root/kernel/sched/fair.c

DEFINITIONS

1. arch_asym_cpu_priority
2. update_load_add
3. update_load_sub
4. update_load_set
5. get_update_sysctl_factor
6. update_sysctl
7. sched_init_granularity
8. __update_inv_weight
9. __calc_delta
10. task_of
11. task_cfs_rq
12. cfs_rq_of
13. group_cfs_rq
14. cfs_rq_tg_path
15. list_add_leaf_cfs_rq
16. list_del_leaf_cfs_rq
17. assert_list_leaf_cfs_rq
18. is_same_group
19. parent_entity
20. find_matching_se
21. task_of
22. task_cfs_rq
23. cfs_rq_of
24. group_cfs_rq
25. cfs_rq_tg_path
26. list_add_leaf_cfs_rq
27. list_del_leaf_cfs_rq
28. assert_list_leaf_cfs_rq
29. parent_entity
30. find_matching_se
31. max_vruntime
32. min_vruntime
33. entity_before
34. update_min_vruntime
35. __enqueue_entity
36. __dequeue_entity
37. __pick_first_entity
38. __pick_next_entity
39. __pick_last_entity
40. sched_proc_update_handler
41. calc_delta_fair
42. __sched_period
43. sched_slice
44. sched_vslice
45. init_entity_runnable_average
46. post_init_entity_util_avg
47. init_entity_runnable_average
48. post_init_entity_util_avg
49. update_tg_load_avg
50. update_curr
51. update_curr_fair
52. update_stats_wait_start
53. update_stats_wait_end
54. update_stats_enqueue_sleeper
55. update_stats_enqueue
56. update_stats_dequeue
57. update_stats_curr_start
58. deref_task_numa_group
59. deref_curr_numa_group
60. task_nr_scan_windows
61. task_scan_min
62. task_scan_start
63. task_scan_max
64. account_numa_enqueue
65. account_numa_dequeue
66. task_numa_group_id
67. task_faults_idx
68. task_faults
69. group_faults
70. group_faults_cpu
71. group_faults_priv
72. group_faults_shared
73. numa_is_active_node
74. score_nearby_nodes
75. task_weight
76. group_weight
77. should_numa_migrate_memory
78. update_numa_stats
79. task_numa_assign
80. load_too_imbalanced
81. task_numa_compare
82. task_numa_find_cpu
83. task_numa_migrate
84. numa_migrate_preferred
85. numa_group_count_active_nodes
86. update_task_scan_period
87. numa_get_avg_runtime
88. preferred_group_nid
89. task_numa_placement
90. get_numa_group
91. put_numa_group
92. task_numa_group
93. task_numa_free
94. task_numa_fault
95. reset_ptenuma_scan
96. task_numa_work
97. init_numa_balancing
98. task_tick_numa
99. update_scan_period
100. task_tick_numa
101. account_numa_enqueue
102. account_numa_dequeue
103. update_scan_period
104. account_entity_enqueue
105. account_entity_dequeue
106. enqueue_runnable_load_avg
107. dequeue_runnable_load_avg
108. enqueue_load_avg
109. dequeue_load_avg
110. enqueue_runnable_load_avg
111. dequeue_runnable_load_avg
112. enqueue_load_avg
113. dequeue_load_avg
114. reweight_entity
115. reweight_task
116. calc_group_shares
117. calc_group_runnable
118. update_cfs_group
119. update_cfs_group
120. cfs_rq_util_change
121. update_tg_load_avg
122. set_task_rq_fair
123. update_tg_cfs_util
124. update_tg_cfs_runnable
125. add_tg_cfs_propagate
126. propagate_entity_load_avg
127. skip_blocked_update
128. update_tg_load_avg
129. propagate_entity_load_avg
130. add_tg_cfs_propagate
131. update_cfs_rq_load_avg
132. attach_entity_load_avg
133. detach_entity_load_avg
134. update_load_avg
135. cfs_rq_last_update_time
136. cfs_rq_last_update_time
137. sync_entity_load_avg
138. remove_entity_load_avg
139. cfs_rq_runnable_load_avg
140. cfs_rq_load_avg
141. task_util
142. _task_util_est
143. task_util_est
144. util_est_enqueue
145. within_margin
146. util_est_dequeue
147. task_fits_capacity
148. update_misfit_status
149. update_load_avg
150. remove_entity_load_avg
151. attach_entity_load_avg
152. detach_entity_load_avg
153. idle_balance
154. util_est_enqueue
155. util_est_dequeue
156. update_misfit_status
157. check_spread
158. place_entity
159. check_schedstat_required
160. enqueue_entity
161. __clear_buddies_last
162. __clear_buddies_next
163. __clear_buddies_skip
164. clear_buddies
165. dequeue_entity
166. check_preempt_tick
167. set_next_entity
168. pick_next_entity
169. put_prev_entity
170. entity_tick
171. cfs_bandwidth_used
172. cfs_bandwidth_usage_inc
173. cfs_bandwidth_usage_dec
174. cfs_bandwidth_used
175. cfs_bandwidth_usage_inc
176. cfs_bandwidth_usage_dec
177. default_cfs_period
178. sched_cfs_bandwidth_slice
179. __refill_cfs_bandwidth_runtime
180. tg_cfs_bandwidth
181. assign_cfs_rq_runtime
182. __account_cfs_rq_runtime
183. account_cfs_rq_runtime
184. cfs_rq_throttled
185. throttled_hierarchy
186. throttled_lb_pair
187. tg_unthrottle_up
188. tg_throttle_down
189. throttle_cfs_rq
190. unthrottle_cfs_rq
191. distribute_cfs_runtime
192. do_sched_cfs_period_timer
193. runtime_refresh_within
194. start_cfs_slack_bandwidth
195. __return_cfs_rq_runtime
196. return_cfs_rq_runtime
197. do_sched_cfs_slack_timer
198. check_enqueue_throttle
199. sync_throttle
200. check_cfs_rq_runtime
201. sched_cfs_slack_timer
202. sched_cfs_period_timer
203. init_cfs_bandwidth
204. init_cfs_rq_runtime
205. start_cfs_bandwidth
206. destroy_cfs_bandwidth
207. update_runtime_enabled
208. unthrottle_offline_cfs_rqs
209. cfs_bandwidth_used
210. account_cfs_rq_runtime
211. check_cfs_rq_runtime
212. check_enqueue_throttle
213. sync_throttle
214. return_cfs_rq_runtime
215. cfs_rq_throttled
216. throttled_hierarchy
217. throttled_lb_pair
218. init_cfs_bandwidth
219. init_cfs_rq_runtime
220. tg_cfs_bandwidth
221. destroy_cfs_bandwidth
222. update_runtime_enabled
223. unthrottle_offline_cfs_rqs
224. hrtick_start_fair
225. hrtick_update
226. hrtick_start_fair
227. hrtick_update
228. cpu_overutilized
229. update_overutilized_status
230. update_overutilized_status
231. enqueue_task_fair
232. dequeue_task_fair
233. sched_idle_cpu
234. cpu_runnable_load
235. capacity_of
236. cpu_avg_load_per_task
237. record_wakee
238. wake_wide
239. wake_affine_idle
240. wake_affine_weight
241. wake_affine
242. capacity_spare_without
243. find_idlest_group
244. find_idlest_group_cpu
245. find_idlest_cpu
246. set_idle_cores
247. test_idle_cores
248. __update_idle_core
249. select_idle_core
250. select_idle_smt
251. select_idle_core
252. select_idle_smt
253. select_idle_cpu
254. select_idle_sibling
255. cpu_util
256. cpu_util_without
257. wake_cap
258. cpu_util_next
259. compute_energy
260. find_energy_efficient_cpu
261. select_task_rq_fair
262. migrate_task_rq_fair
263. task_dead_fair
264. balance_fair
265. wakeup_gran
266. wakeup_preempt_entity
267. set_last_buddy
268. set_next_buddy
269. set_skip_buddy
270. check_preempt_wakeup
271. pick_next_task_fair
272. put_prev_task_fair
273. yield_task_fair
274. yield_to_task_fair
275. task_hot
276. migrate_degrades_locality
277. migrate_degrades_locality
278. can_migrate_task
279. detach_task
280. detach_one_task
281. detach_tasks
282. attach_task
283. attach_one_task
284. attach_tasks
285. cfs_rq_has_blocked
286. others_have_blocked
287. update_blocked_load_status
288. cfs_rq_has_blocked
289. others_have_blocked
290. update_blocked_load_status
291. __update_blocked_others
292. cfs_rq_is_decayed
293. __update_blocked_fair
294. update_cfs_rq_h_load
295. task_h_load
296. __update_blocked_fair
297. task_h_load
298. update_blocked_averages
299. init_sd_lb_stats
300. scale_rt_capacity
301. update_cpu_capacity
302. update_group_capacity
303. check_cpu_capacity
304. check_misfit_status
305. sg_imbalanced
306. group_has_capacity
307. group_is_overloaded
308. group_smaller_min_cpu_capacity
309. group_smaller_max_cpu_capacity
310. group_classify
311. update_nohz_stats
312. update_sg_lb_stats
313. update_sd_pick_busiest
314. fbq_classify_group
315. fbq_classify_rq
316. fbq_classify_group
317. fbq_classify_rq
318. update_sd_lb_stats
319. check_asym_packing
320. fix_small_imbalance
321. calculate_imbalance
322. find_busiest_group
323. find_busiest_queue
324. asym_active_balance
325. voluntary_active_balance
326. need_active_balance
327. should_we_balance
328. load_balance
329. get_sd_balance_interval
330. update_next_balance
331. active_load_balance_cpu_stop
332. update_max_interval
333. rebalance_domains
334. on_null_domain
335. find_new_ilb
336. kick_ilb
337. nohz_balancer_kick
338. set_cpu_sd_state_busy
339. nohz_balance_exit_idle
340. set_cpu_sd_state_idle
341. nohz_balance_enter_idle
342. _nohz_idle_balance
343. nohz_idle_balance
344. nohz_newidle_balance
345. nohz_balancer_kick
346. nohz_idle_balance
347. nohz_newidle_balance
348. newidle_balance
349. run_rebalance_domains
350. trigger_load_balance
351. rq_online_fair
352. rq_offline_fair
353. task_tick_fair
354. task_fork_fair
355. prio_changed_fair
356. vruntime_normalized
357. propagate_entity_cfs_rq
358. propagate_entity_cfs_rq
359. detach_entity_cfs_rq
360. attach_entity_cfs_rq
361. detach_task_cfs_rq
362. attach_task_cfs_rq
363. switched_from_fair
364. switched_to_fair
365. set_next_task_fair
366. init_cfs_rq
367. task_set_group_fair
368. task_move_group_fair
369. task_change_group_fair
370. free_fair_sched_group
371. alloc_fair_sched_group
372. online_fair_sched_group
373. unregister_fair_sched_group
374. init_tg_cfs_entry
375. sched_group_set_shares
376. free_fair_sched_group
377. alloc_fair_sched_group
378. online_fair_sched_group
379. unregister_fair_sched_group
380. get_rr_interval_fair
381. print_cfs_stats
382. show_numa_stats
383. init_sched_fair_class
384. sched_trace_cfs_rq_avg
385. sched_trace_cfs_rq_path
386. sched_trace_cfs_rq_cpu
387. sched_trace_rq_avg_rt
388. sched_trace_rq_avg_dl
389. sched_trace_rq_avg_irq
390. sched_trace_rq_cpu
391. sched_trace_rd_span

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
   4  *
   5  *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
   6  *
   7  *  Interactivity improvements by Mike Galbraith
   8  *  (C) 2007 Mike Galbraith <efault@gmx.de>
   9  *
  10  *  Various enhancements by Dmitry Adamushko.
  11  *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
  12  *
  13  *  Group scheduling enhancements by Srivatsa Vaddagiri
  14  *  Copyright IBM Corporation, 2007
  15  *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
  16  *
  17  *  Scaled math optimizations by Thomas Gleixner
  18  *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
  19  *
  20  *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
  21  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
  22  */
  23 #include "sched.h"
  24 
  25 #include <trace/events/sched.h>
  26 
  27 /*
  28  * Targeted preemption latency for CPU-bound tasks:
  29  *
  30  * NOTE: this latency value is not the same as the concept of
  31  * 'timeslice length' - timeslices in CFS are of variable length
  32  * and have no persistent notion like in traditional, time-slice
  33  * based scheduling concepts.
  34  *
  35  * (to see the precise effective timeslice length of your workload,
  36  *  run vmstat and monitor the context-switches (cs) field)
  37  *
  38  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
  39  */
  40 unsigned int sysctl_sched_latency                       = 6000000ULL;
  41 static unsigned int normalized_sysctl_sched_latency     = 6000000ULL;
  42 
  43 /*
  44  * The initial- and re-scaling of tunables is configurable
  45  *
  46  * Options are:
  47  *
  48  *   SCHED_TUNABLESCALING_NONE - unscaled, always *1
  49  *   SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
  50  *   SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
  51  *
  52  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
  53  */
  54 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
  55 
  56 /*
  57  * Minimal preemption granularity for CPU-bound tasks:
  58  *
  59  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
  60  */
  61 unsigned int sysctl_sched_min_granularity                       = 750000ULL;
  62 static unsigned int normalized_sysctl_sched_min_granularity     = 750000ULL;
  63 
  64 /*
  65  * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
  66  */
  67 static unsigned int sched_nr_latency = 8;
  68 
  69 /*
  70  * After fork, child runs first. If set to 0 (default) then
  71  * parent will (try to) run first.
  72  */
  73 unsigned int sysctl_sched_child_runs_first __read_mostly;
  74 
  75 /*
  76  * SCHED_OTHER wake-up granularity.
  77  *
  78  * This option delays the preemption effects of decoupled workloads
  79  * and reduces their over-scheduling. Synchronous workloads will still
  80  * have immediate wakeup/sleep latencies.
  81  *
  82  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
  83  */
  84 unsigned int sysctl_sched_wakeup_granularity                    = 1000000UL;
  85 static unsigned int normalized_sysctl_sched_wakeup_granularity  = 1000000UL;
  86 
  87 const_debug unsigned int sysctl_sched_migration_cost    = 500000UL;
  88 
  89 #ifdef CONFIG_SMP
  90 /*
  91  * For asym packing, by default the lower numbered CPU has higher priority.
  92  */
  93 int __weak arch_asym_cpu_priority(int cpu)
  94 {
  95         return -cpu;
  96 }
  97 
  98 /*
  99  * The margin used when comparing utilization with CPU capacity.
 100  *
 101  * (default: ~20%)
 102  */
 103 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
 104 
 105 #endif
 106 
 107 #ifdef CONFIG_CFS_BANDWIDTH
 108 /*
 109  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
 110  * each time a cfs_rq requests quota.
 111  *
 112  * Note: in the case that the slice exceeds the runtime remaining (either due
 113  * to consumption or the quota being specified to be smaller than the slice)
 114  * we will always only issue the remaining available time.
 115  *
 116  * (default: 5 msec, units: microseconds)
 117  */
 118 unsigned int sysctl_sched_cfs_bandwidth_slice           = 5000UL;
 119 #endif
 120 
 121 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
 122 {
 123         lw->weight += inc;
 124         lw->inv_weight = 0;
 125 }
 126 
 127 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
 128 {
 129         lw->weight -= dec;
 130         lw->inv_weight = 0;
 131 }
 132 
 133 static inline void update_load_set(struct load_weight *lw, unsigned long w)
 134 {
 135         lw->weight = w;
 136         lw->inv_weight = 0;
 137 }
 138 
 139 /*
 140  * Increase the granularity value when there are more CPUs,
 141  * because with more CPUs the 'effective latency' as visible
 142  * to users decreases. But the relationship is not linear,
 143  * so pick a second-best guess by going with the log2 of the
 144  * number of CPUs.
 145  *
 146  * This idea comes from the SD scheduler of Con Kolivas:
 147  */
 148 static unsigned int get_update_sysctl_factor(void)
 149 {
 150         unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
 151         unsigned int factor;
 152 
 153         switch (sysctl_sched_tunable_scaling) {
 154         case SCHED_TUNABLESCALING_NONE:
 155                 factor = 1;
 156                 break;
 157         case SCHED_TUNABLESCALING_LINEAR:
 158                 factor = cpus;
 159                 break;
 160         case SCHED_TUNABLESCALING_LOG:
 161         default:
 162                 factor = 1 + ilog2(cpus);
 163                 break;
 164         }
 165 
 166         return factor;
 167 }
 168 
 169 static void update_sysctl(void)
 170 {
 171         unsigned int factor = get_update_sysctl_factor();
 172 
 173 #define SET_SYSCTL(name) \
 174         (sysctl_##name = (factor) * normalized_sysctl_##name)
 175         SET_SYSCTL(sched_min_granularity);
 176         SET_SYSCTL(sched_latency);
 177         SET_SYSCTL(sched_wakeup_granularity);
 178 #undef SET_SYSCTL
 179 }
 180 
 181 void sched_init_granularity(void)
 182 {
 183         update_sysctl();
 184 }
 185 
 186 #define WMULT_CONST     (~0U)
 187 #define WMULT_SHIFT     32
 188 
 189 static void __update_inv_weight(struct load_weight *lw)
 190 {
 191         unsigned long w;
 192 
 193         if (likely(lw->inv_weight))
 194                 return;
 195 
 196         w = scale_load_down(lw->weight);
 197 
 198         if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
 199                 lw->inv_weight = 1;
 200         else if (unlikely(!w))
 201                 lw->inv_weight = WMULT_CONST;
 202         else
 203                 lw->inv_weight = WMULT_CONST / w;
 204 }
 205 
 206 /*
 207  * delta_exec * weight / lw.weight
 208  *   OR
 209  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
 210  *
 211  * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
 212  * we're guaranteed shift stays positive because inv_weight is guaranteed to
 213  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
 214  *
 215  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
 216  * weight/lw.weight <= 1, and therefore our shift will also be positive.
 217  */
 218 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
 219 {
 220         u64 fact = scale_load_down(weight);
 221         int shift = WMULT_SHIFT;
 222 
 223         __update_inv_weight(lw);
 224 
 225         if (unlikely(fact >> 32)) {
 226                 while (fact >> 32) {
 227                         fact >>= 1;
 228                         shift--;
 229                 }
 230         }
 231 
 232         /* hint to use a 32x32->64 mul */
 233         fact = (u64)(u32)fact * lw->inv_weight;
 234 
 235         while (fact >> 32) {
 236                 fact >>= 1;
 237                 shift--;
 238         }
 239 
 240         return mul_u64_u32_shr(delta_exec, fact, shift);
 241 }
 242 
 243 
 244 const struct sched_class fair_sched_class;
 245 
 246 /**************************************************************
 247  * CFS operations on generic schedulable entities:
 248  */
 249 
 250 #ifdef CONFIG_FAIR_GROUP_SCHED
 251 static inline struct task_struct *task_of(struct sched_entity *se)
 252 {
 253         SCHED_WARN_ON(!entity_is_task(se));
 254         return container_of(se, struct task_struct, se);
 255 }
 256 
 257 /* Walk up scheduling entities hierarchy */
 258 #define for_each_sched_entity(se) \
 259                 for (; se; se = se->parent)
 260 
 261 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
 262 {
 263         return p->se.cfs_rq;
 264 }
 265 
 266 /* runqueue on which this entity is (to be) queued */
 267 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
 268 {
 269         return se->cfs_rq;
 270 }
 271 
 272 /* runqueue "owned" by this group */
 273 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
 274 {
 275         return grp->my_q;
 276 }
 277 
 278 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
 279 {
 280         if (!path)
 281                 return;
 282 
 283         if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
 284                 autogroup_path(cfs_rq->tg, path, len);
 285         else if (cfs_rq && cfs_rq->tg->css.cgroup)
 286                 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
 287         else
 288                 strlcpy(path, "(null)", len);
 289 }
 290 
 291 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
 292 {
 293         struct rq *rq = rq_of(cfs_rq);
 294         int cpu = cpu_of(rq);
 295 
 296         if (cfs_rq->on_list)
 297                 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
 298 
 299         cfs_rq->on_list = 1;
 300 
 301         /*
 302          * Ensure we either appear before our parent (if already
 303          * enqueued) or force our parent to appear after us when it is
 304          * enqueued. The fact that we always enqueue bottom-up
 305          * reduces this to two cases and a special case for the root
 306          * cfs_rq. Furthermore, it also means that we will always reset
 307          * tmp_alone_branch either when the branch is connected
 308          * to a tree or when we reach the top of the tree
 309          */
 310         if (cfs_rq->tg->parent &&
 311             cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
 312                 /*
 313                  * If parent is already on the list, we add the child
 314                  * just before. Thanks to circular linked property of
 315                  * the list, this means to put the child at the tail
 316                  * of the list that starts by parent.
 317                  */
 318                 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
 319                         &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
 320                 /*
 321                  * The branch is now connected to its tree so we can
 322                  * reset tmp_alone_branch to the beginning of the
 323                  * list.
 324                  */
 325                 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
 326                 return true;
 327         }
 328 
 329         if (!cfs_rq->tg->parent) {
 330                 /*
 331                  * cfs rq without parent should be put
 332                  * at the tail of the list.
 333                  */
 334                 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
 335                         &rq->leaf_cfs_rq_list);
 336                 /*
 337                  * We have reach the top of a tree so we can reset
 338                  * tmp_alone_branch to the beginning of the list.
 339                  */
 340                 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
 341                 return true;
 342         }
 343 
 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 begin of the branch
 348          * where we will add parent.
 349          */
 350         list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
 351         /*
 352          * update tmp_alone_branch to points to the new begin
 353          * of the branch
 354          */
 355         rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
 356         return false;
 357 }
 358 
 359 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
 360 {
 361         if (cfs_rq->on_list) {
 362                 struct rq *rq = rq_of(cfs_rq);
 363 
 364                 /*
 365                  * With cfs_rq being unthrottled/throttled during an enqueue,
 366                  * it can happen the tmp_alone_branch points the a leaf that
 367                  * we finally want to del. In this case, tmp_alone_branch moves
 368                  * to the prev element but it will point to rq->leaf_cfs_rq_list
 369                  * at the end of the enqueue.
 370                  */
 371                 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
 372                         rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
 373 
 374                 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
 375                 cfs_rq->on_list = 0;
 376         }
 377 }
 378 
 379 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
 380 {
 381         SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
 382 }
 383 
 384 /* Iterate thr' all leaf cfs_rq's on a runqueue */
 385 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)                      \
 386         list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list,    \
 387                                  leaf_cfs_rq_list)
 388 
 389 /* Do the two (enqueued) entities belong to the same group ? */
 390 static inline struct cfs_rq *
 391 is_same_group(struct sched_entity *se, struct sched_entity *pse)
 392 {
 393         if (se->cfs_rq == pse->cfs_rq)
 394                 return se->cfs_rq;
 395 
 396         return NULL;
 397 }
 398 
 399 static inline struct sched_entity *parent_entity(struct sched_entity *se)
 400 {
 401         return se->parent;
 402 }
 403 
 404 static void
 405 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
 406 {
 407         int se_depth, pse_depth;
 408 
 409         /*
 410          * preemption test can be made between sibling entities who are in the
 411          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
 412          * both tasks until we find their ancestors who are siblings of common
 413          * parent.
 414          */
 415 
 416         /* First walk up until both entities are at same depth */
 417         se_depth = (*se)->depth;
 418         pse_depth = (*pse)->depth;
 419 
 420         while (se_depth > pse_depth) {
 421                 se_depth--;
 422                 *se = parent_entity(*se);
 423         }
 424 
 425         while (pse_depth > se_depth) {
 426                 pse_depth--;
 427                 *pse = parent_entity(*pse);
 428         }
 429 
 430         while (!is_same_group(*se, *pse)) {
 431                 *se = parent_entity(*se);
 432                 *pse = parent_entity(*pse);
 433         }
 434 }
 435 
 436 #else   /* !CONFIG_FAIR_GROUP_SCHED */
 437 
 438 static inline struct task_struct *task_of(struct sched_entity *se)
 439 {
 440         return container_of(se, struct task_struct, se);
 441 }
 442 
 443 #define for_each_sched_entity(se) \
 444                 for (; se; se = NULL)
 445 
 446 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
 447 {
 448         return &task_rq(p)->cfs;
 449 }
 450 
 451 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
 452 {
 453         struct task_struct *p = task_of(se);
 454         struct rq *rq = task_rq(p);
 455 
 456         return &rq->cfs;
 457 }
 458 
 459 /* runqueue "owned" by this group */
 460 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
 461 {
 462         return NULL;
 463 }
 464 
 465 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
 466 {
 467         if (path)
 468                 strlcpy(path, "(null)", len);
 469 }
 470 
 471 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
 472 {
 473         return true;
 474 }
 475 
 476 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
 477 {
 478 }
 479 
 480 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
 481 {
 482 }
 483 
 484 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)      \
 485                 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
 486 
 487 static inline struct sched_entity *parent_entity(struct sched_entity *se)
 488 {
 489         return NULL;
 490 }
 491 
 492 static inline void
 493 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
 494 {
 495 }
 496 
 497 #endif  /* CONFIG_FAIR_GROUP_SCHED */
 498 
 499 static __always_inline
 500 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
 501 
 502 /**************************************************************
 503  * Scheduling class tree data structure manipulation methods:
 504  */
 505 
 506 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
 507 {
 508         s64 delta = (s64)(vruntime - max_vruntime);
 509         if (delta > 0)
 510                 max_vruntime = vruntime;
 511 
 512         return max_vruntime;
 513 }
 514 
 515 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
 516 {
 517         s64 delta = (s64)(vruntime - min_vruntime);
 518         if (delta < 0)
 519                 min_vruntime = vruntime;
 520 
 521         return min_vruntime;
 522 }
 523 
 524 static inline int entity_before(struct sched_entity *a,
 525                                 struct sched_entity *b)
 526 {
 527         return (s64)(a->vruntime - b->vruntime) < 0;
 528 }
 529 
 530 static void update_min_vruntime(struct cfs_rq *cfs_rq)
 531 {
 532         struct sched_entity *curr = cfs_rq->curr;
 533         struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
 534 
 535         u64 vruntime = cfs_rq->min_vruntime;
 536 
 537         if (curr) {
 538                 if (curr->on_rq)
 539                         vruntime = curr->vruntime;
 540                 else
 541                         curr = NULL;
 542         }
 543 
 544         if (leftmost) { /* non-empty tree */
 545                 struct sched_entity *se;
 546                 se = rb_entry(leftmost, struct sched_entity, run_node);
 547 
 548                 if (!curr)
 549                         vruntime = se->vruntime;
 550                 else
 551                         vruntime = min_vruntime(vruntime, se->vruntime);
 552         }
 553 
 554         /* ensure we never gain time by being placed backwards. */
 555         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
 556 #ifndef CONFIG_64BIT
 557         smp_wmb();
 558         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
 559 #endif
 560 }
 561 
 562 /*
 563  * Enqueue an entity into the rb-tree:
 564  */
 565 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
 566 {
 567         struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
 568         struct rb_node *parent = NULL;
 569         struct sched_entity *entry;
 570         bool leftmost = true;
 571 
 572         /*
 573          * Find the right place in the rbtree:
 574          */
 575         while (*link) {
 576                 parent = *link;
 577                 entry = rb_entry(parent, struct sched_entity, run_node);
 578                 /*
 579                  * We dont care about collisions. Nodes with
 580                  * the same key stay together.
 581                  */
 582                 if (entity_before(se, entry)) {
 583                         link = &parent->rb_left;
 584                 } else {
 585                         link = &parent->rb_right;
 586                         leftmost = false;
 587                 }
 588         }
 589 
 590         rb_link_node(&se->run_node, parent, link);
 591         rb_insert_color_cached(&se->run_node,
 592                                &cfs_rq->tasks_timeline, leftmost);
 593 }
 594 
 595 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
 596 {
 597         rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
 598 }
 599 
 600 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
 601 {
 602         struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
 603 
 604         if (!left)
 605                 return NULL;
 606 
 607         return rb_entry(left, struct sched_entity, run_node);
 608 }
 609 
 610 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
 611 {
 612         struct rb_node *next = rb_next(&se->run_node);
 613 
 614         if (!next)
 615                 return NULL;
 616 
 617         return rb_entry(next, struct sched_entity, run_node);
 618 }
 619 
 620 #ifdef CONFIG_SCHED_DEBUG
 621 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
 622 {
 623         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
 624 
 625         if (!last)
 626                 return NULL;
 627 
 628         return rb_entry(last, struct sched_entity, run_node);
 629 }
 630 
 631 /**************************************************************
 632  * Scheduling class statistics methods:
 633  */
 634 
 635 int sched_proc_update_handler(struct ctl_table *table, int write,
 636                 void __user *buffer, size_t *lenp,
 637                 loff_t *ppos)
 638 {
 639         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 640         unsigned int factor = get_update_sysctl_factor();
 641 
 642         if (ret || !write)
 643                 return ret;
 644 
 645         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
 646                                         sysctl_sched_min_granularity);
 647 
 648 #define WRT_SYSCTL(name) \
 649         (normalized_sysctl_##name = sysctl_##name / (factor))
 650         WRT_SYSCTL(sched_min_granularity);
 651         WRT_SYSCTL(sched_latency);
 652         WRT_SYSCTL(sched_wakeup_granularity);
 653 #undef WRT_SYSCTL
 654 
 655         return 0;
 656 }
 657 #endif
 658 
 659 /*
 660  * delta /= w
 661  */
 662 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
 663 {
 664         if (unlikely(se->load.weight != NICE_0_LOAD))
 665                 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
 666 
 667         return delta;
 668 }
 669 
 670 /*
 671  * The idea is to set a period in which each task runs once.
 672  *
 673  * When there are too many tasks (sched_nr_latency) we have to stretch
 674  * this period because otherwise the slices get too small.
 675  *
 676  * p = (nr <= nl) ? l : l*nr/nl
 677  */
 678 static u64 __sched_period(unsigned long nr_running)
 679 {
 680         if (unlikely(nr_running > sched_nr_latency))
 681                 return nr_running * sysctl_sched_min_granularity;
 682         else
 683                 return sysctl_sched_latency;
 684 }
 685 
 686 /*
 687  * We calculate the wall-time slice from the period by taking a part
 688  * proportional to the weight.
 689  *
 690  * s = p*P[w/rw]
 691  */
 692 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
 693 {
 694         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
 695 
 696         for_each_sched_entity(se) {
 697                 struct load_weight *load;
 698                 struct load_weight lw;
 699 
 700                 cfs_rq = cfs_rq_of(se);
 701                 load = &cfs_rq->load;
 702 
 703                 if (unlikely(!se->on_rq)) {
 704                         lw = cfs_rq->load;
 705 
 706                         update_load_add(&lw, se->load.weight);
 707                         load = &lw;
 708                 }
 709                 slice = __calc_delta(slice, se->load.weight, load);
 710         }
 711         return slice;
 712 }
 713 
 714 /*
 715  * We calculate the vruntime slice of a to-be-inserted task.
 716  *
 717  * vs = s/w
 718  */
 719 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
 720 {
 721         return calc_delta_fair(sched_slice(cfs_rq, se), se);
 722 }
 723 
 724 #include "pelt.h"
 725 #ifdef CONFIG_SMP
 726 
 727 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
 728 static unsigned long task_h_load(struct task_struct *p);
 729 static unsigned long capacity_of(int cpu);
 730 
 731 /* Give new sched_entity start runnable values to heavy its load in infant time */
 732 void init_entity_runnable_average(struct sched_entity *se)
 733 {
 734         struct sched_avg *sa = &se->avg;
 735 
 736         memset(sa, 0, sizeof(*sa));
 737 
 738         /*
 739          * Tasks are initialized with full load to be seen as heavy tasks until
 740          * they get a chance to stabilize to their real load level.
 741          * Group entities are initialized with zero load to reflect the fact that
 742          * nothing has been attached to the task group yet.
 743          */
 744         if (entity_is_task(se))
 745                 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
 746 
 747         se->runnable_weight = se->load.weight;
 748 
 749         /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
 750 }
 751 
 752 static void attach_entity_cfs_rq(struct sched_entity *se);
 753 
 754 /*
 755  * With new tasks being created, their initial util_avgs are extrapolated
 756  * based on the cfs_rq's current util_avg:
 757  *
 758  *   util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
 759  *
 760  * However, in many cases, the above util_avg does not give a desired
 761  * value. Moreover, the sum of the util_avgs may be divergent, such
 762  * as when the series is a harmonic series.
 763  *
 764  * To solve this problem, we also cap the util_avg of successive tasks to
 765  * only 1/2 of the left utilization budget:
 766  *
 767  *   util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
 768  *
 769  * where n denotes the nth task and cpu_scale the CPU capacity.
 770  *
 771  * For example, for a CPU with 1024 of capacity, a simplest series from
 772  * the beginning would be like:
 773  *
 774  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
 775  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
 776  *
 777  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
 778  * if util_avg > util_avg_cap.
 779  */
 780 void post_init_entity_util_avg(struct task_struct *p)
 781 {
 782         struct sched_entity *se = &p->se;
 783         struct cfs_rq *cfs_rq = cfs_rq_of(se);
 784         struct sched_avg *sa = &se->avg;
 785         long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
 786         long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
 787 
 788         if (cap > 0) {
 789                 if (cfs_rq->avg.util_avg != 0) {
 790                         sa->util_avg  = cfs_rq->avg.util_avg * se->load.weight;
 791                         sa->util_avg /= (cfs_rq->avg.load_avg + 1);
 792 
 793                         if (sa->util_avg > cap)
 794                                 sa->util_avg = cap;
 795                 } else {
 796                         sa->util_avg = cap;
 797                 }
 798         }
 799 
 800         if (p->sched_class != &fair_sched_class) {
 801                 /*
 802                  * For !fair tasks do:
 803                  *
 804                 update_cfs_rq_load_avg(now, cfs_rq);
 805                 attach_entity_load_avg(cfs_rq, se, 0);
 806                 switched_from_fair(rq, p);
 807                  *
 808                  * such that the next switched_to_fair() has the
 809                  * expected state.
 810                  */
 811                 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
 812                 return;
 813         }
 814 
 815         attach_entity_cfs_rq(se);
 816 }
 817 
 818 #else /* !CONFIG_SMP */
 819 void init_entity_runnable_average(struct sched_entity *se)
 820 {
 821 }
 822 void post_init_entity_util_avg(struct task_struct *p)
 823 {
 824 }
 825 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
 826 {
 827 }
 828 #endif /* CONFIG_SMP */
 829 
 830 /*
 831  * Update the current task's runtime statistics.
 832  */
 833 static void update_curr(struct cfs_rq *cfs_rq)
 834 {
 835         struct sched_entity *curr = cfs_rq->curr;
 836         u64 now = rq_clock_task(rq_of(cfs_rq));
 837         u64 delta_exec;
 838 
 839         if (unlikely(!curr))
 840                 return;
 841 
 842         delta_exec = now - curr->exec_start;
 843         if (unlikely((s64)delta_exec <= 0))
 844                 return;
 845 
 846         curr->exec_start = now;
 847 
 848         schedstat_set(curr->statistics.exec_max,
 849                       max(delta_exec, curr->statistics.exec_max));
 850 
 851         curr->sum_exec_runtime += delta_exec;
 852         schedstat_add(cfs_rq->exec_clock, delta_exec);
 853 
 854         curr->vruntime += calc_delta_fair(delta_exec, curr);
 855         update_min_vruntime(cfs_rq);
 856 
 857         if (entity_is_task(curr)) {
 858                 struct task_struct *curtask = task_of(curr);
 859 
 860                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
 861                 cgroup_account_cputime(curtask, delta_exec);
 862                 account_group_exec_runtime(curtask, delta_exec);
 863         }
 864 
 865         account_cfs_rq_runtime(cfs_rq, delta_exec);
 866 }
 867 
 868 static void update_curr_fair(struct rq *rq)
 869 {
 870         update_curr(cfs_rq_of(&rq->curr->se));
 871 }
 872 
 873 static inline void
 874 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
 875 {
 876         u64 wait_start, prev_wait_start;
 877 
 878         if (!schedstat_enabled())
 879                 return;
 880 
 881         wait_start = rq_clock(rq_of(cfs_rq));
 882         prev_wait_start = schedstat_val(se->statistics.wait_start);
 883 
 884         if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
 885             likely(wait_start > prev_wait_start))
 886                 wait_start -= prev_wait_start;
 887 
 888         __schedstat_set(se->statistics.wait_start, wait_start);
 889 }
 890 
 891 static inline void
 892 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
 893 {
 894         struct task_struct *p;
 895         u64 delta;
 896 
 897         if (!schedstat_enabled())
 898                 return;
 899 
 900         delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
 901 
 902         if (entity_is_task(se)) {
 903                 p = task_of(se);
 904                 if (task_on_rq_migrating(p)) {
 905                         /*
 906                          * Preserve migrating task's wait time so wait_start
 907                          * time stamp can be adjusted to accumulate wait time
 908                          * prior to migration.
 909                          */
 910                         __schedstat_set(se->statistics.wait_start, delta);
 911                         return;
 912                 }
 913                 trace_sched_stat_wait(p, delta);
 914         }
 915 
 916         __schedstat_set(se->statistics.wait_max,
 917                       max(schedstat_val(se->statistics.wait_max), delta));
 918         __schedstat_inc(se->statistics.wait_count);
 919         __schedstat_add(se->statistics.wait_sum, delta);
 920         __schedstat_set(se->statistics.wait_start, 0);
 921 }
 922 
 923 static inline void
 924 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
 925 {
 926         struct task_struct *tsk = NULL;
 927         u64 sleep_start, block_start;
 928 
 929         if (!schedstat_enabled())
 930                 return;
 931 
 932         sleep_start = schedstat_val(se->statistics.sleep_start);
 933         block_start = schedstat_val(se->statistics.block_start);
 934 
 935         if (entity_is_task(se))
 936                 tsk = task_of(se);
 937 
 938         if (sleep_start) {
 939                 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
 940 
 941                 if ((s64)delta < 0)
 942                         delta = 0;
 943 
 944                 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
 945                         __schedstat_set(se->statistics.sleep_max, delta);
 946 
 947                 __schedstat_set(se->statistics.sleep_start, 0);
 948                 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
 949 
 950                 if (tsk) {
 951                         account_scheduler_latency(tsk, delta >> 10, 1);
 952                         trace_sched_stat_sleep(tsk, delta);
 953                 }
 954         }
 955         if (block_start) {
 956                 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
 957 
 958                 if ((s64)delta < 0)
 959                         delta = 0;
 960 
 961                 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
 962                         __schedstat_set(se->statistics.block_max, delta);
 963 
 964                 __schedstat_set(se->statistics.block_start, 0);
 965                 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
 966 
 967                 if (tsk) {
 968                         if (tsk->in_iowait) {
 969                                 __schedstat_add(se->statistics.iowait_sum, delta);
 970                                 __schedstat_inc(se->statistics.iowait_count);
 971                                 trace_sched_stat_iowait(tsk, delta);
 972                         }
 973 
 974                         trace_sched_stat_blocked(tsk, delta);
 975 
 976                         /*
 977                          * Blocking time is in units of nanosecs, so shift by
 978                          * 20 to get a milliseconds-range estimation of the
 979                          * amount of time that the task spent sleeping:
 980                          */
 981                         if (unlikely(prof_on == SLEEP_PROFILING)) {
 982                                 profile_hits(SLEEP_PROFILING,
 983                                                 (void *)get_wchan(tsk),
 984                                                 delta >> 20);
 985                         }
 986                         account_scheduler_latency(tsk, delta >> 10, 0);
 987                 }
 988         }
 989 }
 990 
 991 /*
 992  * Task is being enqueued - update stats:
 993  */
 994 static inline void
 995 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
 996 {
 997         if (!schedstat_enabled())
 998                 return;
 999 
1000         /*
1001          * Are we enqueueing a waiting task? (for current tasks
1002          * a dequeue/enqueue event is a NOP)
1003          */
1004         if (se != cfs_rq->curr)
1005                 update_stats_wait_start(cfs_rq, se);
1006 
1007         if (flags & ENQUEUE_WAKEUP)
1008                 update_stats_enqueue_sleeper(cfs_rq, se);
1009 }
1010 
1011 static inline void
1012 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1013 {
1014 
1015         if (!schedstat_enabled())
1016                 return;
1017 
1018         /*
1019          * Mark the end of the wait period if dequeueing a
1020          * waiting task:
1021          */
1022         if (se != cfs_rq->curr)
1023                 update_stats_wait_end(cfs_rq, se);
1024 
1025         if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1026                 struct task_struct *tsk = task_of(se);
1027 
1028                 if (tsk->state & TASK_INTERRUPTIBLE)
1029                         __schedstat_set(se->statistics.sleep_start,
1030                                       rq_clock(rq_of(cfs_rq)));
1031                 if (tsk->state & TASK_UNINTERRUPTIBLE)
1032                         __schedstat_set(se->statistics.block_start,
1033                                       rq_clock(rq_of(cfs_rq)));
1034         }
1035 }
1036 
1037 /*
1038  * We are picking a new current task - update its stats:
1039  */
1040 static inline void
1041 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1042 {
1043         /*
1044          * We are starting a new run period:
1045          */
1046         se->exec_start = rq_clock_task(rq_of(cfs_rq));
1047 }
1048 
1049 /**************************************************
1050  * Scheduling class queueing methods:
1051  */
1052 
1053 #ifdef CONFIG_NUMA_BALANCING
1054 /*
1055  * Approximate time to scan a full NUMA task in ms. The task scan period is
1056  * calculated based on the tasks virtual memory size and
1057  * numa_balancing_scan_size.
1058  */
1059 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1060 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1061 
1062 /* Portion of address space to scan in MB */
1063 unsigned int sysctl_numa_balancing_scan_size = 256;
1064 
1065 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1066 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1067 
1068 struct numa_group {
1069         refcount_t refcount;
1070 
1071         spinlock_t lock; /* nr_tasks, tasks */
1072         int nr_tasks;
1073         pid_t gid;
1074         int active_nodes;
1075 
1076         struct rcu_head rcu;
1077         unsigned long total_faults;
1078         unsigned long max_faults_cpu;
1079         /*
1080          * Faults_cpu is used to decide whether memory should move
1081          * towards the CPU. As a consequence, these stats are weighted
1082          * more by CPU use than by memory faults.
1083          */
1084         unsigned long *faults_cpu;
1085         unsigned long faults[0];
1086 };
1087 
1088 /*
1089  * For functions that can be called in multiple contexts that permit reading
1090  * ->numa_group (see struct task_struct for locking rules).
1091  */
1092 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1093 {
1094         return rcu_dereference_check(p->numa_group, p == current ||
1095                 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1096 }
1097 
1098 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1099 {
1100         return rcu_dereference_protected(p->numa_group, p == current);
1101 }
1102 
1103 static inline unsigned long group_faults_priv(struct numa_group *ng);
1104 static inline unsigned long group_faults_shared(struct numa_group *ng);
1105 
1106 static unsigned int task_nr_scan_windows(struct task_struct *p)
1107 {
1108         unsigned long rss = 0;
1109         unsigned long nr_scan_pages;
1110 
1111         /*
1112          * Calculations based on RSS as non-present and empty pages are skipped
1113          * by the PTE scanner and NUMA hinting faults should be trapped based
1114          * on resident pages
1115          */
1116         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1117         rss = get_mm_rss(p->mm);
1118         if (!rss)
1119                 rss = nr_scan_pages;
1120 
1121         rss = round_up(rss, nr_scan_pages);
1122         return rss / nr_scan_pages;
1123 }
1124 
1125 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1126 #define MAX_SCAN_WINDOW 2560
1127 
1128 static unsigned int task_scan_min(struct task_struct *p)
1129 {
1130         unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1131         unsigned int scan, floor;
1132         unsigned int windows = 1;
1133 
1134         if (scan_size < MAX_SCAN_WINDOW)
1135                 windows = MAX_SCAN_WINDOW / scan_size;
1136         floor = 1000 / windows;
1137 
1138         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1139         return max_t(unsigned int, floor, scan);
1140 }
1141 
1142 static unsigned int task_scan_start(struct task_struct *p)
1143 {
1144         unsigned long smin = task_scan_min(p);
1145         unsigned long period = smin;
1146         struct numa_group *ng;
1147 
1148         /* Scale the maximum scan period with the amount of shared memory. */
1149         rcu_read_lock();
1150         ng = rcu_dereference(p->numa_group);
1151         if (ng) {
1152                 unsigned long shared = group_faults_shared(ng);
1153                 unsigned long private = group_faults_priv(ng);
1154 
1155                 period *= refcount_read(&ng->refcount);
1156                 period *= shared + 1;
1157                 period /= private + shared + 1;
1158         }
1159         rcu_read_unlock();
1160 
1161         return max(smin, period);
1162 }
1163 
1164 static unsigned int task_scan_max(struct task_struct *p)
1165 {
1166         unsigned long smin = task_scan_min(p);
1167         unsigned long smax;
1168         struct numa_group *ng;
1169 
1170         /* Watch for min being lower than max due to floor calculations */
1171         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1172 
1173         /* Scale the maximum scan period with the amount of shared memory. */
1174         ng = deref_curr_numa_group(p);
1175         if (ng) {
1176                 unsigned long shared = group_faults_shared(ng);
1177                 unsigned long private = group_faults_priv(ng);
1178                 unsigned long period = smax;
1179 
1180                 period *= refcount_read(&ng->refcount);
1181                 period *= shared + 1;
1182                 period /= private + shared + 1;
1183 
1184                 smax = max(smax, period);
1185         }
1186 
1187         return max(smin, smax);
1188 }
1189 
1190 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1191 {
1192         rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1193         rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1194 }
1195 
1196 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1197 {
1198         rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1199         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1200 }
1201 
1202 /* Shared or private faults. */
1203 #define NR_NUMA_HINT_FAULT_TYPES 2
1204 
1205 /* Memory and CPU locality */
1206 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1207 
1208 /* Averaged statistics, and temporary buffers. */
1209 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1210 
1211 pid_t task_numa_group_id(struct task_struct *p)
1212 {
1213         struct numa_group *ng;
1214         pid_t gid = 0;
1215 
1216         rcu_read_lock();
1217         ng = rcu_dereference(p->numa_group);
1218         if (ng)
1219                 gid = ng->gid;
1220         rcu_read_unlock();
1221 
1222         return gid;
1223 }
1224 
1225 /*
1226  * The averaged statistics, shared & private, memory & CPU,
1227  * occupy the first half of the array. The second half of the
1228  * array is for current counters, which are averaged into the
1229  * first set by task_numa_placement.
1230  */
1231 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1232 {
1233         return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1234 }
1235 
1236 static inline unsigned long task_faults(struct task_struct *p, int nid)
1237 {
1238         if (!p->numa_faults)
1239                 return 0;
1240 
1241         return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1242                 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1243 }
1244 
1245 static inline unsigned long group_faults(struct task_struct *p, int nid)
1246 {
1247         struct numa_group *ng = deref_task_numa_group(p);
1248 
1249         if (!ng)
1250                 return 0;
1251 
1252         return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1253                 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1254 }
1255 
1256 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1257 {
1258         return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1259                 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1260 }
1261 
1262 static inline unsigned long group_faults_priv(struct numa_group *ng)
1263 {
1264         unsigned long faults = 0;
1265         int node;
1266 
1267         for_each_online_node(node) {
1268                 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1269         }
1270 
1271         return faults;
1272 }
1273 
1274 static inline unsigned long group_faults_shared(struct numa_group *ng)
1275 {
1276         unsigned long faults = 0;
1277         int node;
1278 
1279         for_each_online_node(node) {
1280                 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1281         }
1282 
1283         return faults;
1284 }
1285 
1286 /*
1287  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1288  * considered part of a numa group's pseudo-interleaving set. Migrations
1289  * between these nodes are slowed down, to allow things to settle down.
1290  */
1291 #define ACTIVE_NODE_FRACTION 3
1292 
1293 static bool numa_is_active_node(int nid, struct numa_group *ng)
1294 {
1295         return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1296 }
1297 
1298 /* Handle placement on systems where not all nodes are directly connected. */
1299 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1300                                         int maxdist, bool task)
1301 {
1302         unsigned long score = 0;
1303         int node;
1304 
1305         /*
1306          * All nodes are directly connected, and the same distance
1307          * from each other. No need for fancy placement algorithms.
1308          */
1309         if (sched_numa_topology_type == NUMA_DIRECT)
1310                 return 0;
1311 
1312         /*
1313          * This code is called for each node, introducing N^2 complexity,
1314          * which should be ok given the number of nodes rarely exceeds 8.
1315          */
1316         for_each_online_node(node) {
1317                 unsigned long faults;
1318                 int dist = node_distance(nid, node);
1319 
1320                 /*
1321                  * The furthest away nodes in the system are not interesting
1322                  * for placement; nid was already counted.
1323                  */
1324                 if (dist == sched_max_numa_distance || node == nid)
1325                         continue;
1326 
1327                 /*
1328                  * On systems with a backplane NUMA topology, compare groups
1329                  * of nodes, and move tasks towards the group with the most
1330                  * memory accesses. When comparing two nodes at distance
1331                  * "hoplimit", only nodes closer by than "hoplimit" are part
1332                  * of each group. Skip other nodes.
1333                  */
1334                 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1335                                         dist >= maxdist)
1336                         continue;
1337 
1338                 /* Add up the faults from nearby nodes. */
1339                 if (task)
1340                         faults = task_faults(p, node);
1341                 else
1342                         faults = group_faults(p, node);
1343 
1344                 /*
1345                  * On systems with a glueless mesh NUMA topology, there are
1346                  * no fixed "groups of nodes". Instead, nodes that are not
1347                  * directly connected bounce traffic through intermediate
1348                  * nodes; a numa_group can occupy any set of nodes.
1349                  * The further away a node is, the less the faults count.
1350                  * This seems to result in good task placement.
1351                  */
1352                 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1353                         faults *= (sched_max_numa_distance - dist);
1354                         faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1355                 }
1356 
1357                 score += faults;
1358         }
1359 
1360         return score;
1361 }
1362 
1363 /*
1364  * These return the fraction of accesses done by a particular task, or
1365  * task group, on a particular numa node.  The group weight is given a
1366  * larger multiplier, in order to group tasks together that are almost
1367  * evenly spread out between numa nodes.
1368  */
1369 static inline unsigned long task_weight(struct task_struct *p, int nid,
1370                                         int dist)
1371 {
1372         unsigned long faults, total_faults;
1373 
1374         if (!p->numa_faults)
1375                 return 0;
1376 
1377         total_faults = p->total_numa_faults;
1378 
1379         if (!total_faults)
1380                 return 0;
1381 
1382         faults = task_faults(p, nid);
1383         faults += score_nearby_nodes(p, nid, dist, true);
1384 
1385         return 1000 * faults / total_faults;
1386 }
1387 
1388 static inline unsigned long group_weight(struct task_struct *p, int nid,
1389                                          int dist)
1390 {
1391         struct numa_group *ng = deref_task_numa_group(p);
1392         unsigned long faults, total_faults;
1393 
1394         if (!ng)
1395                 return 0;
1396 
1397         total_faults = ng->total_faults;
1398 
1399         if (!total_faults)
1400                 return 0;
1401 
1402         faults = group_faults(p, nid);
1403         faults += score_nearby_nodes(p, nid, dist, false);
1404 
1405         return 1000 * faults / total_faults;
1406 }
1407 
1408 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1409                                 int src_nid, int dst_cpu)
1410 {
1411         struct numa_group *ng = deref_curr_numa_group(p);
1412         int dst_nid = cpu_to_node(dst_cpu);
1413         int last_cpupid, this_cpupid;
1414 
1415         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1416         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1417 
1418         /*
1419          * Allow first faults or private faults to migrate immediately early in
1420          * the lifetime of a task. The magic number 4 is based on waiting for
1421          * two full passes of the "multi-stage node selection" test that is
1422          * executed below.
1423          */
1424         if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1425             (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1426                 return true;
1427 
1428         /*
1429          * Multi-stage node selection is used in conjunction with a periodic
1430          * migration fault to build a temporal task<->page relation. By using
1431          * a two-stage filter we remove short/unlikely relations.
1432          *
1433          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1434          * a task's usage of a particular page (n_p) per total usage of this
1435          * page (n_t) (in a given time-span) to a probability.
1436          *
1437          * Our periodic faults will sample this probability and getting the
1438          * same result twice in a row, given these samples are fully
1439          * independent, is then given by P(n)^2, provided our sample period
1440          * is sufficiently short compared to the usage pattern.
1441          *
1442          * This quadric squishes small probabilities, making it less likely we
1443          * act on an unlikely task<->page relation.
1444          */
1445         if (!cpupid_pid_unset(last_cpupid) &&
1446                                 cpupid_to_nid(last_cpupid) != dst_nid)
1447                 return false;
1448 
1449         /* Always allow migrate on private faults */
1450         if (cpupid_match_pid(p, last_cpupid))
1451                 return true;
1452 
1453         /* A shared fault, but p->numa_group has not been set up yet. */
1454         if (!ng)
1455                 return true;
1456 
1457         /*
1458          * Destination node is much more heavily used than the source
1459          * node? Allow migration.
1460          */
1461         if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1462                                         ACTIVE_NODE_FRACTION)
1463                 return true;
1464 
1465         /*
1466          * Distribute memory according to CPU & memory use on each node,
1467          * with 3/4 hysteresis to avoid unnecessary memory migrations:
1468          *
1469          * faults_cpu(dst)   3   faults_cpu(src)
1470          * --------------- * - > ---------------
1471          * faults_mem(dst)   4   faults_mem(src)
1472          */
1473         return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1474                group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1475 }
1476 
1477 static unsigned long cpu_runnable_load(struct rq *rq);
1478 
1479 /* Cached statistics for all CPUs within a node */
1480 struct numa_stats {
1481         unsigned long load;
1482 
1483         /* Total compute capacity of CPUs on a node */
1484         unsigned long compute_capacity;
1485 };
1486 
1487 /*
1488  * XXX borrowed from update_sg_lb_stats
1489  */
1490 static void update_numa_stats(struct numa_stats *ns, int nid)
1491 {
1492         int cpu;
1493 
1494         memset(ns, 0, sizeof(*ns));
1495         for_each_cpu(cpu, cpumask_of_node(nid)) {
1496                 struct rq *rq = cpu_rq(cpu);
1497 
1498                 ns->load += cpu_runnable_load(rq);
1499                 ns->compute_capacity += capacity_of(cpu);
1500         }
1501 
1502 }
1503 
1504 struct task_numa_env {
1505         struct task_struct *p;
1506 
1507         int src_cpu, src_nid;
1508         int dst_cpu, dst_nid;
1509 
1510         struct numa_stats src_stats, dst_stats;
1511 
1512         int imbalance_pct;
1513         int dist;
1514 
1515         struct task_struct *best_task;
1516         long best_imp;
1517         int best_cpu;
1518 };
1519 
1520 static void task_numa_assign(struct task_numa_env *env,
1521                              struct task_struct *p, long imp)
1522 {
1523         struct rq *rq = cpu_rq(env->dst_cpu);
1524 
1525         /* Bail out if run-queue part of active NUMA balance. */
1526         if (xchg(&rq->numa_migrate_on, 1))
1527                 return;
1528 
1529         /*
1530          * Clear previous best_cpu/rq numa-migrate flag, since task now
1531          * found a better CPU to move/swap.
1532          */
1533         if (env->best_cpu != -1) {
1534                 rq = cpu_rq(env->best_cpu);
1535                 WRITE_ONCE(rq->numa_migrate_on, 0);
1536         }
1537 
1538         if (env->best_task)
1539                 put_task_struct(env->best_task);
1540         if (p)
1541                 get_task_struct(p);
1542 
1543         env->best_task = p;
1544         env->best_imp = imp;
1545         env->best_cpu = env->dst_cpu;
1546 }
1547 
1548 static bool load_too_imbalanced(long src_load, long dst_load,
1549                                 struct task_numa_env *env)
1550 {
1551         long imb, old_imb;
1552         long orig_src_load, orig_dst_load;
1553         long src_capacity, dst_capacity;
1554 
1555         /*
1556          * The load is corrected for the CPU capacity available on each node.
1557          *
1558          * src_load        dst_load
1559          * ------------ vs ---------
1560          * src_capacity    dst_capacity
1561          */
1562         src_capacity = env->src_stats.compute_capacity;
1563         dst_capacity = env->dst_stats.compute_capacity;
1564 
1565         imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1566 
1567         orig_src_load = env->src_stats.load;
1568         orig_dst_load = env->dst_stats.load;
1569 
1570         old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1571 
1572         /* Would this change make things worse? */
1573         return (imb > old_imb);
1574 }
1575 
1576 /*
1577  * Maximum NUMA importance can be 1998 (2*999);
1578  * SMALLIMP @ 30 would be close to 1998/64.
1579  * Used to deter task migration.
1580  */
1581 #define SMALLIMP        30
1582 
1583 /*
1584  * This checks if the overall compute and NUMA accesses of the system would
1585  * be improved if the source tasks was migrated to the target dst_cpu taking
1586  * into account that it might be best if task running on the dst_cpu should
1587  * be exchanged with the source task
1588  */
1589 static void task_numa_compare(struct task_numa_env *env,
1590                               long taskimp, long groupimp, bool maymove)
1591 {
1592         struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1593         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1594         long imp = p_ng ? groupimp : taskimp;
1595         struct task_struct *cur;
1596         long src_load, dst_load;
1597         int dist = env->dist;
1598         long moveimp = imp;
1599         long load;
1600 
1601         if (READ_ONCE(dst_rq->numa_migrate_on))
1602                 return;
1603 
1604         rcu_read_lock();
1605         cur = rcu_dereference(dst_rq->curr);
1606         if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1607                 cur = NULL;
1608 
1609         /*
1610          * Because we have preemption enabled we can get migrated around and
1611          * end try selecting ourselves (current == env->p) as a swap candidate.
1612          */
1613         if (cur == env->p)
1614                 goto unlock;
1615 
1616         if (!cur) {
1617                 if (maymove && moveimp >= env->best_imp)
1618                         goto assign;
1619                 else
1620                         goto unlock;
1621         }
1622 
1623         /*
1624          * "imp" is the fault differential for the source task between the
1625          * source and destination node. Calculate the total differential for
1626          * the source task and potential destination task. The more negative
1627          * the value is, the more remote accesses that would be expected to
1628          * be incurred if the tasks were swapped.
1629          */
1630         /* Skip this swap candidate if cannot move to the source cpu */
1631         if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1632                 goto unlock;
1633 
1634         /*
1635          * If dst and source tasks are in the same NUMA group, or not
1636          * in any group then look only at task weights.
1637          */
1638         cur_ng = rcu_dereference(cur->numa_group);
1639         if (cur_ng == p_ng) {
1640                 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1641                       task_weight(cur, env->dst_nid, dist);
1642                 /*
1643                  * Add some hysteresis to prevent swapping the
1644                  * tasks within a group over tiny differences.
1645                  */
1646                 if (cur_ng)
1647                         imp -= imp / 16;
1648         } else {
1649                 /*
1650                  * Compare the group weights. If a task is all by itself
1651                  * (not part of a group), use the task weight instead.
1652                  */
1653                 if (cur_ng && p_ng)
1654                         imp += group_weight(cur, env->src_nid, dist) -
1655                                group_weight(cur, env->dst_nid, dist);
1656                 else
1657                         imp += task_weight(cur, env->src_nid, dist) -
1658                                task_weight(cur, env->dst_nid, dist);
1659         }
1660 
1661         if (maymove && moveimp > imp && moveimp > env->best_imp) {
1662                 imp = moveimp;
1663                 cur = NULL;
1664                 goto assign;
1665         }
1666 
1667         /*
1668          * If the NUMA importance is less than SMALLIMP,
1669          * task migration might only result in ping pong
1670          * of tasks and also hurt performance due to cache
1671          * misses.
1672          */
1673         if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1674                 goto unlock;
1675 
1676         /*
1677          * In the overloaded case, try and keep the load balanced.
1678          */
1679         load = task_h_load(env->p) - task_h_load(cur);
1680         if (!load)
1681                 goto assign;
1682 
1683         dst_load = env->dst_stats.load + load;
1684         src_load = env->src_stats.load - load;
1685 
1686         if (load_too_imbalanced(src_load, dst_load, env))
1687                 goto unlock;
1688 
1689 assign:
1690         /*
1691          * One idle CPU per node is evaluated for a task numa move.
1692          * Call select_idle_sibling to maybe find a better one.
1693          */
1694         if (!cur) {
1695                 /*
1696                  * select_idle_siblings() uses an per-CPU cpumask that
1697                  * can be used from IRQ context.
1698                  */
1699                 local_irq_disable();
1700                 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1701                                                    env->dst_cpu);
1702                 local_irq_enable();
1703         }
1704 
1705         task_numa_assign(env, cur, imp);
1706 unlock:
1707         rcu_read_unlock();
1708 }
1709 
1710 static void task_numa_find_cpu(struct task_numa_env *env,
1711                                 long taskimp, long groupimp)
1712 {
1713         long src_load, dst_load, load;
1714         bool maymove = false;
1715         int cpu;
1716 
1717         load = task_h_load(env->p);
1718         dst_load = env->dst_stats.load + load;
1719         src_load = env->src_stats.load - load;
1720 
1721         /*
1722          * If the improvement from just moving env->p direction is better
1723          * than swapping tasks around, check if a move is possible.
1724          */
1725         maymove = !load_too_imbalanced(src_load, dst_load, env);
1726 
1727         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1728                 /* Skip this CPU if the source task cannot migrate */
1729                 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1730                         continue;
1731 
1732                 env->dst_cpu = cpu;
1733                 task_numa_compare(env, taskimp, groupimp, maymove);
1734         }
1735 }
1736 
1737 static int task_numa_migrate(struct task_struct *p)
1738 {
1739         struct task_numa_env env = {
1740                 .p = p,
1741 
1742                 .src_cpu = task_cpu(p),
1743                 .src_nid = task_node(p),
1744 
1745                 .imbalance_pct = 112,
1746 
1747                 .best_task = NULL,
1748                 .best_imp = 0,
1749                 .best_cpu = -1,
1750         };
1751         unsigned long taskweight, groupweight;
1752         struct sched_domain *sd;
1753         long taskimp, groupimp;
1754         struct numa_group *ng;
1755         struct rq *best_rq;
1756         int nid, ret, dist;
1757 
1758         /*
1759          * Pick the lowest SD_NUMA domain, as that would have the smallest
1760          * imbalance and would be the first to start moving tasks about.
1761          *
1762          * And we want to avoid any moving of tasks about, as that would create
1763          * random movement of tasks -- counter the numa conditions we're trying
1764          * to satisfy here.
1765          */
1766         rcu_read_lock();
1767         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1768         if (sd)
1769                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1770         rcu_read_unlock();
1771 
1772         /*
1773          * Cpusets can break the scheduler domain tree into smaller
1774          * balance domains, some of which do not cross NUMA boundaries.
1775          * Tasks that are "trapped" in such domains cannot be migrated
1776          * elsewhere, so there is no point in (re)trying.
1777          */
1778         if (unlikely(!sd)) {
1779                 sched_setnuma(p, task_node(p));
1780                 return -EINVAL;
1781         }
1782 
1783         env.dst_nid = p->numa_preferred_nid;
1784         dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1785         taskweight = task_weight(p, env.src_nid, dist);
1786         groupweight = group_weight(p, env.src_nid, dist);
1787         update_numa_stats(&env.src_stats, env.src_nid);
1788         taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1789         groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1790         update_numa_stats(&env.dst_stats, env.dst_nid);
1791 
1792         /* Try to find a spot on the preferred nid. */
1793         task_numa_find_cpu(&env, taskimp, groupimp);
1794 
1795         /*
1796          * Look at other nodes in these cases:
1797          * - there is no space available on the preferred_nid
1798          * - the task is part of a numa_group that is interleaved across
1799          *   multiple NUMA nodes; in order to better consolidate the group,
1800          *   we need to check other locations.
1801          */
1802         ng = deref_curr_numa_group(p);
1803         if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1804                 for_each_online_node(nid) {
1805                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1806                                 continue;
1807 
1808                         dist = node_distance(env.src_nid, env.dst_nid);
1809                         if (sched_numa_topology_type == NUMA_BACKPLANE &&
1810                                                 dist != env.dist) {
1811                                 taskweight = task_weight(p, env.src_nid, dist);
1812                                 groupweight = group_weight(p, env.src_nid, dist);
1813                         }
1814 
1815                         /* Only consider nodes where both task and groups benefit */
1816                         taskimp = task_weight(p, nid, dist) - taskweight;
1817                         groupimp = group_weight(p, nid, dist) - groupweight;
1818                         if (taskimp < 0 && groupimp < 0)
1819                                 continue;
1820 
1821                         env.dist = dist;
1822                         env.dst_nid = nid;
1823                         update_numa_stats(&env.dst_stats, env.dst_nid);
1824                         task_numa_find_cpu(&env, taskimp, groupimp);
1825                 }
1826         }
1827 
1828         /*
1829          * If the task is part of a workload that spans multiple NUMA nodes,
1830          * and is migrating into one of the workload's active nodes, remember
1831          * this node as the task's preferred numa node, so the workload can
1832          * settle down.
1833          * A task that migrated to a second choice node will be better off
1834          * trying for a better one later. Do not set the preferred node here.
1835          */
1836         if (ng) {
1837                 if (env.best_cpu == -1)
1838                         nid = env.src_nid;
1839                 else
1840                         nid = cpu_to_node(env.best_cpu);
1841 
1842                 if (nid != p->numa_preferred_nid)
1843                         sched_setnuma(p, nid);
1844         }
1845 
1846         /* No better CPU than the current one was found. */
1847         if (env.best_cpu == -1)
1848                 return -EAGAIN;
1849 
1850         best_rq = cpu_rq(env.best_cpu);
1851         if (env.best_task == NULL) {
1852                 ret = migrate_task_to(p, env.best_cpu);
1853                 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1854                 if (ret != 0)
1855                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1856                 return ret;
1857         }
1858 
1859         ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1860         WRITE_ONCE(best_rq->numa_migrate_on, 0);
1861 
1862         if (ret != 0)
1863                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1864         put_task_struct(env.best_task);
1865         return ret;
1866 }
1867 
1868 /* Attempt to migrate a task to a CPU on the preferred node. */
1869 static void numa_migrate_preferred(struct task_struct *p)
1870 {
1871         unsigned long interval = HZ;
1872 
1873         /* This task has no NUMA fault statistics yet */
1874         if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1875                 return;
1876 
1877         /* Periodically retry migrating the task to the preferred node */
1878         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1879         p->numa_migrate_retry = jiffies + interval;
1880 
1881         /* Success if task is already running on preferred CPU */
1882         if (task_node(p) == p->numa_preferred_nid)
1883                 return;
1884 
1885         /* Otherwise, try migrate to a CPU on the preferred node */
1886         task_numa_migrate(p);
1887 }
1888 
1889 /*
1890  * Find out how many nodes on the workload is actively running on. Do this by
1891  * tracking the nodes from which NUMA hinting faults are triggered. This can
1892  * be different from the set of nodes where the workload's memory is currently
1893  * located.
1894  */
1895 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1896 {
1897         unsigned long faults, max_faults = 0;
1898         int nid, active_nodes = 0;
1899 
1900         for_each_online_node(nid) {
1901                 faults = group_faults_cpu(numa_group, nid);
1902                 if (faults > max_faults)
1903                         max_faults = faults;
1904         }
1905 
1906         for_each_online_node(nid) {
1907                 faults = group_faults_cpu(numa_group, nid);
1908                 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1909                         active_nodes++;
1910         }
1911 
1912         numa_group->max_faults_cpu = max_faults;
1913         numa_group->active_nodes = active_nodes;
1914 }
1915 
1916 /*
1917  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1918  * increments. The more local the fault statistics are, the higher the scan
1919  * period will be for the next scan window. If local/(local+remote) ratio is
1920  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1921  * the scan period will decrease. Aim for 70% local accesses.
1922  */
1923 #define NUMA_PERIOD_SLOTS 10
1924 #define NUMA_PERIOD_THRESHOLD 7
1925 
1926 /*
1927  * Increase the scan period (slow down scanning) if the majority of
1928  * our memory is already on our local node, or if the majority of
1929  * the page accesses are shared with other processes.
1930  * Otherwise, decrease the scan period.
1931  */
1932 static void update_task_scan_period(struct task_struct *p,
1933                         unsigned long shared, unsigned long private)
1934 {
1935         unsigned int period_slot;
1936         int lr_ratio, ps_ratio;
1937         int diff;
1938 
1939         unsigned long remote = p->numa_faults_locality[0];
1940         unsigned long local = p->numa_faults_locality[1];
1941 
1942         /*
1943          * If there were no record hinting faults then either the task is
1944          * completely idle or all activity is areas that are not of interest
1945          * to automatic numa balancing. Related to that, if there were failed
1946          * migration then it implies we are migrating too quickly or the local
1947          * node is overloaded. In either case, scan slower
1948          */
1949         if (local + shared == 0 || p->numa_faults_locality[2]) {
1950                 p->numa_scan_period = min(p->numa_scan_period_max,
1951                         p->numa_scan_period << 1);
1952 
1953                 p->mm->numa_next_scan = jiffies +
1954                         msecs_to_jiffies(p->numa_scan_period);
1955 
1956                 return;
1957         }
1958 
1959         /*
1960          * Prepare to scale scan period relative to the current period.
1961          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
1962          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1963          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1964          */
1965         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1966         lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1967         ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1968 
1969         if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1970                 /*
1971                  * Most memory accesses are local. There is no need to
1972                  * do fast NUMA scanning, since memory is already local.
1973                  */
1974                 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1975                 if (!slot)
1976                         slot = 1;
1977                 diff = slot * period_slot;
1978         } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1979                 /*
1980                  * Most memory accesses are shared with other tasks.
1981                  * There is no point in continuing fast NUMA scanning,
1982                  * since other tasks may just move the memory elsewhere.
1983                  */
1984                 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1985                 if (!slot)
1986                         slot = 1;
1987                 diff = slot * period_slot;
1988         } else {
1989                 /*
1990                  * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1991                  * yet they are not on the local NUMA node. Speed up
1992                  * NUMA scanning to get the memory moved over.
1993                  */
1994                 int ratio = max(lr_ratio, ps_ratio);
1995                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1996         }
1997 
1998         p->numa_scan_period = clamp(p->numa_scan_period + diff,
1999                         task_scan_min(p), task_scan_max(p));
2000         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2001 }
2002 
2003 /*
2004  * Get the fraction of time the task has been running since the last
2005  * NUMA placement cycle. The scheduler keeps similar statistics, but
2006  * decays those on a 32ms period, which is orders of magnitude off
2007  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2008  * stats only if the task is so new there are no NUMA statistics yet.
2009  */
2010 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2011 {
2012         u64 runtime, delta, now;
2013         /* Use the start of this time slice to avoid calculations. */
2014         now = p->se.exec_start;
2015         runtime = p->se.sum_exec_runtime;
2016 
2017         if (p->last_task_numa_placement) {
2018                 delta = runtime - p->last_sum_exec_runtime;
2019                 *period = now - p->last_task_numa_placement;
2020 
2021                 /* Avoid time going backwards, prevent potential divide error: */
2022                 if (unlikely((s64)*period < 0))
2023                         *period = 0;
2024         } else {
2025                 delta = p->se.avg.load_sum;
2026                 *period = LOAD_AVG_MAX;
2027         }
2028 
2029         p->last_sum_exec_runtime = runtime;
2030         p->last_task_numa_placement = now;
2031 
2032         return delta;
2033 }
2034 
2035 /*
2036  * Determine the preferred nid for a task in a numa_group. This needs to
2037  * be done in a way that produces consistent results with group_weight,
2038  * otherwise workloads might not converge.
2039  */
2040 static int preferred_group_nid(struct task_struct *p, int nid)
2041 {
2042         nodemask_t nodes;
2043         int dist;
2044 
2045         /* Direct connections between all NUMA nodes. */
2046         if (sched_numa_topology_type == NUMA_DIRECT)
2047                 return nid;
2048 
2049         /*
2050          * On a system with glueless mesh NUMA topology, group_weight
2051          * scores nodes according to the number of NUMA hinting faults on
2052          * both the node itself, and on nearby nodes.
2053          */
2054         if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2055                 unsigned long score, max_score = 0;
2056                 int node, max_node = nid;
2057 
2058                 dist = sched_max_numa_distance;
2059 
2060                 for_each_online_node(node) {
2061                         score = group_weight(p, node, dist);
2062                         if (score > max_score) {
2063                                 max_score = score;
2064                                 max_node = node;
2065                         }
2066                 }
2067                 return max_node;
2068         }
2069 
2070         /*
2071          * Finding the preferred nid in a system with NUMA backplane
2072          * interconnect topology is more involved. The goal is to locate
2073          * tasks from numa_groups near each other in the system, and
2074          * untangle workloads from different sides of the system. This requires
2075          * searching down the hierarchy of node groups, recursively searching
2076          * inside the highest scoring group of nodes. The nodemask tricks
2077          * keep the complexity of the search down.
2078          */
2079         nodes = node_online_map;
2080         for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2081                 unsigned long max_faults = 0;
2082                 nodemask_t max_group = NODE_MASK_NONE;
2083                 int a, b;
2084 
2085                 /* Are there nodes at this distance from each other? */
2086                 if (!find_numa_distance(dist))
2087                         continue;
2088 
2089                 for_each_node_mask(a, nodes) {
2090                         unsigned long faults = 0;
2091                         nodemask_t this_group;
2092                         nodes_clear(this_group);
2093 
2094                         /* Sum group's NUMA faults; includes a==b case. */
2095                         for_each_node_mask(b, nodes) {
2096                                 if (node_distance(a, b) < dist) {
2097                                         faults += group_faults(p, b);
2098                                         node_set(b, this_group);
2099                                         node_clear(b, nodes);
2100                                 }
2101                         }
2102 
2103                         /* Remember the top group. */
2104                         if (faults > max_faults) {
2105                                 max_faults = faults;
2106                                 max_group = this_group;
2107                                 /*
2108                                  * subtle: at the smallest distance there is
2109                                  * just one node left in each "group", the
2110                                  * winner is the preferred nid.
2111                                  */
2112                                 nid = a;
2113                         }
2114                 }
2115                 /* Next round, evaluate the nodes within max_group. */
2116                 if (!max_faults)
2117                         break;
2118                 nodes = max_group;
2119         }
2120         return nid;
2121 }
2122 
2123 static void task_numa_placement(struct task_struct *p)
2124 {
2125         int seq, nid, max_nid = NUMA_NO_NODE;
2126         unsigned long max_faults = 0;
2127         unsigned long fault_types[2] = { 0, 0 };
2128         unsigned long total_faults;
2129         u64 runtime, period;
2130         spinlock_t *group_lock = NULL;
2131         struct numa_group *ng;
2132 
2133         /*
2134          * The p->mm->numa_scan_seq field gets updated without
2135          * exclusive access. Use READ_ONCE() here to ensure
2136          * that the field is read in a single access:
2137          */
2138         seq = READ_ONCE(p->mm->numa_scan_seq);
2139         if (p->numa_scan_seq == seq)
2140                 return;
2141         p->numa_scan_seq = seq;
2142         p->numa_scan_period_max = task_scan_max(p);
2143 
2144         total_faults = p->numa_faults_locality[0] +
2145                        p->numa_faults_locality[1];
2146         runtime = numa_get_avg_runtime(p, &period);
2147 
2148         /* If the task is part of a group prevent parallel updates to group stats */
2149         ng = deref_curr_numa_group(p);
2150         if (ng) {
2151                 group_lock = &ng->lock;
2152                 spin_lock_irq(group_lock);
2153         }
2154 
2155         /* Find the node with the highest number of faults */
2156         for_each_online_node(nid) {
2157                 /* Keep track of the offsets in numa_faults array */
2158                 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2159                 unsigned long faults = 0, group_faults = 0;
2160                 int priv;
2161 
2162                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2163                         long diff, f_diff, f_weight;
2164 
2165                         mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2166                         membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2167                         cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2168                         cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2169 
2170                         /* Decay existing window, copy faults since last scan */
2171                         diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2172                         fault_types[priv] += p->numa_faults[membuf_idx];
2173                         p->numa_faults[membuf_idx] = 0;
2174 
2175                         /*
2176                          * Normalize the faults_from, so all tasks in a group
2177                          * count according to CPU use, instead of by the raw
2178                          * number of faults. Tasks with little runtime have
2179                          * little over-all impact on throughput, and thus their
2180                          * faults are less important.
2181                          */
2182                         f_weight = div64_u64(runtime << 16, period + 1);
2183                         f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2184                                    (total_faults + 1);
2185                         f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2186                         p->numa_faults[cpubuf_idx] = 0;
2187 
2188                         p->numa_faults[mem_idx] += diff;
2189                         p->numa_faults[cpu_idx] += f_diff;
2190                         faults += p->numa_faults[mem_idx];
2191                         p->total_numa_faults += diff;
2192                         if (ng) {
2193                                 /*
2194                                  * safe because we can only change our own group
2195                                  *
2196                                  * mem_idx represents the offset for a given
2197                                  * nid and priv in a specific region because it
2198                                  * is at the beginning of the numa_faults array.
2199                                  */
2200                                 ng->faults[mem_idx] += diff;
2201                                 ng->faults_cpu[mem_idx] += f_diff;
2202                                 ng->total_faults += diff;
2203                                 group_faults += ng->faults[mem_idx];
2204                         }
2205                 }
2206 
2207                 if (!ng) {
2208                         if (faults > max_faults) {
2209                                 max_faults = faults;
2210                                 max_nid = nid;
2211                         }
2212                 } else if (group_faults > max_faults) {
2213                         max_faults = group_faults;
2214                         max_nid = nid;
2215                 }
2216         }
2217 
2218         if (ng) {
2219                 numa_group_count_active_nodes(ng);
2220                 spin_unlock_irq(group_lock);
2221                 max_nid = preferred_group_nid(p, max_nid);
2222         }
2223 
2224         if (max_faults) {
2225                 /* Set the new preferred node */
2226                 if (max_nid != p->numa_preferred_nid)
2227                         sched_setnuma(p, max_nid);
2228         }
2229 
2230         update_task_scan_period(p, fault_types[0], fault_types[1]);
2231 }
2232 
2233 static inline int get_numa_group(struct numa_group *grp)
2234 {
2235         return refcount_inc_not_zero(&grp->refcount);
2236 }
2237 
2238 static inline void put_numa_group(struct numa_group *grp)
2239 {
2240         if (refcount_dec_and_test(&grp->refcount))
2241                 kfree_rcu(grp, rcu);
2242 }
2243 
2244 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2245                         int *priv)
2246 {
2247         struct numa_group *grp, *my_grp;
2248         struct task_struct *tsk;
2249         bool join = false;
2250         int cpu = cpupid_to_cpu(cpupid);
2251         int i;
2252 
2253         if (unlikely(!deref_curr_numa_group(p))) {
2254                 unsigned int size = sizeof(struct numa_group) +
2255                                     4*nr_node_ids*sizeof(unsigned long);
2256 
2257                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2258                 if (!grp)
2259                         return;
2260 
2261                 refcount_set(&grp->refcount, 1);
2262                 grp->active_nodes = 1;
2263                 grp->max_faults_cpu = 0;
2264                 spin_lock_init(&grp->lock);
2265                 grp->gid = p->pid;
2266                 /* Second half of the array tracks nids where faults happen */
2267                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2268                                                 nr_node_ids;
2269 
2270                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2271                         grp->faults[i] = p->numa_faults[i];
2272 
2273                 grp->total_faults = p->total_numa_faults;
2274 
2275                 grp->nr_tasks++;
2276                 rcu_assign_pointer(p->numa_group, grp);
2277         }
2278 
2279         rcu_read_lock();
2280         tsk = READ_ONCE(cpu_rq(cpu)->curr);
2281 
2282         if (!cpupid_match_pid(tsk, cpupid))
2283                 goto no_join;
2284 
2285         grp = rcu_dereference(tsk->numa_group);
2286         if (!grp)
2287                 goto no_join;
2288 
2289         my_grp = deref_curr_numa_group(p);
2290         if (grp == my_grp)
2291                 goto no_join;
2292 
2293         /*
2294          * Only join the other group if its bigger; if we're the bigger group,
2295          * the other task will join us.
2296          */
2297         if (my_grp->nr_tasks > grp->nr_tasks)
2298                 goto no_join;
2299 
2300         /*
2301          * Tie-break on the grp address.
2302          */
2303         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2304                 goto no_join;
2305 
2306         /* Always join threads in the same process. */
2307         if (tsk->mm == current->mm)
2308                 join = true;
2309 
2310         /* Simple filter to avoid false positives due to PID collisions */
2311         if (flags & TNF_SHARED)
2312                 join = true;
2313 
2314         /* Update priv based on whether false sharing was detected */
2315         *priv = !join;
2316 
2317         if (join && !get_numa_group(grp))
2318                 goto no_join;
2319 
2320         rcu_read_unlock();
2321 
2322         if (!join)
2323                 return;
2324 
2325         BUG_ON(irqs_disabled());
2326         double_lock_irq(&my_grp->lock, &grp->lock);
2327 
2328         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2329                 my_grp->faults[i] -= p->numa_faults[i];
2330                 grp->faults[i] += p->numa_faults[i];
2331         }
2332         my_grp->total_faults -= p->total_numa_faults;
2333         grp->total_faults += p->total_numa_faults;
2334 
2335         my_grp->nr_tasks--;
2336         grp->nr_tasks++;
2337 
2338         spin_unlock(&my_grp->lock);
2339         spin_unlock_irq(&grp->lock);
2340 
2341         rcu_assign_pointer(p->numa_group, grp);
2342 
2343         put_numa_group(my_grp);
2344         return;
2345 
2346 no_join:
2347         rcu_read_unlock();
2348         return;
2349 }
2350 
2351 /*
2352  * Get rid of NUMA staticstics associated with a task (either current or dead).
2353  * If @final is set, the task is dead and has reached refcount zero, so we can
2354  * safely free all relevant data structures. Otherwise, there might be
2355  * concurrent reads from places like load balancing and procfs, and we should
2356  * reset the data back to default state without freeing ->numa_faults.
2357  */
2358 void task_numa_free(struct task_struct *p, bool final)
2359 {
2360         /* safe: p either is current or is being freed by current */
2361         struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2362         unsigned long *numa_faults = p->numa_faults;
2363         unsigned long flags;
2364         int i;
2365 
2366         if (!numa_faults)
2367                 return;
2368 
2369         if (grp) {
2370                 spin_lock_irqsave(&grp->lock, flags);
2371                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2372                         grp->faults[i] -= p->numa_faults[i];
2373                 grp->total_faults -= p->total_numa_faults;
2374 
2375                 grp->nr_tasks--;
2376                 spin_unlock_irqrestore(&grp->lock, flags);
2377                 RCU_INIT_POINTER(p->numa_group, NULL);
2378                 put_numa_group(grp);
2379         }
2380 
2381         if (final) {
2382                 p->numa_faults = NULL;
2383                 kfree(numa_faults);
2384         } else {
2385                 p->total_numa_faults = 0;
2386                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2387                         numa_faults[i] = 0;
2388         }
2389 }
2390 
2391 /*
2392  * Got a PROT_NONE fault for a page on @node.
2393  */
2394 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2395 {
2396         struct task_struct *p = current;
2397         bool migrated = flags & TNF_MIGRATED;
2398         int cpu_node = task_node(current);
2399         int local = !!(flags & TNF_FAULT_LOCAL);
2400         struct numa_group *ng;
2401         int priv;
2402 
2403         if (!static_branch_likely(&sched_numa_balancing))
2404                 return;
2405 
2406         /* for example, ksmd faulting in a user's mm */
2407         if (!p->mm)
2408                 return;
2409 
2410         /* Allocate buffer to track faults on a per-node basis */
2411         if (unlikely(!p->numa_faults)) {
2412                 int size = sizeof(*p->numa_faults) *
2413                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2414 
2415                 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2416                 if (!p->numa_faults)
2417                         return;
2418 
2419                 p->total_numa_faults = 0;
2420                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2421         }
2422 
2423         /*
2424          * 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);
2431                 if (!priv && !(flags & TNF_NO_GROUP))
2432                         task_numa_group(p, last_cpupid, flags, &priv);
2433         }
2434 
2435         /*
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          */
2441         ng = deref_curr_numa_group(p);
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))
2445                 local = 1;
2446 
2447         /*
2448          * Retry to migrate task to preferred node periodically, in case it
2449          * previously failed, or the scheduler moved us.
2450          */
2451         if (time_after(jiffies, p->numa_migrate_retry)) {
2452                 task_numa_placement(p);
2453                 numa_migrate_preferred(p);
2454         }
2455 
2456         if (migrated)
2457                 p->numa_pages_migrated += pages;
2458         if (flags & TNF_MIGRATE_FAIL)
2459                 p->numa_faults_locality[2] += pages;
2460 
2461         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;
2463         p->numa_faults_locality[local] += pages;
2464 }
2465 
2466 static void reset_ptenuma_scan(struct task_struct *p)
2467 {
2468         /*
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          */
2476         WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2477         p->mm->numa_scan_offset = 0;
2478 }
2479 
2480 /*
2481  * The expensive part of numa migration is done from task_work context.
2482  * Triggered from task_tick_numa().
2483  */
2484 static void 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;
2489         u64 runtime = p->se.sum_exec_runtime;
2490         struct vm_area_struct *vma;
2491         unsigned long start, end;
2492         unsigned long nr_pte_updates = 0;
2493         long pages, virtpages;
2494 
2495         SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2496 
2497         work->next = work;
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 
2509         if (!mm->numa_next_scan) {
2510                 mm->numa_next_scan = now +
2511                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2512         }
2513 
2514         /*
2515          * Enforce maximal scan/migration frequency..
2516          */
2517         migrate = mm->numa_next_scan;
2518         if (time_before(now, migrate))
2519                 return;
2520 
2521         if (p->numa_scan_period == 0) {
2522                 p->numa_scan_period_max = task_scan_max(p);
2523                 p->numa_scan_period = task_scan_start(p);
2524         }
2525 
2526         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2527         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2528                 return;
2529 
2530         /*
2531          * 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 
2536         start = mm->numa_scan_offset;
2537         pages = sysctl_numa_balancing_scan_size;
2538         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2539         virtpages = pages * 8;     /* Scan up to this much virtual space */
2540         if (!pages)
2541                 return;
2542 
2543 
2544         if (!down_read_trylock(&mm->mmap_sem))
2545                 return;
2546         vma = find_vma(mm, start);
2547         if (!vma) {
2548                 reset_ptenuma_scan(p);
2549                 start = 0;
2550                 vma = mm->mmap;
2551         }
2552         for (; vma; vma = vma->vm_next) {
2553                 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2554                         is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2555                         continue;
2556                 }
2557 
2558                 /*
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 
2568                 /*
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 
2575                 do {
2576                         start = max(start, vma->vm_start);
2577                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2578                         end = min(end, vma->vm_end);
2579                         nr_pte_updates = change_prot_numa(vma, start, end);
2580 
2581                         /*
2582                          * 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.
2588                          */
2589                         if (nr_pte_updates)
2590                                 pages -= (end - start) >> PAGE_SHIFT;
2591                         virtpages -= (end - start) >> PAGE_SHIFT;
2592 
2593                         start = end;
2594                         if (pages <= 0 || virtpages <= 0)
2595                                 goto out;
2596 
2597                         cond_resched();
2598                 } while (end != vma->vm_end);
2599         }
2600 
2601 out:
2602         /*
2603          * 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.
2607          */
2608         if (vma)
2609                 mm->numa_scan_offset = start;
2610         else
2611                 reset_ptenuma_scan(p);
2612         up_read(&mm->mmap_sem);
2613 
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         }
2624 }
2625 
2626 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2627 {
2628         int mm_users = 0;
2629         struct mm_struct *mm = p->mm;
2630 
2631         if (mm) {
2632                 mm_users = atomic_read(&mm->mm_users);
2633                 if (mm_users == 1) {
2634                         mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2635                         mm->numa_scan_seq = 0;
2636                 }
2637         }
2638         p->node_stamp                   = 0;
2639         p->numa_scan_seq                = mm ? mm->numa_scan_seq : 0;
2640         p->numa_scan_period             = sysctl_numa_balancing_scan_delay;
2641         /* Protect against double add, see task_tick_numa and task_numa_work */
2642         p->numa_work.next               = &p->numa_work;
2643         p->numa_faults                  = NULL;
2644         RCU_INIT_POINTER(p->numa_group, NULL);
2645         p->last_task_numa_placement     = 0;
2646         p->last_sum_exec_runtime        = 0;
2647 
2648         init_task_work(&p->numa_work, task_numa_work);
2649 
2650         /* New address space, reset the preferred nid */
2651         if (!(clone_flags & CLONE_VM)) {
2652                 p->numa_preferred_nid = NUMA_NO_NODE;
2653                 return;
2654         }
2655 
2656         /*
2657          * New thread, keep existing numa_preferred_nid which should be copied
2658          * already by arch_dup_task_struct but stagger when scans start.
2659          */
2660         if (mm) {
2661                 unsigned int delay;
2662 
2663                 delay = min_t(unsigned int, task_scan_max(current),
2664                         current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2665                 delay += 2 * TICK_NSEC;
2666                 p->node_stamp = delay;
2667         }
2668 }
2669 
2670 /*
2671  * Drive the periodic memory faults..
2672  */
2673 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2674 {
2675         struct callback_head *work = &curr->numa_work;
2676         u64 period, now;
2677 
2678         /*
2679          * We don't care about NUMA placement if we don't have memory.
2680          */
2681         if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2682                 return;
2683 
2684         /*
2685          * Using runtime rather than walltime has the dual advantage that
2686          * we (mostly) drive the selection from busy threads and that the
2687          * task needs to have done some actual work before we bother with
2688          * NUMA placement.
2689          */
2690         now = curr->se.sum_exec_runtime;
2691         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2692 
2693         if (now > curr->node_stamp + period) {
2694                 if (!curr->node_stamp)
2695                         curr->numa_scan_period = task_scan_start(curr);
2696                 curr->node_stamp += period;
2697 
2698                 if (!time_before(jiffies, curr->mm->numa_next_scan))
2699                         task_work_add(curr, work, true);
2700         }
2701 }
2702 
2703 static void update_scan_period(struct task_struct *p, int new_cpu)
2704 {
2705         int src_nid = cpu_to_node(task_cpu(p));
2706         int dst_nid = cpu_to_node(new_cpu);
2707 
2708         if (!static_branch_likely(&sched_numa_balancing))
2709                 return;
2710 
2711         if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2712                 return;
2713 
2714         if (src_nid == dst_nid)
2715                 return;
2716 
2717         /*
2718          * Allow resets if faults have been trapped before one scan
2719          * has completed. This is most likely due to a new task that
2720          * is pulled cross-node due to wakeups or load balancing.
2721          */
2722         if (p->numa_scan_seq) {
2723                 /*
2724                  * Avoid scan adjustments if moving to the preferred
2725                  * node or if the task was not previously running on
2726                  * the preferred node.
2727                  */
2728                 if (dst_nid == p->numa_preferred_nid ||
2729                     (p->numa_preferred_nid != NUMA_NO_NODE &&
2730                         src_nid != p->numa_preferred_nid))
2731                         return;
2732         }
2733 
2734         p->numa_scan_period = task_scan_start(p);
2735 }
2736 
2737 #else
2738 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2739 {
2740 }
2741 
2742 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2743 {
2744 }
2745 
2746 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2747 {
2748 }
2749 
2750 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2751 {
2752 }
2753 
2754 #endif /* CONFIG_NUMA_BALANCING */
2755 
2756 static void
2757 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2758 {
2759         update_load_add(&cfs_rq->load, se->load.weight);
2760 #ifdef CONFIG_SMP
2761         if (entity_is_task(se)) {
2762                 struct rq *rq = rq_of(cfs_rq);
2763 
2764                 account_numa_enqueue(rq, task_of(se));
2765                 list_add(&se->group_node, &rq->cfs_tasks);
2766         }
2767 #endif
2768         cfs_rq->nr_running++;
2769 }
2770 
2771 static void
2772 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2773 {
2774         update_load_sub(&cfs_rq->load, se->load.weight);
2775 #ifdef CONFIG_SMP
2776         if (entity_is_task(se)) {
2777                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2778                 list_del_init(&se->group_node);
2779         }
2780 #endif
2781         cfs_rq->nr_running--;
2782 }
2783 
2784 /*
2785  * Signed add and clamp on underflow.
2786  *
2787  * Explicitly do a load-store to ensure the intermediate value never hits
2788  * memory. This allows lockless observations without ever seeing the negative
2789  * values.
2790  */
2791 #define add_positive(_ptr, _val) do {                           \
2792         typeof(_ptr) ptr = (_ptr);                              \
2793         typeof(_val) val = (_val);                              \
2794         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
2795                                                                 \
2796         res = var + val;                                        \
2797                                                                 \
2798         if (val < 0 && res > var)                               \
2799                 res = 0;                                        \
2800                                                                 \
2801         WRITE_ONCE(*ptr, res);                                  \
2802 } while (0)
2803 
2804 /*
2805  * Unsigned subtract and clamp on underflow.
2806  *
2807  * Explicitly do a load-store to ensure the intermediate value never hits
2808  * memory. This allows lockless observations without ever seeing the negative
2809  * values.
2810  */
2811 #define sub_positive(_ptr, _val) do {                           \
2812         typeof(_ptr) ptr = (_ptr);                              \
2813         typeof(*ptr) val = (_val);                              \
2814         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
2815         res = var - val;                                        \
2816         if (res > var)                                          \
2817                 res = 0;                                        \
2818         WRITE_ONCE(*ptr, res);                                  \
2819 } while (0)
2820 
2821 /*
2822  * Remove and clamp on negative, from a local variable.
2823  *
2824  * A variant of sub_positive(), which does not use explicit load-store
2825  * and is thus optimized for local variable updates.
2826  */
2827 #define lsub_positive(_ptr, _val) do {                          \
2828         typeof(_ptr) ptr = (_ptr);                              \
2829         *ptr -= min_t(typeof(*ptr), *ptr, _val);                \
2830 } while (0)
2831 
2832 #ifdef CONFIG_SMP
2833 static inline void
2834 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2835 {
2836         cfs_rq->runnable_weight += se->runnable_weight;
2837 
2838         cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2839         cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2840 }
2841 
2842 static inline void
2843 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2844 {
2845         cfs_rq->runnable_weight -= se->runnable_weight;
2846 
2847         sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2848         sub_positive(&cfs_rq->avg.runnable_load_sum,
2849                      se_runnable(se) * se->avg.runnable_load_sum);
2850 }
2851 
2852 static inline void
2853 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2854 {
2855         cfs_rq->avg.load_avg += se->avg.load_avg;
2856         cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2857 }
2858 
2859 static inline void
2860 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2861 {
2862         sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2863         sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2864 }
2865 #else
2866 static inline void
2867 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2868 static inline void
2869 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2870 static inline void
2871 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2872 static inline void
2873 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2874 #endif
2875 
2876 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2877                             unsigned long weight, unsigned long runnable)
2878 {
2879         if (se->on_rq) {
2880                 /* commit outstanding execution time */
2881                 if (cfs_rq->curr == se)
2882                         update_curr(cfs_rq);
2883                 account_entity_dequeue(cfs_rq, se);
2884                 dequeue_runnable_load_avg(cfs_rq, se);
2885         }
2886         dequeue_load_avg(cfs_rq, se);
2887 
2888         se->runnable_weight = runnable;
2889         update_load_set(&se->load, weight);
2890 
2891 #ifdef CONFIG_SMP
2892         do {
2893                 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2894 
2895                 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2896                 se->avg.runnable_load_avg =
2897                         div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2898         } while (0);
2899 #endif
2900 
2901         enqueue_load_avg(cfs_rq, se);
2902         if (se->on_rq) {
2903                 account_entity_enqueue(cfs_rq, se);
2904                 enqueue_runnable_load_avg(cfs_rq, se);
2905         }
2906 }
2907 
2908 void reweight_task(struct task_struct *p, int prio)
2909 {
2910         struct sched_entity *se = &p->se;
2911         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2912         struct load_weight *load = &se->load;
2913         unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2914 
2915         reweight_entity(cfs_rq, se, weight, weight);
2916         load->inv_weight = sched_prio_to_wmult[prio];
2917 }
2918 
2919 #ifdef CONFIG_FAIR_GROUP_SCHED
2920 #ifdef CONFIG_SMP
2921 /*
2922  * All this does is approximate the hierarchical proportion which includes that
2923  * global sum we all love to hate.
2924  *
2925  * That is, the weight of a group entity, is the proportional share of the
2926  * group weight based on the group runqueue weights. That is:
2927  *
2928  *                     tg->weight * grq->load.weight
2929  *   ge->load.weight = -----------------------------               (1)
2930  *                        \Sum grq->load.weight
2931  *
2932  * Now, because computing that sum is prohibitively expensive to compute (been
2933  * there, done that) we approximate it with this average stuff. The average
2934  * moves slower and therefore the approximation is cheaper and more stable.
2935  *
2936  * So instead of the above, we substitute:
2937  *
2938  *   grq->load.weight -> grq->avg.load_avg                         (2)
2939  *
2940  * which yields the following:
2941  *
2942  *                     tg->weight * grq->avg.load_avg
2943  *   ge->load.weight = ------------------------------              (3)
2944  *                              tg->load_avg
2945  *
2946  * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2947  *
2948  * That is shares_avg, and it is right (given the approximation (2)).
2949  *
2950  * The problem with it is that because the average is slow -- it was designed
2951  * to be exactly that of course -- this leads to transients in boundary
2952  * conditions. In specific, the case where the group was idle and we start the
2953  * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2954  * yielding bad latency etc..
2955  *
2956  * Now, in that special case (1) reduces to:
2957  *
2958  *                     tg->weight * grq->load.weight
2959  *   ge->load.weight = ----------------------------- = tg->weight   (4)
2960  *                          grp->load.weight
2961  *
2962  * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2963  *
2964  * So what we do is modify our approximation (3) to approach (4) in the (near)
2965  * UP case, like:
2966  *
2967  *   ge->load.weight =
2968  *
2969  *              tg->weight * grq->load.weight
2970  *     ---------------------------------------------------         (5)
2971  *     tg->load_avg - grq->avg.load_avg + grq->load.weight
2972  *
2973  * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2974  * we need to use grq->avg.load_avg as its lower bound, which then gives:
2975  *
2976  *
2977  *                     tg->weight * grq->load.weight
2978  *   ge->load.weight = -----------------------------               (6)
2979  *                              tg_load_avg'
2980  *
2981  * Where:
2982  *
2983  *   tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2984  *                  max(grq->load.weight, grq->avg.load_avg)
2985  *
2986  * And that is shares_weight and is icky. In the (near) UP case it approaches
2987  * (4) while in the normal case it approaches (3). It consistently
2988  * overestimates the ge->load.weight and therefore:
2989  *
2990  *   \Sum ge->load.weight >= tg->weight
2991  *
2992  * hence icky!
2993  */
2994 static long calc_group_shares(struct cfs_rq *cfs_rq)
2995 {
2996         long tg_weight, tg_shares, load, shares;
2997         struct task_group *tg = cfs_rq->tg;
2998 
2999         tg_shares = READ_ONCE(tg->shares);
3000 
3001         load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3002 
3003         tg_weight = atomic_long_read(&tg->load_avg);
3004 
3005         /* Ensure tg_weight >= load */
3006         tg_weight -= cfs_rq->tg_load_avg_contrib;
3007         tg_weight += load;
3008 
3009         shares = (tg_shares * load);
3010         if (tg_weight)
3011                 shares /= tg_weight;
3012 
3013         /*
3014          * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3015          * of a group with small tg->shares value. It is a floor value which is
3016          * assigned as a minimum load.weight to the sched_entity representing
3017          * the group on a CPU.
3018          *
3019          * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3020          * on an 8-core system with 8 tasks each runnable on one CPU shares has
3021          * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3022          * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3023          * instead of 0.
3024          */
3025         return clamp_t(long, shares, MIN_SHARES, tg_shares);
3026 }
3027 
3028 /*
3029  * This calculates the effective runnable weight for a group entity based on
3030  * the group entity weight calculated above.
3031  *
3032  * Because of the above approximation (2), our group entity weight is
3033  * an load_avg based ratio (3). This means that it includes blocked load and
3034  * does not represent the runnable weight.
3035  *
3036  * Approximate the group entity's runnable weight per ratio from the group
3037  * runqueue:
3038  *
3039  *                                           grq->avg.runnable_load_avg
3040  *   ge->runnable_weight = ge->load.weight * -------------------------- (7)
3041  *                                               grq->avg.load_avg
3042  *
3043  * However, analogous to above, since the avg numbers are slow, this leads to
3044  * transients in the from-idle case. Instead we use:
3045  *
3046  *   ge->runnable_weight = ge->load.weight *
3047  *
3048  *              max(grq->avg.runnable_load_avg, grq->runnable_weight)
3049  *              -----------------------------------------------------   (8)
3050  *                    max(grq->avg.load_avg, grq->load.weight)
3051  *
3052  * Where these max() serve both to use the 'instant' values to fix the slow
3053  * from-idle and avoid the /0 on to-idle, similar to (6).
3054  */
3055 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
3056 {
3057         long runnable, load_avg;
3058 
3059         load_avg = max(cfs_rq->avg.load_avg,
3060                        scale_load_down(cfs_rq->load.weight));
3061 
3062         runnable = max(cfs_rq->avg.runnable_load_avg,
3063                        scale_load_down(cfs_rq->runnable_weight));
3064 
3065         runnable *= shares;
3066         if (load_avg)
3067                 runnable /= load_avg;
3068 
3069         return clamp_t(long, runnable, MIN_SHARES, shares);
3070 }
3071 #endif /* CONFIG_SMP */
3072 
3073 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3074 
3075 /*
3076  * Recomputes the group entity based on the current state of its group
3077  * runqueue.
3078  */
3079 static void update_cfs_group(struct sched_entity *se)
3080 {
3081         struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3082         long shares, runnable;
3083 
3084         if (!gcfs_rq)
3085                 return;
3086 
3087         if (throttled_hierarchy(gcfs_rq))
3088                 return;
3089 
3090 #ifndef CONFIG_SMP
3091         runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3092 
3093         if (likely(se->load.weight == shares))
3094                 return;
3095 #else
3096         shares   = calc_group_shares(gcfs_rq);
3097         runnable = calc_group_runnable(gcfs_rq, shares);
3098 #endif
3099 
3100         reweight_entity(cfs_rq_of(se), se, shares, runnable);
3101 }
3102 
3103 #else /* CONFIG_FAIR_GROUP_SCHED */
3104 static inline void update_cfs_group(struct sched_entity *se)
3105 {
3106 }
3107 #endif /* CONFIG_FAIR_GROUP_SCHED */
3108 
3109 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3110 {
3111         struct rq *rq = rq_of(cfs_rq);
3112 
3113         if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3114                 /*
3115                  * There are a few boundary cases this might miss but it should
3116                  * get called often enough that that should (hopefully) not be
3117                  * a real problem.
3118                  *
3119                  * It will not get called when we go idle, because the idle
3120                  * thread is a different class (!fair), nor will the utilization
3121                  * number include things like RT tasks.
3122                  *
3123                  * As is, the util number is not freq-invariant (we'd have to
3124                  * implement arch_scale_freq_capacity() for that).
3125                  *
3126                  * See cpu_util().
3127                  */
3128                 cpufreq_update_util(rq, flags);
3129         }
3130 }
3131 
3132 #ifdef CONFIG_SMP
3133 #ifdef CONFIG_FAIR_GROUP_SCHED
3134 /**
3135  * update_tg_load_avg - update the tg's load avg
3136  * @cfs_rq: the cfs_rq whose avg changed
3137  * @force: update regardless of how small the difference
3138  *
3139  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3140  * However, because tg->load_avg is a global value there are performance
3141  * considerations.
3142  *
3143  * In order to avoid having to look at the other cfs_rq's, we use a
3144  * differential update where we store the last value we propagated. This in
3145  * turn allows skipping updates if the differential is 'small'.
3146  *
3147  * Updating tg's load_avg is necessary before update_cfs_share().
3148  */
3149 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3150 {
3151         long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3152 
3153         /*
3154          * No need to update load_avg for root_task_group as it is not used.
3155          */
3156         if (cfs_rq->tg == &root_task_group)
3157                 return;
3158 
3159         if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3160                 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3161                 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3162         }
3163 }
3164 
3165 /*
3166  * Called within set_task_rq() right before setting a task's CPU. The
3167  * caller only guarantees p->pi_lock is held; no other assumptions,
3168  * including the state of rq->lock, should be made.
3169  */
3170 void set_task_rq_fair(struct sched_entity *se,
3171                       struct cfs_rq *prev, struct cfs_rq *next)
3172 {
3173         u64 p_last_update_time;
3174         u64 n_last_update_time;
3175 
3176         if (!sched_feat(ATTACH_AGE_LOAD))
3177                 return;
3178 
3179         /*
3180          * We are supposed to update the task to "current" time, then its up to
3181          * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3182          * getting what current time is, so simply throw away the out-of-date
3183          * time. This will result in the wakee task is less decayed, but giving
3184          * the wakee more load sounds not bad.
3185          */
3186         if (!(se->avg.last_update_time && prev))
3187                 return;
3188 
3189 #ifndef CONFIG_64BIT
3190         {
3191                 u64 p_last_update_time_copy;
3192                 u64 n_last_update_time_copy;
3193 
3194                 do {
3195                         p_last_update_time_copy = prev->load_last_update_time_copy;
3196                         n_last_update_time_copy = next->load_last_update_time_copy;
3197 
3198                         smp_rmb();
3199 
3200                         p_last_update_time = prev->avg.last_update_time;
3201                         n_last_update_time = next->avg.last_update_time;
3202 
3203                 } while (p_last_update_time != p_last_update_time_copy ||
3204                          n_last_update_time != n_last_update_time_copy);
3205         }
3206 #else
3207         p_last_update_time = prev->avg.last_update_time;
3208         n_last_update_time = next->avg.last_update_time;
3209 #endif
3210         __update_load_avg_blocked_se(p_last_update_time, se);
3211         se->avg.last_update_time = n_last_update_time;
3212 }
3213 
3214 
3215 /*
3216  * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3217  * propagate its contribution. The key to this propagation is the invariant
3218  * that for each group:
3219  *
3220  *   ge->avg == grq->avg                                                (1)
3221  *
3222  * _IFF_ we look at the pure running and runnable sums. Because they
3223  * represent the very same entity, just at different points in the hierarchy.
3224  *
3225  * Per the above update_tg_cfs_util() is trivial and simply copies the running
3226  * sum over (but still wrong, because the group entity and group rq do not have
3227  * their PELT windows aligned).
3228  *
3229  * However, update_tg_cfs_runnable() is more complex. So we have:
3230  *
3231  *   ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg          (2)
3232  *
3233  * And since, like util, the runnable part should be directly transferable,
3234  * the following would _appear_ to be the straight forward approach:
3235  *
3236  *   grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg       (3)
3237  *
3238  * And per (1) we have:
3239  *
3240  *   ge->avg.runnable_avg == grq->avg.runnable_avg
3241  *
3242  * Which gives:
3243  *
3244  *                      ge->load.weight * grq->avg.load_avg
3245  *   ge->avg.load_avg = -----------------------------------             (4)
3246  *                               grq->load.weight
3247  *
3248  * Except that is wrong!
3249  *
3250  * Because while for entities historical weight is not important and we
3251  * really only care about our future and therefore can consider a pure
3252  * runnable sum, runqueues can NOT do this.
3253  *
3254  * We specifically want runqueues to have a load_avg that includes
3255  * historical weights. Those represent the blocked load, the load we expect
3256  * to (shortly) return to us. This only works by keeping the weights as
3257  * integral part of the sum. We therefore cannot decompose as per (3).
3258  *
3259  * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3260  * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3261  * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3262  * runnable section of these tasks overlap (or not). If they were to perfectly
3263  * align the rq as a whole would be runnable 2/3 of the time. If however we
3264  * always have at least 1 runnable task, the rq as a whole is always runnable.
3265  *
3266  * So we'll have to approximate.. :/
3267  *
3268  * Given the constraint:
3269  *
3270  *   ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3271  *
3272  * We can construct a rule that adds runnable to a rq by assuming minimal
3273  * overlap.
3274  *
3275  * On removal, we'll assume each task is equally runnable; which yields:
3276  *
3277  *   grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3278  *
3279  * XXX: only do this for the part of runnable > running ?
3280  *
3281  */
3282 
3283 static inline void
3284 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3285 {
3286         long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3287 
3288         /* Nothing to update */
3289         if (!delta)
3290                 return;
3291 
3292         /*
3293          * The relation between sum and avg is:
3294          *
3295          *   LOAD_AVG_MAX - 1024 + sa->period_contrib
3296          *
3297          * however, the PELT windows are not aligned between grq and gse.
3298          */
3299 
3300         /* Set new sched_entity's utilization */
3301         se->avg.util_avg = gcfs_rq->avg.util_avg;
3302         se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3303 
3304         /* Update parent cfs_rq utilization */
3305         add_positive(&cfs_rq->avg.util_avg, delta);
3306         cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3307 }
3308 
3309 static inline void
3310 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3311 {
3312         long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3313         unsigned long runnable_load_avg, load_avg;
3314         u64 runnable_load_sum, load_sum = 0;
3315         s64 delta_sum;
3316 
3317         if (!runnable_sum)
3318                 return;
3319 
3320         gcfs_rq->prop_runnable_sum = 0;
3321 
3322         if (runnable_sum >= 0) {
3323                 /*
3324                  * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3325                  * the CPU is saturated running == runnable.
3326                  */
3327                 runnable_sum += se->avg.load_sum;
3328                 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3329         } else {
3330                 /*
3331                  * Estimate the new unweighted runnable_sum of the gcfs_rq by
3332                  * assuming all tasks are equally runnable.
3333                  */
3334                 if (scale_load_down(gcfs_rq->load.weight)) {
3335                         load_sum = div_s64(gcfs_rq->avg.load_sum,
3336                                 scale_load_down(gcfs_rq->load.weight));
3337                 }
3338 
3339                 /* But make sure to not inflate se's runnable */
3340                 runnable_sum = min(se->avg.load_sum, load_sum);
3341         }
3342 
3343         /*
3344          * runnable_sum can't be lower than running_sum
3345          * Rescale running sum to be in the same range as runnable sum
3346          * running_sum is in [0 : LOAD_AVG_MAX <<  SCHED_CAPACITY_SHIFT]
3347          * runnable_sum is in [0 : LOAD_AVG_MAX]
3348          */
3349         running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3350         runnable_sum = max(runnable_sum, running_sum);
3351 
3352         load_sum = (s64)se_weight(se) * runnable_sum;
3353         load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3354 
3355         delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3356         delta_avg = load_avg - se->avg.load_avg;
3357 
3358         se->avg.load_sum = runnable_sum;
3359         se->avg.load_avg = load_avg;
3360         add_positive(&cfs_rq->avg.load_avg, delta_avg);
3361         add_positive(&cfs_rq->avg.load_sum, delta_sum);
3362 
3363         runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3364         runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3365         delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3366         delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3367 
3368         se->avg.runnable_load_sum = runnable_sum;
3369         se->avg.runnable_load_avg = runnable_load_avg;
3370 
3371         if (se->on_rq) {
3372                 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3373                 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3374         }
3375 }
3376 
3377 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3378 {
3379         cfs_rq->propagate = 1;
3380         cfs_rq->prop_runnable_sum += runnable_sum;
3381 }
3382 
3383 /* Update task and its cfs_rq load average */
3384 static inline int propagate_entity_load_avg(struct sched_entity *se)
3385 {
3386         struct cfs_rq *cfs_rq, *gcfs_rq;
3387 
3388         if (entity_is_task(se))
3389                 return 0;
3390 
3391         gcfs_rq = group_cfs_rq(se);
3392         if (!gcfs_rq->propagate)
3393                 return 0;
3394 
3395         gcfs_rq->propagate = 0;
3396 
3397         cfs_rq = cfs_rq_of(se);
3398 
3399         add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3400 
3401         update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3402         update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3403 
3404         trace_pelt_cfs_tp(cfs_rq);
3405         trace_pelt_se_tp(se);
3406 
3407         return 1;
3408 }
3409 
3410 /*
3411  * Check if we need to update the load and the utilization of a blocked
3412  * group_entity:
3413  */
3414 static inline bool skip_blocked_update(struct sched_entity *se)
3415 {
3416         struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3417 
3418         /*
3419          * If sched_entity still have not zero load or utilization, we have to
3420          * decay it:
3421          */
3422         if (se->avg.load_avg || se->avg.util_avg)
3423                 return false;
3424 
3425         /*
3426          * If there is a pending propagation, we have to update the load and
3427          * the utilization of the sched_entity:
3428          */
3429         if (gcfs_rq->propagate)
3430                 return false;
3431 
3432         /*
3433          * Otherwise, the load and the utilization of the sched_entity is
3434          * already zero and there is no pending propagation, so it will be a
3435          * waste of time to try to decay it:
3436          */
3437         return true;
3438 }
3439 
3440 #else /* CONFIG_FAIR_GROUP_SCHED */
3441 
3442 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3443 
3444 static inline int propagate_entity_load_avg(struct sched_entity *se)
3445 {
3446         return 0;
3447 }
3448 
3449 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3450 
3451 #endif /* CONFIG_FAIR_GROUP_SCHED */
3452 
3453 /**
3454  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3455  * @now: current time, as per cfs_rq_clock_pelt()
3456  * @cfs_rq: cfs_rq to update
3457  *
3458  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3459  * avg. The immediate corollary is that all (fair) tasks must be attached, see
3460  * post_init_entity_util_avg().
3461  *
3462  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3463  *
3464  * Returns true if the load decayed or we removed load.
3465  *
3466  * Since both these conditions indicate a changed cfs_rq->avg.load we should
3467  * call update_tg_load_avg() when this function returns true.
3468  */
3469 static inline int
3470 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3471 {
3472         unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3473         struct sched_avg *sa = &cfs_rq->avg;
3474         int decayed = 0;
3475 
3476         if (cfs_rq->removed.nr) {
3477                 unsigned long r;
3478                 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3479 
3480                 raw_spin_lock(&cfs_rq->removed.lock);
3481                 swap(cfs_rq->removed.util_avg, removed_util);
3482                 swap(cfs_rq->removed.load_avg, removed_load);
3483                 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3484                 cfs_rq->removed.nr = 0;
3485                 raw_spin_unlock(&cfs_rq->removed.lock);
3486 
3487                 r = removed_load;
3488                 sub_positive(&sa->load_avg, r);
3489                 sub_positive(&sa->load_sum, r * divider);
3490 
3491                 r = removed_util;
3492                 sub_positive(&sa->util_avg, r);
3493                 sub_positive(&sa->util_sum, r * divider);
3494 
3495                 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3496 
3497                 decayed = 1;
3498         }
3499 
3500         decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3501 
3502 #ifndef CONFIG_64BIT
3503         smp_wmb();
3504         cfs_rq->load_last_update_time_copy = sa->last_update_time;
3505 #endif
3506 
3507         return decayed;
3508 }
3509 
3510 /**
3511  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3512  * @cfs_rq: cfs_rq to attach to
3513  * @se: sched_entity to attach
3514  * @flags: migration hints
3515  *
3516  * Must call update_cfs_rq_load_avg() before this, since we rely on
3517  * cfs_rq->avg.last_update_time being current.
3518  */
3519 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3520 {
3521         u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3522 
3523         /*
3524          * When we attach the @se to the @cfs_rq, we must align the decay
3525          * window because without that, really weird and wonderful things can
3526          * happen.
3527          *
3528          * XXX illustrate
3529          */
3530         se->avg.last_update_time = cfs_rq->avg.last_update_time;
3531         se->avg.period_contrib = cfs_rq->avg.period_contrib;
3532 
3533         /*
3534          * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3535          * period_contrib. This isn't strictly correct, but since we're
3536          * entirely outside of the PELT hierarchy, nobody cares if we truncate
3537          * _sum a little.
3538          */
3539         se->avg.util_sum = se->avg.util_avg * divider;
3540 
3541         se->avg.load_sum = divider;
3542         if (se_weight(se)) {
3543                 se->avg.load_sum =
3544                         div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3545         }
3546 
3547         se->avg.runnable_load_sum = se->avg.load_sum;
3548 
3549         enqueue_load_avg(cfs_rq, se);
3550         cfs_rq->avg.util_avg += se->avg.util_avg;
3551         cfs_rq->avg.util_sum += se->avg.util_sum;
3552 
3553         add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3554 
3555         cfs_rq_util_change(cfs_rq, flags);
3556 
3557         trace_pelt_cfs_tp(cfs_rq);
3558 }
3559 
3560 /**
3561  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3562  * @cfs_rq: cfs_rq to detach from
3563  * @se: sched_entity to detach
3564  *
3565  * Must call update_cfs_rq_load_avg() before this, since we rely on
3566  * cfs_rq->avg.last_update_time being current.
3567  */
3568 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3569 {
3570         dequeue_load_avg(cfs_rq, se);
3571         sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3572         sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3573 
3574         add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3575 
3576         cfs_rq_util_change(cfs_rq, 0);
3577 
3578         trace_pelt_cfs_tp(cfs_rq);
3579 }
3580 
3581 /*
3582  * Optional action to be done while updating the load average
3583  */
3584 #define UPDATE_TG       0x1
3585 #define SKIP_AGE_LOAD   0x2
3586 #define DO_ATTACH       0x4
3587 
3588 /* Update task and its cfs_rq load average */
3589 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3590 {
3591         u64 now = cfs_rq_clock_pelt(cfs_rq);
3592         int decayed;
3593 
3594         /*
3595          * Track task load average for carrying it to new CPU after migrated, and
3596          * track group sched_entity load average for task_h_load calc in migration
3597          */
3598         if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3599                 __update_load_avg_se(now, cfs_rq, se);
3600 
3601         decayed  = update_cfs_rq_load_avg(now, cfs_rq);
3602         decayed |= propagate_entity_load_avg(se);
3603 
3604         if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3605 
3606                 /*
3607                  * DO_ATTACH means we're here from enqueue_entity().
3608                  * !last_update_time means we've passed through
3609                  * migrate_task_rq_fair() indicating we migrated.
3610                  *
3611                  * IOW we're enqueueing a task on a new CPU.
3612                  */
3613                 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3614                 update_tg_load_avg(cfs_rq, 0);
3615 
3616         } else if (decayed) {
3617                 cfs_rq_util_change(cfs_rq, 0);
3618 
3619                 if (flags & UPDATE_TG)
3620                         update_tg_load_avg(cfs_rq, 0);
3621         }
3622 }
3623 
3624 #ifndef CONFIG_64BIT
3625 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3626 {
3627         u64 last_update_time_copy;
3628         u64 last_update_time;
3629 
3630         do {
3631                 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3632                 smp_rmb();
3633                 last_update_time = cfs_rq->avg.last_update_time;
3634         } while (last_update_time != last_update_time_copy);
3635 
3636         return last_update_time;
3637 }
3638 #else
3639 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3640 {
3641         return cfs_rq->avg.last_update_time;
3642 }
3643 #endif
3644 
3645 /*
3646  * Synchronize entity load avg of dequeued entity without locking
3647  * the previous rq.
3648  */
3649 static void sync_entity_load_avg(struct sched_entity *se)
3650 {
3651         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3652         u64 last_update_time;
3653 
3654         last_update_time = cfs_rq_last_update_time(cfs_rq);
3655         __update_load_avg_blocked_se(last_update_time, se);
3656 }
3657 
3658 /*
3659  * Task first catches up with cfs_rq, and then subtract
3660  * itself from the cfs_rq (task must be off the queue now).
3661  */
3662 static void remove_entity_load_avg(struct sched_entity *se)
3663 {
3664         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3665         unsigned long flags;
3666 
3667         /*
3668          * tasks cannot exit without having gone through wake_up_new_task() ->
3669          * post_init_entity_util_avg() which will have added things to the
3670          * cfs_rq, so we can remove unconditionally.
3671          */
3672 
3673         sync_entity_load_avg(se);
3674 
3675         raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3676         ++cfs_rq->removed.nr;
3677         cfs_rq->removed.util_avg        += se->avg.util_avg;
3678         cfs_rq->removed.load_avg        += se->avg.load_avg;
3679         cfs_rq->removed.runnable_sum    += se->avg.load_sum; /* == runnable_sum */
3680         raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3681 }
3682 
3683 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3684 {
3685         return cfs_rq->avg.runnable_load_avg;
3686 }
3687 
3688 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3689 {
3690         return cfs_rq->avg.load_avg;
3691 }
3692 
3693 static inline unsigned long task_util(struct task_struct *p)
3694 {
3695         return READ_ONCE(p->se.avg.util_avg);
3696 }
3697 
3698 static inline unsigned long _task_util_est(struct task_struct *p)
3699 {
3700         struct util_est ue = READ_ONCE(p->se.avg.util_est);
3701 
3702         return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3703 }
3704 
3705 static inline unsigned long task_util_est(struct task_struct *p)
3706 {
3707         return max(task_util(p), _task_util_est(p));
3708 }
3709 
3710 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3711                                     struct task_struct *p)
3712 {
3713         unsigned int enqueued;
3714 
3715         if (!sched_feat(UTIL_EST))
3716                 return;
3717 
3718         /* Update root cfs_rq's estimated utilization */
3719         enqueued  = cfs_rq->avg.util_est.enqueued;
3720         enqueued += _task_util_est(p);
3721         WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3722 }
3723 
3724 /*
3725  * Check if a (signed) value is within a specified (unsigned) margin,
3726  * based on the observation that:
3727  *
3728  *     abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3729  *
3730  * NOTE: this only works when value + maring < INT_MAX.
3731  */
3732 static inline bool within_margin(int value, int margin)
3733 {
3734         return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3735 }
3736 
3737 static void
3738 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3739 {
3740         long last_ewma_diff;
3741         struct util_est ue;
3742         int cpu;
3743 
3744         if (!sched_feat(UTIL_EST))
3745                 return;
3746 
3747         /* Update root cfs_rq's estimated utilization */
3748         ue.enqueued  = cfs_rq->avg.util_est.enqueued;
3749         ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3750         WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3751 
3752         /*
3753          * Skip update of task's estimated utilization when the task has not
3754          * yet completed an activation, e.g. being migrated.
3755          */
3756         if (!task_sleep)
3757                 return;
3758 
3759         /*
3760          * If the PELT values haven't changed since enqueue time,
3761          * skip the util_est update.
3762          */
3763         ue = p->se.avg.util_est;
3764         if (ue.enqueued & UTIL_AVG_UNCHANGED)
3765                 return;
3766 
3767         /*
3768          * Skip update of task's estimated utilization when its EWMA is
3769          * already ~1% close to its last activation value.
3770          */
3771         ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3772         last_ewma_diff = ue.enqueued - ue.ewma;
3773         if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3774                 return;
3775 
3776         /*
3777          * To avoid overestimation of actual task utilization, skip updates if
3778          * we cannot grant there is idle time in this CPU.
3779          */
3780         cpu = cpu_of(rq_of(cfs_rq));
3781         if (task_util(p) > capacity_orig_of(cpu))
3782                 return;
3783 
3784         /*
3785          * Update Task's estimated utilization
3786          *
3787          * When *p completes an activation we can consolidate another sample
3788          * of the task size. This is done by storing the current PELT value
3789          * as ue.enqueued and by using this value to update the Exponential
3790          * Weighted Moving Average (EWMA):
3791          *
3792          *  ewma(t) = w *  task_util(p) + (1-w) * ewma(t-1)
3793          *          = w *  task_util(p) +         ewma(t-1)  - w * ewma(t-1)
3794          *          = w * (task_util(p) -         ewma(t-1)) +     ewma(t-1)
3795          *          = w * (      last_ewma_diff            ) +     ewma(t-1)
3796          *          = w * (last_ewma_diff  +  ewma(t-1) / w)
3797          *
3798          * Where 'w' is the weight of new samples, which is configured to be
3799          * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3800          */
3801         ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3802         ue.ewma  += last_ewma_diff;
3803         ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3804         WRITE_ONCE(p->se.avg.util_est, ue);
3805 }
3806 
3807 static inline int task_fits_capacity(struct task_struct *p, long capacity)
3808 {
3809         return fits_capacity(task_util_est(p), capacity);
3810 }
3811 
3812 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3813 {
3814         if (!static_branch_unlikely(&sched_asym_cpucapacity))
3815                 return;
3816 
3817         if (!p) {
3818                 rq->misfit_task_load = 0;
3819                 return;
3820         }
3821 
3822         if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3823                 rq->misfit_task_load = 0;
3824                 return;
3825         }
3826 
3827         rq->misfit_task_load = task_h_load(p);
3828 }
3829 
3830 #else /* CONFIG_SMP */
3831 
3832 #define UPDATE_TG       0x0
3833 #define SKIP_AGE_LOAD   0x0
3834 #define DO_ATTACH       0x0
3835 
3836 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3837 {
3838         cfs_rq_util_change(cfs_rq, 0);
3839 }
3840 
3841 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3842 
3843 static inline void
3844 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3845 static inline void
3846 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3847 
3848 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3849 {
3850         return 0;
3851 }
3852 
3853 static inline void
3854 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3855 
3856 static inline void
3857 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3858                  bool task_sleep) {}
3859 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3860 
3861 #endif /* CONFIG_SMP */
3862 
3863 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3864 {
3865 #ifdef CONFIG_SCHED_DEBUG
3866         s64 d = se->vruntime - cfs_rq->min_vruntime;
3867 
3868         if (d < 0)
3869                 d = -d;
3870 
3871         if (d > 3*sysctl_sched_latency)
3872                 schedstat_inc(cfs_rq->nr_spread_over);
3873 #endif
3874 }
3875 
3876 static void
3877 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3878 {
3879         u64 vruntime = cfs_rq->min_vruntime;
3880 
3881         /*
3882          * The 'current' period is already promised to the current tasks,
3883          * however the extra weight of the new task will slow them down a
3884          * little, place the new task so that it fits in the slot that
3885          * stays open at the end.
3886          */
3887         if (initial && sched_feat(START_DEBIT))
3888                 vruntime += sched_vslice(cfs_rq, se);
3889 
3890         /* sleeps up to a single latency don't count. */
3891         if (!initial) {
3892                 unsigned long thresh = sysctl_sched_latency;
3893 
3894                 /*
3895                  * Halve their sleep time's effect, to allow
3896                  * for a gentler effect of sleepers:
3897                  */
3898                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3899                         thresh >>= 1;
3900 
3901                 vruntime -= thresh;
3902         }
3903 
3904         /* ensure we never gain time by being placed backwards. */
3905         se->vruntime = max_vruntime(se->vruntime, vruntime);
3906 }
3907 
3908 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3909 
3910 static inline void check_schedstat_required(void)
3911 {
3912 #ifdef CONFIG_SCHEDSTATS
3913         if (schedstat_enabled())
3914                 return;
3915 
3916         /* Force schedstat enabled if a dependent tracepoint is active */
3917         if (trace_sched_stat_wait_enabled()    ||
3918                         trace_sched_stat_sleep_enabled()   ||
3919                         trace_sched_stat_iowait_enabled()  ||
3920                         trace_sched_stat_blocked_enabled() ||
3921                         trace_sched_stat_runtime_enabled())  {
3922                 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3923                              "stat_blocked and stat_runtime require the "
3924                              "kernel parameter schedstats=enable or "
3925                              "kernel.sched_schedstats=1\n");
3926         }
3927 #endif
3928 }
3929 
3930 static inline bool cfs_bandwidth_used(void);
3931 
3932 /*
3933  * MIGRATION
3934  *
3935  *      dequeue
3936  *        update_curr()
3937  *          update_min_vruntime()
3938  *        vruntime -= min_vruntime
3939  *
3940  *      enqueue
3941  *        update_curr()
3942  *          update_min_vruntime()
3943  *        vruntime += min_vruntime
3944  *
3945  * this way the vruntime transition between RQs is done when both
3946  * min_vruntime are up-to-date.
3947  *
3948  * WAKEUP (remote)
3949  *
3950  *      ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3951  *        vruntime -= min_vruntime
3952  *
3953  *      enqueue
3954  *        update_curr()
3955  *          update_min_vruntime()
3956  *        vruntime += min_vruntime
3957  *
3958  * this way we don't have the most up-to-date min_vruntime on the originating
3959  * CPU and an up-to-date min_vruntime on the destination CPU.
3960  */
3961 
3962 static void
3963 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3964 {
3965         bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3966         bool curr = cfs_rq->curr == se;
3967 
3968         /*
3969          * If we're the current task, we must renormalise before calling
3970          * update_curr().
3971          */
3972         if (renorm && curr)
3973                 se->vruntime += cfs_rq->min_vruntime;
3974 
3975         update_curr(cfs_rq);
3976 
3977         /*
3978          * Otherwise, renormalise after, such that we're placed at the current
3979          * moment in time, instead of some random moment in the past. Being
3980          * placed in the past could significantly boost this task to the
3981          * fairness detriment of existing tasks.
3982          */
3983         if (renorm && !curr)
3984                 se->vruntime += cfs_rq->min_vruntime;
3985 
3986         /*
3987          * When enqueuing a sched_entity, we must:
3988          *   - Update loads to have both entity and cfs_rq synced with now.
3989          *   - Add its load to cfs_rq->runnable_avg
3990          *   - For group_entity, update its weight to reflect the new share of
3991          *     its group cfs_rq
3992          *   - Add its new weight to cfs_rq->load.weight
3993          */
3994         update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3995         update_cfs_group(se);
3996         enqueue_runnable_load_avg(cfs_rq, se);
3997         account_entity_enqueue(cfs_rq, se);
3998 
3999         if (flags & ENQUEUE_WAKEUP)
4000                 place_entity(cfs_rq, se, 0);
4001 
4002         check_schedstat_required();
4003         update_stats_enqueue(cfs_rq, se, flags);
4004         check_spread(cfs_rq, se);
4005         if (!curr)
4006                 __enqueue_entity(cfs_rq, se);
4007         se->on_rq = 1;
4008 
4009         /*
4010          * When bandwidth control is enabled, cfs might have been removed
4011          * because of a parent been throttled but cfs->nr_running > 1. Try to
4012          * add it unconditionnally.
4013          */
4014         if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
4015                 list_add_leaf_cfs_rq(cfs_rq);
4016 
4017         if (cfs_rq->nr_running == 1)
4018                 check_enqueue_throttle(cfs_rq);
4019 }
4020 
4021 static void __clear_buddies_last(struct sched_entity *se)
4022 {
4023         for_each_sched_entity(se) {
4024                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4025                 if (cfs_rq->last != se)
4026                         break;
4027 
4028                 cfs_rq->last = NULL;
4029         }
4030 }
4031 
4032 static void __clear_buddies_next(struct sched_entity *se)
4033 {
4034         for_each_sched_entity(se) {
4035                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4036                 if (cfs_rq->next != se)
4037                         break;
4038 
4039                 cfs_rq->next = NULL;
4040         }
4041 }
4042 
4043 static void __clear_buddies_skip(struct sched_entity *se)
4044 {
4045         for_each_sched_entity(se) {
4046                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4047                 if (cfs_rq->skip != se)
4048                         break;
4049 
4050                 cfs_rq->skip = NULL;
4051         }
4052 }
4053 
4054 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4055 {
4056         if (cfs_rq->last == se)
4057                 __clear_buddies_last(se);
4058 
4059         if (cfs_rq->next == se)
4060                 __clear_buddies_next(se);
4061 
4062         if (cfs_rq->skip == se)
4063                 __clear_buddies_skip(se);
4064 }
4065 
4066 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4067 
4068 static void
4069 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4070 {
4071         /*
4072          * Update run-time statistics of the 'current'.
4073          */
4074         update_curr(cfs_rq);
4075 
4076         /*
4077          * When dequeuing a sched_entity, we must:
4078          *   - Update loads to have both entity and cfs_rq synced with now.
4079          *   - Subtract its load from the cfs_rq->runnable_avg.
4080          *   - Subtract its previous weight from cfs_rq->load.weight.
4081          *   - For group entity, update its weight to reflect the new share
4082          *     of its group cfs_rq.
4083          */
4084         update_load_avg(cfs_rq, se, UPDATE_TG);
4085         dequeue_runnable_load_avg(cfs_rq, se);
4086 
4087         update_stats_dequeue(cfs_rq, se, flags);
4088 
4089         clear_buddies(cfs_rq, se);
4090 
4091         if (se != cfs_rq->curr)
4092                 __dequeue_entity(cfs_rq, se);
4093         se->on_rq = 0;
4094         account_entity_dequeue(cfs_rq, se);
4095 
4096         /*
4097          * Normalize after update_curr(); which will also have moved
4098          * min_vruntime if @se is the one holding it back. But before doing
4099          * update_min_vruntime() again, which will discount @se's position and
4100          * can move min_vruntime forward still more.
4101          */
4102         if (!(flags & DEQUEUE_SLEEP))
4103                 se->vruntime -= cfs_rq->min_vruntime;
4104 
4105         /* return excess runtime on last dequeue */
4106         return_cfs_rq_runtime(cfs_rq);
4107 
4108         update_cfs_group(se);
4109 
4110         /*
4111          * Now advance min_vruntime if @se was the entity holding it back,
4112          * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4113          * put back on, and if we advance min_vruntime, we'll be placed back
4114          * further than we started -- ie. we'll be penalized.
4115          */
4116         if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4117                 update_min_vruntime(cfs_rq);
4118 }
4119 
4120 /*
4121  * Preempt the current task with a newly woken task if needed:
4122  */
4123 static void
4124 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4125 {
4126         unsigned long ideal_runtime, delta_exec;
4127         struct sched_entity *se;
4128         s64 delta;
4129 
4130         ideal_runtime = sched_slice(cfs_rq, curr);
4131         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4132         if (delta_exec > ideal_runtime) {
4133                 resched_curr(rq_of(cfs_rq));
4134                 /*
4135                  * The current task ran long enough, ensure it doesn't get
4136                  * re-elected due to buddy favours.
4137                  */
4138                 clear_buddies(cfs_rq, curr);
4139                 return;
4140         }
4141 
4142         /*
4143          * Ensure that a task that missed wakeup preemption by a
4144          * narrow margin doesn't have to wait for a full slice.
4145          * This also mitigates buddy induced latencies under load.
4146          */
4147         if (delta_exec < sysctl_sched_min_granularity)
4148                 return;
4149 
4150         se = __pick_first_entity(cfs_rq);
4151         delta = curr->vruntime - se->vruntime;
4152 
4153         if (delta < 0)
4154                 return;
4155 
4156         if (delta > ideal_runtime)
4157                 resched_curr(rq_of(cfs_rq));
4158 }
4159 
4160 static void
4161 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4162 {
4163         /* 'current' is not kept within the tree. */
4164         if (se->on_rq) {
4165                 /*
4166                  * Any task has to be enqueued before it get to execute on
4167                  * a CPU. So account for the time it spent waiting on the
4168                  * runqueue.
4169                  */
4170                 update_stats_wait_end(cfs_rq, se);
4171                 __dequeue_entity(cfs_rq, se);
4172                 update_load_avg(cfs_rq, se, UPDATE_TG);
4173         }
4174 
4175         update_stats_curr_start(cfs_rq, se);
4176         cfs_rq->curr = se;
4177 
4178         /*
4179          * Track our maximum slice length, if the CPU's load is at
4180          * least twice that of our own weight (i.e. dont track it
4181          * when there are only lesser-weight tasks around):
4182          */
4183         if (schedstat_enabled() &&
4184             rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4185                 schedstat_set(se->statistics.slice_max,
4186                         max((u64)schedstat_val(se->statistics.slice_max),
4187                             se->sum_exec_runtime - se->prev_sum_exec_runtime));
4188         }
4189 
4190         se->prev_sum_exec_runtime = se->sum_exec_runtime;
4191 }
4192 
4193 static int
4194 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4195 
4196 /*
4197  * Pick the next process, keeping these things in mind, in this order:
4198  * 1) keep things fair between processes/task groups
4199  * 2) pick the "next" process, since someone really wants that to run
4200  * 3) pick the "last" process, for cache locality
4201  * 4) do not run the "skip" process, if something else is available
4202  */
4203 static struct sched_entity *
4204 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4205 {
4206         struct sched_entity *left = __pick_first_entity(cfs_rq);
4207         struct sched_entity *se;
4208 
4209         /*
4210          * If curr is set we have to see if its left of the leftmost entity
4211          * still in the tree, provided there was anything in the tree at all.
4212          */
4213         if (!left || (curr && entity_before(curr, left)))
4214                 left = curr;
4215 
4216         se = left; /* ideally we run the leftmost entity */
4217 
4218         /*
4219          * Avoid running the skip buddy, if running something else can
4220          * be done without getting too unfair.
4221          */
4222         if (cfs_rq->skip == se) {
4223                 struct sched_entity *second;
4224 
4225                 if (se == curr) {
4226                         second = __pick_first_entity(cfs_rq);
4227                 } else {
4228                         second = __pick_next_entity(se);
4229                         if (!second || (curr && entity_before(curr, second)))
4230                                 second = curr;
4231                 }
4232 
4233                 if (second && wakeup_preempt_entity(second, left) < 1)
4234                         se = second;
4235         }
4236 
4237         /*
4238          * Prefer last buddy, try to return the CPU to a preempted task.
4239          */
4240         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4241                 se = cfs_rq->last;
4242 
4243         /*
4244          * Someone really wants this to run. If it's not unfair, run it.
4245          */
4246         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4247                 se = cfs_rq->next;
4248 
4249         clear_buddies(cfs_rq, se);
4250 
4251         return se;
4252 }
4253 
4254 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4255 
4256 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4257 {
4258         /*
4259          * If still on the runqueue then deactivate_task()
4260          * was not called and update_curr() has to be done:
4261          */
4262         if (prev->on_rq)
4263                 update_curr(cfs_rq);
4264 
4265         /* throttle cfs_rqs exceeding runtime */
4266         check_cfs_rq_runtime(cfs_rq);
4267 
4268         check_spread(cfs_rq, prev);
4269 
4270         if (prev->on_rq) {
4271                 update_stats_wait_start(cfs_rq, prev);
4272                 /* Put 'current' back into the tree. */
4273                 __enqueue_entity(cfs_rq, prev);
4274                 /* in !on_rq case, update occurred at dequeue */
4275                 update_load_avg(cfs_rq, prev, 0);
4276         }
4277         cfs_rq->curr = NULL;
4278 }
4279 
4280 static void
4281 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4282 {
4283         /*
4284          * Update run-time statistics of the 'current'.
4285          */
4286         update_curr(cfs_rq);
4287 
4288         /*
4289          * Ensure that runnable average is periodically updated.
4290          */
4291         update_load_avg(cfs_rq, curr, UPDATE_TG);
4292         update_cfs_group(curr);
4293 
4294 #ifdef CONFIG_SCHED_HRTICK
4295         /*
4296          * queued ticks are scheduled to match the slice, so don't bother
4297          * validating it and just reschedule.
4298          */
4299         if (queued) {
4300                 resched_curr(rq_of(cfs_rq));
4301                 return;
4302         }
4303         /*
4304          * don't let the period tick interfere with the hrtick preemption
4305          */
4306         if (!sched_feat(DOUBLE_TICK) &&
4307                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4308                 return;
4309 #endif
4310 
4311         if (cfs_rq->nr_running > 1)
4312                 check_preempt_tick(cfs_rq, curr);
4313 }
4314 
4315 
4316 /**************************************************
4317  * CFS bandwidth control machinery
4318  */
4319 
4320 #ifdef CONFIG_CFS_BANDWIDTH
4321 
4322 #ifdef CONFIG_JUMP_LABEL
4323 static struct static_key __cfs_bandwidth_used;
4324 
4325 static inline bool cfs_bandwidth_used(void)
4326 {
4327         return static_key_false(&__cfs_bandwidth_used);
4328 }
4329 
4330 void cfs_bandwidth_usage_inc(void)
4331 {
4332         static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4333 }
4334 
4335 void cfs_bandwidth_usage_dec(void)
4336 {
4337         static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4338 }
4339 #else /* CONFIG_JUMP_LABEL */
4340 static bool cfs_bandwidth_used(void)
4341 {
4342         return true;
4343 }
4344 
4345 void cfs_bandwidth_usage_inc(void) {}
4346 void cfs_bandwidth_usage_dec(void) {}
4347 #endif /* CONFIG_JUMP_LABEL */
4348 
4349 /*
4350  * default period for cfs group bandwidth.
4351  * default: 0.1s, units: nanoseconds
4352  */
4353 static inline u64 default_cfs_period(void)
4354 {
4355         return 100000000ULL;
4356 }
4357 
4358 static inline u64 sched_cfs_bandwidth_slice(void)
4359 {
4360         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4361 }
4362 
4363 /*
4364  * Replenish runtime according to assigned quota. We use sched_clock_cpu
4365  * directly instead of rq->clock to avoid adding additional synchronization
4366  * around rq->lock.
4367  *
4368  * requires cfs_b->lock
4369  */
4370 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4371 {
4372         if (cfs_b->quota != RUNTIME_INF)
4373                 cfs_b->runtime = cfs_b->quota;
4374 }
4375 
4376 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4377 {
4378         return &tg->cfs_bandwidth;
4379 }
4380 
4381 /* returns 0 on failure to allocate runtime */
4382 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4383 {
4384         struct task_group *tg = cfs_rq->tg;
4385         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4386         u64 amount = 0, min_amount;
4387 
4388         /* note: this is a positive sum as runtime_remaining <= 0 */
4389         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4390 
4391         raw_spin_lock(&cfs_b->lock);
4392         if (cfs_b->quota == RUNTIME_INF)
4393                 amount = min_amount;
4394         else {
4395                 start_cfs_bandwidth(cfs_b);
4396 
4397                 if (cfs_b->runtime > 0) {
4398                         amount = min(cfs_b->runtime, min_amount);
4399                         cfs_b->runtime -= amount;
4400                         cfs_b->idle = 0;
4401                 }
4402         }
4403         raw_spin_unlock(&cfs_b->lock);
4404 
4405         cfs_rq->runtime_remaining += amount;
4406 
4407         return cfs_rq->runtime_remaining > 0;
4408 }
4409 
4410 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4411 {
4412         /* dock delta_exec before expiring quota (as it could span periods) */
4413         cfs_rq->runtime_remaining -= delta_exec;
4414 
4415         if (likely(cfs_rq->runtime_remaining > 0))
4416                 return;
4417 
4418         if (cfs_rq->throttled)
4419                 return;
4420         /*
4421          * if we're unable to extend our runtime we resched so that the active
4422          * hierarchy can be throttled
4423          */
4424         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4425                 resched_curr(rq_of(cfs_rq));
4426 }
4427 
4428 static __always_inline
4429 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4430 {
4431         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4432                 return;
4433 
4434         __account_cfs_rq_runtime(cfs_rq, delta_exec);
4435 }
4436 
4437 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4438 {
4439         return cfs_bandwidth_used() && cfs_rq->throttled;
4440 }
4441 
4442 /* check whether cfs_rq, or any parent, is throttled */
4443 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4444 {
4445         return cfs_bandwidth_used() && cfs_rq->throttle_count;
4446 }
4447 
4448 /*
4449  * Ensure that neither of the group entities corresponding to src_cpu or
4450  * dest_cpu are members of a throttled hierarchy when performing group
4451  * load-balance operations.
4452  */
4453 static inline int throttled_lb_pair(struct task_group *tg,
4454                                     int src_cpu, int dest_cpu)
4455 {
4456         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4457 
4458         src_cfs_rq = tg->cfs_rq[src_cpu];
4459         dest_cfs_rq = tg->cfs_rq[dest_cpu];
4460 
4461         return throttled_hierarchy(src_cfs_rq) ||
4462                throttled_hierarchy(dest_cfs_rq);
4463 }
4464 
4465 static int tg_unthrottle_up(struct task_group *tg, void *data)
4466 {
4467         struct rq *rq = data;
4468         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4469 
4470         cfs_rq->throttle_count--;
4471         if (!cfs_rq->throttle_count) {
4472                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4473                                              cfs_rq->throttled_clock_task;
4474 
4475                 /* Add cfs_rq with already running entity in the list */
4476                 if (cfs_rq->nr_running >= 1)
4477                         list_add_leaf_cfs_rq(cfs_rq);
4478         }
4479 
4480         return 0;
4481 }
4482 
4483 static int tg_throttle_down(struct task_group *tg, void *data)
4484 {
4485         struct rq *rq = data;
4486         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4487 
4488         /* group is entering throttled state, stop time */
4489         if (!cfs_rq->throttle_count) {
4490                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4491                 list_del_leaf_cfs_rq(cfs_rq);
4492         }
4493         cfs_rq->throttle_count++;
4494 
4495         return 0;
4496 }
4497 
4498 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4499 {
4500         struct rq *rq = rq_of(cfs_rq);
4501         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4502         struct sched_entity *se;
4503         long task_delta, idle_task_delta, dequeue = 1;
4504         bool empty;
4505 
4506         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4507 
4508         /* freeze hierarchy runnable averages while throttled */
4509         rcu_read_lock();
4510         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4511         rcu_read_unlock();
4512 
4513         task_delta = cfs_rq->h_nr_running;
4514         idle_task_delta = cfs_rq->idle_h_nr_running;
4515         for_each_sched_entity(se) {
4516                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4517                 /* throttled entity or throttle-on-deactivate */
4518                 if (!se->on_rq)
4519                         break;
4520 
4521                 if (dequeue)
4522                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4523                 qcfs_rq->h_nr_running -= task_delta;
4524                 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4525 
4526                 if (qcfs_rq->load.weight)
4527                         dequeue = 0;
4528         }
4529 
4530         if (!se)
4531                 sub_nr_running(rq, task_delta);
4532 
4533         cfs_rq->throttled = 1;
4534         cfs_rq->throttled_clock = rq_clock(rq);
4535         raw_spin_lock(&cfs_b->lock);
4536         empty = list_empty(&cfs_b->throttled_cfs_rq);
4537 
4538         /*
4539          * Add to the _head_ of the list, so that an already-started
4540          * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4541          * not running add to the tail so that later runqueues don't get starved.
4542          */
4543         if (cfs_b->distribute_running)
4544                 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4545         else
4546                 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4547 
4548         /*
4549          * If we're the first throttled task, make sure the bandwidth
4550          * timer is running.
4551          */
4552         if (empty)
4553                 start_cfs_bandwidth(cfs_b);
4554 
4555         raw_spin_unlock(&cfs_b->lock);
4556 }
4557 
4558 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4559 {
4560         struct rq *rq = rq_of(cfs_rq);
4561         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4562         struct sched_entity *se;
4563         int enqueue = 1;
4564         long task_delta, idle_task_delta;
4565 
4566         se = cfs_rq->tg->se[cpu_of(rq)];
4567 
4568         cfs_rq->throttled = 0;
4569 
4570         update_rq_clock(rq);
4571 
4572         raw_spin_lock(&cfs_b->lock);
4573         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4574         list_del_rcu(&cfs_rq->throttled_list);
4575         raw_spin_unlock(&cfs_b->lock);
4576 
4577         /* update hierarchical throttle state */
4578         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4579 
4580         if (!cfs_rq->load.weight)
4581                 return;
4582 
4583         task_delta = cfs_rq->h_nr_running;
4584         idle_task_delta = cfs_rq->idle_h_nr_running;
4585         for_each_sched_entity(se) {
4586                 if (se->on_rq)
4587                         enqueue = 0;
4588 
4589                 cfs_rq = cfs_rq_of(se);
4590                 if (enqueue)
4591                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4592                 cfs_rq->h_nr_running += task_delta;
4593                 cfs_rq->idle_h_nr_running += idle_task_delta;
4594 
4595                 if (cfs_rq_throttled(cfs_rq))
4596                         break;
4597         }
4598 
4599         if (!se)
4600                 add_nr_running(rq, task_delta);
4601 
4602         /*
4603          * The cfs_rq_throttled() breaks in the above iteration can result in
4604          * incomplete leaf list maintenance, resulting in triggering the
4605          * assertion below.
4606          */
4607         for_each_sched_entity(se) {
4608                 cfs_rq = cfs_rq_of(se);
4609 
4610                 list_add_leaf_cfs_rq(cfs_rq);
4611         }
4612 
4613         assert_list_leaf_cfs_rq(rq);
4614 
4615         /* Determine whether we need to wake up potentially idle CPU: */
4616         if (rq->curr == rq->idle && rq->cfs.nr_running)
4617                 resched_curr(rq);
4618 }
4619 
4620 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4621 {
4622         struct cfs_rq *cfs_rq;
4623         u64 runtime;
4624         u64 starting_runtime = remaining;
4625 
4626         rcu_read_lock();
4627         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4628                                 throttled_list) {
4629                 struct rq *rq = rq_of(cfs_rq);
4630                 struct rq_flags rf;
4631 
4632                 rq_lock_irqsave(rq, &rf);
4633                 if (!cfs_rq_throttled(cfs_rq))
4634                         goto next;
4635 
4636                 /* By the above check, this should never be true */
4637                 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4638 
4639                 runtime = -cfs_rq->runtime_remaining + 1;
4640                 if (runtime > remaining)
4641                         runtime = remaining;
4642                 remaining -= runtime;
4643 
4644                 cfs_rq->runtime_remaining += runtime;
4645 
4646                 /* we check whether we're throttled above */
4647                 if (cfs_rq->runtime_remaining > 0)
4648                         unthrottle_cfs_rq(cfs_rq);
4649 
4650 next:
4651                 rq_unlock_irqrestore(rq, &rf);
4652 
4653                 if (!remaining)
4654                         break;
4655         }
4656         rcu_read_unlock();
4657 
4658         return starting_runtime - remaining;
4659 }
4660 
4661 /*
4662  * Responsible for refilling a task_group's bandwidth and unthrottling its
4663  * cfs_rqs as appropriate. If there has been no activity within the last
4664  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4665  * used to track this state.
4666  */
4667 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4668 {
4669         u64 runtime;
4670         int throttled;
4671 
4672         /* no need to continue the timer with no bandwidth constraint */
4673         if (cfs_b->quota == RUNTIME_INF)
4674                 goto out_deactivate;
4675 
4676         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4677         cfs_b->nr_periods += overrun;
4678 
4679         /*
4680          * idle depends on !throttled (for the case of a large deficit), and if
4681          * we're going inactive then everything else can be deferred
4682          */
4683         if (cfs_b->idle && !throttled)
4684                 goto out_deactivate;
4685 
4686         __refill_cfs_bandwidth_runtime(cfs_b);
4687 
4688         if (!throttled) {
4689                 /* mark as potentially idle for the upcoming period */
4690                 cfs_b->idle = 1;
4691                 return 0;
4692         }
4693 
4694         /* account preceding periods in which throttling occurred */
4695         cfs_b->nr_throttled += overrun;
4696 
4697         /*
4698          * This check is repeated as we are holding onto the new bandwidth while
4699          * we unthrottle. This can potentially race with an unthrottled group
4700          * trying to acquire new bandwidth from the global pool. This can result
4701          * in us over-using our runtime if it is all used during this loop, but
4702          * only by limited amounts in that extreme case.
4703          */
4704         while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4705                 runtime = cfs_b->runtime;
4706                 cfs_b->distribute_running = 1;
4707                 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4708                 /* we can't nest cfs_b->lock while distributing bandwidth */
4709                 runtime = distribute_cfs_runtime(cfs_b, runtime);
4710                 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4711 
4712                 cfs_b->distribute_running = 0;
4713                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4714 
4715                 lsub_positive(&cfs_b->runtime, runtime);
4716         }
4717 
4718         /*
4719          * While we are ensured activity in the period following an
4720          * unthrottle, this also covers the case in which the new bandwidth is
4721          * insufficient to cover the existing bandwidth deficit.  (Forcing the
4722          * timer to remain active while there are any throttled entities.)
4723          */
4724         cfs_b->idle = 0;
4725 
4726         return 0;
4727 
4728 out_deactivate:
4729         return 1;
4730 }
4731 
4732 /* a cfs_rq won't donate quota below this amount */
4733 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4734 /* minimum remaining period time to redistribute slack quota */
4735 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4736 /* how long we wait to gather additional slack before distributing */
4737 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4738 
4739 /*
4740  * Are we near the end of the current quota period?
4741  *
4742  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4743  * hrtimer base being cleared by hrtimer_start. In the case of
4744  * migrate_hrtimers, base is never cleared, so we are fine.
4745  */
4746 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4747 {
4748         struct hrtimer *refresh_timer = &cfs_b->period_timer;
4749         u64 remaining;
4750 
4751         /* if the call-back is running a quota refresh is already occurring */
4752         if (hrtimer_callback_running(refresh_timer))
4753                 return 1;
4754 
4755         /* is a quota refresh about to occur? */
4756         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4757         if (remaining < min_expire)
4758                 return 1;
4759 
4760         return 0;
4761 }
4762 
4763 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4764 {
4765         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4766 
4767         /* if there's a quota refresh soon don't bother with slack */
4768         if (runtime_refresh_within(cfs_b, min_left))
4769                 return;
4770 
4771         /* don't push forwards an existing deferred unthrottle */
4772         if (cfs_b->slack_started)
4773                 return;
4774         cfs_b->slack_started = true;
4775 
4776         hrtimer_start(&cfs_b->slack_timer,
4777                         ns_to_ktime(cfs_bandwidth_slack_period),
4778                         HRTIMER_MODE_REL);
4779 }
4780 
4781 /* we know any runtime found here is valid as update_curr() precedes return */
4782 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4783 {
4784         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4785         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4786 
4787         if (slack_runtime <= 0)
4788                 return;
4789 
4790         raw_spin_lock(&cfs_b->lock);
4791         if (cfs_b->quota != RUNTIME_INF) {
4792                 cfs_b->runtime += slack_runtime;
4793 
4794                 /* we are under rq->lock, defer unthrottling using a timer */
4795                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4796                     !list_empty(&cfs_b->throttled_cfs_rq))
4797                         start_cfs_slack_bandwidth(cfs_b);
4798         }
4799         raw_spin_unlock(&cfs_b->lock);
4800 
4801         /* even if it's not valid for return we don't want to try again */
4802         cfs_rq->runtime_remaining -= slack_runtime;
4803 }
4804 
4805 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4806 {
4807         if (!cfs_bandwidth_used())
4808                 return;
4809 
4810         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4811                 return;
4812 
4813         __return_cfs_rq_runtime(cfs_rq);
4814 }
4815 
4816 /*
4817  * This is done with a timer (instead of inline with bandwidth return) since
4818  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4819  */
4820 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4821 {
4822         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4823         unsigned long flags;
4824 
4825         /* confirm we're still not at a refresh boundary */
4826         raw_spin_lock_irqsave(&cfs_b->lock, flags);
4827         cfs_b->slack_started = false;
4828         if (cfs_b->distribute_running) {
4829                 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4830                 return;
4831         }
4832 
4833         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4834                 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4835                 return;
4836         }
4837 
4838         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4839                 runtime = cfs_b->runtime;
4840 
4841         if (runtime)
4842                 cfs_b->distribute_running = 1;
4843 
4844         raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4845 
4846         if (!runtime)
4847                 return;
4848 
4849         runtime = distribute_cfs_runtime(cfs_b, runtime);
4850 
4851         raw_spin_lock_irqsave(&cfs_b->lock, flags);
4852         lsub_positive(&cfs_b->runtime, runtime);
4853         cfs_b->distribute_running = 0;
4854         raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4855 }
4856 
4857 /*
4858  * When a group wakes up we want to make sure that its quota is not already
4859  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4860  * runtime as update_curr() throttling can not not trigger until it's on-rq.
4861  */
4862 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4863 {
4864         if (!cfs_bandwidth_used())
4865                 return;
4866 
4867         /* an active group must be handled by the update_curr()->put() path */
4868         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4869                 return;
4870 
4871         /* ensure the group is not already throttled */
4872         if (cfs_rq_throttled(cfs_rq))
4873                 return;
4874 
4875         /* update runtime allocation */
4876         account_cfs_rq_runtime(cfs_rq, 0);
4877         if (cfs_rq->runtime_remaining <= 0)
4878                 throttle_cfs_rq(cfs_rq);
4879 }
4880 
4881 static void sync_throttle(struct task_group *tg, int cpu)
4882 {
4883         struct cfs_rq *pcfs_rq, *cfs_rq;
4884 
4885         if (!cfs_bandwidth_used())
4886                 return;
4887 
4888         if (!tg->parent)
4889                 return;
4890 
4891         cfs_rq = tg->cfs_rq[cpu];
4892         pcfs_rq = tg->parent->cfs_rq[cpu];
4893 
4894         cfs_rq->throttle_count = pcfs_rq->throttle_count;
4895         cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4896 }
4897 
4898 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4899 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4900 {
4901         if (!cfs_bandwidth_used())
4902                 return false;
4903 
4904         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4905                 return false;
4906 
4907         /*
4908          * it's possible for a throttled entity to be forced into a running
4909          * state (e.g. set_curr_task), in this case we're finished.
4910          */
4911         if (cfs_rq_throttled(cfs_rq))
4912                 return true;
4913 
4914         throttle_cfs_rq(cfs_rq);
4915         return true;
4916 }
4917 
4918 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4919 {
4920         struct cfs_bandwidth *cfs_b =
4921                 container_of(timer, struct cfs_bandwidth, slack_timer);
4922 
4923         do_sched_cfs_slack_timer(cfs_b);
4924 
4925         return HRTIMER_NORESTART;
4926 }
4927 
4928 extern const u64 max_cfs_quota_period;
4929 
4930 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4931 {
4932         struct cfs_bandwidth *cfs_b =
4933                 container_of(timer, struct cfs_bandwidth, period_timer);
4934         unsigned long flags;
4935         int overrun;
4936         int idle = 0;
4937         int count = 0;
4938 
4939         raw_spin_lock_irqsave(&cfs_b->lock, flags);
4940         for (;;) {
4941                 overrun = hrtimer_forward_now(timer, cfs_b->period);
4942                 if (!overrun)
4943                         break;
4944 
4945                 if (++count > 3) {
4946                         u64 new, old = ktime_to_ns(cfs_b->period);
4947 
4948                         /*
4949                          * Grow period by a factor of 2 to avoid losing precision.
4950                          * Precision loss in the quota/period ratio can cause __cfs_schedulable
4951                          * to fail.
4952                          */
4953                         new = old * 2;
4954                         if (new < max_cfs_quota_period) {
4955                                 cfs_b->period = ns_to_ktime(new);
4956                                 cfs_b->quota *= 2;
4957 
4958                                 pr_warn_ratelimited(
4959         "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4960                                         smp_processor_id(),
4961                                         div_u64(new, NSEC_PER_USEC),
4962                                         div_u64(cfs_b->quota, NSEC_PER_USEC));
4963                         } else {
4964                                 pr_warn_ratelimited(
4965         "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4966                                         smp_processor_id(),
4967                                         div_u64(old, NSEC_PER_USEC),
4968                                         div_u64(cfs_b->quota, NSEC_PER_USEC));
4969                         }
4970 
4971                         /* reset count so we don't come right back in here */
4972                         count = 0;
4973                 }
4974 
4975                 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4976         }
4977         if (idle)
4978                 cfs_b->period_active = 0;
4979         raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4980 
4981         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4982 }
4983 
4984 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4985 {
4986         raw_spin_lock_init(&cfs_b->lock);
4987         cfs_b->runtime = 0;
4988         cfs_b->quota = RUNTIME_INF;
4989         cfs_b->period = ns_to_ktime(default_cfs_period());
4990 
4991         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4992         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4993         cfs_b->period_timer.function = sched_cfs_period_timer;
4994         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4995         cfs_b->slack_timer.function = sched_cfs_slack_timer;
4996         cfs_b->distribute_running = 0;
4997         cfs_b->slack_started = false;
4998 }
4999 
5000 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5001 {
5002         cfs_rq->runtime_enabled = 0;
5003         INIT_LIST_HEAD(&cfs_rq->throttled_list);
5004 }
5005 
5006 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5007 {
5008         lockdep_assert_held(&cfs_b->lock);
5009 
5010         if (cfs_b->period_active)
5011                 return;
5012 
5013         cfs_b->period_active = 1;
5014         hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5015         hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5016 }
5017 
5018 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5019 {
5020         /* init_cfs_bandwidth() was not called */
5021         if (!cfs_b->throttled_cfs_rq.next)
5022                 return;
5023 
5024         hrtimer_cancel(&cfs_b->period_timer);
5025         hrtimer_cancel(&cfs_b->slack_timer);
5026 }
5027 
5028 /*
5029  * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5030  *
5031  * The race is harmless, since modifying bandwidth settings of unhooked group
5032  * bits doesn't do much.
5033  */
5034 
5035 /* cpu online calback */
5036 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5037 {
5038         struct task_group *tg;
5039 
5040         lockdep_assert_held(&rq->lock);
5041 
5042         rcu_read_lock();
5043         list_for_each_entry_rcu(tg, &task_groups, list) {
5044                 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5045                 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5046 
5047                 raw_spin_lock(&cfs_b->lock);
5048                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5049                 raw_spin_unlock(&cfs_b->lock);
5050         }
5051         rcu_read_unlock();
5052 }
5053 
5054 /* cpu offline callback */
5055 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5056 {
5057         struct task_group *tg;
5058 
5059         lockdep_assert_held(&rq->lock);
5060 
5061         rcu_read_lock();
5062         list_for_each_entry_rcu(tg, &task_groups, list) {
5063                 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5064 
5065                 if (!cfs_rq->runtime_enabled)
5066                         continue;
5067 
5068                 /*
5069                  * clock_task is not advancing so we just need to make sure
5070                  * there's some valid quota amount
5071                  */
5072                 cfs_rq->runtime_remaining = 1;
5073                 /*
5074                  * Offline rq is schedulable till CPU is completely disabled
5075                  * in take_cpu_down(), so we prevent new cfs throttling here.
5076                  */
5077                 cfs_rq->runtime_enabled = 0;
5078 
5079                 if (cfs_rq_throttled(cfs_rq))
5080                         unthrottle_cfs_rq(cfs_rq);
5081         }
5082         rcu_read_unlock();
5083 }
5084 
5085 #else /* CONFIG_CFS_BANDWIDTH */
5086 
5087 static inline bool cfs_bandwidth_used(void)
5088 {
5089         return false;
5090 }
5091 
5092 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5093 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5094 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5095 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5096 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5097 
5098 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5099 {
5100         return 0;
5101 }
5102 
5103 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5104 {
5105         return 0;
5106 }
5107 
5108 static inline int throttled_lb_pair(struct task_group *tg,
5109                                     int src_cpu, int dest_cpu)
5110 {
5111         return 0;
5112 }
5113 
5114 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5115 
5116 #ifdef CONFIG_FAIR_GROUP_SCHED
5117 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5118 #endif
5119 
5120 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5121 {
5122         return NULL;
5123 }
5124 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5125 static inline void update_runtime_enabled(struct rq *rq) {}
5126 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5127 
5128 #endif /* CONFIG_CFS_BANDWIDTH */
5129 
5130 /**************************************************
5131  * CFS operations on tasks:
5132  */
5133 
5134 #ifdef CONFIG_SCHED_HRTICK
5135 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5136 {
5137         struct sched_entity *se = &p->se;
5138         struct cfs_rq *cfs_rq = cfs_rq_of(se);
5139 
5140         SCHED_WARN_ON(task_rq(p) != rq);
5141 
5142         if (rq->cfs.h_nr_running > 1) {
5143                 u64 slice = sched_slice(cfs_rq, se);
5144                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5145                 s64 delta = slice - ran;
5146 
5147                 if (delta < 0) {
5148                         if (rq->curr == p)
5149                                 resched_curr(rq);
5150                         return;
5151                 }
5152                 hrtick_start(rq, delta);
5153         }
5154 }
5155 
5156 /*
5157  * called from enqueue/dequeue and updates the hrtick when the
5158  * current task is from our class and nr_running is low enough
5159  * to matter.
5160  */
5161 static void hrtick_update(struct rq *rq)
5162 {
5163         struct task_struct *curr = rq->curr;
5164 
5165         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5166                 return;
5167 
5168         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5169                 hrtick_start_fair(rq, curr);
5170 }
5171 #else /* !CONFIG_SCHED_HRTICK */
5172 static inline void
5173 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5174 {
5175 }
5176 
5177 static inline void hrtick_update(struct rq *rq)
5178 {
5179 }
5180 #endif
5181 
5182 #ifdef CONFIG_SMP
5183 static inline unsigned long cpu_util(int cpu);
5184 
5185 static inline bool cpu_overutilized(int cpu)
5186 {
5187         return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5188 }
5189 
5190 static inline void update_overutilized_status(struct rq *rq)
5191 {
5192         if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5193                 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5194                 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5195         }
5196 }
5197 #else
5198 static inline void update_overutilized_status(struct rq *rq) { }
5199 #endif
5200 
5201 /*
5202  * The enqueue_task method is called before nr_running is
5203  * increased. Here we update the fair scheduling stats and
5204  * then put the task into the rbtree:
5205  */
5206 static void
5207 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5208 {
5209         struct cfs_rq *cfs_rq;
5210         struct sched_entity *se = &p->se;
5211         int idle_h_nr_running = task_has_idle_policy(p);
5212 
5213         /*
5214          * The code below (indirectly) updates schedutil which looks at
5215          * the cfs_rq utilization to select a frequency.
5216          * Let's add the task's estimated utilization to the cfs_rq's
5217          * estimated utilization, before we update schedutil.
5218          */
5219         util_est_enqueue(&rq->cfs, p);
5220 
5221         /*
5222          * If in_iowait is set, the code below may not trigger any cpufreq
5223          * utilization updates, so do it here explicitly with the IOWAIT flag
5224          * passed.
5225          */
5226         if (p->in_iowait)
5227                 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5228 
5229         for_each_sched_entity(se) {
5230                 if (se->on_rq)
5231                         break;
5232                 cfs_rq = cfs_rq_of(se);
5233                 enqueue_entity(cfs_rq, se, flags);
5234 
5235                 cfs_rq->h_nr_running++;
5236                 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5237 
5238                 /* end evaluation on encountering a throttled cfs_rq */
5239                 if (cfs_rq_throttled(cfs_rq))
5240                         goto enqueue_throttle;
5241 
5242                 flags = ENQUEUE_WAKEUP;
5243         }
5244 
5245         for_each_sched_entity(se) {
5246                 cfs_rq = cfs_rq_of(se);
5247 
5248                 update_load_avg(cfs_rq, se, UPDATE_TG);
5249                 update_cfs_group(se);
5250 
5251                 cfs_rq->h_nr_running++;
5252                 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5253 
5254                 /* end evaluation on encountering a throttled cfs_rq */
5255                 if (cfs_rq_throttled(cfs_rq))
5256                         goto enqueue_throttle;
5257 
5258                /*
5259                 * One parent has been throttled and cfs_rq removed from the
5260                 * list. Add it back to not break the leaf list.
5261                 */
5262                if (throttled_hierarchy(cfs_rq))
5263                        list_add_leaf_cfs_rq(cfs_rq);
5264         }
5265 
5266 enqueue_throttle:
5267         if (!se) {
5268                 add_nr_running(rq, 1);
5269                 /*
5270                  * Since new tasks are assigned an initial util_avg equal to
5271                  * half of the spare capacity of their CPU, tiny tasks have the
5272                  * ability to cross the overutilized threshold, which will
5273                  * result in the load balancer ruining all the task placement
5274                  * done by EAS. As a way to mitigate that effect, do not account
5275                  * for the first enqueue operation of new tasks during the
5276                  * overutilized flag detection.
5277                  *
5278                  * A better way of solving this problem would be to wait for
5279                  * the PELT signals of tasks to converge before taking them
5280                  * into account, but that is not straightforward to implement,
5281                  * and the following generally works well enough in practice.
5282                  */
5283                 if (flags & ENQUEUE_WAKEUP)
5284                         update_overutilized_status(rq);
5285 
5286         }
5287 
5288         if (cfs_bandwidth_used()) {
5289                 /*
5290                  * When bandwidth control is enabled; the cfs_rq_throttled()
5291                  * breaks in the above iteration can result in incomplete
5292                  * leaf list maintenance, resulting in triggering the assertion
5293                  * below.
5294                  */
5295                 for_each_sched_entity(se) {
5296                         cfs_rq = cfs_rq_of(se);
5297 
5298                         if (list_add_leaf_cfs_rq(cfs_rq))
5299                                 break;
5300                 }
5301         }
5302 
5303         assert_list_leaf_cfs_rq(rq);
5304 
5305         hrtick_update(rq);
5306 }
5307 
5308 static void set_next_buddy(struct sched_entity *se);
5309 
5310 /*
5311  * The dequeue_task method is called before nr_running is
5312  * decreased. We remove the task from the rbtree and
5313  * update the fair scheduling stats:
5314  */
5315 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5316 {
5317         struct cfs_rq *cfs_rq;
5318         struct sched_entity *se = &p->se;
5319         int task_sleep = flags & DEQUEUE_SLEEP;
5320         int idle_h_nr_running = task_has_idle_policy(p);
5321 
5322         for_each_sched_entity(se) {
5323                 cfs_rq = cfs_rq_of(se);
5324                 dequeue_entity(cfs_rq, se, flags);
5325 
5326                 cfs_rq->h_nr_running--;
5327                 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5328 
5329                 /* end evaluation on encountering a throttled cfs_rq */
5330                 if (cfs_rq_throttled(cfs_rq))
5331                         goto dequeue_throttle;
5332 
5333                 /* Don't dequeue parent if it has other entities besides us */
5334                 if (cfs_rq->load.weight) {
5335                         /* Avoid re-evaluating load for this entity: */
5336                         se = parent_entity(se);
5337                         /*
5338                          * Bias pick_next to pick a task from this cfs_rq, as
5339                          * p is sleeping when it is within its sched_slice.
5340                          */
5341                         if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5342                                 set_next_buddy(se);
5343                         break;
5344                 }
5345                 flags |= DEQUEUE_SLEEP;
5346         }
5347 
5348         for_each_sched_entity(se) {
5349                 cfs_rq = cfs_rq_of(se);
5350 
5351                 update_load_avg(cfs_rq, se, UPDATE_TG);
5352                 update_cfs_group(se);
5353 
5354                 cfs_rq->h_nr_running--;
5355                 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5356 
5357                 /* end evaluation on encountering a throttled cfs_rq */
5358                 if (cfs_rq_throttled(cfs_rq))
5359                         goto dequeue_throttle;
5360 
5361         }
5362 
5363 dequeue_throttle:
5364         if (!se)
5365                 sub_nr_running(rq, 1);
5366 
5367         util_est_dequeue(&rq->cfs, p, task_sleep);
5368         hrtick_update(rq);
5369 }
5370 
5371 #ifdef CONFIG_SMP
5372 
5373 /* Working cpumask for: load_balance, load_balance_newidle. */
5374 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5375 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5376 
5377 #ifdef CONFIG_NO_HZ_COMMON
5378 
5379 static struct {
5380         cpumask_var_t idle_cpus_mask;
5381         atomic_t nr_cpus;
5382         int has_blocked;                /* Idle CPUS has blocked load */
5383         unsigned long next_balance;     /* in jiffy units */
5384         unsigned long next_blocked;     /* Next update of blocked load in jiffies */
5385 } nohz ____cacheline_aligned;
5386 
5387 #endif /* CONFIG_NO_HZ_COMMON */
5388 
5389 /* CPU only has SCHED_IDLE tasks enqueued */
5390 static int sched_idle_cpu(int cpu)
5391 {
5392         struct rq *rq = cpu_rq(cpu);
5393 
5394         return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5395                         rq->nr_running);
5396 }
5397 
5398 static unsigned long cpu_runnable_load(struct rq *rq)
5399 {
5400         return cfs_rq_runnable_load_avg(&rq->cfs);
5401 }
5402 
5403 static unsigned long capacity_of(int cpu)
5404 {
5405         return cpu_rq(cpu)->cpu_capacity;
5406 }
5407 
5408 static unsigned long cpu_avg_load_per_task(int cpu)
5409 {
5410         struct rq *rq = cpu_rq(cpu);
5411         unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5412         unsigned long load_avg = cpu_runnable_load(rq);
5413 
5414         if (nr_running)
5415                 return load_avg / nr_running;
5416 
5417         return 0;
5418 }
5419 
5420 static void record_wakee(struct task_struct *p)
5421 {
5422         /*
5423          * Only decay a single time; tasks that have less then 1 wakeup per
5424          * jiffy will not have built up many flips.
5425          */
5426         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5427                 current->wakee_flips >>= 1;
5428                 current->wakee_flip_decay_ts = jiffies;
5429         }
5430 
5431         if (current->last_wakee != p) {
5432                 current->last_wakee = p;
5433                 current->wakee_flips++;
5434         }
5435 }
5436 
5437 /*
5438  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5439  *
5440  * A waker of many should wake a different task than the one last awakened
5441  * at a frequency roughly N times higher than one of its wakees.
5442  *
5443  * In order to determine whether we should let the load spread vs consolidating
5444  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5445  * partner, and a factor of lls_size higher frequency in the other.
5446  *
5447  * With both conditions met, we can be relatively sure that the relationship is
5448  * non-monogamous, with partner count exceeding socket size.
5449  *
5450  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5451  * whatever is irrelevant, spread criteria is apparent partner count exceeds
5452  * socket size.
5453  */
5454 static int wake_wide(struct task_struct *p)
5455 {
5456         unsigned int master = current->wakee_flips;
5457         unsigned int slave = p->wakee_flips;
5458         int factor = this_cpu_read(sd_llc_size);
5459 
5460         if (master < slave)
5461                 swap(master, slave);
5462         if (slave < factor || master < slave * factor)
5463                 return 0;
5464         return 1;
5465 }
5466 
5467 /*
5468  * The purpose of wake_affine() is to quickly determine on which CPU we can run
5469  * soonest. For the purpose of speed we only consider the waking and previous
5470  * CPU.
5471  *
5472  * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5473  *                      cache-affine and is (or will be) idle.
5474  *
5475  * wake_affine_weight() - considers the weight to reflect the average
5476  *                        scheduling latency of the CPUs. This seems to work
5477  *                        for the overloaded case.
5478  */
5479 static int
5480 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5481 {
5482         /*
5483          * If this_cpu is idle, it implies the wakeup is from interrupt
5484          * context. Only allow the move if cache is shared. Otherwise an
5485          * interrupt intensive workload could force all tasks onto one
5486          * node depending on the IO topology or IRQ affinity settings.
5487          *
5488          * If the prev_cpu is idle and cache affine then avoid a migration.
5489          * There is no guarantee that the cache hot data from an interrupt
5490          * is more important than cache hot data on the prev_cpu and from
5491          * a cpufreq perspective, it's better to have higher utilisation
5492          * on one CPU.
5493          */
5494         if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5495                 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5496 
5497         if (sync && cpu_rq(this_cpu)->nr_running == 1)
5498                 return this_cpu;
5499 
5500         return nr_cpumask_bits;
5501 }
5502 
5503 static int
5504 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5505                    int this_cpu, int prev_cpu, int sync)
5506 {
5507         s64 this_eff_load, prev_eff_load;
5508         unsigned long task_load;
5509 
5510         this_eff_load = cpu_runnable_load(cpu_rq(this_cpu));
5511 
5512         if (sync) {
5513                 unsigned long current_load = task_h_load(current);
5514 
5515                 if (current_load > this_eff_load)
5516                         return this_cpu;
5517 
5518                 this_eff_load -= current_load;
5519         }
5520 
5521         task_load = task_h_load(p);
5522 
5523         this_eff_load += task_load;
5524         if (sched_feat(WA_BIAS))
5525                 this_eff_load *= 100;
5526         this_eff_load *= capacity_of(prev_cpu);
5527 
5528         prev_eff_load = cpu_runnable_load(cpu_rq(prev_cpu));
5529         prev_eff_load -= task_load;
5530         if (sched_feat(WA_BIAS))
5531                 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5532         prev_eff_load *= capacity_of(this_cpu);
5533 
5534         /*
5535          * If sync, adjust the weight of prev_eff_load such that if
5536          * prev_eff == this_eff that select_idle_sibling() will consider
5537          * stacking the wakee on top of the waker if no other CPU is
5538          * idle.
5539          */
5540         if (sync)
5541                 prev_eff_load += 1;
5542 
5543         return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5544 }
5545 
5546 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5547                        int this_cpu, int prev_cpu, int sync)
5548 {
5549         int target = nr_cpumask_bits;
5550 
5551         if (sched_feat(WA_IDLE))
5552                 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5553 
5554         if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5555                 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5556 
5557         schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5558         if (target == nr_cpumask_bits)
5559                 return prev_cpu;
5560 
5561         schedstat_inc(sd->ttwu_move_affine);
5562         schedstat_inc(p->se.statistics.nr_wakeups_affine);
5563         return target;
5564 }
5565 
5566 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5567 
5568 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5569 {
5570         return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5571 }
5572 
5573 /*
5574  * find_idlest_group finds and returns the least busy CPU group within the
5575  * domain.
5576  *
5577  * Assumes p is allowed on at least one CPU in sd.
5578  */
5579 static struct sched_group *
5580 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5581                   int this_cpu, int sd_flag)
5582 {
5583         struct sched_group *idlest = NULL, *group = sd->groups;
5584         struct sched_group *most_spare_sg = NULL;
5585         unsigned long min_runnable_load = ULONG_MAX;
5586         unsigned long this_runnable_load = ULONG_MAX;
5587         unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5588         unsigned long most_spare = 0, this_spare = 0;
5589         int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5590         unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5591                                 (sd->imbalance_pct-100) / 100;
5592 
5593         do {
5594                 unsigned long load, avg_load, runnable_load;
5595                 unsigned long spare_cap, max_spare_cap;
5596                 int local_group;
5597                 int i;
5598 
5599                 /* Skip over this group if it has no CPUs allowed */
5600                 if (!cpumask_intersects(sched_group_span(group),
5601                                         p->cpus_ptr))
5602                         continue;
5603 
5604                 local_group = cpumask_test_cpu(this_cpu,
5605                                                sched_group_span(group));
5606 
5607                 /*
5608                  * Tally up the load of all CPUs in the group and find
5609                  * the group containing the CPU with most spare capacity.
5610                  */
5611                 avg_load = 0;
5612                 runnable_load = 0;
5613                 max_spare_cap = 0;
5614 
5615                 for_each_cpu(i, sched_group_span(group)) {
5616                         load = cpu_runnable_load(cpu_rq(i));
5617                         runnable_load += load;
5618 
5619                         avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5620 
5621                         spare_cap = capacity_spare_without(i, p);
5622 
5623                         if (spare_cap > max_spare_cap)
5624                                 max_spare_cap = spare_cap;
5625                 }
5626 
5627                 /* Adjust by relative CPU capacity of the group */
5628                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5629                                         group->sgc->capacity;
5630                 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5631                                         group->sgc->capacity;
5632 
5633                 if (local_group) {
5634                         this_runnable_load = runnable_load;
5635                         this_avg_load = avg_load;
5636                         this_spare = max_spare_cap;
5637                 } else {
5638                         if (min_runnable_load > (runnable_load + imbalance)) {
5639                                 /*
5640                                  * The runnable load is significantly smaller
5641                                  * so we can pick this new CPU:
5642                                  */
5643                                 min_runnable_load = runnable_load;
5644                                 min_avg_load = avg_load;
5645                                 idlest = group;
5646                         } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5647                                    (100*min_avg_load > imbalance_scale*avg_load)) {
5648                                 /*
5649                                  * The runnable loads are close so take the
5650                                  * blocked load into account through avg_load:
5651                                  */
5652                                 min_avg_load = avg_load;
5653                                 idlest = group;
5654                         }
5655 
5656                         if (most_spare < max_spare_cap) {
5657                                 most_spare = max_spare_cap;
5658                                 most_spare_sg = group;
5659                         }
5660                 }
5661         } while (group = group->next, group != sd->groups);
5662 
5663         /*
5664          * The cross-over point between using spare capacity or least load
5665          * is too conservative for high utilization tasks on partially
5666          * utilized systems if we require spare_capacity > task_util(p),
5667          * so we allow for some task stuffing by using
5668          * spare_capacity > task_util(p)/2.
5669          *
5670          * Spare capacity can't be used for fork because the utilization has
5671          * not been set yet, we must first select a rq to compute the initial
5672          * utilization.
5673          */
5674         if (sd_flag & SD_BALANCE_FORK)
5675                 goto skip_spare;
5676 
5677         if (this_spare > task_util(p) / 2 &&
5678             imbalance_scale*this_spare > 100*most_spare)
5679                 return NULL;
5680 
5681         if (most_spare > task_util(p) / 2)
5682                 return most_spare_sg;
5683 
5684 skip_spare:
5685         if (!idlest)
5686                 return NULL;
5687 
5688         /*
5689          * When comparing groups across NUMA domains, it's possible for the
5690          * local domain to be very lightly loaded relative to the remote
5691          * domains but "imbalance" skews the comparison making remote CPUs
5692          * look much more favourable. When considering cross-domain, add
5693          * imbalance to the runnable load on the remote node and consider
5694          * staying local.
5695          */
5696         if ((sd->flags & SD_NUMA) &&
5697             min_runnable_load + imbalance >= this_runnable_load)
5698                 return NULL;
5699 
5700         if (min_runnable_load > (this_runnable_load + imbalance))
5701                 return NULL;
5702 
5703         if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5704              (100*this_avg_load < imbalance_scale*min_avg_load))
5705                 return NULL;
5706 
5707         return idlest;
5708 }
5709 
5710 /*
5711  * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5712  */
5713 static int
5714 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5715 {
5716         unsigned long load, min_load = ULONG_MAX;
5717         unsigned int min_exit_latency = UINT_MAX;
5718         u64 latest_idle_timestamp = 0;
5719         int least_loaded_cpu = this_cpu;
5720         int shallowest_idle_cpu = -1, si_cpu = -1;
5721         int i;
5722 
5723         /* Check if we have any choice: */
5724         if (group->group_weight == 1)
5725                 return cpumask_first(sched_group_span(group));
5726 
5727         /* Traverse only the allowed CPUs */
5728         for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5729                 if (available_idle_cpu(i)) {
5730                         struct rq *rq = cpu_rq(i);
5731                         struct cpuidle_state *idle = idle_get_state(rq);
5732                         if (idle && idle->exit_latency < min_exit_latency) {
5733                                 /*
5734                                  * We give priority to a CPU whose idle state
5735                                  * has the smallest exit latency irrespective
5736                                  * of any idle timestamp.
5737                                  */
5738                                 min_exit_latency = idle->exit_latency;
5739                                 latest_idle_timestamp = rq->idle_stamp;
5740                                 shallowest_idle_cpu = i;
5741                         } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5742                                    rq->idle_stamp > latest_idle_timestamp) {
5743                                 /*
5744                                  * If equal or no active idle state, then
5745                                  * the most recently idled CPU might have
5746                                  * a warmer cache.
5747                                  */
5748                                 latest_idle_timestamp = rq->idle_stamp;
5749                                 shallowest_idle_cpu = i;
5750                         }
5751                 } else if (shallowest_idle_cpu == -1 && si_cpu == -1) {
5752                         if (sched_idle_cpu(i)) {
5753                                 si_cpu = i;
5754                                 continue;
5755                         }
5756 
5757                         load = cpu_runnable_load(cpu_rq(i));
5758                         if (load < min_load) {
5759                                 min_load = load;
5760                                 least_loaded_cpu = i;
5761                         }
5762                 }
5763         }
5764 
5765         if (shallowest_idle_cpu != -1)
5766                 return shallowest_idle_cpu;
5767         if (si_cpu != -1)
5768                 return si_cpu;
5769         return least_loaded_cpu;
5770 }
5771 
5772 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5773                                   int cpu, int prev_cpu, int sd_flag)
5774 {
5775         int new_cpu = cpu;
5776 
5777         if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5778                 return prev_cpu;
5779 
5780         /*
5781          * We need task's util for capacity_spare_without, sync it up to
5782          * prev_cpu's last_update_time.
5783          */
5784         if (!(sd_flag & SD_BALANCE_FORK))
5785                 sync_entity_load_avg(&p->se);
5786 
5787         while (sd) {
5788                 struct sched_group *group;
5789                 struct sched_domain *tmp;
5790                 int weight;
5791 
5792                 if (!(sd->flags & sd_flag)) {
5793                         sd = sd->child;
5794                         continue;
5795                 }
5796 
5797                 group = find_idlest_group(sd, p, cpu, sd_flag);
5798                 if (!group) {
5799                         sd = sd->child;
5800                         continue;
5801                 }
5802 
5803                 new_cpu = find_idlest_group_cpu(group, p, cpu);
5804                 if (new_cpu == cpu) {
5805                         /* Now try balancing at a lower domain level of 'cpu': */
5806                         sd = sd->child;
5807                         continue;
5808                 }
5809 
5810                 /* Now try balancing at a lower domain level of 'new_cpu': */
5811                 cpu = new_cpu;
5812                 weight = sd->span_weight;
5813                 sd = NULL;
5814                 for_each_domain(cpu, tmp) {
5815                         if (weight <= tmp->span_weight)
5816                                 break;
5817                         if (tmp->flags & sd_flag)
5818                                 sd = tmp;
5819                 }
5820         }
5821 
5822         return new_cpu;
5823 }
5824 
5825 #ifdef CONFIG_SCHED_SMT
5826 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5827 EXPORT_SYMBOL_GPL(sched_smt_present);
5828 
5829 static inline void set_idle_cores(int cpu, int val)
5830 {
5831         struct sched_domain_shared *sds;
5832 
5833         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5834         if (sds)
5835                 WRITE_ONCE(sds->has_idle_cores, val);
5836 }
5837 
5838 static inline bool test_idle_cores(int cpu, bool def)
5839 {
5840         struct sched_domain_shared *sds;
5841 
5842         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5843         if (sds)
5844                 return READ_ONCE(sds->has_idle_cores);
5845 
5846         return def;
5847 }
5848 
5849 /*
5850  * Scans the local SMT mask to see if the entire core is idle, and records this
5851  * information in sd_llc_shared->has_idle_cores.
5852  *
5853  * Since SMT siblings share all cache levels, inspecting this limited remote
5854  * state should be fairly cheap.
5855  */
5856 void __update_idle_core(struct rq *rq)
5857 {
5858         int core = cpu_of(rq);
5859         int cpu;
5860 
5861         rcu_read_lock();
5862         if (test_idle_cores(core, true))
5863                 goto unlock;
5864 
5865         for_each_cpu(cpu, cpu_smt_mask(core)) {
5866                 if (cpu == core)
5867                         continue;
5868 
5869                 if (!available_idle_cpu(cpu))
5870                         goto unlock;
5871         }
5872 
5873         set_idle_cores(core, 1);
5874 unlock:
5875         rcu_read_unlock();
5876 }
5877 
5878 /*
5879  * Scan the entire LLC domain for idle cores; this dynamically switches off if
5880  * there are no idle cores left in the system; tracked through
5881  * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5882  */
5883 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5884 {
5885         struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5886         int core, cpu;
5887 
5888         if (!static_branch_likely(&sched_smt_present))
5889                 return -1;
5890 
5891         if (!test_idle_cores(target, false))
5892                 return -1;
5893 
5894         cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5895 
5896         for_each_cpu_wrap(core, cpus, target) {
5897                 bool idle = true;
5898 
5899                 for_each_cpu(cpu, cpu_smt_mask(core)) {
5900                         __cpumask_clear_cpu(cpu, cpus);
5901                         if (!available_idle_cpu(cpu))
5902                                 idle = false;
5903                 }
5904 
5905                 if (idle)
5906                         return core;
5907         }
5908 
5909         /*
5910          * Failed to find an idle core; stop looking for one.
5911          */
5912         set_idle_cores(target, 0);
5913 
5914         return -1;
5915 }
5916 
5917 /*
5918  * Scan the local SMT mask for idle CPUs.
5919  */
5920 static int select_idle_smt(struct task_struct *p, int target)
5921 {
5922         int cpu, si_cpu = -1;
5923 
5924         if (!static_branch_likely(&sched_smt_present))
5925                 return -1;
5926 
5927         for_each_cpu(cpu, cpu_smt_mask(target)) {
5928                 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5929                         continue;
5930                 if (available_idle_cpu(cpu))
5931                         return cpu;
5932                 if (si_cpu == -1 && sched_idle_cpu(cpu))
5933                         si_cpu = cpu;
5934         }
5935 
5936         return si_cpu;
5937 }
5938 
5939 #else /* CONFIG_SCHED_SMT */
5940 
5941 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5942 {
5943         return -1;
5944 }
5945 
5946 static inline int select_idle_smt(struct task_struct *p, int target)
5947 {
5948         return -1;
5949 }
5950 
5951 #endif /* CONFIG_SCHED_SMT */
5952 
5953 /*
5954  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5955  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5956  * average idle time for this rq (as found in rq->avg_idle).
5957  */
5958 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5959 {
5960         struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5961         struct sched_domain *this_sd;
5962         u64 avg_cost, avg_idle;
5963         u64 time, cost;
5964         s64 delta;
5965         int this = smp_processor_id();
5966         int cpu, nr = INT_MAX, si_cpu = -1;
5967 
5968         this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5969         if (!this_sd)
5970                 return -1;
5971 
5972         /*
5973          * Due to large variance we need a large fuzz factor; hackbench in
5974          * particularly is sensitive here.
5975          */
5976         avg_idle = this_rq()->avg_idle / 512;
5977         avg_cost = this_sd->avg_scan_cost + 1;
5978 
5979         if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5980                 return -1;
5981 
5982         if (sched_feat(SIS_PROP)) {
5983                 u64 span_avg = sd->span_weight * avg_idle;
5984                 if (span_avg > 4*avg_cost)
5985                         nr = div_u64(span_avg, avg_cost);
5986                 else
5987                         nr = 4;
5988         }
5989 
5990         time = cpu_clock(this);
5991 
5992         cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5993 
5994         for_each_cpu_wrap(cpu, cpus, target) {
5995                 if (!--nr)
5996                         return si_cpu;
5997                 if (available_idle_cpu(cpu))
5998                         break;
5999                 if (si_cpu == -1 && sched_idle_cpu(cpu))
6000                         si_cpu = cpu;
6001         }
6002 
6003         time = cpu_clock(this) - time;
6004         cost = this_sd->avg_scan_cost;
6005         delta = (s64)(time - cost) / 8;
6006         this_sd->avg_scan_cost += delta;
6007 
6008         return cpu;
6009 }
6010 
6011 /*
6012  * Try and locate an idle core/thread in the LLC cache domain.
6013  */
6014 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6015 {
6016         struct sched_domain *sd;
6017         int i, recent_used_cpu;
6018 
6019         if (available_idle_cpu(target) || sched_idle_cpu(target))
6020                 return target;
6021 
6022         /*
6023          * If the previous CPU is cache affine and idle, don't be stupid:
6024          */
6025         if (prev != target && cpus_share_cache(prev, target) &&
6026             (available_idle_cpu(prev) || sched_idle_cpu(prev)))
6027                 return prev;
6028 
6029         /* Check a recently used CPU as a potential idle candidate: */
6030         recent_used_cpu = p->recent_used_cpu;
6031         if (recent_used_cpu != prev &&
6032             recent_used_cpu != target &&
6033             cpus_share_cache(recent_used_cpu, target) &&
6034             (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6035             cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
6036                 /*
6037                  * Replace recent_used_cpu with prev as it is a potential
6038                  * candidate for the next wake:
6039                  */
6040                 p->recent_used_cpu = prev;
6041                 return recent_used_cpu;
6042         }
6043 
6044         sd = rcu_dereference(per_cpu(sd_llc, target));
6045         if (!sd)
6046                 return target;
6047 
6048         i = select_idle_core(p, sd, target);
6049         if ((unsigned)i < nr_cpumask_bits)
6050                 return i;
6051 
6052         i = select_idle_cpu(p, sd, target);
6053         if ((unsigned)i < nr_cpumask_bits)
6054                 return i;
6055 
6056         i = select_idle_smt(p, target);
6057         if ((unsigned)i < nr_cpumask_bits)
6058                 return i;
6059 
6060         return target;
6061 }
6062 
6063 /**
6064  * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6065  * @cpu: the CPU to get the utilization of
6066  *
6067  * The unit of the return value must be the one of capacity so we can compare
6068  * the utilization with the capacity of the CPU that is available for CFS task
6069  * (ie cpu_capacity).
6070  *
6071  * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6072  * recent utilization of currently non-runnable tasks on a CPU. It represents
6073  * the amount of utilization of a CPU in the range [0..capacity_orig] where
6074  * capacity_orig is the cpu_capacity available at the highest frequency
6075  * (arch_scale_freq_capacity()).
6076  * The utilization of a CPU converges towards a sum equal to or less than the
6077  * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6078  * the running time on this CPU scaled by capacity_curr.
6079  *
6080  * The estimated utilization of a CPU is defined to be the maximum between its
6081  * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6082  * currently RUNNABLE on that CPU.
6083  * This allows to properly represent the expected utilization of a CPU which
6084  * has just got a big task running since a long sleep period. At the same time
6085  * however it preserves the benefits of the "blocked utilization" in
6086  * describing the potential for other tasks waking up on the same CPU.
6087  *
6088  * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6089  * higher than capacity_orig because of unfortunate rounding in
6090  * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6091  * the average stabilizes with the new running time. We need to check that the
6092  * utilization stays within the range of [0..capacity_orig] and cap it if
6093  * necessary. Without utilization capping, a group could be seen as overloaded
6094  * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6095  * available capacity. We allow utilization to overshoot capacity_curr (but not
6096  * capacity_orig) as it useful for predicting the capacity required after task
6097  * migrations (scheduler-driven DVFS).
6098  *
6099  * Return: the (estimated) utilization for the specified CPU
6100  */
6101 static inline unsigned long cpu_util(int cpu)
6102 {
6103         struct cfs_rq *cfs_rq;
6104         unsigned int util;
6105 
6106         cfs_rq = &cpu_rq(cpu)->cfs;
6107         util = READ_ONCE(cfs_rq->avg.util_avg);
6108 
6109         if (sched_feat(UTIL_EST))
6110                 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6111 
6112         return min_t(unsigned long, util, capacity_orig_of(cpu));
6113 }
6114 
6115 /*
6116  * cpu_util_without: compute cpu utilization without any contributions from *p
6117  * @cpu: the CPU which utilization is requested
6118  * @p: the task which utilization should be discounted
6119  *
6120  * The utilization of a CPU is defined by the utilization of tasks currently
6121  * enqueued on that CPU as well as tasks which are currently sleeping after an
6122  * execution on that CPU.
6123  *
6124  * This method returns the utilization of the specified CPU by discounting the
6125  * utilization of the specified task, whenever the task is currently
6126  * contributing to the CPU utilization.
6127  */
6128 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6129 {
6130         struct cfs_rq *cfs_rq;
6131         unsigned int util;
6132 
6133         /* Task has no contribution or is new */
6134         if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6135                 return cpu_util(cpu);
6136 
6137         cfs_rq = &cpu_rq(cpu)->cfs;
6138         util = READ_ONCE(cfs_rq->avg.util_avg);
6139 
6140         /* Discount task's util from CPU's util */
6141         lsub_positive(&util, task_util(p));
6142 
6143         /*
6144          * Covered cases:
6145          *
6146          * a) if *p is the only task sleeping on this CPU, then:
6147          *      cpu_util (== task_util) > util_est (== 0)
6148          *    and thus we return:
6149          *      cpu_util_without = (cpu_util - task_util) = 0
6150          *
6151          * b) if other tasks are SLEEPING on this CPU, which is now exiting
6152          *    IDLE, then:
6153          *      cpu_util >= task_util
6154          *      cpu_util > util_est (== 0)
6155          *    and thus we discount *p's blocked utilization to return:
6156          *      cpu_util_without = (cpu_util - task_util) >= 0
6157          *
6158          * c) if other tasks are RUNNABLE on that CPU and
6159          *      util_est > cpu_util
6160          *    then we use util_est since it returns a more restrictive
6161          *    estimation of the spare capacity on that CPU, by just
6162          *    considering the expected utilization of tasks already
6163          *    runnable on that CPU.
6164          *
6165          * Cases a) and b) are covered by the above code, while case c) is
6166          * covered by the following code when estimated utilization is
6167          * enabled.
6168          */
6169         if (sched_feat(UTIL_EST)) {
6170                 unsigned int estimated =
6171                         READ_ONCE(cfs_rq->avg.util_est.enqueued);
6172 
6173                 /*
6174                  * Despite the following checks we still have a small window
6175                  * for a possible race, when an execl's select_task_rq_fair()
6176                  * races with LB's detach_task():
6177                  *
6178                  *   detach_task()
6179                  *     p->on_rq = TASK_ON_RQ_MIGRATING;
6180                  *     ---------------------------------- A
6181                  *     deactivate_task()                   \
6182                  *       dequeue_task()                     + RaceTime
6183                  *         util_est_dequeue()              /
6184                  *     ---------------------------------- B
6185                  *
6186                  * The additional check on "current == p" it's required to
6187                  * properly fix the execl regression and it helps in further
6188                  * reducing the chances for the above race.
6189                  */
6190                 if (unlikely(task_on_rq_queued(p) || current == p))
6191                         lsub_positive(&estimated, _task_util_est(p));
6192 
6193                 util = max(util, estimated);
6194         }
6195 
6196         /*
6197          * Utilization (estimated) can exceed the CPU capacity, thus let's
6198          * clamp to the maximum CPU capacity to ensure consistency with
6199          * the cpu_util call.
6200          */
6201         return min_t(unsigned long, util, capacity_orig_of(cpu));
6202 }
6203 
6204 /*
6205  * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6206  * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6207  *
6208  * In that case WAKE_AFFINE doesn't make sense and we'll let
6209  * BALANCE_WAKE sort things out.
6210  */
6211 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6212 {
6213         long min_cap, max_cap;
6214 
6215         if (!static_branch_unlikely(&sched_asym_cpucapacity))
6216                 return 0;
6217 
6218         min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6219         max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6220 
6221         /* Minimum capacity is close to max, no need to abort wake_affine */
6222         if (max_cap - min_cap < max_cap >> 3)
6223                 return 0;
6224 
6225         /* Bring task utilization in sync with prev_cpu */
6226         sync_entity_load_avg(&p->se);
6227 
6228         return !task_fits_capacity(p, min_cap);
6229 }
6230 
6231 /*
6232  * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6233  * to @dst_cpu.
6234  */
6235 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6236 {
6237         struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6238         unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6239 
6240         /*
6241          * If @p migrates from @cpu to another, remove its contribution. Or,
6242          * if @p migrates from another CPU to @cpu, add its contribution. In
6243          * the other cases, @cpu is not impacted by the migration, so the
6244          * util_avg should already be correct.
6245          */
6246         if (task_cpu(p) == cpu && dst_cpu != cpu)
6247                 sub_positive(&util, task_util(p));
6248         else if (task_cpu(p) != cpu && dst_cpu == cpu)
6249                 util += task_util(p);
6250 
6251         if (sched_feat(UTIL_EST)) {
6252                 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6253 
6254                 /*
6255                  * During wake-up, the task isn't enqueued yet and doesn't
6256                  * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6257                  * so just add it (if needed) to "simulate" what will be
6258                  * cpu_util() after the task has been enqueued.
6259                  */
6260                 if (dst_cpu == cpu)
6261                         util_est += _task_util_est(p);
6262 
6263                 util = max(util, util_est);
6264         }
6265 
6266         return min(util, capacity_orig_of(cpu));
6267 }
6268 
6269 /*
6270  * compute_energy(): Estimates the energy that @pd would consume if @p was
6271  * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6272  * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6273  * to compute what would be the energy if we decided to actually migrate that
6274  * task.
6275  */
6276 static long
6277 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6278 {
6279         struct cpumask *pd_mask = perf_domain_span(pd);
6280         unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6281         unsigned long max_util = 0, sum_util = 0;
6282         int cpu;
6283 
6284         /*
6285          * The capacity state of CPUs of the current rd can be driven by CPUs
6286          * of another rd if they belong to the same pd. So, account for the
6287          * utilization of these CPUs too by masking pd with cpu_online_mask
6288          * instead of the rd span.
6289          *
6290          * If an entire pd is outside of the current rd, it will not appear in
6291          * its pd list and will not be accounted by compute_energy().
6292          */
6293         for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6294                 unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu);
6295                 struct task_struct *tsk = cpu == dst_cpu ? p : NULL;
6296 
6297                 /*
6298                  * Busy time computation: utilization clamping is not
6299                  * required since the ratio (sum_util / cpu_capacity)
6300                  * is already enough to scale the EM reported power
6301                  * consumption at the (eventually clamped) cpu_capacity.
6302                  */
6303                 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6304                                                ENERGY_UTIL, NULL);
6305 
6306                 /*
6307                  * Performance domain frequency: utilization clamping
6308                  * must be considered since it affects the selection
6309                  * of the performance domain frequency.
6310                  * NOTE: in case RT tasks are running, by default the
6311                  * FREQUENCY_UTIL's utilization can be max OPP.
6312                  */
6313                 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6314                                               FREQUENCY_UTIL, tsk);
6315                 max_util = max(max_util, cpu_util);
6316         }
6317 
6318         return em_pd_energy(pd->em_pd, max_util, sum_util);
6319 }
6320 
6321 /*
6322  * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6323  * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6324  * spare capacity in each performance domain and uses it as a potential
6325  * candidate to execute the task. Then, it uses the Energy Model to figure
6326  * out which of the CPU candidates is the most energy-efficient.
6327  *
6328  * The rationale for this heuristic is as follows. In a performance domain,
6329  * all the most energy efficient CPU candidates (according to the Energy
6330  * Model) are those for which we'll request a low frequency. When there are
6331  * several CPUs for which the frequency request will be the same, we don't
6332  * have enough data to break the tie between them, because the Energy Model
6333  * only includes active power costs. With this model, if we assume that
6334  * frequency requests follow utilization (e.g. using schedutil), the CPU with
6335  * the maximum spare capacity in a performance domain is guaranteed to be among
6336  * the best candidates of the performance domain.
6337  *
6338  * In practice, it could be preferable from an energy standpoint to pack
6339  * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6340  * but that could also hurt our chances to go cluster idle, and we have no
6341  * ways to tell with the current Energy Model if this is actually a good
6342  * idea or not. So, find_energy_efficient_cpu() basically favors
6343  * cluster-packing, and spreading inside a cluster. That should at least be
6344  * a good thing for latency, and this is consistent with the idea that most
6345  * of the energy savings of EAS come from the asymmetry of the system, and
6346  * not so much from breaking the tie between identical CPUs. That's also the
6347  * reason why EAS is enabled in the topology code only for systems where
6348  * SD_ASYM_CPUCAPACITY is set.
6349  *
6350  * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6351  * they don't have any useful utilization data yet and it's not possible to
6352  * forecast their impact on energy consumption. Consequently, they will be
6353  * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6354  * to be energy-inefficient in some use-cases. The alternative would be to
6355  * bias new tasks towards specific types of CPUs first, or to try to infer
6356  * their util_avg from the parent task, but those heuristics could hurt
6357  * other use-cases too. So, until someone finds a better way to solve this,
6358  * let's keep things simple by re-using the existing slow path.
6359  */
6360 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6361 {
6362         unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6363         struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6364         unsigned long cpu_cap, util, base_energy = 0;
6365         int cpu, best_energy_cpu = prev_cpu;
6366         struct sched_domain *sd;
6367         struct perf_domain *pd;
6368 
6369         rcu_read_lock();
6370         pd = rcu_dereference(rd->pd);
6371         if (!pd || READ_ONCE(rd->overutilized))
6372                 goto fail;
6373 
6374         /*
6375          * Energy-aware wake-up happens on the lowest sched_domain starting
6376          * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6377          */
6378         sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6379         while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6380                 sd = sd->parent;
6381         if (!sd)
6382                 goto fail;
6383 
6384         sync_entity_load_avg(&p->se);
6385         if (!task_util_est(p))
6386                 goto unlock;
6387 
6388         for (; pd; pd = pd->next) {
6389                 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6390                 unsigned long base_energy_pd;
6391                 int max_spare_cap_cpu = -1;
6392 
6393                 /* Compute the 'base' energy of the pd, without @p */
6394                 base_energy_pd = compute_energy(p, -1, pd);
6395                 base_energy += base_energy_pd;
6396 
6397                 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6398                         if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6399                                 continue;
6400 
6401                         /* Skip CPUs that will be overutilized. */
6402                         util = cpu_util_next(cpu, p, cpu);
6403                         cpu_cap = capacity_of(cpu);
6404                         if (!fits_capacity(util, cpu_cap))
6405                                 continue;
6406 
6407                         /* Always use prev_cpu as a candidate. */
6408                         if (cpu == prev_cpu) {
6409                                 prev_delta = compute_energy(p, prev_cpu, pd);
6410                                 prev_delta -= base_energy_pd;
6411                                 best_delta = min(best_delta, prev_delta);
6412                         }
6413 
6414                         /*
6415                          * Find the CPU with the maximum spare capacity in
6416                          * the performance domain
6417                          */
6418                         spare_cap = cpu_cap - util;
6419                         if (spare_cap > max_spare_cap) {
6420                                 max_spare_cap = spare_cap;
6421                                 max_spare_cap_cpu = cpu;
6422                         }
6423                 }
6424 
6425                 /* Evaluate the energy impact of using this CPU. */
6426                 if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) {
6427                         cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6428                         cur_delta -= base_energy_pd;
6429                         if (cur_delta < best_delta) {
6430                                 best_delta = cur_delta;
6431                                 best_energy_cpu = max_spare_cap_cpu;
6432                         }
6433                 }
6434         }
6435 unlock:
6436         rcu_read_unlock();
6437 
6438         /*
6439          * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6440          * least 6% of the energy used by prev_cpu.
6441          */
6442         if (prev_delta == ULONG_MAX)
6443                 return best_energy_cpu;
6444 
6445         if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6446                 return best_energy_cpu;
6447 
6448         return prev_cpu;
6449 
6450 fail:
6451         rcu_read_unlock();
6452 
6453         return -1;
6454 }
6455 
6456 /*
6457  * select_task_rq_fair: Select target runqueue for the waking task in domains
6458  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6459  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6460  *
6461  * Balances load by selecting the idlest CPU in the idlest group, or under
6462  * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6463  *
6464  * Returns the target CPU number.
6465  *
6466  * preempt must be disabled.
6467  */
6468 static int
6469 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6470 {
6471         struct sched_domain *tmp, *sd = NULL;
6472         int cpu = smp_processor_id();
6473         int new_cpu = prev_cpu;
6474         int want_affine = 0;
6475         int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6476 
6477         if (sd_flag & SD_BALANCE_WAKE) {
6478                 record_wakee(p);
6479 
6480                 if (sched_energy_enabled()) {
6481                         new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6482                         if (new_cpu >= 0)
6483                                 return new_cpu;
6484                         new_cpu = prev_cpu;
6485                 }
6486 
6487                 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6488                               cpumask_test_cpu(cpu, p->cpus_ptr);
6489         }
6490 
6491         rcu_read_lock();
6492         for_each_domain(cpu, tmp) {
6493                 if (!(tmp->flags & SD_LOAD_BALANCE))
6494                         break;
6495 
6496                 /*
6497                  * If both 'cpu' and 'prev_cpu' are part of this domain,
6498                  * cpu is a valid SD_WAKE_AFFINE target.
6499                  */
6500                 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6501                     cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6502                         if (cpu != prev_cpu)
6503                                 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6504 
6505                         sd = NULL; /* Prefer wake_affine over balance flags */
6506                         break;
6507                 }
6508 
6509                 if (tmp->flags & sd_flag)
6510                         sd = tmp;
6511                 else if (!want_affine)
6512                         break;
6513         }
6514 
6515         if (unlikely(sd)) {
6516                 /* Slow path */
6517                 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6518         } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6519                 /* Fast path */
6520 
6521                 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6522 
6523                 if (want_affine)
6524                         current->recent_used_cpu = cpu;
6525         }
6526         rcu_read_unlock();
6527 
6528         return new_cpu;
6529 }
6530 
6531 static void detach_entity_cfs_rq(struct sched_entity *se);
6532 
6533 /*
6534  * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6535  * cfs_rq_of(p) references at time of call are still valid and identify the
6536  * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6537  */
6538 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6539 {
6540         /*
6541          * As blocked tasks retain absolute vruntime the migration needs to
6542          * deal with this by subtracting the old and adding the new
6543          * min_vruntime -- the latter is done by enqueue_entity() when placing
6544          * the task on the new runqueue.
6545          */
6546         if (p->state == TASK_WAKING) {
6547                 struct sched_entity *se = &p->se;
6548                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6549                 u64 min_vruntime;
6550 
6551 #ifndef CONFIG_64BIT
6552                 u64 min_vruntime_copy;
6553 
6554                 do {
6555                         min_vruntime_copy = cfs_rq->min_vruntime_copy;
6556                         smp_rmb();
6557                         min_vruntime = cfs_rq->min_vruntime;
6558                 } while (min_vruntime != min_vruntime_copy);
6559 #else
6560                 min_vruntime = cfs_rq->min_vruntime;
6561 #endif
6562 
6563                 se->vruntime -= min_vruntime;
6564         }
6565 
6566         if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6567                 /*
6568                  * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6569                  * rq->lock and can modify state directly.
6570                  */
6571                 lockdep_assert_held(&task_rq(p)->lock);
6572                 detach_entity_cfs_rq(&p->se);
6573 
6574         } else {
6575                 /*
6576                  * We are supposed to update the task to "current" time, then
6577                  * its up to date and ready to go to new CPU/cfs_rq. But we
6578                  * have difficulty in getting what current time is, so simply
6579                  * throw away the out-of-date time. This will result in the
6580                  * wakee task is less decayed, but giving the wakee more load
6581                  * sounds not bad.
6582                  */
6583                 remove_entity_load_avg(&p->se);
6584         }
6585 
6586         /* Tell new CPU we are migrated */
6587         p->se.avg.last_update_time = 0;
6588 
6589         /* We have migrated, no longer consider this task hot */
6590         p->se.exec_start = 0;
6591 
6592         update_scan_period(p, new_cpu);
6593 }
6594 
6595 static void task_dead_fair(struct task_struct *p)
6596 {
6597         remove_entity_load_avg(&p->se);
6598 }
6599 
6600 static int
6601 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6602 {
6603         if (rq->nr_running)
6604                 return 1;
6605 
6606         return newidle_balance(rq, rf) != 0;
6607 }
6608 #endif /* CONFIG_SMP */
6609 
6610 static unsigned long wakeup_gran(struct sched_entity *se)
6611 {
6612         unsigned long gran = sysctl_sched_wakeup_granularity;
6613 
6614         /*
6615          * Since its curr running now, convert the gran from real-time
6616          * to virtual-time in his units.
6617          *
6618          * By using 'se' instead of 'curr' we penalize light tasks, so
6619          * they get preempted easier. That is, if 'se' < 'curr' then
6620          * the resulting gran will be larger, therefore penalizing the
6621          * lighter, if otoh 'se' > 'curr' then the resulting gran will
6622          * be smaller, again penalizing the lighter task.
6623          *
6624          * This is especially important for buddies when the leftmost
6625          * task is higher priority than the buddy.
6626          */
6627         return calc_delta_fair(gran, se);
6628 }
6629 
6630 /*
6631  * Should 'se' preempt 'curr'.
6632  *
6633  *             |s1
6634  *        |s2
6635  *   |s3
6636  *         g
6637  *      |<--->|c
6638  *
6639  *  w(c, s1) = -1
6640  *  w(c, s2) =  0
6641  *  w(c, s3) =  1
6642  *
6643  */
6644 static int
6645 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6646 {
6647         s64 gran, vdiff = curr->vruntime - se->vruntime;
6648 
6649         if (vdiff <= 0)
6650                 return -1;
6651 
6652         gran = wakeup_gran(se);
6653         if (vdiff > gran)
6654                 return 1;
6655 
6656         return 0;
6657 }
6658 
6659 static void set_last_buddy(struct sched_entity *se)
6660 {
6661         if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6662                 return;
6663 
6664         for_each_sched_entity(se) {
6665                 if (SCHED_WARN_ON(!se->on_rq))
6666                         return;
6667                 cfs_rq_of(se)->last = se;
6668         }
6669 }
6670 
6671 static void set_next_buddy(struct sched_entity *se)
6672 {
6673         if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6674                 return;
6675 
6676         for_each_sched_entity(se) {
6677                 if (SCHED_WARN_ON(!se->on_rq))
6678                         return;
6679                 cfs_rq_of(se)->next = se;
6680         }
6681 }
6682 
6683 static void set_skip_buddy(struct sched_entity *se)
6684 {
6685         for_each_sched_entity(se)
6686                 cfs_rq_of(se)->skip = se;
6687 }
6688 
6689 /*
6690  * Preempt the current task with a newly woken task if needed:
6691  */
6692 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6693 {
6694         struct task_struct *curr = rq->curr;
6695         struct sched_entity *se = &curr->se, *pse = &p->se;
6696         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6697         int scale = cfs_rq->nr_running >= sched_nr_latency;
6698         int next_buddy_marked = 0;
6699 
6700         if (unlikely(se == pse))
6701                 return;
6702 
6703         /*
6704          * This is possible from callers such as attach_tasks(), in which we
6705          * unconditionally check_prempt_curr() after an enqueue (which may have
6706          * lead to a throttle).  This both saves work and prevents false
6707          * next-buddy nomination below.
6708          */
6709         if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6710                 return;
6711 
6712         if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6713                 set_next_buddy(pse);
6714                 next_buddy_marked = 1;
6715         }
6716 
6717         /*
6718          * We can come here with TIF_NEED_RESCHED already set from new task
6719          * wake up path.
6720          *
6721          * Note: this also catches the edge-case of curr being in a throttled
6722          * group (e.g. via set_curr_task), since update_curr() (in the
6723          * enqueue of curr) will have resulted in resched being set.  This
6724          * prevents us from potentially nominating it as a false LAST_BUDDY
6725          * below.
6726          */
6727         if (test_tsk_need_resched(curr))
6728                 return;
6729 
6730         /* Idle tasks are by definition preempted by non-idle tasks. */
6731         if (unlikely(task_has_idle_policy(curr)) &&
6732             likely(!task_has_idle_policy(p)))
6733                 goto preempt;
6734 
6735         /*
6736          * Batch and idle tasks do not preempt non-idle tasks (their preemption
6737          * is driven by the tick):
6738          */
6739         if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6740                 return;
6741 
6742         find_matching_se(&se, &pse);
6743         update_curr(cfs_rq_of(se));
6744         BUG_ON(!pse);
6745         if (wakeup_preempt_entity(se, pse) == 1) {
6746                 /*
6747                  * Bias pick_next to pick the sched entity that is
6748                  * triggering this preemption.
6749                  */
6750                 if (!next_buddy_marked)
6751                         set_next_buddy(pse);
6752                 goto preempt;
6753         }
6754 
6755         return;
6756 
6757 preempt:
6758         resched_curr(rq);
6759         /*
6760          * Only set the backward buddy when the current task is still
6761          * on the rq. This can happen when a wakeup gets interleaved
6762          * with schedule on the ->pre_schedule() or idle_balance()
6763          * point, either of which can * drop the rq lock.
6764          *
6765          * Also, during early boot the idle thread is in the fair class,
6766          * for obvious reasons its a bad idea to schedule back to it.
6767          */
6768         if (unlikely(!se->on_rq || curr == rq->idle))
6769                 return;
6770 
6771         if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6772                 set_last_buddy(se);
6773 }
6774 
6775 static struct task_struct *
6776 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6777 {
6778         struct cfs_rq *cfs_rq = &rq->cfs;
6779         struct sched_entity *se;
6780         struct task_struct *p;
6781         int new_tasks;
6782 
6783 again:
6784         if (!sched_fair_runnable(rq))
6785                 goto idle;
6786 
6787 #ifdef CONFIG_FAIR_GROUP_SCHED
6788         if (!prev || prev->sched_class != &fair_sched_class)
6789                 goto simple;
6790 
6791         /*
6792          * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6793          * likely that a next task is from the same cgroup as the current.
6794          *
6795          * Therefore attempt to avoid putting and setting the entire cgroup
6796          * hierarchy, only change the part that actually changes.
6797          */
6798 
6799         do {
6800                 struct sched_entity *curr = cfs_rq->curr;
6801 
6802                 /*
6803                  * Since we got here without doing put_prev_entity() we also
6804                  * have to consider cfs_rq->curr. If it is still a runnable
6805                  * entity, update_curr() will update its vruntime, otherwise
6806                  * forget we've ever seen it.
6807                  */
6808                 if (curr) {
6809                         if (curr->on_rq)
6810                                 update_curr(cfs_rq);
6811                         else
6812                                 curr = NULL;
6813 
6814                         /*
6815                          * This call to check_cfs_rq_runtime() will do the
6816                          * throttle and dequeue its entity in the parent(s).
6817                          * Therefore the nr_running test will indeed
6818                          * be correct.
6819                          */
6820                         if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6821                                 cfs_rq = &rq->cfs;
6822 
6823                                 if (!cfs_rq->nr_running)
6824                                         goto idle;
6825 
6826                                 goto simple;
6827                         }
6828                 }
6829 
6830                 se = pick_next_entity(cfs_rq, curr);
6831                 cfs_rq = group_cfs_rq(se);
6832         } while (cfs_rq);
6833 
6834         p = task_of(se);
6835 
6836         /*
6837          * Since we haven't yet done put_prev_entity and if the selected task
6838          * is a different task than we started out with, try and touch the
6839          * least amount of cfs_rqs.
6840          */
6841         if (prev != p) {
6842                 struct sched_entity *pse = &prev->se;
6843 
6844                 while (!(cfs_rq = is_same_group(se, pse))) {
6845                         int se_depth = se->depth;
6846                         int pse_depth = pse->depth;
6847 
6848                         if (se_depth <= pse_depth) {
6849                                 put_prev_entity(cfs_rq_of(pse), pse);
6850                                 pse = parent_entity(pse);
6851                         }
6852                         if (se_depth >= pse_depth) {
6853                                 set_next_entity(cfs_rq_of(se), se);
6854                                 se = parent_entity(se);
6855                         }
6856                 }
6857 
6858                 put_prev_entity(cfs_rq, pse);
6859                 set_next_entity(cfs_rq, se);
6860         }
6861 
6862         goto done;
6863 simple:
6864 #endif
6865         if (prev)
6866                 put_prev_task(rq, prev);
6867 
6868         do {
6869                 se = pick_next_entity(cfs_rq, NULL);
6870                 set_next_entity(cfs_rq, se);
6871                 cfs_rq = group_cfs_rq(se);
6872         } while (cfs_rq);
6873 
6874         p = task_of(se);
6875 
6876 done: __maybe_unused;
6877 #ifdef CONFIG_SMP
6878         /*
6879          * Move the next running task to the front of
6880          * the list, so our cfs_tasks list becomes MRU
6881          * one.
6882          */
6883         list_move(&p->se.group_node, &rq->cfs_tasks);
6884 #endif
6885 
6886         if (hrtick_enabled(rq))
6887                 hrtick_start_fair(rq, p);
6888 
6889         update_misfit_status(p, rq);
6890 
6891         return p;
6892 
6893 idle:
6894         if (!rf)
6895                 return NULL;
6896 
6897         new_tasks = newidle_balance(rq, rf);
6898 
6899         /*
6900          * Because newidle_balance() releases (and re-acquires) rq->lock, it is
6901          * possible for any higher priority task to appear. In that case we
6902          * must re-start the pick_next_entity() loop.
6903          */
6904         if (new_tasks < 0)
6905                 return RETRY_TASK;
6906 
6907         if (new_tasks > 0)
6908                 goto again;
6909 
6910         /*
6911          * rq is about to be idle, check if we need to update the
6912          * lost_idle_time of clock_pelt
6913          */
6914         update_idle_rq_clock_pelt(rq);
6915 
6916         return NULL;
6917 }
6918 
6919 /*
6920  * Account for a descheduled task:
6921  */
6922 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6923 {
6924         struct sched_entity *se = &prev->se;
6925         struct cfs_rq *cfs_rq;
6926 
6927         for_each_sched_entity(se) {
6928                 cfs_rq = cfs_rq_of(se);
6929                 put_prev_entity(cfs_rq, se);
6930         }
6931 }
6932 
6933 /*
6934  * sched_yield() is very simple
6935  *
6936  * The magic of dealing with the ->skip buddy is in pick_next_entity.
6937  */
6938 static void yield_task_fair(struct rq *rq)
6939 {
6940         struct task_struct *curr = rq->curr;
6941         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6942         struct sched_entity *se = &curr->se;
6943 
6944         /*
6945          * Are we the only task in the tree?
6946          */
6947         if (unlikely(rq->nr_running == 1))
6948                 return;
6949 
6950         clear_buddies(cfs_rq, se);
6951 
6952         if (curr->policy != SCHED_BATCH) {
6953                 update_rq_clock(rq);
6954                 /*
6955                  * Update run-time statistics of the 'current'.
6956                  */
6957                 update_curr(cfs_rq);
6958                 /*
6959                  * Tell update_rq_clock() that we've just updated,
6960                  * so we don't do microscopic update in schedule()
6961                  * and double the fastpath cost.
6962                  */
6963                 rq_clock_skip_update(rq);
6964         }
6965 
6966         set_skip_buddy(se);
6967 }
6968 
6969 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6970 {
6971         struct sched_entity *se = &p->se;
6972 
6973         /* throttled hierarchies are not runnable */
6974         if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6975                 return false;
6976 
6977         /* Tell the scheduler that we'd really like pse to run next. */
6978         set_next_buddy(se);
6979 
6980         yield_task_fair(rq);
6981 
6982         return true;
6983 }
6984 
6985 #ifdef CONFIG_SMP
6986 /**************************************************
6987  * Fair scheduling class load-balancing methods.
6988  *
6989  * BASICS
6990  *
6991  * The purpose of load-balancing is to achieve the same basic fairness the
6992  * per-CPU scheduler provides, namely provide a proportional amount of compute
6993  * time to each task. This is expressed in the following equation:
6994  *
6995  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
6996  *
6997  * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6998  * W_i,0 is defined as:
6999  *
7000  *   W_i,0 = \Sum_j w_i,j                                             (2)
7001  *
7002  * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7003  * is derived from the nice value as per sched_prio_to_weight[].
7004  *
7005  * The weight average is an exponential decay average of the instantaneous
7006  * weight:
7007  *
7008  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
7009  *
7010  * C_i is the compute capacity of CPU i, typically it is the
7011  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7012  * can also include other factors [XXX].
7013  *
7014  * To achieve this balance we define a measure of imbalance which follows
7015  * directly from (1):
7016  *
7017  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
7018  *
7019  * We them move tasks around to minimize the imbalance. In the continuous
7020  * function space it is obvious this converges, in the discrete case we get
7021  * a few fun cases generally called infeasible weight scenarios.
7022  *
7023  * [XXX expand on:
7024  *     - infeasible weights;
7025  *     - local vs global optima in the discrete case. ]
7026  *
7027  *
7028  * SCHED DOMAINS
7029  *
7030  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7031  * for all i,j solution, we create a tree of CPUs that follows the hardware
7032  * topology where each level pairs two lower groups (or better). This results
7033  * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7034  * tree to only the first of the previous level and we decrease the frequency
7035  * of load-balance at each level inv. proportional to the number of CPUs in
7036  * the groups.
7037  *
7038  * This yields:
7039  *
7040  *     log_2 n     1     n
7041  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
7042  *     i = 0      2^i   2^i
7043  *                               `- size of each group
7044  *         |         |     `- number of CPUs doing load-balance
7045  *         |         `- freq
7046  *         `- sum over all levels
7047  *
7048  * Coupled with a limit on how many tasks we can migrate every balance pass,
7049  * this makes (5) the runtime complexity of the balancer.
7050  *
7051  * An important property here is that each CPU is still (indirectly) connected
7052  * to every other CPU in at most O(log n) steps:
7053  *
7054  * The adjacency matrix of the resulting graph is given by:
7055  *
7056  *             log_2 n
7057  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
7058  *             k = 0
7059  *
7060  * And you'll find that:
7061  *
7062  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
7063  *
7064  * Showing there's indeed a path between every CPU in at most O(log n) steps.
7065  * The task movement gives a factor of O(m), giving a convergence complexity
7066  * of:
7067  *
7068  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
7069  *
7070  *
7071  * WORK CONSERVING
7072  *
7073  * In order to avoid CPUs going idle while there's still work to do, new idle
7074  * balancing is more aggressive and has the newly idle CPU iterate up the domain
7075  * tree itself instead of relying on other CPUs to bring it work.
7076  *
7077  * This adds some complexity to both (5) and (8) but it reduces the total idle
7078  * time.
7079  *
7080  * [XXX more?]
7081  *
7082  *
7083  * CGROUPS
7084  *
7085  * Cgroups make a horror show out of (2), instead of a simple sum we get:
7086  *
7087  *                                s_k,i
7088  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
7089  *                                 S_k
7090  *
7091  * Where
7092  *
7093  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
7094  *
7095  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7096  *
7097  * The big problem is S_k, its a global sum needed to compute a local (W_i)
7098  * property.
7099  *
7100  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7101  *      rewrite all of this once again.]
7102  */
7103 
7104 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7105 
7106 enum fbq_type { regular, remote, all };
7107 
7108 enum group_type {
7109         group_other = 0,
7110         group_misfit_task,
7111         group_imbalanced,
7112         group_overloaded,
7113 };
7114 
7115 #define LBF_ALL_PINNED  0x01
7116 #define LBF_NEED_BREAK  0x02
7117 #define LBF_DST_PINNED  0x04
7118 #define LBF_SOME_PINNED 0x08
7119 #define LBF_NOHZ_STATS  0x10
7120 #define LBF_NOHZ_AGAIN  0x20
7121 
7122 struct lb_env {
7123         struct sched_domain     *sd;
7124 
7125         struct rq               *src_rq;
7126         int                     src_cpu;
7127 
7128         int                     dst_cpu;
7129         struct rq               *dst_rq;
7130 
7131         struct cpumask          *dst_grpmask;
7132         int                     new_dst_cpu;
7133         enum cpu_idle_type      idle;
7134         long                    imbalance;
7135         /* The set of CPUs under consideration for load-balancing */
7136         struct cpumask          *cpus;
7137 
7138         unsigned int            flags;
7139 
7140         unsigned int            loop;
7141         unsigned int            loop_break;
7142         unsigned int            loop_max;
7143 
7144         enum fbq_type           fbq_type;
7145         enum group_type         src_grp_type;
7146         struct list_head        tasks;
7147 };
7148 
7149 /*
7150  * Is this task likely cache-hot:
7151  */
7152 static int task_hot(struct task_struct *p, struct lb_env *env)
7153 {
7154         s64 delta;
7155 
7156         lockdep_assert_held(&env->src_rq->lock);
7157 
7158         if (p->sched_class != &fair_sched_class)
7159                 return 0;
7160 
7161         if (unlikely(task_has_idle_policy(p)))
7162                 return 0;
7163 
7164         /*
7165          * Buddy candidates are cache hot:
7166          */
7167         if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7168                         (&p->se == cfs_rq_of(&p->se)->next ||
7169                          &p->se == cfs_rq_of(&p->se)->last))
7170                 return 1;
7171 
7172         if (sysctl_sched_migration_cost == -1)
7173                 return 1;
7174         if (sysctl_sched_migration_cost == 0)
7175                 return 0;
7176 
7177         delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7178 
7179         return delta < (s64)sysctl_sched_migration_cost;
7180 }
7181 
7182 #ifdef CONFIG_NUMA_BALANCING
7183 /*
7184  * Returns 1, if task migration degrades locality
7185  * Returns 0, if task migration improves locality i.e migration preferred.
7186  * Returns -1, if task migration is not affected by locality.
7187  */
7188 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7189 {
7190         struct numa_group *numa_group = rcu_dereference(p->numa_group);
7191         unsigned long src_weight, dst_weight;
7192         int src_nid, dst_nid, dist;
7193 
7194         if (!static_branch_likely(&sched_numa_balancing))
7195                 return -1;
7196 
7197         if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7198                 return -1;
7199 
7200         src_nid = cpu_to_node(env->src_cpu);
7201         dst_nid = cpu_to_node(env->dst_cpu);
7202 
7203         if (src_nid == dst_nid)
7204                 return -1;
7205 
7206         /* Migrating away from the preferred node is always bad. */
7207         if (src_nid == p->numa_preferred_nid) {
7208                 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7209                         return 1;
7210                 else
7211                         return -1;
7212         }
7213 
7214         /* Encourage migration to the preferred node. */
7215         if (dst_nid == p->numa_preferred_nid)
7216                 return 0;
7217 
7218         /* Leaving a core idle is often worse than degrading locality. */
7219         if (env->idle == CPU_IDLE)
7220                 return -1;
7221 
7222         dist = node_distance(src_nid, dst_nid);
7223         if (numa_group) {
7224                 src_weight = group_weight(p, src_nid, dist);
7225                 dst_weight = group_weight(p, dst_nid, dist);
7226         } else {
7227                 src_weight = task_weight(p, src_nid, dist);
7228                 dst_weight = task_weight(p, dst_nid, dist);
7229         }
7230 
7231         return dst_weight < src_weight;
7232 }
7233 
7234 #else
7235 static inline int migrate_degrades_locality(struct task_struct *p,
7236                                              struct lb_env *env)
7237 {
7238         return -1;
7239 }
7240 #endif
7241 
7242 /*
7243  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7244  */
7245 static
7246 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7247 {
7248         int tsk_cache_hot;
7249 
7250         lockdep_assert_held(&env->src_rq->lock);
7251 
7252         /*
7253          * We do not migrate tasks that are:
7254          * 1) throttled_lb_pair, or
7255          * 2) cannot be migrated to this CPU due to cpus_ptr, or
7256          * 3) running (obviously), or
7257          * 4) are cache-hot on their current CPU.
7258          */
7259         if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7260                 return 0;
7261 
7262         if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7263                 int cpu;
7264 
7265                 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7266 
7267                 env->flags |= LBF_SOME_PINNED;
7268 
7269                 /*
7270                  * Remember if this task can be migrated to any other CPU in
7271                  * our sched_group. We may want to revisit it if we couldn't
7272                  * meet load balance goals by pulling other tasks on src_cpu.
7273                  *
7274                  * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7275                  * already computed one in current iteration.
7276                  */
7277                 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7278                         return 0;
7279 
7280                 /* Prevent to re-select dst_cpu via env's CPUs: */
7281                 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7282                         if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7283                                 env->flags |= LBF_DST_PINNED;
7284                                 env->new_dst_cpu = cpu;
7285                                 break;
7286                         }
7287                 }
7288 
7289                 return 0;
7290         }
7291 
7292         /* Record that we found atleast one task that could run on dst_cpu */
7293         env->flags &= ~LBF_ALL_PINNED;
7294 
7295         if (task_running(env->src_rq, p)) {
7296                 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7297                 return 0;
7298         }
7299 
7300         /*
7301          * Aggressive migration if:
7302          * 1) destination numa is preferred
7303          * 2) task is cache cold, or
7304          * 3) too many balance attempts have failed.
7305          */
7306         tsk_cache_hot = migrate_degrades_locality(p, env);
7307         if (tsk_cache_hot == -1)
7308                 tsk_cache_hot = task_hot(p, env);
7309 
7310         if (tsk_cache_hot <= 0 ||
7311             env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7312                 if (tsk_cache_hot == 1) {
7313                         schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7314                         schedstat_inc(p->se.statistics.nr_forced_migrations);
7315                 }
7316                 return 1;
7317         }
7318 
7319         schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7320         return 0;
7321 }
7322 
7323 /*
7324  * detach_task() -- detach the task for the migration specified in env
7325  */
7326 static void detach_task(struct task_struct *p, struct lb_env *env)
7327 {
7328         lockdep_assert_held(&env->src_rq->lock);
7329 
7330         deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7331         set_task_cpu(p, env->dst_cpu);
7332 }
7333 
7334 /*
7335  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7336  * part of active balancing operations within "domain".
7337  *
7338  * Returns a task if successful and NULL otherwise.
7339  */
7340 static struct task_struct *detach_one_task(struct lb_env *env)
7341 {
7342         struct task_struct *p;
7343 
7344         lockdep_assert_held(&env->src_rq->lock);
7345 
7346         list_for_each_entry_reverse(p,
7347                         &env->src_rq->cfs_tasks, se.group_node) {
7348                 if (!can_migrate_task(p, env))
7349                         continue;
7350 
7351                 detach_task(p, env);
7352 
7353                 /*
7354                  * Right now, this is only the second place where
7355                  * lb_gained[env->idle] is updated (other is detach_tasks)
7356                  * so we can safely collect stats here rather than
7357                  * inside detach_tasks().
7358                  */
7359                 schedstat_inc(env->sd->lb_gained[env->idle]);
7360                 return p;
7361         }
7362         return NULL;
7363 }
7364 
7365 static const unsigned int sched_nr_migrate_break = 32;
7366 
7367 /*
7368  * detach_tasks() -- tries to detach up to imbalance runnable load from
7369  * busiest_rq, as part of a balancing operation within domain "sd".
7370  *
7371  * Returns number of detached tasks if successful and 0 otherwise.
7372  */
7373 static int detach_tasks(struct lb_env *env)
7374 {
7375         struct list_head *tasks = &env->src_rq->cfs_tasks;
7376         struct task_struct *p;
7377         unsigned long load;
7378         int detached = 0;
7379 
7380         lockdep_assert_held(&env->src_rq->lock);
7381 
7382         if (env->imbalance <= 0)
7383                 return 0;
7384 
7385         while (!list_empty(tasks)) {
7386                 /*
7387                  * We don't want to steal all, otherwise we may be treated likewise,
7388                  * which could at worst lead to a livelock crash.
7389                  */
7390                 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7391                         break;
7392 
7393                 p = list_last_entry(tasks, struct task_struct, se.group_node);
7394 
7395                 env->loop++;
7396                 /* We've more or less seen every task there is, call it quits */
7397                 if (env->loop > env->loop_max)
7398                         break;
7399 
7400                 /* take a breather every nr_migrate tasks */
7401                 if (env->loop > env->loop_break) {
7402                         env->loop_break += sched_nr_migrate_break;
7403                         env->flags |= LBF_NEED_BREAK;
7404                         break;
7405                 }
7406 
7407                 if (!can_migrate_task(p, env))
7408                         goto next;
7409 
7410                 load = task_h_load(p);
7411 
7412                 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7413                         goto next;
7414 
7415                 if ((load / 2) > env->imbalance)
7416                         goto next;
7417 
7418                 detach_task(p, env);
7419                 list_add(&p->se.group_node, &env->tasks);
7420 
7421                 detached++;
7422                 env->imbalance -= load;
7423 
7424 #ifdef CONFIG_PREEMPTION
7425                 /*
7426                  * NEWIDLE balancing is a source of latency, so preemptible
7427                  * kernels will stop after the first task is detached to minimize
7428                  * the critical section.
7429                  */
7430                 if (env->idle == CPU_NEWLY_IDLE)
7431                         break;
7432 #endif
7433 
7434                 /*
7435                  * We only want to steal up to the prescribed amount of
7436                  * runnable load.
7437                  */
7438                 if (env->imbalance <= 0)
7439                         break;
7440 
7441                 continue;
7442 next:
7443                 list_move(&p->se.group_node, tasks);
7444         }
7445 
7446         /*
7447          * Right now, this is one of only two places we collect this stat
7448          * so we can safely collect detach_one_task() stats here rather
7449          * than inside detach_one_task().
7450          */
7451         schedstat_add(env->sd->lb_gained[env->idle], detached);
7452 
7453         return detached;
7454 }
7455 
7456 /*
7457  * attach_task() -- attach the task detached by detach_task() to its new rq.
7458  */
7459 static void attach_task(struct rq *rq, struct task_struct *p)
7460 {
7461         lockdep_assert_held(&rq->lock);
7462 
7463         BUG_ON(task_rq(p) != rq);
7464         activate_task(rq, p, ENQUEUE_NOCLOCK);
7465         check_preempt_curr(rq, p, 0);
7466 }
7467 
7468 /*
7469  * attach_one_task() -- attaches the task returned from detach_one_task() to
7470  * its new rq.
7471  */
7472 static void attach_one_task(struct rq *rq, struct task_struct *p)
7473 {
7474         struct rq_flags rf;
7475 
7476         rq_lock(rq, &rf);
7477         update_rq_clock(rq);
7478         attach_task(rq, p);
7479         rq_unlock(rq, &rf);
7480 }
7481 
7482 /*
7483  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7484  * new rq.
7485  */
7486 static void attach_tasks(struct lb_env *env)
7487 {
7488         struct list_head *tasks = &env->tasks;
7489         struct task_struct *p;
7490         struct rq_flags rf;
7491 
7492         rq_lock(env->dst_rq, &rf);
7493         update_rq_clock(env->dst_rq);
7494 
7495         while (!list_empty(tasks)) {
7496                 p = list_first_entry(tasks, struct task_struct, se.group_node);
7497                 list_del_init(&p->se.group_node);
7498 
7499                 attach_task(env->dst_rq, p);
7500         }
7501 
7502         rq_unlock(env->dst_rq, &rf);
7503 }
7504 
7505 #ifdef CONFIG_NO_HZ_COMMON
7506 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7507 {
7508         if (cfs_rq->avg.load_avg)
7509                 return true;
7510 
7511         if (cfs_rq->avg.util_avg)
7512                 return true;
7513 
7514         return false;
7515 }
7516 
7517 static inline bool others_have_blocked(struct rq *rq)
7518 {
7519         if (READ_ONCE(rq->avg_rt.util_avg))
7520                 return true;
7521 
7522         if (READ_ONCE(rq->avg_dl.util_avg))
7523                 return true;
7524 
7525 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7526         if (READ_ONCE(rq->avg_irq.util_avg))
7527                 return true;
7528 #endif
7529 
7530         return false;
7531 }
7532 
7533 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7534 {
7535         rq->last_blocked_load_update_tick = jiffies;
7536 
7537         if (!has_blocked)
7538                 rq->has_blocked_load = 0;
7539 }
7540 #else
7541 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7542 static inline bool others_have_blocked(struct rq *rq) { return false; }
7543 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7544 #endif
7545 
7546 static bool __update_blocked_others(struct rq *rq, bool *done)
7547 {
7548         const struct sched_class *curr_class;
7549         u64 now = rq_clock_pelt(rq);
7550         bool decayed;
7551 
7552         /*
7553          * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7554          * DL and IRQ signals have been updated before updating CFS.
7555          */
7556         curr_class = rq->curr->sched_class;
7557 
7558         decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
7559                   update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
7560                   update_irq_load_avg(rq, 0);
7561 
7562         if (others_have_blocked(rq))
7563                 *done = false;
7564 
7565         return decayed;
7566 }
7567 
7568 #ifdef CONFIG_FAIR_GROUP_SCHED
7569 
7570 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7571 {
7572         if (cfs_rq->load.weight)
7573                 return false;
7574 
7575         if (cfs_rq->avg.load_sum)
7576                 return false;
7577 
7578         if (cfs_rq->avg.util_sum)
7579                 return false;
7580 
7581         if (cfs_rq->avg.runnable_load_sum)
7582                 return false;
7583 
7584         return true;
7585 }
7586 
7587 static bool __update_blocked_fair(struct rq *rq, bool *done)
7588 {
7589         struct cfs_rq *cfs_rq, *pos;
7590         bool decayed = false;
7591         int cpu = cpu_of(rq);
7592 
7593         /*
7594          * Iterates the task_group tree in a bottom up fashion, see
7595          * list_add_leaf_cfs_rq() for details.
7596          */
7597         for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7598                 struct sched_entity *se;
7599 
7600                 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
7601                         update_tg_load_avg(cfs_rq, 0);
7602 
7603                         if (cfs_rq == &rq->cfs)
7604                                 decayed = true;
7605                 }
7606 
7607                 /* Propagate pending load changes to the parent, if any: */
7608                 se = cfs_rq->tg->se[cpu];
7609                 if (se && !skip_blocked_update(se))
7610                         update_load_avg(cfs_rq_of(se), se, 0);
7611 
7612                 /*
7613                  * There can be a lot of idle CPU cgroups.  Don't let fully
7614                  * decayed cfs_rqs linger on the list.
7615                  */
7616                 if (cfs_rq_is_decayed(cfs_rq))
7617                         list_del_leaf_cfs_rq(cfs_rq);
7618 
7619                 /* Don't need periodic decay once load/util_avg are null */
7620                 if (cfs_rq_has_blocked(cfs_rq))
7621                         *done = false;
7622         }
7623 
7624         return decayed;
7625 }
7626 
7627 /*
7628  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7629  * This needs to be done in a top-down fashion because the load of a child
7630  * group is a fraction of its parents load.
7631  */
7632 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7633 {
7634         struct rq *rq = rq_of(cfs_rq);
7635         struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7636         unsigned long now = jiffies;
7637         unsigned long load;
7638 
7639         if (cfs_rq->last_h_load_update == now)
7640                 return;
7641 
7642         WRITE_ONCE(cfs_rq->h_load_next, NULL);
7643         for_each_sched_entity(se) {
7644                 cfs_rq = cfs_rq_of(se);
7645                 WRITE_ONCE(cfs_rq->h_load_next, se);
7646                 if (cfs_rq->last_h_load_update == now)
7647                         break;
7648         }
7649 
7650         if (!se) {
7651                 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7652                 cfs_rq->last_h_load_update = now;
7653         }
7654 
7655         while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7656                 load = cfs_rq->h_load;
7657                 load = div64_ul(load * se->avg.load_avg,
7658                         cfs_rq_load_avg(cfs_rq) + 1);
7659                 cfs_rq = group_cfs_rq(se);
7660                 cfs_rq->h_load = load;
7661                 cfs_rq->last_h_load_update = now;
7662         }
7663 }
7664 
7665 static unsigned long task_h_load(struct task_struct *p)
7666 {
7667         struct cfs_rq *cfs_rq = task_cfs_rq(p);
7668 
7669         update_cfs_rq_h_load(cfs_rq);
7670         return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7671                         cfs_rq_load_avg(cfs_rq) + 1);
7672 }
7673 #else
7674 static bool __update_blocked_fair(struct rq *rq, bool *done)
7675 {
7676         struct cfs_rq *cfs_rq = &rq->cfs;
7677         bool decayed;
7678 
7679         decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7680         if (cfs_rq_has_blocked(cfs_rq))
7681                 *done = false;
7682 
7683         return decayed;
7684 }
7685 
7686 static unsigned long task_h_load(struct task_struct *p)
7687 {
7688         return p->se.avg.load_avg;
7689 }
7690 #endif
7691 
7692 static void update_blocked_averages(int cpu)
7693 {
7694         bool decayed = false, done = true;
7695         struct rq *rq = cpu_rq(cpu);
7696         struct rq_flags rf;
7697 
7698         rq_lock_irqsave(rq, &rf);
7699         update_rq_clock(rq);
7700 
7701         decayed |= __update_blocked_others(rq, &done);
7702         decayed |= __update_blocked_fair(rq, &done);
7703 
7704         update_blocked_load_status(rq, !done);
7705         if (decayed)
7706                 cpufreq_update_util(rq, 0);
7707         rq_unlock_irqrestore(rq, &rf);
7708 }
7709 
7710 /********** Helpers for find_busiest_group ************************/
7711 
7712 /*
7713  * sg_lb_stats - stats of a sched_group required for load_balancing
7714  */
7715 struct sg_lb_stats {
7716         unsigned long avg_load; /*Avg load across the CPUs of the group */
7717         unsigned long group_load; /* Total load over the CPUs of the group */
7718         unsigned long load_per_task;
7719         unsigned long group_capacity;
7720         unsigned long group_util; /* Total utilization of the group */
7721         unsigned int sum_nr_running; /* Nr tasks running in the group */
7722         unsigned int idle_cpus;
7723         unsigned int group_weight;
7724         enum group_type group_type;
7725         int group_no_capacity;
7726         unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7727 #ifdef CONFIG_NUMA_BALANCING
7728         unsigned int nr_numa_running;
7729         unsigned int nr_preferred_running;
7730 #endif
7731 };
7732 
7733 /*
7734  * sd_lb_stats - Structure to store the statistics of a sched_domain
7735  *               during load balancing.
7736  */
7737 struct sd_lb_stats {
7738         struct sched_group *busiest;    /* Busiest group in this sd */
7739         struct sched_group *local;      /* Local group in this sd */
7740         unsigned long total_running;
7741         unsigned long total_load;       /* Total load of all groups in sd */
7742         unsigned long total_capacity;   /* Total capacity of all groups in sd */
7743         unsigned long avg_load; /* Average load across all groups in sd */
7744 
7745         struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7746         struct sg_lb_stats local_stat;  /* Statistics of the local group */
7747 };
7748 
7749 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7750 {
7751         /*
7752          * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7753          * local_stat because update_sg_lb_stats() does a full clear/assignment.
7754          * We must however clear busiest_stat::avg_load because
7755          * update_sd_pick_busiest() reads this before assignment.
7756          */
7757         *sds = (struct sd_lb_stats){
7758                 .busiest = NULL,
7759                 .local = NULL,
7760                 .total_running = 0UL,
7761                 .total_load = 0UL,
7762                 .total_capacity = 0UL,
7763                 .busiest_stat = {
7764                         .avg_load = 0UL,
7765                         .sum_nr_running = 0,
7766                         .group_type = group_other,
7767                 },
7768         };
7769 }
7770 
7771 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7772 {
7773         struct rq *rq = cpu_rq(cpu);
7774         unsigned long max = arch_scale_cpu_capacity(cpu);
7775         unsigned long used, free;
7776         unsigned long irq;
7777 
7778         irq = cpu_util_irq(rq);
7779 
7780         if (unlikely(irq >= max))
7781                 return 1;
7782 
7783         used = READ_ONCE(rq->avg_rt.util_avg);
7784         used += READ_ONCE(rq->avg_dl.util_avg);
7785 
7786         if (unlikely(used >= max))
7787                 return 1;
7788 
7789         free = max - used;
7790 
7791         return scale_irq_capacity(free, irq, max);
7792 }
7793 
7794 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7795 {
7796         unsigned long capacity = scale_rt_capacity(sd, cpu);
7797         struct sched_group *sdg = sd->groups;
7798 
7799         cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7800 
7801         if (!capacity)
7802                 capacity = 1;
7803 
7804         cpu_rq(cpu)->cpu_capacity = capacity;
7805         sdg->sgc->capacity = capacity;
7806         sdg->sgc->min_capacity = capacity;
7807         sdg->sgc->max_capacity = capacity;
7808 }
7809 
7810 void update_group_capacity(struct sched_domain *sd, int cpu)
7811 {
7812         struct sched_domain *child = sd->child;
7813         struct sched_group *group, *sdg = sd->groups;
7814         unsigned long capacity, min_capacity, max_capacity;
7815         unsigned long interval;
7816 
7817         interval = msecs_to_jiffies(sd->balance_interval);
7818         interval = clamp(interval, 1UL, max_load_balance_interval);
7819         sdg->sgc->next_update = jiffies + interval;
7820 
7821         if (!child) {
7822                 update_cpu_capacity(sd, cpu);
7823                 return;
7824         }
7825 
7826         capacity = 0;
7827         min_capacity = ULONG_MAX;
7828         max_capacity = 0;
7829 
7830         if (child->flags & SD_OVERLAP) {
7831                 /*
7832                  * SD_OVERLAP domains cannot assume that child groups
7833                  * span the current group.
7834                  */
7835 
7836                 for_each_cpu(cpu, sched_group_span(sdg)) {
7837                         struct sched_group_capacity *sgc;
7838                         struct rq *rq = cpu_rq(cpu);
7839 
7840                         /*
7841                          * build_sched_domains() -> init_sched_groups_capacity()
7842                          * gets here before we've attached the domains to the
7843                          * runqueues.
7844                          *
7845                          * Use capacity_of(), which is set irrespective of domains
7846                          * in update_cpu_capacity().
7847                          *
7848                          * This avoids capacity from being 0 and
7849                          * causing divide-by-zero issues on boot.
7850                          */
7851                         if (unlikely(!rq->sd)) {
7852                                 capacity += capacity_of(cpu);
7853                         } else {
7854                                 sgc = rq->sd->groups->sgc;
7855                                 capacity += sgc->capacity;
7856                         }
7857 
7858                         min_capacity = min(capacity, min_capacity);
7859                         max_capacity = max(capacity, max_capacity);
7860                 }
7861         } else  {
7862                 /*
7863                  * !SD_OVERLAP domains can assume that child groups
7864                  * span the current group.
7865                  */
7866 
7867                 group = child->groups;
7868                 do {
7869                         struct sched_group_capacity *sgc = group->sgc;
7870 
7871                         capacity += sgc->capacity;
7872                         min_capacity = min(sgc->min_capacity, min_capacity);
7873                         max_capacity = max(sgc->max_capacity, max_capacity);
7874                         group = group->next;
7875                 } while (group != child->groups);
7876         }
7877 
7878         sdg->sgc->capacity = capacity;
7879         sdg->sgc->min_capacity = min_capacity;
7880         sdg->sgc->max_capacity = max_capacity;
7881 }
7882 
7883 /*
7884  * Check whether the capacity of the rq has been noticeably reduced by side
7885  * activity. The imbalance_pct is used for the threshold.
7886  * Return true is the capacity is reduced
7887  */
7888 static inline int
7889 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7890 {
7891         return ((rq->cpu_capacity * sd->imbalance_pct) <
7892                                 (rq->cpu_capacity_orig * 100));
7893 }
7894 
7895 /*
7896  * Check whether a rq has a misfit task and if it looks like we can actually
7897  * help that task: we can migrate the task to a CPU of higher capacity, or
7898  * the task's current CPU is heavily pressured.
7899  */
7900 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7901 {
7902         return rq->misfit_task_load &&
7903                 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7904                  check_cpu_capacity(rq, sd));
7905 }
7906 
7907 /*
7908  * Group imbalance indicates (and tries to solve) the problem where balancing
7909  * groups is inadequate due to ->cpus_ptr constraints.
7910  *
7911  * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7912  * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7913  * Something like:
7914  *
7915  *      { 0 1 2 3 } { 4 5 6 7 }
7916  *              *     * * *
7917  *
7918  * If we were to balance group-wise we'd place two tasks in the first group and
7919  * two tasks in the second group. Clearly this is undesired as it will overload
7920  * cpu 3 and leave one of the CPUs in the second group unused.
7921  *
7922  * The current solution to this issue is detecting the skew in the first group
7923  * by noticing the lower domain failed to reach balance and had difficulty
7924  * moving tasks due to affinity constraints.
7925  *
7926  * When this is so detected; this group becomes a candidate for busiest; see
7927  * update_sd_pick_busiest(). And calculate_imbalance() and
7928  * find_busiest_group() avoid some of the usual balance conditions to allow it
7929  * to create an effective group imbalance.
7930  *
7931  * This is a somewhat tricky proposition since the next run might not find the
7932  * group imbalance and decide the groups need to be balanced again. A most
7933  * subtle and fragile situation.
7934  */
7935 
7936 static inline int sg_imbalanced(struct sched_group *group)
7937 {
7938         return group->sgc->imbalance;
7939 }
7940 
7941 /*
7942  * group_has_capacity returns true if the group has spare capacity that could
7943  * be used by some tasks.
7944  * We consider that a group has spare capacity if the  * number of task is
7945  * smaller than the number of CPUs or if the utilization is lower than the
7946  * available capacity for CFS tasks.
7947  * For the latter, we use a threshold to stabilize the state, to take into
7948  * account the variance of the tasks' load and to return true if the available
7949  * capacity in meaningful for the load balancer.
7950  * As an example, an available capacity of 1% can appear but it doesn't make
7951  * any benefit for the load balance.
7952  */
7953 static inline bool
7954 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7955 {
7956         if (sgs->sum_nr_running < sgs->group_weight)
7957                 return true;
7958 
7959         if ((sgs->group_capacity * 100) >
7960                         (sgs->group_util * env->sd->imbalance_pct))
7961                 return true;
7962 
7963         return false;
7964 }
7965 
7966 /*
7967  *  group_is_overloaded returns true if the group has more tasks than it can
7968  *  handle.
7969  *  group_is_overloaded is not equals to !group_has_capacity because a group
7970  *  with the exact right number of tasks, has no more spare capacity but is not
7971  *  overloaded so both group_has_capacity and group_is_overloaded return
7972  *  false.
7973  */
7974 static inline bool
7975 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7976 {
7977         if (sgs->sum_nr_running <= sgs->group_weight)
7978                 return false;
7979 
7980         if ((sgs->group_capacity * 100) <
7981                         (sgs->group_util * env->sd->imbalance_pct))
7982                 return true;
7983 
7984         return false;
7985 }
7986 
7987 /*
7988  * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
7989  * per-CPU capacity than sched_group ref.
7990  */
7991 static inline bool
7992 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7993 {
7994         return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
7995 }
7996 
7997 /*
7998  * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
7999  * per-CPU capacity_orig than sched_group ref.
8000  */
8001 static inline bool
8002 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8003 {
8004         return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
8005 }
8006 
8007 static inline enum
8008 group_type group_classify(struct sched_group *group,
8009                           struct sg_lb_stats *sgs)
8010 {
8011         if (sgs->group_no_capacity)
8012                 return group_overloaded;
8013 
8014         if (sg_imbalanced(group))
8015                 return group_imbalanced;
8016 
8017         if (sgs->group_misfit_task_load)
8018                 return group_misfit_task;
8019 
8020         return group_other;
8021 }
8022 
8023 static bool update_nohz_stats(struct rq *rq, bool force)
8024 {
8025 #ifdef CONFIG_NO_HZ_COMMON
8026         unsigned int cpu = rq->cpu;
8027 
8028         if (!rq->has_blocked_load)
8029                 return false;
8030 
8031         if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8032                 return false;
8033 
8034         if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8035                 return true;
8036 
8037         update_blocked_averages(cpu);
8038 
8039         return rq->has_blocked_load;
8040 #else
8041         return false;
8042 #endif
8043 }
8044 
8045 /**
8046  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8047  * @env: The load balancing environment.
8048  * @group: sched_group whose statistics are to be updated.
8049  * @sgs: variable to hold the statistics for this group.
8050  * @sg_status: Holds flag indicating the status of the sched_group
8051  */
8052 static inline void update_sg_lb_stats(struct lb_env *env,
8053                                       struct sched_group *group,
8054                                       struct sg_lb_stats *sgs,
8055                                       int *sg_status)
8056 {
8057         int i, nr_running;
8058 
8059         memset(sgs, 0, sizeof(*sgs));
8060 
8061         for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8062                 struct rq *rq = cpu_rq(i);
8063 
8064                 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8065                         env->flags |= LBF_NOHZ_AGAIN;
8066 
8067                 sgs->group_load += cpu_runnable_load(rq);
8068                 sgs->group_util += cpu_util(i);
8069                 sgs->sum_nr_running += rq->cfs.h_nr_running;
8070 
8071                 nr_running = rq->nr_running;
8072                 if (nr_running > 1)
8073                         *sg_status |= SG_OVERLOAD;
8074 
8075                 if (cpu_overutilized(i))
8076                         *sg_status |= SG_OVERUTILIZED;
8077 
8078 #ifdef CONFIG_NUMA_BALANCING
8079                 sgs->nr_numa_running += rq->nr_numa_running;
8080                 sgs->nr_preferred_running += rq->nr_preferred_running;
8081 #endif
8082                 /*
8083                  * No need to call idle_cpu() if nr_running is not 0
8084                  */
8085                 if (!nr_running && idle_cpu(i))
8086                         sgs->idle_cpus++;
8087 
8088                 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8089                     sgs->group_misfit_task_load < rq->misfit_task_load) {
8090                         sgs->group_misfit_task_load = rq->misfit_task_load;
8091                         *sg_status |= SG_OVERLOAD;
8092                 }
8093         }
8094 
8095         /* Adjust by relative CPU capacity of the group */
8096         sgs->group_capacity = group->sgc->capacity;
8097         sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8098 
8099         if (sgs->sum_nr_running)
8100                 sgs->load_per_task = sgs->group_load / sgs->sum_nr_running;
8101 
8102         sgs->group_weight = group->group_weight;
8103 
8104         sgs->group_no_capacity = group_is_overloaded(env, sgs);
8105         sgs->group_type = group_classify(group, sgs);
8106 }
8107 
8108 /**
8109  * update_sd_pick_busiest - return 1 on busiest group
8110  * @env: The load balancing environment.
8111  * @sds: sched_domain statistics
8112  * @sg: sched_group candidate to be checked for being the busiest
8113  * @sgs: sched_group statistics
8114  *
8115  * Determine if @sg is a busier group than the previously selected
8116  * busiest group.
8117  *
8118  * Return: %true if @sg is a busier group than the previously selected
8119  * busiest group. %false otherwise.
8120  */
8121 static bool update_sd_pick_busiest(struct lb_env *env,
8122                                    struct sd_lb_stats *sds,
8123                                    struct sched_group *sg,
8124                                    struct sg_lb_stats *sgs)
8125 {
8126         struct sg_lb_stats *busiest = &sds->busiest_stat;
8127 
8128         /*
8129          * Don't try to pull misfit tasks we can't help.
8130          * We can use max_capacity here as reduction in capacity on some
8131          * CPUs in the group should either be possible to resolve
8132          * internally or be covered by avg_load imbalance (eventually).
8133          */
8134         if (sgs->group_type == group_misfit_task &&
8135             (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8136              !group_has_capacity(env, &sds->local_stat)))
8137                 return false;
8138 
8139         if (sgs->group_type > busiest->group_type)
8140                 return true;
8141 
8142         if (sgs->group_type < busiest->group_type)
8143                 return false;
8144 
8145         if (sgs->avg_load <= busiest->avg_load)
8146                 return false;
8147 
8148         if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8149                 goto asym_packing;
8150 
8151         /*
8152          * Candidate sg has no more than one task per CPU and
8153          * has higher per-CPU capacity. Migrating tasks to less
8154          * capable CPUs may harm throughput. Maximize throughput,
8155          * power/energy consequences are not considered.
8156          */
8157         if (sgs->sum_nr_running <= sgs->group_weight &&
8158             group_smaller_min_cpu_capacity(sds->local, sg))
8159                 return false;
8160 
8161         /*
8162          * If we have more than one misfit sg go with the biggest misfit.
8163          */
8164         if (sgs->group_type == group_misfit_task &&
8165             sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8166                 return false;
8167 
8168 asym_packing:
8169         /* This is the busiest node in its class. */
8170         if (!(env->sd->flags & SD_ASYM_PACKING))
8171                 return true;
8172 
8173         /* No ASYM_PACKING if target CPU is already busy */
8174         if (env->idle == CPU_NOT_IDLE)
8175                 return true;
8176         /*
8177          * ASYM_PACKING needs to move all the work to the highest
8178          * prority CPUs in the group, therefore mark all groups
8179          * of lower priority than ourself as busy.
8180          */
8181         if (sgs->sum_nr_running &&
8182             sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8183                 if (!sds->busiest)
8184                         return true;
8185 
8186                 /* Prefer to move from lowest priority CPU's work */
8187                 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8188                                       sg->asym_prefer_cpu))
8189                         return true;
8190         }
8191 
8192         return false;
8193 }
8194 
8195 #ifdef CONFIG_NUMA_BALANCING
8196 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8197 {
8198         if (sgs->sum_nr_running > sgs->nr_numa_running)
8199                 return regular;
8200         if (sgs->sum_nr_running > sgs->nr_preferred_running)
8201                 return remote;
8202         return all;
8203 }
8204 
8205 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8206 {
8207         if (rq->nr_running > rq->nr_numa_running)
8208                 return regular;
8209         if (rq->nr_running > rq->nr_preferred_running)
8210                 return remote;
8211         return all;
8212 }
8213 #else
8214 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8215 {
8216         return all;
8217 }
8218 
8219 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8220 {
8221         return regular;
8222 }
8223 #endif /* CONFIG_NUMA_BALANCING */
8224 
8225 /**
8226  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8227  * @env: The load balancing environment.
8228  * @sds: variable to hold the statistics for this sched_domain.
8229  */
8230 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8231 {
8232         struct sched_domain *child = env->sd->child;
8233         struct sched_group *sg = env->sd->groups;
8234         struct sg_lb_stats *local = &sds->local_stat;
8235         struct sg_lb_stats tmp_sgs;
8236         bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8237         int sg_status = 0;
8238 
8239 #ifdef CONFIG_NO_HZ_COMMON
8240         if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8241                 env->flags |= LBF_NOHZ_STATS;
8242 #endif
8243 
8244         do {
8245                 struct sg_lb_stats *sgs = &tmp_sgs;
8246                 int local_group;
8247 
8248                 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8249                 if (local_group) {
8250                         sds->local = sg;
8251                         sgs = local;
8252 
8253                         if (env->idle != CPU_NEWLY_IDLE ||
8254                             time_after_eq(jiffies, sg->sgc->next_update))
8255                                 update_group_capacity(env->sd, env->dst_cpu);
8256                 }
8257 
8258                 update_sg_lb_stats(env, sg, sgs, &sg_status);
8259 
8260                 if (local_group)
8261                         goto next_group;
8262 
8263                 /*
8264                  * In case the child domain prefers tasks go to siblings
8265                  * first, lower the sg capacity so that we'll try
8266                  * and move all the excess tasks away. We lower the capacity
8267                  * of a group only if the local group has the capacity to fit
8268                  * these excess tasks. The extra check prevents the case where
8269                  * you always pull from the heaviest group when it is already
8270                  * under-utilized (possible with a large weight task outweighs
8271                  * the tasks on the system).
8272                  */
8273                 if (prefer_sibling && sds->local &&
8274                     group_has_capacity(env, local) &&
8275                     (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8276                         sgs->group_no_capacity = 1;
8277                         sgs->group_type = group_classify(sg, sgs);
8278                 }
8279 
8280                 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8281                         sds->busiest = sg;
8282                         sds->busiest_stat = *sgs;
8283                 }
8284 
8285 next_group:
8286                 /* Now, start updating sd_lb_stats */
8287                 sds->total_running += sgs->sum_nr_running;
8288                 sds->total_load += sgs->group_load;
8289                 sds->total_capacity += sgs->group_capacity;
8290 
8291                 sg = sg->next;
8292         } while (sg != env->sd->groups);
8293 
8294 #ifdef CONFIG_NO_HZ_COMMON
8295         if ((env->flags & LBF_NOHZ_AGAIN) &&
8296             cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8297 
8298                 WRITE_ONCE(nohz.next_blocked,
8299                            jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8300         }
8301 #endif
8302 
8303         if (env->sd->flags & SD_NUMA)
8304                 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8305 
8306         if (!env->sd->parent) {
8307                 struct root_domain *rd = env->dst_rq->rd;
8308 
8309                 /* update overload indicator if we are at root domain */
8310                 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8311 
8312                 /* Update over-utilization (tipping point, U >= 0) indicator */
8313                 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8314                 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8315         } else if (sg_status & SG_OVERUTILIZED) {
8316                 struct root_domain *rd = env->dst_rq->rd;
8317 
8318                 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8319                 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8320         }
8321 }
8322 
8323 /**
8324  * check_asym_packing - Check to see if the group is packed into the
8325  *                      sched domain.
8326  *
8327  * This is primarily intended to used at the sibling level.  Some
8328  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
8329  * case of POWER7, it can move to lower SMT modes only when higher
8330  * threads are idle.  When in lower SMT modes, the threads will
8331  * perform better since they share less core resources.  Hence when we
8332  * have idle threads, we want them to be the higher ones.
8333  *
8334  * This packing function is run on idle threads.  It checks to see if
8335  * the busiest CPU in this domain (core in the P7 case) has a higher
8336  * CPU number than the packing function is being run on.  Here we are
8337  * assuming lower CPU number will be equivalent to lower a SMT thread
8338  * number.
8339  *
8340  * Return: 1 when packing is required and a task should be moved to
8341  * this CPU.  The amount of the imbalance is returned in env->imbalance.
8342  *
8343  * @env: The load balancing environment.
8344  * @sds: Statistics of the sched_domain which is to be packed
8345  */
8346 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8347 {
8348         int busiest_cpu;
8349 
8350         if (!(env->sd->flags & SD_ASYM_PACKING))
8351                 return 0;
8352 
8353         if (env->idle == CPU_NOT_IDLE)
8354                 return 0;
8355 
8356         if (!sds->busiest)
8357                 return 0;
8358 
8359         busiest_cpu = sds->busiest->asym_prefer_cpu;
8360         if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8361                 return 0;
8362 
8363         env->imbalance = sds->busiest_stat.group_load;
8364 
8365         return 1;
8366 }
8367 
8368 /**
8369  * fix_small_imbalance - Calculate the minor imbalance that exists
8370  *                      amongst the groups of a sched_domain, during
8371  *                      load balancing.
8372  * @env: The load balancing environment.
8373  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8374  */
8375 static inline
8376 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8377 {
8378         unsigned long tmp, capa_now = 0, capa_move = 0;
8379         unsigned int imbn = 2;
8380         unsigned long scaled_busy_load_per_task;
8381         struct sg_lb_stats *local, *busiest;
8382 
8383         local = &sds->local_stat;
8384         busiest = &sds->busiest_stat;
8385 
8386         if (!local->sum_nr_running)
8387                 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8388         else if (busiest->load_per_task > local->load_per_task)
8389                 imbn = 1;
8390 
8391         scaled_busy_load_per_task =
8392                 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8393                 busiest->group_capacity;
8394 
8395         if (busiest->avg_load + scaled_busy_load_per_task >=
8396             local->avg_load + (scaled_busy_load_per_task * imbn)) {
8397                 env->imbalance = busiest->load_per_task;
8398                 return;
8399         }
8400 
8401         /*
8402          * OK, we don't have enough imbalance to justify moving tasks,
8403          * however we may be able to increase total CPU capacity used by
8404          * moving them.
8405          */
8406 
8407         capa_now += busiest->group_capacity *
8408                         min(busiest->load_per_task, busiest->avg_load);
8409         capa_now += local->group_capacity *
8410                         min(local->load_per_task, local->avg_load);
8411         capa_now /= SCHED_CAPACITY_SCALE;
8412 
8413         /* Amount of load we'd subtract */
8414         if (busiest->avg_load > scaled_busy_load_per_task) {
8415                 capa_move += busiest->group_capacity *
8416                             min(busiest->load_per_task,
8417                                 busiest->avg_load - scaled_busy_load_per_task);
8418         }
8419 
8420         /* Amount of load we'd add */
8421         if (busiest->avg_load * busiest->group_capacity <
8422             busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8423                 tmp = (busiest->avg_load * busiest->group_capacity) /
8424                       local->group_capacity;
8425         } else {
8426                 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8427                       local->group_capacity;
8428         }
8429         capa_move += local->group_capacity *
8430                     min(local->load_per_task, local->avg_load + tmp);
8431         capa_move /= SCHED_CAPACITY_SCALE;
8432 
8433         /* Move if we gain throughput */
8434         if (capa_move > capa_now)
8435                 env->imbalance = busiest->load_per_task;
8436 }
8437 
8438 /**
8439  * calculate_imbalance - Calculate the amount of imbalance present within the
8440  *                       groups of a given sched_domain during load balance.
8441  * @env: load balance environment
8442  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8443  */
8444 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8445 {
8446         unsigned long max_pull, load_above_capacity = ~0UL;
8447         struct sg_lb_stats *local, *busiest;
8448 
8449         local = &sds->local_stat;
8450         busiest = &sds->busiest_stat;
8451 
8452         if (busiest->group_type == group_imbalanced) {
8453                 /*
8454                  * In the group_imb case we cannot rely on group-wide averages
8455                  * to ensure CPU-load equilibrium, look at wider averages. XXX
8456                  */
8457                 busiest->load_per_task =
8458                         min(busiest->load_per_task, sds->avg_load);
8459         }
8460 
8461         /*
8462          * Avg load of busiest sg can be less and avg load of local sg can
8463          * be greater than avg load across all sgs of sd because avg load
8464          * factors in sg capacity and sgs with smaller group_type are
8465          * skipped when updating the busiest sg:
8466          */
8467         if (busiest->group_type != group_misfit_task &&
8468             (busiest->avg_load <= sds->avg_load ||
8469              local->avg_load >= sds->avg_load)) {
8470                 env->imbalance = 0;
8471                 return fix_small_imbalance(env, sds);
8472         }
8473 
8474         /*
8475          * If there aren't any idle CPUs, avoid creating some.
8476          */
8477         if (busiest->group_type == group_overloaded &&
8478             local->group_type   == group_overloaded) {
8479                 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8480                 if (load_above_capacity > busiest->group_capacity) {
8481                         load_above_capacity -= busiest->group_capacity;
8482                         load_above_capacity *= scale_load_down(NICE_0_LOAD);
8483                         load_above_capacity /= busiest->group_capacity;
8484                 } else
8485                         load_above_capacity = ~0UL;
8486         }
8487 
8488         /*
8489          * We're trying to get all the CPUs to the average_load, so we don't
8490          * want to push ourselves above the average load, nor do we wish to
8491          * reduce the max loaded CPU below the average load. At the same time,
8492          * we also don't want to reduce the group load below the group
8493          * capacity. Thus we look for the minimum possible imbalance.
8494          */
8495         max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8496 
8497         /* How much load to actually move to equalise the imbalance */
8498         env->imbalance = min(
8499                 max_pull * busiest->group_capacity,
8500                 (sds->avg_load - local->avg_load) * local->group_capacity
8501         ) / SCHED_CAPACITY_SCALE;
8502 
8503         /* Boost imbalance to allow misfit task to be balanced. */
8504         if (busiest->group_type == group_misfit_task) {
8505                 env->imbalance = max_t(long, env->imbalance,
8506                                        busiest->group_misfit_task_load);
8507         }
8508 
8509         /*
8510          * if *imbalance is less than the average load per runnable task
8511          * there is no guarantee that any tasks will be moved so we'll have
8512          * a think about bumping its value to force at least one task to be
8513          * moved
8514          */
8515         if (env->imbalance < busiest->load_per_task)
8516                 return fix_small_imbalance(env, sds);
8517 }
8518 
8519 /******* find_busiest_group() helpers end here *********************/
8520 
8521 /**
8522  * find_busiest_group - Returns the busiest group within the sched_domain
8523  * if there is an imbalance.
8524  *
8525  * Also calculates the amount of runnable load which should be moved
8526  * to restore balance.
8527  *
8528  * @env: The load balancing environment.
8529  *
8530  * Return:      - The busiest group if imbalance exists.
8531  */
8532 static struct sched_group *find_busiest_group(struct lb_env *env)
8533 {
8534         struct sg_lb_stats *local, *busiest;
8535         struct sd_lb_stats sds;
8536 
8537         init_sd_lb_stats(&sds);
8538 
8539         /*
8540          * Compute the various statistics relavent for load balancing at
8541          * this level.
8542          */
8543         update_sd_lb_stats(env, &sds);
8544 
8545         if (sched_energy_enabled()) {
8546                 struct root_domain *rd = env->dst_rq->rd;
8547 
8548                 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8549                         goto out_balanced;
8550         }
8551 
8552         local = &sds.local_stat;
8553         busiest = &sds.busiest_stat;
8554 
8555         /* ASYM feature bypasses nice load balance check */
8556         if (check_asym_packing(env, &sds))
8557                 return sds.busiest;
8558 
8559         /* There is no busy sibling group to pull tasks from */
8560         if (!sds.busiest || busiest->sum_nr_running == 0)
8561                 goto out_balanced;
8562 
8563         /* XXX broken for overlapping NUMA groups */
8564         sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8565                                                 / sds.total_capacity;
8566 
8567         /*
8568          * If the busiest group is imbalanced the below checks don't
8569          * work because they assume all things are equal, which typically
8570          * isn't true due to cpus_ptr constraints and the like.
8571          */
8572         if (busiest->group_type == group_imbalanced)
8573                 goto force_balance;
8574 
8575         /*
8576          * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8577          * capacities from resulting in underutilization due to avg_load.
8578          */
8579         if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8580             busiest->group_no_capacity)
8581                 goto force_balance;
8582 
8583         /* Misfit tasks should be dealt with regardless of the avg load */
8584         if (busiest->group_type == group_misfit_task)
8585                 goto force_balance;
8586 
8587         /*
8588          * If the local group is busier than the selected busiest group
8589          * don't try and pull any tasks.
8590          */
8591         if (local->avg_load >= busiest->avg_load)
8592                 goto out_balanced;
8593 
8594         /*
8595          * Don't pull any tasks if this group is already above the domain
8596          * average load.
8597          */
8598         if (local->avg_load >= sds.avg_load)
8599                 goto out_balanced;
8600 
8601         if (env->idle == CPU_IDLE) {
8602                 /*
8603                  * This CPU is idle. If the busiest group is not overloaded
8604                  * and there is no imbalance between this and busiest group
8605                  * wrt idle CPUs, it is balanced. The imbalance becomes
8606                  * significant if the diff is greater than 1 otherwise we
8607                  * might end up to just move the imbalance on another group
8608                  */
8609                 if ((busiest->group_type != group_overloaded) &&
8610                                 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8611                         goto out_balanced;
8612         } else {
8613                 /*
8614                  * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8615                  * imbalance_pct to be conservative.
8616                  */
8617                 if (100 * busiest->avg_load <=
8618                                 env->sd->imbalance_pct * local->avg_load)
8619                         goto out_balanced;
8620         }
8621 
8622 force_balance:
8623         /* Looks like there is an imbalance. Compute it */
8624         env->src_grp_type = busiest->group_type;
8625         calculate_imbalance(env, &sds);
8626         return env->imbalance ? sds.busiest : NULL;
8627 
8628 out_balanced:
8629         env->imbalance = 0;
8630         return NULL;
8631 }
8632 
8633 /*
8634  * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8635  */
8636 static struct rq *find_busiest_queue(struct lb_env *env,
8637                                      struct sched_group *group)
8638 {
8639         struct rq *busiest = NULL, *rq;
8640         unsigned long busiest_load = 0, busiest_capacity = 1;
8641         int i;
8642 
8643         for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8644                 unsigned long capacity, load;
8645                 enum fbq_type rt;
8646 
8647                 rq = cpu_rq(i);
8648                 rt = fbq_classify_rq(rq);
8649 
8650                 /*
8651                  * We classify groups/runqueues into three groups:
8652                  *  - regular: there are !numa tasks
8653                  *  - remote:  there are numa tasks that run on the 'wrong' node
8654                  *  - all:     there is no distinction
8655                  *
8656                  * In order to avoid migrating ideally placed numa tasks,
8657                  * ignore those when there's better options.
8658                  *
8659                  * If we ignore the actual busiest queue to migrate another
8660                  * task, the next balance pass can still reduce the busiest
8661                  * queue by moving tasks around inside the node.
8662                  *
8663                  * If we cannot move enough load due to this classification
8664                  * the next pass will adjust the group classification and
8665                  * allow migration of more tasks.
8666                  *
8667                  * Both cases only affect the total convergence complexity.
8668                  */
8669                 if (rt > env->fbq_type)
8670                         continue;
8671 
8672                 /*
8673                  * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8674                  * seek the "biggest" misfit task.
8675                  */
8676                 if (env->src_grp_type == group_misfit_task) {
8677                         if (rq->misfit_task_load > busiest_load) {
8678                                 busiest_load = rq->misfit_task_load;
8679                                 busiest = rq;
8680                         }
8681 
8682                         continue;
8683                 }
8684 
8685                 capacity = capacity_of(i);
8686 
8687                 /*
8688                  * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8689                  * eventually lead to active_balancing high->low capacity.
8690                  * Higher per-CPU capacity is considered better than balancing
8691                  * average load.
8692                  */
8693                 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8694                     capacity_of(env->dst_cpu) < capacity &&
8695                     rq->nr_running == 1)
8696                         continue;
8697 
8698                 load = cpu_runnable_load(rq);
8699 
8700                 /*
8701                  * When comparing with imbalance, use cpu_runnable_load()
8702                  * which is not scaled with the CPU capacity.
8703                  */
8704 
8705                 if (rq->nr_running == 1 && load > env->imbalance &&
8706                     !check_cpu_capacity(rq, env->sd))
8707                         continue;
8708 
8709                 /*
8710                  * For the load comparisons with the other CPU's, consider
8711                  * the cpu_runnable_load() scaled with the CPU capacity, so
8712                  * that the load can be moved away from the CPU that is
8713                  * potentially running at a lower capacity.
8714                  *
8715                  * Thus we're looking for max(load_i / capacity_i), crosswise
8716                  * multiplication to rid ourselves of the division works out
8717                  * to: load_i * capacity_j > load_j * capacity_i;  where j is
8718                  * our previous maximum.
8719                  */
8720                 if (load * busiest_capacity > busiest_load * capacity) {
8721                         busiest_load = load;
8722                         busiest_capacity = capacity;
8723                         busiest = rq;
8724                 }
8725         }
8726 
8727         return busiest;
8728 }
8729 
8730 /*
8731  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8732  * so long as it is large enough.
8733  */
8734 #define MAX_PINNED_INTERVAL     512
8735 
8736 static inline bool
8737 asym_active_balance(struct lb_env *env)
8738 {
8739         /*
8740          * ASYM_PACKING needs to force migrate tasks from busy but
8741          * lower priority CPUs in order to pack all tasks in the
8742          * highest priority CPUs.
8743          */
8744         return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8745                sched_asym_prefer(env->dst_cpu, env->src_cpu);
8746 }
8747 
8748 static inline bool
8749 voluntary_active_balance(struct lb_env *env)
8750 {
8751         struct sched_domain *sd = env->sd;
8752 
8753         if (asym_active_balance(env))
8754                 return 1;
8755 
8756         /*
8757          * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8758          * It's worth migrating the task if the src_cpu's capacity is reduced
8759          * because of other sched_class or IRQs if more capacity stays
8760          * available on dst_cpu.
8761          */
8762         if ((env->idle != CPU_NOT_IDLE) &&
8763             (env->src_rq->cfs.h_nr_running == 1)) {
8764                 if ((check_cpu_capacity(env->src_rq, sd)) &&
8765                     (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8766                         return 1;
8767         }
8768 
8769         if (env->src_grp_type == group_misfit_task)
8770                 return 1;
8771 
8772         return 0;
8773 }
8774 
8775 static int need_active_balance(struct lb_env *env)
8776 {
8777         struct sched_domain *sd = env->sd;
8778 
8779         if (voluntary_active_balance(env))
8780                 return 1;
8781 
8782         return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8783 }
8784 
8785 static int active_load_balance_cpu_stop(void *data);
8786 
8787 static int should_we_balance(struct lb_env *env)
8788 {
8789         struct sched_group *sg = env->sd->groups;
8790         int cpu, balance_cpu = -1;
8791 
8792         /*
8793          * Ensure the balancing environment is consistent; can happen
8794          * when the softirq triggers 'during' hotplug.
8795          */
8796         if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8797                 return 0;
8798 
8799         /*
8800          * In the newly idle case, we will allow all the CPUs
8801          * to do the newly idle load balance.
8802          */
8803         if (env->idle == CPU_NEWLY_IDLE)
8804                 return 1;
8805 
8806         /* Try to find first idle CPU */
8807         for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8808                 if (!idle_cpu(cpu))
8809                         continue;
8810 
8811                 balance_cpu = cpu;
8812                 break;
8813         }
8814 
8815         if (balance_cpu == -1)
8816                 balance_cpu = group_balance_cpu(sg);
8817 
8818         /*
8819          * First idle CPU or the first CPU(busiest) in this sched group
8820          * is eligible for doing load balancing at this and above domains.
8821          */
8822         return balance_cpu == env->dst_cpu;
8823 }
8824 
8825 /*
8826  * Check this_cpu to ensure it is balanced within domain. Attempt to move
8827  * tasks if there is an imbalance.
8828  */
8829 static int load_balance(int this_cpu, struct rq *this_rq,
8830                         struct sched_domain *sd, enum cpu_idle_type idle,
8831                         int *continue_balancing)
8832 {
8833         int ld_moved, cur_ld_moved, active_balance = 0;
8834         struct sched_domain *sd_parent = sd->parent;
8835         struct sched_group *group;
8836         struct rq *busiest;
8837         struct rq_flags rf;
8838         struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8839 
8840         struct lb_env env = {
8841                 .sd             = sd,
8842                 .dst_cpu        = this_cpu,
8843                 .dst_rq         = this_rq,
8844                 .dst_grpmask    = sched_group_span(sd->groups),
8845                 .idle           = idle,
8846                 .loop_break     = sched_nr_migrate_break,
8847                 .cpus           = cpus,
8848                 .fbq_type       = all,
8849                 .tasks          = LIST_HEAD_INIT(env.tasks),
8850         };
8851 
8852         cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8853 
8854         schedstat_inc(sd->lb_count[idle]);
8855 
8856 redo:
8857         if (!should_we_balance(&env)) {
8858                 *continue_balancing = 0;
8859                 goto out_balanced;
8860         }
8861 
8862         group = find_busiest_group(&env);
8863         if (!group) {
8864                 schedstat_inc(sd->lb_nobusyg[idle]);
8865                 goto out_balanced;
8866         }
8867 
8868         busiest = find_busiest_queue(&env, group);
8869         if (!busiest) {
8870                 schedstat_inc(sd->lb_nobusyq[idle]);
8871                 goto out_balanced;
8872         }
8873 
8874         BUG_ON(busiest == env.dst_rq);
8875 
8876         schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8877 
8878         env.src_cpu = busiest->cpu;
8879         env.src_rq = busiest;
8880 
8881         ld_moved = 0;
8882         if (busiest->nr_running > 1) {
8883                 /*
8884                  * Attempt to move tasks. If find_busiest_group has found
8885                  * an imbalance but busiest->nr_running <= 1, the group is
8886                  * still unbalanced. ld_moved simply stays zero, so it is
8887                  * correctly treated as an imbalance.
8888                  */
8889                 env.flags |= LBF_ALL_PINNED;
8890                 env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
8891 
8892 more_balance:
8893                 rq_lock_irqsave(busiest, &rf);
8894                 update_rq_clock(busiest);
8895 
8896                 /*
8897                  * cur_ld_moved - load moved in current iteration
8898                  * ld_moved     - cumulative load moved across iterations
8899                  */
8900                 cur_ld_moved = detach_tasks(&env);
8901 
8902                 /*
8903                  * We've detached some tasks from busiest_rq. Every
8904                  * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8905                  * unlock busiest->lock, and we are able to be sure
8906                  * that nobody can manipulate the tasks in parallel.
8907                  * See task_rq_lock() family for the details.
8908                  */
8909 
8910                 rq_unlock(busiest, &rf);
8911 
8912                 if (cur_ld_moved) {
8913                         attach_tasks(&env);
8914                         ld_moved += cur_ld_moved;
8915                 }
8916 
8917                 local_irq_restore(rf.flags);
8918 
8919                 if (env.flags & LBF_NEED_BREAK) {
8920                         env.flags &= ~LBF_NEED_BREAK;
8921                         goto more_balance;
8922                 }
8923 
8924                 /*
8925                  * Revisit (affine) tasks on src_cpu that couldn't be moved to
8926                  * us and move them to an alternate dst_cpu in our sched_group
8927                  * where they can run. The upper limit on how many times we
8928                  * iterate on same src_cpu is dependent on number of CPUs in our
8929                  * sched_group.
8930                  *
8931                  * This changes load balance semantics a bit on who can move
8932                  * load to a given_cpu. In addition to the given_cpu itself
8933                  * (or a ilb_cpu acting on its behalf where given_cpu is
8934                  * nohz-idle), we now have balance_cpu in a position to move
8935                  * load to given_cpu. In rare situations, this may cause
8936                  * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8937                  * _independently_ and at _same_ time to move some load to
8938                  * given_cpu) causing exceess load to be moved to given_cpu.
8939                  * This however should not happen so much in practice and
8940                  * moreover subsequent load balance cycles should correct the
8941                  * excess load moved.
8942                  */
8943                 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8944 
8945                         /* Prevent to re-select dst_cpu via env's CPUs */
8946                         __cpumask_clear_cpu(env.dst_cpu, env.cpus);
8947 
8948                         env.dst_rq       = cpu_rq(env.new_dst_cpu);
8949                         env.dst_cpu      = env.new_dst_cpu;
8950                         env.flags       &= ~LBF_DST_PINNED;
8951                         env.loop         = 0;
8952                         env.loop_break   = sched_nr_migrate_break;
8953 
8954                         /*
8955                          * Go back to "more_balance" rather than "redo" since we
8956                          * need to continue with same src_cpu.
8957                          */
8958                         goto more_balance;
8959                 }
8960 
8961                 /*
8962                  * We failed to reach balance because of affinity.
8963                  */
8964                 if (sd_parent) {
8965                         int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8966 
8967                         if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8968                                 *group_imbalance = 1;
8969                 }
8970 
8971                 /* All tasks on this runqueue were pinned by CPU affinity */
8972                 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8973                         __cpumask_clear_cpu(cpu_of(busiest), cpus);
8974                         /*
8975                          * Attempting to continue load balancing at the current
8976                          * sched_domain level only makes sense if there are
8977                          * active CPUs remaining as possible busiest CPUs to
8978                          * pull load from which are not contained within the
8979                          * destination group that is receiving any migrated
8980                          * load.
8981                          */
8982                         if (!cpumask_subset(cpus, env.dst_grpmask)) {
8983                                 env.loop = 0;
8984                                 env.loop_break = sched_nr_migrate_break;
8985                                 goto redo;
8986                         }
8987                         goto out_all_pinned;
8988                 }
8989         }
8990 
8991         if (!ld_moved) {
8992                 schedstat_inc(sd->lb_failed[idle]);
8993                 /*
8994                  * Increment the failure counter only on periodic balance.
8995                  * We do not want newidle balance, which can be very
8996                  * frequent, pollute the failure counter causing
8997                  * excessive cache_hot migrations and active balances.
8998                  */
8999                 if (idle != CPU_NEWLY_IDLE)
9000                         sd->nr_balance_failed++;
9001 
9002                 if (need_active_balance(&env)) {
9003                         unsigned long flags;
9004 
9005                         raw_spin_lock_irqsave(&busiest->lock, flags);
9006 
9007                         /*
9008                          * Don't kick the active_load_balance_cpu_stop,
9009                          * if the curr task on busiest CPU can't be
9010                          * moved to this_cpu:
9011                          */
9012                         if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9013                                 raw_spin_unlock_irqrestore(&busiest->lock,
9014                                                             flags);
9015                                 env.flags |= LBF_ALL_PINNED;
9016                                 goto out_one_pinned;
9017                         }
9018 
9019                         /*
9020                          * ->active_balance synchronizes accesses to
9021                          * ->active_balance_work.  Once set, it's cleared
9022                          * only after active load balance is finished.
9023                          */
9024                         if (!busiest->active_balance) {
9025                                 busiest->active_balance = 1;
9026                                 busiest->push_cpu = this_cpu;
9027                                 active_balance = 1;
9028                         }
9029                         raw_spin_unlock_irqrestore(&busiest->lock, flags);
9030 
9031                         if (active_balance) {
9032                                 stop_one_cpu_nowait(cpu_of(busiest),
9033                                         active_load_balance_cpu_stop, busiest,
9034                                         &busiest->active_balance_work);
9035                         }
9036 
9037                         /* We've kicked active balancing, force task migration. */
9038                         sd->nr_balance_failed = sd->cache_nice_tries+1;
9039                 }
9040         } else
9041                 sd->nr_balance_failed = 0;
9042 
9043         if (likely(!active_balance) || voluntary_active_balance(&env)) {
9044                 /* We were unbalanced, so reset the balancing interval */
9045                 sd->balance_interval = sd->min_interval;
9046         } else {
9047                 /*
9048                  * If we've begun active balancing, start to back off. This
9049                  * case may not be covered by the all_pinned logic if there
9050                  * is only 1 task on the busy runqueue (because we don't call
9051                  * detach_tasks).
9052                  */
9053                 if (sd->balance_interval < sd->max_interval)
9054                         sd->balance_interval *= 2;
9055         }
9056 
9057         goto out;
9058 
9059 out_balanced:
9060         /*
9061          * We reach balance although we may have faced some affinity
9062          * constraints. Clear the imbalance flag only if other tasks got
9063          * a chance to move and fix the imbalance.
9064          */
9065         if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9066                 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9067 
9068                 if (*group_imbalance)
9069                         *group_imbalance = 0;
9070         }
9071 
9072 out_all_pinned:
9073         /*
9074          * We reach balance because all tasks are pinned at this level so
9075          * we can't migrate them. Let the imbalance flag set so parent level
9076          * can try to migrate them.
9077          */
9078         schedstat_inc(sd->lb_balanced[idle]);
9079 
9080         sd->nr_balance_failed = 0;
9081 
9082 out_one_pinned:
9083         ld_moved = 0;
9084 
9085         /*
9086          * newidle_balance() disregards balance intervals, so we could
9087          * repeatedly reach this code, which would lead to balance_interval
9088          * skyrocketting in a short amount of time. Skip the balance_interval
9089          * increase logic to avoid that.
9090          */
9091         if (env.idle == CPU_NEWLY_IDLE)
9092                 goto out;
9093 
9094         /* tune up the balancing interval */
9095         if ((env.flags & LBF_ALL_PINNED &&
9096              sd->balance_interval < MAX_PINNED_INTERVAL) ||
9097             sd->balance_interval < sd->max_interval)
9098                 sd->balance_interval *= 2;
9099 out:
9100         return ld_moved;
9101 }
9102 
9103 static inline unsigned long
9104 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9105 {
9106         unsigned long interval = sd->balance_interval;
9107 
9108         if (cpu_busy)
9109                 interval *= sd->busy_factor;
9110 
9111         /* scale ms to jiffies */
9112         interval = msecs_to_jiffies(interval);
9113         interval = clamp(interval, 1UL, max_load_balance_interval);
9114 
9115         return interval;
9116 }
9117 
9118 static inline void
9119 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9120 {
9121         unsigned long interval, next;
9122 
9123         /* used by idle balance, so cpu_busy = 0 */
9124         interval = get_sd_balance_interval(sd, 0);
9125         next = sd->last_balance + interval;
9126 
9127         if (time_after(*next_balance, next))
9128                 *next_balance = next;
9129 }
9130 
9131 /*
9132  * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9133  * running tasks off the busiest CPU onto idle CPUs. It requires at
9134  * least 1 task to be running on each physical CPU where possible, and
9135  * avoids physical / logical imbalances.
9136  */
9137 static int active_load_balance_cpu_stop(void *data)
9138 {
9139         struct rq *busiest_rq = data;
9140         int busiest_cpu = cpu_of(busiest_rq);
9141         int target_cpu = busiest_rq->push_cpu;
9142         struct rq *target_rq = cpu_rq(target_cpu);
9143         struct sched_domain *sd;
9144         struct task_struct *p = NULL;
9145         struct rq_flags rf;
9146 
9147         rq_lock_irq(busiest_rq, &rf);
9148         /*
9149          * Between queueing the stop-work and running it is a hole in which
9150          * CPUs can become inactive. We should not move tasks from or to
9151          * inactive CPUs.
9152          */
9153         if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9154                 goto out_unlock;
9155 
9156         /* Make sure the requested CPU hasn't gone down in the meantime: */
9157         if (unlikely(busiest_cpu != smp_processor_id() ||
9158                      !busiest_rq->active_balance))
9159                 goto out_unlock;
9160 
9161         /* Is there any task to move? */
9162         if (busiest_rq->nr_running <= 1)
9163                 goto out_unlock;
9164 
9165         /*
9166          * This condition is "impossible", if it occurs
9167          * we need to fix it. Originally reported by
9168          * Bjorn Helgaas on a 128-CPU setup.
9169          */
9170         BUG_ON(busiest_rq == target_rq);
9171 
9172         /* Search for an sd spanning us and the target CPU. */
9173         rcu_read_lock();
9174         for_each_domain(target_cpu, sd) {
9175                 if ((sd->flags & SD_LOAD_BALANCE) &&
9176                     cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9177                                 break;
9178         }
9179 
9180         if (likely(sd)) {
9181                 struct lb_env env = {
9182                         .sd             = sd,
9183                         .dst_cpu        = target_cpu,
9184                         .dst_rq         = target_rq,
9185                         .src_cpu        = busiest_rq->cpu,
9186                         .src_rq         = busiest_rq,
9187                         .idle           = CPU_IDLE,
9188                         /*
9189                          * can_migrate_task() doesn't need to compute new_dst_cpu
9190                          * for active balancing. Since we have CPU_IDLE, but no
9191                          * @dst_grpmask we need to make that test go away with lying
9192                          * about DST_PINNED.
9193                          */
9194                         .flags          = LBF_DST_PINNED,
9195                 };
9196 
9197                 schedstat_inc(sd->alb_count);
9198                 update_rq_clock(busiest_rq);
9199 
9200                 p = detach_one_task(&env);
9201                 if (p) {
9202                         schedstat_inc(sd->alb_pushed);
9203                         /* Active balancing done, reset the failure counter. */
9204                         sd->nr_balance_failed = 0;
9205                 } else {
9206                         schedstat_inc(sd->alb_failed);
9207                 }
9208         }
9209         rcu_read_unlock();
9210 out_unlock:
9211         busiest_rq->active_balance = 0;
9212         rq_unlock(busiest_rq, &rf);
9213 
9214         if (p)
9215                 attach_one_task(target_rq, p);
9216 
9217         local_irq_enable();
9218 
9219         return 0;
9220 }
9221 
9222 static DEFINE_SPINLOCK(balancing);
9223 
9224 /*
9225  * Scale the max load_balance interval with the number of CPUs in the system.
9226  * This trades load-balance latency on larger machines for less cross talk.
9227  */
9228 void update_max_interval(void)
9229 {
9230         max_load_balance_interval = HZ*num_online_cpus()/10;
9231 }
9232 
9233 /*
9234  * It checks each scheduling domain to see if it is due to be balanced,
9235  * and initiates a balancing operation if so.
9236  *
9237  * Balancing parameters are set up in init_sched_domains.
9238  */
9239 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9240 {
9241         int continue_balancing = 1;
9242         int cpu = rq->cpu;
9243         unsigned long interval;
9244         struct sched_domain *sd;
9245         /* Earliest time when we have to do rebalance again */
9246         unsigned long next_balance = jiffies + 60*HZ;
9247         int update_next_balance = 0;
9248         int need_serialize, need_decay = 0;
9249         u64 max_cost = 0;
9250 
9251         rcu_read_lock();
9252         for_each_domain(cpu, sd) {
9253                 /*
9254                  * Decay the newidle max times here because this is a regular
9255                  * visit to all the domains. Decay ~1% per second.
9256                  */
9257                 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9258                         sd->max_newidle_lb_cost =
9259                                 (sd->max_newidle_lb_cost * 253) / 256;
9260                         sd->next_decay_max_lb_cost = jiffies + HZ;
9261                         need_decay = 1;
9262                 }
9263                 max_cost += sd->max_newidle_lb_cost;
9264 
9265                 if (!(sd->flags & SD_LOAD_BALANCE))
9266                         continue;
9267 
9268                 /*
9269                  * Stop the load balance at this level. There is another
9270                  * CPU in our sched group which is doing load balancing more
9271                  * actively.
9272                  */
9273                 if (!continue_balancing) {
9274                         if (need_decay)
9275                                 continue;
9276                         break;
9277                 }
9278 
9279                 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9280 
9281                 need_serialize = sd->flags & SD_SERIALIZE;
9282                 if (need_serialize) {
9283                         if (!spin_trylock(&balancing))
9284                                 goto out;
9285                 }
9286 
9287                 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9288                         if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9289                                 /*
9290                                  * The LBF_DST_PINNED logic could have changed
9291                                  * env->dst_cpu, so we can't know our idle
9292                                  * state even if we migrated tasks. Update it.
9293                                  */
9294                                 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9295                         }
9296                         sd->last_balance = jiffies;
9297                         interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9298                 }
9299                 if (need_serialize)
9300                         spin_unlock(&balancing);
9301 out:
9302                 if (time_after(next_balance, sd->last_balance + interval)) {
9303                         next_balance = sd->last_balance + interval;
9304                         update_next_balance = 1;
9305                 }
9306         }
9307         if (need_decay) {
9308                 /*
9309                  * Ensure the rq-wide value also decays but keep it at a
9310                  * reasonable floor to avoid funnies with rq->avg_idle.
9311                  */
9312                 rq->max_idle_balance_cost =
9313                         max((u64)sysctl_sched_migration_cost, max_cost);
9314         }
9315         rcu_read_unlock();
9316 
9317         /*
9318          * next_balance will be updated only when there is a need.
9319          * When the cpu is attached to null domain for ex, it will not be
9320          * updated.
9321          */
9322         if (likely(update_next_balance)) {
9323                 rq->next_balance = next_balance;
9324 
9325 #ifdef CONFIG_NO_HZ_COMMON
9326                 /*
9327                  * If this CPU has been elected to perform the nohz idle
9328                  * balance. Other idle CPUs have already rebalanced with
9329                  * nohz_idle_balance() and nohz.next_balance has been
9330                  * updated accordingly. This CPU is now running the idle load
9331                  * balance for itself and we need to update the
9332                  * nohz.next_balance accordingly.
9333                  */
9334                 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9335                         nohz.next_balance = rq->next_balance;
9336 #endif
9337         }
9338 }
9339 
9340 static inline int on_null_domain(struct rq *rq)
9341 {
9342         return unlikely(!rcu_dereference_sched(rq->sd));
9343 }
9344 
9345 #ifdef CONFIG_NO_HZ_COMMON
9346 /*
9347  * idle load balancing details
9348  * - When one of the busy CPUs notice that there may be an idle rebalancing
9349  *   needed, they will kick the idle load balancer, which then does idle
9350  *   load balancing for all the idle CPUs.
9351  * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9352  *   anywhere yet.
9353  */
9354 
9355 static inline int find_new_ilb(void)
9356 {
9357         int ilb;
9358 
9359         for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9360                               housekeeping_cpumask(HK_FLAG_MISC)) {
9361                 if (idle_cpu(ilb))
9362                         return ilb;
9363         }
9364 
9365         return nr_cpu_ids;
9366 }
9367 
9368 /*
9369  * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9370  * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9371  */
9372 static void kick_ilb(unsigned int flags)
9373 {
9374         int ilb_cpu;
9375 
9376         nohz.next_balance++;
9377 
9378         ilb_cpu = find_new_ilb();
9379 
9380         if (ilb_cpu >= nr_cpu_ids)
9381                 return;
9382 
9383         flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9384         if (flags & NOHZ_KICK_MASK)
9385                 return;
9386 
9387         /*
9388          * Use smp_send_reschedule() instead of resched_cpu().
9389          * This way we generate a sched IPI on the target CPU which
9390          * is idle. And the softirq performing nohz idle load balance
9391          * will be run before returning from the IPI.
9392          */
9393         smp_send_reschedule(ilb_cpu);
9394 }
9395 
9396 /*
9397  * Current decision point for kicking the idle load balancer in the presence
9398  * of idle CPUs in the system.
9399  */
9400 static void nohz_balancer_kick(struct rq *rq)
9401 {
9402         unsigned long now = jiffies;
9403         struct sched_domain_shared *sds;
9404         struct sched_domain *sd;
9405         int nr_busy, i, cpu = rq->cpu;
9406         unsigned int flags = 0;
9407 
9408         if (unlikely(rq->idle_balance))
9409                 return;
9410 
9411         /*
9412          * We may be recently in ticked or tickless idle mode. At the first
9413          * busy tick after returning from idle, we will update the busy stats.
9414          */
9415         nohz_balance_exit_idle(rq);
9416 
9417         /*
9418          * None are in tickless mode and hence no need for NOHZ idle load
9419          * balancing.
9420          */
9421         if (likely(!atomic_read(&nohz.nr_cpus)))
9422                 return;
9423 
9424         if (READ_ONCE(nohz.has_blocked) &&
9425             time_after(now, READ_ONCE(nohz.next_blocked)))
9426                 flags = NOHZ_STATS_KICK;
9427 
9428         if (time_before(now, nohz.next_balance))
9429                 goto out;
9430 
9431         if (rq->nr_running >= 2) {
9432                 flags = NOHZ_KICK_MASK;
9433                 goto out;
9434         }
9435 
9436         rcu_read_lock();
9437 
9438         sd = rcu_dereference(rq->sd);
9439         if (sd) {
9440                 /*
9441                  * If there's a CFS task and the current CPU has reduced
9442                  * capacity; kick the ILB to see if there's a better CPU to run
9443                  * on.
9444                  */
9445                 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9446                         flags = NOHZ_KICK_MASK;
9447                         goto unlock;
9448                 }
9449         }
9450 
9451         sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9452         if (sd) {
9453                 /*
9454                  * When ASYM_PACKING; see if there's a more preferred CPU
9455                  * currently idle; in which case, kick the ILB to move tasks
9456                  * around.
9457                  */
9458                 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9459                         if (sched_asym_prefer(i, cpu)) {
9460                                 flags = NOHZ_KICK_MASK;
9461                                 goto unlock;
9462                         }
9463                 }
9464         }
9465 
9466         sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9467         if (sd) {
9468                 /*
9469                  * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9470                  * to run the misfit task on.
9471                  */
9472                 if (check_misfit_status(rq, sd)) {
9473                         flags = NOHZ_KICK_MASK;
9474                         goto unlock;
9475                 }
9476 
9477                 /*
9478                  * For asymmetric systems, we do not want to nicely balance
9479                  * cache use, instead we want to embrace asymmetry and only
9480                  * ensure tasks have enough CPU capacity.
9481                  *
9482                  * Skip the LLC logic because it's not relevant in that case.
9483                  */
9484                 goto unlock;
9485         }
9486 
9487         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9488         if (sds) {
9489                 /*
9490                  * If there is an imbalance between LLC domains (IOW we could
9491                  * increase the overall cache use), we need some less-loaded LLC
9492                  * domain to pull some load. Likewise, we may need to spread
9493                  * load within the current LLC domain (e.g. packed SMT cores but
9494                  * other CPUs are idle). We can't really know from here how busy
9495                  * the others are - so just get a nohz balance going if it looks
9496                  * like this LLC domain has tasks we could move.
9497                  */
9498                 nr_busy = atomic_read(&sds->nr_busy_cpus);
9499                 if (nr_busy > 1) {
9500                         flags = NOHZ_KICK_MASK;
9501                         goto unlock;
9502                 }
9503         }
9504 unlock:
9505         rcu_read_unlock();
9506 out:
9507         if (flags)
9508                 kick_ilb(flags);
9509 }
9510 
9511 static void set_cpu_sd_state_busy(int cpu)
9512 {
9513         struct sched_domain *sd;
9514 
9515         rcu_read_lock();
9516         sd = rcu_dereference(per_cpu(sd_llc, cpu));
9517 
9518         if (!sd || !sd->nohz_idle)
9519                 goto unlock;
9520         sd->nohz_idle = 0;
9521 
9522         atomic_inc(&sd->shared->nr_busy_cpus);
9523 unlock:
9524         rcu_read_unlock();
9525 }
9526 
9527 void nohz_balance_exit_idle(struct rq *rq)
9528 {
9529         SCHED_WARN_ON(rq != this_rq());
9530 
9531         if (likely(!rq->nohz_tick_stopped))
9532                 return;
9533 
9534         rq->nohz_tick_stopped = 0;
9535         cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9536         atomic_dec(&nohz.nr_cpus);
9537 
9538         set_cpu_sd_state_busy(rq->cpu);
9539 }
9540 
9541 static void set_cpu_sd_state_idle(int cpu)
9542 {
9543         struct sched_domain *sd;
9544 
9545         rcu_read_lock();
9546         sd = rcu_dereference(per_cpu(sd_llc, cpu));
9547 
9548         if (!sd || sd->nohz_idle)
9549                 goto unlock;
9550         sd->nohz_idle = 1;
9551 
9552         atomic_dec(&sd->shared->nr_busy_cpus);
9553 unlock:
9554         rcu_read_unlock();
9555 }
9556 
9557 /*
9558  * This routine will record that the CPU is going idle with tick stopped.
9559  * This info will be used in performing idle load balancing in the future.
9560  */
9561 void nohz_balance_enter_idle(int cpu)
9562 {
9563         struct rq *rq = cpu_rq(cpu);
9564 
9565         SCHED_WARN_ON(cpu != smp_processor_id());
9566 
9567         /* If this CPU is going down, then nothing needs to be done: */
9568         if (!cpu_active(cpu))
9569                 return;
9570 
9571         /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9572         if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9573                 return;
9574 
9575         /*
9576          * Can be set safely without rq->lock held
9577          * If a clear happens, it will have evaluated last additions because
9578          * rq->lock is held during the check and the clear
9579          */
9580         rq->has_blocked_load = 1;
9581 
9582         /*
9583          * The tick is still stopped but load could have been added in the
9584          * meantime. We set the nohz.has_blocked flag to trig a check of the
9585          * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9586          * of nohz.has_blocked can only happen after checking the new load
9587          */
9588         if (rq->nohz_tick_stopped)
9589                 goto out;
9590 
9591         /* If we're a completely isolated CPU, we don't play: */
9592         if (on_null_domain(rq))
9593                 return;
9594 
9595         rq->nohz_tick_stopped = 1;
9596 
9597         cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9598         atomic_inc(&nohz.nr_cpus);
9599 
9600         /*
9601          * Ensures that if nohz_idle_balance() fails to observe our
9602          * @idle_cpus_mask store, it must observe the @has_blocked
9603          * store.
9604          */
9605         smp_mb__after_atomic();
9606 
9607         set_cpu_sd_state_idle(cpu);
9608 
9609 out:
9610         /*
9611          * Each time a cpu enter idle, we assume that it has blocked load and
9612          * enable the periodic update of the load of idle cpus
9613          */
9614         WRITE_ONCE(nohz.has_blocked, 1);
9615 }
9616 
9617 /*
9618  * Internal function that runs load balance for all idle cpus. The load balance
9619  * can be a simple update of blocked load or a complete load balance with
9620  * tasks movement depending of flags.
9621  * The function returns false if the loop has stopped before running
9622  * through all idle CPUs.
9623  */
9624 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9625                                enum cpu_idle_type idle)
9626 {
9627         /* Earliest time when we have to do rebalance again */
9628         unsigned long now = jiffies;
9629         unsigned long next_balance = now + 60*HZ;
9630         bool has_blocked_load = false;
9631         int update_next_balance = 0;
9632         int this_cpu = this_rq->cpu;
9633         int balance_cpu;
9634         int ret = false;
9635         struct rq *rq;
9636 
9637         SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9638 
9639         /*
9640          * We assume there will be no idle load after this update and clear
9641          * the has_blocked flag. If a cpu enters idle in the mean time, it will
9642          * set the has_blocked flag and trig another update of idle load.
9643          * Because a cpu that becomes idle, is added to idle_cpus_mask before
9644          * setting the flag, we are sure to not clear the state and not
9645          * check the load of an idle cpu.
9646          */
9647         WRITE_ONCE(nohz.has_blocked, 0);
9648 
9649         /*
9650          * Ensures that if we miss the CPU, we must see the has_blocked
9651          * store from nohz_balance_enter_idle().
9652          */
9653         smp_mb();
9654 
9655         for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9656                 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9657                         continue;
9658 
9659                 /*
9660                  * If this CPU gets work to do, stop the load balancing
9661                  * work being done for other CPUs. Next load
9662                  * balancing owner will pick it up.
9663                  */
9664                 if (need_resched()) {
9665                         has_blocked_load = true;
9666                         goto abort;
9667                 }
9668 
9669                 rq = cpu_rq(balance_cpu);
9670 
9671                 has_blocked_load |= update_nohz_stats(rq, true);
9672 
9673                 /*
9674                  * If time for next balance is due,
9675                  * do the balance.
9676                  */
9677                 if (time_after_eq(jiffies, rq->next_balance)) {
9678                         struct rq_flags rf;
9679 
9680                         rq_lock_irqsave(rq, &rf);
9681                         update_rq_clock(rq);
9682                         rq_unlock_irqrestore(rq, &rf);
9683 
9684                         if (flags & NOHZ_BALANCE_KICK)
9685                                 rebalance_domains(rq, CPU_IDLE);
9686                 }
9687 
9688                 if (time_after(next_balance, rq->next_balance)) {
9689                         next_balance = rq->next_balance;
9690                         update_next_balance = 1;
9691                 }
9692         }
9693 
9694         /* Newly idle CPU doesn't need an update */
9695         if (idle != CPU_NEWLY_IDLE) {
9696                 update_blocked_averages(this_cpu);
9697                 has_blocked_load |= this_rq->has_blocked_load;
9698         }
9699 
9700         if (flags & NOHZ_BALANCE_KICK)
9701                 rebalance_domains(this_rq, CPU_IDLE);
9702 
9703         WRITE_ONCE(nohz.next_blocked,
9704                 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9705 
9706         /* The full idle balance loop has been done */
9707         ret = true;
9708 
9709 abort:
9710         /* There is still blocked load, enable periodic update */
9711         if (has_blocked_load)
9712                 WRITE_ONCE(nohz.has_blocked, 1);
9713 
9714         /*
9715          * next_balance will be updated only when there is a need.
9716          * When the CPU is attached to null domain for ex, it will not be
9717          * updated.
9718          */
9719         if (likely(update_next_balance))
9720                 nohz.next_balance = next_balance;
9721 
9722         return ret;
9723 }
9724 
9725 /*
9726  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9727  * rebalancing for all the cpus for whom scheduler ticks are stopped.
9728  */
9729 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9730 {
9731         int this_cpu = this_rq->cpu;
9732         unsigned int flags;
9733 
9734         if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9735                 return false;
9736 
9737         if (idle != CPU_IDLE) {
9738                 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9739                 return false;
9740         }
9741 
9742         /* could be _relaxed() */
9743         flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9744         if (!(flags & NOHZ_KICK_MASK))
9745                 return false;
9746 
9747         _nohz_idle_balance(this_rq, flags, idle);
9748 
9749         return true;
9750 }
9751 
9752 static void nohz_newidle_balance(struct rq *this_rq)
9753 {
9754         int this_cpu = this_rq->cpu;
9755 
9756         /*
9757          * This CPU doesn't want to be disturbed by scheduler
9758          * housekeeping
9759          */
9760         if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9761                 return;
9762 
9763         /* Will wake up very soon. No time for doing anything else*/
9764         if (this_rq->avg_idle < sysctl_sched_migration_cost)
9765                 return;
9766 
9767         /* Don't need to update blocked load of idle CPUs*/
9768         if (!READ_ONCE(nohz.has_blocked) ||
9769             time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9770                 return;
9771 
9772         raw_spin_unlock(&this_rq->lock);
9773         /*
9774          * This CPU is going to be idle and blocked load of idle CPUs
9775          * need to be updated. Run the ilb locally as it is a good
9776          * candidate for ilb instead of waking up another idle CPU.
9777          * Kick an normal ilb if we failed to do the update.
9778          */
9779         if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9780                 kick_ilb(NOHZ_STATS_KICK);
9781         raw_spin_lock(&this_rq->lock);
9782 }
9783 
9784 #else /* !CONFIG_NO_HZ_COMMON */
9785 static inline void nohz_balancer_kick(struct rq *rq) { }
9786 
9787 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9788 {
9789         return false;
9790 }
9791 
9792 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9793 #endif /* CONFIG_NO_HZ_COMMON */
9794 
9795 /*
9796  * idle_balance is called by schedule() if this_cpu is about to become
9797  * idle. Attempts to pull tasks from other CPUs.
9798  */
9799 int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
9800 {
9801         unsigned long next_balance = jiffies + HZ;
9802         int this_cpu = this_rq->cpu;
9803         struct sched_domain *sd;
9804         int pulled_task = 0;
9805         u64 curr_cost = 0;
9806 
9807         update_misfit_status(NULL, this_rq);
9808         /*
9809          * We must set idle_stamp _before_ calling idle_balance(), such that we
9810          * measure the duration of idle_balance() as idle time.
9811          */
9812         this_rq->idle_stamp = rq_clock(this_rq);
9813 
9814         /*
9815          * Do not pull tasks towards !active CPUs...
9816          */
9817         if (!cpu_active(this_cpu))
9818                 return 0;
9819 
9820         /*
9821          * This is OK, because current is on_cpu, which avoids it being picked
9822          * for load-balance and preemption/IRQs are still disabled avoiding
9823          * further scheduler activity on it and we're being very careful to
9824          * re-start the picking loop.
9825          */
9826         rq_unpin_lock(this_rq, rf);
9827 
9828         if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9829             !READ_ONCE(this_rq->rd->overload)) {
9830 
9831                 rcu_read_lock();
9832                 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9833                 if (sd)
9834                         update_next_balance(sd, &next_balance);
9835                 rcu_read_unlock();
9836 
9837                 nohz_newidle_balance(this_rq);
9838 
9839                 goto out;
9840         }
9841 
9842         raw_spin_unlock(&this_rq->lock);
9843 
9844         update_blocked_averages(this_cpu);
9845         rcu_read_lock();
9846         for_each_domain(this_cpu, sd) {
9847                 int continue_balancing = 1;
9848                 u64 t0, domain_cost;
9849 
9850                 if (!(sd->flags & SD_LOAD_BALANCE))
9851                         continue;
9852 
9853                 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9854                         update_next_balance(sd, &next_balance);
9855                         break;
9856                 }
9857 
9858                 if (sd->flags & SD_BALANCE_NEWIDLE) {
9859                         t0 = sched_clock_cpu(this_cpu);
9860 
9861                         pulled_task = load_balance(this_cpu, this_rq,
9862                                                    sd, CPU_NEWLY_IDLE,
9863                                                    &continue_balancing);
9864 
9865                         domain_cost = sched_clock_cpu(this_cpu) - t0;
9866                         if (domain_cost > sd->max_newidle_lb_cost)
9867                                 sd->max_newidle_lb_cost = domain_cost;
9868 
9869                         curr_cost += domain_cost;
9870                 }
9871 
9872                 update_next_balance(sd, &next_balance);
9873 
9874                 /*
9875                  * Stop searching for tasks to pull if there are
9876                  * now runnable tasks on this rq.
9877                  */
9878                 if (pulled_task || this_rq->nr_running > 0)
9879                         break;
9880         }
9881         rcu_read_unlock();
9882 
9883         raw_spin_lock(&this_rq->lock);
9884 
9885         if (curr_cost > this_rq->max_idle_balance_cost)
9886                 this_rq->max_idle_balance_cost = curr_cost;
9887 
9888 out:
9889         /*
9890          * While browsing the domains, we released the rq lock, a task could
9891          * have been enqueued in the meantime. Since we're not going idle,
9892          * pretend we pulled a task.
9893          */
9894         if (this_rq->cfs.h_nr_running && !pulled_task)
9895                 pulled_task = 1;
9896 
9897         /* Move the next balance forward */
9898         if (time_after(this_rq->next_balance, next_balance))
9899                 this_rq->next_balance = next_balance;
9900 
9901         /* Is there a task of a high priority class? */
9902         if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9903                 pulled_task = -1;
9904 
9905         if (pulled_task)
9906                 this_rq->idle_stamp = 0;
9907 
9908         rq_repin_lock(this_rq, rf);
9909 
9910         return pulled_task;
9911 }
9912 
9913 /*
9914  * run_rebalance_domains is triggered when needed from the scheduler tick.
9915  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9916  */
9917 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9918 {
9919         struct rq *this_rq = this_rq();
9920         enum cpu_idle_type idle = this_rq->idle_balance ?
9921                                                 CPU_IDLE : CPU_NOT_IDLE;
9922 
9923         /*
9924          * If this CPU has a pending nohz_balance_kick, then do the
9925          * balancing on behalf of the other idle CPUs whose ticks are
9926          * stopped. Do nohz_idle_balance *before* rebalance_domains to
9927          * give the idle CPUs a chance to load balance. Else we may
9928          * load balance only within the local sched_domain hierarchy
9929          * and abort nohz_idle_balance altogether if we pull some load.
9930          */
9931         if (nohz_idle_balance(this_rq, idle))
9932                 return;
9933 
9934         /* normal load balance */
9935         update_blocked_averages(this_rq->cpu);
9936         rebalance_domains(this_rq, idle);
9937 }
9938 
9939 /*
9940  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9941  */
9942 void trigger_load_balance(struct rq *rq)
9943 {
9944         /* Don't need to rebalance while attached to NULL domain */
9945         if (unlikely(on_null_domain(rq)))
9946                 return;
9947 
9948         if (time_after_eq(jiffies, rq->next_balance))
9949                 raise_softirq(SCHED_SOFTIRQ);
9950 
9951         nohz_balancer_kick(rq);
9952 }
9953 
9954 static void rq_online_fair(struct rq *rq)
9955 {
9956         update_sysctl();
9957 
9958         update_runtime_enabled(rq);
9959 }
9960 
9961 static void rq_offline_fair(struct rq *rq)
9962 {
9963         update_sysctl();
9964 
9965         /* Ensure any throttled groups are reachable by pick_next_task */
9966         unthrottle_offline_cfs_rqs(rq);
9967 }
9968 
9969 #endif /* CONFIG_SMP */
9970 
9971 /*
9972  * scheduler tick hitting a task of our scheduling class.
9973  *
9974  * NOTE: This function can be called remotely by the tick offload that
9975  * goes along full dynticks. Therefore no local assumption can be made
9976  * and everything must be accessed through the @rq and @curr passed in
9977  * parameters.
9978  */
9979 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9980 {
9981         struct cfs_rq *cfs_rq;
9982         struct sched_entity *se = &curr->se;
9983 
9984         for_each_sched_entity(se) {
9985                 cfs_rq = cfs_rq_of(se);
9986                 entity_tick(cfs_rq, se, queued);
9987         }
9988 
9989         if (static_branch_unlikely(&sched_numa_balancing))
9990                 task_tick_numa(rq, curr);
9991 
9992         update_misfit_status(curr, rq);
9993         update_overutilized_status(task_rq(curr));
9994 }
9995 
9996 /*
9997  * called on fork with the child task as argument from the parent's context
9998  *  - child not yet on the tasklist
9999  *  - preemption disabled
10000  */
10001 static void task_fork_fair(struct task_struct *p)
10002 {
10003         struct cfs_rq *cfs_rq;
10004         struct sched_entity *se = &p->se, *curr;
10005         struct rq *rq = this_rq();
10006         struct rq_flags rf;
10007 
10008         rq_lock(rq, &rf);
10009         update_rq_clock(rq);
10010 
10011         cfs_rq = task_cfs_rq(current);
10012         curr = cfs_rq->curr;
10013         if (curr) {
10014                 update_curr(cfs_rq);
10015                 se->vruntime = curr->vruntime;
10016         }
10017         place_entity(cfs_rq, se, 1);
10018 
10019         if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10020                 /*
10021                  * Upon rescheduling, sched_class::put_prev_task() will place
10022                  * 'current' within the tree based on its new key value.
10023                  */
10024                 swap(curr->vruntime, se->vruntime);
10025                 resched_curr(rq);
10026         }
10027 
10028         se->vruntime -= cfs_rq->min_vruntime;
10029         rq_unlock(rq, &rf);
10030 }
10031 
10032 /*
10033  * Priority of the task has changed. Check to see if we preempt
10034  * the current task.
10035  */
10036 static void
10037 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10038 {
10039         if (!task_on_rq_queued(p))
10040                 return;
10041 
10042         /*
10043          * Reschedule if we are currently running on this runqueue and
10044          * our priority decreased, or if we are not currently running on
10045          * this runqueue and our priority is higher than the current's
10046          */
10047         if (rq->curr == p) {
10048                 if (p->prio > oldprio)
10049                         resched_curr(rq);
10050         } else
10051                 check_preempt_curr(rq, p, 0);
10052 }
10053 
10054 static inline bool vruntime_normalized(struct task_struct *p)
10055 {
10056         struct sched_entity *se = &p->se;
10057 
10058         /*
10059          * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10060          * the dequeue_entity(.flags=0) will already have normalized the
10061          * vruntime.
10062          */
10063         if (p->on_rq)
10064                 return true;
10065 
10066         /*
10067          * When !on_rq, vruntime of the task has usually NOT been normalized.
10068          * But there are some cases where it has already been normalized:
10069          *
10070          * - A forked child which is waiting for being woken up by
10071          *   wake_up_new_task().
10072          * - A task which has been woken up by try_to_wake_up() and
10073          *   waiting for actually being woken up by sched_ttwu_pending().
10074          */
10075         if (!se->sum_exec_runtime ||
10076             (p->state == TASK_WAKING && p->sched_remote_wakeup))
10077                 return true;
10078 
10079         return false;
10080 }
10081 
10082 #ifdef CONFIG_FAIR_GROUP_SCHED
10083 /*
10084  * Propagate the changes of the sched_entity across the tg tree to make it
10085  * visible to the root
10086  */
10087 static void propagate_entity_cfs_rq(struct sched_entity *se)
10088 {
10089         struct cfs_rq *cfs_rq;
10090 
10091         /* Start to propagate at parent */
10092         se = se->parent;
10093 
10094         for_each_sched_entity(se) {
10095                 cfs_rq = cfs_rq_of(se);
10096 
10097                 if (cfs_rq_throttled(cfs_rq))
10098                         break;
10099 
10100                 update_load_avg(cfs_rq, se, UPDATE_TG);
10101         }
10102 }
10103 #else
10104 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10105 #endif
10106 
10107 static void detach_entity_cfs_rq(struct sched_entity *se)
10108 {
10109         struct cfs_rq *cfs_rq = cfs_rq_of(se);
10110 
10111         /* Catch up with the cfs_rq and remove our load when we leave */
10112         update_load_avg(cfs_rq, se, 0);
10113         detach_entity_load_avg(cfs_rq, se);
10114         update_tg_load_avg(cfs_rq, false);
10115         propagate_entity_cfs_rq(se);
10116 }
10117 
10118 static void attach_entity_cfs_rq(struct sched_entity *se)
10119 {
10120         struct cfs_rq *cfs_rq = cfs_rq_of(se);
10121 
10122 #ifdef CONFIG_FAIR_GROUP_SCHED
10123         /*
10124          * Since the real-depth could have been changed (only FAIR
10125          * class maintain depth value), reset depth properly.
10126          */
10127         se->depth = se->parent ? se->parent->depth + 1 : 0;
10128 #endif
10129 
10130         /* Synchronize entity with its cfs_rq */
10131         update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10132         attach_entity_load_avg(cfs_rq, se, 0);
10133         update_tg_load_avg(cfs_rq, false);
10134         propagate_entity_cfs_rq(se);
10135 }
10136 
10137 static void detach_task_cfs_rq(struct task_struct *p)
10138 {
10139         struct sched_entity *se = &p->se;
10140         struct cfs_rq *cfs_rq = cfs_rq_of(se);
10141 
10142         if (!vruntime_normalized(p)) {
10143                 /*
10144                  * Fix up our vruntime so that the current sleep doesn't
10145                  * cause 'unlimited' sleep bonus.
10146                  */
10147                 place_entity(cfs_rq, se, 0);
10148                 se->vruntime -= cfs_rq->min_vruntime;
10149         }
10150 
10151         detach_entity_cfs_rq(se);
10152 }
10153 
10154 static void attach_task_cfs_rq(struct task_struct *p)
10155 {
10156         struct sched_entity *se = &p->se;
10157         struct cfs_rq *cfs_rq = cfs_rq_of(se);
10158 
10159         attach_entity_cfs_rq(se);
10160 
10161         if (!vruntime_normalized(p))
10162                 se->vruntime += cfs_rq->min_vruntime;
10163 }
10164 
10165 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10166 {
10167         detach_task_cfs_rq(p);
10168 }
10169 
10170 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10171 {
10172         attach_task_cfs_rq(p);
10173 
10174         if (task_on_rq_queued(p)) {
10175                 /*
10176                  * We were most likely switched from sched_rt, so
10177                  * kick off the schedule if running, otherwise just see
10178                  * if we can still preempt the current task.
10179                  */
10180                 if (rq->curr == p)
10181                         resched_curr(rq);
10182                 else
10183                         check_preempt_curr(rq, p, 0);
10184         }
10185 }
10186 
10187 /* Account for a task changing its policy or group.
10188  *
10189  * This routine is mostly called to set cfs_rq->curr field when a task
10190  * migrates between groups/classes.
10191  */
10192 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
10193 {
10194         struct sched_entity *se = &p->se;
10195 
10196 #ifdef CONFIG_SMP
10197         if (task_on_rq_queued(p)) {
10198                 /*
10199                  * Move the next running task to the front of the list, so our
10200                  * cfs_tasks list becomes MRU one.
10201                  */
10202                 list_move(&se->group_node, &rq->cfs_tasks);
10203         }
10204 #endif
10205 
10206         for_each_sched_entity(se) {
10207                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10208 
10209                 set_next_entity(cfs_rq, se);
10210                 /* ensure bandwidth has been allocated on our new cfs_rq */
10211                 account_cfs_rq_runtime(cfs_rq, 0);
10212         }
10213 }
10214 
10215 void init_cfs_rq(struct cfs_rq *cfs_rq)
10216 {
10217         cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10218         cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10219 #ifndef CONFIG_64BIT
10220         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10221 #endif
10222 #ifdef CONFIG_SMP
10223         raw_spin_lock_init(&cfs_rq->removed.lock);
10224 #endif
10225 }
10226 
10227 #ifdef CONFIG_FAIR_GROUP_SCHED
10228 static void task_set_group_fair(struct task_struct *p)
10229 {
10230         struct sched_entity *se = &p->se;
10231 
10232         set_task_rq(p, task_cpu(p));
10233         se->depth = se->parent ? se->parent->depth + 1 : 0;
10234 }
10235 
10236 static void task_move_group_fair(struct task_struct *p)
10237 {
10238         detach_task_cfs_rq(p);
10239         set_task_rq(p, task_cpu(p));
10240 
10241 #ifdef CONFIG_SMP
10242         /* Tell se's cfs_rq has been changed -- migrated */
10243         p->se.avg.last_update_time = 0;
10244 #endif
10245         attach_task_cfs_rq(p);
10246 }
10247 
10248 static void task_change_group_fair(struct task_struct *p, int type)
10249 {
10250         switch (type) {
10251         case TASK_SET_GROUP:
10252                 task_set_group_fair(p);
10253                 break;
10254 
10255         case TASK_MOVE_GROUP:
10256                 task_move_group_fair(p);
10257                 break;
10258         }
10259 }
10260 
10261 void free_fair_sched_group(struct task_group *tg)
10262 {
10263         int i;
10264 
10265         destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10266 
10267         for_each_possible_cpu(i) {
10268                 if (tg->cfs_rq)
10269                         kfree(tg->cfs_rq[i]);
10270                 if (tg->se)
10271                         kfree(tg->se[i]);
10272         }
10273 
10274         kfree(tg->cfs_rq);
10275         kfree(tg->se);
10276 }
10277 
10278 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10279 {
10280         struct sched_entity *se;
10281         struct cfs_rq *cfs_rq;
10282         int i;
10283 
10284         tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10285         if (!tg->cfs_rq)
10286                 goto err;
10287         tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10288         if (!tg->se)
10289                 goto err;
10290 
10291         tg->shares = NICE_0_LOAD;
10292 
10293         init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10294 
10295         for_each_possible_cpu(i) {
10296                 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10297                                       GFP_KERNEL, cpu_to_node(i));
10298                 if (!cfs_rq)
10299                         goto err;
10300 
10301                 se = kzalloc_node(sizeof(struct sched_entity),
10302                                   GFP_KERNEL, cpu_to_node(i));
10303                 if (!se)
10304                         goto err_free_rq;
10305 
10306                 init_cfs_rq(cfs_rq);
10307                 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10308                 init_entity_runnable_average(se);
10309         }
10310 
10311         return 1;
10312 
10313 err_free_rq:
10314         kfree(cfs_rq);
10315 err:
10316         return 0;
10317 }
10318 
10319 void online_fair_sched_group(struct task_group *tg)
10320 {
10321         struct sched_entity *se;
10322         struct rq_flags rf;
10323         struct rq *rq;
10324         int i;
10325 
10326         for_each_possible_cpu(i) {
10327                 rq = cpu_rq(i);
10328                 se = tg->se[i];
10329                 rq_lock_irq(rq, &rf);
10330                 update_rq_clock(rq);
10331                 attach_entity_cfs_rq(se);
10332                 sync_throttle(tg, i);
10333                 rq_unlock_irq(rq, &rf);
10334         }
10335 }
10336 
10337 void unregister_fair_sched_group(struct task_group *tg)
10338 {
10339         unsigned long flags;
10340         struct rq *rq;
10341         int cpu;
10342 
10343         for_each_possible_cpu(cpu) {
10344                 if (tg->se[cpu])
10345                         remove_entity_load_avg(tg->se[cpu]);
10346 
10347                 /*
10348                  * Only empty task groups can be destroyed; so we can speculatively
10349                  * check on_list without danger of it being re-added.
10350                  */
10351                 if (!tg->cfs_rq[cpu]->on_list)
10352                         continue;
10353 
10354                 rq = cpu_rq(cpu);
10355 
10356                 raw_spin_lock_irqsave(&rq->lock, flags);
10357                 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10358                 raw_spin_unlock_irqrestore(&rq->lock, flags);
10359         }
10360 }
10361 
10362 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10363                         struct sched_entity *se, int cpu,
10364                         struct sched_entity *parent)
10365 {
10366         struct rq *rq = cpu_rq(cpu);
10367 
10368         cfs_rq->tg = tg;
10369         cfs_rq->rq = rq;
10370         init_cfs_rq_runtime(cfs_rq);
10371 
10372         tg->cfs_rq[cpu] = cfs_rq;
10373         tg->se[cpu] = se;
10374 
10375         /* se could be NULL for root_task_group */
10376         if (!se)
10377                 return;
10378 
10379         if (!parent) {
10380                 se->cfs_rq = &rq->cfs;
10381                 se->depth = 0;
10382         } else {
10383                 se->cfs_rq = parent->my_q;
10384                 se->depth = parent->depth + 1;
10385         }
10386 
10387         se->my_q = cfs_rq;
10388         /* guarantee group entities always have weight */
10389         update_load_set(&se->load, NICE_0_LOAD);
10390         se->parent = parent;
10391 }
10392 
10393 static DEFINE_MUTEX(shares_mutex);
10394 
10395 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10396 {
10397         int i;
10398 
10399         /*
10400          * We can't change the weight of the root cgroup.
10401          */
10402         if (!tg->se[0])
10403                 return -EINVAL;
10404 
10405         shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10406 
10407         mutex_lock(&shares_mutex);
10408         if (tg->shares == shares)
10409                 goto done;
10410 
10411         tg->shares = shares;
10412         for_each_possible_cpu(i) {
10413                 struct rq *rq = cpu_rq(i);
10414                 struct sched_entity *se = tg->se[i];
10415                 struct rq_flags rf;
10416 
10417                 /* Propagate contribution to hierarchy */
10418                 rq_lock_irqsave(rq, &rf);
10419                 update_rq_clock(rq);
10420                 for_each_sched_entity(se) {
10421                         update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10422                         update_cfs_group(se);
10423                 }
10424                 rq_unlock_irqrestore(rq, &rf);
10425         }
10426 
10427 done:
10428         mutex_unlock(&shares_mutex);
10429         return 0;
10430 }
10431 #else /* CONFIG_FAIR_GROUP_SCHED */
10432 
10433 void free_fair_sched_group(struct task_group *tg) { }
10434 
10435 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10436 {
10437         return 1;
10438 }
10439 
10440 void online_fair_sched_group(struct task_group *tg) { }
10441 
10442 void unregister_fair_sched_group(struct task_group *tg) { }
10443 
10444 #endif /* CONFIG_FAIR_GROUP_SCHED */
10445 
10446 
10447 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10448 {
10449         struct sched_entity *se = &task->se;
10450         unsigned int rr_interval = 0;
10451 
10452         /*
10453          * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10454          * idle runqueue:
10455          */
10456         if (rq->cfs.load.weight)
10457                 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10458 
10459         return rr_interval;
10460 }
10461 
10462 /*
10463  * All the scheduling class methods:
10464  */
10465 const struct sched_class fair_sched_class = {
10466         .next                   = &idle_sched_class,
10467         .enqueue_task           = enqueue_task_fair,
10468         .dequeue_task           = dequeue_task_fair,
10469         .yield_task             = yield_task_fair,
10470         .yield_to_task          = yield_to_task_fair,
10471 
10472         .check_preempt_curr     = check_preempt_wakeup,
10473 
10474         .pick_next_task         = pick_next_task_fair,
10475         .put_prev_task          = put_prev_task_fair,
10476         .set_next_task          = set_next_task_fair,
10477 
10478 #ifdef CONFIG_SMP
10479         .balance                = balance_fair,
10480         .select_task_rq         = select_task_rq_fair,
10481         .migrate_task_rq        = migrate_task_rq_fair,
10482 
10483         .rq_online              = rq_online_fair,
10484         .rq_offline             = rq_offline_fair,
10485 
10486         .task_dead              = task_dead_fair,
10487         .set_cpus_allowed       = set_cpus_allowed_common,
10488 #endif
10489 
10490         .task_tick              = task_tick_fair,
10491         .task_fork              = task_fork_fair,
10492 
10493         .prio_changed           = prio_changed_fair,
10494         .switched_from          = switched_from_fair,
10495         .switched_to            = switched_to_fair,
10496 
10497         .get_rr_interval        = get_rr_interval_fair,
10498 
10499         .update_curr            = update_curr_fair,
10500 
10501 #ifdef CONFIG_FAIR_GROUP_SCHED
10502         .task_change_group      = task_change_group_fair,
10503 #endif
10504 
10505 #ifdef CONFIG_UCLAMP_TASK
10506         .uclamp_enabled         = 1,
10507 #endif
10508 };
10509 
10510 #ifdef CONFIG_SCHED_DEBUG
10511 void print_cfs_stats(struct seq_file *m, int cpu)
10512 {
10513         struct cfs_rq *cfs_rq, *pos;
10514 
10515         rcu_read_lock();
10516         for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10517                 print_cfs_rq(m, cpu, cfs_rq);
10518         rcu_read_unlock();
10519 }
10520 
10521 #ifdef CONFIG_NUMA_BALANCING
10522 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10523 {
10524         int node;
10525         unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10526         struct numa_group *ng;
10527 
10528         rcu_read_lock();
10529         ng = rcu_dereference(p->numa_group);
10530         for_each_online_node(node) {
10531                 if (p->numa_faults) {
10532                         tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10533                         tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10534                 }
10535                 if (ng) {
10536                         gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10537                         gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10538                 }
10539                 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10540         }
10541         rcu_read_unlock();
10542 }
10543 #endif /* CONFIG_NUMA_BALANCING */
10544 #endif /* CONFIG_SCHED_DEBUG */
10545 
10546 __init void init_sched_fair_class(void)
10547 {
10548 #ifdef CONFIG_SMP
10549         open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10550 
10551 #ifdef CONFIG_NO_HZ_COMMON
10552         nohz.next_balance = jiffies;
10553         nohz.next_blocked = jiffies;
10554         zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10555 #endif
10556 #endif /* SMP */
10557 
10558 }
10559 
10560 /*
10561  * Helper functions to facilitate extracting info from tracepoints.
10562  */
10563 
10564 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
10565 {
10566 #ifdef CONFIG_SMP
10567         return cfs_rq ? &cfs_rq->avg : NULL;
10568 #else
10569         return NULL;
10570 #endif
10571 }
10572 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
10573 
10574 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
10575 {
10576         if (!cfs_rq) {
10577                 if (str)
10578                         strlcpy(str, "(null)", len);
10579                 else
10580                         return NULL;
10581         }
10582 
10583         cfs_rq_tg_path(cfs_rq, str, len);
10584         return str;
10585 }
10586 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
10587 
10588 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
10589 {
10590         return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
10591 }
10592 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
10593 
10594 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
10595 {
10596 #ifdef CONFIG_SMP
10597         return rq ? &rq->avg_rt : NULL;
10598 #else
10599         return NULL;
10600 #endif
10601 }
10602 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
10603 
10604 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
10605 {
10606 #ifdef CONFIG_SMP
10607         return rq ? &rq->avg_dl : NULL;
10608 #else
10609         return NULL;
10610 #endif
10611 }
10612 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
10613 
10614 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
10615 {
10616 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
10617         return rq ? &rq->avg_irq : NULL;
10618 #else
10619         return NULL;
10620 #endif
10621 }
10622 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
10623 
10624 int sched_trace_rq_cpu(struct rq *rq)
10625 {
10626         return rq ? cpu_of(rq) : -1;
10627 }
10628 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
10629 
10630 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
10631 {
10632 #ifdef CONFIG_SMP
10633         return rd ? rd->span : NULL;
10634 #else
10635         return NULL;
10636 #endif
10637 }
10638 EXPORT_SYMBOL_GPL(sched_trace_rd_span);