2015-10-19 13:49:30 +02005031
5032static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5033static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5034 { 0, 0, 0, 0, 0, 0, 0, 0 },
5035 { 64, 32, 8, 0, 0, 0, 0, 0 },
5036 { 96, 72, 40, 12, 1, 0, 0, 0 },
5037 { 112, 98, 75, 43, 15, 1, 0, 0 },
5038 { 120, 112, 98, 76, 45, 16, 2, 0 }
5039};
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005040
5041/*
5042 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5043 * would be when CPU is idle and so we just decay the old load without
5044 * adding any new load.
5045 */
5046static unsigned long
5047decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5048{
5049 int j = 0;
5050
5051 if (!missed_updates)
5052 return load;
5053
5054 if (missed_updates >= degrade_zero_ticks[idx])
5055 return 0;
5056
5057 if (idx == 1)
5058 return load >> missed_updates;
5059
5060 while (missed_updates) {
5061 if (missed_updates % 2)
5062 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5063
5064 missed_updates >>= 1;
5065 j++;
5066 }
5067 return load;
5068}
Frederic Weisbecker9fd81dd2016-04-19 17:36:51 +02005069#endif /* CONFIG_NO_HZ_COMMON */
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005070
Byungchul Park59543272015-10-14 18:47:35 +09005071/**
Frederic Weisbeckercee1afc2016-04-13 15:56:50 +02005072 * __cpu_load_update - update the rq->cpu_load[] statistics
Byungchul Park59543272015-10-14 18:47:35 +09005073 * @this_rq: The rq to update statistics for
5074 * @this_load: The current load
5075 * @pending_updates: The number of missed updates
Byungchul Park59543272015-10-14 18:47:35 +09005076 *
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005077 * Update rq->cpu_load[] statistics. This function is usually called every
Byungchul Park59543272015-10-14 18:47:35 +09005078 * scheduler tick (TICK_NSEC).
5079 *
5080 * This function computes a decaying average:
5081 *
5082 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5083 *
5084 * Because of NOHZ it might not get called on every tick which gives need for
5085 * the @pending_updates argument.
5086 *
5087 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5088 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5089 * = A * (A * load[i]_n-2 + B) + B
5090 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5091 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5092 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5093 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5094 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5095 *
5096 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5097 * any change in load would have resulted in the tick being turned back on.
5098 *
5099 * For regular NOHZ, this reduces to:
5100 *
5101 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5102 *
5103 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005104 * term.
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005105 */
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005106static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5107 unsigned long pending_updates)
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005108{
Frederic Weisbecker9fd81dd2016-04-19 17:36:51 +02005109 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005110 int i, scale;
5111
5112 this_rq->nr_load_updates++;
5113
5114 /* Update our load: */
5115 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5116 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5117 unsigned long old_load, new_load;
5118
5119 /* scale is effectively 1 << i now, and >> i divides by scale */
5120
Byungchul Park7400d3b2016-01-15 16:07:49 +09005121 old_load = this_rq->cpu_load[i];
Frederic Weisbecker9fd81dd2016-04-19 17:36:51 +02005122#ifdef CONFIG_NO_HZ_COMMON
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005123 old_load = decay_load_missed(old_load, pending_updates - 1, i);
Byungchul Park7400d3b2016-01-15 16:07:49 +09005124 if (tickless_load) {
5125 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5126 /*
5127 * old_load can never be a negative value because a
5128 * decayed tickless_load cannot be greater than the
5129 * original tickless_load.
5130 */
5131 old_load += tickless_load;
5132 }
Frederic Weisbecker9fd81dd2016-04-19 17:36:51 +02005133#endif
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005134 new_load = this_load;
5135 /*
5136 * Round up the averaging division if load is increasing. This
5137 * prevents us from getting stuck on 9 if the load is 10, for
5138 * example.
5139 */
5140 if (new_load > old_load)
5141 new_load += scale - 1;
5142
5143 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5144 }
5145
5146 sched_avg_update(this_rq);
5147}
5148
Yuyang Du7ea241a2015-07-15 08:04:42 +08005149/* Used instead of source_load when we know the type == 0 */
Viresh Kumarc7132dd2017-05-24 10:59:54 +05305150static unsigned long weighted_cpuload(struct rq *rq)
Yuyang Du7ea241a2015-07-15 08:04:42 +08005151{
Viresh Kumarc7132dd2017-05-24 10:59:54 +05305152 return cfs_rq_runnable_load_avg(&rq->cfs);
Yuyang Du7ea241a2015-07-15 08:04:42 +08005153}
5154
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005155#ifdef CONFIG_NO_HZ_COMMON
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005156/*
5157 * There is no sane way to deal with nohz on smp when using jiffies because the
5158 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
5159 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5160 *
5161 * Therefore we need to avoid the delta approach from the regular tick when
5162 * possible since that would seriously skew the load calculation. This is why we
5163 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5164 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5165 * loop exit, nohz_idle_balance, nohz full exit...)
5166 *
5167 * This means we might still be one tick off for nohz periods.
5168 */
5169
5170static void cpu_load_update_nohz(struct rq *this_rq,
5171 unsigned long curr_jiffies,
5172 unsigned long load)
Frederic Weisbeckerbe68a682016-01-13 17:01:29 +01005173{
5174 unsigned long pending_updates;
5175
5176 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5177 if (pending_updates) {
5178 this_rq->last_load_update_tick = curr_jiffies;
5179 /*
5180 * In the regular NOHZ case, we were idle, this means load 0.
5181 * In the NOHZ_FULL case, we were non-idle, we should consider
5182 * its weighted load.
5183 */
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005184 cpu_load_update(this_rq, load, pending_updates);
Frederic Weisbeckerbe68a682016-01-13 17:01:29 +01005185 }
5186}
5187
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005188/*
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005189 * Called from nohz_idle_balance() to update the load ratings before doing the
5190 * idle balance.
5191 */
Frederic Weisbeckercee1afc2016-04-13 15:56:50 +02005192static void cpu_load_update_idle(struct rq *this_rq)
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005193{
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005194 /*
5195 * bail if there's load or we're actually up-to-date.
5196 */
Viresh Kumarc7132dd2017-05-24 10:59:54 +05305197 if (weighted_cpuload(this_rq))
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005198 return;
5199
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005200 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005201}
5202
5203/*
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005204 * Record CPU load on nohz entry so we know the tickless load to account
5205 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5206 * than other cpu_load[idx] but it should be fine as cpu_load readers
5207 * shouldn't rely into synchronized cpu_load[*] updates.
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005208 */
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005209void cpu_load_update_nohz_start(void)
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005210{
5211 struct rq *this_rq = this_rq();
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005212
5213 /*
5214 * This is all lockless but should be fine. If weighted_cpuload changes
5215 * concurrently we'll exit nohz. And cpu_load write can race with
5216 * cpu_load_update_idle() but both updater would be writing the same.
5217 */
Viresh Kumarc7132dd2017-05-24 10:59:54 +05305218 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005219}
5220
5221/*
5222 * Account the tickless load in the end of a nohz frame.
5223 */
5224void cpu_load_update_nohz_stop(void)
5225{
Jason Low316c1608d2015-04-28 13:00:20 -07005226 unsigned long curr_jiffies = READ_ONCE(jiffies);
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005227 struct rq *this_rq = this_rq();
5228 unsigned long load;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02005229 struct rq_flags rf;
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005230
5231 if (curr_jiffies == this_rq->last_load_update_tick)
5232 return;
5233
Viresh Kumarc7132dd2017-05-24 10:59:54 +05305234 load = weighted_cpuload(this_rq);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02005235 rq_lock(this_rq, &rf);
Matt Flemingb52fad22016-05-03 20:46:54 +01005236 update_rq_clock(this_rq);
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005237 cpu_load_update_nohz(this_rq, curr_jiffies, load);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02005238 rq_unlock(this_rq, &rf);
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005239}
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005240#else /* !CONFIG_NO_HZ_COMMON */
5241static inline void cpu_load_update_nohz(struct rq *this_rq,
5242 unsigned long curr_jiffies,
5243 unsigned long load) { }
5244#endif /* CONFIG_NO_HZ_COMMON */
5245
5246static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5247{
Frederic Weisbecker9fd81dd2016-04-19 17:36:51 +02005248#ifdef CONFIG_NO_HZ_COMMON
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005249 /* See the mess around cpu_load_update_nohz(). */
5250 this_rq->last_load_update_tick = READ_ONCE(jiffies);
Frederic Weisbecker9fd81dd2016-04-19 17:36:51 +02005251#endif
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005252 cpu_load_update(this_rq, load, 1);
5253}
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005254
5255/*
5256 * Called from scheduler_tick()
5257 */
Frederic Weisbeckercee1afc2016-04-13 15:56:50 +02005258void cpu_load_update_active(struct rq *this_rq)
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005259{
Viresh Kumarc7132dd2017-05-24 10:59:54 +05305260 unsigned long load = weighted_cpuload(this_rq);
Frederic Weisbecker1f419062016-04-13 15:56:51 +02005261
5262 if (tick_nohz_tick_stopped())
5263 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5264 else
5265 cpu_load_update_periodic(this_rq, load);
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005266}
5267
Peter Zijlstra029632f2011-10-25 10:00:11 +02005268/*
5269 * Return a low guess at the load of a migration-source cpu weighted
5270 * according to the scheduling class and "nice" value.
5271 *
5272 * We want to under-estimate the load of migration sources, to
5273 * balance conservatively.
5274 */
5275static unsigned long source_load(int cpu, int type)
5276{
5277 struct rq *rq = cpu_rq(cpu);
Viresh Kumarc7132dd2017-05-24 10:59:54 +05305278 unsigned long total = weighted_cpuload(rq);
Peter Zijlstra029632f2011-10-25 10:00:11 +02005279
5280 if (type == 0 || !sched_feat(LB_BIAS))
5281 return total;
5282
5283 return min(rq->cpu_load[type-1], total);
5284}
5285
5286/*
5287 * Return a high guess at the load of a migration-target cpu weighted
5288 * according to the scheduling class and "nice" value.
5289 */
5290static unsigned long target_load(int cpu, int type)
5291{
5292 struct rq *rq = cpu_rq(cpu);
Viresh Kumarc7132dd2017-05-24 10:59:54 +05305293 unsigned long total = weighted_cpuload(rq);
Peter Zijlstra029632f2011-10-25 10:00:11 +02005294
5295 if (type == 0 || !sched_feat(LB_BIAS))
5296 return total;
5297
5298 return max(rq->cpu_load[type-1], total);
5299}
5300
Nicolas Pitreced549f2014-05-26 18:19:38 -04005301static unsigned long capacity_of(int cpu)
Peter Zijlstra029632f2011-10-25 10:00:11 +02005302{
Nicolas Pitreced549f2014-05-26 18:19:38 -04005303 return cpu_rq(cpu)->cpu_capacity;
Peter Zijlstra029632f2011-10-25 10:00:11 +02005304}
5305
Vincent Guittotca6d75e2015-02-27 16:54:09 +01005306static unsigned long capacity_orig_of(int cpu)
5307{
5308 return cpu_rq(cpu)->cpu_capacity_orig;
5309}
5310
Peter Zijlstra029632f2011-10-25 10:00:11 +02005311static unsigned long cpu_avg_load_per_task(int cpu)
5312{
5313 struct rq *rq = cpu_rq(cpu);
Jason Low316c1608d2015-04-28 13:00:20 -07005314 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
Viresh Kumarc7132dd2017-05-24 10:59:54 +05305315 unsigned long load_avg = weighted_cpuload(rq);
Peter Zijlstra029632f2011-10-25 10:00:11 +02005316
5317 if (nr_running)
Alex Shib92486c2013-06-20 10:18:50 +08005318 return load_avg / nr_running;
Peter Zijlstra029632f2011-10-25 10:00:11 +02005319
5320 return 0;
5321}
5322
Peter Zijlstrac58d25f2016-05-12 09:19:59 +02005323static void record_wakee(struct task_struct *p)
5324{
5325 /*
5326 * Only decay a single time; tasks that have less then 1 wakeup per
5327 * jiffy will not have built up many flips.
5328 */
5329 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5330 current->wakee_flips >>= 1;
5331 current->wakee_flip_decay_ts = jiffies;
5332 }
5333
5334 if (current->last_wakee != p) {
5335 current->last_wakee = p;
5336 current->wakee_flips++;
5337 }
5338}
5339
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02005340/*
5341 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
Peter Zijlstrac58d25f2016-05-12 09:19:59 +02005342 *
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02005343 * A waker of many should wake a different task than the one last awakened
Peter Zijlstrac58d25f2016-05-12 09:19:59 +02005344 * at a frequency roughly N times higher than one of its wakees.
5345 *
5346 * In order to determine whether we should let the load spread vs consolidating
5347 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5348 * partner, and a factor of lls_size higher frequency in the other.
5349 *
5350 * With both conditions met, we can be relatively sure that the relationship is
5351 * non-monogamous, with partner count exceeding socket size.
5352 *
5353 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5354 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5355 * socket size.
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02005356 */
Michael Wang62470412013-07-04 12:55:51 +08005357static int wake_wide(struct task_struct *p)
5358{
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02005359 unsigned int master = current->wakee_flips;
5360 unsigned int slave = p->wakee_flips;
Peter Zijlstra7d9ffa82013-07-04 12:56:46 +08005361 int factor = this_cpu_read(sd_llc_size);
Michael Wang62470412013-07-04 12:55:51 +08005362
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02005363 if (master < slave)
5364 swap(master, slave);
5365 if (slave < factor || master < slave * factor)
5366 return 0;
5367 return 1;
Michael Wang62470412013-07-04 12:55:51 +08005368}
5369
Peter Zijlstra90001d62017-07-31 17:50:05 +02005370struct llc_stats {
5371 unsigned long nr_running;
5372 unsigned long load;
5373 unsigned long capacity;
5374 int has_capacity;
5375};
5376
5377static bool get_llc_stats(struct llc_stats *stats, int cpu)
5378{
5379 struct sched_domain_shared *sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5380
5381 if (!sds)
5382 return false;
5383
5384 stats->nr_running = READ_ONCE(sds->nr_running);
5385 stats->load = READ_ONCE(sds->load);
5386 stats->capacity = READ_ONCE(sds->capacity);
5387 stats->has_capacity = stats->nr_running < per_cpu(sd_llc_size, cpu);
5388
5389 return true;
5390}
5391
5392/*
5393 * Can a task be moved from prev_cpu to this_cpu without causing a load
5394 * imbalance that would trigger the load balancer?
5395 *
5396 * Since we're running on 'stale' values, we might in fact create an imbalance
5397 * but recomputing these values is expensive, as that'd mean iteration 2 cache
5398 * domains worth of CPUs.
5399 */
5400static bool
5401wake_affine_llc(struct sched_domain *sd, struct task_struct *p,
5402 int this_cpu, int prev_cpu, int sync)
5403{
5404 struct llc_stats prev_stats, this_stats;
5405 s64 this_eff_load, prev_eff_load;
5406 unsigned long task_load;
5407
5408 if (!get_llc_stats(&prev_stats, prev_cpu) ||
5409 !get_llc_stats(&this_stats, this_cpu))
5410 return false;
5411
5412 /*
5413 * If sync wakeup then subtract the (maximum possible)
5414 * effect of the currently running task from the load
5415 * of the current LLC.
5416 */
5417 if (sync) {
5418 unsigned long current_load = task_h_load(current);
5419
5420 /* in this case load hits 0 and this LLC is considered 'idle' */
5421 if (current_load > this_stats.load)
5422 return true;
5423
5424 this_stats.load -= current_load;
5425 }
5426
5427 /*
5428 * The has_capacity stuff is not SMT aware, but by trying to balance
5429 * the nr_running on both ends we try and fill the domain at equal
5430 * rates, thereby first consuming cores before siblings.
5431 */
5432
5433 /* if the old cache has capacity, stay there */
5434 if (prev_stats.has_capacity && prev_stats.nr_running < this_stats.nr_running+1)
5435 return false;
5436
5437 /* if this cache has capacity, come here */
5438 if (this_stats.has_capacity && this_stats.nr_running < prev_stats.nr_running+1)
5439 return true;
5440
5441 /*
5442 * Check to see if we can move the load without causing too much
5443 * imbalance.
5444 */
5445 task_load = task_h_load(p);
5446
5447 this_eff_load = 100;
5448 this_eff_load *= prev_stats.capacity;
5449
5450 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5451 prev_eff_load *= this_stats.capacity;
5452
5453 this_eff_load *= this_stats.load + task_load;
5454 prev_eff_load *= prev_stats.load - task_load;
5455
5456 return this_eff_load <= prev_eff_load;
5457}
5458
Morten Rasmussen772bd008c2016-06-22 18:03:13 +01005459static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5460 int prev_cpu, int sync)
Ingo Molnar098fb9d2008-03-16 20:36:10 +01005461{
Rik van Riel3fed3822017-06-23 12:55:29 -04005462 int this_cpu = smp_processor_id();
Peter Zijlstra90001d62017-07-31 17:50:05 +02005463 bool affine;
Ingo Molnar098fb9d2008-03-16 20:36:10 +01005464
5465 /*
Peter Zijlstra90001d62017-07-31 17:50:05 +02005466 * Default to no affine wakeups; wake_affine() should not effect a task
5467 * placement the load-balancer feels inclined to undo. The conservative
5468 * option is therefore to not move tasks when they wake up.
Rik van Riel7d894e62017-06-23 12:55:28 -04005469 */
Peter Zijlstra90001d62017-07-31 17:50:05 +02005470 affine = false;
5471
5472 /*
5473 * If the wakeup is across cache domains, try to evaluate if movement
5474 * makes sense, otherwise rely on select_idle_siblings() to do
5475 * placement inside the cache domain.
5476 */
5477 if (!cpus_share_cache(prev_cpu, this_cpu))
5478 affine = wake_affine_llc(sd, p, this_cpu, prev_cpu, sync);
Mike Galbraithb3137bc2008-05-29 11:11:41 +02005479
Josh Poimboeufae928822016-06-17 12:43:24 -05005480 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
Rik van Riel3fed3822017-06-23 12:55:29 -04005481 if (affine) {
5482 schedstat_inc(sd->ttwu_move_affine);
5483 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5484 }
Mike Galbraithb3137bc2008-05-29 11:11:41 +02005485
Rik van Riel3fed3822017-06-23 12:55:29 -04005486 return affine;
Ingo Molnar098fb9d2008-03-16 20:36:10 +01005487}
5488
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005489static inline int task_util(struct task_struct *p);
5490static int cpu_util_wake(int cpu, struct task_struct *p);
5491
5492static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5493{
5494 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5495}
5496
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005497/*
5498 * find_idlest_group finds and returns the least busy CPU group within the
5499 * domain.
5500 */
5501static struct sched_group *
Peter Zijlstra78e7ed52009-09-03 13:16:51 +02005502find_idlest_group(struct sched_domain *sd, struct task_struct *p,
Vincent Guittotc44f2a02013-10-18 13:52:21 +02005503 int this_cpu, int sd_flag)
Gregory Haskinse7693a32008-01-25 21:08:09 +01005504{
Andi Kleenb3bd3de2010-08-10 14:17:51 -07005505 struct sched_group *idlest = NULL, *group = sd->groups;
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005506 struct sched_group *most_spare_sg = NULL;
Vincent Guittot6b947802016-12-08 17:56:54 +01005507 unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5508 unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005509 unsigned long most_spare = 0, this_spare = 0;
Vincent Guittotc44f2a02013-10-18 13:52:21 +02005510 int load_idx = sd->forkexec_idx;
Vincent Guittot6b947802016-12-08 17:56:54 +01005511 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5512 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5513 (sd->imbalance_pct-100) / 100;
Gregory Haskinse7693a32008-01-25 21:08:09 +01005514
Vincent Guittotc44f2a02013-10-18 13:52:21 +02005515 if (sd_flag & SD_BALANCE_WAKE)
5516 load_idx = sd->wake_idx;
5517
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005518 do {
Vincent Guittot6b947802016-12-08 17:56:54 +01005519 unsigned long load, avg_load, runnable_load;
5520 unsigned long spare_cap, max_spare_cap;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005521 int local_group;
5522 int i;
Gregory Haskinse7693a32008-01-25 21:08:09 +01005523
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005524 /* Skip over this group if it has no CPUs allowed */
Peter Zijlstraae4df9d2017-05-01 11:03:12 +02005525 if (!cpumask_intersects(sched_group_span(group),
Ingo Molnar0c98d342017-02-05 15:38:10 +01005526 &p->cpus_allowed))
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005527 continue;
5528
5529 local_group = cpumask_test_cpu(this_cpu,
Peter Zijlstraae4df9d2017-05-01 11:03:12 +02005530 sched_group_span(group));
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005531
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005532 /*
5533 * Tally up the load of all CPUs in the group and find
5534 * the group containing the CPU with most spare capacity.
5535 */
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005536 avg_load = 0;
Vincent Guittot6b947802016-12-08 17:56:54 +01005537 runnable_load = 0;
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005538 max_spare_cap = 0;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005539
Peter Zijlstraae4df9d2017-05-01 11:03:12 +02005540 for_each_cpu(i, sched_group_span(group)) {
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005541 /* Bias balancing toward cpus of our domain */
5542 if (local_group)
5543 load = source_load(i, load_idx);
5544 else
5545 load = target_load(i, load_idx);
5546
Vincent Guittot6b947802016-12-08 17:56:54 +01005547 runnable_load += load;
5548
5549 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005550
5551 spare_cap = capacity_spare_wake(i, p);
5552
5553 if (spare_cap > max_spare_cap)
5554 max_spare_cap = spare_cap;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005555 }
5556
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04005557 /* Adjust by relative CPU capacity of the group */
Vincent Guittot6b947802016-12-08 17:56:54 +01005558 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5559 group->sgc->capacity;
5560 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5561 group->sgc->capacity;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005562
5563 if (local_group) {
Vincent Guittot6b947802016-12-08 17:56:54 +01005564 this_runnable_load = runnable_load;
5565 this_avg_load = avg_load;
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005566 this_spare = max_spare_cap;
5567 } else {
Vincent Guittot6b947802016-12-08 17:56:54 +01005568 if (min_runnable_load > (runnable_load + imbalance)) {
5569 /*
5570 * The runnable load is significantly smaller
5571 * so we can pick this new cpu
5572 */
5573 min_runnable_load = runnable_load;
5574 min_avg_load = avg_load;
5575 idlest = group;
5576 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5577 (100*min_avg_load > imbalance_scale*avg_load)) {
5578 /*
5579 * The runnable loads are close so take the
5580 * blocked load into account through avg_load.
5581 */
5582 min_avg_load = avg_load;
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005583 idlest = group;
5584 }
5585
5586 if (most_spare < max_spare_cap) {
5587 most_spare = max_spare_cap;
5588 most_spare_sg = group;
5589 }
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005590 }
5591 } while (group = group->next, group != sd->groups);
5592
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005593 /*
5594 * The cross-over point between using spare capacity or least load
5595 * is too conservative for high utilization tasks on partially
5596 * utilized systems if we require spare_capacity > task_util(p),
5597 * so we allow for some task stuffing by using
5598 * spare_capacity > task_util(p)/2.
Vincent Guittotf519a3f2016-12-08 17:56:53 +01005599 *
5600 * Spare capacity can't be used for fork because the utilization has
5601 * not been set yet, we must first select a rq to compute the initial
5602 * utilization.
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005603 */
Vincent Guittotf519a3f2016-12-08 17:56:53 +01005604 if (sd_flag & SD_BALANCE_FORK)
5605 goto skip_spare;
5606
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005607 if (this_spare > task_util(p) / 2 &&
Vincent Guittot6b947802016-12-08 17:56:54 +01005608 imbalance_scale*this_spare > 100*most_spare)
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005609 return NULL;
Vincent Guittot6b947802016-12-08 17:56:54 +01005610
5611 if (most_spare > task_util(p) / 2)
Morten Rasmussen6a0b19c2016-10-14 14:41:08 +01005612 return most_spare_sg;
5613
Vincent Guittotf519a3f2016-12-08 17:56:53 +01005614skip_spare:
Vincent Guittot6b947802016-12-08 17:56:54 +01005615 if (!idlest)
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005616 return NULL;
Vincent Guittot6b947802016-12-08 17:56:54 +01005617
5618 if (min_runnable_load > (this_runnable_load + imbalance))
5619 return NULL;
5620
5621 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5622 (100*this_avg_load < imbalance_scale*min_avg_load))
5623 return NULL;
5624
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005625 return idlest;
5626}
5627
5628/*
5629 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5630 */
5631static int
5632find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5633{
5634 unsigned long load, min_load = ULONG_MAX;
Nicolas Pitre83a0a962014-09-04 11:32:10 -04005635 unsigned int min_exit_latency = UINT_MAX;
5636 u64 latest_idle_timestamp = 0;
5637 int least_loaded_cpu = this_cpu;
5638 int shallowest_idle_cpu = -1;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005639 int i;
5640
Morten Rasmusseneaecf412016-06-22 18:03:14 +01005641 /* Check if we have any choice: */
5642 if (group->group_weight == 1)
Peter Zijlstraae4df9d2017-05-01 11:03:12 +02005643 return cpumask_first(sched_group_span(group));
Morten Rasmusseneaecf412016-06-22 18:03:14 +01005644
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005645 /* Traverse only the allowed CPUs */
Peter Zijlstraae4df9d2017-05-01 11:03:12 +02005646 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
Nicolas Pitre83a0a962014-09-04 11:32:10 -04005647 if (idle_cpu(i)) {
5648 struct rq *rq = cpu_rq(i);
5649 struct cpuidle_state *idle = idle_get_state(rq);
5650 if (idle && idle->exit_latency < min_exit_latency) {
5651 /*
5652 * We give priority to a CPU whose idle state
5653 * has the smallest exit latency irrespective
5654 * of any idle timestamp.
5655 */
5656 min_exit_latency = idle->exit_latency;
5657 latest_idle_timestamp = rq->idle_stamp;
5658 shallowest_idle_cpu = i;
5659 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5660 rq->idle_stamp > latest_idle_timestamp) {
5661 /*
5662 * If equal or no active idle state, then
5663 * the most recently idled CPU might have
5664 * a warmer cache.
5665 */
5666 latest_idle_timestamp = rq->idle_stamp;
5667 shallowest_idle_cpu = i;
5668 }
Yao Dongdong9f967422014-10-28 04:08:06 +00005669 } else if (shallowest_idle_cpu == -1) {
Viresh Kumarc7132dd2017-05-24 10:59:54 +05305670 load = weighted_cpuload(cpu_rq(i));
Nicolas Pitre83a0a962014-09-04 11:32:10 -04005671 if (load < min_load || (load == min_load && i == this_cpu)) {
5672 min_load = load;
5673 least_loaded_cpu = i;
5674 }
Gregory Haskinse7693a32008-01-25 21:08:09 +01005675 }
5676 }
5677
Nicolas Pitre83a0a962014-09-04 11:32:10 -04005678 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005679}
Gregory Haskinse7693a32008-01-25 21:08:09 +01005680
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005681#ifdef CONFIG_SCHED_SMT
5682
5683static inline void set_idle_cores(int cpu, int val)
5684{
5685 struct sched_domain_shared *sds;
5686
5687 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5688 if (sds)
5689 WRITE_ONCE(sds->has_idle_cores, val);
5690}
5691
5692static inline bool test_idle_cores(int cpu, bool def)
5693{
5694 struct sched_domain_shared *sds;
5695
5696 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5697 if (sds)
5698 return READ_ONCE(sds->has_idle_cores);
5699
5700 return def;
5701}
5702
5703/*
5704 * Scans the local SMT mask to see if the entire core is idle, and records this
5705 * information in sd_llc_shared->has_idle_cores.
5706 *
5707 * Since SMT siblings share all cache levels, inspecting this limited remote
5708 * state should be fairly cheap.
5709 */
Peter Zijlstra1b568f02016-05-09 10:38:41 +02005710void __update_idle_core(struct rq *rq)
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005711{
5712 int core = cpu_of(rq);
5713 int cpu;
5714
5715 rcu_read_lock();
5716 if (test_idle_cores(core, true))
5717 goto unlock;
5718
5719 for_each_cpu(cpu, cpu_smt_mask(core)) {
5720 if (cpu == core)
5721 continue;
5722
5723 if (!idle_cpu(cpu))
5724 goto unlock;
5725 }
5726
5727 set_idle_cores(core, 1);
5728unlock:
5729 rcu_read_unlock();
5730}
5731
5732/*
5733 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5734 * there are no idle cores left in the system; tracked through
5735 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5736 */
5737static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5738{
5739 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
Peter Zijlstrac743f0a2017-04-14 14:20:05 +02005740 int core, cpu;
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005741
Peter Zijlstra1b568f02016-05-09 10:38:41 +02005742 if (!static_branch_likely(&sched_smt_present))
5743 return -1;
5744
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005745 if (!test_idle_cores(target, false))
5746 return -1;
5747
Ingo Molnar0c98d342017-02-05 15:38:10 +01005748 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005749
Peter Zijlstrac743f0a2017-04-14 14:20:05 +02005750 for_each_cpu_wrap(core, cpus, target) {
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005751 bool idle = true;
5752
5753 for_each_cpu(cpu, cpu_smt_mask(core)) {
5754 cpumask_clear_cpu(cpu, cpus);
5755 if (!idle_cpu(cpu))
5756 idle = false;
5757 }
5758
5759 if (idle)
5760 return core;
5761 }
5762
5763 /*
5764 * Failed to find an idle core; stop looking for one.
5765 */
5766 set_idle_cores(target, 0);
5767
5768 return -1;
5769}
5770
5771/*
5772 * Scan the local SMT mask for idle CPUs.
5773 */
5774static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5775{
5776 int cpu;
5777
Peter Zijlstra1b568f02016-05-09 10:38:41 +02005778 if (!static_branch_likely(&sched_smt_present))
5779 return -1;
5780
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005781 for_each_cpu(cpu, cpu_smt_mask(target)) {
Ingo Molnar0c98d342017-02-05 15:38:10 +01005782 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005783 continue;
5784 if (idle_cpu(cpu))
5785 return cpu;
5786 }
5787
5788 return -1;
5789}
5790
5791#else /* CONFIG_SCHED_SMT */
5792
5793static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5794{
5795 return -1;
5796}
5797
5798static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5799{
5800 return -1;
5801}
5802
5803#endif /* CONFIG_SCHED_SMT */
5804
5805/*
5806 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5807 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5808 * average idle time for this rq (as found in rq->avg_idle).
5809 */
5810static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5811{
Wanpeng Li9cfb38a2016-10-09 08:04:03 +08005812 struct sched_domain *this_sd;
Peter Zijlstra1ad3aaf2017-05-17 12:53:50 +02005813 u64 avg_cost, avg_idle;
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005814 u64 time, cost;
5815 s64 delta;
Peter Zijlstra1ad3aaf2017-05-17 12:53:50 +02005816 int cpu, nr = INT_MAX;
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005817
Wanpeng Li9cfb38a2016-10-09 08:04:03 +08005818 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5819 if (!this_sd)
5820 return -1;
5821
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005822 /*
5823 * Due to large variance we need a large fuzz factor; hackbench in
5824 * particularly is sensitive here.
5825 */
Peter Zijlstra1ad3aaf2017-05-17 12:53:50 +02005826 avg_idle = this_rq()->avg_idle / 512;
5827 avg_cost = this_sd->avg_scan_cost + 1;
5828
5829 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005830 return -1;
5831
Peter Zijlstra1ad3aaf2017-05-17 12:53:50 +02005832 if (sched_feat(SIS_PROP)) {
5833 u64 span_avg = sd->span_weight * avg_idle;
5834 if (span_avg > 4*avg_cost)
5835 nr = div_u64(span_avg, avg_cost);
5836 else
5837 nr = 4;
5838 }
5839
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005840 time = local_clock();
5841
Peter Zijlstrac743f0a2017-04-14 14:20:05 +02005842 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
Peter Zijlstra1ad3aaf2017-05-17 12:53:50 +02005843 if (!--nr)
5844 return -1;
Ingo Molnar0c98d342017-02-05 15:38:10 +01005845 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005846 continue;
5847 if (idle_cpu(cpu))
5848 break;
5849 }
5850
5851 time = local_clock() - time;
5852 cost = this_sd->avg_scan_cost;
5853 delta = (s64)(time - cost) / 8;
5854 this_sd->avg_scan_cost += delta;
5855
5856 return cpu;
5857}
5858
5859/*
5860 * Try and locate an idle core/thread in the LLC cache domain.
Peter Zijlstraa50bde52009-11-12 15:55:28 +01005861 */
Morten Rasmussen772bd008c2016-06-22 18:03:13 +01005862static int select_idle_sibling(struct task_struct *p, int prev, int target)
Peter Zijlstraa50bde52009-11-12 15:55:28 +01005863{
Suresh Siddha99bd5e22010-03-31 16:47:45 -07005864 struct sched_domain *sd;
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005865 int i;
Mike Galbraithe0a79f52013-01-28 12:19:25 +01005866
5867 if (idle_cpu(target))
5868 return target;
Peter Zijlstraa50bde52009-11-12 15:55:28 +01005869
5870 /*
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005871 * If the previous cpu is cache affine and idle, don't be stupid.
Peter Zijlstraa50bde52009-11-12 15:55:28 +01005872 */
Morten Rasmussen772bd008c2016-06-22 18:03:13 +01005873 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5874 return prev;
Peter Zijlstraa50bde52009-11-12 15:55:28 +01005875
Peter Zijlstra518cd622011-12-07 15:07:31 +01005876 sd = rcu_dereference(per_cpu(sd_llc, target));
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005877 if (!sd)
5878 return target;
Morten Rasmussen772bd008c2016-06-22 18:03:13 +01005879
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005880 i = select_idle_core(p, sd, target);
5881 if ((unsigned)i < nr_cpumask_bits)
Gregory Haskinse7693a32008-01-25 21:08:09 +01005882 return i;
Ingo Molnar098fb9d2008-03-16 20:36:10 +01005883
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005884 i = select_idle_cpu(p, sd, target);
5885 if ((unsigned)i < nr_cpumask_bits)
5886 return i;
Mike Galbraith970e1782012-06-12 05:18:32 +02005887
Peter Zijlstra10e2f1a2016-05-09 10:38:05 +02005888 i = select_idle_smt(p, sd, target);
5889 if ((unsigned)i < nr_cpumask_bits)
5890 return i;
Linus Torvalds37407ea2012-09-16 12:29:43 -07005891
Peter Zijlstraa50bde52009-11-12 15:55:28 +01005892 return target;
5893}
Dietmar Eggemann231678b2015-08-14 17:23:13 +01005894
Vincent Guittot8bb5b002015-03-04 08:48:47 +01005895/*
Dietmar Eggemann9e91d612015-08-14 17:23:12 +01005896 * cpu_util returns the amount of capacity of a CPU that is used by CFS
Vincent Guittot8bb5b002015-03-04 08:48:47 +01005897 * tasks. The unit of the return value must be the one of capacity so we can
Dietmar Eggemann9e91d612015-08-14 17:23:12 +01005898 * compare the utilization with the capacity of the CPU that is available for
5899 * CFS task (ie cpu_capacity).
Dietmar Eggemann231678b2015-08-14 17:23:13 +01005900 *
5901 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5902 * recent utilization of currently non-runnable tasks on a CPU. It represents
5903 * the amount of utilization of a CPU in the range [0..capacity_orig] where
5904 * capacity_orig is the cpu_capacity available at the highest frequency
5905 * (arch_scale_freq_capacity()).
5906 * The utilization of a CPU converges towards a sum equal to or less than the
5907 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5908 * the running time on this CPU scaled by capacity_curr.
5909 *
5910 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5911 * higher than capacity_orig because of unfortunate rounding in
5912 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5913 * the average stabilizes with the new running time. We need to check that the
5914 * utilization stays within the range of [0..capacity_orig] and cap it if
5915 * necessary. Without utilization capping, a group could be seen as overloaded
5916 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5917 * available capacity. We allow utilization to overshoot capacity_curr (but not
5918 * capacity_orig) as it useful for predicting the capacity required after task
5919 * migrations (scheduler-driven DVFS).
Vincent Guittot8bb5b002015-03-04 08:48:47 +01005920 */
Dietmar Eggemann9e91d612015-08-14 17:23:12 +01005921static int cpu_util(int cpu)
Vincent Guittot8bb5b002015-03-04 08:48:47 +01005922{
Dietmar Eggemann9e91d612015-08-14 17:23:12 +01005923 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
Vincent Guittot8bb5b002015-03-04 08:48:47 +01005924 unsigned long capacity = capacity_orig_of(cpu);
5925
Dietmar Eggemann231678b2015-08-14 17:23:13 +01005926 return (util >= capacity) ? capacity : util;
Vincent Guittot8bb5b002015-03-04 08:48:47 +01005927}
Peter Zijlstraa50bde52009-11-12 15:55:28 +01005928
Morten Rasmussen32731632016-07-25 14:34:26 +01005929static inline int task_util(struct task_struct *p)
5930{
5931 return p->se.avg.util_avg;
5932}
5933
5934/*
Morten Rasmussen104cb162016-10-14 14:41:07 +01005935 * cpu_util_wake: Compute cpu utilization with any contributions from
5936 * the waking task p removed.
5937 */
5938static int cpu_util_wake(int cpu, struct task_struct *p)
5939{
5940 unsigned long util, capacity;
5941
5942 /* Task has no contribution or is new */
5943 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5944 return cpu_util(cpu);
5945
5946 capacity = capacity_orig_of(cpu);
5947 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5948
5949 return (util >= capacity) ? capacity : util;
5950}
5951
5952/*
Morten Rasmussen32731632016-07-25 14:34:26 +01005953 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5954 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5955 *
5956 * In that case WAKE_AFFINE doesn't make sense and we'll let
5957 * BALANCE_WAKE sort things out.
5958 */
5959static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5960{
5961 long min_cap, max_cap;
5962
5963 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5964 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5965
5966 /* Minimum capacity is close to max, no need to abort wake_affine */
5967 if (max_cap - min_cap < max_cap >> 3)
5968 return 0;
5969
Morten Rasmussen104cb162016-10-14 14:41:07 +01005970 /* Bring task utilization in sync with prev_cpu */
5971 sync_entity_load_avg(&p->se);
5972
Morten Rasmussen32731632016-07-25 14:34:26 +01005973 return min_cap * 1024 < task_util(p) * capacity_margin;
5974}
5975
Peter Zijlstraa50bde52009-11-12 15:55:28 +01005976/*
Morten Rasmussende91b9c2014-02-18 14:14:24 +00005977 * select_task_rq_fair: Select target runqueue for the waking task in domains
5978 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5979 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005980 *
Morten Rasmussende91b9c2014-02-18 14:14:24 +00005981 * Balances load by selecting the idlest cpu in the idlest group, or under
5982 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005983 *
Morten Rasmussende91b9c2014-02-18 14:14:24 +00005984 * Returns the target cpu number.
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005985 *
5986 * preempt must be disabled.
5987 */
Peter Zijlstra0017d732010-03-24 18:34:10 +01005988static int
Peter Zijlstraac66f542013-10-07 11:29:16 +01005989select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
Peter Zijlstraaaee1202009-09-10 13:36:25 +02005990{
Peter Zijlstra29cd8ba2009-09-17 09:01:14 +02005991 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
Peter Zijlstrac88d5912009-09-10 13:50:02 +02005992 int cpu = smp_processor_id();
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02005993 int new_cpu = prev_cpu;
Suresh Siddha99bd5e22010-03-31 16:47:45 -07005994 int want_affine = 0;
Peter Zijlstra5158f4e2009-09-16 13:46:59 +02005995 int sync = wake_flags & WF_SYNC;
Gregory Haskinse7693a32008-01-25 21:08:09 +01005996
Peter Zijlstrac58d25f2016-05-12 09:19:59 +02005997 if (sd_flag & SD_BALANCE_WAKE) {
5998 record_wakee(p);
Morten Rasmussen32731632016-07-25 14:34:26 +01005999 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
Ingo Molnar0c98d342017-02-05 15:38:10 +01006000 && cpumask_test_cpu(cpu, &p->cpus_allowed);
Peter Zijlstrac58d25f2016-05-12 09:19:59 +02006001 }
Gregory Haskinse7693a32008-01-25 21:08:09 +01006002
Peter Zijlstradce840a2011-04-07 14:09:50 +02006003 rcu_read_lock();
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006004 for_each_domain(cpu, tmp) {
Peter Zijlstrae4f42882009-12-16 18:04:34 +01006005 if (!(tmp->flags & SD_LOAD_BALANCE))
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02006006 break;
Peter Zijlstrae4f42882009-12-16 18:04:34 +01006007
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006008 /*
Suresh Siddha99bd5e22010-03-31 16:47:45 -07006009 * If both cpu and prev_cpu are part of this domain,
6010 * cpu is a valid SD_WAKE_AFFINE target.
Peter Zijlstrafe3bcfe2009-11-12 15:55:29 +01006011 */
Suresh Siddha99bd5e22010-03-31 16:47:45 -07006012 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6013 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6014 affine_sd = tmp;
Alex Shif03542a2012-07-26 08:55:34 +08006015 break;
Peter Zijlstrac88d5912009-09-10 13:50:02 +02006016 }
6017
Alex Shif03542a2012-07-26 08:55:34 +08006018 if (tmp->flags & sd_flag)
Peter Zijlstra29cd8ba2009-09-17 09:01:14 +02006019 sd = tmp;
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02006020 else if (!want_affine)
6021 break;
Peter Zijlstrac88d5912009-09-10 13:50:02 +02006022 }
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006023
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02006024 if (affine_sd) {
6025 sd = NULL; /* Prefer wake_affine over balance flags */
Rik van Riel7d894e62017-06-23 12:55:28 -04006026 if (cpu == prev_cpu)
6027 goto pick_cpu;
6028
6029 if (wake_affine(affine_sd, p, prev_cpu, sync))
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02006030 new_cpu = cpu;
Mike Galbraith8b911ac2010-03-11 17:17:16 +01006031 }
Peter Zijlstra3b640892009-09-16 13:44:33 +02006032
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02006033 if (!sd) {
Rik van Riel7d894e62017-06-23 12:55:28 -04006034 pick_cpu:
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02006035 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
Morten Rasmussen772bd008c2016-06-22 18:03:13 +01006036 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
Mike Galbraith63b0e9e2015-07-14 17:39:50 +02006037
6038 } else while (sd) {
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006039 struct sched_group *group;
Peter Zijlstrac88d5912009-09-10 13:50:02 +02006040 int weight;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006041
Peter Zijlstra0763a662009-09-14 19:37:39 +02006042 if (!(sd->flags & sd_flag)) {
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006043 sd = sd->child;
6044 continue;
6045 }
6046
Vincent Guittotc44f2a02013-10-18 13:52:21 +02006047 group = find_idlest_group(sd, p, cpu, sd_flag);
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006048 if (!group) {
6049 sd = sd->child;
6050 continue;
6051 }
6052
Peter Zijlstrad7c33c42009-09-11 12:45:38 +02006053 new_cpu = find_idlest_cpu(group, p, cpu);
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006054 if (new_cpu == -1 || new_cpu == cpu) {
6055 /* Now try balancing at a lower domain level of cpu */
6056 sd = sd->child;
6057 continue;
6058 }
6059
6060 /* Now try balancing at a lower domain level of new_cpu */
6061 cpu = new_cpu;
Peter Zijlstra669c55e2010-04-16 14:59:29 +02006062 weight = sd->span_weight;
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006063 sd = NULL;
6064 for_each_domain(cpu, tmp) {
Peter Zijlstra669c55e2010-04-16 14:59:29 +02006065 if (weight <= tmp->span_weight)
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006066 break;
Peter Zijlstra0763a662009-09-14 19:37:39 +02006067 if (tmp->flags & sd_flag)
Peter Zijlstraaaee1202009-09-10 13:36:25 +02006068 sd = tmp;
6069 }
6070 /* while loop will break here if sd == NULL */
Gregory Haskinse7693a32008-01-25 21:08:09 +01006071 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02006072 rcu_read_unlock();
Gregory Haskinse7693a32008-01-25 21:08:09 +01006073
Peter Zijlstrac88d5912009-09-10 13:50:02 +02006074 return new_cpu;
Gregory Haskinse7693a32008-01-25 21:08:09 +01006075}
Paul Turner0a74bef2012-10-04 13:18:30 +02006076
6077/*
6078 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6079 * cfs_rq_of(p) references at time of call are still valid and identify the
Byungchul Park525628c2015-11-18 09:34:59 +09006080 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
Paul Turner0a74bef2012-10-04 13:18:30 +02006081 */
xiaofeng.yan5a4fd032015-09-23 14:55:59 +08006082static void migrate_task_rq_fair(struct task_struct *p)
Paul Turner0a74bef2012-10-04 13:18:30 +02006083{
Paul Turneraff3e492012-10-04 13:18:30 +02006084 /*
Peter Zijlstra59efa0b2016-05-10 18:24:37 +02006085 * As blocked tasks retain absolute vruntime the migration needs to
6086 * deal with this by subtracting the old and adding the new
6087 * min_vruntime -- the latter is done by enqueue_entity() when placing
6088 * the task on the new runqueue.
6089 */
6090 if (p->state == TASK_WAKING) {
6091 struct sched_entity *se = &p->se;
6092 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6093 u64 min_vruntime;
6094
6095#ifndef CONFIG_64BIT
6096 u64 min_vruntime_copy;
6097
6098 do {
6099 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6100 smp_rmb();
6101 min_vruntime = cfs_rq->min_vruntime;
6102 } while (min_vruntime != min_vruntime_copy);
6103#else
6104 min_vruntime = cfs_rq->min_vruntime;
6105#endif
6106
6107 se->vruntime -= min_vruntime;
6108 }
6109
6110 /*
Yuyang Du9d89c252015-07-15 08:04:37 +08006111 * We are supposed to update the task to "current" time, then its up to date
6112 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6113 * what current time is, so simply throw away the out-of-date time. This
6114 * will result in the wakee task is less decayed, but giving the wakee more
6115 * load sounds not bad.
Paul Turneraff3e492012-10-04 13:18:30 +02006116 */
Yuyang Du9d89c252015-07-15 08:04:37 +08006117 remove_entity_load_avg(&p->se);
6118
6119 /* Tell new CPU we are migrated */
6120 p->se.avg.last_update_time = 0;
Ben Segall3944a922014-05-15 15:59:20 -07006121
6122 /* We have migrated, no longer consider this task hot */
Yuyang Du9d89c252015-07-15 08:04:37 +08006123 p->se.exec_start = 0;
Paul Turner0a74bef2012-10-04 13:18:30 +02006124}
Yuyang Du12695572015-07-15 08:04:40 +08006125
6126static void task_dead_fair(struct task_struct *p)
6127{
6128 remove_entity_load_avg(&p->se);
6129}
Gregory Haskinse7693a32008-01-25 21:08:09 +01006130#endif /* CONFIG_SMP */
6131
Peter Zijlstrae52fb7c2009-01-14 12:39:19 +01006132static unsigned long
6133wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
Peter Zijlstra0bbd3332008-04-19 19:44:57 +02006134{
6135 unsigned long gran = sysctl_sched_wakeup_granularity;
6136
6137 /*
Peter Zijlstrae52fb7c2009-01-14 12:39:19 +01006138 * Since its curr running now, convert the gran from real-time
6139 * to virtual-time in his units.
Mike Galbraith13814d42010-03-11 17:17:04 +01006140 *
6141 * By using 'se' instead of 'curr' we penalize light tasks, so
6142 * they get preempted easier. That is, if 'se' < 'curr' then
6143 * the resulting gran will be larger, therefore penalizing the
6144 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6145 * be smaller, again penalizing the lighter task.
6146 *
6147 * This is especially important for buddies when the leftmost
6148 * task is higher priority than the buddy.
Peter Zijlstra0bbd3332008-04-19 19:44:57 +02006149 */
Shaohua Lif4ad9bd2011-04-08 12:53:09 +08006150 return calc_delta_fair(gran, se);
Peter Zijlstra0bbd3332008-04-19 19:44:57 +02006151}
6152
6153/*
Peter Zijlstra464b7522008-10-24 11:06:15 +02006154 * Should 'se' preempt 'curr'.
6155 *
6156 * |s1
6157 * |s2
6158 * |s3
6159 * g
6160 * |<--->|c
6161 *
6162 * w(c, s1) = -1
6163 * w(c, s2) = 0
6164 * w(c, s3) = 1
6165 *
6166 */
6167static int
6168wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6169{
6170 s64 gran, vdiff = curr->vruntime - se->vruntime;
6171
6172 if (vdiff <= 0)
6173 return -1;
6174
Peter Zijlstrae52fb7c2009-01-14 12:39:19 +01006175 gran = wakeup_gran(curr, se);
Peter Zijlstra464b7522008-10-24 11:06:15 +02006176 if (vdiff > gran)
6177 return 1;
6178
6179 return 0;
6180}
6181
Peter Zijlstra02479092008-11-04 21:25:10 +01006182static void set_last_buddy(struct sched_entity *se)
6183{
Venkatesh Pallipadi69c80f32011-04-13 18:21:09 -07006184 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6185 return;
6186
Daniel Axtensc5ae3662017-05-11 06:11:39 +10006187 for_each_sched_entity(se) {
6188 if (SCHED_WARN_ON(!se->on_rq))
6189 return;
Venkatesh Pallipadi69c80f32011-04-13 18:21:09 -07006190 cfs_rq_of(se)->last = se;
Daniel Axtensc5ae3662017-05-11 06:11:39 +10006191 }
Peter Zijlstra02479092008-11-04 21:25:10 +01006192}
6193
6194static void set_next_buddy(struct sched_entity *se)
6195{
Venkatesh Pallipadi69c80f32011-04-13 18:21:09 -07006196 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6197 return;
6198
Daniel Axtensc5ae3662017-05-11 06:11:39 +10006199 for_each_sched_entity(se) {
6200 if (SCHED_WARN_ON(!se->on_rq))
6201 return;
Venkatesh Pallipadi69c80f32011-04-13 18:21:09 -07006202 cfs_rq_of(se)->next = se;
Daniel Axtensc5ae3662017-05-11 06:11:39 +10006203 }
Peter Zijlstra02479092008-11-04 21:25:10 +01006204}
6205
Rik van Rielac53db52011-02-01 09:51:03 -05006206static void set_skip_buddy(struct sched_entity *se)
6207{
Venkatesh Pallipadi69c80f32011-04-13 18:21:09 -07006208 for_each_sched_entity(se)
6209 cfs_rq_of(se)->skip = se;
Rik van Rielac53db52011-02-01 09:51:03 -05006210}
6211
Peter Zijlstra464b7522008-10-24 11:06:15 +02006212/*
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006213 * Preempt the current task with a newly woken task if needed:
6214 */
Peter Zijlstra5a9b86f2009-09-16 13:47:58 +02006215static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006216{
6217 struct task_struct *curr = rq->curr;
Srivatsa Vaddagiri8651a862007-10-15 17:00:12 +02006218 struct sched_entity *se = &curr->se, *pse = &p->se;
Mike Galbraith03e89e42008-12-16 08:45:30 +01006219 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
Mike Galbraithf685cea2009-10-23 23:09:22 +02006220 int scale = cfs_rq->nr_running >= sched_nr_latency;
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07006221 int next_buddy_marked = 0;
Mike Galbraith03e89e42008-12-16 08:45:30 +01006222
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01006223 if (unlikely(se == pse))
6224 return;
6225
Paul Turner5238cdd2011-07-21 09:43:37 -07006226 /*
Kirill Tkhai163122b2014-08-20 13:48:29 +04006227 * This is possible from callers such as attach_tasks(), in which we
Paul Turner5238cdd2011-07-21 09:43:37 -07006228 * unconditionally check_prempt_curr() after an enqueue (which may have
6229 * lead to a throttle). This both saves work and prevents false
6230 * next-buddy nomination below.
6231 */
6232 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6233 return;
6234
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07006235 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
Mike Galbraith3cb63d52009-09-11 12:01:17 +02006236 set_next_buddy(pse);
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07006237 next_buddy_marked = 1;
6238 }
Peter Zijlstra57fdc262008-09-23 15:33:45 +02006239
Bharata B Raoaec0a512008-08-28 14:42:49 +05306240 /*
6241 * We can come here with TIF_NEED_RESCHED already set from new task
6242 * wake up path.
Paul Turner5238cdd2011-07-21 09:43:37 -07006243 *
6244 * Note: this also catches the edge-case of curr being in a throttled
6245 * group (e.g. via set_curr_task), since update_curr() (in the
6246 * enqueue of curr) will have resulted in resched being set. This
6247 * prevents us from potentially nominating it as a false LAST_BUDDY
6248 * below.
Bharata B Raoaec0a512008-08-28 14:42:49 +05306249 */
6250 if (test_tsk_need_resched(curr))
6251 return;
6252
Darren Harta2f5c9a2011-02-22 13:04:33 -08006253 /* Idle tasks are by definition preempted by non-idle tasks. */
6254 if (unlikely(curr->policy == SCHED_IDLE) &&
6255 likely(p->policy != SCHED_IDLE))
6256 goto preempt;
6257
Ingo Molnar91c234b2007-10-15 17:00:18 +02006258 /*
Darren Harta2f5c9a2011-02-22 13:04:33 -08006259 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6260 * is driven by the tick):
Ingo Molnar91c234b2007-10-15 17:00:18 +02006261 */
Ingo Molnar8ed92e52012-10-14 14:28:50 +02006262 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
Ingo Molnar91c234b2007-10-15 17:00:18 +02006263 return;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006264
Peter Zijlstra3a7e73a2009-11-28 18:51:02 +01006265 find_matching_se(&se, &pse);
Paul Turner9bbd7372011-07-05 19:07:21 -07006266 update_curr(cfs_rq_of(se));
Peter Zijlstra3a7e73a2009-11-28 18:51:02 +01006267 BUG_ON(!pse);
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07006268 if (wakeup_preempt_entity(se, pse) == 1) {
6269 /*
6270 * Bias pick_next to pick the sched entity that is
6271 * triggering this preemption.
6272 */
6273 if (!next_buddy_marked)
6274 set_next_buddy(pse);
Peter Zijlstra3a7e73a2009-11-28 18:51:02 +01006275 goto preempt;
Venkatesh Pallipadi2f368252011-04-14 10:30:53 -07006276 }
Jupyung Leea65ac742009-11-17 18:51:40 +09006277
Peter Zijlstra3a7e73a2009-11-28 18:51:02 +01006278 return;
6279
6280preempt:
Kirill Tkhai88751252014-06-29 00:03:57 +04006281 resched_curr(rq);
Peter Zijlstra3a7e73a2009-11-28 18:51:02 +01006282 /*
6283 * Only set the backward buddy when the current task is still
6284 * on the rq. This can happen when a wakeup gets interleaved
6285 * with schedule on the ->pre_schedule() or idle_balance()
6286 * point, either of which can * drop the rq lock.
6287 *
6288 * Also, during early boot the idle thread is in the fair class,
6289 * for obvious reasons its a bad idea to schedule back to it.
6290 */
6291 if (unlikely(!se->on_rq || curr == rq->idle))
6292 return;
6293
6294 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6295 set_last_buddy(se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006296}
6297
Peter Zijlstra606dba22012-02-11 06:05:00 +01006298static struct task_struct *
Matt Flemingd8ac8972016-09-21 14:38:10 +01006299pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006300{
6301 struct cfs_rq *cfs_rq = &rq->cfs;
6302 struct sched_entity *se;
Peter Zijlstra678d5712012-02-11 06:05:00 +01006303 struct task_struct *p;
Peter Zijlstra37e117c2014-02-14 12:25:08 +01006304 int new_tasks;
Peter Zijlstra678d5712012-02-11 06:05:00 +01006305
Peter Zijlstra6e831252014-02-11 16:11:48 +01006306again:
Peter Zijlstra678d5712012-02-11 06:05:00 +01006307 if (!cfs_rq->nr_running)
Peter Zijlstra38033c32014-01-23 20:32:21 +01006308 goto idle;
Peter Zijlstra678d5712012-02-11 06:05:00 +01006309
Viresh Kumar9674f5c2017-05-24 10:59:55 +05306310#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra3f1d2a32014-02-12 10:49:30 +01006311 if (prev->sched_class != &fair_sched_class)
Peter Zijlstra678d5712012-02-11 06:05:00 +01006312 goto simple;
6313
6314 /*
6315 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6316 * likely that a next task is from the same cgroup as the current.
6317 *
6318 * Therefore attempt to avoid putting and setting the entire cgroup
6319 * hierarchy, only change the part that actually changes.
6320 */
6321
6322 do {
6323 struct sched_entity *curr = cfs_rq->curr;
6324
6325 /*
6326 * Since we got here without doing put_prev_entity() we also
6327 * have to consider cfs_rq->curr. If it is still a runnable
6328 * entity, update_curr() will update its vruntime, otherwise
6329 * forget we've ever seen it.
6330 */
Ben Segall54d27362015-04-06 15:28:10 -07006331 if (curr) {
6332 if (curr->on_rq)
6333 update_curr(cfs_rq);
6334 else
6335 curr = NULL;
Peter Zijlstra678d5712012-02-11 06:05:00 +01006336
Ben Segall54d27362015-04-06 15:28:10 -07006337 /*
6338 * This call to check_cfs_rq_runtime() will do the
6339 * throttle and dequeue its entity in the parent(s).
Viresh Kumar9674f5c2017-05-24 10:59:55 +05306340 * Therefore the nr_running test will indeed
Ben Segall54d27362015-04-06 15:28:10 -07006341 * be correct.
6342 */
Viresh Kumar9674f5c2017-05-24 10:59:55 +05306343 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6344 cfs_rq = &rq->cfs;
6345
6346 if (!cfs_rq->nr_running)
6347 goto idle;
6348
Ben Segall54d27362015-04-06 15:28:10 -07006349 goto simple;
Viresh Kumar9674f5c2017-05-24 10:59:55 +05306350 }
Ben Segall54d27362015-04-06 15:28:10 -07006351 }
Peter Zijlstra678d5712012-02-11 06:05:00 +01006352
6353 se = pick_next_entity(cfs_rq, curr);
6354 cfs_rq = group_cfs_rq(se);
6355 } while (cfs_rq);
6356
6357 p = task_of(se);
6358
6359 /*
6360 * Since we haven't yet done put_prev_entity and if the selected task
6361 * is a different task than we started out with, try and touch the
6362 * least amount of cfs_rqs.
6363 */
6364 if (prev != p) {
6365 struct sched_entity *pse = &prev->se;
6366
6367 while (!(cfs_rq = is_same_group(se, pse))) {
6368 int se_depth = se->depth;
6369 int pse_depth = pse->depth;
6370
6371 if (se_depth <= pse_depth) {
6372 put_prev_entity(cfs_rq_of(pse), pse);
6373 pse = parent_entity(pse);
6374 }
6375 if (se_depth >= pse_depth) {
6376 set_next_entity(cfs_rq_of(se), se);
6377 se = parent_entity(se);
6378 }
6379 }
6380
6381 put_prev_entity(cfs_rq, pse);
6382 set_next_entity(cfs_rq, se);
6383 }
6384
6385 if (hrtick_enabled(rq))
6386 hrtick_start_fair(rq, p);
6387
6388 return p;
6389simple:
Peter Zijlstra678d5712012-02-11 06:05:00 +01006390#endif
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006391
Peter Zijlstra3f1d2a32014-02-12 10:49:30 +01006392 put_prev_task(rq, prev);
Peter Zijlstra606dba22012-02-11 06:05:00 +01006393
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006394 do {
Peter Zijlstra678d5712012-02-11 06:05:00 +01006395 se = pick_next_entity(cfs_rq, NULL);
Peter Zijlstraf4b67552008-11-04 21:25:07 +01006396 set_next_entity(cfs_rq, se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006397 cfs_rq = group_cfs_rq(se);
6398 } while (cfs_rq);
6399
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01006400 p = task_of(se);
Peter Zijlstra678d5712012-02-11 06:05:00 +01006401
Mike Galbraithb39e66e2011-11-22 15:20:07 +01006402 if (hrtick_enabled(rq))
6403 hrtick_start_fair(rq, p);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01006404
6405 return p;
Peter Zijlstra38033c32014-01-23 20:32:21 +01006406
6407idle:
Matt Fleming46f69fa2016-09-21 14:38:12 +01006408 new_tasks = idle_balance(rq, rf);
6409
Peter Zijlstra37e117c2014-02-14 12:25:08 +01006410 /*
6411 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6412 * possible for any higher priority task to appear. In that case we
6413 * must re-start the pick_next_entity() loop.
6414 */
Kirill Tkhaie4aa3582014-03-06 13:31:55 +04006415 if (new_tasks < 0)
Peter Zijlstra37e117c2014-02-14 12:25:08 +01006416 return RETRY_TASK;
6417
Kirill Tkhaie4aa3582014-03-06 13:31:55 +04006418 if (new_tasks > 0)
Peter Zijlstra38033c32014-01-23 20:32:21 +01006419 goto again;
Peter Zijlstra38033c32014-01-23 20:32:21 +01006420
6421 return NULL;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006422}
6423
6424/*
6425 * Account for a descheduled task:
6426 */
Ingo Molnar31ee5292007-08-09 11:16:49 +02006427static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006428{
6429 struct sched_entity *se = &prev->se;
6430 struct cfs_rq *cfs_rq;
6431
6432 for_each_sched_entity(se) {
6433 cfs_rq = cfs_rq_of(se);
Ingo Molnarab6cde22007-08-09 11:16:48 +02006434 put_prev_entity(cfs_rq, se);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006435 }
6436}
6437
Rik van Rielac53db52011-02-01 09:51:03 -05006438/*
6439 * sched_yield() is very simple
6440 *
6441 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6442 */
6443static void yield_task_fair(struct rq *rq)
6444{
6445 struct task_struct *curr = rq->curr;
6446 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6447 struct sched_entity *se = &curr->se;
6448
6449 /*
6450 * Are we the only task in the tree?
6451 */
6452 if (unlikely(rq->nr_running == 1))
6453 return;
6454
6455 clear_buddies(cfs_rq, se);
6456
6457 if (curr->policy != SCHED_BATCH) {
6458 update_rq_clock(rq);
6459 /*
6460 * Update run-time statistics of the 'current'.
6461 */
6462 update_curr(cfs_rq);
Mike Galbraith916671c2011-11-22 15:21:26 +01006463 /*
6464 * Tell update_rq_clock() that we've just updated,
6465 * so we don't do microscopic update in schedule()
6466 * and double the fastpath cost.
6467 */
Peter Zijlstra9edfbfe2015-01-05 11:18:11 +01006468 rq_clock_skip_update(rq, true);
Rik van Rielac53db52011-02-01 09:51:03 -05006469 }
6470
6471 set_skip_buddy(se);
6472}
6473
Mike Galbraithd95f4122011-02-01 09:50:51 -05006474static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6475{
6476 struct sched_entity *se = &p->se;
6477
Paul Turner5238cdd2011-07-21 09:43:37 -07006478 /* throttled hierarchies are not runnable */
6479 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
Mike Galbraithd95f4122011-02-01 09:50:51 -05006480 return false;
6481
6482 /* Tell the scheduler that we'd really like pse to run next. */
6483 set_next_buddy(se);
6484
Mike Galbraithd95f4122011-02-01 09:50:51 -05006485 yield_task_fair(rq);
6486
6487 return true;
6488}
6489
Peter Williams681f3e62007-10-24 18:23:51 +02006490#ifdef CONFIG_SMP
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006491/**************************************************
Peter Zijlstrae9c84cb2012-07-03 13:53:26 +02006492 * Fair scheduling class load-balancing methods.
6493 *
6494 * BASICS
6495 *
6496 * The purpose of load-balancing is to achieve the same basic fairness the
6497 * per-cpu scheduler provides, namely provide a proportional amount of compute
6498 * time to each task. This is expressed in the following equation:
6499 *
6500 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6501 *
6502 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6503 * W_i,0 is defined as:
6504 *
6505 * W_i,0 = \Sum_j w_i,j (2)
6506 *
6507 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
Yuyang Du1c3de5e2016-03-30 07:07:51 +08006508 * is derived from the nice value as per sched_prio_to_weight[].
Peter Zijlstrae9c84cb2012-07-03 13:53:26 +02006509 *
6510 * The weight average is an exponential decay average of the instantaneous
6511 * weight:
6512 *
6513 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6514 *
Nicolas Pitreced549f2014-05-26 18:19:38 -04006515 * C_i is the compute capacity of cpu i, typically it is the
Peter Zijlstrae9c84cb2012-07-03 13:53:26 +02006516 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6517 * can also include other factors [XXX].
6518 *
6519 * To achieve this balance we define a measure of imbalance which follows
6520 * directly from (1):
6521 *
Nicolas Pitreced549f2014-05-26 18:19:38 -04006522 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
Peter Zijlstrae9c84cb2012-07-03 13:53:26 +02006523 *
6524 * We them move tasks around to minimize the imbalance. In the continuous
6525 * function space it is obvious this converges, in the discrete case we get
6526 * a few fun cases generally called infeasible weight scenarios.
6527 *
6528 * [XXX expand on:
6529 * - infeasible weights;
6530 * - local vs global optima in the discrete case. ]
6531 *
6532 *
6533 * SCHED DOMAINS
6534 *
6535 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6536 * for all i,j solution, we create a tree of cpus that follows the hardware
6537 * topology where each level pairs two lower groups (or better). This results
6538 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6539 * tree to only the first of the previous level and we decrease the frequency
6540 * of load-balance at each level inv. proportional to the number of cpus in
6541 * the groups.
6542 *
6543 * This yields:
6544 *
6545 * log_2 n 1 n
6546 * \Sum { --- * --- * 2^i } = O(n) (5)
6547 * i = 0 2^i 2^i
6548 * `- size of each group
6549 * | | `- number of cpus doing load-balance
6550 * | `- freq
6551 * `- sum over all levels
6552 *
6553 * Coupled with a limit on how many tasks we can migrate every balance pass,
6554 * this makes (5) the runtime complexity of the balancer.
6555 *
6556 * An important property here is that each CPU is still (indirectly) connected
6557 * to every other cpu in at most O(log n) steps:
6558 *
6559 * The adjacency matrix of the resulting graph is given by:
6560 *
Byungchul Park97a71422015-07-05 18:33:48 +09006561 * log_2 n
Peter Zijlstrae9c84cb2012-07-03 13:53:26 +02006562 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6563 * k = 0
6564 *
6565 * And you'll find that:
6566 *
6567 * A^(log_2 n)_i,j != 0 for all i,j (7)
6568 *
6569 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6570 * The task movement gives a factor of O(m), giving a convergence complexity
6571 * of:
6572 *
6573 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6574 *
6575 *
6576 * WORK CONSERVING
6577 *
6578 * In order to avoid CPUs going idle while there's still work to do, new idle
6579 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6580 * tree itself instead of relying on other CPUs to bring it work.
6581 *
6582 * This adds some complexity to both (5) and (8) but it reduces the total idle
6583 * time.
6584 *
6585 * [XXX more?]
6586 *
6587 *
6588 * CGROUPS
6589 *
6590 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6591 *
6592 * s_k,i
6593 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6594 * S_k
6595 *
6596 * Where
6597 *
6598 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6599 *
6600 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6601 *
6602 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6603 * property.
6604 *
6605 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6606 * rewrite all of this once again.]
Byungchul Park97a71422015-07-05 18:33:48 +09006607 */
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02006608
Hiroshi Shimamotoed387b72012-01-31 11:40:32 +09006609static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6610
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01006611enum fbq_type { regular, remote, all };
6612
Peter Zijlstraddcdf6e2012-02-22 19:27:40 +01006613#define LBF_ALL_PINNED 0x01
Peter Zijlstra367456c2012-02-20 21:49:09 +01006614#define LBF_NEED_BREAK 0x02
Peter Zijlstra62633222013-08-19 12:41:09 +02006615#define LBF_DST_PINNED 0x04
6616#define LBF_SOME_PINNED 0x08
Peter Zijlstraddcdf6e2012-02-22 19:27:40 +01006617
6618struct lb_env {
6619 struct sched_domain *sd;
6620
Peter Zijlstraddcdf6e2012-02-22 19:27:40 +01006621 struct rq *src_rq;
Prashanth Nageshappa85c1e7d2012-06-19 17:47:34 +05306622 int src_cpu;
Peter Zijlstraddcdf6e2012-02-22 19:27:40 +01006623
6624 int dst_cpu;
6625 struct rq *dst_rq;
6626
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05306627 struct cpumask *dst_grpmask;
6628 int new_dst_cpu;
Peter Zijlstraddcdf6e2012-02-22 19:27:40 +01006629 enum cpu_idle_type idle;
Peter Zijlstrabd939f42012-05-02 14:20:37 +02006630 long imbalance;
Michael Wangb94031302012-07-12 16:10:13 +08006631 /* The set of CPUs under consideration for load-balancing */
6632 struct cpumask *cpus;
6633
Peter Zijlstraddcdf6e2012-02-22 19:27:40 +01006634 unsigned int flags;
Peter Zijlstra367456c2012-02-20 21:49:09 +01006635
6636 unsigned int loop;
6637 unsigned int loop_break;
6638 unsigned int loop_max;
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01006639
6640 enum fbq_type fbq_type;
Kirill Tkhai163122b2014-08-20 13:48:29 +04006641 struct list_head tasks;
Peter Zijlstraddcdf6e2012-02-22 19:27:40 +01006642};
6643
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006644/*
Peter Zijlstra029632f2011-10-25 10:00:11 +02006645 * Is this task likely cache-hot:
6646 */
Hillf Danton5d5e2b12014-06-10 10:58:43 +02006647static int task_hot(struct task_struct *p, struct lb_env *env)
Peter Zijlstra029632f2011-10-25 10:00:11 +02006648{
6649 s64 delta;
6650
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006651 lockdep_assert_held(&env->src_rq->lock);
6652
Peter Zijlstra029632f2011-10-25 10:00:11 +02006653 if (p->sched_class != &fair_sched_class)
6654 return 0;
6655
6656 if (unlikely(p->policy == SCHED_IDLE))
6657 return 0;
6658
6659 /*
6660 * Buddy candidates are cache hot:
6661 */
Hillf Danton5d5e2b12014-06-10 10:58:43 +02006662 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
Peter Zijlstra029632f2011-10-25 10:00:11 +02006663 (&p->se == cfs_rq_of(&p->se)->next ||
6664 &p->se == cfs_rq_of(&p->se)->last))
6665 return 1;
6666
6667 if (sysctl_sched_migration_cost == -1)
6668 return 1;
6669 if (sysctl_sched_migration_cost == 0)
6670 return 0;
6671
Hillf Danton5d5e2b12014-06-10 10:58:43 +02006672 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
Peter Zijlstra029632f2011-10-25 10:00:11 +02006673
6674 return delta < (s64)sysctl_sched_migration_cost;
6675}
6676
Mel Gorman3a7053b2013-10-07 11:29:00 +01006677#ifdef CONFIG_NUMA_BALANCING
Rik van Rielc1ceac62015-05-14 22:59:36 -04006678/*
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306679 * Returns 1, if task migration degrades locality
6680 * Returns 0, if task migration improves locality i.e migration preferred.
6681 * Returns -1, if task migration is not affected by locality.
Rik van Rielc1ceac62015-05-14 22:59:36 -04006682 */
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306683static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
Mel Gorman3a7053b2013-10-07 11:29:00 +01006684{
Rik van Rielb1ad0652014-05-15 13:03:06 -04006685 struct numa_group *numa_group = rcu_dereference(p->numa_group);
Rik van Rielc1ceac62015-05-14 22:59:36 -04006686 unsigned long src_faults, dst_faults;
Mel Gorman3a7053b2013-10-07 11:29:00 +01006687 int src_nid, dst_nid;
6688
Srikar Dronamraju2a595722015-08-11 21:54:21 +05306689 if (!static_branch_likely(&sched_numa_balancing))
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306690 return -1;
6691
Srikar Dronamrajuc3b9bc52015-08-11 16:30:12 +05306692 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306693 return -1;
Mel Gorman7a0f3082013-10-07 11:29:01 +01006694
6695 src_nid = cpu_to_node(env->src_cpu);
6696 dst_nid = cpu_to_node(env->dst_cpu);
6697
Mel Gorman83e1d2c2013-10-07 11:29:27 +01006698 if (src_nid == dst_nid)
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306699 return -1;
Mel Gorman7a0f3082013-10-07 11:29:01 +01006700
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306701 /* Migrating away from the preferred node is always bad. */
6702 if (src_nid == p->numa_preferred_nid) {
6703 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6704 return 1;
6705 else
6706 return -1;
6707 }
Mel Gorman83e1d2c2013-10-07 11:29:27 +01006708
Rik van Rielc1ceac62015-05-14 22:59:36 -04006709 /* Encourage migration to the preferred node. */
6710 if (dst_nid == p->numa_preferred_nid)
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306711 return 0;
Rik van Rielc1ceac62015-05-14 22:59:36 -04006712
Rik van Riel739294f2017-06-23 12:55:27 -04006713 /* Leaving a core idle is often worse than degrading locality. */
6714 if (env->idle != CPU_NOT_IDLE)
6715 return -1;
6716
Rik van Rielc1ceac62015-05-14 22:59:36 -04006717 if (numa_group) {
6718 src_faults = group_faults(p, src_nid);
6719 dst_faults = group_faults(p, dst_nid);
6720 } else {
6721 src_faults = task_faults(p, src_nid);
6722 dst_faults = task_faults(p, dst_nid);
6723 }
6724
6725 return dst_faults < src_faults;
Mel Gorman7a0f3082013-10-07 11:29:01 +01006726}
6727
Mel Gorman3a7053b2013-10-07 11:29:00 +01006728#else
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306729static inline int migrate_degrades_locality(struct task_struct *p,
Mel Gorman3a7053b2013-10-07 11:29:00 +01006730 struct lb_env *env)
6731{
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306732 return -1;
Mel Gorman7a0f3082013-10-07 11:29:01 +01006733}
Mel Gorman3a7053b2013-10-07 11:29:00 +01006734#endif
6735
Peter Zijlstra029632f2011-10-25 10:00:11 +02006736/*
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006737 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6738 */
6739static
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01006740int can_migrate_task(struct task_struct *p, struct lb_env *env)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006741{
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306742 int tsk_cache_hot;
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006743
6744 lockdep_assert_held(&env->src_rq->lock);
6745
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006746 /*
6747 * We do not migrate tasks that are:
Joonsoo Kimd3198082013-04-23 17:27:40 +09006748 * 1) throttled_lb_pair, or
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006749 * 2) cannot be migrated to this CPU due to cpus_allowed, or
Joonsoo Kimd3198082013-04-23 17:27:40 +09006750 * 3) running (obviously), or
6751 * 4) are cache-hot on their current CPU.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006752 */
Joonsoo Kimd3198082013-04-23 17:27:40 +09006753 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6754 return 0;
6755
Ingo Molnar0c98d342017-02-05 15:38:10 +01006756 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
Joonsoo Kime02e60c2013-04-23 17:27:42 +09006757 int cpu;
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05306758
Josh Poimboeufae928822016-06-17 12:43:24 -05006759 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05306760
Peter Zijlstra62633222013-08-19 12:41:09 +02006761 env->flags |= LBF_SOME_PINNED;
6762
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05306763 /*
6764 * Remember if this task can be migrated to any other cpu in
6765 * our sched_group. We may want to revisit it if we couldn't
6766 * meet load balance goals by pulling other tasks on src_cpu.
6767 *
Jeffrey Hugo65a44332017-06-07 13:18:57 -06006768 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
6769 * already computed one in current iteration.
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05306770 */
Jeffrey Hugo65a44332017-06-07 13:18:57 -06006771 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05306772 return 0;
6773
Joonsoo Kime02e60c2013-04-23 17:27:42 +09006774 /* Prevent to re-select dst_cpu via env's cpus */
6775 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
Ingo Molnar0c98d342017-02-05 15:38:10 +01006776 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
Peter Zijlstra62633222013-08-19 12:41:09 +02006777 env->flags |= LBF_DST_PINNED;
Joonsoo Kime02e60c2013-04-23 17:27:42 +09006778 env->new_dst_cpu = cpu;
6779 break;
6780 }
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05306781 }
Joonsoo Kime02e60c2013-04-23 17:27:42 +09006782
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006783 return 0;
6784 }
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05306785
6786 /* Record that we found atleast one task that could run on dst_cpu */
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01006787 env->flags &= ~LBF_ALL_PINNED;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006788
Peter Zijlstraddcdf6e2012-02-22 19:27:40 +01006789 if (task_running(env->src_rq, p)) {
Josh Poimboeufae928822016-06-17 12:43:24 -05006790 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006791 return 0;
6792 }
6793
6794 /*
6795 * Aggressive migration if:
Mel Gorman3a7053b2013-10-07 11:29:00 +01006796 * 1) destination numa is preferred
6797 * 2) task is cache cold, or
6798 * 3) too many balance attempts have failed.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006799 */
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306800 tsk_cache_hot = migrate_degrades_locality(p, env);
6801 if (tsk_cache_hot == -1)
6802 tsk_cache_hot = task_hot(p, env);
Mel Gorman3a7053b2013-10-07 11:29:00 +01006803
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306804 if (tsk_cache_hot <= 0 ||
Kirill Tkhai7a96c232014-09-22 22:36:12 +04006805 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
Srikar Dronamraju2a1ed242015-06-16 17:25:59 +05306806 if (tsk_cache_hot == 1) {
Josh Poimboeufae928822016-06-17 12:43:24 -05006807 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6808 schedstat_inc(p->se.statistics.nr_forced_migrations);
Mel Gorman3a7053b2013-10-07 11:29:00 +01006809 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006810 return 1;
6811 }
6812
Josh Poimboeufae928822016-06-17 12:43:24 -05006813 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
Zhang Hang4e2dcb72013-04-10 14:04:55 +08006814 return 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006815}
6816
Peter Zijlstra897c3952009-12-17 17:45:42 +01006817/*
Kirill Tkhai163122b2014-08-20 13:48:29 +04006818 * detach_task() -- detach the task for the migration specified in env
Peter Zijlstra897c3952009-12-17 17:45:42 +01006819 */
Kirill Tkhai163122b2014-08-20 13:48:29 +04006820static void detach_task(struct task_struct *p, struct lb_env *env)
6821{
6822 lockdep_assert_held(&env->src_rq->lock);
6823
Kirill Tkhai163122b2014-08-20 13:48:29 +04006824 p->on_rq = TASK_ON_RQ_MIGRATING;
Peter Zijlstra5704ac02017-02-21 17:15:21 +01006825 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
Kirill Tkhai163122b2014-08-20 13:48:29 +04006826 set_task_cpu(p, env->dst_cpu);
6827}
6828
6829/*
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006830 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
Peter Zijlstra897c3952009-12-17 17:45:42 +01006831 * part of active balancing operations within "domain".
Peter Zijlstra897c3952009-12-17 17:45:42 +01006832 *
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006833 * Returns a task if successful and NULL otherwise.
Peter Zijlstra897c3952009-12-17 17:45:42 +01006834 */
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006835static struct task_struct *detach_one_task(struct lb_env *env)
Peter Zijlstra897c3952009-12-17 17:45:42 +01006836{
6837 struct task_struct *p, *n;
Peter Zijlstra897c3952009-12-17 17:45:42 +01006838
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006839 lockdep_assert_held(&env->src_rq->lock);
6840
Peter Zijlstra367456c2012-02-20 21:49:09 +01006841 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
Peter Zijlstra367456c2012-02-20 21:49:09 +01006842 if (!can_migrate_task(p, env))
6843 continue;
Peter Zijlstra897c3952009-12-17 17:45:42 +01006844
Kirill Tkhai163122b2014-08-20 13:48:29 +04006845 detach_task(p, env);
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006846
Peter Zijlstra367456c2012-02-20 21:49:09 +01006847 /*
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006848 * Right now, this is only the second place where
Kirill Tkhai163122b2014-08-20 13:48:29 +04006849 * lb_gained[env->idle] is updated (other is detach_tasks)
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006850 * so we can safely collect stats here rather than
Kirill Tkhai163122b2014-08-20 13:48:29 +04006851 * inside detach_tasks().
Peter Zijlstra367456c2012-02-20 21:49:09 +01006852 */
Josh Poimboeufae928822016-06-17 12:43:24 -05006853 schedstat_inc(env->sd->lb_gained[env->idle]);
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006854 return p;
Peter Zijlstra897c3952009-12-17 17:45:42 +01006855 }
Kirill Tkhaie5673f22014-08-20 13:48:01 +04006856 return NULL;
Peter Zijlstra897c3952009-12-17 17:45:42 +01006857}
6858
Peter Zijlstraeb953082012-04-17 13:38:40 +02006859static const unsigned int sched_nr_migrate_break = 32;
6860
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006861/*
Kirill Tkhai163122b2014-08-20 13:48:29 +04006862 * detach_tasks() -- tries to detach up to imbalance weighted load from
6863 * busiest_rq, as part of a balancing operation within domain "sd".
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006864 *
Kirill Tkhai163122b2014-08-20 13:48:29 +04006865 * Returns number of detached tasks if successful and 0 otherwise.
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006866 */
Kirill Tkhai163122b2014-08-20 13:48:29 +04006867static int detach_tasks(struct lb_env *env)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006868{
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006869 struct list_head *tasks = &env->src_rq->cfs_tasks;
6870 struct task_struct *p;
Peter Zijlstra367456c2012-02-20 21:49:09 +01006871 unsigned long load;
Kirill Tkhai163122b2014-08-20 13:48:29 +04006872 int detached = 0;
6873
6874 lockdep_assert_held(&env->src_rq->lock);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006875
Peter Zijlstrabd939f42012-05-02 14:20:37 +02006876 if (env->imbalance <= 0)
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006877 return 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006878
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006879 while (!list_empty(tasks)) {
Yuyang Du985d3a42015-07-06 06:11:51 +08006880 /*
6881 * We don't want to steal all, otherwise we may be treated likewise,
6882 * which could at worst lead to a livelock crash.
6883 */
6884 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6885 break;
6886
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006887 p = list_first_entry(tasks, struct task_struct, se.group_node);
6888
Peter Zijlstra367456c2012-02-20 21:49:09 +01006889 env->loop++;
6890 /* We've more or less seen every task there is, call it quits */
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006891 if (env->loop > env->loop_max)
Peter Zijlstra367456c2012-02-20 21:49:09 +01006892 break;
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006893
6894 /* take a breather every nr_migrate tasks */
Peter Zijlstra367456c2012-02-20 21:49:09 +01006895 if (env->loop > env->loop_break) {
Peter Zijlstraeb953082012-04-17 13:38:40 +02006896 env->loop_break += sched_nr_migrate_break;
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01006897 env->flags |= LBF_NEED_BREAK;
Peter Zijlstraee00e662009-12-17 17:25:20 +01006898 break;
Peter Zijlstraa195f002011-09-22 15:30:18 +02006899 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006900
Joonsoo Kimd3198082013-04-23 17:27:40 +09006901 if (!can_migrate_task(p, env))
Peter Zijlstra367456c2012-02-20 21:49:09 +01006902 goto next;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006903
Peter Zijlstra367456c2012-02-20 21:49:09 +01006904 load = task_h_load(p);
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006905
Peter Zijlstraeb953082012-04-17 13:38:40 +02006906 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
Peter Zijlstra367456c2012-02-20 21:49:09 +01006907 goto next;
6908
Peter Zijlstrabd939f42012-05-02 14:20:37 +02006909 if ((load / 2) > env->imbalance)
Peter Zijlstra367456c2012-02-20 21:49:09 +01006910 goto next;
6911
Kirill Tkhai163122b2014-08-20 13:48:29 +04006912 detach_task(p, env);
6913 list_add(&p->se.group_node, &env->tasks);
6914
6915 detached++;
Peter Zijlstrabd939f42012-05-02 14:20:37 +02006916 env->imbalance -= load;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006917
6918#ifdef CONFIG_PREEMPT
Peter Zijlstraee00e662009-12-17 17:25:20 +01006919 /*
6920 * NEWIDLE balancing is a source of latency, so preemptible
Kirill Tkhai163122b2014-08-20 13:48:29 +04006921 * kernels will stop after the first task is detached to minimize
Peter Zijlstraee00e662009-12-17 17:25:20 +01006922 * the critical section.
6923 */
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006924 if (env->idle == CPU_NEWLY_IDLE)
Peter Zijlstraee00e662009-12-17 17:25:20 +01006925 break;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006926#endif
6927
Peter Zijlstraee00e662009-12-17 17:25:20 +01006928 /*
6929 * We only want to steal up to the prescribed amount of
6930 * weighted load.
6931 */
Peter Zijlstrabd939f42012-05-02 14:20:37 +02006932 if (env->imbalance <= 0)
Peter Zijlstraee00e662009-12-17 17:25:20 +01006933 break;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006934
Peter Zijlstra367456c2012-02-20 21:49:09 +01006935 continue;
6936next:
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006937 list_move_tail(&p->se.group_node, tasks);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006938 }
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01006939
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006940 /*
Kirill Tkhai163122b2014-08-20 13:48:29 +04006941 * Right now, this is one of only two places we collect this stat
6942 * so we can safely collect detach_one_task() stats here rather
6943 * than inside detach_one_task().
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006944 */
Josh Poimboeufae928822016-06-17 12:43:24 -05006945 schedstat_add(env->sd->lb_gained[env->idle], detached);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006946
Kirill Tkhai163122b2014-08-20 13:48:29 +04006947 return detached;
6948}
6949
6950/*
6951 * attach_task() -- attach the task detached by detach_task() to its new rq.
6952 */
6953static void attach_task(struct rq *rq, struct task_struct *p)
6954{
6955 lockdep_assert_held(&rq->lock);
6956
6957 BUG_ON(task_rq(p) != rq);
Peter Zijlstra5704ac02017-02-21 17:15:21 +01006958 activate_task(rq, p, ENQUEUE_NOCLOCK);
Joonwoo Park3ea94de2015-11-12 19:38:54 -08006959 p->on_rq = TASK_ON_RQ_QUEUED;
Kirill Tkhai163122b2014-08-20 13:48:29 +04006960 check_preempt_curr(rq, p, 0);
6961}
6962
6963/*
6964 * attach_one_task() -- attaches the task returned from detach_one_task() to
6965 * its new rq.
6966 */
6967static void attach_one_task(struct rq *rq, struct task_struct *p)
6968{
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02006969 struct rq_flags rf;
6970
6971 rq_lock(rq, &rf);
Peter Zijlstra5704ac02017-02-21 17:15:21 +01006972 update_rq_clock(rq);
Kirill Tkhai163122b2014-08-20 13:48:29 +04006973 attach_task(rq, p);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02006974 rq_unlock(rq, &rf);
Kirill Tkhai163122b2014-08-20 13:48:29 +04006975}
6976
6977/*
6978 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6979 * new rq.
6980 */
6981static void attach_tasks(struct lb_env *env)
6982{
6983 struct list_head *tasks = &env->tasks;
6984 struct task_struct *p;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02006985 struct rq_flags rf;
Kirill Tkhai163122b2014-08-20 13:48:29 +04006986
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02006987 rq_lock(env->dst_rq, &rf);
Peter Zijlstra5704ac02017-02-21 17:15:21 +01006988 update_rq_clock(env->dst_rq);
Kirill Tkhai163122b2014-08-20 13:48:29 +04006989
6990 while (!list_empty(tasks)) {
6991 p = list_first_entry(tasks, struct task_struct, se.group_node);
6992 list_del_init(&p->se.group_node);
6993
6994 attach_task(env->dst_rq, p);
6995 }
6996
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02006997 rq_unlock(env->dst_rq, &rf);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01006998}
6999
Peter Zijlstra230059de2009-12-17 17:47:12 +01007000#ifdef CONFIG_FAIR_GROUP_SCHED
Tejun Heoa9e7f652017-04-25 17:43:50 -07007001
7002static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7003{
7004 if (cfs_rq->load.weight)
7005 return false;
7006
7007 if (cfs_rq->avg.load_sum)
7008 return false;
7009
7010 if (cfs_rq->avg.util_sum)
7011 return false;
7012
7013 if (cfs_rq->runnable_load_sum)
7014 return false;
7015
7016 return true;
7017}
7018
Paul Turner48a16752012-10-04 13:18:31 +02007019static void update_blocked_averages(int cpu)
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08007020{
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08007021 struct rq *rq = cpu_rq(cpu);
Tejun Heoa9e7f652017-04-25 17:43:50 -07007022 struct cfs_rq *cfs_rq, *pos;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02007023 struct rq_flags rf;
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08007024
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02007025 rq_lock_irqsave(rq, &rf);
Paul Turner48a16752012-10-04 13:18:31 +02007026 update_rq_clock(rq);
Yuyang Du9d89c252015-07-15 08:04:37 +08007027
Peter Zijlstra9763b672011-07-13 13:09:25 +02007028 /*
7029 * Iterates the task_group tree in a bottom up fashion, see
7030 * list_add_leaf_cfs_rq() for details.
7031 */
Tejun Heoa9e7f652017-04-25 17:43:50 -07007032 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
Vincent Guittotbc427892017-03-17 14:47:22 +01007033 struct sched_entity *se;
7034
Yuyang Du9d89c252015-07-15 08:04:37 +08007035 /* throttled entities do not contribute to load */
7036 if (throttled_hierarchy(cfs_rq))
7037 continue;
Paul Turner48a16752012-10-04 13:18:31 +02007038
Viresh Kumar3a123bb2017-05-24 10:59:56 +05307039 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
Yuyang Du9d89c252015-07-15 08:04:37 +08007040 update_tg_load_avg(cfs_rq, 0);
Vincent Guittot4e516072016-11-08 10:53:46 +01007041
Vincent Guittotbc427892017-03-17 14:47:22 +01007042 /* Propagate pending load changes to the parent, if any: */
7043 se = cfs_rq->tg->se[cpu];
7044 if (se && !skip_blocked_update(se))
7045 update_load_avg(se, 0);
Tejun Heoa9e7f652017-04-25 17:43:50 -07007046
7047 /*
7048 * There can be a lot of idle CPU cgroups. Don't let fully
7049 * decayed cfs_rqs linger on the list.
7050 */
7051 if (cfs_rq_is_decayed(cfs_rq))
7052 list_del_leaf_cfs_rq(cfs_rq);
Yuyang Du9d89c252015-07-15 08:04:37 +08007053 }
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02007054 rq_unlock_irqrestore(rq, &rf);
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08007055}
7056
Peter Zijlstra9763b672011-07-13 13:09:25 +02007057/*
Vladimir Davydov68520792013-07-15 17:49:19 +04007058 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
Peter Zijlstra9763b672011-07-13 13:09:25 +02007059 * This needs to be done in a top-down fashion because the load of a child
7060 * group is a fraction of its parents load.
7061 */
Vladimir Davydov68520792013-07-15 17:49:19 +04007062static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
Peter Zijlstra9763b672011-07-13 13:09:25 +02007063{
Vladimir Davydov68520792013-07-15 17:49:19 +04007064 struct rq *rq = rq_of(cfs_rq);
7065 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
Peter Zijlstraa35b6462012-08-08 21:46:40 +02007066 unsigned long now = jiffies;
Vladimir Davydov68520792013-07-15 17:49:19 +04007067 unsigned long load;
Peter Zijlstraa35b6462012-08-08 21:46:40 +02007068
Vladimir Davydov68520792013-07-15 17:49:19 +04007069 if (cfs_rq->last_h_load_update == now)
Peter Zijlstraa35b6462012-08-08 21:46:40 +02007070 return;
7071
Vladimir Davydov68520792013-07-15 17:49:19 +04007072 cfs_rq->h_load_next = NULL;
7073 for_each_sched_entity(se) {
7074 cfs_rq = cfs_rq_of(se);
7075 cfs_rq->h_load_next = se;
7076 if (cfs_rq->last_h_load_update == now)
7077 break;
7078 }
Peter Zijlstraa35b6462012-08-08 21:46:40 +02007079
Vladimir Davydov68520792013-07-15 17:49:19 +04007080 if (!se) {
Yuyang Du7ea241a2015-07-15 08:04:42 +08007081 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
Vladimir Davydov68520792013-07-15 17:49:19 +04007082 cfs_rq->last_h_load_update = now;
7083 }
7084
7085 while ((se = cfs_rq->h_load_next) != NULL) {
7086 load = cfs_rq->h_load;
Yuyang Du7ea241a2015-07-15 08:04:42 +08007087 load = div64_ul(load * se->avg.load_avg,
7088 cfs_rq_load_avg(cfs_rq) + 1);
Vladimir Davydov68520792013-07-15 17:49:19 +04007089 cfs_rq = group_cfs_rq(se);
7090 cfs_rq->h_load = load;
7091 cfs_rq->last_h_load_update = now;
7092 }
Peter Zijlstra9763b672011-07-13 13:09:25 +02007093}
7094
Peter Zijlstra367456c2012-02-20 21:49:09 +01007095static unsigned long task_h_load(struct task_struct *p)
Peter Zijlstra230059de2009-12-17 17:47:12 +01007096{
Peter Zijlstra367456c2012-02-20 21:49:09 +01007097 struct cfs_rq *cfs_rq = task_cfs_rq(p);
Peter Zijlstra230059de2009-12-17 17:47:12 +01007098
Vladimir Davydov68520792013-07-15 17:49:19 +04007099 update_cfs_rq_h_load(cfs_rq);
Yuyang Du9d89c252015-07-15 08:04:37 +08007100 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
Yuyang Du7ea241a2015-07-15 08:04:42 +08007101 cfs_rq_load_avg(cfs_rq) + 1);
Peter Zijlstra230059de2009-12-17 17:47:12 +01007102}
7103#else
Paul Turner48a16752012-10-04 13:18:31 +02007104static inline void update_blocked_averages(int cpu)
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08007105{
Vincent Guittot6c1d47c2015-07-15 08:04:38 +08007106 struct rq *rq = cpu_rq(cpu);
7107 struct cfs_rq *cfs_rq = &rq->cfs;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02007108 struct rq_flags rf;
Vincent Guittot6c1d47c2015-07-15 08:04:38 +08007109
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02007110 rq_lock_irqsave(rq, &rf);
Vincent Guittot6c1d47c2015-07-15 08:04:38 +08007111 update_rq_clock(rq);
Viresh Kumar3a123bb2017-05-24 10:59:56 +05307112 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02007113 rq_unlock_irqrestore(rq, &rf);
Peter Zijlstra9e3081c2010-11-15 15:47:02 -08007114}
7115
Peter Zijlstra367456c2012-02-20 21:49:09 +01007116static unsigned long task_h_load(struct task_struct *p)
7117{
Yuyang Du9d89c252015-07-15 08:04:37 +08007118 return p->se.avg.load_avg;
Peter Zijlstra230059de2009-12-17 17:47:12 +01007119}
7120#endif
7121
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007122/********** Helpers for find_busiest_group ************************/
Rik van Rielcaeb1782014-07-28 14:16:28 -04007123
7124enum group_type {
7125 group_other = 0,
7126 group_imbalanced,
7127 group_overloaded,
7128};
7129
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007130/*
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007131 * sg_lb_stats - stats of a sched_group required for load_balancing
7132 */
7133struct sg_lb_stats {
7134 unsigned long avg_load; /*Avg load across the CPUs of the group */
7135 unsigned long group_load; /* Total load over the CPUs of the group */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007136 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007137 unsigned long load_per_task;
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007138 unsigned long group_capacity;
Dietmar Eggemann9e91d612015-08-14 17:23:12 +01007139 unsigned long group_util; /* Total utilization of the group */
Peter Zijlstra147c5fc2013-08-19 15:22:57 +02007140 unsigned int sum_nr_running; /* Nr tasks running in the group */
Peter Zijlstra147c5fc2013-08-19 15:22:57 +02007141 unsigned int idle_cpus;
7142 unsigned int group_weight;
Rik van Rielcaeb1782014-07-28 14:16:28 -04007143 enum group_type group_type;
Vincent Guittotea678212015-02-27 16:54:11 +01007144 int group_no_capacity;
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01007145#ifdef CONFIG_NUMA_BALANCING
7146 unsigned int nr_numa_running;
7147 unsigned int nr_preferred_running;
7148#endif
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007149};
7150
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007151/*
7152 * sd_lb_stats - Structure to store the statistics of a sched_domain
7153 * during load balancing.
7154 */
7155struct sd_lb_stats {
7156 struct sched_group *busiest; /* Busiest group in this sd */
7157 struct sched_group *local; /* Local group in this sd */
Peter Zijlstra90001d62017-07-31 17:50:05 +02007158 unsigned long total_running;
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007159 unsigned long total_load; /* Total load of all groups in sd */
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007160 unsigned long total_capacity; /* Total capacity of all groups in sd */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007161 unsigned long avg_load; /* Average load across all groups in sd */
7162
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007163 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
Peter Zijlstra147c5fc2013-08-19 15:22:57 +02007164 struct sg_lb_stats local_stat; /* Statistics of the local group */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007165};
7166
Peter Zijlstra147c5fc2013-08-19 15:22:57 +02007167static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7168{
7169 /*
7170 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7171 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7172 * We must however clear busiest_stat::avg_load because
7173 * update_sd_pick_busiest() reads this before assignment.
7174 */
7175 *sds = (struct sd_lb_stats){
7176 .busiest = NULL,
7177 .local = NULL,
Peter Zijlstra90001d62017-07-31 17:50:05 +02007178 .total_running = 0UL,
Peter Zijlstra147c5fc2013-08-19 15:22:57 +02007179 .total_load = 0UL,
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007180 .total_capacity = 0UL,
Peter Zijlstra147c5fc2013-08-19 15:22:57 +02007181 .busiest_stat = {
7182 .avg_load = 0UL,
Rik van Rielcaeb1782014-07-28 14:16:28 -04007183 .sum_nr_running = 0,
7184 .group_type = group_other,
Peter Zijlstra147c5fc2013-08-19 15:22:57 +02007185 },
7186 };
7187}
7188
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007189/**
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007190 * get_sd_load_idx - Obtain the load index for a given sched domain.
7191 * @sd: The sched_domain whose load_idx is to be obtained.
Kamalesh Babulaled1b7732013-10-13 23:06:15 +05307192 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
Yacine Belkadie69f6182013-07-12 20:45:47 +02007193 *
7194 * Return: The load index.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007195 */
7196static inline int get_sd_load_idx(struct sched_domain *sd,
7197 enum cpu_idle_type idle)
7198{
7199 int load_idx;
7200
7201 switch (idle) {
7202 case CPU_NOT_IDLE:
7203 load_idx = sd->busy_idx;
7204 break;
7205
7206 case CPU_NEWLY_IDLE:
7207 load_idx = sd->newidle_idx;
7208 break;
7209 default:
7210 load_idx = sd->idle_idx;
7211 break;
7212 }
7213
7214 return load_idx;
7215}
7216
Nicolas Pitreced549f2014-05-26 18:19:38 -04007217static unsigned long scale_rt_capacity(int cpu)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007218{
7219 struct rq *rq = cpu_rq(cpu);
Vincent Guittotb5b48602015-02-27 16:54:08 +01007220 u64 total, used, age_stamp, avg;
Peter Zijlstracadefd32014-02-27 10:40:35 +01007221 s64 delta;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007222
Peter Zijlstrab654f7d2012-05-22 14:04:28 +02007223 /*
7224 * Since we're reading these variables without serialization make sure
7225 * we read them once before doing sanity checks on them.
7226 */
Jason Low316c1608d2015-04-28 13:00:20 -07007227 age_stamp = READ_ONCE(rq->age_stamp);
7228 avg = READ_ONCE(rq->rt_avg);
Peter Zijlstracebde6d2015-01-05 11:18:10 +01007229 delta = __rq_clock_broken(rq) - age_stamp;
Venkatesh Pallipadiaa483802010-10-04 17:03:22 -07007230
Peter Zijlstracadefd32014-02-27 10:40:35 +01007231 if (unlikely(delta < 0))
7232 delta = 0;
7233
7234 total = sched_avg_period() + delta;
Peter Zijlstrab654f7d2012-05-22 14:04:28 +02007235
Vincent Guittotb5b48602015-02-27 16:54:08 +01007236 used = div_u64(avg, total);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007237
Vincent Guittotb5b48602015-02-27 16:54:08 +01007238 if (likely(used < SCHED_CAPACITY_SCALE))
7239 return SCHED_CAPACITY_SCALE - used;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007240
Vincent Guittotb5b48602015-02-27 16:54:08 +01007241 return 1;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007242}
7243
Nicolas Pitreced549f2014-05-26 18:19:38 -04007244static void update_cpu_capacity(struct sched_domain *sd, int cpu)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007245{
Morten Rasmussen8cd56012015-08-14 17:23:10 +01007246 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007247 struct sched_group *sdg = sd->groups;
7248
Vincent Guittotca6d75e2015-02-27 16:54:09 +01007249 cpu_rq(cpu)->cpu_capacity_orig = capacity;
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10007250
Nicolas Pitreced549f2014-05-26 18:19:38 -04007251 capacity *= scale_rt_capacity(cpu);
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04007252 capacity >>= SCHED_CAPACITY_SHIFT;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007253
Nicolas Pitreced549f2014-05-26 18:19:38 -04007254 if (!capacity)
7255 capacity = 1;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007256
Nicolas Pitreced549f2014-05-26 18:19:38 -04007257 cpu_rq(cpu)->cpu_capacity = capacity;
7258 sdg->sgc->capacity = capacity;
Morten Rasmussenbf475ce2016-10-14 14:41:09 +01007259 sdg->sgc->min_capacity = capacity;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007260}
7261
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007262void update_group_capacity(struct sched_domain *sd, int cpu)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007263{
7264 struct sched_domain *child = sd->child;
7265 struct sched_group *group, *sdg = sd->groups;
Morten Rasmussenbf475ce2016-10-14 14:41:09 +01007266 unsigned long capacity, min_capacity;
Vincent Guittot4ec44122011-12-12 20:21:08 +01007267 unsigned long interval;
7268
7269 interval = msecs_to_jiffies(sd->balance_interval);
7270 interval = clamp(interval, 1UL, max_load_balance_interval);
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007271 sdg->sgc->next_update = jiffies + interval;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007272
7273 if (!child) {
Nicolas Pitreced549f2014-05-26 18:19:38 -04007274 update_cpu_capacity(sd, cpu);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007275 return;
7276 }
7277
Vincent Guittotdc7ff762015-03-03 11:35:03 +01007278 capacity = 0;
Morten Rasmussenbf475ce2016-10-14 14:41:09 +01007279 min_capacity = ULONG_MAX;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007280
Peter Zijlstra74a5ce22012-05-23 18:00:43 +02007281 if (child->flags & SD_OVERLAP) {
7282 /*
7283 * SD_OVERLAP domains cannot assume that child groups
7284 * span the current group.
7285 */
7286
Peter Zijlstraae4df9d2017-05-01 11:03:12 +02007287 for_each_cpu(cpu, sched_group_span(sdg)) {
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007288 struct sched_group_capacity *sgc;
Srikar Dronamraju9abf24d2013-11-12 22:11:26 +05307289 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra863bffc2013-08-28 11:44:39 +02007290
Srikar Dronamraju9abf24d2013-11-12 22:11:26 +05307291 /*
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007292 * build_sched_domains() -> init_sched_groups_capacity()
Srikar Dronamraju9abf24d2013-11-12 22:11:26 +05307293 * gets here before we've attached the domains to the
7294 * runqueues.
7295 *
Nicolas Pitreced549f2014-05-26 18:19:38 -04007296 * Use capacity_of(), which is set irrespective of domains
7297 * in update_cpu_capacity().
Srikar Dronamraju9abf24d2013-11-12 22:11:26 +05307298 *
Vincent Guittotdc7ff762015-03-03 11:35:03 +01007299 * This avoids capacity from being 0 and
Srikar Dronamraju9abf24d2013-11-12 22:11:26 +05307300 * causing divide-by-zero issues on boot.
Srikar Dronamraju9abf24d2013-11-12 22:11:26 +05307301 */
7302 if (unlikely(!rq->sd)) {
Nicolas Pitreced549f2014-05-26 18:19:38 -04007303 capacity += capacity_of(cpu);
Morten Rasmussenbf475ce2016-10-14 14:41:09 +01007304 } else {
7305 sgc = rq->sd->groups->sgc;
7306 capacity += sgc->capacity;
Srikar Dronamraju9abf24d2013-11-12 22:11:26 +05307307 }
7308
Morten Rasmussenbf475ce2016-10-14 14:41:09 +01007309 min_capacity = min(capacity, min_capacity);
Peter Zijlstra863bffc2013-08-28 11:44:39 +02007310 }
Peter Zijlstra74a5ce22012-05-23 18:00:43 +02007311 } else {
7312 /*
7313 * !SD_OVERLAP domains can assume that child groups
7314 * span the current group.
Byungchul Park97a71422015-07-05 18:33:48 +09007315 */
Peter Zijlstra74a5ce22012-05-23 18:00:43 +02007316
7317 group = child->groups;
7318 do {
Morten Rasmussenbf475ce2016-10-14 14:41:09 +01007319 struct sched_group_capacity *sgc = group->sgc;
7320
7321 capacity += sgc->capacity;
7322 min_capacity = min(sgc->min_capacity, min_capacity);
Peter Zijlstra74a5ce22012-05-23 18:00:43 +02007323 group = group->next;
7324 } while (group != child->groups);
7325 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007326
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007327 sdg->sgc->capacity = capacity;
Morten Rasmussenbf475ce2016-10-14 14:41:09 +01007328 sdg->sgc->min_capacity = min_capacity;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007329}
7330
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10007331/*
Vincent Guittotea678212015-02-27 16:54:11 +01007332 * Check whether the capacity of the rq has been noticeably reduced by side
7333 * activity. The imbalance_pct is used for the threshold.
7334 * Return true is the capacity is reduced
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10007335 */
7336static inline int
Vincent Guittotea678212015-02-27 16:54:11 +01007337check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10007338{
Vincent Guittotea678212015-02-27 16:54:11 +01007339 return ((rq->cpu_capacity * sd->imbalance_pct) <
7340 (rq->cpu_capacity_orig * 100));
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10007341}
7342
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007343/*
7344 * Group imbalance indicates (and tries to solve) the problem where balancing
Ingo Molnar0c98d342017-02-05 15:38:10 +01007345 * groups is inadequate due to ->cpus_allowed constraints.
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007346 *
7347 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7348 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7349 * Something like:
7350 *
Ingo Molnar2b4d5b22016-11-23 07:37:00 +01007351 * { 0 1 2 3 } { 4 5 6 7 }
7352 * * * * *
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007353 *
7354 * If we were to balance group-wise we'd place two tasks in the first group and
7355 * two tasks in the second group. Clearly this is undesired as it will overload
7356 * cpu 3 and leave one of the cpus in the second group unused.
7357 *
7358 * The current solution to this issue is detecting the skew in the first group
Peter Zijlstra62633222013-08-19 12:41:09 +02007359 * by noticing the lower domain failed to reach balance and had difficulty
7360 * moving tasks due to affinity constraints.
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007361 *
7362 * When this is so detected; this group becomes a candidate for busiest; see
Kamalesh Babulaled1b7732013-10-13 23:06:15 +05307363 * update_sd_pick_busiest(). And calculate_imbalance() and
Peter Zijlstra62633222013-08-19 12:41:09 +02007364 * find_busiest_group() avoid some of the usual balance conditions to allow it
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007365 * to create an effective group imbalance.
7366 *
7367 * This is a somewhat tricky proposition since the next run might not find the
7368 * group imbalance and decide the groups need to be balanced again. A most
7369 * subtle and fragile situation.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007370 */
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007371
Peter Zijlstra62633222013-08-19 12:41:09 +02007372static inline int sg_imbalanced(struct sched_group *group)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007373{
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007374 return group->sgc->imbalance;
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007375}
7376
Peter Zijlstrab37d9312013-08-28 11:50:34 +02007377/*
Vincent Guittotea678212015-02-27 16:54:11 +01007378 * group_has_capacity returns true if the group has spare capacity that could
7379 * be used by some tasks.
7380 * We consider that a group has spare capacity if the * number of task is
Dietmar Eggemann9e91d612015-08-14 17:23:12 +01007381 * smaller than the number of CPUs or if the utilization is lower than the
7382 * available capacity for CFS tasks.
Vincent Guittotea678212015-02-27 16:54:11 +01007383 * For the latter, we use a threshold to stabilize the state, to take into
7384 * account the variance of the tasks' load and to return true if the available
7385 * capacity in meaningful for the load balancer.
7386 * As an example, an available capacity of 1% can appear but it doesn't make
7387 * any benefit for the load balance.
Peter Zijlstrab37d9312013-08-28 11:50:34 +02007388 */
Vincent Guittotea678212015-02-27 16:54:11 +01007389static inline bool
7390group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
Peter Zijlstrab37d9312013-08-28 11:50:34 +02007391{
Vincent Guittotea678212015-02-27 16:54:11 +01007392 if (sgs->sum_nr_running < sgs->group_weight)
7393 return true;
Peter Zijlstrab37d9312013-08-28 11:50:34 +02007394
Vincent Guittotea678212015-02-27 16:54:11 +01007395 if ((sgs->group_capacity * 100) >
Dietmar Eggemann9e91d612015-08-14 17:23:12 +01007396 (sgs->group_util * env->sd->imbalance_pct))
Vincent Guittotea678212015-02-27 16:54:11 +01007397 return true;
Peter Zijlstrab37d9312013-08-28 11:50:34 +02007398
Vincent Guittotea678212015-02-27 16:54:11 +01007399 return false;
Peter Zijlstrab37d9312013-08-28 11:50:34 +02007400}
7401
Vincent Guittotea678212015-02-27 16:54:11 +01007402/*
7403 * group_is_overloaded returns true if the group has more tasks than it can
7404 * handle.
7405 * group_is_overloaded is not equals to !group_has_capacity because a group
7406 * with the exact right number of tasks, has no more spare capacity but is not
7407 * overloaded so both group_has_capacity and group_is_overloaded return
7408 * false.
7409 */
7410static inline bool
7411group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
Rik van Rielcaeb1782014-07-28 14:16:28 -04007412{
Vincent Guittotea678212015-02-27 16:54:11 +01007413 if (sgs->sum_nr_running <= sgs->group_weight)
7414 return false;
7415
7416 if ((sgs->group_capacity * 100) <
Dietmar Eggemann9e91d612015-08-14 17:23:12 +01007417 (sgs->group_util * env->sd->imbalance_pct))
Vincent Guittotea678212015-02-27 16:54:11 +01007418 return true;
7419
7420 return false;
7421}
7422
Morten Rasmussen9e0994c2016-10-14 14:41:10 +01007423/*
7424 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7425 * per-CPU capacity than sched_group ref.
7426 */
7427static inline bool
7428group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7429{
7430 return sg->sgc->min_capacity * capacity_margin <
7431 ref->sgc->min_capacity * 1024;
7432}
7433
Leo Yan79a89f92015-09-15 18:56:45 +08007434static inline enum
7435group_type group_classify(struct sched_group *group,
7436 struct sg_lb_stats *sgs)
Vincent Guittotea678212015-02-27 16:54:11 +01007437{
7438 if (sgs->group_no_capacity)
Rik van Rielcaeb1782014-07-28 14:16:28 -04007439 return group_overloaded;
7440
7441 if (sg_imbalanced(group))
7442 return group_imbalanced;
7443
7444 return group_other;
7445}
7446
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007447/**
7448 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7449 * @env: The load balancing environment.
7450 * @group: sched_group whose statistics are to be updated.
7451 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7452 * @local_group: Does group contain this_cpu.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007453 * @sgs: variable to hold the statistics for this group.
Masanari Iidacd3bd4e2014-07-28 12:38:06 +09007454 * @overload: Indicate more than one runnable task for any CPU.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007455 */
7456static inline void update_sg_lb_stats(struct lb_env *env,
7457 struct sched_group *group, int load_idx,
Tim Chen4486edd2014-06-23 12:16:49 -07007458 int local_group, struct sg_lb_stats *sgs,
7459 bool *overload)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007460{
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007461 unsigned long load;
Waiman Longa426f992015-11-25 14:09:38 -05007462 int i, nr_running;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007463
Peter Zijlstrab72ff132013-08-28 10:32:32 +02007464 memset(sgs, 0, sizeof(*sgs));
7465
Peter Zijlstraae4df9d2017-05-01 11:03:12 +02007466 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007467 struct rq *rq = cpu_rq(i);
7468
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007469 /* Bias balancing toward cpus of our domain */
Peter Zijlstra62633222013-08-19 12:41:09 +02007470 if (local_group)
Peter Zijlstra04f733b2012-05-11 00:12:02 +02007471 load = target_load(i, load_idx);
Peter Zijlstra62633222013-08-19 12:41:09 +02007472 else
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007473 load = source_load(i, load_idx);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007474
7475 sgs->group_load += load;
Dietmar Eggemann9e91d612015-08-14 17:23:12 +01007476 sgs->group_util += cpu_util(i);
Vincent Guittot65fdac02014-08-26 13:06:46 +02007477 sgs->sum_nr_running += rq->cfs.h_nr_running;
Tim Chen4486edd2014-06-23 12:16:49 -07007478
Waiman Longa426f992015-11-25 14:09:38 -05007479 nr_running = rq->nr_running;
7480 if (nr_running > 1)
Tim Chen4486edd2014-06-23 12:16:49 -07007481 *overload = true;
7482
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01007483#ifdef CONFIG_NUMA_BALANCING
7484 sgs->nr_numa_running += rq->nr_numa_running;
7485 sgs->nr_preferred_running += rq->nr_preferred_running;
7486#endif
Viresh Kumarc7132dd2017-05-24 10:59:54 +05307487 sgs->sum_weighted_load += weighted_cpuload(rq);
Waiman Longa426f992015-11-25 14:09:38 -05007488 /*
7489 * No need to call idle_cpu() if nr_running is not 0
7490 */
7491 if (!nr_running && idle_cpu(i))
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07007492 sgs->idle_cpus++;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007493 }
7494
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007495 /* Adjust by relative CPU capacity of the group */
7496 sgs->group_capacity = group->sgc->capacity;
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04007497 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007498
Suresh Siddhadd5feea2010-02-23 16:13:52 -08007499 if (sgs->sum_nr_running)
Peter Zijlstra38d0f772013-08-15 19:47:56 +02007500 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007501
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07007502 sgs->group_weight = group->group_weight;
Peter Zijlstrab37d9312013-08-28 11:50:34 +02007503
Vincent Guittotea678212015-02-27 16:54:11 +01007504 sgs->group_no_capacity = group_is_overloaded(env, sgs);
Leo Yan79a89f92015-09-15 18:56:45 +08007505 sgs->group_type = group_classify(group, sgs);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007506}
7507
7508/**
Michael Neuling532cb4c2010-06-08 14:57:02 +10007509 * update_sd_pick_busiest - return 1 on busiest group
Randy Dunlapcd968912012-06-08 13:18:33 -07007510 * @env: The load balancing environment.
Michael Neuling532cb4c2010-06-08 14:57:02 +10007511 * @sds: sched_domain statistics
7512 * @sg: sched_group candidate to be checked for being the busiest
Michael Neulingb6b12292010-06-10 12:06:21 +10007513 * @sgs: sched_group statistics
Michael Neuling532cb4c2010-06-08 14:57:02 +10007514 *
7515 * Determine if @sg is a busier group than the previously selected
7516 * busiest group.
Yacine Belkadie69f6182013-07-12 20:45:47 +02007517 *
7518 * Return: %true if @sg is a busier group than the previously selected
7519 * busiest group. %false otherwise.
Michael Neuling532cb4c2010-06-08 14:57:02 +10007520 */
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007521static bool update_sd_pick_busiest(struct lb_env *env,
Michael Neuling532cb4c2010-06-08 14:57:02 +10007522 struct sd_lb_stats *sds,
7523 struct sched_group *sg,
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007524 struct sg_lb_stats *sgs)
Michael Neuling532cb4c2010-06-08 14:57:02 +10007525{
Rik van Rielcaeb1782014-07-28 14:16:28 -04007526 struct sg_lb_stats *busiest = &sds->busiest_stat;
Michael Neuling532cb4c2010-06-08 14:57:02 +10007527
Rik van Rielcaeb1782014-07-28 14:16:28 -04007528 if (sgs->group_type > busiest->group_type)
Michael Neuling532cb4c2010-06-08 14:57:02 +10007529 return true;
7530
Rik van Rielcaeb1782014-07-28 14:16:28 -04007531 if (sgs->group_type < busiest->group_type)
7532 return false;
7533
7534 if (sgs->avg_load <= busiest->avg_load)
7535 return false;
7536
Morten Rasmussen9e0994c2016-10-14 14:41:10 +01007537 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7538 goto asym_packing;
7539
7540 /*
7541 * Candidate sg has no more than one task per CPU and
7542 * has higher per-CPU capacity. Migrating tasks to less
7543 * capable CPUs may harm throughput. Maximize throughput,
7544 * power/energy consequences are not considered.
7545 */
7546 if (sgs->sum_nr_running <= sgs->group_weight &&
7547 group_smaller_cpu_capacity(sds->local, sg))
7548 return false;
7549
7550asym_packing:
Rik van Rielcaeb1782014-07-28 14:16:28 -04007551 /* This is the busiest node in its class. */
7552 if (!(env->sd->flags & SD_ASYM_PACKING))
Michael Neuling532cb4c2010-06-08 14:57:02 +10007553 return true;
7554
Srikar Dronamraju1f621e02016-04-06 18:47:40 +05307555 /* No ASYM_PACKING if target cpu is already busy */
7556 if (env->idle == CPU_NOT_IDLE)
7557 return true;
Michael Neuling532cb4c2010-06-08 14:57:02 +10007558 /*
Tim Chenafe06ef2016-11-22 12:23:53 -08007559 * ASYM_PACKING needs to move all the work to the highest
7560 * prority CPUs in the group, therefore mark all groups
7561 * of lower priority than ourself as busy.
Michael Neuling532cb4c2010-06-08 14:57:02 +10007562 */
Tim Chenafe06ef2016-11-22 12:23:53 -08007563 if (sgs->sum_nr_running &&
7564 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
Michael Neuling532cb4c2010-06-08 14:57:02 +10007565 if (!sds->busiest)
7566 return true;
7567
Tim Chenafe06ef2016-11-22 12:23:53 -08007568 /* Prefer to move from lowest priority cpu's work */
7569 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7570 sg->asym_prefer_cpu))
Michael Neuling532cb4c2010-06-08 14:57:02 +10007571 return true;
7572 }
7573
7574 return false;
7575}
7576
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01007577#ifdef CONFIG_NUMA_BALANCING
7578static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7579{
7580 if (sgs->sum_nr_running > sgs->nr_numa_running)
7581 return regular;
7582 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7583 return remote;
7584 return all;
7585}
7586
7587static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7588{
7589 if (rq->nr_running > rq->nr_numa_running)
7590 return regular;
7591 if (rq->nr_running > rq->nr_preferred_running)
7592 return remote;
7593 return all;
7594}
7595#else
7596static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7597{
7598 return all;
7599}
7600
7601static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7602{
7603 return regular;
7604}
7605#endif /* CONFIG_NUMA_BALANCING */
7606
Michael Neuling532cb4c2010-06-08 14:57:02 +10007607/**
Hui Kang461819a2011-10-11 23:00:59 -04007608 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
Randy Dunlapcd968912012-06-08 13:18:33 -07007609 * @env: The load balancing environment.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007610 * @sds: variable to hold the statistics for this sched_domain.
7611 */
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01007612static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007613{
Peter Zijlstra90001d62017-07-31 17:50:05 +02007614 struct sched_domain_shared *shared = env->sd->shared;
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007615 struct sched_domain *child = env->sd->child;
7616 struct sched_group *sg = env->sd->groups;
Srikar Dronamraju05b40e02017-03-22 23:27:50 +05307617 struct sg_lb_stats *local = &sds->local_stat;
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007618 struct sg_lb_stats tmp_sgs;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007619 int load_idx, prefer_sibling = 0;
Tim Chen4486edd2014-06-23 12:16:49 -07007620 bool overload = false;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007621
7622 if (child && child->flags & SD_PREFER_SIBLING)
7623 prefer_sibling = 1;
7624
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007625 load_idx = get_sd_load_idx(env->sd, env->idle);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007626
7627 do {
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007628 struct sg_lb_stats *sgs = &tmp_sgs;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007629 int local_group;
7630
Peter Zijlstraae4df9d2017-05-01 11:03:12 +02007631 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007632 if (local_group) {
7633 sds->local = sg;
Srikar Dronamraju05b40e02017-03-22 23:27:50 +05307634 sgs = local;
Peter Zijlstrab72ff132013-08-28 10:32:32 +02007635
7636 if (env->idle != CPU_NEWLY_IDLE ||
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007637 time_after_eq(jiffies, sg->sgc->next_update))
7638 update_group_capacity(env->sd, env->dst_cpu);
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007639 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007640
Tim Chen4486edd2014-06-23 12:16:49 -07007641 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7642 &overload);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007643
Peter Zijlstrab72ff132013-08-28 10:32:32 +02007644 if (local_group)
7645 goto next_group;
7646
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007647 /*
7648 * In case the child domain prefers tasks go to siblings
Vincent Guittotea678212015-02-27 16:54:11 +01007649 * first, lower the sg capacity so that we'll try
Nikhil Rao75dd3212010-10-15 13:12:30 -07007650 * and move all the excess tasks away. We lower the capacity
7651 * of a group only if the local group has the capacity to fit
Vincent Guittotea678212015-02-27 16:54:11 +01007652 * these excess tasks. The extra check prevents the case where
7653 * you always pull from the heaviest group when it is already
7654 * under-utilized (possible with a large weight task outweighs
7655 * the tasks on the system).
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007656 */
Peter Zijlstrab72ff132013-08-28 10:32:32 +02007657 if (prefer_sibling && sds->local &&
Srikar Dronamraju05b40e02017-03-22 23:27:50 +05307658 group_has_capacity(env, local) &&
7659 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
Vincent Guittotea678212015-02-27 16:54:11 +01007660 sgs->group_no_capacity = 1;
Leo Yan79a89f92015-09-15 18:56:45 +08007661 sgs->group_type = group_classify(sg, sgs);
Wanpeng Licb0b9f22014-11-05 07:44:50 +08007662 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007663
Peter Zijlstrab72ff132013-08-28 10:32:32 +02007664 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
Michael Neuling532cb4c2010-06-08 14:57:02 +10007665 sds->busiest = sg;
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007666 sds->busiest_stat = *sgs;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007667 }
7668
Peter Zijlstrab72ff132013-08-28 10:32:32 +02007669next_group:
7670 /* Now, start updating sd_lb_stats */
Peter Zijlstra90001d62017-07-31 17:50:05 +02007671 sds->total_running += sgs->sum_nr_running;
Peter Zijlstrab72ff132013-08-28 10:32:32 +02007672 sds->total_load += sgs->group_load;
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007673 sds->total_capacity += sgs->group_capacity;
Peter Zijlstrab72ff132013-08-28 10:32:32 +02007674
Michael Neuling532cb4c2010-06-08 14:57:02 +10007675 sg = sg->next;
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007676 } while (sg != env->sd->groups);
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01007677
7678 if (env->sd->flags & SD_NUMA)
7679 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
Tim Chen4486edd2014-06-23 12:16:49 -07007680
7681 if (!env->sd->parent) {
7682 /* update overload indicator if we are at root domain */
7683 if (env->dst_rq->rd->overload != overload)
7684 env->dst_rq->rd->overload = overload;
7685 }
7686
Peter Zijlstra90001d62017-07-31 17:50:05 +02007687 if (!shared)
7688 return;
7689
7690 /*
7691 * Since these are sums over groups they can contain some CPUs
7692 * multiple times for the NUMA domains.
7693 *
7694 * Currently only wake_affine_llc() and find_busiest_group()
7695 * uses these numbers, only the last is affected by this problem.
7696 *
7697 * XXX fix that.
7698 */
7699 WRITE_ONCE(shared->nr_running, sds->total_running);
7700 WRITE_ONCE(shared->load, sds->total_load);
7701 WRITE_ONCE(shared->capacity, sds->total_capacity);
Michael Neuling532cb4c2010-06-08 14:57:02 +10007702}
7703
Michael Neuling532cb4c2010-06-08 14:57:02 +10007704/**
7705 * check_asym_packing - Check to see if the group is packed into the
Masanari Iida0ba42a52017-03-07 20:48:02 +09007706 * sched domain.
Michael Neuling532cb4c2010-06-08 14:57:02 +10007707 *
7708 * This is primarily intended to used at the sibling level. Some
7709 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7710 * case of POWER7, it can move to lower SMT modes only when higher
7711 * threads are idle. When in lower SMT modes, the threads will
7712 * perform better since they share less core resources. Hence when we
7713 * have idle threads, we want them to be the higher ones.
7714 *
7715 * This packing function is run on idle threads. It checks to see if
7716 * the busiest CPU in this domain (core in the P7 case) has a higher
7717 * CPU number than the packing function is being run on. Here we are
7718 * assuming lower CPU number will be equivalent to lower a SMT thread
7719 * number.
7720 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02007721 * Return: 1 when packing is required and a task should be moved to
Michael Neulingb6b12292010-06-10 12:06:21 +10007722 * this CPU. The amount of the imbalance is returned in *imbalance.
7723 *
Randy Dunlapcd968912012-06-08 13:18:33 -07007724 * @env: The load balancing environment.
Michael Neuling532cb4c2010-06-08 14:57:02 +10007725 * @sds: Statistics of the sched_domain which is to be packed
Michael Neuling532cb4c2010-06-08 14:57:02 +10007726 */
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007727static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
Michael Neuling532cb4c2010-06-08 14:57:02 +10007728{
7729 int busiest_cpu;
7730
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007731 if (!(env->sd->flags & SD_ASYM_PACKING))
Michael Neuling532cb4c2010-06-08 14:57:02 +10007732 return 0;
7733
Srikar Dronamraju1f621e02016-04-06 18:47:40 +05307734 if (env->idle == CPU_NOT_IDLE)
7735 return 0;
7736
Michael Neuling532cb4c2010-06-08 14:57:02 +10007737 if (!sds->busiest)
7738 return 0;
7739
Tim Chenafe06ef2016-11-22 12:23:53 -08007740 busiest_cpu = sds->busiest->asym_prefer_cpu;
7741 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
Michael Neuling532cb4c2010-06-08 14:57:02 +10007742 return 0;
7743
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007744 env->imbalance = DIV_ROUND_CLOSEST(
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007745 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04007746 SCHED_CAPACITY_SCALE);
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007747
Michael Neuling532cb4c2010-06-08 14:57:02 +10007748 return 1;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007749}
7750
7751/**
7752 * fix_small_imbalance - Calculate the minor imbalance that exists
7753 * amongst the groups of a sched_domain, during
7754 * load balancing.
Randy Dunlapcd968912012-06-08 13:18:33 -07007755 * @env: The load balancing environment.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007756 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007757 */
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007758static inline
7759void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007760{
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007761 unsigned long tmp, capa_now = 0, capa_move = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007762 unsigned int imbn = 2;
Suresh Siddhadd5feea2010-02-23 16:13:52 -08007763 unsigned long scaled_busy_load_per_task;
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007764 struct sg_lb_stats *local, *busiest;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007765
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007766 local = &sds->local_stat;
7767 busiest = &sds->busiest_stat;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007768
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007769 if (!local->sum_nr_running)
7770 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7771 else if (busiest->load_per_task > local->load_per_task)
7772 imbn = 1;
Suresh Siddhadd5feea2010-02-23 16:13:52 -08007773
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007774 scaled_busy_load_per_task =
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04007775 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007776 busiest->group_capacity;
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007777
Vladimir Davydov3029ede2013-09-15 17:49:14 +04007778 if (busiest->avg_load + scaled_busy_load_per_task >=
7779 local->avg_load + (scaled_busy_load_per_task * imbn)) {
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007780 env->imbalance = busiest->load_per_task;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007781 return;
7782 }
7783
7784 /*
7785 * OK, we don't have enough imbalance to justify moving tasks,
Nicolas Pitreced549f2014-05-26 18:19:38 -04007786 * however we may be able to increase total CPU capacity used by
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007787 * moving them.
7788 */
7789
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007790 capa_now += busiest->group_capacity *
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007791 min(busiest->load_per_task, busiest->avg_load);
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007792 capa_now += local->group_capacity *
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007793 min(local->load_per_task, local->avg_load);
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04007794 capa_now /= SCHED_CAPACITY_SCALE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007795
7796 /* Amount of load we'd subtract */
Vincent Guittota2cd4262014-03-11 17:26:06 +01007797 if (busiest->avg_load > scaled_busy_load_per_task) {
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007798 capa_move += busiest->group_capacity *
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007799 min(busiest->load_per_task,
Vincent Guittota2cd4262014-03-11 17:26:06 +01007800 busiest->avg_load - scaled_busy_load_per_task);
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007801 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007802
7803 /* Amount of load we'd add */
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007804 if (busiest->avg_load * busiest->group_capacity <
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04007805 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007806 tmp = (busiest->avg_load * busiest->group_capacity) /
7807 local->group_capacity;
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007808 } else {
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04007809 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007810 local->group_capacity;
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007811 }
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007812 capa_move += local->group_capacity *
Peter Zijlstra3ae11c92013-08-15 20:37:48 +02007813 min(local->load_per_task, local->avg_load + tmp);
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04007814 capa_move /= SCHED_CAPACITY_SCALE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007815
7816 /* Move if we gain throughput */
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007817 if (capa_move > capa_now)
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007818 env->imbalance = busiest->load_per_task;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007819}
7820
7821/**
7822 * calculate_imbalance - Calculate the amount of imbalance present within the
7823 * groups of a given sched_domain during load balance.
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007824 * @env: load balance environment
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007825 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007826 */
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007827static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007828{
Suresh Siddhadd5feea2010-02-23 16:13:52 -08007829 unsigned long max_pull, load_above_capacity = ~0UL;
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007830 struct sg_lb_stats *local, *busiest;
Suresh Siddhadd5feea2010-02-23 16:13:52 -08007831
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007832 local = &sds->local_stat;
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007833 busiest = &sds->busiest_stat;
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007834
Rik van Rielcaeb1782014-07-28 14:16:28 -04007835 if (busiest->group_type == group_imbalanced) {
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007836 /*
7837 * In the group_imb case we cannot rely on group-wide averages
7838 * to ensure cpu-load equilibrium, look at wider averages. XXX
7839 */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007840 busiest->load_per_task =
7841 min(busiest->load_per_task, sds->avg_load);
Suresh Siddhadd5feea2010-02-23 16:13:52 -08007842 }
7843
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007844 /*
Dietmar Eggemann885e5422016-04-29 20:32:39 +01007845 * Avg load of busiest sg can be less and avg load of local sg can
7846 * be greater than avg load across all sgs of sd because avg load
7847 * factors in sg capacity and sgs with smaller group_type are
7848 * skipped when updating the busiest sg:
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007849 */
Vladimir Davydovb1885552013-09-15 17:49:13 +04007850 if (busiest->avg_load <= sds->avg_load ||
7851 local->avg_load >= sds->avg_load) {
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007852 env->imbalance = 0;
7853 return fix_small_imbalance(env, sds);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007854 }
7855
Peter Zijlstra9a5d9ba2014-07-29 17:15:11 +02007856 /*
7857 * If there aren't any idle cpus, avoid creating some.
7858 */
7859 if (busiest->group_type == group_overloaded &&
7860 local->group_type == group_overloaded) {
Peter Zijlstra1be0eb22016-05-06 12:21:23 +02007861 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
Morten Rasmussencfa10332016-04-29 20:32:40 +01007862 if (load_above_capacity > busiest->group_capacity) {
Vincent Guittotea678212015-02-27 16:54:11 +01007863 load_above_capacity -= busiest->group_capacity;
Dietmar Eggemann26656212016-08-10 11:27:27 +01007864 load_above_capacity *= scale_load_down(NICE_0_LOAD);
Morten Rasmussencfa10332016-04-29 20:32:40 +01007865 load_above_capacity /= busiest->group_capacity;
7866 } else
Vincent Guittotea678212015-02-27 16:54:11 +01007867 load_above_capacity = ~0UL;
Suresh Siddhadd5feea2010-02-23 16:13:52 -08007868 }
7869
7870 /*
7871 * We're trying to get all the cpus to the average_load, so we don't
7872 * want to push ourselves above the average load, nor do we wish to
7873 * reduce the max loaded cpu below the average load. At the same time,
Dietmar Eggemann0a9b23c2016-04-29 20:32:38 +01007874 * we also don't want to reduce the group load below the group
7875 * capacity. Thus we look for the minimum possible imbalance.
Suresh Siddhadd5feea2010-02-23 16:13:52 -08007876 */
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007877 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007878
7879 /* How much load to actually move to equalise the imbalance */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007880 env->imbalance = min(
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04007881 max_pull * busiest->group_capacity,
7882 (sds->avg_load - local->avg_load) * local->group_capacity
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04007883 ) / SCHED_CAPACITY_SCALE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007884
7885 /*
7886 * if *imbalance is less than the average load per runnable task
Lucas De Marchi25985ed2011-03-30 22:57:33 -03007887 * there is no guarantee that any tasks will be moved so we'll have
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007888 * a think about bumping its value to force at least one task to be
7889 * moved
7890 */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007891 if (env->imbalance < busiest->load_per_task)
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007892 return fix_small_imbalance(env, sds);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007893}
Nikhil Raofab47622010-10-15 13:12:29 -07007894
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007895/******* find_busiest_group() helpers end here *********************/
7896
7897/**
7898 * find_busiest_group - Returns the busiest group within the sched_domain
Dietmar Eggemann0a9b23c2016-04-29 20:32:38 +01007899 * if there is an imbalance.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007900 *
7901 * Also calculates the amount of weighted load which should be moved
7902 * to restore balance.
7903 *
Randy Dunlapcd968912012-06-08 13:18:33 -07007904 * @env: The load balancing environment.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007905 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02007906 * Return: - The busiest group if imbalance exists.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007907 */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007908static struct sched_group *find_busiest_group(struct lb_env *env)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007909{
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007910 struct sg_lb_stats *local, *busiest;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007911 struct sd_lb_stats sds;
7912
Peter Zijlstra147c5fc2013-08-19 15:22:57 +02007913 init_sd_lb_stats(&sds);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007914
7915 /*
7916 * Compute the various statistics relavent for load balancing at
7917 * this level.
7918 */
Joonsoo Kim23f0d202013-08-06 17:36:42 +09007919 update_sd_lb_stats(env, &sds);
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007920 local = &sds.local_stat;
7921 busiest = &sds.busiest_stat;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007922
Vincent Guittotea678212015-02-27 16:54:11 +01007923 /* ASYM feature bypasses nice load balance check */
Srikar Dronamraju1f621e02016-04-06 18:47:40 +05307924 if (check_asym_packing(env, &sds))
Michael Neuling532cb4c2010-06-08 14:57:02 +10007925 return sds.busiest;
7926
Peter Zijlstracc57aa82011-02-21 18:55:32 +01007927 /* There is no busy sibling group to pull tasks from */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007928 if (!sds.busiest || busiest->sum_nr_running == 0)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007929 goto out_balanced;
7930
Peter Zijlstra90001d62017-07-31 17:50:05 +02007931 /* XXX broken for overlapping NUMA groups */
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04007932 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7933 / sds.total_capacity;
Ken Chenb0432d82011-04-07 17:23:22 -07007934
Peter Zijlstra866ab432011-02-21 18:56:47 +01007935 /*
7936 * If the busiest group is imbalanced the below checks don't
Peter Zijlstra30ce5da2013-08-15 20:29:29 +02007937 * work because they assume all things are equal, which typically
Peter Zijlstra866ab432011-02-21 18:56:47 +01007938 * isn't true due to cpus_allowed constraints and the like.
7939 */
Rik van Rielcaeb1782014-07-28 14:16:28 -04007940 if (busiest->group_type == group_imbalanced)
Peter Zijlstra866ab432011-02-21 18:56:47 +01007941 goto force_balance;
7942
Peter Zijlstracc57aa82011-02-21 18:55:32 +01007943 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
Vincent Guittotea678212015-02-27 16:54:11 +01007944 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7945 busiest->group_no_capacity)
Nikhil Raofab47622010-10-15 13:12:29 -07007946 goto force_balance;
7947
Peter Zijlstracc57aa82011-02-21 18:55:32 +01007948 /*
Zhihui Zhang9c58c792014-09-20 21:24:36 -04007949 * If the local group is busier than the selected busiest group
Peter Zijlstracc57aa82011-02-21 18:55:32 +01007950 * don't try and pull any tasks.
7951 */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007952 if (local->avg_load >= busiest->avg_load)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007953 goto out_balanced;
7954
Peter Zijlstracc57aa82011-02-21 18:55:32 +01007955 /*
7956 * Don't pull any tasks if this group is already above the domain
7957 * average load.
7958 */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007959 if (local->avg_load >= sds.avg_load)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007960 goto out_balanced;
7961
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007962 if (env->idle == CPU_IDLE) {
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07007963 /*
Vincent Guittot43f4d662014-10-01 15:38:55 +02007964 * This cpu is idle. If the busiest group is not overloaded
7965 * and there is no imbalance between this and busiest group
7966 * wrt idle cpus, it is balanced. The imbalance becomes
7967 * significant if the diff is greater than 1 otherwise we
7968 * might end up to just move the imbalance on another group
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07007969 */
Vincent Guittot43f4d662014-10-01 15:38:55 +02007970 if ((busiest->group_type != group_overloaded) &&
7971 (local->idle_cpus <= (busiest->idle_cpus + 1)))
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07007972 goto out_balanced;
Peter Zijlstrac186faf2011-02-21 18:52:53 +01007973 } else {
7974 /*
7975 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7976 * imbalance_pct to be conservative.
7977 */
Joonsoo Kim56cf5152013-08-06 17:36:43 +09007978 if (100 * busiest->avg_load <=
7979 env->sd->imbalance_pct * local->avg_load)
Peter Zijlstrac186faf2011-02-21 18:52:53 +01007980 goto out_balanced;
Suresh Siddhaaae6d3d2010-09-17 15:02:32 -07007981 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007982
Nikhil Raofab47622010-10-15 13:12:29 -07007983force_balance:
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007984 /* Looks like there is an imbalance. Compute it */
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007985 calculate_imbalance(env, &sds);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007986 return sds.busiest;
7987
7988out_balanced:
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007989 env->imbalance = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007990 return NULL;
7991}
7992
7993/*
7994 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7995 */
Peter Zijlstrabd939f42012-05-02 14:20:37 +02007996static struct rq *find_busiest_queue(struct lb_env *env,
Michael Wangb94031302012-07-12 16:10:13 +08007997 struct sched_group *group)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01007998{
7999 struct rq *busiest = NULL, *rq;
Nicolas Pitreced549f2014-05-26 18:19:38 -04008000 unsigned long busiest_load = 0, busiest_capacity = 1;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008001 int i;
8002
Peter Zijlstraae4df9d2017-05-01 11:03:12 +02008003 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
Vincent Guittotea678212015-02-27 16:54:11 +01008004 unsigned long capacity, wl;
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01008005 enum fbq_type rt;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008006
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01008007 rq = cpu_rq(i);
8008 rt = fbq_classify_rq(rq);
8009
8010 /*
8011 * We classify groups/runqueues into three groups:
8012 * - regular: there are !numa tasks
8013 * - remote: there are numa tasks that run on the 'wrong' node
8014 * - all: there is no distinction
8015 *
8016 * In order to avoid migrating ideally placed numa tasks,
8017 * ignore those when there's better options.
8018 *
8019 * If we ignore the actual busiest queue to migrate another
8020 * task, the next balance pass can still reduce the busiest
8021 * queue by moving tasks around inside the node.
8022 *
8023 * If we cannot move enough load due to this classification
8024 * the next pass will adjust the group classification and
8025 * allow migration of more tasks.
8026 *
8027 * Both cases only affect the total convergence complexity.
8028 */
8029 if (rt > env->fbq_type)
8030 continue;
8031
Nicolas Pitreced549f2014-05-26 18:19:38 -04008032 capacity = capacity_of(i);
Srivatsa Vaddagiri9d5efe02010-06-08 14:57:02 +10008033
Viresh Kumarc7132dd2017-05-24 10:59:54 +05308034 wl = weighted_cpuload(rq);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008035
Thomas Gleixner6e40f5b2010-02-16 16:48:56 +01008036 /*
8037 * When comparing with imbalance, use weighted_cpuload()
Nicolas Pitreced549f2014-05-26 18:19:38 -04008038 * which is not scaled with the cpu capacity.
Thomas Gleixner6e40f5b2010-02-16 16:48:56 +01008039 */
Vincent Guittotea678212015-02-27 16:54:11 +01008040
8041 if (rq->nr_running == 1 && wl > env->imbalance &&
8042 !check_cpu_capacity(rq, env->sd))
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008043 continue;
8044
Thomas Gleixner6e40f5b2010-02-16 16:48:56 +01008045 /*
8046 * For the load comparisons with the other cpu's, consider
Nicolas Pitreced549f2014-05-26 18:19:38 -04008047 * the weighted_cpuload() scaled with the cpu capacity, so
8048 * that the load can be moved away from the cpu that is
8049 * potentially running at a lower capacity.
Joonsoo Kim95a79b82013-08-06 17:36:41 +09008050 *
Nicolas Pitreced549f2014-05-26 18:19:38 -04008051 * Thus we're looking for max(wl_i / capacity_i), crosswise
Joonsoo Kim95a79b82013-08-06 17:36:41 +09008052 * multiplication to rid ourselves of the division works out
Nicolas Pitreced549f2014-05-26 18:19:38 -04008053 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8054 * our previous maximum.
Thomas Gleixner6e40f5b2010-02-16 16:48:56 +01008055 */
Nicolas Pitreced549f2014-05-26 18:19:38 -04008056 if (wl * busiest_capacity > busiest_load * capacity) {
Joonsoo Kim95a79b82013-08-06 17:36:41 +09008057 busiest_load = wl;
Nicolas Pitreced549f2014-05-26 18:19:38 -04008058 busiest_capacity = capacity;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008059 busiest = rq;
8060 }
8061 }
8062
8063 return busiest;
8064}
8065
8066/*
8067 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8068 * so long as it is large enough.
8069 */
8070#define MAX_PINNED_INTERVAL 512
8071
Peter Zijlstrabd939f42012-05-02 14:20:37 +02008072static int need_active_balance(struct lb_env *env)
Peter Zijlstra1af3ed32009-12-23 15:10:31 +01008073{
Peter Zijlstrabd939f42012-05-02 14:20:37 +02008074 struct sched_domain *sd = env->sd;
8075
8076 if (env->idle == CPU_NEWLY_IDLE) {
Michael Neuling532cb4c2010-06-08 14:57:02 +10008077
8078 /*
8079 * ASYM_PACKING needs to force migrate tasks from busy but
Tim Chenafe06ef2016-11-22 12:23:53 -08008080 * lower priority CPUs in order to pack all tasks in the
8081 * highest priority CPUs.
Michael Neuling532cb4c2010-06-08 14:57:02 +10008082 */
Tim Chenafe06ef2016-11-22 12:23:53 -08008083 if ((sd->flags & SD_ASYM_PACKING) &&
8084 sched_asym_prefer(env->dst_cpu, env->src_cpu))
Michael Neuling532cb4c2010-06-08 14:57:02 +10008085 return 1;
Peter Zijlstra1af3ed32009-12-23 15:10:31 +01008086 }
8087
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008088 /*
8089 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8090 * It's worth migrating the task if the src_cpu's capacity is reduced
8091 * because of other sched_class or IRQs if more capacity stays
8092 * available on dst_cpu.
8093 */
8094 if ((env->idle != CPU_NOT_IDLE) &&
8095 (env->src_rq->cfs.h_nr_running == 1)) {
8096 if ((check_cpu_capacity(env->src_rq, sd)) &&
8097 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8098 return 1;
8099 }
8100
Peter Zijlstra1af3ed32009-12-23 15:10:31 +01008101 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8102}
8103
Tejun Heo969c7922010-05-06 18:49:21 +02008104static int active_load_balance_cpu_stop(void *data);
8105
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008106static int should_we_balance(struct lb_env *env)
8107{
8108 struct sched_group *sg = env->sd->groups;
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008109 int cpu, balance_cpu = -1;
8110
8111 /*
8112 * In the newly idle case, we will allow all the cpu's
8113 * to do the newly idle load balance.
8114 */
8115 if (env->idle == CPU_NEWLY_IDLE)
8116 return 1;
8117
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008118 /* Try to find first idle cpu */
Peter Zijlstrae5c14b12017-05-01 10:47:02 +02008119 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
Peter Zijlstraaf218122017-05-01 08:51:05 +02008120 if (!idle_cpu(cpu))
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008121 continue;
8122
8123 balance_cpu = cpu;
8124 break;
8125 }
8126
8127 if (balance_cpu == -1)
8128 balance_cpu = group_balance_cpu(sg);
8129
8130 /*
8131 * First idle cpu or the first cpu(busiest) in this sched group
8132 * is eligible for doing load balancing at this and above domains.
8133 */
Joonsoo Kimb0cff9d2013-09-10 15:54:49 +09008134 return balance_cpu == env->dst_cpu;
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008135}
8136
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008137/*
8138 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8139 * tasks if there is an imbalance.
8140 */
8141static int load_balance(int this_cpu, struct rq *this_rq,
8142 struct sched_domain *sd, enum cpu_idle_type idle,
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008143 int *continue_balancing)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008144{
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05308145 int ld_moved, cur_ld_moved, active_balance = 0;
Peter Zijlstra62633222013-08-19 12:41:09 +02008146 struct sched_domain *sd_parent = sd->parent;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008147 struct sched_group *group;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008148 struct rq *busiest;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008149 struct rq_flags rf;
Christoph Lameter4ba29682014-08-26 19:12:21 -05008150 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008151
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008152 struct lb_env env = {
8153 .sd = sd,
Peter Zijlstraddcdf6e2012-02-22 19:27:40 +01008154 .dst_cpu = this_cpu,
8155 .dst_rq = this_rq,
Peter Zijlstraae4df9d2017-05-01 11:03:12 +02008156 .dst_grpmask = sched_group_span(sd->groups),
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008157 .idle = idle,
Peter Zijlstraeb953082012-04-17 13:38:40 +02008158 .loop_break = sched_nr_migrate_break,
Michael Wangb94031302012-07-12 16:10:13 +08008159 .cpus = cpus,
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01008160 .fbq_type = all,
Kirill Tkhai163122b2014-08-20 13:48:29 +04008161 .tasks = LIST_HEAD_INIT(env.tasks),
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008162 };
8163
Jeffrey Hugo65a44332017-06-07 13:18:57 -06008164 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008165
Josh Poimboeufae928822016-06-17 12:43:24 -05008166 schedstat_inc(sd->lb_count[idle]);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008167
8168redo:
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008169 if (!should_we_balance(&env)) {
8170 *continue_balancing = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008171 goto out_balanced;
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008172 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008173
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008174 group = find_busiest_group(&env);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008175 if (!group) {
Josh Poimboeufae928822016-06-17 12:43:24 -05008176 schedstat_inc(sd->lb_nobusyg[idle]);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008177 goto out_balanced;
8178 }
8179
Michael Wangb94031302012-07-12 16:10:13 +08008180 busiest = find_busiest_queue(&env, group);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008181 if (!busiest) {
Josh Poimboeufae928822016-06-17 12:43:24 -05008182 schedstat_inc(sd->lb_nobusyq[idle]);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008183 goto out_balanced;
8184 }
8185
Michael Wang78feefc2012-08-06 16:41:59 +08008186 BUG_ON(busiest == env.dst_rq);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008187
Josh Poimboeufae928822016-06-17 12:43:24 -05008188 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008189
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008190 env.src_cpu = busiest->cpu;
8191 env.src_rq = busiest;
8192
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008193 ld_moved = 0;
8194 if (busiest->nr_running > 1) {
8195 /*
8196 * Attempt to move tasks. If find_busiest_group has found
8197 * an imbalance but busiest->nr_running <= 1, the group is
8198 * still unbalanced. ld_moved simply stays zero, so it is
8199 * correctly treated as an imbalance.
8200 */
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008201 env.flags |= LBF_ALL_PINNED;
Peter Zijlstrac82513e2012-04-26 13:12:27 +02008202 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008203
Peter Zijlstra5d6523e2012-03-10 00:07:36 +01008204more_balance:
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008205 rq_lock_irqsave(busiest, &rf);
Peter Zijlstra3bed5e22016-10-03 16:35:32 +02008206 update_rq_clock(busiest);
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05308207
8208 /*
8209 * cur_ld_moved - load moved in current iteration
8210 * ld_moved - cumulative load moved across iterations
8211 */
Kirill Tkhai163122b2014-08-20 13:48:29 +04008212 cur_ld_moved = detach_tasks(&env);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008213
8214 /*
Kirill Tkhai163122b2014-08-20 13:48:29 +04008215 * We've detached some tasks from busiest_rq. Every
8216 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8217 * unlock busiest->lock, and we are able to be sure
8218 * that nobody can manipulate the tasks in parallel.
8219 * See task_rq_lock() family for the details.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008220 */
Kirill Tkhai163122b2014-08-20 13:48:29 +04008221
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008222 rq_unlock(busiest, &rf);
Kirill Tkhai163122b2014-08-20 13:48:29 +04008223
8224 if (cur_ld_moved) {
8225 attach_tasks(&env);
8226 ld_moved += cur_ld_moved;
8227 }
8228
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008229 local_irq_restore(rf.flags);
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05308230
Joonsoo Kimf1cd0852013-04-23 17:27:37 +09008231 if (env.flags & LBF_NEED_BREAK) {
8232 env.flags &= ~LBF_NEED_BREAK;
8233 goto more_balance;
8234 }
8235
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05308236 /*
8237 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8238 * us and move them to an alternate dst_cpu in our sched_group
8239 * where they can run. The upper limit on how many times we
8240 * iterate on same src_cpu is dependent on number of cpus in our
8241 * sched_group.
8242 *
8243 * This changes load balance semantics a bit on who can move
8244 * load to a given_cpu. In addition to the given_cpu itself
8245 * (or a ilb_cpu acting on its behalf where given_cpu is
8246 * nohz-idle), we now have balance_cpu in a position to move
8247 * load to given_cpu. In rare situations, this may cause
8248 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8249 * _independently_ and at _same_ time to move some load to
8250 * given_cpu) causing exceess load to be moved to given_cpu.
8251 * This however should not happen so much in practice and
8252 * moreover subsequent load balance cycles should correct the
8253 * excess load moved.
8254 */
Peter Zijlstra62633222013-08-19 12:41:09 +02008255 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05308256
Vladimir Davydov7aff2e32013-09-15 21:30:13 +04008257 /* Prevent to re-select dst_cpu via env's cpus */
8258 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8259
Michael Wang78feefc2012-08-06 16:41:59 +08008260 env.dst_rq = cpu_rq(env.new_dst_cpu);
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05308261 env.dst_cpu = env.new_dst_cpu;
Peter Zijlstra62633222013-08-19 12:41:09 +02008262 env.flags &= ~LBF_DST_PINNED;
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05308263 env.loop = 0;
8264 env.loop_break = sched_nr_migrate_break;
Joonsoo Kime02e60c2013-04-23 17:27:42 +09008265
Srivatsa Vaddagiri88b8dac2012-06-19 17:43:15 +05308266 /*
8267 * Go back to "more_balance" rather than "redo" since we
8268 * need to continue with same src_cpu.
8269 */
8270 goto more_balance;
8271 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008272
Peter Zijlstra62633222013-08-19 12:41:09 +02008273 /*
8274 * We failed to reach balance because of affinity.
8275 */
8276 if (sd_parent) {
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04008277 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
Peter Zijlstra62633222013-08-19 12:41:09 +02008278
Vincent Guittotafdeee02014-08-26 13:06:44 +02008279 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
Peter Zijlstra62633222013-08-19 12:41:09 +02008280 *group_imbalance = 1;
Peter Zijlstra62633222013-08-19 12:41:09 +02008281 }
8282
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008283 /* All tasks on this runqueue were pinned by CPU affinity */
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008284 if (unlikely(env.flags & LBF_ALL_PINNED)) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008285 cpumask_clear_cpu(cpu_of(busiest), cpus);
Jeffrey Hugo65a44332017-06-07 13:18:57 -06008286 /*
8287 * Attempting to continue load balancing at the current
8288 * sched_domain level only makes sense if there are
8289 * active CPUs remaining as possible busiest CPUs to
8290 * pull load from which are not contained within the
8291 * destination group that is receiving any migrated
8292 * load.
8293 */
8294 if (!cpumask_subset(cpus, env.dst_grpmask)) {
Prashanth Nageshappabbf18b12012-06-19 17:52:07 +05308295 env.loop = 0;
8296 env.loop_break = sched_nr_migrate_break;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008297 goto redo;
Prashanth Nageshappabbf18b12012-06-19 17:52:07 +05308298 }
Vincent Guittotafdeee02014-08-26 13:06:44 +02008299 goto out_all_pinned;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008300 }
8301 }
8302
8303 if (!ld_moved) {
Josh Poimboeufae928822016-06-17 12:43:24 -05008304 schedstat_inc(sd->lb_failed[idle]);
Venkatesh Pallipadi58b26c42010-09-10 18:19:17 -07008305 /*
8306 * Increment the failure counter only on periodic balance.
8307 * We do not want newidle balance, which can be very
8308 * frequent, pollute the failure counter causing
8309 * excessive cache_hot migrations and active balances.
8310 */
8311 if (idle != CPU_NEWLY_IDLE)
8312 sd->nr_balance_failed++;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008313
Peter Zijlstrabd939f42012-05-02 14:20:37 +02008314 if (need_active_balance(&env)) {
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008315 unsigned long flags;
8316
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008317 raw_spin_lock_irqsave(&busiest->lock, flags);
8318
Tejun Heo969c7922010-05-06 18:49:21 +02008319 /* don't kick the active_load_balance_cpu_stop,
8320 * if the curr task on busiest cpu can't be
8321 * moved to this_cpu
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008322 */
Ingo Molnar0c98d342017-02-05 15:38:10 +01008323 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008324 raw_spin_unlock_irqrestore(&busiest->lock,
8325 flags);
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008326 env.flags |= LBF_ALL_PINNED;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008327 goto out_one_pinned;
8328 }
8329
Tejun Heo969c7922010-05-06 18:49:21 +02008330 /*
8331 * ->active_balance synchronizes accesses to
8332 * ->active_balance_work. Once set, it's cleared
8333 * only after active load balance is finished.
8334 */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008335 if (!busiest->active_balance) {
8336 busiest->active_balance = 1;
8337 busiest->push_cpu = this_cpu;
8338 active_balance = 1;
8339 }
8340 raw_spin_unlock_irqrestore(&busiest->lock, flags);
Tejun Heo969c7922010-05-06 18:49:21 +02008341
Peter Zijlstrabd939f42012-05-02 14:20:37 +02008342 if (active_balance) {
Tejun Heo969c7922010-05-06 18:49:21 +02008343 stop_one_cpu_nowait(cpu_of(busiest),
8344 active_load_balance_cpu_stop, busiest,
8345 &busiest->active_balance_work);
Peter Zijlstrabd939f42012-05-02 14:20:37 +02008346 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008347
Srikar Dronamrajud02c071182016-03-23 17:54:44 +05308348 /* We've kicked active balancing, force task migration. */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008349 sd->nr_balance_failed = sd->cache_nice_tries+1;
8350 }
8351 } else
8352 sd->nr_balance_failed = 0;
8353
8354 if (likely(!active_balance)) {
8355 /* We were unbalanced, so reset the balancing interval */
8356 sd->balance_interval = sd->min_interval;
8357 } else {
8358 /*
8359 * If we've begun active balancing, start to back off. This
8360 * case may not be covered by the all_pinned logic if there
8361 * is only 1 task on the busy runqueue (because we don't call
Kirill Tkhai163122b2014-08-20 13:48:29 +04008362 * detach_tasks).
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008363 */
8364 if (sd->balance_interval < sd->max_interval)
8365 sd->balance_interval *= 2;
8366 }
8367
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008368 goto out;
8369
8370out_balanced:
Vincent Guittotafdeee02014-08-26 13:06:44 +02008371 /*
8372 * We reach balance although we may have faced some affinity
8373 * constraints. Clear the imbalance flag if it was set.
8374 */
8375 if (sd_parent) {
8376 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8377
8378 if (*group_imbalance)
8379 *group_imbalance = 0;
8380 }
8381
8382out_all_pinned:
8383 /*
8384 * We reach balance because all tasks are pinned at this level so
8385 * we can't migrate them. Let the imbalance flag set so parent level
8386 * can try to migrate them.
8387 */
Josh Poimboeufae928822016-06-17 12:43:24 -05008388 schedstat_inc(sd->lb_balanced[idle]);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008389
8390 sd->nr_balance_failed = 0;
8391
8392out_one_pinned:
8393 /* tune up the balancing interval */
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008394 if (((env.flags & LBF_ALL_PINNED) &&
Peter Zijlstra5b54b562011-09-22 15:23:13 +02008395 sd->balance_interval < MAX_PINNED_INTERVAL) ||
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008396 (sd->balance_interval < sd->max_interval))
8397 sd->balance_interval *= 2;
8398
Venkatesh Pallipadi46e49b32011-02-14 14:38:50 -08008399 ld_moved = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008400out:
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008401 return ld_moved;
8402}
8403
Jason Low52a08ef2014-05-08 17:49:22 -07008404static inline unsigned long
8405get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8406{
8407 unsigned long interval = sd->balance_interval;
8408
8409 if (cpu_busy)
8410 interval *= sd->busy_factor;
8411
8412 /* scale ms to jiffies */
8413 interval = msecs_to_jiffies(interval);
8414 interval = clamp(interval, 1UL, max_load_balance_interval);
8415
8416 return interval;
8417}
8418
8419static inline void
Leo Yan31851a92016-08-05 14:31:29 +08008420update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
Jason Low52a08ef2014-05-08 17:49:22 -07008421{
8422 unsigned long interval, next;
8423
Leo Yan31851a92016-08-05 14:31:29 +08008424 /* used by idle balance, so cpu_busy = 0 */
8425 interval = get_sd_balance_interval(sd, 0);
Jason Low52a08ef2014-05-08 17:49:22 -07008426 next = sd->last_balance + interval;
8427
8428 if (time_after(*next_balance, next))
8429 *next_balance = next;
8430}
8431
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008432/*
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008433 * idle_balance is called by schedule() if this_cpu is about to become
8434 * idle. Attempts to pull tasks from other CPUs.
8435 */
Matt Fleming46f69fa2016-09-21 14:38:12 +01008436static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008437{
Jason Low52a08ef2014-05-08 17:49:22 -07008438 unsigned long next_balance = jiffies + HZ;
8439 int this_cpu = this_rq->cpu;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008440 struct sched_domain *sd;
8441 int pulled_task = 0;
Jason Low9bd721c2013-09-13 11:26:52 -07008442 u64 curr_cost = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008443
Peter Zijlstra6e831252014-02-11 16:11:48 +01008444 /*
8445 * We must set idle_stamp _before_ calling idle_balance(), such that we
8446 * measure the duration of idle_balance() as idle time.
8447 */
8448 this_rq->idle_stamp = rq_clock(this_rq);
8449
Matt Fleming46f69fa2016-09-21 14:38:12 +01008450 /*
8451 * This is OK, because current is on_cpu, which avoids it being picked
8452 * for load-balance and preemption/IRQs are still disabled avoiding
8453 * further scheduler activity on it and we're being very careful to
8454 * re-start the picking loop.
8455 */
8456 rq_unpin_lock(this_rq, rf);
8457
Tim Chen4486edd2014-06-23 12:16:49 -07008458 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
8459 !this_rq->rd->overload) {
Jason Low52a08ef2014-05-08 17:49:22 -07008460 rcu_read_lock();
8461 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8462 if (sd)
Leo Yan31851a92016-08-05 14:31:29 +08008463 update_next_balance(sd, &next_balance);
Jason Low52a08ef2014-05-08 17:49:22 -07008464 rcu_read_unlock();
8465
Peter Zijlstra6e831252014-02-11 16:11:48 +01008466 goto out;
Jason Low52a08ef2014-05-08 17:49:22 -07008467 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008468
Peter Zijlstraf492e122009-12-23 15:29:42 +01008469 raw_spin_unlock(&this_rq->lock);
8470
Paul Turner48a16752012-10-04 13:18:31 +02008471 update_blocked_averages(this_cpu);
Peter Zijlstradce840a2011-04-07 14:09:50 +02008472 rcu_read_lock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008473 for_each_domain(this_cpu, sd) {
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008474 int continue_balancing = 1;
Jason Low9bd721c2013-09-13 11:26:52 -07008475 u64 t0, domain_cost;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008476
8477 if (!(sd->flags & SD_LOAD_BALANCE))
8478 continue;
8479
Jason Low52a08ef2014-05-08 17:49:22 -07008480 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
Leo Yan31851a92016-08-05 14:31:29 +08008481 update_next_balance(sd, &next_balance);
Jason Low9bd721c2013-09-13 11:26:52 -07008482 break;
Jason Low52a08ef2014-05-08 17:49:22 -07008483 }
Jason Low9bd721c2013-09-13 11:26:52 -07008484
Peter Zijlstraf492e122009-12-23 15:29:42 +01008485 if (sd->flags & SD_BALANCE_NEWIDLE) {
Jason Low9bd721c2013-09-13 11:26:52 -07008486 t0 = sched_clock_cpu(this_cpu);
8487
Peter Zijlstraf492e122009-12-23 15:29:42 +01008488 pulled_task = load_balance(this_cpu, this_rq,
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008489 sd, CPU_NEWLY_IDLE,
8490 &continue_balancing);
Jason Low9bd721c2013-09-13 11:26:52 -07008491
8492 domain_cost = sched_clock_cpu(this_cpu) - t0;
8493 if (domain_cost > sd->max_newidle_lb_cost)
8494 sd->max_newidle_lb_cost = domain_cost;
8495
8496 curr_cost += domain_cost;
Peter Zijlstraf492e122009-12-23 15:29:42 +01008497 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008498
Leo Yan31851a92016-08-05 14:31:29 +08008499 update_next_balance(sd, &next_balance);
Jason Low39a4d9c2014-04-23 18:30:35 -07008500
8501 /*
8502 * Stop searching for tasks to pull if there are
8503 * now runnable tasks on this rq.
8504 */
8505 if (pulled_task || this_rq->nr_running > 0)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008506 break;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008507 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02008508 rcu_read_unlock();
Peter Zijlstraf492e122009-12-23 15:29:42 +01008509
8510 raw_spin_lock(&this_rq->lock);
8511
Jason Low0e5b5332014-04-28 15:45:54 -07008512 if (curr_cost > this_rq->max_idle_balance_cost)
8513 this_rq->max_idle_balance_cost = curr_cost;
8514
Daniel Lezcanoe5fc6612014-01-17 10:04:02 +01008515 /*
Jason Low0e5b5332014-04-28 15:45:54 -07008516 * While browsing the domains, we released the rq lock, a task could
8517 * have been enqueued in the meantime. Since we're not going idle,
8518 * pretend we pulled a task.
Daniel Lezcanoe5fc6612014-01-17 10:04:02 +01008519 */
Jason Low0e5b5332014-04-28 15:45:54 -07008520 if (this_rq->cfs.h_nr_running && !pulled_task)
Peter Zijlstra6e831252014-02-11 16:11:48 +01008521 pulled_task = 1;
Daniel Lezcanoe5fc6612014-01-17 10:04:02 +01008522
Peter Zijlstra6e831252014-02-11 16:11:48 +01008523out:
Jason Low52a08ef2014-05-08 17:49:22 -07008524 /* Move the next balance forward */
8525 if (time_after(this_rq->next_balance, next_balance))
8526 this_rq->next_balance = next_balance;
8527
Kirill Tkhaie4aa3582014-03-06 13:31:55 +04008528 /* Is there a task of a high priority class? */
Kirill Tkhai46383642014-03-15 02:15:07 +04008529 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
Kirill Tkhaie4aa3582014-03-06 13:31:55 +04008530 pulled_task = -1;
8531
Dietmar Eggemann38c6ade2015-10-20 13:04:41 +01008532 if (pulled_task)
Peter Zijlstra6e831252014-02-11 16:11:48 +01008533 this_rq->idle_stamp = 0;
8534
Matt Fleming46f69fa2016-09-21 14:38:12 +01008535 rq_repin_lock(this_rq, rf);
8536
Daniel Lezcano3c4017c2014-01-17 10:04:03 +01008537 return pulled_task;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008538}
8539
8540/*
Tejun Heo969c7922010-05-06 18:49:21 +02008541 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8542 * running tasks off the busiest CPU onto idle CPUs. It requires at
8543 * least 1 task to be running on each physical CPU where possible, and
8544 * avoids physical / logical imbalances.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008545 */
Tejun Heo969c7922010-05-06 18:49:21 +02008546static int active_load_balance_cpu_stop(void *data)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008547{
Tejun Heo969c7922010-05-06 18:49:21 +02008548 struct rq *busiest_rq = data;
8549 int busiest_cpu = cpu_of(busiest_rq);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008550 int target_cpu = busiest_rq->push_cpu;
Tejun Heo969c7922010-05-06 18:49:21 +02008551 struct rq *target_rq = cpu_rq(target_cpu);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008552 struct sched_domain *sd;
Kirill Tkhaie5673f22014-08-20 13:48:01 +04008553 struct task_struct *p = NULL;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008554 struct rq_flags rf;
Tejun Heo969c7922010-05-06 18:49:21 +02008555
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008556 rq_lock_irq(busiest_rq, &rf);
Tejun Heo969c7922010-05-06 18:49:21 +02008557
8558 /* make sure the requested cpu hasn't gone down in the meantime */
8559 if (unlikely(busiest_cpu != smp_processor_id() ||
8560 !busiest_rq->active_balance))
8561 goto out_unlock;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008562
8563 /* Is there any task to move? */
8564 if (busiest_rq->nr_running <= 1)
Tejun Heo969c7922010-05-06 18:49:21 +02008565 goto out_unlock;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008566
8567 /*
8568 * This condition is "impossible", if it occurs
8569 * we need to fix it. Originally reported by
8570 * Bjorn Helgaas on a 128-cpu setup.
8571 */
8572 BUG_ON(busiest_rq == target_rq);
8573
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008574 /* Search for an sd spanning us and the target CPU. */
Peter Zijlstradce840a2011-04-07 14:09:50 +02008575 rcu_read_lock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008576 for_each_domain(target_cpu, sd) {
8577 if ((sd->flags & SD_LOAD_BALANCE) &&
8578 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8579 break;
8580 }
8581
8582 if (likely(sd)) {
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008583 struct lb_env env = {
8584 .sd = sd,
Peter Zijlstraddcdf6e2012-02-22 19:27:40 +01008585 .dst_cpu = target_cpu,
8586 .dst_rq = target_rq,
8587 .src_cpu = busiest_rq->cpu,
8588 .src_rq = busiest_rq,
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008589 .idle = CPU_IDLE,
Jeffrey Hugo65a44332017-06-07 13:18:57 -06008590 /*
8591 * can_migrate_task() doesn't need to compute new_dst_cpu
8592 * for active balancing. Since we have CPU_IDLE, but no
8593 * @dst_grpmask we need to make that test go away with lying
8594 * about DST_PINNED.
8595 */
8596 .flags = LBF_DST_PINNED,
Peter Zijlstra8e45cb52012-02-22 12:47:19 +01008597 };
8598
Josh Poimboeufae928822016-06-17 12:43:24 -05008599 schedstat_inc(sd->alb_count);
Peter Zijlstra3bed5e22016-10-03 16:35:32 +02008600 update_rq_clock(busiest_rq);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008601
Kirill Tkhaie5673f22014-08-20 13:48:01 +04008602 p = detach_one_task(&env);
Srikar Dronamrajud02c071182016-03-23 17:54:44 +05308603 if (p) {
Josh Poimboeufae928822016-06-17 12:43:24 -05008604 schedstat_inc(sd->alb_pushed);
Srikar Dronamrajud02c071182016-03-23 17:54:44 +05308605 /* Active balancing done, reset the failure counter. */
8606 sd->nr_balance_failed = 0;
8607 } else {
Josh Poimboeufae928822016-06-17 12:43:24 -05008608 schedstat_inc(sd->alb_failed);
Srikar Dronamrajud02c071182016-03-23 17:54:44 +05308609 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008610 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02008611 rcu_read_unlock();
Tejun Heo969c7922010-05-06 18:49:21 +02008612out_unlock:
8613 busiest_rq->active_balance = 0;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008614 rq_unlock(busiest_rq, &rf);
Kirill Tkhaie5673f22014-08-20 13:48:01 +04008615
8616 if (p)
8617 attach_one_task(target_rq, p);
8618
8619 local_irq_enable();
8620
Tejun Heo969c7922010-05-06 18:49:21 +02008621 return 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008622}
8623
Mike Galbraithd987fc72011-12-05 10:01:47 +01008624static inline int on_null_domain(struct rq *rq)
8625{
8626 return unlikely(!rcu_dereference_sched(rq->sd));
8627}
8628
Frederic Weisbecker3451d022011-08-10 23:21:01 +02008629#ifdef CONFIG_NO_HZ_COMMON
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008630/*
8631 * idle load balancing details
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008632 * - When one of the busy CPUs notice that there may be an idle rebalancing
8633 * needed, they will kick the idle load balancer, which then does idle
8634 * load balancing for all the idle CPUs.
8635 */
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008636static struct {
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008637 cpumask_var_t idle_cpus_mask;
Suresh Siddha0b005cf2011-12-01 17:07:34 -08008638 atomic_t nr_cpus;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008639 unsigned long next_balance; /* in jiffy units */
8640} nohz ____cacheline_aligned;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008641
Daniel Lezcano3dd03372014-01-06 12:34:41 +01008642static inline int find_new_ilb(void)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008643{
Suresh Siddha0b005cf2011-12-01 17:07:34 -08008644 int ilb = cpumask_first(nohz.idle_cpus_mask);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008645
Suresh Siddha786d6dc2011-12-01 17:07:35 -08008646 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8647 return ilb;
8648
8649 return nr_cpu_ids;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008650}
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008651
8652/*
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008653 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8654 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8655 * CPU (if there is one).
8656 */
Daniel Lezcano0aeeeeb2014-01-06 12:34:42 +01008657static void nohz_balancer_kick(void)
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008658{
8659 int ilb_cpu;
8660
8661 nohz.next_balance++;
8662
Daniel Lezcano3dd03372014-01-06 12:34:41 +01008663 ilb_cpu = find_new_ilb();
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008664
Suresh Siddha0b005cf2011-12-01 17:07:34 -08008665 if (ilb_cpu >= nr_cpu_ids)
8666 return;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008667
Suresh Siddhacd490c52011-12-06 11:26:34 -08008668 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
Suresh Siddha1c792db2011-12-01 17:07:32 -08008669 return;
8670 /*
8671 * Use smp_send_reschedule() instead of resched_cpu().
8672 * This way we generate a sched IPI on the target cpu which
8673 * is idle. And the softirq performing nohz idle load balance
8674 * will be run before returning from the IPI.
8675 */
8676 smp_send_reschedule(ilb_cpu);
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008677 return;
8678}
8679
Thomas Gleixner20a5c8c2016-03-10 12:54:20 +01008680void nohz_balance_exit_idle(unsigned int cpu)
Suresh Siddha71325962012-01-19 18:28:57 -08008681{
8682 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
Mike Galbraithd987fc72011-12-05 10:01:47 +01008683 /*
8684 * Completely isolated CPUs don't ever set, so we must test.
8685 */
8686 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8687 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8688 atomic_dec(&nohz.nr_cpus);
8689 }
Suresh Siddha71325962012-01-19 18:28:57 -08008690 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8691 }
8692}
8693
Suresh Siddha69e1e812011-12-01 17:07:33 -08008694static inline void set_cpu_sd_state_busy(void)
8695{
8696 struct sched_domain *sd;
Preeti U Murthy37dc6b52013-10-30 08:42:52 +05308697 int cpu = smp_processor_id();
Suresh Siddha69e1e812011-12-01 17:07:33 -08008698
Suresh Siddha69e1e812011-12-01 17:07:33 -08008699 rcu_read_lock();
Peter Zijlstra0e369d72016-05-09 10:38:01 +02008700 sd = rcu_dereference(per_cpu(sd_llc, cpu));
Vincent Guittot25f55d92013-04-23 16:59:02 +02008701
8702 if (!sd || !sd->nohz_idle)
8703 goto unlock;
8704 sd->nohz_idle = 0;
8705
Peter Zijlstra0e369d72016-05-09 10:38:01 +02008706 atomic_inc(&sd->shared->nr_busy_cpus);
Vincent Guittot25f55d92013-04-23 16:59:02 +02008707unlock:
Suresh Siddha69e1e812011-12-01 17:07:33 -08008708 rcu_read_unlock();
8709}
8710
8711void set_cpu_sd_state_idle(void)
8712{
8713 struct sched_domain *sd;
Preeti U Murthy37dc6b52013-10-30 08:42:52 +05308714 int cpu = smp_processor_id();
Suresh Siddha69e1e812011-12-01 17:07:33 -08008715
Suresh Siddha69e1e812011-12-01 17:07:33 -08008716 rcu_read_lock();
Peter Zijlstra0e369d72016-05-09 10:38:01 +02008717 sd = rcu_dereference(per_cpu(sd_llc, cpu));
Vincent Guittot25f55d92013-04-23 16:59:02 +02008718
8719 if (!sd || sd->nohz_idle)
8720 goto unlock;
8721 sd->nohz_idle = 1;
8722
Peter Zijlstra0e369d72016-05-09 10:38:01 +02008723 atomic_dec(&sd->shared->nr_busy_cpus);
Vincent Guittot25f55d92013-04-23 16:59:02 +02008724unlock:
Suresh Siddha69e1e812011-12-01 17:07:33 -08008725 rcu_read_unlock();
8726}
8727
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008728/*
Alex Shic1cc0172012-09-10 15:10:58 +08008729 * This routine will record that the cpu is going idle with tick stopped.
Suresh Siddha0b005cf2011-12-01 17:07:34 -08008730 * This info will be used in performing idle load balancing in the future.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008731 */
Alex Shic1cc0172012-09-10 15:10:58 +08008732void nohz_balance_enter_idle(int cpu)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008733{
Suresh Siddha71325962012-01-19 18:28:57 -08008734 /*
8735 * If this cpu is going down, then nothing needs to be done.
8736 */
8737 if (!cpu_active(cpu))
8738 return;
8739
Frederic Weisbecker387bc8b2017-06-19 04:12:02 +02008740 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
8741 if (!is_housekeeping_cpu(cpu))
8742 return;
8743
Alex Shic1cc0172012-09-10 15:10:58 +08008744 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8745 return;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008746
Mike Galbraithd987fc72011-12-05 10:01:47 +01008747 /*
8748 * If we're a completely isolated CPU, we don't play.
8749 */
8750 if (on_null_domain(cpu_rq(cpu)))
8751 return;
8752
Alex Shic1cc0172012-09-10 15:10:58 +08008753 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8754 atomic_inc(&nohz.nr_cpus);
8755 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008756}
8757#endif
8758
8759static DEFINE_SPINLOCK(balancing);
8760
Peter Zijlstra49c022e2011-04-05 10:14:25 +02008761/*
8762 * Scale the max load_balance interval with the number of CPUs in the system.
8763 * This trades load-balance latency on larger machines for less cross talk.
8764 */
Peter Zijlstra029632f2011-10-25 10:00:11 +02008765void update_max_interval(void)
Peter Zijlstra49c022e2011-04-05 10:14:25 +02008766{
8767 max_load_balance_interval = HZ*num_online_cpus()/10;
8768}
8769
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008770/*
8771 * It checks each scheduling domain to see if it is due to be balanced,
8772 * and initiates a balancing operation if so.
8773 *
Libinb9b08532013-04-01 19:14:01 +08008774 * Balancing parameters are set up in init_sched_domains.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008775 */
Daniel Lezcanof7ed0a82014-01-06 12:34:43 +01008776static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008777{
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008778 int continue_balancing = 1;
Daniel Lezcanof7ed0a82014-01-06 12:34:43 +01008779 int cpu = rq->cpu;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008780 unsigned long interval;
Peter Zijlstra04f733b2012-05-11 00:12:02 +02008781 struct sched_domain *sd;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008782 /* Earliest time when we have to do rebalance again */
8783 unsigned long next_balance = jiffies + 60*HZ;
8784 int update_next_balance = 0;
Jason Lowf48627e2013-09-13 11:26:53 -07008785 int need_serialize, need_decay = 0;
8786 u64 max_cost = 0;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008787
Paul Turner48a16752012-10-04 13:18:31 +02008788 update_blocked_averages(cpu);
Peter Zijlstra2069dd72010-11-15 15:47:00 -08008789
Peter Zijlstradce840a2011-04-07 14:09:50 +02008790 rcu_read_lock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008791 for_each_domain(cpu, sd) {
Jason Lowf48627e2013-09-13 11:26:53 -07008792 /*
8793 * Decay the newidle max times here because this is a regular
8794 * visit to all the domains. Decay ~1% per second.
8795 */
8796 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8797 sd->max_newidle_lb_cost =
8798 (sd->max_newidle_lb_cost * 253) / 256;
8799 sd->next_decay_max_lb_cost = jiffies + HZ;
8800 need_decay = 1;
8801 }
8802 max_cost += sd->max_newidle_lb_cost;
8803
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008804 if (!(sd->flags & SD_LOAD_BALANCE))
8805 continue;
8806
Jason Lowf48627e2013-09-13 11:26:53 -07008807 /*
8808 * Stop the load balance at this level. There is another
8809 * CPU in our sched group which is doing load balancing more
8810 * actively.
8811 */
8812 if (!continue_balancing) {
8813 if (need_decay)
8814 continue;
8815 break;
8816 }
8817
Jason Low52a08ef2014-05-08 17:49:22 -07008818 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008819
8820 need_serialize = sd->flags & SD_SERIALIZE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008821 if (need_serialize) {
8822 if (!spin_trylock(&balancing))
8823 goto out;
8824 }
8825
8826 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Joonsoo Kim23f0d202013-08-06 17:36:42 +09008827 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008828 /*
Peter Zijlstra62633222013-08-19 12:41:09 +02008829 * The LBF_DST_PINNED logic could have changed
Joonsoo Kimde5eb2d2013-04-23 17:27:38 +09008830 * env->dst_cpu, so we can't know our idle
8831 * state even if we migrated tasks. Update it.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008832 */
Joonsoo Kimde5eb2d2013-04-23 17:27:38 +09008833 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008834 }
8835 sd->last_balance = jiffies;
Jason Low52a08ef2014-05-08 17:49:22 -07008836 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008837 }
8838 if (need_serialize)
8839 spin_unlock(&balancing);
8840out:
8841 if (time_after(next_balance, sd->last_balance + interval)) {
8842 next_balance = sd->last_balance + interval;
8843 update_next_balance = 1;
8844 }
Jason Lowf48627e2013-09-13 11:26:53 -07008845 }
8846 if (need_decay) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008847 /*
Jason Lowf48627e2013-09-13 11:26:53 -07008848 * Ensure the rq-wide value also decays but keep it at a
8849 * reasonable floor to avoid funnies with rq->avg_idle.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008850 */
Jason Lowf48627e2013-09-13 11:26:53 -07008851 rq->max_idle_balance_cost =
8852 max((u64)sysctl_sched_migration_cost, max_cost);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008853 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02008854 rcu_read_unlock();
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008855
8856 /*
8857 * next_balance will be updated only when there is a need.
8858 * When the cpu is attached to null domain for ex, it will not be
8859 * updated.
8860 */
Vincent Guittotc5afb6a2015-08-03 11:55:50 +02008861 if (likely(update_next_balance)) {
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008862 rq->next_balance = next_balance;
Vincent Guittotc5afb6a2015-08-03 11:55:50 +02008863
8864#ifdef CONFIG_NO_HZ_COMMON
8865 /*
8866 * If this CPU has been elected to perform the nohz idle
8867 * balance. Other idle CPUs have already rebalanced with
8868 * nohz_idle_balance() and nohz.next_balance has been
8869 * updated accordingly. This CPU is now running the idle load
8870 * balance for itself and we need to update the
8871 * nohz.next_balance accordingly.
8872 */
8873 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8874 nohz.next_balance = rq->next_balance;
8875#endif
8876 }
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01008877}
8878
Frederic Weisbecker3451d022011-08-10 23:21:01 +02008879#ifdef CONFIG_NO_HZ_COMMON
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008880/*
Frederic Weisbecker3451d022011-08-10 23:21:01 +02008881 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008882 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8883 */
Daniel Lezcano208cb162014-01-06 12:34:44 +01008884static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008885{
Daniel Lezcano208cb162014-01-06 12:34:44 +01008886 int this_cpu = this_rq->cpu;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008887 struct rq *rq;
8888 int balance_cpu;
Vincent Guittotc5afb6a2015-08-03 11:55:50 +02008889 /* Earliest time when we have to do rebalance again */
8890 unsigned long next_balance = jiffies + 60*HZ;
8891 int update_next_balance = 0;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008892
Suresh Siddha1c792db2011-12-01 17:07:32 -08008893 if (idle != CPU_IDLE ||
8894 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8895 goto end;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008896
8897 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
Suresh Siddha8a6d42d2011-12-06 11:19:37 -08008898 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008899 continue;
8900
8901 /*
8902 * If this cpu gets work to do, stop the load balancing
8903 * work being done for other cpus. Next load
8904 * balancing owner will pick it up.
8905 */
Suresh Siddha1c792db2011-12-01 17:07:32 -08008906 if (need_resched())
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008907 break;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008908
Vincent Guittot5ed4f1d2012-09-13 06:11:26 +02008909 rq = cpu_rq(balance_cpu);
8910
Tim Chened61bbc2014-05-20 14:39:27 -07008911 /*
8912 * If time for next balance is due,
8913 * do the balance.
8914 */
8915 if (time_after_eq(jiffies, rq->next_balance)) {
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008916 struct rq_flags rf;
8917
8918 rq_lock_irq(rq, &rf);
Tim Chened61bbc2014-05-20 14:39:27 -07008919 update_rq_clock(rq);
Frederic Weisbeckercee1afc2016-04-13 15:56:50 +02008920 cpu_load_update_idle(rq);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008921 rq_unlock_irq(rq, &rf);
8922
Tim Chened61bbc2014-05-20 14:39:27 -07008923 rebalance_domains(rq, CPU_IDLE);
8924 }
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008925
Vincent Guittotc5afb6a2015-08-03 11:55:50 +02008926 if (time_after(next_balance, rq->next_balance)) {
8927 next_balance = rq->next_balance;
8928 update_next_balance = 1;
8929 }
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008930 }
Vincent Guittotc5afb6a2015-08-03 11:55:50 +02008931
8932 /*
8933 * next_balance will be updated only when there is a need.
8934 * When the CPU is attached to null domain for ex, it will not be
8935 * updated.
8936 */
8937 if (likely(update_next_balance))
8938 nohz.next_balance = next_balance;
Suresh Siddha1c792db2011-12-01 17:07:32 -08008939end:
8940 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008941}
8942
8943/*
Suresh Siddha0b005cf2011-12-01 17:07:34 -08008944 * Current heuristic for kicking the idle load balancer in the presence
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008945 * of an idle cpu in the system.
Suresh Siddha0b005cf2011-12-01 17:07:34 -08008946 * - This rq has more than one task.
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008947 * - This rq has at least one CFS task and the capacity of the CPU is
8948 * significantly reduced because of RT tasks or IRQs.
8949 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8950 * multiple busy cpu.
Suresh Siddha0b005cf2011-12-01 17:07:34 -08008951 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8952 * domain span are idle.
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008953 */
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008954static inline bool nohz_kick_needed(struct rq *rq)
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008955{
8956 unsigned long now = jiffies;
Peter Zijlstra0e369d72016-05-09 10:38:01 +02008957 struct sched_domain_shared *sds;
Suresh Siddha0b005cf2011-12-01 17:07:34 -08008958 struct sched_domain *sd;
Tim Chenafe06ef2016-11-22 12:23:53 -08008959 int nr_busy, i, cpu = rq->cpu;
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008960 bool kick = false;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008961
Daniel Lezcano4a725622014-01-06 12:34:39 +01008962 if (unlikely(rq->idle_balance))
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008963 return false;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008964
Suresh Siddha1c792db2011-12-01 17:07:32 -08008965 /*
8966 * We may be recently in ticked or tickless idle mode. At the first
8967 * busy tick after returning from idle, we will update the busy stats.
8968 */
Suresh Siddha69e1e812011-12-01 17:07:33 -08008969 set_cpu_sd_state_busy();
Alex Shic1cc0172012-09-10 15:10:58 +08008970 nohz_balance_exit_idle(cpu);
Suresh Siddha0b005cf2011-12-01 17:07:34 -08008971
8972 /*
8973 * None are in tickless mode and hence no need for NOHZ idle load
8974 * balancing.
8975 */
8976 if (likely(!atomic_read(&nohz.nr_cpus)))
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008977 return false;
Suresh Siddha1c792db2011-12-01 17:07:32 -08008978
8979 if (time_before(now, nohz.next_balance))
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008980 return false;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008981
Suresh Siddha0b005cf2011-12-01 17:07:34 -08008982 if (rq->nr_running >= 2)
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008983 return true;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07008984
Peter Zijlstra067491b2011-12-07 14:32:08 +01008985 rcu_read_lock();
Peter Zijlstra0e369d72016-05-09 10:38:01 +02008986 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
8987 if (sds) {
8988 /*
8989 * XXX: write a coherent comment on why we do this.
8990 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
8991 */
8992 nr_busy = atomic_read(&sds->nr_busy_cpus);
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01008993 if (nr_busy > 1) {
8994 kick = true;
8995 goto unlock;
8996 }
8997
8998 }
8999
9000 sd = rcu_dereference(rq->sd);
9001 if (sd) {
9002 if ((rq->cfs.h_nr_running >= 1) &&
9003 check_cpu_capacity(rq, sd)) {
9004 kick = true;
9005 goto unlock;
9006 }
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07009007 }
Preeti U Murthy37dc6b52013-10-30 08:42:52 +05309008
9009 sd = rcu_dereference(per_cpu(sd_asym, cpu));
Tim Chenafe06ef2016-11-22 12:23:53 -08009010 if (sd) {
9011 for_each_cpu(i, sched_domain_span(sd)) {
9012 if (i == cpu ||
9013 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9014 continue;
Preeti U Murthy37dc6b52013-10-30 08:42:52 +05309015
Tim Chenafe06ef2016-11-22 12:23:53 -08009016 if (sched_asym_prefer(i, cpu)) {
9017 kick = true;
9018 goto unlock;
9019 }
9020 }
9021 }
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01009022unlock:
Peter Zijlstra067491b2011-12-07 14:32:08 +01009023 rcu_read_unlock();
Vincent Guittot1aaf90a2015-02-27 16:54:14 +01009024 return kick;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07009025}
9026#else
Daniel Lezcano208cb162014-01-06 12:34:44 +01009027static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07009028#endif
9029
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009030/*
9031 * run_rebalance_domains is triggered when needed from the scheduler tick.
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07009032 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009033 */
Emese Revfy0766f782016-06-20 20:42:34 +02009034static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009035{
Daniel Lezcano208cb162014-01-06 12:34:44 +01009036 struct rq *this_rq = this_rq();
Suresh Siddha6eb57e02011-10-03 15:09:01 -07009037 enum cpu_idle_type idle = this_rq->idle_balance ?
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009038 CPU_IDLE : CPU_NOT_IDLE;
9039
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009040 /*
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07009041 * If this cpu has a pending nohz_balance_kick, then do the
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009042 * balancing on behalf of the other idle cpus whose ticks are
Preeti U Murthyd4573c32015-03-26 18:32:44 +05309043 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9044 * give the idle cpus a chance to load balance. Else we may
9045 * load balance only within the local sched_domain hierarchy
9046 * and abort nohz_idle_balance altogether if we pull some load.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009047 */
Daniel Lezcano208cb162014-01-06 12:34:44 +01009048 nohz_idle_balance(this_rq, idle);
Preeti U Murthyd4573c32015-03-26 18:32:44 +05309049 rebalance_domains(this_rq, idle);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009050}
9051
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009052/*
9053 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009054 */
Daniel Lezcano7caff662014-01-06 12:34:38 +01009055void trigger_load_balance(struct rq *rq)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009056{
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009057 /* Don't need to rebalance while attached to NULL domain */
Daniel Lezcanoc7260992014-01-06 12:34:45 +01009058 if (unlikely(on_null_domain(rq)))
9059 return;
9060
9061 if (time_after_eq(jiffies, rq->next_balance))
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009062 raise_softirq(SCHED_SOFTIRQ);
Frederic Weisbecker3451d022011-08-10 23:21:01 +02009063#ifdef CONFIG_NO_HZ_COMMON
Daniel Lezcanoc7260992014-01-06 12:34:45 +01009064 if (nohz_kick_needed(rq))
Daniel Lezcano0aeeeeb2014-01-06 12:34:42 +01009065 nohz_balancer_kick();
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07009066#endif
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01009067}
9068
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01009069static void rq_online_fair(struct rq *rq)
9070{
9071 update_sysctl();
Kirill Tkhai0e59bda2014-06-25 12:19:42 +04009072
9073 update_runtime_enabled(rq);
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01009074}
9075
9076static void rq_offline_fair(struct rq *rq)
9077{
9078 update_sysctl();
Peter Boonstoppela4c96ae2012-08-09 15:34:47 -07009079
9080 /* Ensure any throttled groups are reachable by pick_next_task */
9081 unthrottle_offline_cfs_rqs(rq);
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01009082}
9083
Dhaval Giani55e12e52008-06-24 23:39:43 +05309084#endif /* CONFIG_SMP */
Peter Williamse1d14842007-10-24 18:23:51 +02009085
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009086/*
9087 * scheduler tick hitting a task of our scheduling class:
9088 */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01009089static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009090{
9091 struct cfs_rq *cfs_rq;
9092 struct sched_entity *se = &curr->se;
9093
9094 for_each_sched_entity(se) {
9095 cfs_rq = cfs_rq_of(se);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01009096 entity_tick(cfs_rq, se, queued);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009097 }
Ben Segall18bf2802012-10-04 12:51:20 +02009098
Srikar Dronamrajub52da862015-10-02 07:48:25 +05309099 if (static_branch_unlikely(&sched_numa_balancing))
Peter Zijlstracbee9f82012-10-25 14:16:43 +02009100 task_tick_numa(rq, curr);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009101}
9102
9103/*
Peter Zijlstracd29fe62009-11-27 17:32:46 +01009104 * called on fork with the child task as argument from the parent's context
9105 * - child not yet on the tasklist
9106 * - preemption disabled
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009107 */
Peter Zijlstracd29fe62009-11-27 17:32:46 +01009108static void task_fork_fair(struct task_struct *p)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009109{
Daisuke Nishimura4fc420c2011-12-15 14:36:55 +09009110 struct cfs_rq *cfs_rq;
9111 struct sched_entity *se = &p->se, *curr;
Peter Zijlstracd29fe62009-11-27 17:32:46 +01009112 struct rq *rq = this_rq();
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009113 struct rq_flags rf;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009114
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009115 rq_lock(rq, &rf);
Peter Zijlstra861d0342010-08-19 13:31:43 +02009116 update_rq_clock(rq);
9117
Daisuke Nishimura4fc420c2011-12-15 14:36:55 +09009118 cfs_rq = task_cfs_rq(current);
9119 curr = cfs_rq->curr;
Peter Zijlstrae210bff2016-06-16 18:51:48 +02009120 if (curr) {
9121 update_curr(cfs_rq);
Mike Galbraithb5d9d732009-09-08 11:12:28 +02009122 se->vruntime = curr->vruntime;
Peter Zijlstrae210bff2016-06-16 18:51:48 +02009123 }
Peter Zijlstraaeb73b02007-10-15 17:00:05 +02009124 place_entity(cfs_rq, se, 1);
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +02009125
Peter Zijlstracd29fe62009-11-27 17:32:46 +01009126 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
Dmitry Adamushko87fefa32007-10-15 17:00:08 +02009127 /*
Ingo Molnaredcb60a2007-10-15 17:00:08 +02009128 * Upon rescheduling, sched_class::put_prev_task() will place
9129 * 'current' within the tree based on its new key value.
9130 */
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +02009131 swap(curr->vruntime, se->vruntime);
Kirill Tkhai88751252014-06-29 00:03:57 +04009132 resched_curr(rq);
Peter Zijlstra4d78e7b2007-10-15 17:00:04 +02009133 }
9134
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01009135 se->vruntime -= cfs_rq->min_vruntime;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009136 rq_unlock(rq, &rf);
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009137}
9138
Steven Rostedtcb469842008-01-25 21:08:22 +01009139/*
9140 * Priority of the task has changed. Check to see if we preempt
9141 * the current task.
9142 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01009143static void
9144prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
Steven Rostedtcb469842008-01-25 21:08:22 +01009145{
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04009146 if (!task_on_rq_queued(p))
Peter Zijlstrada7a7352011-01-17 17:03:27 +01009147 return;
9148
Steven Rostedtcb469842008-01-25 21:08:22 +01009149 /*
9150 * Reschedule if we are currently running on this runqueue and
9151 * our priority decreased, or if we are not currently running on
9152 * this runqueue and our priority is higher than the current's
9153 */
Peter Zijlstrada7a7352011-01-17 17:03:27 +01009154 if (rq->curr == p) {
Steven Rostedtcb469842008-01-25 21:08:22 +01009155 if (p->prio > oldprio)
Kirill Tkhai88751252014-06-29 00:03:57 +04009156 resched_curr(rq);
Steven Rostedtcb469842008-01-25 21:08:22 +01009157 } else
Peter Zijlstra15afe092008-09-20 23:38:02 +02009158 check_preempt_curr(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01009159}
9160
Byungchul Parkdaa59402015-08-20 20:22:00 +09009161static inline bool vruntime_normalized(struct task_struct *p)
9162{
9163 struct sched_entity *se = &p->se;
9164
9165 /*
9166 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9167 * the dequeue_entity(.flags=0) will already have normalized the
9168 * vruntime.
9169 */
9170 if (p->on_rq)
9171 return true;
9172
9173 /*
9174 * When !on_rq, vruntime of the task has usually NOT been normalized.
9175 * But there are some cases where it has already been normalized:
9176 *
9177 * - A forked child which is waiting for being woken up by
9178 * wake_up_new_task().
9179 * - A task which has been woken up by try_to_wake_up() and
9180 * waiting for actually being woken up by sched_ttwu_pending().
9181 */
9182 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9183 return true;
9184
9185 return false;
9186}
9187
Vincent Guittot09a43ac2016-11-08 10:53:45 +01009188#ifdef CONFIG_FAIR_GROUP_SCHED
9189/*
9190 * Propagate the changes of the sched_entity across the tg tree to make it
9191 * visible to the root
9192 */
9193static void propagate_entity_cfs_rq(struct sched_entity *se)
9194{
9195 struct cfs_rq *cfs_rq;
9196
9197 /* Start to propagate at parent */
9198 se = se->parent;
9199
9200 for_each_sched_entity(se) {
9201 cfs_rq = cfs_rq_of(se);
9202
9203 if (cfs_rq_throttled(cfs_rq))
9204 break;
9205
9206 update_load_avg(se, UPDATE_TG);
9207 }
9208}
9209#else
9210static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9211#endif
9212
Vincent Guittotdf217912016-11-08 10:53:42 +01009213static void detach_entity_cfs_rq(struct sched_entity *se)
Peter Zijlstrada7a7352011-01-17 17:03:27 +01009214{
Peter Zijlstrada7a7352011-01-17 17:03:27 +01009215 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9216
Yuyang Du9d89c252015-07-15 08:04:37 +08009217 /* Catch up with the cfs_rq and remove our load when we leave */
Vincent Guittotd31b1a62016-11-08 10:53:44 +01009218 update_load_avg(se, 0);
Byungchul Parka05e8c52015-08-20 20:21:56 +09009219 detach_entity_load_avg(cfs_rq, se);
Peter Zijlstra7c3edd22016-07-13 10:56:25 +02009220 update_tg_load_avg(cfs_rq, false);
Vincent Guittot09a43ac2016-11-08 10:53:45 +01009221 propagate_entity_cfs_rq(se);
Peter Zijlstrada7a7352011-01-17 17:03:27 +01009222}
9223
Vincent Guittotdf217912016-11-08 10:53:42 +01009224static void attach_entity_cfs_rq(struct sched_entity *se)
Steven Rostedtcb469842008-01-25 21:08:22 +01009225{
Byungchul Parkdaa59402015-08-20 20:22:00 +09009226 struct cfs_rq *cfs_rq = cfs_rq_of(se);
Byungchul Park7855a352015-08-10 18:02:55 +09009227
9228#ifdef CONFIG_FAIR_GROUP_SCHED
Michael wangeb7a59b2014-02-20 11:14:53 +08009229 /*
9230 * Since the real-depth could have been changed (only FAIR
9231 * class maintain depth value), reset depth properly.
9232 */
9233 se->depth = se->parent ? se->parent->depth + 1 : 0;
9234#endif
Byungchul Park7855a352015-08-10 18:02:55 +09009235
Vincent Guittotdf217912016-11-08 10:53:42 +01009236 /* Synchronize entity with its cfs_rq */
Vincent Guittotd31b1a62016-11-08 10:53:44 +01009237 update_load_avg(se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
Byungchul Parkdaa59402015-08-20 20:22:00 +09009238 attach_entity_load_avg(cfs_rq, se);
Peter Zijlstra7c3edd22016-07-13 10:56:25 +02009239 update_tg_load_avg(cfs_rq, false);
Vincent Guittot09a43ac2016-11-08 10:53:45 +01009240 propagate_entity_cfs_rq(se);
Vincent Guittotdf217912016-11-08 10:53:42 +01009241}
9242
9243static void detach_task_cfs_rq(struct task_struct *p)
9244{
9245 struct sched_entity *se = &p->se;
9246 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9247
9248 if (!vruntime_normalized(p)) {
9249 /*
9250 * Fix up our vruntime so that the current sleep doesn't
9251 * cause 'unlimited' sleep bonus.
9252 */
9253 place_entity(cfs_rq, se, 0);
9254 se->vruntime -= cfs_rq->min_vruntime;
9255 }
9256
9257 detach_entity_cfs_rq(se);
9258}
9259
9260static void attach_task_cfs_rq(struct task_struct *p)
9261{
9262 struct sched_entity *se = &p->se;
9263 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9264
9265 attach_entity_cfs_rq(se);
Byungchul Park6efdb102015-08-20 20:21:59 +09009266
Byungchul Parkdaa59402015-08-20 20:22:00 +09009267 if (!vruntime_normalized(p))
9268 se->vruntime += cfs_rq->min_vruntime;
9269}
Byungchul Park7855a352015-08-10 18:02:55 +09009270
Byungchul Parkdaa59402015-08-20 20:22:00 +09009271static void switched_from_fair(struct rq *rq, struct task_struct *p)
9272{
9273 detach_task_cfs_rq(p);
9274}
9275
9276static void switched_to_fair(struct rq *rq, struct task_struct *p)
9277{
9278 attach_task_cfs_rq(p);
9279
9280 if (task_on_rq_queued(p)) {
Byungchul Park7855a352015-08-10 18:02:55 +09009281 /*
Byungchul Parkdaa59402015-08-20 20:22:00 +09009282 * We were most likely switched from sched_rt, so
9283 * kick off the schedule if running, otherwise just see
9284 * if we can still preempt the current task.
Byungchul Park7855a352015-08-10 18:02:55 +09009285 */
Byungchul Parkdaa59402015-08-20 20:22:00 +09009286 if (rq->curr == p)
9287 resched_curr(rq);
9288 else
9289 check_preempt_curr(rq, p, 0);
Byungchul Park7855a352015-08-10 18:02:55 +09009290 }
Steven Rostedtcb469842008-01-25 21:08:22 +01009291}
9292
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02009293/* Account for a task changing its policy or group.
9294 *
9295 * This routine is mostly called to set cfs_rq->curr field when a task
9296 * migrates between groups/classes.
9297 */
9298static void set_curr_task_fair(struct rq *rq)
9299{
9300 struct sched_entity *se = &rq->curr->se;
9301
Paul Turnerec12cb72011-07-21 09:43:30 -07009302 for_each_sched_entity(se) {
9303 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9304
9305 set_next_entity(cfs_rq, se);
9306 /* ensure bandwidth has been allocated on our new cfs_rq */
9307 account_cfs_rq_runtime(cfs_rq, 0);
9308 }
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02009309}
9310
Peter Zijlstra029632f2011-10-25 10:00:11 +02009311void init_cfs_rq(struct cfs_rq *cfs_rq)
9312{
9313 cfs_rq->tasks_timeline = RB_ROOT;
Peter Zijlstra029632f2011-10-25 10:00:11 +02009314 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9315#ifndef CONFIG_64BIT
9316 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9317#endif
Alex Shi141965c2013-06-26 13:05:39 +08009318#ifdef CONFIG_SMP
Vincent Guittot09a43ac2016-11-08 10:53:45 +01009319#ifdef CONFIG_FAIR_GROUP_SCHED
9320 cfs_rq->propagate_avg = 0;
9321#endif
Yuyang Du9d89c252015-07-15 08:04:37 +08009322 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9323 atomic_long_set(&cfs_rq->removed_util_avg, 0);
Paul Turner9ee474f2012-10-04 13:18:30 +02009324#endif
Peter Zijlstra029632f2011-10-25 10:00:11 +02009325}
9326
Peter Zijlstra810b3812008-02-29 15:21:01 -05009327#ifdef CONFIG_FAIR_GROUP_SCHED
Vincent Guittotea86cb42016-06-17 13:38:55 +02009328static void task_set_group_fair(struct task_struct *p)
9329{
9330 struct sched_entity *se = &p->se;
9331
9332 set_task_rq(p, task_cpu(p));
9333 se->depth = se->parent ? se->parent->depth + 1 : 0;
9334}
9335
Peter Zijlstrabc54da22015-08-31 17:13:55 +02009336static void task_move_group_fair(struct task_struct *p)
Peter Zijlstra810b3812008-02-29 15:21:01 -05009337{
Byungchul Parkdaa59402015-08-20 20:22:00 +09009338 detach_task_cfs_rq(p);
Peter Zijlstrab2b5ce02010-10-15 15:24:15 +02009339 set_task_rq(p, task_cpu(p));
Byungchul Park6efdb102015-08-20 20:21:59 +09009340
9341#ifdef CONFIG_SMP
9342 /* Tell se's cfs_rq has been changed -- migrated */
9343 p->se.avg.last_update_time = 0;
9344#endif
Byungchul Parkdaa59402015-08-20 20:22:00 +09009345 attach_task_cfs_rq(p);
Peter Zijlstra810b3812008-02-29 15:21:01 -05009346}
Peter Zijlstra029632f2011-10-25 10:00:11 +02009347
Vincent Guittotea86cb42016-06-17 13:38:55 +02009348static void task_change_group_fair(struct task_struct *p, int type)
9349{
9350 switch (type) {
9351 case TASK_SET_GROUP:
9352 task_set_group_fair(p);
9353 break;
9354
9355 case TASK_MOVE_GROUP:
9356 task_move_group_fair(p);
9357 break;
9358 }
9359}
9360
Peter Zijlstra029632f2011-10-25 10:00:11 +02009361void free_fair_sched_group(struct task_group *tg)
9362{
9363 int i;
9364
9365 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9366
9367 for_each_possible_cpu(i) {
9368 if (tg->cfs_rq)
9369 kfree(tg->cfs_rq[i]);
Peter Zijlstra6fe1f342016-01-21 22:24:16 +01009370 if (tg->se)
Peter Zijlstra029632f2011-10-25 10:00:11 +02009371 kfree(tg->se[i]);
9372 }
9373
9374 kfree(tg->cfs_rq);
9375 kfree(tg->se);
9376}
9377
9378int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9379{
Peter Zijlstra029632f2011-10-25 10:00:11 +02009380 struct sched_entity *se;
Peter Zijlstrab7fa30c2016-06-09 15:07:50 +02009381 struct cfs_rq *cfs_rq;
Peter Zijlstra029632f2011-10-25 10:00:11 +02009382 int i;
9383
9384 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9385 if (!tg->cfs_rq)
9386 goto err;
9387 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9388 if (!tg->se)
9389 goto err;
9390
9391 tg->shares = NICE_0_LOAD;
9392
9393 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9394
9395 for_each_possible_cpu(i) {
9396 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9397 GFP_KERNEL, cpu_to_node(i));
9398 if (!cfs_rq)
9399 goto err;
9400
9401 se = kzalloc_node(sizeof(struct sched_entity),
9402 GFP_KERNEL, cpu_to_node(i));
9403 if (!se)
9404 goto err_free_rq;
9405
9406 init_cfs_rq(cfs_rq);
9407 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
Yuyang Du540247f2015-07-15 08:04:39 +08009408 init_entity_runnable_average(se);
Peter Zijlstra029632f2011-10-25 10:00:11 +02009409 }
9410
9411 return 1;
9412
9413err_free_rq:
9414 kfree(cfs_rq);
9415err:
9416 return 0;
9417}
9418
Peter Zijlstra8663e242016-06-22 14:58:02 +02009419void online_fair_sched_group(struct task_group *tg)
9420{
9421 struct sched_entity *se;
9422 struct rq *rq;
9423 int i;
9424
9425 for_each_possible_cpu(i) {
9426 rq = cpu_rq(i);
9427 se = tg->se[i];
9428
9429 raw_spin_lock_irq(&rq->lock);
Peter Zijlstra4126bad2016-10-03 16:20:59 +02009430 update_rq_clock(rq);
Vincent Guittotd0326692016-11-08 10:53:47 +01009431 attach_entity_cfs_rq(se);
Peter Zijlstra55e16d32016-06-22 15:14:26 +02009432 sync_throttle(tg, i);
Peter Zijlstra8663e242016-06-22 14:58:02 +02009433 raw_spin_unlock_irq(&rq->lock);
9434 }
9435}
9436
Peter Zijlstra6fe1f342016-01-21 22:24:16 +01009437void unregister_fair_sched_group(struct task_group *tg)
Peter Zijlstra029632f2011-10-25 10:00:11 +02009438{
Peter Zijlstra029632f2011-10-25 10:00:11 +02009439 unsigned long flags;
Peter Zijlstra6fe1f342016-01-21 22:24:16 +01009440 struct rq *rq;
9441 int cpu;
Peter Zijlstra029632f2011-10-25 10:00:11 +02009442
Peter Zijlstra6fe1f342016-01-21 22:24:16 +01009443 for_each_possible_cpu(cpu) {
9444 if (tg->se[cpu])
9445 remove_entity_load_avg(tg->se[cpu]);
Peter Zijlstra029632f2011-10-25 10:00:11 +02009446
Peter Zijlstra6fe1f342016-01-21 22:24:16 +01009447 /*
9448 * Only empty task groups can be destroyed; so we can speculatively
9449 * check on_list without danger of it being re-added.
9450 */
9451 if (!tg->cfs_rq[cpu]->on_list)
9452 continue;
9453
9454 rq = cpu_rq(cpu);
9455
9456 raw_spin_lock_irqsave(&rq->lock, flags);
9457 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9458 raw_spin_unlock_irqrestore(&rq->lock, flags);
9459 }
Peter Zijlstra029632f2011-10-25 10:00:11 +02009460}
9461
9462void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9463 struct sched_entity *se, int cpu,
9464 struct sched_entity *parent)
9465{
9466 struct rq *rq = cpu_rq(cpu);
9467
9468 cfs_rq->tg = tg;
9469 cfs_rq->rq = rq;
Peter Zijlstra029632f2011-10-25 10:00:11 +02009470 init_cfs_rq_runtime(cfs_rq);
9471
9472 tg->cfs_rq[cpu] = cfs_rq;
9473 tg->se[cpu] = se;
9474
9475 /* se could be NULL for root_task_group */
9476 if (!se)
9477 return;
9478
Peter Zijlstrafed14d42012-02-11 06:05:00 +01009479 if (!parent) {
Peter Zijlstra029632f2011-10-25 10:00:11 +02009480 se->cfs_rq = &rq->cfs;
Peter Zijlstrafed14d42012-02-11 06:05:00 +01009481 se->depth = 0;
9482 } else {
Peter Zijlstra029632f2011-10-25 10:00:11 +02009483 se->cfs_rq = parent->my_q;
Peter Zijlstrafed14d42012-02-11 06:05:00 +01009484 se->depth = parent->depth + 1;
9485 }
Peter Zijlstra029632f2011-10-25 10:00:11 +02009486
9487 se->my_q = cfs_rq;
Paul Turner0ac9b1c2013-10-16 11:16:27 -07009488 /* guarantee group entities always have weight */
9489 update_load_set(&se->load, NICE_0_LOAD);
Peter Zijlstra029632f2011-10-25 10:00:11 +02009490 se->parent = parent;
9491}
9492
9493static DEFINE_MUTEX(shares_mutex);
9494
9495int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9496{
9497 int i;
Peter Zijlstra029632f2011-10-25 10:00:11 +02009498
9499 /*
9500 * We can't change the weight of the root cgroup.
9501 */
9502 if (!tg->se[0])
9503 return -EINVAL;
9504
9505 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9506
9507 mutex_lock(&shares_mutex);
9508 if (tg->shares == shares)
9509 goto done;
9510
9511 tg->shares = shares;
9512 for_each_possible_cpu(i) {
9513 struct rq *rq = cpu_rq(i);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009514 struct sched_entity *se = tg->se[i];
9515 struct rq_flags rf;
Peter Zijlstra029632f2011-10-25 10:00:11 +02009516
Peter Zijlstra029632f2011-10-25 10:00:11 +02009517 /* Propagate contribution to hierarchy */
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009518 rq_lock_irqsave(rq, &rf);
Frederic Weisbecker71b1da42013-04-12 01:50:59 +02009519 update_rq_clock(rq);
Vincent Guittot89ee0482016-12-21 16:50:26 +01009520 for_each_sched_entity(se) {
9521 update_load_avg(se, UPDATE_TG);
9522 update_cfs_shares(se);
9523 }
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009524 rq_unlock_irqrestore(rq, &rf);
Peter Zijlstra029632f2011-10-25 10:00:11 +02009525 }
9526
9527done:
9528 mutex_unlock(&shares_mutex);
9529 return 0;
9530}
9531#else /* CONFIG_FAIR_GROUP_SCHED */
9532
9533void free_fair_sched_group(struct task_group *tg) { }
9534
9535int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9536{
9537 return 1;
9538}
9539
Peter Zijlstra8663e242016-06-22 14:58:02 +02009540void online_fair_sched_group(struct task_group *tg) { }
9541
Peter Zijlstra6fe1f342016-01-21 22:24:16 +01009542void unregister_fair_sched_group(struct task_group *tg) { }
Peter Zijlstra029632f2011-10-25 10:00:11 +02009543
9544#endif /* CONFIG_FAIR_GROUP_SCHED */
9545
Peter Zijlstra810b3812008-02-29 15:21:01 -05009546
H Hartley Sweeten6d686f42010-01-13 20:21:52 -07009547static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
Peter Williams0d721ce2009-09-21 01:31:53 +00009548{
9549 struct sched_entity *se = &task->se;
Peter Williams0d721ce2009-09-21 01:31:53 +00009550 unsigned int rr_interval = 0;
9551
9552 /*
9553 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9554 * idle runqueue:
9555 */
Peter Williams0d721ce2009-09-21 01:31:53 +00009556 if (rq->cfs.load.weight)
Zhu Yanhaia59f4e02013-01-08 12:56:52 +08009557 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
Peter Williams0d721ce2009-09-21 01:31:53 +00009558
9559 return rr_interval;
9560}
9561
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009562/*
9563 * All the scheduling class methods:
9564 */
Peter Zijlstra029632f2011-10-25 10:00:11 +02009565const struct sched_class fair_sched_class = {
Ingo Molnar5522d5d2007-10-15 17:00:12 +02009566 .next = &idle_sched_class,
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009567 .enqueue_task = enqueue_task_fair,
9568 .dequeue_task = dequeue_task_fair,
9569 .yield_task = yield_task_fair,
Mike Galbraithd95f4122011-02-01 09:50:51 -05009570 .yield_to_task = yield_to_task_fair,
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009571
Ingo Molnar2e09bf52007-10-15 17:00:05 +02009572 .check_preempt_curr = check_preempt_wakeup,
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009573
9574 .pick_next_task = pick_next_task_fair,
9575 .put_prev_task = put_prev_task_fair,
9576
Peter Williams681f3e62007-10-24 18:23:51 +02009577#ifdef CONFIG_SMP
Li Zefan4ce72a22008-10-22 15:25:26 +08009578 .select_task_rq = select_task_rq_fair,
Paul Turner0a74bef2012-10-04 13:18:30 +02009579 .migrate_task_rq = migrate_task_rq_fair,
Alex Shi141965c2013-06-26 13:05:39 +08009580
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01009581 .rq_online = rq_online_fair,
9582 .rq_offline = rq_offline_fair,
Peter Zijlstra88ec22d2009-12-16 18:04:41 +01009583
Yuyang Du12695572015-07-15 08:04:40 +08009584 .task_dead = task_dead_fair,
Peter Zijlstrac5b28032015-05-15 17:43:35 +02009585 .set_cpus_allowed = set_cpus_allowed_common,
Peter Williams681f3e62007-10-24 18:23:51 +02009586#endif
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009587
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02009588 .set_curr_task = set_curr_task_fair,
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009589 .task_tick = task_tick_fair,
Peter Zijlstracd29fe62009-11-27 17:32:46 +01009590 .task_fork = task_fork_fair,
Steven Rostedtcb469842008-01-25 21:08:22 +01009591
9592 .prio_changed = prio_changed_fair,
Peter Zijlstrada7a7352011-01-17 17:03:27 +01009593 .switched_from = switched_from_fair,
Steven Rostedtcb469842008-01-25 21:08:22 +01009594 .switched_to = switched_to_fair,
Peter Zijlstra810b3812008-02-29 15:21:01 -05009595
Peter Williams0d721ce2009-09-21 01:31:53 +00009596 .get_rr_interval = get_rr_interval_fair,
9597
Stanislaw Gruszka6e998912014-11-12 16:58:44 +01009598 .update_curr = update_curr_fair,
9599
Peter Zijlstra810b3812008-02-29 15:21:01 -05009600#ifdef CONFIG_FAIR_GROUP_SCHED
Vincent Guittotea86cb42016-06-17 13:38:55 +02009601 .task_change_group = task_change_group_fair,
Peter Zijlstra810b3812008-02-29 15:21:01 -05009602#endif
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009603};
9604
9605#ifdef CONFIG_SCHED_DEBUG
Peter Zijlstra029632f2011-10-25 10:00:11 +02009606void print_cfs_stats(struct seq_file *m, int cpu)
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009607{
Tejun Heoa9e7f652017-04-25 17:43:50 -07009608 struct cfs_rq *cfs_rq, *pos;
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009609
Peter Zijlstra5973e5b2008-01-25 21:08:34 +01009610 rcu_read_lock();
Tejun Heoa9e7f652017-04-25 17:43:50 -07009611 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
Ingo Molnar5cef9ec2007-08-09 11:16:47 +02009612 print_cfs_rq(m, cpu, cfs_rq);
Peter Zijlstra5973e5b2008-01-25 21:08:34 +01009613 rcu_read_unlock();
Ingo Molnarbf0f6f22007-07-09 18:51:58 +02009614}
Srikar Dronamraju397f2372015-06-25 22:51:43 +05309615
9616#ifdef CONFIG_NUMA_BALANCING
9617void show_numa_stats(struct task_struct *p, struct seq_file *m)
9618{
9619 int node;
9620 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9621
9622 for_each_online_node(node) {
9623 if (p->numa_faults) {
9624 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9625 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9626 }
9627 if (p->numa_group) {
9628 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9629 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9630 }
9631 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9632 }
9633}
9634#endif /* CONFIG_NUMA_BALANCING */
9635#endif /* CONFIG_SCHED_DEBUG */
Peter Zijlstra029632f2011-10-25 10:00:11 +02009636
9637__init void init_sched_fair_class(void)
9638{
9639#ifdef CONFIG_SMP
9640 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9641
Frederic Weisbecker3451d022011-08-10 23:21:01 +02009642#ifdef CONFIG_NO_HZ_COMMON
Diwakar Tundlam554ceca2012-03-07 14:44:26 -08009643 nohz.next_balance = jiffies;
Peter Zijlstra029632f2011-10-25 10:00:11 +02009644 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
Peter Zijlstra029632f2011-10-25 10:00:11 +02009645#endif
9646#endif /* SMP */
9647
9648}