Loading...
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 * Scheduler internal types and methods:
4 */
5#ifndef _KERNEL_SCHED_SCHED_H
6#define _KERNEL_SCHED_SCHED_H
7
8#include <linux/sched/affinity.h>
9#include <linux/sched/autogroup.h>
10#include <linux/sched/cpufreq.h>
11#include <linux/sched/deadline.h>
12#include <linux/sched.h>
13#include <linux/sched/loadavg.h>
14#include <linux/sched/mm.h>
15#include <linux/sched/rseq_api.h>
16#include <linux/sched/signal.h>
17#include <linux/sched/smt.h>
18#include <linux/sched/stat.h>
19#include <linux/sched/sysctl.h>
20#include <linux/sched/task_flags.h>
21#include <linux/sched/task.h>
22#include <linux/sched/topology.h>
23
24#include <linux/atomic.h>
25#include <linux/bitmap.h>
26#include <linux/bug.h>
27#include <linux/capability.h>
28#include <linux/cgroup_api.h>
29#include <linux/cgroup.h>
30#include <linux/context_tracking.h>
31#include <linux/cpufreq.h>
32#include <linux/cpumask_api.h>
33#include <linux/ctype.h>
34#include <linux/file.h>
35#include <linux/fs_api.h>
36#include <linux/hrtimer_api.h>
37#include <linux/interrupt.h>
38#include <linux/irq_work.h>
39#include <linux/jiffies.h>
40#include <linux/kref_api.h>
41#include <linux/kthread.h>
42#include <linux/ktime_api.h>
43#include <linux/lockdep_api.h>
44#include <linux/lockdep.h>
45#include <linux/minmax.h>
46#include <linux/mm.h>
47#include <linux/module.h>
48#include <linux/mutex_api.h>
49#include <linux/plist.h>
50#include <linux/poll.h>
51#include <linux/proc_fs.h>
52#include <linux/profile.h>
53#include <linux/psi.h>
54#include <linux/rcupdate.h>
55#include <linux/seq_file.h>
56#include <linux/seqlock.h>
57#include <linux/softirq.h>
58#include <linux/spinlock_api.h>
59#include <linux/static_key.h>
60#include <linux/stop_machine.h>
61#include <linux/syscalls_api.h>
62#include <linux/syscalls.h>
63#include <linux/tick.h>
64#include <linux/topology.h>
65#include <linux/types.h>
66#include <linux/u64_stats_sync_api.h>
67#include <linux/uaccess.h>
68#include <linux/wait_api.h>
69#include <linux/wait_bit.h>
70#include <linux/workqueue_api.h>
71
72#include <trace/events/power.h>
73#include <trace/events/sched.h>
74
75#include "../workqueue_internal.h"
76
77#ifdef CONFIG_PARAVIRT
78# include <asm/paravirt.h>
79# include <asm/paravirt_api_clock.h>
80#endif
81
82#include "cpupri.h"
83#include "cpudeadline.h"
84
85#ifdef CONFIG_SCHED_DEBUG
86# define SCHED_WARN_ON(x) WARN_ONCE(x, #x)
87#else
88# define SCHED_WARN_ON(x) ({ (void)(x), 0; })
89#endif
90
91struct rq;
92struct cpuidle_state;
93
94/* task_struct::on_rq states: */
95#define TASK_ON_RQ_QUEUED 1
96#define TASK_ON_RQ_MIGRATING 2
97
98extern __read_mostly int scheduler_running;
99
100extern unsigned long calc_load_update;
101extern atomic_long_t calc_load_tasks;
102
103extern void calc_global_load_tick(struct rq *this_rq);
104extern long calc_load_fold_active(struct rq *this_rq, long adjust);
105
106extern void call_trace_sched_update_nr_running(struct rq *rq, int count);
107
108extern int sysctl_sched_rt_period;
109extern int sysctl_sched_rt_runtime;
110extern int sched_rr_timeslice;
111
112/*
113 * Helpers for converting nanosecond timing to jiffy resolution
114 */
115#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
116
117/*
118 * Increase resolution of nice-level calculations for 64-bit architectures.
119 * The extra resolution improves shares distribution and load balancing of
120 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
121 * hierarchies, especially on larger systems. This is not a user-visible change
122 * and does not change the user-interface for setting shares/weights.
123 *
124 * We increase resolution only if we have enough bits to allow this increased
125 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
126 * are pretty high and the returns do not justify the increased costs.
127 *
128 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to
129 * increase coverage and consistency always enable it on 64-bit platforms.
130 */
131#ifdef CONFIG_64BIT
132# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT)
133# define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT)
134# define scale_load_down(w) \
135({ \
136 unsigned long __w = (w); \
137 if (__w) \
138 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \
139 __w; \
140})
141#else
142# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT)
143# define scale_load(w) (w)
144# define scale_load_down(w) (w)
145#endif
146
147/*
148 * Task weight (visible to users) and its load (invisible to users) have
149 * independent resolution, but they should be well calibrated. We use
150 * scale_load() and scale_load_down(w) to convert between them. The
151 * following must be true:
152 *
153 * scale_load(sched_prio_to_weight[NICE_TO_PRIO(0)-MAX_RT_PRIO]) == NICE_0_LOAD
154 *
155 */
156#define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT)
157
158/*
159 * Single value that decides SCHED_DEADLINE internal math precision.
160 * 10 -> just above 1us
161 * 9 -> just above 0.5us
162 */
163#define DL_SCALE 10
164
165/*
166 * Single value that denotes runtime == period, ie unlimited time.
167 */
168#define RUNTIME_INF ((u64)~0ULL)
169
170static inline int idle_policy(int policy)
171{
172 return policy == SCHED_IDLE;
173}
174static inline int fair_policy(int policy)
175{
176 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
177}
178
179static inline int rt_policy(int policy)
180{
181 return policy == SCHED_FIFO || policy == SCHED_RR;
182}
183
184static inline int dl_policy(int policy)
185{
186 return policy == SCHED_DEADLINE;
187}
188static inline bool valid_policy(int policy)
189{
190 return idle_policy(policy) || fair_policy(policy) ||
191 rt_policy(policy) || dl_policy(policy);
192}
193
194static inline int task_has_idle_policy(struct task_struct *p)
195{
196 return idle_policy(p->policy);
197}
198
199static inline int task_has_rt_policy(struct task_struct *p)
200{
201 return rt_policy(p->policy);
202}
203
204static inline int task_has_dl_policy(struct task_struct *p)
205{
206 return dl_policy(p->policy);
207}
208
209#define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
210
211static inline void update_avg(u64 *avg, u64 sample)
212{
213 s64 diff = sample - *avg;
214 *avg += diff / 8;
215}
216
217/*
218 * Shifting a value by an exponent greater *or equal* to the size of said value
219 * is UB; cap at size-1.
220 */
221#define shr_bound(val, shift) \
222 (val >> min_t(typeof(shift), shift, BITS_PER_TYPE(typeof(val)) - 1))
223
224/*
225 * !! For sched_setattr_nocheck() (kernel) only !!
226 *
227 * This is actually gross. :(
228 *
229 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE
230 * tasks, but still be able to sleep. We need this on platforms that cannot
231 * atomically change clock frequency. Remove once fast switching will be
232 * available on such platforms.
233 *
234 * SUGOV stands for SchedUtil GOVernor.
235 */
236#define SCHED_FLAG_SUGOV 0x10000000
237
238#define SCHED_DL_FLAGS (SCHED_FLAG_RECLAIM | SCHED_FLAG_DL_OVERRUN | SCHED_FLAG_SUGOV)
239
240static inline bool dl_entity_is_special(const struct sched_dl_entity *dl_se)
241{
242#ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
243 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV);
244#else
245 return false;
246#endif
247}
248
249/*
250 * Tells if entity @a should preempt entity @b.
251 */
252static inline bool dl_entity_preempt(const struct sched_dl_entity *a,
253 const struct sched_dl_entity *b)
254{
255 return dl_entity_is_special(a) ||
256 dl_time_before(a->deadline, b->deadline);
257}
258
259/*
260 * This is the priority-queue data structure of the RT scheduling class:
261 */
262struct rt_prio_array {
263 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
264 struct list_head queue[MAX_RT_PRIO];
265};
266
267struct rt_bandwidth {
268 /* nests inside the rq lock: */
269 raw_spinlock_t rt_runtime_lock;
270 ktime_t rt_period;
271 u64 rt_runtime;
272 struct hrtimer rt_period_timer;
273 unsigned int rt_period_active;
274};
275
276static inline int dl_bandwidth_enabled(void)
277{
278 return sysctl_sched_rt_runtime >= 0;
279}
280
281/*
282 * To keep the bandwidth of -deadline tasks under control
283 * we need some place where:
284 * - store the maximum -deadline bandwidth of each cpu;
285 * - cache the fraction of bandwidth that is currently allocated in
286 * each root domain;
287 *
288 * This is all done in the data structure below. It is similar to the
289 * one used for RT-throttling (rt_bandwidth), with the main difference
290 * that, since here we are only interested in admission control, we
291 * do not decrease any runtime while the group "executes", neither we
292 * need a timer to replenish it.
293 *
294 * With respect to SMP, bandwidth is given on a per root domain basis,
295 * meaning that:
296 * - bw (< 100%) is the deadline bandwidth of each CPU;
297 * - total_bw is the currently allocated bandwidth in each root domain;
298 */
299struct dl_bw {
300 raw_spinlock_t lock;
301 u64 bw;
302 u64 total_bw;
303};
304
305extern void init_dl_bw(struct dl_bw *dl_b);
306extern int sched_dl_global_validate(void);
307extern void sched_dl_do_global(void);
308extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr);
309extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr);
310extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr);
311extern bool __checkparam_dl(const struct sched_attr *attr);
312extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr);
313extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial);
314extern int dl_bw_check_overflow(int cpu);
315
316/*
317 * SCHED_DEADLINE supports servers (nested scheduling) with the following
318 * interface:
319 *
320 * dl_se::rq -- runqueue we belong to.
321 *
322 * dl_se::server_has_tasks() -- used on bandwidth enforcement; we 'stop' the
323 * server when it runs out of tasks to run.
324 *
325 * dl_se::server_pick() -- nested pick_next_task(); we yield the period if this
326 * returns NULL.
327 *
328 * dl_server_update() -- called from update_curr_common(), propagates runtime
329 * to the server.
330 *
331 * dl_server_start()
332 * dl_server_stop() -- start/stop the server when it has (no) tasks.
333 *
334 * dl_server_init() -- initializes the server.
335 */
336extern void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec);
337extern void dl_server_start(struct sched_dl_entity *dl_se);
338extern void dl_server_stop(struct sched_dl_entity *dl_se);
339extern void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
340 dl_server_has_tasks_f has_tasks,
341 dl_server_pick_f pick);
342
343#ifdef CONFIG_CGROUP_SCHED
344
345struct cfs_rq;
346struct rt_rq;
347
348extern struct list_head task_groups;
349
350struct cfs_bandwidth {
351#ifdef CONFIG_CFS_BANDWIDTH
352 raw_spinlock_t lock;
353 ktime_t period;
354 u64 quota;
355 u64 runtime;
356 u64 burst;
357 u64 runtime_snap;
358 s64 hierarchical_quota;
359
360 u8 idle;
361 u8 period_active;
362 u8 slack_started;
363 struct hrtimer period_timer;
364 struct hrtimer slack_timer;
365 struct list_head throttled_cfs_rq;
366
367 /* Statistics: */
368 int nr_periods;
369 int nr_throttled;
370 int nr_burst;
371 u64 throttled_time;
372 u64 burst_time;
373#endif
374};
375
376/* Task group related information */
377struct task_group {
378 struct cgroup_subsys_state css;
379
380#ifdef CONFIG_FAIR_GROUP_SCHED
381 /* schedulable entities of this group on each CPU */
382 struct sched_entity **se;
383 /* runqueue "owned" by this group on each CPU */
384 struct cfs_rq **cfs_rq;
385 unsigned long shares;
386
387 /* A positive value indicates that this is a SCHED_IDLE group. */
388 int idle;
389
390#ifdef CONFIG_SMP
391 /*
392 * load_avg can be heavily contended at clock tick time, so put
393 * it in its own cacheline separated from the fields above which
394 * will also be accessed at each tick.
395 */
396 atomic_long_t load_avg ____cacheline_aligned;
397#endif
398#endif
399
400#ifdef CONFIG_RT_GROUP_SCHED
401 struct sched_rt_entity **rt_se;
402 struct rt_rq **rt_rq;
403
404 struct rt_bandwidth rt_bandwidth;
405#endif
406
407 struct rcu_head rcu;
408 struct list_head list;
409
410 struct task_group *parent;
411 struct list_head siblings;
412 struct list_head children;
413
414#ifdef CONFIG_SCHED_AUTOGROUP
415 struct autogroup *autogroup;
416#endif
417
418 struct cfs_bandwidth cfs_bandwidth;
419
420#ifdef CONFIG_UCLAMP_TASK_GROUP
421 /* The two decimal precision [%] value requested from user-space */
422 unsigned int uclamp_pct[UCLAMP_CNT];
423 /* Clamp values requested for a task group */
424 struct uclamp_se uclamp_req[UCLAMP_CNT];
425 /* Effective clamp values used for a task group */
426 struct uclamp_se uclamp[UCLAMP_CNT];
427#endif
428
429};
430
431#ifdef CONFIG_FAIR_GROUP_SCHED
432#define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
433
434/*
435 * A weight of 0 or 1 can cause arithmetics problems.
436 * A weight of a cfs_rq is the sum of weights of which entities
437 * are queued on this cfs_rq, so a weight of a entity should not be
438 * too large, so as the shares value of a task group.
439 * (The default weight is 1024 - so there's no practical
440 * limitation from this.)
441 */
442#define MIN_SHARES (1UL << 1)
443#define MAX_SHARES (1UL << 18)
444#endif
445
446typedef int (*tg_visitor)(struct task_group *, void *);
447
448extern int walk_tg_tree_from(struct task_group *from,
449 tg_visitor down, tg_visitor up, void *data);
450
451/*
452 * Iterate the full tree, calling @down when first entering a node and @up when
453 * leaving it for the final time.
454 *
455 * Caller must hold rcu_lock or sufficient equivalent.
456 */
457static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
458{
459 return walk_tg_tree_from(&root_task_group, down, up, data);
460}
461
462extern int tg_nop(struct task_group *tg, void *data);
463
464#ifdef CONFIG_FAIR_GROUP_SCHED
465extern void free_fair_sched_group(struct task_group *tg);
466extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
467extern void online_fair_sched_group(struct task_group *tg);
468extern void unregister_fair_sched_group(struct task_group *tg);
469#else
470static inline void free_fair_sched_group(struct task_group *tg) { }
471static inline int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
472{
473 return 1;
474}
475static inline void online_fair_sched_group(struct task_group *tg) { }
476static inline void unregister_fair_sched_group(struct task_group *tg) { }
477#endif
478
479extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
480 struct sched_entity *se, int cpu,
481 struct sched_entity *parent);
482extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent);
483
484extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
485extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
486extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
487extern bool cfs_task_bw_constrained(struct task_struct *p);
488
489extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
490 struct sched_rt_entity *rt_se, int cpu,
491 struct sched_rt_entity *parent);
492extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us);
493extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us);
494extern long sched_group_rt_runtime(struct task_group *tg);
495extern long sched_group_rt_period(struct task_group *tg);
496extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
497
498extern struct task_group *sched_create_group(struct task_group *parent);
499extern void sched_online_group(struct task_group *tg,
500 struct task_group *parent);
501extern void sched_destroy_group(struct task_group *tg);
502extern void sched_release_group(struct task_group *tg);
503
504extern void sched_move_task(struct task_struct *tsk);
505
506#ifdef CONFIG_FAIR_GROUP_SCHED
507extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
508
509extern int sched_group_set_idle(struct task_group *tg, long idle);
510
511#ifdef CONFIG_SMP
512extern void set_task_rq_fair(struct sched_entity *se,
513 struct cfs_rq *prev, struct cfs_rq *next);
514#else /* !CONFIG_SMP */
515static inline void set_task_rq_fair(struct sched_entity *se,
516 struct cfs_rq *prev, struct cfs_rq *next) { }
517#endif /* CONFIG_SMP */
518#endif /* CONFIG_FAIR_GROUP_SCHED */
519
520#else /* CONFIG_CGROUP_SCHED */
521
522struct cfs_bandwidth { };
523static inline bool cfs_task_bw_constrained(struct task_struct *p) { return false; }
524
525#endif /* CONFIG_CGROUP_SCHED */
526
527extern void unregister_rt_sched_group(struct task_group *tg);
528extern void free_rt_sched_group(struct task_group *tg);
529extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
530
531/*
532 * u64_u32_load/u64_u32_store
533 *
534 * Use a copy of a u64 value to protect against data race. This is only
535 * applicable for 32-bits architectures.
536 */
537#ifdef CONFIG_64BIT
538# define u64_u32_load_copy(var, copy) var
539# define u64_u32_store_copy(var, copy, val) (var = val)
540#else
541# define u64_u32_load_copy(var, copy) \
542({ \
543 u64 __val, __val_copy; \
544 do { \
545 __val_copy = copy; \
546 /* \
547 * paired with u64_u32_store_copy(), ordering access \
548 * to var and copy. \
549 */ \
550 smp_rmb(); \
551 __val = var; \
552 } while (__val != __val_copy); \
553 __val; \
554})
555# define u64_u32_store_copy(var, copy, val) \
556do { \
557 typeof(val) __val = (val); \
558 var = __val; \
559 /* \
560 * paired with u64_u32_load_copy(), ordering access to var and \
561 * copy. \
562 */ \
563 smp_wmb(); \
564 copy = __val; \
565} while (0)
566#endif
567# define u64_u32_load(var) u64_u32_load_copy(var, var##_copy)
568# define u64_u32_store(var, val) u64_u32_store_copy(var, var##_copy, val)
569
570/* CFS-related fields in a runqueue */
571struct cfs_rq {
572 struct load_weight load;
573 unsigned int nr_running;
574 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */
575 unsigned int idle_nr_running; /* SCHED_IDLE */
576 unsigned int idle_h_nr_running; /* SCHED_IDLE */
577
578 s64 avg_vruntime;
579 u64 avg_load;
580
581 u64 exec_clock;
582 u64 min_vruntime;
583#ifdef CONFIG_SCHED_CORE
584 unsigned int forceidle_seq;
585 u64 min_vruntime_fi;
586#endif
587
588#ifndef CONFIG_64BIT
589 u64 min_vruntime_copy;
590#endif
591
592 struct rb_root_cached tasks_timeline;
593
594 /*
595 * 'curr' points to currently running entity on this cfs_rq.
596 * It is set to NULL otherwise (i.e when none are currently running).
597 */
598 struct sched_entity *curr;
599 struct sched_entity *next;
600
601#ifdef CONFIG_SCHED_DEBUG
602 unsigned int nr_spread_over;
603#endif
604
605#ifdef CONFIG_SMP
606 /*
607 * CFS load tracking
608 */
609 struct sched_avg avg;
610#ifndef CONFIG_64BIT
611 u64 last_update_time_copy;
612#endif
613 struct {
614 raw_spinlock_t lock ____cacheline_aligned;
615 int nr;
616 unsigned long load_avg;
617 unsigned long util_avg;
618 unsigned long runnable_avg;
619 } removed;
620
621#ifdef CONFIG_FAIR_GROUP_SCHED
622 u64 last_update_tg_load_avg;
623 unsigned long tg_load_avg_contrib;
624 long propagate;
625 long prop_runnable_sum;
626
627 /*
628 * h_load = weight * f(tg)
629 *
630 * Where f(tg) is the recursive weight fraction assigned to
631 * this group.
632 */
633 unsigned long h_load;
634 u64 last_h_load_update;
635 struct sched_entity *h_load_next;
636#endif /* CONFIG_FAIR_GROUP_SCHED */
637#endif /* CONFIG_SMP */
638
639#ifdef CONFIG_FAIR_GROUP_SCHED
640 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */
641
642 /*
643 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
644 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
645 * (like users, containers etc.)
646 *
647 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU.
648 * This list is used during load balance.
649 */
650 int on_list;
651 struct list_head leaf_cfs_rq_list;
652 struct task_group *tg; /* group that "owns" this runqueue */
653
654 /* Locally cached copy of our task_group's idle value */
655 int idle;
656
657#ifdef CONFIG_CFS_BANDWIDTH
658 int runtime_enabled;
659 s64 runtime_remaining;
660
661 u64 throttled_pelt_idle;
662#ifndef CONFIG_64BIT
663 u64 throttled_pelt_idle_copy;
664#endif
665 u64 throttled_clock;
666 u64 throttled_clock_pelt;
667 u64 throttled_clock_pelt_time;
668 u64 throttled_clock_self;
669 u64 throttled_clock_self_time;
670 int throttled;
671 int throttle_count;
672 struct list_head throttled_list;
673 struct list_head throttled_csd_list;
674#endif /* CONFIG_CFS_BANDWIDTH */
675#endif /* CONFIG_FAIR_GROUP_SCHED */
676};
677
678static inline int rt_bandwidth_enabled(void)
679{
680 return sysctl_sched_rt_runtime >= 0;
681}
682
683/* RT IPI pull logic requires IRQ_WORK */
684#if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP)
685# define HAVE_RT_PUSH_IPI
686#endif
687
688/* Real-Time classes' related field in a runqueue: */
689struct rt_rq {
690 struct rt_prio_array active;
691 unsigned int rt_nr_running;
692 unsigned int rr_nr_running;
693#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
694 struct {
695 int curr; /* highest queued rt task prio */
696#ifdef CONFIG_SMP
697 int next; /* next highest */
698#endif
699 } highest_prio;
700#endif
701#ifdef CONFIG_SMP
702 int overloaded;
703 struct plist_head pushable_tasks;
704
705#endif /* CONFIG_SMP */
706 int rt_queued;
707
708 int rt_throttled;
709 u64 rt_time;
710 u64 rt_runtime;
711 /* Nests inside the rq lock: */
712 raw_spinlock_t rt_runtime_lock;
713
714#ifdef CONFIG_RT_GROUP_SCHED
715 unsigned int rt_nr_boosted;
716
717 struct rq *rq;
718 struct task_group *tg;
719#endif
720};
721
722static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq)
723{
724 return rt_rq->rt_queued && rt_rq->rt_nr_running;
725}
726
727/* Deadline class' related fields in a runqueue */
728struct dl_rq {
729 /* runqueue is an rbtree, ordered by deadline */
730 struct rb_root_cached root;
731
732 unsigned int dl_nr_running;
733
734#ifdef CONFIG_SMP
735 /*
736 * Deadline values of the currently executing and the
737 * earliest ready task on this rq. Caching these facilitates
738 * the decision whether or not a ready but not running task
739 * should migrate somewhere else.
740 */
741 struct {
742 u64 curr;
743 u64 next;
744 } earliest_dl;
745
746 int overloaded;
747
748 /*
749 * Tasks on this rq that can be pushed away. They are kept in
750 * an rb-tree, ordered by tasks' deadlines, with caching
751 * of the leftmost (earliest deadline) element.
752 */
753 struct rb_root_cached pushable_dl_tasks_root;
754#else
755 struct dl_bw dl_bw;
756#endif
757 /*
758 * "Active utilization" for this runqueue: increased when a
759 * task wakes up (becomes TASK_RUNNING) and decreased when a
760 * task blocks
761 */
762 u64 running_bw;
763
764 /*
765 * Utilization of the tasks "assigned" to this runqueue (including
766 * the tasks that are in runqueue and the tasks that executed on this
767 * CPU and blocked). Increased when a task moves to this runqueue, and
768 * decreased when the task moves away (migrates, changes scheduling
769 * policy, or terminates).
770 * This is needed to compute the "inactive utilization" for the
771 * runqueue (inactive utilization = this_bw - running_bw).
772 */
773 u64 this_bw;
774 u64 extra_bw;
775
776 /*
777 * Maximum available bandwidth for reclaiming by SCHED_FLAG_RECLAIM
778 * tasks of this rq. Used in calculation of reclaimable bandwidth(GRUB).
779 */
780 u64 max_bw;
781
782 /*
783 * Inverse of the fraction of CPU utilization that can be reclaimed
784 * by the GRUB algorithm.
785 */
786 u64 bw_ratio;
787};
788
789#ifdef CONFIG_FAIR_GROUP_SCHED
790/* An entity is a task if it doesn't "own" a runqueue */
791#define entity_is_task(se) (!se->my_q)
792
793static inline void se_update_runnable(struct sched_entity *se)
794{
795 if (!entity_is_task(se))
796 se->runnable_weight = se->my_q->h_nr_running;
797}
798
799static inline long se_runnable(struct sched_entity *se)
800{
801 if (entity_is_task(se))
802 return !!se->on_rq;
803 else
804 return se->runnable_weight;
805}
806
807#else
808#define entity_is_task(se) 1
809
810static inline void se_update_runnable(struct sched_entity *se) {}
811
812static inline long se_runnable(struct sched_entity *se)
813{
814 return !!se->on_rq;
815}
816#endif
817
818#ifdef CONFIG_SMP
819/*
820 * XXX we want to get rid of these helpers and use the full load resolution.
821 */
822static inline long se_weight(struct sched_entity *se)
823{
824 return scale_load_down(se->load.weight);
825}
826
827
828static inline bool sched_asym_prefer(int a, int b)
829{
830 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b);
831}
832
833struct perf_domain {
834 struct em_perf_domain *em_pd;
835 struct perf_domain *next;
836 struct rcu_head rcu;
837};
838
839/* Scheduling group status flags */
840#define SG_OVERLOAD 0x1 /* More than one runnable task on a CPU. */
841#define SG_OVERUTILIZED 0x2 /* One or more CPUs are over-utilized. */
842
843/*
844 * We add the notion of a root-domain which will be used to define per-domain
845 * variables. Each exclusive cpuset essentially defines an island domain by
846 * fully partitioning the member CPUs from any other cpuset. Whenever a new
847 * exclusive cpuset is created, we also create and attach a new root-domain
848 * object.
849 *
850 */
851struct root_domain {
852 atomic_t refcount;
853 atomic_t rto_count;
854 struct rcu_head rcu;
855 cpumask_var_t span;
856 cpumask_var_t online;
857
858 /*
859 * Indicate pullable load on at least one CPU, e.g:
860 * - More than one runnable task
861 * - Running task is misfit
862 */
863 int overload;
864
865 /* Indicate one or more cpus over-utilized (tipping point) */
866 int overutilized;
867
868 /*
869 * The bit corresponding to a CPU gets set here if such CPU has more
870 * than one runnable -deadline task (as it is below for RT tasks).
871 */
872 cpumask_var_t dlo_mask;
873 atomic_t dlo_count;
874 struct dl_bw dl_bw;
875 struct cpudl cpudl;
876
877 /*
878 * Indicate whether a root_domain's dl_bw has been checked or
879 * updated. It's monotonously increasing value.
880 *
881 * Also, some corner cases, like 'wrap around' is dangerous, but given
882 * that u64 is 'big enough'. So that shouldn't be a concern.
883 */
884 u64 visit_gen;
885
886#ifdef HAVE_RT_PUSH_IPI
887 /*
888 * For IPI pull requests, loop across the rto_mask.
889 */
890 struct irq_work rto_push_work;
891 raw_spinlock_t rto_lock;
892 /* These are only updated and read within rto_lock */
893 int rto_loop;
894 int rto_cpu;
895 /* These atomics are updated outside of a lock */
896 atomic_t rto_loop_next;
897 atomic_t rto_loop_start;
898#endif
899 /*
900 * The "RT overload" flag: it gets set if a CPU has more than
901 * one runnable RT task.
902 */
903 cpumask_var_t rto_mask;
904 struct cpupri cpupri;
905
906 unsigned long max_cpu_capacity;
907
908 /*
909 * NULL-terminated list of performance domains intersecting with the
910 * CPUs of the rd. Protected by RCU.
911 */
912 struct perf_domain __rcu *pd;
913};
914
915extern void init_defrootdomain(void);
916extern int sched_init_domains(const struct cpumask *cpu_map);
917extern void rq_attach_root(struct rq *rq, struct root_domain *rd);
918extern void sched_get_rd(struct root_domain *rd);
919extern void sched_put_rd(struct root_domain *rd);
920
921#ifdef HAVE_RT_PUSH_IPI
922extern void rto_push_irq_work_func(struct irq_work *work);
923#endif
924#endif /* CONFIG_SMP */
925
926#ifdef CONFIG_UCLAMP_TASK
927/*
928 * struct uclamp_bucket - Utilization clamp bucket
929 * @value: utilization clamp value for tasks on this clamp bucket
930 * @tasks: number of RUNNABLE tasks on this clamp bucket
931 *
932 * Keep track of how many tasks are RUNNABLE for a given utilization
933 * clamp value.
934 */
935struct uclamp_bucket {
936 unsigned long value : bits_per(SCHED_CAPACITY_SCALE);
937 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE);
938};
939
940/*
941 * struct uclamp_rq - rq's utilization clamp
942 * @value: currently active clamp values for a rq
943 * @bucket: utilization clamp buckets affecting a rq
944 *
945 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values.
946 * A clamp value is affecting a rq when there is at least one task RUNNABLE
947 * (or actually running) with that value.
948 *
949 * There are up to UCLAMP_CNT possible different clamp values, currently there
950 * are only two: minimum utilization and maximum utilization.
951 *
952 * All utilization clamping values are MAX aggregated, since:
953 * - for util_min: we want to run the CPU at least at the max of the minimum
954 * utilization required by its currently RUNNABLE tasks.
955 * - for util_max: we want to allow the CPU to run up to the max of the
956 * maximum utilization allowed by its currently RUNNABLE tasks.
957 *
958 * Since on each system we expect only a limited number of different
959 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track
960 * the metrics required to compute all the per-rq utilization clamp values.
961 */
962struct uclamp_rq {
963 unsigned int value;
964 struct uclamp_bucket bucket[UCLAMP_BUCKETS];
965};
966
967DECLARE_STATIC_KEY_FALSE(sched_uclamp_used);
968#endif /* CONFIG_UCLAMP_TASK */
969
970struct rq;
971struct balance_callback {
972 struct balance_callback *next;
973 void (*func)(struct rq *rq);
974};
975
976/*
977 * This is the main, per-CPU runqueue data structure.
978 *
979 * Locking rule: those places that want to lock multiple runqueues
980 * (such as the load balancing or the thread migration code), lock
981 * acquire operations must be ordered by ascending &runqueue.
982 */
983struct rq {
984 /* runqueue lock: */
985 raw_spinlock_t __lock;
986
987 unsigned int nr_running;
988#ifdef CONFIG_NUMA_BALANCING
989 unsigned int nr_numa_running;
990 unsigned int nr_preferred_running;
991 unsigned int numa_migrate_on;
992#endif
993#ifdef CONFIG_NO_HZ_COMMON
994#ifdef CONFIG_SMP
995 unsigned long last_blocked_load_update_tick;
996 unsigned int has_blocked_load;
997 call_single_data_t nohz_csd;
998#endif /* CONFIG_SMP */
999 unsigned int nohz_tick_stopped;
1000 atomic_t nohz_flags;
1001#endif /* CONFIG_NO_HZ_COMMON */
1002
1003#ifdef CONFIG_SMP
1004 unsigned int ttwu_pending;
1005#endif
1006 u64 nr_switches;
1007
1008#ifdef CONFIG_UCLAMP_TASK
1009 /* Utilization clamp values based on CPU's RUNNABLE tasks */
1010 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned;
1011 unsigned int uclamp_flags;
1012#define UCLAMP_FLAG_IDLE 0x01
1013#endif
1014
1015 struct cfs_rq cfs;
1016 struct rt_rq rt;
1017 struct dl_rq dl;
1018
1019#ifdef CONFIG_FAIR_GROUP_SCHED
1020 /* list of leaf cfs_rq on this CPU: */
1021 struct list_head leaf_cfs_rq_list;
1022 struct list_head *tmp_alone_branch;
1023#endif /* CONFIG_FAIR_GROUP_SCHED */
1024
1025 /*
1026 * This is part of a global counter where only the total sum
1027 * over all CPUs matters. A task can increase this counter on
1028 * one CPU and if it got migrated afterwards it may decrease
1029 * it on another CPU. Always updated under the runqueue lock:
1030 */
1031 unsigned int nr_uninterruptible;
1032
1033 struct task_struct __rcu *curr;
1034 struct task_struct *idle;
1035 struct task_struct *stop;
1036 unsigned long next_balance;
1037 struct mm_struct *prev_mm;
1038
1039 unsigned int clock_update_flags;
1040 u64 clock;
1041 /* Ensure that all clocks are in the same cache line */
1042 u64 clock_task ____cacheline_aligned;
1043 u64 clock_pelt;
1044 unsigned long lost_idle_time;
1045 u64 clock_pelt_idle;
1046 u64 clock_idle;
1047#ifndef CONFIG_64BIT
1048 u64 clock_pelt_idle_copy;
1049 u64 clock_idle_copy;
1050#endif
1051
1052 atomic_t nr_iowait;
1053
1054#ifdef CONFIG_SCHED_DEBUG
1055 u64 last_seen_need_resched_ns;
1056 int ticks_without_resched;
1057#endif
1058
1059#ifdef CONFIG_MEMBARRIER
1060 int membarrier_state;
1061#endif
1062
1063#ifdef CONFIG_SMP
1064 struct root_domain *rd;
1065 struct sched_domain __rcu *sd;
1066
1067 unsigned long cpu_capacity;
1068
1069 struct balance_callback *balance_callback;
1070
1071 unsigned char nohz_idle_balance;
1072 unsigned char idle_balance;
1073
1074 unsigned long misfit_task_load;
1075
1076 /* For active balancing */
1077 int active_balance;
1078 int push_cpu;
1079 struct cpu_stop_work active_balance_work;
1080
1081 /* CPU of this runqueue: */
1082 int cpu;
1083 int online;
1084
1085 struct list_head cfs_tasks;
1086
1087 struct sched_avg avg_rt;
1088 struct sched_avg avg_dl;
1089#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
1090 struct sched_avg avg_irq;
1091#endif
1092#ifdef CONFIG_SCHED_THERMAL_PRESSURE
1093 struct sched_avg avg_thermal;
1094#endif
1095 u64 idle_stamp;
1096 u64 avg_idle;
1097
1098 /* This is used to determine avg_idle's max value */
1099 u64 max_idle_balance_cost;
1100
1101#ifdef CONFIG_HOTPLUG_CPU
1102 struct rcuwait hotplug_wait;
1103#endif
1104#endif /* CONFIG_SMP */
1105
1106#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1107 u64 prev_irq_time;
1108#endif
1109#ifdef CONFIG_PARAVIRT
1110 u64 prev_steal_time;
1111#endif
1112#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1113 u64 prev_steal_time_rq;
1114#endif
1115
1116 /* calc_load related fields */
1117 unsigned long calc_load_update;
1118 long calc_load_active;
1119
1120#ifdef CONFIG_SCHED_HRTICK
1121#ifdef CONFIG_SMP
1122 call_single_data_t hrtick_csd;
1123#endif
1124 struct hrtimer hrtick_timer;
1125 ktime_t hrtick_time;
1126#endif
1127
1128#ifdef CONFIG_SCHEDSTATS
1129 /* latency stats */
1130 struct sched_info rq_sched_info;
1131 unsigned long long rq_cpu_time;
1132 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
1133
1134 /* sys_sched_yield() stats */
1135 unsigned int yld_count;
1136
1137 /* schedule() stats */
1138 unsigned int sched_count;
1139 unsigned int sched_goidle;
1140
1141 /* try_to_wake_up() stats */
1142 unsigned int ttwu_count;
1143 unsigned int ttwu_local;
1144#endif
1145
1146#ifdef CONFIG_CPU_IDLE
1147 /* Must be inspected within a rcu lock section */
1148 struct cpuidle_state *idle_state;
1149#endif
1150
1151#ifdef CONFIG_SMP
1152 unsigned int nr_pinned;
1153#endif
1154 unsigned int push_busy;
1155 struct cpu_stop_work push_work;
1156
1157#ifdef CONFIG_SCHED_CORE
1158 /* per rq */
1159 struct rq *core;
1160 struct task_struct *core_pick;
1161 unsigned int core_enabled;
1162 unsigned int core_sched_seq;
1163 struct rb_root core_tree;
1164
1165 /* shared state -- careful with sched_core_cpu_deactivate() */
1166 unsigned int core_task_seq;
1167 unsigned int core_pick_seq;
1168 unsigned long core_cookie;
1169 unsigned int core_forceidle_count;
1170 unsigned int core_forceidle_seq;
1171 unsigned int core_forceidle_occupation;
1172 u64 core_forceidle_start;
1173#endif
1174
1175 /* Scratch cpumask to be temporarily used under rq_lock */
1176 cpumask_var_t scratch_mask;
1177
1178#if defined(CONFIG_CFS_BANDWIDTH) && defined(CONFIG_SMP)
1179 call_single_data_t cfsb_csd;
1180 struct list_head cfsb_csd_list;
1181#endif
1182};
1183
1184#ifdef CONFIG_FAIR_GROUP_SCHED
1185
1186/* CPU runqueue to which this cfs_rq is attached */
1187static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1188{
1189 return cfs_rq->rq;
1190}
1191
1192#else
1193
1194static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1195{
1196 return container_of(cfs_rq, struct rq, cfs);
1197}
1198#endif
1199
1200static inline int cpu_of(struct rq *rq)
1201{
1202#ifdef CONFIG_SMP
1203 return rq->cpu;
1204#else
1205 return 0;
1206#endif
1207}
1208
1209#define MDF_PUSH 0x01
1210
1211static inline bool is_migration_disabled(struct task_struct *p)
1212{
1213#ifdef CONFIG_SMP
1214 return p->migration_disabled;
1215#else
1216 return false;
1217#endif
1218}
1219
1220DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1221
1222#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
1223#define this_rq() this_cpu_ptr(&runqueues)
1224#define task_rq(p) cpu_rq(task_cpu(p))
1225#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
1226#define raw_rq() raw_cpu_ptr(&runqueues)
1227
1228struct sched_group;
1229#ifdef CONFIG_SCHED_CORE
1230static inline struct cpumask *sched_group_span(struct sched_group *sg);
1231
1232DECLARE_STATIC_KEY_FALSE(__sched_core_enabled);
1233
1234static inline bool sched_core_enabled(struct rq *rq)
1235{
1236 return static_branch_unlikely(&__sched_core_enabled) && rq->core_enabled;
1237}
1238
1239static inline bool sched_core_disabled(void)
1240{
1241 return !static_branch_unlikely(&__sched_core_enabled);
1242}
1243
1244/*
1245 * Be careful with this function; not for general use. The return value isn't
1246 * stable unless you actually hold a relevant rq->__lock.
1247 */
1248static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1249{
1250 if (sched_core_enabled(rq))
1251 return &rq->core->__lock;
1252
1253 return &rq->__lock;
1254}
1255
1256static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1257{
1258 if (rq->core_enabled)
1259 return &rq->core->__lock;
1260
1261 return &rq->__lock;
1262}
1263
1264bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
1265 bool fi);
1266void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
1267
1268/*
1269 * Helpers to check if the CPU's core cookie matches with the task's cookie
1270 * when core scheduling is enabled.
1271 * A special case is that the task's cookie always matches with CPU's core
1272 * cookie if the CPU is in an idle core.
1273 */
1274static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1275{
1276 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1277 if (!sched_core_enabled(rq))
1278 return true;
1279
1280 return rq->core->core_cookie == p->core_cookie;
1281}
1282
1283static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1284{
1285 bool idle_core = true;
1286 int cpu;
1287
1288 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1289 if (!sched_core_enabled(rq))
1290 return true;
1291
1292 for_each_cpu(cpu, cpu_smt_mask(cpu_of(rq))) {
1293 if (!available_idle_cpu(cpu)) {
1294 idle_core = false;
1295 break;
1296 }
1297 }
1298
1299 /*
1300 * A CPU in an idle core is always the best choice for tasks with
1301 * cookies.
1302 */
1303 return idle_core || rq->core->core_cookie == p->core_cookie;
1304}
1305
1306static inline bool sched_group_cookie_match(struct rq *rq,
1307 struct task_struct *p,
1308 struct sched_group *group)
1309{
1310 int cpu;
1311
1312 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1313 if (!sched_core_enabled(rq))
1314 return true;
1315
1316 for_each_cpu_and(cpu, sched_group_span(group), p->cpus_ptr) {
1317 if (sched_core_cookie_match(cpu_rq(cpu), p))
1318 return true;
1319 }
1320 return false;
1321}
1322
1323static inline bool sched_core_enqueued(struct task_struct *p)
1324{
1325 return !RB_EMPTY_NODE(&p->core_node);
1326}
1327
1328extern void sched_core_enqueue(struct rq *rq, struct task_struct *p);
1329extern void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags);
1330
1331extern void sched_core_get(void);
1332extern void sched_core_put(void);
1333
1334#else /* !CONFIG_SCHED_CORE */
1335
1336static inline bool sched_core_enabled(struct rq *rq)
1337{
1338 return false;
1339}
1340
1341static inline bool sched_core_disabled(void)
1342{
1343 return true;
1344}
1345
1346static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1347{
1348 return &rq->__lock;
1349}
1350
1351static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1352{
1353 return &rq->__lock;
1354}
1355
1356static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1357{
1358 return true;
1359}
1360
1361static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1362{
1363 return true;
1364}
1365
1366static inline bool sched_group_cookie_match(struct rq *rq,
1367 struct task_struct *p,
1368 struct sched_group *group)
1369{
1370 return true;
1371}
1372#endif /* CONFIG_SCHED_CORE */
1373
1374static inline void lockdep_assert_rq_held(struct rq *rq)
1375{
1376 lockdep_assert_held(__rq_lockp(rq));
1377}
1378
1379extern void raw_spin_rq_lock_nested(struct rq *rq, int subclass);
1380extern bool raw_spin_rq_trylock(struct rq *rq);
1381extern void raw_spin_rq_unlock(struct rq *rq);
1382
1383static inline void raw_spin_rq_lock(struct rq *rq)
1384{
1385 raw_spin_rq_lock_nested(rq, 0);
1386}
1387
1388static inline void raw_spin_rq_lock_irq(struct rq *rq)
1389{
1390 local_irq_disable();
1391 raw_spin_rq_lock(rq);
1392}
1393
1394static inline void raw_spin_rq_unlock_irq(struct rq *rq)
1395{
1396 raw_spin_rq_unlock(rq);
1397 local_irq_enable();
1398}
1399
1400static inline unsigned long _raw_spin_rq_lock_irqsave(struct rq *rq)
1401{
1402 unsigned long flags;
1403 local_irq_save(flags);
1404 raw_spin_rq_lock(rq);
1405 return flags;
1406}
1407
1408static inline void raw_spin_rq_unlock_irqrestore(struct rq *rq, unsigned long flags)
1409{
1410 raw_spin_rq_unlock(rq);
1411 local_irq_restore(flags);
1412}
1413
1414#define raw_spin_rq_lock_irqsave(rq, flags) \
1415do { \
1416 flags = _raw_spin_rq_lock_irqsave(rq); \
1417} while (0)
1418
1419#ifdef CONFIG_SCHED_SMT
1420extern void __update_idle_core(struct rq *rq);
1421
1422static inline void update_idle_core(struct rq *rq)
1423{
1424 if (static_branch_unlikely(&sched_smt_present))
1425 __update_idle_core(rq);
1426}
1427
1428#else
1429static inline void update_idle_core(struct rq *rq) { }
1430#endif
1431
1432#ifdef CONFIG_FAIR_GROUP_SCHED
1433static inline struct task_struct *task_of(struct sched_entity *se)
1434{
1435 SCHED_WARN_ON(!entity_is_task(se));
1436 return container_of(se, struct task_struct, se);
1437}
1438
1439static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
1440{
1441 return p->se.cfs_rq;
1442}
1443
1444/* runqueue on which this entity is (to be) queued */
1445static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1446{
1447 return se->cfs_rq;
1448}
1449
1450/* runqueue "owned" by this group */
1451static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1452{
1453 return grp->my_q;
1454}
1455
1456#else
1457
1458#define task_of(_se) container_of(_se, struct task_struct, se)
1459
1460static inline struct cfs_rq *task_cfs_rq(const struct task_struct *p)
1461{
1462 return &task_rq(p)->cfs;
1463}
1464
1465static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1466{
1467 const struct task_struct *p = task_of(se);
1468 struct rq *rq = task_rq(p);
1469
1470 return &rq->cfs;
1471}
1472
1473/* runqueue "owned" by this group */
1474static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1475{
1476 return NULL;
1477}
1478#endif
1479
1480extern void update_rq_clock(struct rq *rq);
1481
1482/*
1483 * rq::clock_update_flags bits
1484 *
1485 * %RQCF_REQ_SKIP - will request skipping of clock update on the next
1486 * call to __schedule(). This is an optimisation to avoid
1487 * neighbouring rq clock updates.
1488 *
1489 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is
1490 * in effect and calls to update_rq_clock() are being ignored.
1491 *
1492 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been
1493 * made to update_rq_clock() since the last time rq::lock was pinned.
1494 *
1495 * If inside of __schedule(), clock_update_flags will have been
1496 * shifted left (a left shift is a cheap operation for the fast path
1497 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use,
1498 *
1499 * if (rq-clock_update_flags >= RQCF_UPDATED)
1500 *
1501 * to check if %RQCF_UPDATED is set. It'll never be shifted more than
1502 * one position though, because the next rq_unpin_lock() will shift it
1503 * back.
1504 */
1505#define RQCF_REQ_SKIP 0x01
1506#define RQCF_ACT_SKIP 0x02
1507#define RQCF_UPDATED 0x04
1508
1509static inline void assert_clock_updated(struct rq *rq)
1510{
1511 /*
1512 * The only reason for not seeing a clock update since the
1513 * last rq_pin_lock() is if we're currently skipping updates.
1514 */
1515 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
1516}
1517
1518static inline u64 rq_clock(struct rq *rq)
1519{
1520 lockdep_assert_rq_held(rq);
1521 assert_clock_updated(rq);
1522
1523 return rq->clock;
1524}
1525
1526static inline u64 rq_clock_task(struct rq *rq)
1527{
1528 lockdep_assert_rq_held(rq);
1529 assert_clock_updated(rq);
1530
1531 return rq->clock_task;
1532}
1533
1534/**
1535 * By default the decay is the default pelt decay period.
1536 * The decay shift can change the decay period in
1537 * multiples of 32.
1538 * Decay shift Decay period(ms)
1539 * 0 32
1540 * 1 64
1541 * 2 128
1542 * 3 256
1543 * 4 512
1544 */
1545extern int sched_thermal_decay_shift;
1546
1547static inline u64 rq_clock_thermal(struct rq *rq)
1548{
1549 return rq_clock_task(rq) >> sched_thermal_decay_shift;
1550}
1551
1552static inline void rq_clock_skip_update(struct rq *rq)
1553{
1554 lockdep_assert_rq_held(rq);
1555 rq->clock_update_flags |= RQCF_REQ_SKIP;
1556}
1557
1558/*
1559 * See rt task throttling, which is the only time a skip
1560 * request is canceled.
1561 */
1562static inline void rq_clock_cancel_skipupdate(struct rq *rq)
1563{
1564 lockdep_assert_rq_held(rq);
1565 rq->clock_update_flags &= ~RQCF_REQ_SKIP;
1566}
1567
1568/*
1569 * During cpu offlining and rq wide unthrottling, we can trigger
1570 * an update_rq_clock() for several cfs and rt runqueues (Typically
1571 * when using list_for_each_entry_*)
1572 * rq_clock_start_loop_update() can be called after updating the clock
1573 * once and before iterating over the list to prevent multiple update.
1574 * After the iterative traversal, we need to call rq_clock_stop_loop_update()
1575 * to clear RQCF_ACT_SKIP of rq->clock_update_flags.
1576 */
1577static inline void rq_clock_start_loop_update(struct rq *rq)
1578{
1579 lockdep_assert_rq_held(rq);
1580 SCHED_WARN_ON(rq->clock_update_flags & RQCF_ACT_SKIP);
1581 rq->clock_update_flags |= RQCF_ACT_SKIP;
1582}
1583
1584static inline void rq_clock_stop_loop_update(struct rq *rq)
1585{
1586 lockdep_assert_rq_held(rq);
1587 rq->clock_update_flags &= ~RQCF_ACT_SKIP;
1588}
1589
1590struct rq_flags {
1591 unsigned long flags;
1592 struct pin_cookie cookie;
1593#ifdef CONFIG_SCHED_DEBUG
1594 /*
1595 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the
1596 * current pin context is stashed here in case it needs to be
1597 * restored in rq_repin_lock().
1598 */
1599 unsigned int clock_update_flags;
1600#endif
1601};
1602
1603extern struct balance_callback balance_push_callback;
1604
1605/*
1606 * Lockdep annotation that avoids accidental unlocks; it's like a
1607 * sticky/continuous lockdep_assert_held().
1608 *
1609 * This avoids code that has access to 'struct rq *rq' (basically everything in
1610 * the scheduler) from accidentally unlocking the rq if they do not also have a
1611 * copy of the (on-stack) 'struct rq_flags rf'.
1612 *
1613 * Also see Documentation/locking/lockdep-design.rst.
1614 */
1615static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
1616{
1617 rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
1618
1619#ifdef CONFIG_SCHED_DEBUG
1620 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
1621 rf->clock_update_flags = 0;
1622#ifdef CONFIG_SMP
1623 SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
1624#endif
1625#endif
1626}
1627
1628static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
1629{
1630#ifdef CONFIG_SCHED_DEBUG
1631 if (rq->clock_update_flags > RQCF_ACT_SKIP)
1632 rf->clock_update_flags = RQCF_UPDATED;
1633#endif
1634
1635 lockdep_unpin_lock(__rq_lockp(rq), rf->cookie);
1636}
1637
1638static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf)
1639{
1640 lockdep_repin_lock(__rq_lockp(rq), rf->cookie);
1641
1642#ifdef CONFIG_SCHED_DEBUG
1643 /*
1644 * Restore the value we stashed in @rf for this pin context.
1645 */
1646 rq->clock_update_flags |= rf->clock_update_flags;
1647#endif
1648}
1649
1650struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1651 __acquires(rq->lock);
1652
1653struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1654 __acquires(p->pi_lock)
1655 __acquires(rq->lock);
1656
1657static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf)
1658 __releases(rq->lock)
1659{
1660 rq_unpin_lock(rq, rf);
1661 raw_spin_rq_unlock(rq);
1662}
1663
1664static inline void
1665task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1666 __releases(rq->lock)
1667 __releases(p->pi_lock)
1668{
1669 rq_unpin_lock(rq, rf);
1670 raw_spin_rq_unlock(rq);
1671 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
1672}
1673
1674DEFINE_LOCK_GUARD_1(task_rq_lock, struct task_struct,
1675 _T->rq = task_rq_lock(_T->lock, &_T->rf),
1676 task_rq_unlock(_T->rq, _T->lock, &_T->rf),
1677 struct rq *rq; struct rq_flags rf)
1678
1679static inline void
1680rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
1681 __acquires(rq->lock)
1682{
1683 raw_spin_rq_lock_irqsave(rq, rf->flags);
1684 rq_pin_lock(rq, rf);
1685}
1686
1687static inline void
1688rq_lock_irq(struct rq *rq, struct rq_flags *rf)
1689 __acquires(rq->lock)
1690{
1691 raw_spin_rq_lock_irq(rq);
1692 rq_pin_lock(rq, rf);
1693}
1694
1695static inline void
1696rq_lock(struct rq *rq, struct rq_flags *rf)
1697 __acquires(rq->lock)
1698{
1699 raw_spin_rq_lock(rq);
1700 rq_pin_lock(rq, rf);
1701}
1702
1703static inline void
1704rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
1705 __releases(rq->lock)
1706{
1707 rq_unpin_lock(rq, rf);
1708 raw_spin_rq_unlock_irqrestore(rq, rf->flags);
1709}
1710
1711static inline void
1712rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
1713 __releases(rq->lock)
1714{
1715 rq_unpin_lock(rq, rf);
1716 raw_spin_rq_unlock_irq(rq);
1717}
1718
1719static inline void
1720rq_unlock(struct rq *rq, struct rq_flags *rf)
1721 __releases(rq->lock)
1722{
1723 rq_unpin_lock(rq, rf);
1724 raw_spin_rq_unlock(rq);
1725}
1726
1727DEFINE_LOCK_GUARD_1(rq_lock, struct rq,
1728 rq_lock(_T->lock, &_T->rf),
1729 rq_unlock(_T->lock, &_T->rf),
1730 struct rq_flags rf)
1731
1732DEFINE_LOCK_GUARD_1(rq_lock_irq, struct rq,
1733 rq_lock_irq(_T->lock, &_T->rf),
1734 rq_unlock_irq(_T->lock, &_T->rf),
1735 struct rq_flags rf)
1736
1737DEFINE_LOCK_GUARD_1(rq_lock_irqsave, struct rq,
1738 rq_lock_irqsave(_T->lock, &_T->rf),
1739 rq_unlock_irqrestore(_T->lock, &_T->rf),
1740 struct rq_flags rf)
1741
1742static inline struct rq *
1743this_rq_lock_irq(struct rq_flags *rf)
1744 __acquires(rq->lock)
1745{
1746 struct rq *rq;
1747
1748 local_irq_disable();
1749 rq = this_rq();
1750 rq_lock(rq, rf);
1751 return rq;
1752}
1753
1754#ifdef CONFIG_NUMA
1755enum numa_topology_type {
1756 NUMA_DIRECT,
1757 NUMA_GLUELESS_MESH,
1758 NUMA_BACKPLANE,
1759};
1760extern enum numa_topology_type sched_numa_topology_type;
1761extern int sched_max_numa_distance;
1762extern bool find_numa_distance(int distance);
1763extern void sched_init_numa(int offline_node);
1764extern void sched_update_numa(int cpu, bool online);
1765extern void sched_domains_numa_masks_set(unsigned int cpu);
1766extern void sched_domains_numa_masks_clear(unsigned int cpu);
1767extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu);
1768#else
1769static inline void sched_init_numa(int offline_node) { }
1770static inline void sched_update_numa(int cpu, bool online) { }
1771static inline void sched_domains_numa_masks_set(unsigned int cpu) { }
1772static inline void sched_domains_numa_masks_clear(unsigned int cpu) { }
1773static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1774{
1775 return nr_cpu_ids;
1776}
1777#endif
1778
1779#ifdef CONFIG_NUMA_BALANCING
1780/* The regions in numa_faults array from task_struct */
1781enum numa_faults_stats {
1782 NUMA_MEM = 0,
1783 NUMA_CPU,
1784 NUMA_MEMBUF,
1785 NUMA_CPUBUF
1786};
1787extern void sched_setnuma(struct task_struct *p, int node);
1788extern int migrate_task_to(struct task_struct *p, int cpu);
1789extern int migrate_swap(struct task_struct *p, struct task_struct *t,
1790 int cpu, int scpu);
1791extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p);
1792#else
1793static inline void
1794init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1795{
1796}
1797#endif /* CONFIG_NUMA_BALANCING */
1798
1799#ifdef CONFIG_SMP
1800
1801static inline void
1802queue_balance_callback(struct rq *rq,
1803 struct balance_callback *head,
1804 void (*func)(struct rq *rq))
1805{
1806 lockdep_assert_rq_held(rq);
1807
1808 /*
1809 * Don't (re)queue an already queued item; nor queue anything when
1810 * balance_push() is active, see the comment with
1811 * balance_push_callback.
1812 */
1813 if (unlikely(head->next || rq->balance_callback == &balance_push_callback))
1814 return;
1815
1816 head->func = func;
1817 head->next = rq->balance_callback;
1818 rq->balance_callback = head;
1819}
1820
1821#define rcu_dereference_check_sched_domain(p) \
1822 rcu_dereference_check((p), \
1823 lockdep_is_held(&sched_domains_mutex))
1824
1825/*
1826 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1827 * See destroy_sched_domains: call_rcu for details.
1828 *
1829 * The domain tree of any CPU may only be accessed from within
1830 * preempt-disabled sections.
1831 */
1832#define for_each_domain(cpu, __sd) \
1833 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
1834 __sd; __sd = __sd->parent)
1835
1836/* A mask of all the SD flags that have the SDF_SHARED_CHILD metaflag */
1837#define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_SHARED_CHILD)) |
1838static const unsigned int SD_SHARED_CHILD_MASK =
1839#include <linux/sched/sd_flags.h>
18400;
1841#undef SD_FLAG
1842
1843/**
1844 * highest_flag_domain - Return highest sched_domain containing flag.
1845 * @cpu: The CPU whose highest level of sched domain is to
1846 * be returned.
1847 * @flag: The flag to check for the highest sched_domain
1848 * for the given CPU.
1849 *
1850 * Returns the highest sched_domain of a CPU which contains @flag. If @flag has
1851 * the SDF_SHARED_CHILD metaflag, all the children domains also have @flag.
1852 */
1853static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
1854{
1855 struct sched_domain *sd, *hsd = NULL;
1856
1857 for_each_domain(cpu, sd) {
1858 if (sd->flags & flag) {
1859 hsd = sd;
1860 continue;
1861 }
1862
1863 /*
1864 * Stop the search if @flag is known to be shared at lower
1865 * levels. It will not be found further up.
1866 */
1867 if (flag & SD_SHARED_CHILD_MASK)
1868 break;
1869 }
1870
1871 return hsd;
1872}
1873
1874static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
1875{
1876 struct sched_domain *sd;
1877
1878 for_each_domain(cpu, sd) {
1879 if (sd->flags & flag)
1880 break;
1881 }
1882
1883 return sd;
1884}
1885
1886DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
1887DECLARE_PER_CPU(int, sd_llc_size);
1888DECLARE_PER_CPU(int, sd_llc_id);
1889DECLARE_PER_CPU(int, sd_share_id);
1890DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
1891DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
1892DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
1893DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
1894extern struct static_key_false sched_asym_cpucapacity;
1895extern struct static_key_false sched_cluster_active;
1896
1897static __always_inline bool sched_asym_cpucap_active(void)
1898{
1899 return static_branch_unlikely(&sched_asym_cpucapacity);
1900}
1901
1902struct sched_group_capacity {
1903 atomic_t ref;
1904 /*
1905 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity
1906 * for a single CPU.
1907 */
1908 unsigned long capacity;
1909 unsigned long min_capacity; /* Min per-CPU capacity in group */
1910 unsigned long max_capacity; /* Max per-CPU capacity in group */
1911 unsigned long next_update;
1912 int imbalance; /* XXX unrelated to capacity but shared group state */
1913
1914#ifdef CONFIG_SCHED_DEBUG
1915 int id;
1916#endif
1917
1918 unsigned long cpumask[]; /* Balance mask */
1919};
1920
1921struct sched_group {
1922 struct sched_group *next; /* Must be a circular list */
1923 atomic_t ref;
1924
1925 unsigned int group_weight;
1926 unsigned int cores;
1927 struct sched_group_capacity *sgc;
1928 int asym_prefer_cpu; /* CPU of highest priority in group */
1929 int flags;
1930
1931 /*
1932 * The CPUs this group covers.
1933 *
1934 * NOTE: this field is variable length. (Allocated dynamically
1935 * by attaching extra space to the end of the structure,
1936 * depending on how many CPUs the kernel has booted up with)
1937 */
1938 unsigned long cpumask[];
1939};
1940
1941static inline struct cpumask *sched_group_span(struct sched_group *sg)
1942{
1943 return to_cpumask(sg->cpumask);
1944}
1945
1946/*
1947 * See build_balance_mask().
1948 */
1949static inline struct cpumask *group_balance_mask(struct sched_group *sg)
1950{
1951 return to_cpumask(sg->sgc->cpumask);
1952}
1953
1954extern int group_balance_cpu(struct sched_group *sg);
1955
1956#ifdef CONFIG_SCHED_DEBUG
1957void update_sched_domain_debugfs(void);
1958void dirty_sched_domain_sysctl(int cpu);
1959#else
1960static inline void update_sched_domain_debugfs(void)
1961{
1962}
1963static inline void dirty_sched_domain_sysctl(int cpu)
1964{
1965}
1966#endif
1967
1968extern int sched_update_scaling(void);
1969
1970static inline const struct cpumask *task_user_cpus(struct task_struct *p)
1971{
1972 if (!p->user_cpus_ptr)
1973 return cpu_possible_mask; /* &init_task.cpus_mask */
1974 return p->user_cpus_ptr;
1975}
1976#endif /* CONFIG_SMP */
1977
1978#include "stats.h"
1979
1980#if defined(CONFIG_SCHED_CORE) && defined(CONFIG_SCHEDSTATS)
1981
1982extern void __sched_core_account_forceidle(struct rq *rq);
1983
1984static inline void sched_core_account_forceidle(struct rq *rq)
1985{
1986 if (schedstat_enabled())
1987 __sched_core_account_forceidle(rq);
1988}
1989
1990extern void __sched_core_tick(struct rq *rq);
1991
1992static inline void sched_core_tick(struct rq *rq)
1993{
1994 if (sched_core_enabled(rq) && schedstat_enabled())
1995 __sched_core_tick(rq);
1996}
1997
1998#else
1999
2000static inline void sched_core_account_forceidle(struct rq *rq) {}
2001
2002static inline void sched_core_tick(struct rq *rq) {}
2003
2004#endif /* CONFIG_SCHED_CORE && CONFIG_SCHEDSTATS */
2005
2006#ifdef CONFIG_CGROUP_SCHED
2007
2008/*
2009 * Return the group to which this tasks belongs.
2010 *
2011 * We cannot use task_css() and friends because the cgroup subsystem
2012 * changes that value before the cgroup_subsys::attach() method is called,
2013 * therefore we cannot pin it and might observe the wrong value.
2014 *
2015 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
2016 * core changes this before calling sched_move_task().
2017 *
2018 * Instead we use a 'copy' which is updated from sched_move_task() while
2019 * holding both task_struct::pi_lock and rq::lock.
2020 */
2021static inline struct task_group *task_group(struct task_struct *p)
2022{
2023 return p->sched_task_group;
2024}
2025
2026/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
2027static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
2028{
2029#if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
2030 struct task_group *tg = task_group(p);
2031#endif
2032
2033#ifdef CONFIG_FAIR_GROUP_SCHED
2034 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
2035 p->se.cfs_rq = tg->cfs_rq[cpu];
2036 p->se.parent = tg->se[cpu];
2037 p->se.depth = tg->se[cpu] ? tg->se[cpu]->depth + 1 : 0;
2038#endif
2039
2040#ifdef CONFIG_RT_GROUP_SCHED
2041 p->rt.rt_rq = tg->rt_rq[cpu];
2042 p->rt.parent = tg->rt_se[cpu];
2043#endif
2044}
2045
2046#else /* CONFIG_CGROUP_SCHED */
2047
2048static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
2049static inline struct task_group *task_group(struct task_struct *p)
2050{
2051 return NULL;
2052}
2053
2054#endif /* CONFIG_CGROUP_SCHED */
2055
2056static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
2057{
2058 set_task_rq(p, cpu);
2059#ifdef CONFIG_SMP
2060 /*
2061 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2062 * successfully executed on another CPU. We must ensure that updates of
2063 * per-task data have been completed by this moment.
2064 */
2065 smp_wmb();
2066 WRITE_ONCE(task_thread_info(p)->cpu, cpu);
2067 p->wake_cpu = cpu;
2068#endif
2069}
2070
2071/*
2072 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
2073 */
2074#ifdef CONFIG_SCHED_DEBUG
2075# define const_debug __read_mostly
2076#else
2077# define const_debug const
2078#endif
2079
2080#define SCHED_FEAT(name, enabled) \
2081 __SCHED_FEAT_##name ,
2082
2083enum {
2084#include "features.h"
2085 __SCHED_FEAT_NR,
2086};
2087
2088#undef SCHED_FEAT
2089
2090#ifdef CONFIG_SCHED_DEBUG
2091
2092/*
2093 * To support run-time toggling of sched features, all the translation units
2094 * (but core.c) reference the sysctl_sched_features defined in core.c.
2095 */
2096extern const_debug unsigned int sysctl_sched_features;
2097
2098#ifdef CONFIG_JUMP_LABEL
2099#define SCHED_FEAT(name, enabled) \
2100static __always_inline bool static_branch_##name(struct static_key *key) \
2101{ \
2102 return static_key_##enabled(key); \
2103}
2104
2105#include "features.h"
2106#undef SCHED_FEAT
2107
2108extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
2109#define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
2110
2111#else /* !CONFIG_JUMP_LABEL */
2112
2113#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2114
2115#endif /* CONFIG_JUMP_LABEL */
2116
2117#else /* !SCHED_DEBUG */
2118
2119/*
2120 * Each translation unit has its own copy of sysctl_sched_features to allow
2121 * constants propagation at compile time and compiler optimization based on
2122 * features default.
2123 */
2124#define SCHED_FEAT(name, enabled) \
2125 (1UL << __SCHED_FEAT_##name) * enabled |
2126static const_debug __maybe_unused unsigned int sysctl_sched_features =
2127#include "features.h"
2128 0;
2129#undef SCHED_FEAT
2130
2131#define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2132
2133#endif /* SCHED_DEBUG */
2134
2135extern struct static_key_false sched_numa_balancing;
2136extern struct static_key_false sched_schedstats;
2137
2138static inline u64 global_rt_period(void)
2139{
2140 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
2141}
2142
2143static inline u64 global_rt_runtime(void)
2144{
2145 if (sysctl_sched_rt_runtime < 0)
2146 return RUNTIME_INF;
2147
2148 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
2149}
2150
2151static inline int task_current(struct rq *rq, struct task_struct *p)
2152{
2153 return rq->curr == p;
2154}
2155
2156static inline int task_on_cpu(struct rq *rq, struct task_struct *p)
2157{
2158#ifdef CONFIG_SMP
2159 return p->on_cpu;
2160#else
2161 return task_current(rq, p);
2162#endif
2163}
2164
2165static inline int task_on_rq_queued(struct task_struct *p)
2166{
2167 return p->on_rq == TASK_ON_RQ_QUEUED;
2168}
2169
2170static inline int task_on_rq_migrating(struct task_struct *p)
2171{
2172 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING;
2173}
2174
2175/* Wake flags. The first three directly map to some SD flag value */
2176#define WF_EXEC 0x02 /* Wakeup after exec; maps to SD_BALANCE_EXEC */
2177#define WF_FORK 0x04 /* Wakeup after fork; maps to SD_BALANCE_FORK */
2178#define WF_TTWU 0x08 /* Wakeup; maps to SD_BALANCE_WAKE */
2179
2180#define WF_SYNC 0x10 /* Waker goes to sleep after wakeup */
2181#define WF_MIGRATED 0x20 /* Internal use, task got migrated */
2182#define WF_CURRENT_CPU 0x40 /* Prefer to move the wakee to the current CPU. */
2183
2184#ifdef CONFIG_SMP
2185static_assert(WF_EXEC == SD_BALANCE_EXEC);
2186static_assert(WF_FORK == SD_BALANCE_FORK);
2187static_assert(WF_TTWU == SD_BALANCE_WAKE);
2188#endif
2189
2190/*
2191 * To aid in avoiding the subversion of "niceness" due to uneven distribution
2192 * of tasks with abnormal "nice" values across CPUs the contribution that
2193 * each task makes to its run queue's load is weighted according to its
2194 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
2195 * scaled version of the new time slice allocation that they receive on time
2196 * slice expiry etc.
2197 */
2198
2199#define WEIGHT_IDLEPRIO 3
2200#define WMULT_IDLEPRIO 1431655765
2201
2202extern const int sched_prio_to_weight[40];
2203extern const u32 sched_prio_to_wmult[40];
2204
2205/*
2206 * {de,en}queue flags:
2207 *
2208 * DEQUEUE_SLEEP - task is no longer runnable
2209 * ENQUEUE_WAKEUP - task just became runnable
2210 *
2211 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
2212 * are in a known state which allows modification. Such pairs
2213 * should preserve as much state as possible.
2214 *
2215 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
2216 * in the runqueue.
2217 *
2218 * NOCLOCK - skip the update_rq_clock() (avoids double updates)
2219 *
2220 * MIGRATION - p->on_rq == TASK_ON_RQ_MIGRATING (used for DEADLINE)
2221 *
2222 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
2223 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
2224 * ENQUEUE_MIGRATED - the task was migrated during wakeup
2225 *
2226 */
2227
2228#define DEQUEUE_SLEEP 0x01
2229#define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */
2230#define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */
2231#define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */
2232#define DEQUEUE_MIGRATING 0x100 /* Matches ENQUEUE_MIGRATING */
2233
2234#define ENQUEUE_WAKEUP 0x01
2235#define ENQUEUE_RESTORE 0x02
2236#define ENQUEUE_MOVE 0x04
2237#define ENQUEUE_NOCLOCK 0x08
2238
2239#define ENQUEUE_HEAD 0x10
2240#define ENQUEUE_REPLENISH 0x20
2241#ifdef CONFIG_SMP
2242#define ENQUEUE_MIGRATED 0x40
2243#else
2244#define ENQUEUE_MIGRATED 0x00
2245#endif
2246#define ENQUEUE_INITIAL 0x80
2247#define ENQUEUE_MIGRATING 0x100
2248
2249#define RETRY_TASK ((void *)-1UL)
2250
2251struct affinity_context {
2252 const struct cpumask *new_mask;
2253 struct cpumask *user_mask;
2254 unsigned int flags;
2255};
2256
2257extern s64 update_curr_common(struct rq *rq);
2258
2259struct sched_class {
2260
2261#ifdef CONFIG_UCLAMP_TASK
2262 int uclamp_enabled;
2263#endif
2264
2265 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
2266 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
2267 void (*yield_task) (struct rq *rq);
2268 bool (*yield_to_task)(struct rq *rq, struct task_struct *p);
2269
2270 void (*wakeup_preempt)(struct rq *rq, struct task_struct *p, int flags);
2271
2272 struct task_struct *(*pick_next_task)(struct rq *rq);
2273
2274 void (*put_prev_task)(struct rq *rq, struct task_struct *p);
2275 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first);
2276
2277#ifdef CONFIG_SMP
2278 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2279 int (*select_task_rq)(struct task_struct *p, int task_cpu, int flags);
2280
2281 struct task_struct * (*pick_task)(struct rq *rq);
2282
2283 void (*migrate_task_rq)(struct task_struct *p, int new_cpu);
2284
2285 void (*task_woken)(struct rq *this_rq, struct task_struct *task);
2286
2287 void (*set_cpus_allowed)(struct task_struct *p, struct affinity_context *ctx);
2288
2289 void (*rq_online)(struct rq *rq);
2290 void (*rq_offline)(struct rq *rq);
2291
2292 struct rq *(*find_lock_rq)(struct task_struct *p, struct rq *rq);
2293#endif
2294
2295 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued);
2296 void (*task_fork)(struct task_struct *p);
2297 void (*task_dead)(struct task_struct *p);
2298
2299 /*
2300 * The switched_from() call is allowed to drop rq->lock, therefore we
2301 * cannot assume the switched_from/switched_to pair is serialized by
2302 * rq->lock. They are however serialized by p->pi_lock.
2303 */
2304 void (*switched_from)(struct rq *this_rq, struct task_struct *task);
2305 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
2306 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
2307 int oldprio);
2308
2309 unsigned int (*get_rr_interval)(struct rq *rq,
2310 struct task_struct *task);
2311
2312 void (*update_curr)(struct rq *rq);
2313
2314#ifdef CONFIG_FAIR_GROUP_SCHED
2315 void (*task_change_group)(struct task_struct *p);
2316#endif
2317
2318#ifdef CONFIG_SCHED_CORE
2319 int (*task_is_throttled)(struct task_struct *p, int cpu);
2320#endif
2321};
2322
2323static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
2324{
2325 WARN_ON_ONCE(rq->curr != prev);
2326 prev->sched_class->put_prev_task(rq, prev);
2327}
2328
2329static inline void set_next_task(struct rq *rq, struct task_struct *next)
2330{
2331 next->sched_class->set_next_task(rq, next, false);
2332}
2333
2334
2335/*
2336 * Helper to define a sched_class instance; each one is placed in a separate
2337 * section which is ordered by the linker script:
2338 *
2339 * include/asm-generic/vmlinux.lds.h
2340 *
2341 * *CAREFUL* they are laid out in *REVERSE* order!!!
2342 *
2343 * Also enforce alignment on the instance, not the type, to guarantee layout.
2344 */
2345#define DEFINE_SCHED_CLASS(name) \
2346const struct sched_class name##_sched_class \
2347 __aligned(__alignof__(struct sched_class)) \
2348 __section("__" #name "_sched_class")
2349
2350/* Defined in include/asm-generic/vmlinux.lds.h */
2351extern struct sched_class __sched_class_highest[];
2352extern struct sched_class __sched_class_lowest[];
2353
2354#define for_class_range(class, _from, _to) \
2355 for (class = (_from); class < (_to); class++)
2356
2357#define for_each_class(class) \
2358 for_class_range(class, __sched_class_highest, __sched_class_lowest)
2359
2360#define sched_class_above(_a, _b) ((_a) < (_b))
2361
2362extern const struct sched_class stop_sched_class;
2363extern const struct sched_class dl_sched_class;
2364extern const struct sched_class rt_sched_class;
2365extern const struct sched_class fair_sched_class;
2366extern const struct sched_class idle_sched_class;
2367
2368static inline bool sched_stop_runnable(struct rq *rq)
2369{
2370 return rq->stop && task_on_rq_queued(rq->stop);
2371}
2372
2373static inline bool sched_dl_runnable(struct rq *rq)
2374{
2375 return rq->dl.dl_nr_running > 0;
2376}
2377
2378static inline bool sched_rt_runnable(struct rq *rq)
2379{
2380 return rq->rt.rt_queued > 0;
2381}
2382
2383static inline bool sched_fair_runnable(struct rq *rq)
2384{
2385 return rq->cfs.nr_running > 0;
2386}
2387
2388extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2389extern struct task_struct *pick_next_task_idle(struct rq *rq);
2390
2391#define SCA_CHECK 0x01
2392#define SCA_MIGRATE_DISABLE 0x02
2393#define SCA_MIGRATE_ENABLE 0x04
2394#define SCA_USER 0x08
2395
2396#ifdef CONFIG_SMP
2397
2398extern void update_group_capacity(struct sched_domain *sd, int cpu);
2399
2400extern void trigger_load_balance(struct rq *rq);
2401
2402extern void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx);
2403
2404static inline struct task_struct *get_push_task(struct rq *rq)
2405{
2406 struct task_struct *p = rq->curr;
2407
2408 lockdep_assert_rq_held(rq);
2409
2410 if (rq->push_busy)
2411 return NULL;
2412
2413 if (p->nr_cpus_allowed == 1)
2414 return NULL;
2415
2416 if (p->migration_disabled)
2417 return NULL;
2418
2419 rq->push_busy = true;
2420 return get_task_struct(p);
2421}
2422
2423extern int push_cpu_stop(void *arg);
2424
2425#endif
2426
2427#ifdef CONFIG_CPU_IDLE
2428static inline void idle_set_state(struct rq *rq,
2429 struct cpuidle_state *idle_state)
2430{
2431 rq->idle_state = idle_state;
2432}
2433
2434static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2435{
2436 SCHED_WARN_ON(!rcu_read_lock_held());
2437
2438 return rq->idle_state;
2439}
2440#else
2441static inline void idle_set_state(struct rq *rq,
2442 struct cpuidle_state *idle_state)
2443{
2444}
2445
2446static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2447{
2448 return NULL;
2449}
2450#endif
2451
2452extern void schedule_idle(void);
2453asmlinkage void schedule_user(void);
2454
2455extern void sysrq_sched_debug_show(void);
2456extern void sched_init_granularity(void);
2457extern void update_max_interval(void);
2458
2459extern void init_sched_dl_class(void);
2460extern void init_sched_rt_class(void);
2461extern void init_sched_fair_class(void);
2462
2463extern void reweight_task(struct task_struct *p, int prio);
2464
2465extern void resched_curr(struct rq *rq);
2466extern void resched_cpu(int cpu);
2467
2468extern struct rt_bandwidth def_rt_bandwidth;
2469extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
2470extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
2471
2472extern void init_dl_entity(struct sched_dl_entity *dl_se);
2473
2474#define BW_SHIFT 20
2475#define BW_UNIT (1 << BW_SHIFT)
2476#define RATIO_SHIFT 8
2477#define MAX_BW_BITS (64 - BW_SHIFT)
2478#define MAX_BW ((1ULL << MAX_BW_BITS) - 1)
2479unsigned long to_ratio(u64 period, u64 runtime);
2480
2481extern void init_entity_runnable_average(struct sched_entity *se);
2482extern void post_init_entity_util_avg(struct task_struct *p);
2483
2484#ifdef CONFIG_NO_HZ_FULL
2485extern bool sched_can_stop_tick(struct rq *rq);
2486extern int __init sched_tick_offload_init(void);
2487
2488/*
2489 * Tick may be needed by tasks in the runqueue depending on their policy and
2490 * requirements. If tick is needed, lets send the target an IPI to kick it out of
2491 * nohz mode if necessary.
2492 */
2493static inline void sched_update_tick_dependency(struct rq *rq)
2494{
2495 int cpu = cpu_of(rq);
2496
2497 if (!tick_nohz_full_cpu(cpu))
2498 return;
2499
2500 if (sched_can_stop_tick(rq))
2501 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
2502 else
2503 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
2504}
2505#else
2506static inline int sched_tick_offload_init(void) { return 0; }
2507static inline void sched_update_tick_dependency(struct rq *rq) { }
2508#endif
2509
2510static inline void add_nr_running(struct rq *rq, unsigned count)
2511{
2512 unsigned prev_nr = rq->nr_running;
2513
2514 rq->nr_running = prev_nr + count;
2515 if (trace_sched_update_nr_running_tp_enabled()) {
2516 call_trace_sched_update_nr_running(rq, count);
2517 }
2518
2519#ifdef CONFIG_SMP
2520 if (prev_nr < 2 && rq->nr_running >= 2) {
2521 if (!READ_ONCE(rq->rd->overload))
2522 WRITE_ONCE(rq->rd->overload, 1);
2523 }
2524#endif
2525
2526 sched_update_tick_dependency(rq);
2527}
2528
2529static inline void sub_nr_running(struct rq *rq, unsigned count)
2530{
2531 rq->nr_running -= count;
2532 if (trace_sched_update_nr_running_tp_enabled()) {
2533 call_trace_sched_update_nr_running(rq, -count);
2534 }
2535
2536 /* Check if we still need preemption */
2537 sched_update_tick_dependency(rq);
2538}
2539
2540extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
2541extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
2542
2543extern void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags);
2544
2545#ifdef CONFIG_PREEMPT_RT
2546#define SCHED_NR_MIGRATE_BREAK 8
2547#else
2548#define SCHED_NR_MIGRATE_BREAK 32
2549#endif
2550
2551extern const_debug unsigned int sysctl_sched_nr_migrate;
2552extern const_debug unsigned int sysctl_sched_migration_cost;
2553
2554extern unsigned int sysctl_sched_base_slice;
2555
2556#ifdef CONFIG_SCHED_DEBUG
2557extern int sysctl_resched_latency_warn_ms;
2558extern int sysctl_resched_latency_warn_once;
2559
2560extern unsigned int sysctl_sched_tunable_scaling;
2561
2562extern unsigned int sysctl_numa_balancing_scan_delay;
2563extern unsigned int sysctl_numa_balancing_scan_period_min;
2564extern unsigned int sysctl_numa_balancing_scan_period_max;
2565extern unsigned int sysctl_numa_balancing_scan_size;
2566extern unsigned int sysctl_numa_balancing_hot_threshold;
2567#endif
2568
2569#ifdef CONFIG_SCHED_HRTICK
2570
2571/*
2572 * Use hrtick when:
2573 * - enabled by features
2574 * - hrtimer is actually high res
2575 */
2576static inline int hrtick_enabled(struct rq *rq)
2577{
2578 if (!cpu_active(cpu_of(rq)))
2579 return 0;
2580 return hrtimer_is_hres_active(&rq->hrtick_timer);
2581}
2582
2583static inline int hrtick_enabled_fair(struct rq *rq)
2584{
2585 if (!sched_feat(HRTICK))
2586 return 0;
2587 return hrtick_enabled(rq);
2588}
2589
2590static inline int hrtick_enabled_dl(struct rq *rq)
2591{
2592 if (!sched_feat(HRTICK_DL))
2593 return 0;
2594 return hrtick_enabled(rq);
2595}
2596
2597void hrtick_start(struct rq *rq, u64 delay);
2598
2599#else
2600
2601static inline int hrtick_enabled_fair(struct rq *rq)
2602{
2603 return 0;
2604}
2605
2606static inline int hrtick_enabled_dl(struct rq *rq)
2607{
2608 return 0;
2609}
2610
2611static inline int hrtick_enabled(struct rq *rq)
2612{
2613 return 0;
2614}
2615
2616#endif /* CONFIG_SCHED_HRTICK */
2617
2618#ifndef arch_scale_freq_tick
2619static __always_inline
2620void arch_scale_freq_tick(void)
2621{
2622}
2623#endif
2624
2625#ifndef arch_scale_freq_capacity
2626/**
2627 * arch_scale_freq_capacity - get the frequency scale factor of a given CPU.
2628 * @cpu: the CPU in question.
2629 *
2630 * Return: the frequency scale factor normalized against SCHED_CAPACITY_SCALE, i.e.
2631 *
2632 * f_curr
2633 * ------ * SCHED_CAPACITY_SCALE
2634 * f_max
2635 */
2636static __always_inline
2637unsigned long arch_scale_freq_capacity(int cpu)
2638{
2639 return SCHED_CAPACITY_SCALE;
2640}
2641#endif
2642
2643#ifdef CONFIG_SCHED_DEBUG
2644/*
2645 * In double_lock_balance()/double_rq_lock(), we use raw_spin_rq_lock() to
2646 * acquire rq lock instead of rq_lock(). So at the end of these two functions
2647 * we need to call double_rq_clock_clear_update() to clear RQCF_UPDATED of
2648 * rq->clock_update_flags to avoid the WARN_DOUBLE_CLOCK warning.
2649 */
2650static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2)
2651{
2652 rq1->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2653 /* rq1 == rq2 for !CONFIG_SMP, so just clear RQCF_UPDATED once. */
2654#ifdef CONFIG_SMP
2655 rq2->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2656#endif
2657}
2658#else
2659static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2) {}
2660#endif
2661
2662#define DEFINE_LOCK_GUARD_2(name, type, _lock, _unlock, ...) \
2663__DEFINE_UNLOCK_GUARD(name, type, _unlock, type *lock2; __VA_ARGS__) \
2664static inline class_##name##_t class_##name##_constructor(type *lock, type *lock2) \
2665{ class_##name##_t _t = { .lock = lock, .lock2 = lock2 }, *_T = &_t; \
2666 _lock; return _t; }
2667
2668#ifdef CONFIG_SMP
2669
2670static inline bool rq_order_less(struct rq *rq1, struct rq *rq2)
2671{
2672#ifdef CONFIG_SCHED_CORE
2673 /*
2674 * In order to not have {0,2},{1,3} turn into into an AB-BA,
2675 * order by core-id first and cpu-id second.
2676 *
2677 * Notably:
2678 *
2679 * double_rq_lock(0,3); will take core-0, core-1 lock
2680 * double_rq_lock(1,2); will take core-1, core-0 lock
2681 *
2682 * when only cpu-id is considered.
2683 */
2684 if (rq1->core->cpu < rq2->core->cpu)
2685 return true;
2686 if (rq1->core->cpu > rq2->core->cpu)
2687 return false;
2688
2689 /*
2690 * __sched_core_flip() relies on SMT having cpu-id lock order.
2691 */
2692#endif
2693 return rq1->cpu < rq2->cpu;
2694}
2695
2696extern void double_rq_lock(struct rq *rq1, struct rq *rq2);
2697
2698#ifdef CONFIG_PREEMPTION
2699
2700/*
2701 * fair double_lock_balance: Safely acquires both rq->locks in a fair
2702 * way at the expense of forcing extra atomic operations in all
2703 * invocations. This assures that the double_lock is acquired using the
2704 * same underlying policy as the spinlock_t on this architecture, which
2705 * reduces latency compared to the unfair variant below. However, it
2706 * also adds more overhead and therefore may reduce throughput.
2707 */
2708static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2709 __releases(this_rq->lock)
2710 __acquires(busiest->lock)
2711 __acquires(this_rq->lock)
2712{
2713 raw_spin_rq_unlock(this_rq);
2714 double_rq_lock(this_rq, busiest);
2715
2716 return 1;
2717}
2718
2719#else
2720/*
2721 * Unfair double_lock_balance: Optimizes throughput at the expense of
2722 * latency by eliminating extra atomic operations when the locks are
2723 * already in proper order on entry. This favors lower CPU-ids and will
2724 * grant the double lock to lower CPUs over higher ids under contention,
2725 * regardless of entry order into the function.
2726 */
2727static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2728 __releases(this_rq->lock)
2729 __acquires(busiest->lock)
2730 __acquires(this_rq->lock)
2731{
2732 if (__rq_lockp(this_rq) == __rq_lockp(busiest) ||
2733 likely(raw_spin_rq_trylock(busiest))) {
2734 double_rq_clock_clear_update(this_rq, busiest);
2735 return 0;
2736 }
2737
2738 if (rq_order_less(this_rq, busiest)) {
2739 raw_spin_rq_lock_nested(busiest, SINGLE_DEPTH_NESTING);
2740 double_rq_clock_clear_update(this_rq, busiest);
2741 return 0;
2742 }
2743
2744 raw_spin_rq_unlock(this_rq);
2745 double_rq_lock(this_rq, busiest);
2746
2747 return 1;
2748}
2749
2750#endif /* CONFIG_PREEMPTION */
2751
2752/*
2753 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2754 */
2755static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2756{
2757 lockdep_assert_irqs_disabled();
2758
2759 return _double_lock_balance(this_rq, busiest);
2760}
2761
2762static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2763 __releases(busiest->lock)
2764{
2765 if (__rq_lockp(this_rq) != __rq_lockp(busiest))
2766 raw_spin_rq_unlock(busiest);
2767 lock_set_subclass(&__rq_lockp(this_rq)->dep_map, 0, _RET_IP_);
2768}
2769
2770static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
2771{
2772 if (l1 > l2)
2773 swap(l1, l2);
2774
2775 spin_lock(l1);
2776 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2777}
2778
2779static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
2780{
2781 if (l1 > l2)
2782 swap(l1, l2);
2783
2784 spin_lock_irq(l1);
2785 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2786}
2787
2788static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2789{
2790 if (l1 > l2)
2791 swap(l1, l2);
2792
2793 raw_spin_lock(l1);
2794 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2795}
2796
2797static inline void double_raw_unlock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2798{
2799 raw_spin_unlock(l1);
2800 raw_spin_unlock(l2);
2801}
2802
2803DEFINE_LOCK_GUARD_2(double_raw_spinlock, raw_spinlock_t,
2804 double_raw_lock(_T->lock, _T->lock2),
2805 double_raw_unlock(_T->lock, _T->lock2))
2806
2807/*
2808 * double_rq_unlock - safely unlock two runqueues
2809 *
2810 * Note this does not restore interrupts like task_rq_unlock,
2811 * you need to do so manually after calling.
2812 */
2813static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2814 __releases(rq1->lock)
2815 __releases(rq2->lock)
2816{
2817 if (__rq_lockp(rq1) != __rq_lockp(rq2))
2818 raw_spin_rq_unlock(rq2);
2819 else
2820 __release(rq2->lock);
2821 raw_spin_rq_unlock(rq1);
2822}
2823
2824extern void set_rq_online (struct rq *rq);
2825extern void set_rq_offline(struct rq *rq);
2826extern bool sched_smp_initialized;
2827
2828#else /* CONFIG_SMP */
2829
2830/*
2831 * double_rq_lock - safely lock two runqueues
2832 *
2833 * Note this does not disable interrupts like task_rq_lock,
2834 * you need to do so manually before calling.
2835 */
2836static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
2837 __acquires(rq1->lock)
2838 __acquires(rq2->lock)
2839{
2840 WARN_ON_ONCE(!irqs_disabled());
2841 WARN_ON_ONCE(rq1 != rq2);
2842 raw_spin_rq_lock(rq1);
2843 __acquire(rq2->lock); /* Fake it out ;) */
2844 double_rq_clock_clear_update(rq1, rq2);
2845}
2846
2847/*
2848 * double_rq_unlock - safely unlock two runqueues
2849 *
2850 * Note this does not restore interrupts like task_rq_unlock,
2851 * you need to do so manually after calling.
2852 */
2853static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2854 __releases(rq1->lock)
2855 __releases(rq2->lock)
2856{
2857 WARN_ON_ONCE(rq1 != rq2);
2858 raw_spin_rq_unlock(rq1);
2859 __release(rq2->lock);
2860}
2861
2862#endif
2863
2864DEFINE_LOCK_GUARD_2(double_rq_lock, struct rq,
2865 double_rq_lock(_T->lock, _T->lock2),
2866 double_rq_unlock(_T->lock, _T->lock2))
2867
2868extern struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq);
2869extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
2870extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
2871
2872#ifdef CONFIG_SCHED_DEBUG
2873extern bool sched_debug_verbose;
2874
2875extern void print_cfs_stats(struct seq_file *m, int cpu);
2876extern void print_rt_stats(struct seq_file *m, int cpu);
2877extern void print_dl_stats(struct seq_file *m, int cpu);
2878extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
2879extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2880extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
2881
2882extern void resched_latency_warn(int cpu, u64 latency);
2883#ifdef CONFIG_NUMA_BALANCING
2884extern void
2885show_numa_stats(struct task_struct *p, struct seq_file *m);
2886extern void
2887print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
2888 unsigned long tpf, unsigned long gsf, unsigned long gpf);
2889#endif /* CONFIG_NUMA_BALANCING */
2890#else
2891static inline void resched_latency_warn(int cpu, u64 latency) {}
2892#endif /* CONFIG_SCHED_DEBUG */
2893
2894extern void init_cfs_rq(struct cfs_rq *cfs_rq);
2895extern void init_rt_rq(struct rt_rq *rt_rq);
2896extern void init_dl_rq(struct dl_rq *dl_rq);
2897
2898extern void cfs_bandwidth_usage_inc(void);
2899extern void cfs_bandwidth_usage_dec(void);
2900
2901#ifdef CONFIG_NO_HZ_COMMON
2902#define NOHZ_BALANCE_KICK_BIT 0
2903#define NOHZ_STATS_KICK_BIT 1
2904#define NOHZ_NEWILB_KICK_BIT 2
2905#define NOHZ_NEXT_KICK_BIT 3
2906
2907/* Run rebalance_domains() */
2908#define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT)
2909/* Update blocked load */
2910#define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT)
2911/* Update blocked load when entering idle */
2912#define NOHZ_NEWILB_KICK BIT(NOHZ_NEWILB_KICK_BIT)
2913/* Update nohz.next_balance */
2914#define NOHZ_NEXT_KICK BIT(NOHZ_NEXT_KICK_BIT)
2915
2916#define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK | NOHZ_NEXT_KICK)
2917
2918#define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
2919
2920extern void nohz_balance_exit_idle(struct rq *rq);
2921#else
2922static inline void nohz_balance_exit_idle(struct rq *rq) { }
2923#endif
2924
2925#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
2926extern void nohz_run_idle_balance(int cpu);
2927#else
2928static inline void nohz_run_idle_balance(int cpu) { }
2929#endif
2930
2931#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2932struct irqtime {
2933 u64 total;
2934 u64 tick_delta;
2935 u64 irq_start_time;
2936 struct u64_stats_sync sync;
2937};
2938
2939DECLARE_PER_CPU(struct irqtime, cpu_irqtime);
2940
2941/*
2942 * Returns the irqtime minus the softirq time computed by ksoftirqd.
2943 * Otherwise ksoftirqd's sum_exec_runtime is subtracted its own runtime
2944 * and never move forward.
2945 */
2946static inline u64 irq_time_read(int cpu)
2947{
2948 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu);
2949 unsigned int seq;
2950 u64 total;
2951
2952 do {
2953 seq = __u64_stats_fetch_begin(&irqtime->sync);
2954 total = irqtime->total;
2955 } while (__u64_stats_fetch_retry(&irqtime->sync, seq));
2956
2957 return total;
2958}
2959#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2960
2961#ifdef CONFIG_CPU_FREQ
2962DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data);
2963
2964/**
2965 * cpufreq_update_util - Take a note about CPU utilization changes.
2966 * @rq: Runqueue to carry out the update for.
2967 * @flags: Update reason flags.
2968 *
2969 * This function is called by the scheduler on the CPU whose utilization is
2970 * being updated.
2971 *
2972 * It can only be called from RCU-sched read-side critical sections.
2973 *
2974 * The way cpufreq is currently arranged requires it to evaluate the CPU
2975 * performance state (frequency/voltage) on a regular basis to prevent it from
2976 * being stuck in a completely inadequate performance level for too long.
2977 * That is not guaranteed to happen if the updates are only triggered from CFS
2978 * and DL, though, because they may not be coming in if only RT tasks are
2979 * active all the time (or there are RT tasks only).
2980 *
2981 * As a workaround for that issue, this function is called periodically by the
2982 * RT sched class to trigger extra cpufreq updates to prevent it from stalling,
2983 * but that really is a band-aid. Going forward it should be replaced with
2984 * solutions targeted more specifically at RT tasks.
2985 */
2986static inline void cpufreq_update_util(struct rq *rq, unsigned int flags)
2987{
2988 struct update_util_data *data;
2989
2990 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data,
2991 cpu_of(rq)));
2992 if (data)
2993 data->func(data, rq_clock(rq), flags);
2994}
2995#else
2996static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
2997#endif /* CONFIG_CPU_FREQ */
2998
2999#ifdef arch_scale_freq_capacity
3000# ifndef arch_scale_freq_invariant
3001# define arch_scale_freq_invariant() true
3002# endif
3003#else
3004# define arch_scale_freq_invariant() false
3005#endif
3006
3007#ifdef CONFIG_SMP
3008unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
3009 unsigned long *min,
3010 unsigned long *max);
3011
3012unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
3013 unsigned long min,
3014 unsigned long max);
3015
3016
3017/*
3018 * Verify the fitness of task @p to run on @cpu taking into account the
3019 * CPU original capacity and the runtime/deadline ratio of the task.
3020 *
3021 * The function will return true if the original capacity of @cpu is
3022 * greater than or equal to task's deadline density right shifted by
3023 * (BW_SHIFT - SCHED_CAPACITY_SHIFT) and false otherwise.
3024 */
3025static inline bool dl_task_fits_capacity(struct task_struct *p, int cpu)
3026{
3027 unsigned long cap = arch_scale_cpu_capacity(cpu);
3028
3029 return cap >= p->dl.dl_density >> (BW_SHIFT - SCHED_CAPACITY_SHIFT);
3030}
3031
3032static inline unsigned long cpu_bw_dl(struct rq *rq)
3033{
3034 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT;
3035}
3036
3037static inline unsigned long cpu_util_dl(struct rq *rq)
3038{
3039 return READ_ONCE(rq->avg_dl.util_avg);
3040}
3041
3042
3043extern unsigned long cpu_util_cfs(int cpu);
3044extern unsigned long cpu_util_cfs_boost(int cpu);
3045
3046static inline unsigned long cpu_util_rt(struct rq *rq)
3047{
3048 return READ_ONCE(rq->avg_rt.util_avg);
3049}
3050#endif
3051
3052#ifdef CONFIG_UCLAMP_TASK
3053unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id);
3054
3055static inline unsigned long uclamp_rq_get(struct rq *rq,
3056 enum uclamp_id clamp_id)
3057{
3058 return READ_ONCE(rq->uclamp[clamp_id].value);
3059}
3060
3061static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3062 unsigned int value)
3063{
3064 WRITE_ONCE(rq->uclamp[clamp_id].value, value);
3065}
3066
3067static inline bool uclamp_rq_is_idle(struct rq *rq)
3068{
3069 return rq->uclamp_flags & UCLAMP_FLAG_IDLE;
3070}
3071
3072/* Is the rq being capped/throttled by uclamp_max? */
3073static inline bool uclamp_rq_is_capped(struct rq *rq)
3074{
3075 unsigned long rq_util;
3076 unsigned long max_util;
3077
3078 if (!static_branch_likely(&sched_uclamp_used))
3079 return false;
3080
3081 rq_util = cpu_util_cfs(cpu_of(rq)) + cpu_util_rt(rq);
3082 max_util = READ_ONCE(rq->uclamp[UCLAMP_MAX].value);
3083
3084 return max_util != SCHED_CAPACITY_SCALE && rq_util >= max_util;
3085}
3086
3087/*
3088 * When uclamp is compiled in, the aggregation at rq level is 'turned off'
3089 * by default in the fast path and only gets turned on once userspace performs
3090 * an operation that requires it.
3091 *
3092 * Returns true if userspace opted-in to use uclamp and aggregation at rq level
3093 * hence is active.
3094 */
3095static inline bool uclamp_is_used(void)
3096{
3097 return static_branch_likely(&sched_uclamp_used);
3098}
3099#else /* CONFIG_UCLAMP_TASK */
3100static inline unsigned long uclamp_eff_value(struct task_struct *p,
3101 enum uclamp_id clamp_id)
3102{
3103 if (clamp_id == UCLAMP_MIN)
3104 return 0;
3105
3106 return SCHED_CAPACITY_SCALE;
3107}
3108
3109static inline bool uclamp_rq_is_capped(struct rq *rq) { return false; }
3110
3111static inline bool uclamp_is_used(void)
3112{
3113 return false;
3114}
3115
3116static inline unsigned long uclamp_rq_get(struct rq *rq,
3117 enum uclamp_id clamp_id)
3118{
3119 if (clamp_id == UCLAMP_MIN)
3120 return 0;
3121
3122 return SCHED_CAPACITY_SCALE;
3123}
3124
3125static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3126 unsigned int value)
3127{
3128}
3129
3130static inline bool uclamp_rq_is_idle(struct rq *rq)
3131{
3132 return false;
3133}
3134#endif /* CONFIG_UCLAMP_TASK */
3135
3136#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
3137static inline unsigned long cpu_util_irq(struct rq *rq)
3138{
3139 return rq->avg_irq.util_avg;
3140}
3141
3142static inline
3143unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3144{
3145 util *= (max - irq);
3146 util /= max;
3147
3148 return util;
3149
3150}
3151#else
3152static inline unsigned long cpu_util_irq(struct rq *rq)
3153{
3154 return 0;
3155}
3156
3157static inline
3158unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3159{
3160 return util;
3161}
3162#endif
3163
3164#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
3165
3166#define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus)))
3167
3168DECLARE_STATIC_KEY_FALSE(sched_energy_present);
3169
3170static inline bool sched_energy_enabled(void)
3171{
3172 return static_branch_unlikely(&sched_energy_present);
3173}
3174
3175extern struct cpufreq_governor schedutil_gov;
3176
3177#else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */
3178
3179#define perf_domain_span(pd) NULL
3180static inline bool sched_energy_enabled(void) { return false; }
3181
3182#endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
3183
3184#ifdef CONFIG_MEMBARRIER
3185/*
3186 * The scheduler provides memory barriers required by membarrier between:
3187 * - prior user-space memory accesses and store to rq->membarrier_state,
3188 * - store to rq->membarrier_state and following user-space memory accesses.
3189 * In the same way it provides those guarantees around store to rq->curr.
3190 */
3191static inline void membarrier_switch_mm(struct rq *rq,
3192 struct mm_struct *prev_mm,
3193 struct mm_struct *next_mm)
3194{
3195 int membarrier_state;
3196
3197 if (prev_mm == next_mm)
3198 return;
3199
3200 membarrier_state = atomic_read(&next_mm->membarrier_state);
3201 if (READ_ONCE(rq->membarrier_state) == membarrier_state)
3202 return;
3203
3204 WRITE_ONCE(rq->membarrier_state, membarrier_state);
3205}
3206#else
3207static inline void membarrier_switch_mm(struct rq *rq,
3208 struct mm_struct *prev_mm,
3209 struct mm_struct *next_mm)
3210{
3211}
3212#endif
3213
3214#ifdef CONFIG_SMP
3215static inline bool is_per_cpu_kthread(struct task_struct *p)
3216{
3217 if (!(p->flags & PF_KTHREAD))
3218 return false;
3219
3220 if (p->nr_cpus_allowed != 1)
3221 return false;
3222
3223 return true;
3224}
3225#endif
3226
3227extern void swake_up_all_locked(struct swait_queue_head *q);
3228extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);
3229
3230extern int try_to_wake_up(struct task_struct *tsk, unsigned int state, int wake_flags);
3231
3232#ifdef CONFIG_PREEMPT_DYNAMIC
3233extern int preempt_dynamic_mode;
3234extern int sched_dynamic_mode(const char *str);
3235extern void sched_dynamic_update(int mode);
3236#endif
3237
3238#ifdef CONFIG_SCHED_MM_CID
3239
3240#define SCHED_MM_CID_PERIOD_NS (100ULL * 1000000) /* 100ms */
3241#define MM_CID_SCAN_DELAY 100 /* 100ms */
3242
3243extern raw_spinlock_t cid_lock;
3244extern int use_cid_lock;
3245
3246extern void sched_mm_cid_migrate_from(struct task_struct *t);
3247extern void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t);
3248extern void task_tick_mm_cid(struct rq *rq, struct task_struct *curr);
3249extern void init_sched_mm_cid(struct task_struct *t);
3250
3251static inline void __mm_cid_put(struct mm_struct *mm, int cid)
3252{
3253 if (cid < 0)
3254 return;
3255 cpumask_clear_cpu(cid, mm_cidmask(mm));
3256}
3257
3258/*
3259 * The per-mm/cpu cid can have the MM_CID_LAZY_PUT flag set or transition to
3260 * the MM_CID_UNSET state without holding the rq lock, but the rq lock needs to
3261 * be held to transition to other states.
3262 *
3263 * State transitions synchronized with cmpxchg or try_cmpxchg need to be
3264 * consistent across cpus, which prevents use of this_cpu_cmpxchg.
3265 */
3266static inline void mm_cid_put_lazy(struct task_struct *t)
3267{
3268 struct mm_struct *mm = t->mm;
3269 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3270 int cid;
3271
3272 lockdep_assert_irqs_disabled();
3273 cid = __this_cpu_read(pcpu_cid->cid);
3274 if (!mm_cid_is_lazy_put(cid) ||
3275 !try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3276 return;
3277 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3278}
3279
3280static inline int mm_cid_pcpu_unset(struct mm_struct *mm)
3281{
3282 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3283 int cid, res;
3284
3285 lockdep_assert_irqs_disabled();
3286 cid = __this_cpu_read(pcpu_cid->cid);
3287 for (;;) {
3288 if (mm_cid_is_unset(cid))
3289 return MM_CID_UNSET;
3290 /*
3291 * Attempt transition from valid or lazy-put to unset.
3292 */
3293 res = cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, cid, MM_CID_UNSET);
3294 if (res == cid)
3295 break;
3296 cid = res;
3297 }
3298 return cid;
3299}
3300
3301static inline void mm_cid_put(struct mm_struct *mm)
3302{
3303 int cid;
3304
3305 lockdep_assert_irqs_disabled();
3306 cid = mm_cid_pcpu_unset(mm);
3307 if (cid == MM_CID_UNSET)
3308 return;
3309 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3310}
3311
3312static inline int __mm_cid_try_get(struct mm_struct *mm)
3313{
3314 struct cpumask *cpumask;
3315 int cid;
3316
3317 cpumask = mm_cidmask(mm);
3318 /*
3319 * Retry finding first zero bit if the mask is temporarily
3320 * filled. This only happens during concurrent remote-clear
3321 * which owns a cid without holding a rq lock.
3322 */
3323 for (;;) {
3324 cid = cpumask_first_zero(cpumask);
3325 if (cid < nr_cpu_ids)
3326 break;
3327 cpu_relax();
3328 }
3329 if (cpumask_test_and_set_cpu(cid, cpumask))
3330 return -1;
3331 return cid;
3332}
3333
3334/*
3335 * Save a snapshot of the current runqueue time of this cpu
3336 * with the per-cpu cid value, allowing to estimate how recently it was used.
3337 */
3338static inline void mm_cid_snapshot_time(struct rq *rq, struct mm_struct *mm)
3339{
3340 struct mm_cid *pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(rq));
3341
3342 lockdep_assert_rq_held(rq);
3343 WRITE_ONCE(pcpu_cid->time, rq->clock);
3344}
3345
3346static inline int __mm_cid_get(struct rq *rq, struct mm_struct *mm)
3347{
3348 int cid;
3349
3350 /*
3351 * All allocations (even those using the cid_lock) are lock-free. If
3352 * use_cid_lock is set, hold the cid_lock to perform cid allocation to
3353 * guarantee forward progress.
3354 */
3355 if (!READ_ONCE(use_cid_lock)) {
3356 cid = __mm_cid_try_get(mm);
3357 if (cid >= 0)
3358 goto end;
3359 raw_spin_lock(&cid_lock);
3360 } else {
3361 raw_spin_lock(&cid_lock);
3362 cid = __mm_cid_try_get(mm);
3363 if (cid >= 0)
3364 goto unlock;
3365 }
3366
3367 /*
3368 * cid concurrently allocated. Retry while forcing following
3369 * allocations to use the cid_lock to ensure forward progress.
3370 */
3371 WRITE_ONCE(use_cid_lock, 1);
3372 /*
3373 * Set use_cid_lock before allocation. Only care about program order
3374 * because this is only required for forward progress.
3375 */
3376 barrier();
3377 /*
3378 * Retry until it succeeds. It is guaranteed to eventually succeed once
3379 * all newcoming allocations observe the use_cid_lock flag set.
3380 */
3381 do {
3382 cid = __mm_cid_try_get(mm);
3383 cpu_relax();
3384 } while (cid < 0);
3385 /*
3386 * Allocate before clearing use_cid_lock. Only care about
3387 * program order because this is for forward progress.
3388 */
3389 barrier();
3390 WRITE_ONCE(use_cid_lock, 0);
3391unlock:
3392 raw_spin_unlock(&cid_lock);
3393end:
3394 mm_cid_snapshot_time(rq, mm);
3395 return cid;
3396}
3397
3398static inline int mm_cid_get(struct rq *rq, struct mm_struct *mm)
3399{
3400 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3401 struct cpumask *cpumask;
3402 int cid;
3403
3404 lockdep_assert_rq_held(rq);
3405 cpumask = mm_cidmask(mm);
3406 cid = __this_cpu_read(pcpu_cid->cid);
3407 if (mm_cid_is_valid(cid)) {
3408 mm_cid_snapshot_time(rq, mm);
3409 return cid;
3410 }
3411 if (mm_cid_is_lazy_put(cid)) {
3412 if (try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3413 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3414 }
3415 cid = __mm_cid_get(rq, mm);
3416 __this_cpu_write(pcpu_cid->cid, cid);
3417 return cid;
3418}
3419
3420static inline void switch_mm_cid(struct rq *rq,
3421 struct task_struct *prev,
3422 struct task_struct *next)
3423{
3424 /*
3425 * Provide a memory barrier between rq->curr store and load of
3426 * {prev,next}->mm->pcpu_cid[cpu] on rq->curr->mm transition.
3427 *
3428 * Should be adapted if context_switch() is modified.
3429 */
3430 if (!next->mm) { // to kernel
3431 /*
3432 * user -> kernel transition does not guarantee a barrier, but
3433 * we can use the fact that it performs an atomic operation in
3434 * mmgrab().
3435 */
3436 if (prev->mm) // from user
3437 smp_mb__after_mmgrab();
3438 /*
3439 * kernel -> kernel transition does not change rq->curr->mm
3440 * state. It stays NULL.
3441 */
3442 } else { // to user
3443 /*
3444 * kernel -> user transition does not provide a barrier
3445 * between rq->curr store and load of {prev,next}->mm->pcpu_cid[cpu].
3446 * Provide it here.
3447 */
3448 if (!prev->mm) // from kernel
3449 smp_mb();
3450 /*
3451 * user -> user transition guarantees a memory barrier through
3452 * switch_mm() when current->mm changes. If current->mm is
3453 * unchanged, no barrier is needed.
3454 */
3455 }
3456 if (prev->mm_cid_active) {
3457 mm_cid_snapshot_time(rq, prev->mm);
3458 mm_cid_put_lazy(prev);
3459 prev->mm_cid = -1;
3460 }
3461 if (next->mm_cid_active)
3462 next->last_mm_cid = next->mm_cid = mm_cid_get(rq, next->mm);
3463}
3464
3465#else
3466static inline void switch_mm_cid(struct rq *rq, struct task_struct *prev, struct task_struct *next) { }
3467static inline void sched_mm_cid_migrate_from(struct task_struct *t) { }
3468static inline void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t) { }
3469static inline void task_tick_mm_cid(struct rq *rq, struct task_struct *curr) { }
3470static inline void init_sched_mm_cid(struct task_struct *t) { }
3471#endif
3472
3473extern u64 avg_vruntime(struct cfs_rq *cfs_rq);
3474extern int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se);
3475
3476#endif /* _KERNEL_SCHED_SCHED_H */
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 * Scheduler internal types and methods:
4 */
5#ifndef _KERNEL_SCHED_SCHED_H
6#define _KERNEL_SCHED_SCHED_H
7
8#include <linux/sched/affinity.h>
9#include <linux/sched/autogroup.h>
10#include <linux/sched/cpufreq.h>
11#include <linux/sched/deadline.h>
12#include <linux/sched.h>
13#include <linux/sched/loadavg.h>
14#include <linux/sched/mm.h>
15#include <linux/sched/rseq_api.h>
16#include <linux/sched/signal.h>
17#include <linux/sched/smt.h>
18#include <linux/sched/stat.h>
19#include <linux/sched/sysctl.h>
20#include <linux/sched/task_flags.h>
21#include <linux/sched/task.h>
22#include <linux/sched/topology.h>
23
24#include <linux/atomic.h>
25#include <linux/bitmap.h>
26#include <linux/bug.h>
27#include <linux/capability.h>
28#include <linux/cgroup_api.h>
29#include <linux/cgroup.h>
30#include <linux/context_tracking.h>
31#include <linux/cpufreq.h>
32#include <linux/cpumask_api.h>
33#include <linux/ctype.h>
34#include <linux/file.h>
35#include <linux/fs_api.h>
36#include <linux/hrtimer_api.h>
37#include <linux/interrupt.h>
38#include <linux/irq_work.h>
39#include <linux/jiffies.h>
40#include <linux/kref_api.h>
41#include <linux/kthread.h>
42#include <linux/ktime_api.h>
43#include <linux/lockdep_api.h>
44#include <linux/lockdep.h>
45#include <linux/minmax.h>
46#include <linux/mm.h>
47#include <linux/module.h>
48#include <linux/mutex_api.h>
49#include <linux/plist.h>
50#include <linux/poll.h>
51#include <linux/proc_fs.h>
52#include <linux/profile.h>
53#include <linux/psi.h>
54#include <linux/rcupdate.h>
55#include <linux/seq_file.h>
56#include <linux/seqlock.h>
57#include <linux/softirq.h>
58#include <linux/spinlock_api.h>
59#include <linux/static_key.h>
60#include <linux/stop_machine.h>
61#include <linux/syscalls_api.h>
62#include <linux/syscalls.h>
63#include <linux/tick.h>
64#include <linux/topology.h>
65#include <linux/types.h>
66#include <linux/u64_stats_sync_api.h>
67#include <linux/uaccess.h>
68#include <linux/wait_api.h>
69#include <linux/wait_bit.h>
70#include <linux/workqueue_api.h>
71
72#include <trace/events/power.h>
73#include <trace/events/sched.h>
74
75#include "../workqueue_internal.h"
76
77#ifdef CONFIG_PARAVIRT
78# include <asm/paravirt.h>
79# include <asm/paravirt_api_clock.h>
80#endif
81
82#include <asm/barrier.h>
83
84#include "cpupri.h"
85#include "cpudeadline.h"
86
87#ifdef CONFIG_SCHED_DEBUG
88# define SCHED_WARN_ON(x) WARN_ONCE(x, #x)
89#else
90# define SCHED_WARN_ON(x) ({ (void)(x), 0; })
91#endif
92
93struct rq;
94struct cpuidle_state;
95
96/* task_struct::on_rq states: */
97#define TASK_ON_RQ_QUEUED 1
98#define TASK_ON_RQ_MIGRATING 2
99
100extern __read_mostly int scheduler_running;
101
102extern unsigned long calc_load_update;
103extern atomic_long_t calc_load_tasks;
104
105extern void calc_global_load_tick(struct rq *this_rq);
106extern long calc_load_fold_active(struct rq *this_rq, long adjust);
107
108extern void call_trace_sched_update_nr_running(struct rq *rq, int count);
109
110extern int sysctl_sched_rt_period;
111extern int sysctl_sched_rt_runtime;
112extern int sched_rr_timeslice;
113
114/*
115 * Helpers for converting nanosecond timing to jiffy resolution
116 */
117#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
118
119/*
120 * Increase resolution of nice-level calculations for 64-bit architectures.
121 * The extra resolution improves shares distribution and load balancing of
122 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
123 * hierarchies, especially on larger systems. This is not a user-visible change
124 * and does not change the user-interface for setting shares/weights.
125 *
126 * We increase resolution only if we have enough bits to allow this increased
127 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
128 * are pretty high and the returns do not justify the increased costs.
129 *
130 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to
131 * increase coverage and consistency always enable it on 64-bit platforms.
132 */
133#ifdef CONFIG_64BIT
134# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT)
135# define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT)
136# define scale_load_down(w) \
137({ \
138 unsigned long __w = (w); \
139 if (__w) \
140 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \
141 __w; \
142})
143#else
144# define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT)
145# define scale_load(w) (w)
146# define scale_load_down(w) (w)
147#endif
148
149/*
150 * Task weight (visible to users) and its load (invisible to users) have
151 * independent resolution, but they should be well calibrated. We use
152 * scale_load() and scale_load_down(w) to convert between them. The
153 * following must be true:
154 *
155 * scale_load(sched_prio_to_weight[NICE_TO_PRIO(0)-MAX_RT_PRIO]) == NICE_0_LOAD
156 *
157 */
158#define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT)
159
160/*
161 * Single value that decides SCHED_DEADLINE internal math precision.
162 * 10 -> just above 1us
163 * 9 -> just above 0.5us
164 */
165#define DL_SCALE 10
166
167/*
168 * Single value that denotes runtime == period, ie unlimited time.
169 */
170#define RUNTIME_INF ((u64)~0ULL)
171
172static inline int idle_policy(int policy)
173{
174 return policy == SCHED_IDLE;
175}
176static inline int fair_policy(int policy)
177{
178 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
179}
180
181static inline int rt_policy(int policy)
182{
183 return policy == SCHED_FIFO || policy == SCHED_RR;
184}
185
186static inline int dl_policy(int policy)
187{
188 return policy == SCHED_DEADLINE;
189}
190static inline bool valid_policy(int policy)
191{
192 return idle_policy(policy) || fair_policy(policy) ||
193 rt_policy(policy) || dl_policy(policy);
194}
195
196static inline int task_has_idle_policy(struct task_struct *p)
197{
198 return idle_policy(p->policy);
199}
200
201static inline int task_has_rt_policy(struct task_struct *p)
202{
203 return rt_policy(p->policy);
204}
205
206static inline int task_has_dl_policy(struct task_struct *p)
207{
208 return dl_policy(p->policy);
209}
210
211#define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
212
213static inline void update_avg(u64 *avg, u64 sample)
214{
215 s64 diff = sample - *avg;
216 *avg += diff / 8;
217}
218
219/*
220 * Shifting a value by an exponent greater *or equal* to the size of said value
221 * is UB; cap at size-1.
222 */
223#define shr_bound(val, shift) \
224 (val >> min_t(typeof(shift), shift, BITS_PER_TYPE(typeof(val)) - 1))
225
226/*
227 * !! For sched_setattr_nocheck() (kernel) only !!
228 *
229 * This is actually gross. :(
230 *
231 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE
232 * tasks, but still be able to sleep. We need this on platforms that cannot
233 * atomically change clock frequency. Remove once fast switching will be
234 * available on such platforms.
235 *
236 * SUGOV stands for SchedUtil GOVernor.
237 */
238#define SCHED_FLAG_SUGOV 0x10000000
239
240#define SCHED_DL_FLAGS (SCHED_FLAG_RECLAIM | SCHED_FLAG_DL_OVERRUN | SCHED_FLAG_SUGOV)
241
242static inline bool dl_entity_is_special(const struct sched_dl_entity *dl_se)
243{
244#ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
245 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV);
246#else
247 return false;
248#endif
249}
250
251/*
252 * Tells if entity @a should preempt entity @b.
253 */
254static inline bool dl_entity_preempt(const struct sched_dl_entity *a,
255 const struct sched_dl_entity *b)
256{
257 return dl_entity_is_special(a) ||
258 dl_time_before(a->deadline, b->deadline);
259}
260
261/*
262 * This is the priority-queue data structure of the RT scheduling class:
263 */
264struct rt_prio_array {
265 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
266 struct list_head queue[MAX_RT_PRIO];
267};
268
269struct rt_bandwidth {
270 /* nests inside the rq lock: */
271 raw_spinlock_t rt_runtime_lock;
272 ktime_t rt_period;
273 u64 rt_runtime;
274 struct hrtimer rt_period_timer;
275 unsigned int rt_period_active;
276};
277
278static inline int dl_bandwidth_enabled(void)
279{
280 return sysctl_sched_rt_runtime >= 0;
281}
282
283/*
284 * To keep the bandwidth of -deadline tasks under control
285 * we need some place where:
286 * - store the maximum -deadline bandwidth of each cpu;
287 * - cache the fraction of bandwidth that is currently allocated in
288 * each root domain;
289 *
290 * This is all done in the data structure below. It is similar to the
291 * one used for RT-throttling (rt_bandwidth), with the main difference
292 * that, since here we are only interested in admission control, we
293 * do not decrease any runtime while the group "executes", neither we
294 * need a timer to replenish it.
295 *
296 * With respect to SMP, bandwidth is given on a per root domain basis,
297 * meaning that:
298 * - bw (< 100%) is the deadline bandwidth of each CPU;
299 * - total_bw is the currently allocated bandwidth in each root domain;
300 */
301struct dl_bw {
302 raw_spinlock_t lock;
303 u64 bw;
304 u64 total_bw;
305};
306
307extern void init_dl_bw(struct dl_bw *dl_b);
308extern int sched_dl_global_validate(void);
309extern void sched_dl_do_global(void);
310extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr);
311extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr);
312extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr);
313extern bool __checkparam_dl(const struct sched_attr *attr);
314extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr);
315extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial);
316extern int dl_bw_check_overflow(int cpu);
317
318/*
319 * SCHED_DEADLINE supports servers (nested scheduling) with the following
320 * interface:
321 *
322 * dl_se::rq -- runqueue we belong to.
323 *
324 * dl_se::server_has_tasks() -- used on bandwidth enforcement; we 'stop' the
325 * server when it runs out of tasks to run.
326 *
327 * dl_se::server_pick() -- nested pick_next_task(); we yield the period if this
328 * returns NULL.
329 *
330 * dl_server_update() -- called from update_curr_common(), propagates runtime
331 * to the server.
332 *
333 * dl_server_start()
334 * dl_server_stop() -- start/stop the server when it has (no) tasks.
335 *
336 * dl_server_init() -- initializes the server.
337 */
338extern void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec);
339extern void dl_server_start(struct sched_dl_entity *dl_se);
340extern void dl_server_stop(struct sched_dl_entity *dl_se);
341extern void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
342 dl_server_has_tasks_f has_tasks,
343 dl_server_pick_f pick);
344
345#ifdef CONFIG_CGROUP_SCHED
346
347struct cfs_rq;
348struct rt_rq;
349
350extern struct list_head task_groups;
351
352struct cfs_bandwidth {
353#ifdef CONFIG_CFS_BANDWIDTH
354 raw_spinlock_t lock;
355 ktime_t period;
356 u64 quota;
357 u64 runtime;
358 u64 burst;
359 u64 runtime_snap;
360 s64 hierarchical_quota;
361
362 u8 idle;
363 u8 period_active;
364 u8 slack_started;
365 struct hrtimer period_timer;
366 struct hrtimer slack_timer;
367 struct list_head throttled_cfs_rq;
368
369 /* Statistics: */
370 int nr_periods;
371 int nr_throttled;
372 int nr_burst;
373 u64 throttled_time;
374 u64 burst_time;
375#endif
376};
377
378/* Task group related information */
379struct task_group {
380 struct cgroup_subsys_state css;
381
382#ifdef CONFIG_FAIR_GROUP_SCHED
383 /* schedulable entities of this group on each CPU */
384 struct sched_entity **se;
385 /* runqueue "owned" by this group on each CPU */
386 struct cfs_rq **cfs_rq;
387 unsigned long shares;
388
389 /* A positive value indicates that this is a SCHED_IDLE group. */
390 int idle;
391
392#ifdef CONFIG_SMP
393 /*
394 * load_avg can be heavily contended at clock tick time, so put
395 * it in its own cacheline separated from the fields above which
396 * will also be accessed at each tick.
397 */
398 atomic_long_t load_avg ____cacheline_aligned;
399#endif
400#endif
401
402#ifdef CONFIG_RT_GROUP_SCHED
403 struct sched_rt_entity **rt_se;
404 struct rt_rq **rt_rq;
405
406 struct rt_bandwidth rt_bandwidth;
407#endif
408
409 struct rcu_head rcu;
410 struct list_head list;
411
412 struct task_group *parent;
413 struct list_head siblings;
414 struct list_head children;
415
416#ifdef CONFIG_SCHED_AUTOGROUP
417 struct autogroup *autogroup;
418#endif
419
420 struct cfs_bandwidth cfs_bandwidth;
421
422#ifdef CONFIG_UCLAMP_TASK_GROUP
423 /* The two decimal precision [%] value requested from user-space */
424 unsigned int uclamp_pct[UCLAMP_CNT];
425 /* Clamp values requested for a task group */
426 struct uclamp_se uclamp_req[UCLAMP_CNT];
427 /* Effective clamp values used for a task group */
428 struct uclamp_se uclamp[UCLAMP_CNT];
429#endif
430
431};
432
433#ifdef CONFIG_FAIR_GROUP_SCHED
434#define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
435
436/*
437 * A weight of 0 or 1 can cause arithmetics problems.
438 * A weight of a cfs_rq is the sum of weights of which entities
439 * are queued on this cfs_rq, so a weight of a entity should not be
440 * too large, so as the shares value of a task group.
441 * (The default weight is 1024 - so there's no practical
442 * limitation from this.)
443 */
444#define MIN_SHARES (1UL << 1)
445#define MAX_SHARES (1UL << 18)
446#endif
447
448typedef int (*tg_visitor)(struct task_group *, void *);
449
450extern int walk_tg_tree_from(struct task_group *from,
451 tg_visitor down, tg_visitor up, void *data);
452
453/*
454 * Iterate the full tree, calling @down when first entering a node and @up when
455 * leaving it for the final time.
456 *
457 * Caller must hold rcu_lock or sufficient equivalent.
458 */
459static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
460{
461 return walk_tg_tree_from(&root_task_group, down, up, data);
462}
463
464extern int tg_nop(struct task_group *tg, void *data);
465
466#ifdef CONFIG_FAIR_GROUP_SCHED
467extern void free_fair_sched_group(struct task_group *tg);
468extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
469extern void online_fair_sched_group(struct task_group *tg);
470extern void unregister_fair_sched_group(struct task_group *tg);
471#else
472static inline void free_fair_sched_group(struct task_group *tg) { }
473static inline int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
474{
475 return 1;
476}
477static inline void online_fair_sched_group(struct task_group *tg) { }
478static inline void unregister_fair_sched_group(struct task_group *tg) { }
479#endif
480
481extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
482 struct sched_entity *se, int cpu,
483 struct sched_entity *parent);
484extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent);
485
486extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
487extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
488extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
489extern bool cfs_task_bw_constrained(struct task_struct *p);
490
491extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
492 struct sched_rt_entity *rt_se, int cpu,
493 struct sched_rt_entity *parent);
494extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us);
495extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us);
496extern long sched_group_rt_runtime(struct task_group *tg);
497extern long sched_group_rt_period(struct task_group *tg);
498extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
499
500extern struct task_group *sched_create_group(struct task_group *parent);
501extern void sched_online_group(struct task_group *tg,
502 struct task_group *parent);
503extern void sched_destroy_group(struct task_group *tg);
504extern void sched_release_group(struct task_group *tg);
505
506extern void sched_move_task(struct task_struct *tsk);
507
508#ifdef CONFIG_FAIR_GROUP_SCHED
509extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
510
511extern int sched_group_set_idle(struct task_group *tg, long idle);
512
513#ifdef CONFIG_SMP
514extern void set_task_rq_fair(struct sched_entity *se,
515 struct cfs_rq *prev, struct cfs_rq *next);
516#else /* !CONFIG_SMP */
517static inline void set_task_rq_fair(struct sched_entity *se,
518 struct cfs_rq *prev, struct cfs_rq *next) { }
519#endif /* CONFIG_SMP */
520#endif /* CONFIG_FAIR_GROUP_SCHED */
521
522#else /* CONFIG_CGROUP_SCHED */
523
524struct cfs_bandwidth { };
525static inline bool cfs_task_bw_constrained(struct task_struct *p) { return false; }
526
527#endif /* CONFIG_CGROUP_SCHED */
528
529extern void unregister_rt_sched_group(struct task_group *tg);
530extern void free_rt_sched_group(struct task_group *tg);
531extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
532
533/*
534 * u64_u32_load/u64_u32_store
535 *
536 * Use a copy of a u64 value to protect against data race. This is only
537 * applicable for 32-bits architectures.
538 */
539#ifdef CONFIG_64BIT
540# define u64_u32_load_copy(var, copy) var
541# define u64_u32_store_copy(var, copy, val) (var = val)
542#else
543# define u64_u32_load_copy(var, copy) \
544({ \
545 u64 __val, __val_copy; \
546 do { \
547 __val_copy = copy; \
548 /* \
549 * paired with u64_u32_store_copy(), ordering access \
550 * to var and copy. \
551 */ \
552 smp_rmb(); \
553 __val = var; \
554 } while (__val != __val_copy); \
555 __val; \
556})
557# define u64_u32_store_copy(var, copy, val) \
558do { \
559 typeof(val) __val = (val); \
560 var = __val; \
561 /* \
562 * paired with u64_u32_load_copy(), ordering access to var and \
563 * copy. \
564 */ \
565 smp_wmb(); \
566 copy = __val; \
567} while (0)
568#endif
569# define u64_u32_load(var) u64_u32_load_copy(var, var##_copy)
570# define u64_u32_store(var, val) u64_u32_store_copy(var, var##_copy, val)
571
572/* CFS-related fields in a runqueue */
573struct cfs_rq {
574 struct load_weight load;
575 unsigned int nr_running;
576 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */
577 unsigned int idle_nr_running; /* SCHED_IDLE */
578 unsigned int idle_h_nr_running; /* SCHED_IDLE */
579
580 s64 avg_vruntime;
581 u64 avg_load;
582
583 u64 exec_clock;
584 u64 min_vruntime;
585#ifdef CONFIG_SCHED_CORE
586 unsigned int forceidle_seq;
587 u64 min_vruntime_fi;
588#endif
589
590#ifndef CONFIG_64BIT
591 u64 min_vruntime_copy;
592#endif
593
594 struct rb_root_cached tasks_timeline;
595
596 /*
597 * 'curr' points to currently running entity on this cfs_rq.
598 * It is set to NULL otherwise (i.e when none are currently running).
599 */
600 struct sched_entity *curr;
601 struct sched_entity *next;
602
603#ifdef CONFIG_SCHED_DEBUG
604 unsigned int nr_spread_over;
605#endif
606
607#ifdef CONFIG_SMP
608 /*
609 * CFS load tracking
610 */
611 struct sched_avg avg;
612#ifndef CONFIG_64BIT
613 u64 last_update_time_copy;
614#endif
615 struct {
616 raw_spinlock_t lock ____cacheline_aligned;
617 int nr;
618 unsigned long load_avg;
619 unsigned long util_avg;
620 unsigned long runnable_avg;
621 } removed;
622
623#ifdef CONFIG_FAIR_GROUP_SCHED
624 u64 last_update_tg_load_avg;
625 unsigned long tg_load_avg_contrib;
626 long propagate;
627 long prop_runnable_sum;
628
629 /*
630 * h_load = weight * f(tg)
631 *
632 * Where f(tg) is the recursive weight fraction assigned to
633 * this group.
634 */
635 unsigned long h_load;
636 u64 last_h_load_update;
637 struct sched_entity *h_load_next;
638#endif /* CONFIG_FAIR_GROUP_SCHED */
639#endif /* CONFIG_SMP */
640
641#ifdef CONFIG_FAIR_GROUP_SCHED
642 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */
643
644 /*
645 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
646 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
647 * (like users, containers etc.)
648 *
649 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU.
650 * This list is used during load balance.
651 */
652 int on_list;
653 struct list_head leaf_cfs_rq_list;
654 struct task_group *tg; /* group that "owns" this runqueue */
655
656 /* Locally cached copy of our task_group's idle value */
657 int idle;
658
659#ifdef CONFIG_CFS_BANDWIDTH
660 int runtime_enabled;
661 s64 runtime_remaining;
662
663 u64 throttled_pelt_idle;
664#ifndef CONFIG_64BIT
665 u64 throttled_pelt_idle_copy;
666#endif
667 u64 throttled_clock;
668 u64 throttled_clock_pelt;
669 u64 throttled_clock_pelt_time;
670 u64 throttled_clock_self;
671 u64 throttled_clock_self_time;
672 int throttled;
673 int throttle_count;
674 struct list_head throttled_list;
675 struct list_head throttled_csd_list;
676#endif /* CONFIG_CFS_BANDWIDTH */
677#endif /* CONFIG_FAIR_GROUP_SCHED */
678};
679
680static inline int rt_bandwidth_enabled(void)
681{
682 return sysctl_sched_rt_runtime >= 0;
683}
684
685/* RT IPI pull logic requires IRQ_WORK */
686#if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP)
687# define HAVE_RT_PUSH_IPI
688#endif
689
690/* Real-Time classes' related field in a runqueue: */
691struct rt_rq {
692 struct rt_prio_array active;
693 unsigned int rt_nr_running;
694 unsigned int rr_nr_running;
695#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
696 struct {
697 int curr; /* highest queued rt task prio */
698#ifdef CONFIG_SMP
699 int next; /* next highest */
700#endif
701 } highest_prio;
702#endif
703#ifdef CONFIG_SMP
704 int overloaded;
705 struct plist_head pushable_tasks;
706
707#endif /* CONFIG_SMP */
708 int rt_queued;
709
710 int rt_throttled;
711 u64 rt_time;
712 u64 rt_runtime;
713 /* Nests inside the rq lock: */
714 raw_spinlock_t rt_runtime_lock;
715
716#ifdef CONFIG_RT_GROUP_SCHED
717 unsigned int rt_nr_boosted;
718
719 struct rq *rq;
720 struct task_group *tg;
721#endif
722};
723
724static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq)
725{
726 return rt_rq->rt_queued && rt_rq->rt_nr_running;
727}
728
729/* Deadline class' related fields in a runqueue */
730struct dl_rq {
731 /* runqueue is an rbtree, ordered by deadline */
732 struct rb_root_cached root;
733
734 unsigned int dl_nr_running;
735
736#ifdef CONFIG_SMP
737 /*
738 * Deadline values of the currently executing and the
739 * earliest ready task on this rq. Caching these facilitates
740 * the decision whether or not a ready but not running task
741 * should migrate somewhere else.
742 */
743 struct {
744 u64 curr;
745 u64 next;
746 } earliest_dl;
747
748 int overloaded;
749
750 /*
751 * Tasks on this rq that can be pushed away. They are kept in
752 * an rb-tree, ordered by tasks' deadlines, with caching
753 * of the leftmost (earliest deadline) element.
754 */
755 struct rb_root_cached pushable_dl_tasks_root;
756#else
757 struct dl_bw dl_bw;
758#endif
759 /*
760 * "Active utilization" for this runqueue: increased when a
761 * task wakes up (becomes TASK_RUNNING) and decreased when a
762 * task blocks
763 */
764 u64 running_bw;
765
766 /*
767 * Utilization of the tasks "assigned" to this runqueue (including
768 * the tasks that are in runqueue and the tasks that executed on this
769 * CPU and blocked). Increased when a task moves to this runqueue, and
770 * decreased when the task moves away (migrates, changes scheduling
771 * policy, or terminates).
772 * This is needed to compute the "inactive utilization" for the
773 * runqueue (inactive utilization = this_bw - running_bw).
774 */
775 u64 this_bw;
776 u64 extra_bw;
777
778 /*
779 * Maximum available bandwidth for reclaiming by SCHED_FLAG_RECLAIM
780 * tasks of this rq. Used in calculation of reclaimable bandwidth(GRUB).
781 */
782 u64 max_bw;
783
784 /*
785 * Inverse of the fraction of CPU utilization that can be reclaimed
786 * by the GRUB algorithm.
787 */
788 u64 bw_ratio;
789};
790
791#ifdef CONFIG_FAIR_GROUP_SCHED
792/* An entity is a task if it doesn't "own" a runqueue */
793#define entity_is_task(se) (!se->my_q)
794
795static inline void se_update_runnable(struct sched_entity *se)
796{
797 if (!entity_is_task(se))
798 se->runnable_weight = se->my_q->h_nr_running;
799}
800
801static inline long se_runnable(struct sched_entity *se)
802{
803 if (entity_is_task(se))
804 return !!se->on_rq;
805 else
806 return se->runnable_weight;
807}
808
809#else
810#define entity_is_task(se) 1
811
812static inline void se_update_runnable(struct sched_entity *se) {}
813
814static inline long se_runnable(struct sched_entity *se)
815{
816 return !!se->on_rq;
817}
818#endif
819
820#ifdef CONFIG_SMP
821/*
822 * XXX we want to get rid of these helpers and use the full load resolution.
823 */
824static inline long se_weight(struct sched_entity *se)
825{
826 return scale_load_down(se->load.weight);
827}
828
829
830static inline bool sched_asym_prefer(int a, int b)
831{
832 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b);
833}
834
835struct perf_domain {
836 struct em_perf_domain *em_pd;
837 struct perf_domain *next;
838 struct rcu_head rcu;
839};
840
841/* Scheduling group status flags */
842#define SG_OVERLOAD 0x1 /* More than one runnable task on a CPU. */
843#define SG_OVERUTILIZED 0x2 /* One or more CPUs are over-utilized. */
844
845/*
846 * We add the notion of a root-domain which will be used to define per-domain
847 * variables. Each exclusive cpuset essentially defines an island domain by
848 * fully partitioning the member CPUs from any other cpuset. Whenever a new
849 * exclusive cpuset is created, we also create and attach a new root-domain
850 * object.
851 *
852 */
853struct root_domain {
854 atomic_t refcount;
855 atomic_t rto_count;
856 struct rcu_head rcu;
857 cpumask_var_t span;
858 cpumask_var_t online;
859
860 /*
861 * Indicate pullable load on at least one CPU, e.g:
862 * - More than one runnable task
863 * - Running task is misfit
864 */
865 int overload;
866
867 /* Indicate one or more cpus over-utilized (tipping point) */
868 int overutilized;
869
870 /*
871 * The bit corresponding to a CPU gets set here if such CPU has more
872 * than one runnable -deadline task (as it is below for RT tasks).
873 */
874 cpumask_var_t dlo_mask;
875 atomic_t dlo_count;
876 struct dl_bw dl_bw;
877 struct cpudl cpudl;
878
879 /*
880 * Indicate whether a root_domain's dl_bw has been checked or
881 * updated. It's monotonously increasing value.
882 *
883 * Also, some corner cases, like 'wrap around' is dangerous, but given
884 * that u64 is 'big enough'. So that shouldn't be a concern.
885 */
886 u64 visit_gen;
887
888#ifdef HAVE_RT_PUSH_IPI
889 /*
890 * For IPI pull requests, loop across the rto_mask.
891 */
892 struct irq_work rto_push_work;
893 raw_spinlock_t rto_lock;
894 /* These are only updated and read within rto_lock */
895 int rto_loop;
896 int rto_cpu;
897 /* These atomics are updated outside of a lock */
898 atomic_t rto_loop_next;
899 atomic_t rto_loop_start;
900#endif
901 /*
902 * The "RT overload" flag: it gets set if a CPU has more than
903 * one runnable RT task.
904 */
905 cpumask_var_t rto_mask;
906 struct cpupri cpupri;
907
908 unsigned long max_cpu_capacity;
909
910 /*
911 * NULL-terminated list of performance domains intersecting with the
912 * CPUs of the rd. Protected by RCU.
913 */
914 struct perf_domain __rcu *pd;
915};
916
917extern void init_defrootdomain(void);
918extern int sched_init_domains(const struct cpumask *cpu_map);
919extern void rq_attach_root(struct rq *rq, struct root_domain *rd);
920extern void sched_get_rd(struct root_domain *rd);
921extern void sched_put_rd(struct root_domain *rd);
922
923#ifdef HAVE_RT_PUSH_IPI
924extern void rto_push_irq_work_func(struct irq_work *work);
925#endif
926#endif /* CONFIG_SMP */
927
928#ifdef CONFIG_UCLAMP_TASK
929/*
930 * struct uclamp_bucket - Utilization clamp bucket
931 * @value: utilization clamp value for tasks on this clamp bucket
932 * @tasks: number of RUNNABLE tasks on this clamp bucket
933 *
934 * Keep track of how many tasks are RUNNABLE for a given utilization
935 * clamp value.
936 */
937struct uclamp_bucket {
938 unsigned long value : bits_per(SCHED_CAPACITY_SCALE);
939 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE);
940};
941
942/*
943 * struct uclamp_rq - rq's utilization clamp
944 * @value: currently active clamp values for a rq
945 * @bucket: utilization clamp buckets affecting a rq
946 *
947 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values.
948 * A clamp value is affecting a rq when there is at least one task RUNNABLE
949 * (or actually running) with that value.
950 *
951 * There are up to UCLAMP_CNT possible different clamp values, currently there
952 * are only two: minimum utilization and maximum utilization.
953 *
954 * All utilization clamping values are MAX aggregated, since:
955 * - for util_min: we want to run the CPU at least at the max of the minimum
956 * utilization required by its currently RUNNABLE tasks.
957 * - for util_max: we want to allow the CPU to run up to the max of the
958 * maximum utilization allowed by its currently RUNNABLE tasks.
959 *
960 * Since on each system we expect only a limited number of different
961 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track
962 * the metrics required to compute all the per-rq utilization clamp values.
963 */
964struct uclamp_rq {
965 unsigned int value;
966 struct uclamp_bucket bucket[UCLAMP_BUCKETS];
967};
968
969DECLARE_STATIC_KEY_FALSE(sched_uclamp_used);
970#endif /* CONFIG_UCLAMP_TASK */
971
972struct rq;
973struct balance_callback {
974 struct balance_callback *next;
975 void (*func)(struct rq *rq);
976};
977
978/*
979 * This is the main, per-CPU runqueue data structure.
980 *
981 * Locking rule: those places that want to lock multiple runqueues
982 * (such as the load balancing or the thread migration code), lock
983 * acquire operations must be ordered by ascending &runqueue.
984 */
985struct rq {
986 /* runqueue lock: */
987 raw_spinlock_t __lock;
988
989 unsigned int nr_running;
990#ifdef CONFIG_NUMA_BALANCING
991 unsigned int nr_numa_running;
992 unsigned int nr_preferred_running;
993 unsigned int numa_migrate_on;
994#endif
995#ifdef CONFIG_NO_HZ_COMMON
996#ifdef CONFIG_SMP
997 unsigned long last_blocked_load_update_tick;
998 unsigned int has_blocked_load;
999 call_single_data_t nohz_csd;
1000#endif /* CONFIG_SMP */
1001 unsigned int nohz_tick_stopped;
1002 atomic_t nohz_flags;
1003#endif /* CONFIG_NO_HZ_COMMON */
1004
1005#ifdef CONFIG_SMP
1006 unsigned int ttwu_pending;
1007#endif
1008 u64 nr_switches;
1009
1010#ifdef CONFIG_UCLAMP_TASK
1011 /* Utilization clamp values based on CPU's RUNNABLE tasks */
1012 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned;
1013 unsigned int uclamp_flags;
1014#define UCLAMP_FLAG_IDLE 0x01
1015#endif
1016
1017 struct cfs_rq cfs;
1018 struct rt_rq rt;
1019 struct dl_rq dl;
1020
1021#ifdef CONFIG_FAIR_GROUP_SCHED
1022 /* list of leaf cfs_rq on this CPU: */
1023 struct list_head leaf_cfs_rq_list;
1024 struct list_head *tmp_alone_branch;
1025#endif /* CONFIG_FAIR_GROUP_SCHED */
1026
1027 /*
1028 * This is part of a global counter where only the total sum
1029 * over all CPUs matters. A task can increase this counter on
1030 * one CPU and if it got migrated afterwards it may decrease
1031 * it on another CPU. Always updated under the runqueue lock:
1032 */
1033 unsigned int nr_uninterruptible;
1034
1035 struct task_struct __rcu *curr;
1036 struct task_struct *idle;
1037 struct task_struct *stop;
1038 unsigned long next_balance;
1039 struct mm_struct *prev_mm;
1040
1041 unsigned int clock_update_flags;
1042 u64 clock;
1043 /* Ensure that all clocks are in the same cache line */
1044 u64 clock_task ____cacheline_aligned;
1045 u64 clock_pelt;
1046 unsigned long lost_idle_time;
1047 u64 clock_pelt_idle;
1048 u64 clock_idle;
1049#ifndef CONFIG_64BIT
1050 u64 clock_pelt_idle_copy;
1051 u64 clock_idle_copy;
1052#endif
1053
1054 atomic_t nr_iowait;
1055
1056#ifdef CONFIG_SCHED_DEBUG
1057 u64 last_seen_need_resched_ns;
1058 int ticks_without_resched;
1059#endif
1060
1061#ifdef CONFIG_MEMBARRIER
1062 int membarrier_state;
1063#endif
1064
1065#ifdef CONFIG_SMP
1066 struct root_domain *rd;
1067 struct sched_domain __rcu *sd;
1068
1069 unsigned long cpu_capacity;
1070
1071 struct balance_callback *balance_callback;
1072
1073 unsigned char nohz_idle_balance;
1074 unsigned char idle_balance;
1075
1076 unsigned long misfit_task_load;
1077
1078 /* For active balancing */
1079 int active_balance;
1080 int push_cpu;
1081 struct cpu_stop_work active_balance_work;
1082
1083 /* CPU of this runqueue: */
1084 int cpu;
1085 int online;
1086
1087 struct list_head cfs_tasks;
1088
1089 struct sched_avg avg_rt;
1090 struct sched_avg avg_dl;
1091#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
1092 struct sched_avg avg_irq;
1093#endif
1094#ifdef CONFIG_SCHED_THERMAL_PRESSURE
1095 struct sched_avg avg_thermal;
1096#endif
1097 u64 idle_stamp;
1098 u64 avg_idle;
1099
1100 /* This is used to determine avg_idle's max value */
1101 u64 max_idle_balance_cost;
1102
1103#ifdef CONFIG_HOTPLUG_CPU
1104 struct rcuwait hotplug_wait;
1105#endif
1106#endif /* CONFIG_SMP */
1107
1108#ifdef CONFIG_IRQ_TIME_ACCOUNTING
1109 u64 prev_irq_time;
1110#endif
1111#ifdef CONFIG_PARAVIRT
1112 u64 prev_steal_time;
1113#endif
1114#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1115 u64 prev_steal_time_rq;
1116#endif
1117
1118 /* calc_load related fields */
1119 unsigned long calc_load_update;
1120 long calc_load_active;
1121
1122#ifdef CONFIG_SCHED_HRTICK
1123#ifdef CONFIG_SMP
1124 call_single_data_t hrtick_csd;
1125#endif
1126 struct hrtimer hrtick_timer;
1127 ktime_t hrtick_time;
1128#endif
1129
1130#ifdef CONFIG_SCHEDSTATS
1131 /* latency stats */
1132 struct sched_info rq_sched_info;
1133 unsigned long long rq_cpu_time;
1134 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
1135
1136 /* sys_sched_yield() stats */
1137 unsigned int yld_count;
1138
1139 /* schedule() stats */
1140 unsigned int sched_count;
1141 unsigned int sched_goidle;
1142
1143 /* try_to_wake_up() stats */
1144 unsigned int ttwu_count;
1145 unsigned int ttwu_local;
1146#endif
1147
1148#ifdef CONFIG_CPU_IDLE
1149 /* Must be inspected within a rcu lock section */
1150 struct cpuidle_state *idle_state;
1151#endif
1152
1153#ifdef CONFIG_SMP
1154 unsigned int nr_pinned;
1155#endif
1156 unsigned int push_busy;
1157 struct cpu_stop_work push_work;
1158
1159#ifdef CONFIG_SCHED_CORE
1160 /* per rq */
1161 struct rq *core;
1162 struct task_struct *core_pick;
1163 unsigned int core_enabled;
1164 unsigned int core_sched_seq;
1165 struct rb_root core_tree;
1166
1167 /* shared state -- careful with sched_core_cpu_deactivate() */
1168 unsigned int core_task_seq;
1169 unsigned int core_pick_seq;
1170 unsigned long core_cookie;
1171 unsigned int core_forceidle_count;
1172 unsigned int core_forceidle_seq;
1173 unsigned int core_forceidle_occupation;
1174 u64 core_forceidle_start;
1175#endif
1176
1177 /* Scratch cpumask to be temporarily used under rq_lock */
1178 cpumask_var_t scratch_mask;
1179
1180#if defined(CONFIG_CFS_BANDWIDTH) && defined(CONFIG_SMP)
1181 call_single_data_t cfsb_csd;
1182 struct list_head cfsb_csd_list;
1183#endif
1184};
1185
1186#ifdef CONFIG_FAIR_GROUP_SCHED
1187
1188/* CPU runqueue to which this cfs_rq is attached */
1189static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1190{
1191 return cfs_rq->rq;
1192}
1193
1194#else
1195
1196static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1197{
1198 return container_of(cfs_rq, struct rq, cfs);
1199}
1200#endif
1201
1202static inline int cpu_of(struct rq *rq)
1203{
1204#ifdef CONFIG_SMP
1205 return rq->cpu;
1206#else
1207 return 0;
1208#endif
1209}
1210
1211#define MDF_PUSH 0x01
1212
1213static inline bool is_migration_disabled(struct task_struct *p)
1214{
1215#ifdef CONFIG_SMP
1216 return p->migration_disabled;
1217#else
1218 return false;
1219#endif
1220}
1221
1222DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1223
1224#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
1225#define this_rq() this_cpu_ptr(&runqueues)
1226#define task_rq(p) cpu_rq(task_cpu(p))
1227#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
1228#define raw_rq() raw_cpu_ptr(&runqueues)
1229
1230struct sched_group;
1231#ifdef CONFIG_SCHED_CORE
1232static inline struct cpumask *sched_group_span(struct sched_group *sg);
1233
1234DECLARE_STATIC_KEY_FALSE(__sched_core_enabled);
1235
1236static inline bool sched_core_enabled(struct rq *rq)
1237{
1238 return static_branch_unlikely(&__sched_core_enabled) && rq->core_enabled;
1239}
1240
1241static inline bool sched_core_disabled(void)
1242{
1243 return !static_branch_unlikely(&__sched_core_enabled);
1244}
1245
1246/*
1247 * Be careful with this function; not for general use. The return value isn't
1248 * stable unless you actually hold a relevant rq->__lock.
1249 */
1250static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1251{
1252 if (sched_core_enabled(rq))
1253 return &rq->core->__lock;
1254
1255 return &rq->__lock;
1256}
1257
1258static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1259{
1260 if (rq->core_enabled)
1261 return &rq->core->__lock;
1262
1263 return &rq->__lock;
1264}
1265
1266bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
1267 bool fi);
1268void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
1269
1270/*
1271 * Helpers to check if the CPU's core cookie matches with the task's cookie
1272 * when core scheduling is enabled.
1273 * A special case is that the task's cookie always matches with CPU's core
1274 * cookie if the CPU is in an idle core.
1275 */
1276static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1277{
1278 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1279 if (!sched_core_enabled(rq))
1280 return true;
1281
1282 return rq->core->core_cookie == p->core_cookie;
1283}
1284
1285static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1286{
1287 bool idle_core = true;
1288 int cpu;
1289
1290 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1291 if (!sched_core_enabled(rq))
1292 return true;
1293
1294 for_each_cpu(cpu, cpu_smt_mask(cpu_of(rq))) {
1295 if (!available_idle_cpu(cpu)) {
1296 idle_core = false;
1297 break;
1298 }
1299 }
1300
1301 /*
1302 * A CPU in an idle core is always the best choice for tasks with
1303 * cookies.
1304 */
1305 return idle_core || rq->core->core_cookie == p->core_cookie;
1306}
1307
1308static inline bool sched_group_cookie_match(struct rq *rq,
1309 struct task_struct *p,
1310 struct sched_group *group)
1311{
1312 int cpu;
1313
1314 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1315 if (!sched_core_enabled(rq))
1316 return true;
1317
1318 for_each_cpu_and(cpu, sched_group_span(group), p->cpus_ptr) {
1319 if (sched_core_cookie_match(cpu_rq(cpu), p))
1320 return true;
1321 }
1322 return false;
1323}
1324
1325static inline bool sched_core_enqueued(struct task_struct *p)
1326{
1327 return !RB_EMPTY_NODE(&p->core_node);
1328}
1329
1330extern void sched_core_enqueue(struct rq *rq, struct task_struct *p);
1331extern void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags);
1332
1333extern void sched_core_get(void);
1334extern void sched_core_put(void);
1335
1336#else /* !CONFIG_SCHED_CORE */
1337
1338static inline bool sched_core_enabled(struct rq *rq)
1339{
1340 return false;
1341}
1342
1343static inline bool sched_core_disabled(void)
1344{
1345 return true;
1346}
1347
1348static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1349{
1350 return &rq->__lock;
1351}
1352
1353static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1354{
1355 return &rq->__lock;
1356}
1357
1358static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1359{
1360 return true;
1361}
1362
1363static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1364{
1365 return true;
1366}
1367
1368static inline bool sched_group_cookie_match(struct rq *rq,
1369 struct task_struct *p,
1370 struct sched_group *group)
1371{
1372 return true;
1373}
1374#endif /* CONFIG_SCHED_CORE */
1375
1376static inline void lockdep_assert_rq_held(struct rq *rq)
1377{
1378 lockdep_assert_held(__rq_lockp(rq));
1379}
1380
1381extern void raw_spin_rq_lock_nested(struct rq *rq, int subclass);
1382extern bool raw_spin_rq_trylock(struct rq *rq);
1383extern void raw_spin_rq_unlock(struct rq *rq);
1384
1385static inline void raw_spin_rq_lock(struct rq *rq)
1386{
1387 raw_spin_rq_lock_nested(rq, 0);
1388}
1389
1390static inline void raw_spin_rq_lock_irq(struct rq *rq)
1391{
1392 local_irq_disable();
1393 raw_spin_rq_lock(rq);
1394}
1395
1396static inline void raw_spin_rq_unlock_irq(struct rq *rq)
1397{
1398 raw_spin_rq_unlock(rq);
1399 local_irq_enable();
1400}
1401
1402static inline unsigned long _raw_spin_rq_lock_irqsave(struct rq *rq)
1403{
1404 unsigned long flags;
1405 local_irq_save(flags);
1406 raw_spin_rq_lock(rq);
1407 return flags;
1408}
1409
1410static inline void raw_spin_rq_unlock_irqrestore(struct rq *rq, unsigned long flags)
1411{
1412 raw_spin_rq_unlock(rq);
1413 local_irq_restore(flags);
1414}
1415
1416#define raw_spin_rq_lock_irqsave(rq, flags) \
1417do { \
1418 flags = _raw_spin_rq_lock_irqsave(rq); \
1419} while (0)
1420
1421#ifdef CONFIG_SCHED_SMT
1422extern void __update_idle_core(struct rq *rq);
1423
1424static inline void update_idle_core(struct rq *rq)
1425{
1426 if (static_branch_unlikely(&sched_smt_present))
1427 __update_idle_core(rq);
1428}
1429
1430#else
1431static inline void update_idle_core(struct rq *rq) { }
1432#endif
1433
1434#ifdef CONFIG_FAIR_GROUP_SCHED
1435static inline struct task_struct *task_of(struct sched_entity *se)
1436{
1437 SCHED_WARN_ON(!entity_is_task(se));
1438 return container_of(se, struct task_struct, se);
1439}
1440
1441static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
1442{
1443 return p->se.cfs_rq;
1444}
1445
1446/* runqueue on which this entity is (to be) queued */
1447static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1448{
1449 return se->cfs_rq;
1450}
1451
1452/* runqueue "owned" by this group */
1453static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1454{
1455 return grp->my_q;
1456}
1457
1458#else
1459
1460#define task_of(_se) container_of(_se, struct task_struct, se)
1461
1462static inline struct cfs_rq *task_cfs_rq(const struct task_struct *p)
1463{
1464 return &task_rq(p)->cfs;
1465}
1466
1467static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1468{
1469 const struct task_struct *p = task_of(se);
1470 struct rq *rq = task_rq(p);
1471
1472 return &rq->cfs;
1473}
1474
1475/* runqueue "owned" by this group */
1476static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1477{
1478 return NULL;
1479}
1480#endif
1481
1482extern void update_rq_clock(struct rq *rq);
1483
1484/*
1485 * rq::clock_update_flags bits
1486 *
1487 * %RQCF_REQ_SKIP - will request skipping of clock update on the next
1488 * call to __schedule(). This is an optimisation to avoid
1489 * neighbouring rq clock updates.
1490 *
1491 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is
1492 * in effect and calls to update_rq_clock() are being ignored.
1493 *
1494 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been
1495 * made to update_rq_clock() since the last time rq::lock was pinned.
1496 *
1497 * If inside of __schedule(), clock_update_flags will have been
1498 * shifted left (a left shift is a cheap operation for the fast path
1499 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use,
1500 *
1501 * if (rq-clock_update_flags >= RQCF_UPDATED)
1502 *
1503 * to check if %RQCF_UPDATED is set. It'll never be shifted more than
1504 * one position though, because the next rq_unpin_lock() will shift it
1505 * back.
1506 */
1507#define RQCF_REQ_SKIP 0x01
1508#define RQCF_ACT_SKIP 0x02
1509#define RQCF_UPDATED 0x04
1510
1511static inline void assert_clock_updated(struct rq *rq)
1512{
1513 /*
1514 * The only reason for not seeing a clock update since the
1515 * last rq_pin_lock() is if we're currently skipping updates.
1516 */
1517 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
1518}
1519
1520static inline u64 rq_clock(struct rq *rq)
1521{
1522 lockdep_assert_rq_held(rq);
1523 assert_clock_updated(rq);
1524
1525 return rq->clock;
1526}
1527
1528static inline u64 rq_clock_task(struct rq *rq)
1529{
1530 lockdep_assert_rq_held(rq);
1531 assert_clock_updated(rq);
1532
1533 return rq->clock_task;
1534}
1535
1536/**
1537 * By default the decay is the default pelt decay period.
1538 * The decay shift can change the decay period in
1539 * multiples of 32.
1540 * Decay shift Decay period(ms)
1541 * 0 32
1542 * 1 64
1543 * 2 128
1544 * 3 256
1545 * 4 512
1546 */
1547extern int sched_thermal_decay_shift;
1548
1549static inline u64 rq_clock_thermal(struct rq *rq)
1550{
1551 return rq_clock_task(rq) >> sched_thermal_decay_shift;
1552}
1553
1554static inline void rq_clock_skip_update(struct rq *rq)
1555{
1556 lockdep_assert_rq_held(rq);
1557 rq->clock_update_flags |= RQCF_REQ_SKIP;
1558}
1559
1560/*
1561 * See rt task throttling, which is the only time a skip
1562 * request is canceled.
1563 */
1564static inline void rq_clock_cancel_skipupdate(struct rq *rq)
1565{
1566 lockdep_assert_rq_held(rq);
1567 rq->clock_update_flags &= ~RQCF_REQ_SKIP;
1568}
1569
1570/*
1571 * During cpu offlining and rq wide unthrottling, we can trigger
1572 * an update_rq_clock() for several cfs and rt runqueues (Typically
1573 * when using list_for_each_entry_*)
1574 * rq_clock_start_loop_update() can be called after updating the clock
1575 * once and before iterating over the list to prevent multiple update.
1576 * After the iterative traversal, we need to call rq_clock_stop_loop_update()
1577 * to clear RQCF_ACT_SKIP of rq->clock_update_flags.
1578 */
1579static inline void rq_clock_start_loop_update(struct rq *rq)
1580{
1581 lockdep_assert_rq_held(rq);
1582 SCHED_WARN_ON(rq->clock_update_flags & RQCF_ACT_SKIP);
1583 rq->clock_update_flags |= RQCF_ACT_SKIP;
1584}
1585
1586static inline void rq_clock_stop_loop_update(struct rq *rq)
1587{
1588 lockdep_assert_rq_held(rq);
1589 rq->clock_update_flags &= ~RQCF_ACT_SKIP;
1590}
1591
1592struct rq_flags {
1593 unsigned long flags;
1594 struct pin_cookie cookie;
1595#ifdef CONFIG_SCHED_DEBUG
1596 /*
1597 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the
1598 * current pin context is stashed here in case it needs to be
1599 * restored in rq_repin_lock().
1600 */
1601 unsigned int clock_update_flags;
1602#endif
1603};
1604
1605extern struct balance_callback balance_push_callback;
1606
1607/*
1608 * Lockdep annotation that avoids accidental unlocks; it's like a
1609 * sticky/continuous lockdep_assert_held().
1610 *
1611 * This avoids code that has access to 'struct rq *rq' (basically everything in
1612 * the scheduler) from accidentally unlocking the rq if they do not also have a
1613 * copy of the (on-stack) 'struct rq_flags rf'.
1614 *
1615 * Also see Documentation/locking/lockdep-design.rst.
1616 */
1617static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
1618{
1619 rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
1620
1621#ifdef CONFIG_SCHED_DEBUG
1622 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
1623 rf->clock_update_flags = 0;
1624#ifdef CONFIG_SMP
1625 SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
1626#endif
1627#endif
1628}
1629
1630static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
1631{
1632#ifdef CONFIG_SCHED_DEBUG
1633 if (rq->clock_update_flags > RQCF_ACT_SKIP)
1634 rf->clock_update_flags = RQCF_UPDATED;
1635#endif
1636
1637 lockdep_unpin_lock(__rq_lockp(rq), rf->cookie);
1638}
1639
1640static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf)
1641{
1642 lockdep_repin_lock(__rq_lockp(rq), rf->cookie);
1643
1644#ifdef CONFIG_SCHED_DEBUG
1645 /*
1646 * Restore the value we stashed in @rf for this pin context.
1647 */
1648 rq->clock_update_flags |= rf->clock_update_flags;
1649#endif
1650}
1651
1652struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1653 __acquires(rq->lock);
1654
1655struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1656 __acquires(p->pi_lock)
1657 __acquires(rq->lock);
1658
1659static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf)
1660 __releases(rq->lock)
1661{
1662 rq_unpin_lock(rq, rf);
1663 raw_spin_rq_unlock(rq);
1664}
1665
1666static inline void
1667task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1668 __releases(rq->lock)
1669 __releases(p->pi_lock)
1670{
1671 rq_unpin_lock(rq, rf);
1672 raw_spin_rq_unlock(rq);
1673 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
1674}
1675
1676DEFINE_LOCK_GUARD_1(task_rq_lock, struct task_struct,
1677 _T->rq = task_rq_lock(_T->lock, &_T->rf),
1678 task_rq_unlock(_T->rq, _T->lock, &_T->rf),
1679 struct rq *rq; struct rq_flags rf)
1680
1681static inline void
1682rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
1683 __acquires(rq->lock)
1684{
1685 raw_spin_rq_lock_irqsave(rq, rf->flags);
1686 rq_pin_lock(rq, rf);
1687}
1688
1689static inline void
1690rq_lock_irq(struct rq *rq, struct rq_flags *rf)
1691 __acquires(rq->lock)
1692{
1693 raw_spin_rq_lock_irq(rq);
1694 rq_pin_lock(rq, rf);
1695}
1696
1697static inline void
1698rq_lock(struct rq *rq, struct rq_flags *rf)
1699 __acquires(rq->lock)
1700{
1701 raw_spin_rq_lock(rq);
1702 rq_pin_lock(rq, rf);
1703}
1704
1705static inline void
1706rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
1707 __releases(rq->lock)
1708{
1709 rq_unpin_lock(rq, rf);
1710 raw_spin_rq_unlock_irqrestore(rq, rf->flags);
1711}
1712
1713static inline void
1714rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
1715 __releases(rq->lock)
1716{
1717 rq_unpin_lock(rq, rf);
1718 raw_spin_rq_unlock_irq(rq);
1719}
1720
1721static inline void
1722rq_unlock(struct rq *rq, struct rq_flags *rf)
1723 __releases(rq->lock)
1724{
1725 rq_unpin_lock(rq, rf);
1726 raw_spin_rq_unlock(rq);
1727}
1728
1729DEFINE_LOCK_GUARD_1(rq_lock, struct rq,
1730 rq_lock(_T->lock, &_T->rf),
1731 rq_unlock(_T->lock, &_T->rf),
1732 struct rq_flags rf)
1733
1734DEFINE_LOCK_GUARD_1(rq_lock_irq, struct rq,
1735 rq_lock_irq(_T->lock, &_T->rf),
1736 rq_unlock_irq(_T->lock, &_T->rf),
1737 struct rq_flags rf)
1738
1739DEFINE_LOCK_GUARD_1(rq_lock_irqsave, struct rq,
1740 rq_lock_irqsave(_T->lock, &_T->rf),
1741 rq_unlock_irqrestore(_T->lock, &_T->rf),
1742 struct rq_flags rf)
1743
1744static inline struct rq *
1745this_rq_lock_irq(struct rq_flags *rf)
1746 __acquires(rq->lock)
1747{
1748 struct rq *rq;
1749
1750 local_irq_disable();
1751 rq = this_rq();
1752 rq_lock(rq, rf);
1753 return rq;
1754}
1755
1756#ifdef CONFIG_NUMA
1757enum numa_topology_type {
1758 NUMA_DIRECT,
1759 NUMA_GLUELESS_MESH,
1760 NUMA_BACKPLANE,
1761};
1762extern enum numa_topology_type sched_numa_topology_type;
1763extern int sched_max_numa_distance;
1764extern bool find_numa_distance(int distance);
1765extern void sched_init_numa(int offline_node);
1766extern void sched_update_numa(int cpu, bool online);
1767extern void sched_domains_numa_masks_set(unsigned int cpu);
1768extern void sched_domains_numa_masks_clear(unsigned int cpu);
1769extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu);
1770#else
1771static inline void sched_init_numa(int offline_node) { }
1772static inline void sched_update_numa(int cpu, bool online) { }
1773static inline void sched_domains_numa_masks_set(unsigned int cpu) { }
1774static inline void sched_domains_numa_masks_clear(unsigned int cpu) { }
1775static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1776{
1777 return nr_cpu_ids;
1778}
1779#endif
1780
1781#ifdef CONFIG_NUMA_BALANCING
1782/* The regions in numa_faults array from task_struct */
1783enum numa_faults_stats {
1784 NUMA_MEM = 0,
1785 NUMA_CPU,
1786 NUMA_MEMBUF,
1787 NUMA_CPUBUF
1788};
1789extern void sched_setnuma(struct task_struct *p, int node);
1790extern int migrate_task_to(struct task_struct *p, int cpu);
1791extern int migrate_swap(struct task_struct *p, struct task_struct *t,
1792 int cpu, int scpu);
1793extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p);
1794#else
1795static inline void
1796init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1797{
1798}
1799#endif /* CONFIG_NUMA_BALANCING */
1800
1801#ifdef CONFIG_SMP
1802
1803static inline void
1804queue_balance_callback(struct rq *rq,
1805 struct balance_callback *head,
1806 void (*func)(struct rq *rq))
1807{
1808 lockdep_assert_rq_held(rq);
1809
1810 /*
1811 * Don't (re)queue an already queued item; nor queue anything when
1812 * balance_push() is active, see the comment with
1813 * balance_push_callback.
1814 */
1815 if (unlikely(head->next || rq->balance_callback == &balance_push_callback))
1816 return;
1817
1818 head->func = func;
1819 head->next = rq->balance_callback;
1820 rq->balance_callback = head;
1821}
1822
1823#define rcu_dereference_check_sched_domain(p) \
1824 rcu_dereference_check((p), \
1825 lockdep_is_held(&sched_domains_mutex))
1826
1827/*
1828 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1829 * See destroy_sched_domains: call_rcu for details.
1830 *
1831 * The domain tree of any CPU may only be accessed from within
1832 * preempt-disabled sections.
1833 */
1834#define for_each_domain(cpu, __sd) \
1835 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
1836 __sd; __sd = __sd->parent)
1837
1838/* A mask of all the SD flags that have the SDF_SHARED_CHILD metaflag */
1839#define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_SHARED_CHILD)) |
1840static const unsigned int SD_SHARED_CHILD_MASK =
1841#include <linux/sched/sd_flags.h>
18420;
1843#undef SD_FLAG
1844
1845/**
1846 * highest_flag_domain - Return highest sched_domain containing flag.
1847 * @cpu: The CPU whose highest level of sched domain is to
1848 * be returned.
1849 * @flag: The flag to check for the highest sched_domain
1850 * for the given CPU.
1851 *
1852 * Returns the highest sched_domain of a CPU which contains @flag. If @flag has
1853 * the SDF_SHARED_CHILD metaflag, all the children domains also have @flag.
1854 */
1855static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
1856{
1857 struct sched_domain *sd, *hsd = NULL;
1858
1859 for_each_domain(cpu, sd) {
1860 if (sd->flags & flag) {
1861 hsd = sd;
1862 continue;
1863 }
1864
1865 /*
1866 * Stop the search if @flag is known to be shared at lower
1867 * levels. It will not be found further up.
1868 */
1869 if (flag & SD_SHARED_CHILD_MASK)
1870 break;
1871 }
1872
1873 return hsd;
1874}
1875
1876static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
1877{
1878 struct sched_domain *sd;
1879
1880 for_each_domain(cpu, sd) {
1881 if (sd->flags & flag)
1882 break;
1883 }
1884
1885 return sd;
1886}
1887
1888DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
1889DECLARE_PER_CPU(int, sd_llc_size);
1890DECLARE_PER_CPU(int, sd_llc_id);
1891DECLARE_PER_CPU(int, sd_share_id);
1892DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
1893DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
1894DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
1895DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
1896extern struct static_key_false sched_asym_cpucapacity;
1897extern struct static_key_false sched_cluster_active;
1898
1899static __always_inline bool sched_asym_cpucap_active(void)
1900{
1901 return static_branch_unlikely(&sched_asym_cpucapacity);
1902}
1903
1904struct sched_group_capacity {
1905 atomic_t ref;
1906 /*
1907 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity
1908 * for a single CPU.
1909 */
1910 unsigned long capacity;
1911 unsigned long min_capacity; /* Min per-CPU capacity in group */
1912 unsigned long max_capacity; /* Max per-CPU capacity in group */
1913 unsigned long next_update;
1914 int imbalance; /* XXX unrelated to capacity but shared group state */
1915
1916#ifdef CONFIG_SCHED_DEBUG
1917 int id;
1918#endif
1919
1920 unsigned long cpumask[]; /* Balance mask */
1921};
1922
1923struct sched_group {
1924 struct sched_group *next; /* Must be a circular list */
1925 atomic_t ref;
1926
1927 unsigned int group_weight;
1928 unsigned int cores;
1929 struct sched_group_capacity *sgc;
1930 int asym_prefer_cpu; /* CPU of highest priority in group */
1931 int flags;
1932
1933 /*
1934 * The CPUs this group covers.
1935 *
1936 * NOTE: this field is variable length. (Allocated dynamically
1937 * by attaching extra space to the end of the structure,
1938 * depending on how many CPUs the kernel has booted up with)
1939 */
1940 unsigned long cpumask[];
1941};
1942
1943static inline struct cpumask *sched_group_span(struct sched_group *sg)
1944{
1945 return to_cpumask(sg->cpumask);
1946}
1947
1948/*
1949 * See build_balance_mask().
1950 */
1951static inline struct cpumask *group_balance_mask(struct sched_group *sg)
1952{
1953 return to_cpumask(sg->sgc->cpumask);
1954}
1955
1956extern int group_balance_cpu(struct sched_group *sg);
1957
1958#ifdef CONFIG_SCHED_DEBUG
1959void update_sched_domain_debugfs(void);
1960void dirty_sched_domain_sysctl(int cpu);
1961#else
1962static inline void update_sched_domain_debugfs(void)
1963{
1964}
1965static inline void dirty_sched_domain_sysctl(int cpu)
1966{
1967}
1968#endif
1969
1970extern int sched_update_scaling(void);
1971
1972static inline const struct cpumask *task_user_cpus(struct task_struct *p)
1973{
1974 if (!p->user_cpus_ptr)
1975 return cpu_possible_mask; /* &init_task.cpus_mask */
1976 return p->user_cpus_ptr;
1977}
1978#endif /* CONFIG_SMP */
1979
1980#include "stats.h"
1981
1982#if defined(CONFIG_SCHED_CORE) && defined(CONFIG_SCHEDSTATS)
1983
1984extern void __sched_core_account_forceidle(struct rq *rq);
1985
1986static inline void sched_core_account_forceidle(struct rq *rq)
1987{
1988 if (schedstat_enabled())
1989 __sched_core_account_forceidle(rq);
1990}
1991
1992extern void __sched_core_tick(struct rq *rq);
1993
1994static inline void sched_core_tick(struct rq *rq)
1995{
1996 if (sched_core_enabled(rq) && schedstat_enabled())
1997 __sched_core_tick(rq);
1998}
1999
2000#else
2001
2002static inline void sched_core_account_forceidle(struct rq *rq) {}
2003
2004static inline void sched_core_tick(struct rq *rq) {}
2005
2006#endif /* CONFIG_SCHED_CORE && CONFIG_SCHEDSTATS */
2007
2008#ifdef CONFIG_CGROUP_SCHED
2009
2010/*
2011 * Return the group to which this tasks belongs.
2012 *
2013 * We cannot use task_css() and friends because the cgroup subsystem
2014 * changes that value before the cgroup_subsys::attach() method is called,
2015 * therefore we cannot pin it and might observe the wrong value.
2016 *
2017 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
2018 * core changes this before calling sched_move_task().
2019 *
2020 * Instead we use a 'copy' which is updated from sched_move_task() while
2021 * holding both task_struct::pi_lock and rq::lock.
2022 */
2023static inline struct task_group *task_group(struct task_struct *p)
2024{
2025 return p->sched_task_group;
2026}
2027
2028/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
2029static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
2030{
2031#if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
2032 struct task_group *tg = task_group(p);
2033#endif
2034
2035#ifdef CONFIG_FAIR_GROUP_SCHED
2036 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
2037 p->se.cfs_rq = tg->cfs_rq[cpu];
2038 p->se.parent = tg->se[cpu];
2039 p->se.depth = tg->se[cpu] ? tg->se[cpu]->depth + 1 : 0;
2040#endif
2041
2042#ifdef CONFIG_RT_GROUP_SCHED
2043 p->rt.rt_rq = tg->rt_rq[cpu];
2044 p->rt.parent = tg->rt_se[cpu];
2045#endif
2046}
2047
2048#else /* CONFIG_CGROUP_SCHED */
2049
2050static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
2051static inline struct task_group *task_group(struct task_struct *p)
2052{
2053 return NULL;
2054}
2055
2056#endif /* CONFIG_CGROUP_SCHED */
2057
2058static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
2059{
2060 set_task_rq(p, cpu);
2061#ifdef CONFIG_SMP
2062 /*
2063 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2064 * successfully executed on another CPU. We must ensure that updates of
2065 * per-task data have been completed by this moment.
2066 */
2067 smp_wmb();
2068 WRITE_ONCE(task_thread_info(p)->cpu, cpu);
2069 p->wake_cpu = cpu;
2070#endif
2071}
2072
2073/*
2074 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
2075 */
2076#ifdef CONFIG_SCHED_DEBUG
2077# define const_debug __read_mostly
2078#else
2079# define const_debug const
2080#endif
2081
2082#define SCHED_FEAT(name, enabled) \
2083 __SCHED_FEAT_##name ,
2084
2085enum {
2086#include "features.h"
2087 __SCHED_FEAT_NR,
2088};
2089
2090#undef SCHED_FEAT
2091
2092#ifdef CONFIG_SCHED_DEBUG
2093
2094/*
2095 * To support run-time toggling of sched features, all the translation units
2096 * (but core.c) reference the sysctl_sched_features defined in core.c.
2097 */
2098extern const_debug unsigned int sysctl_sched_features;
2099
2100#ifdef CONFIG_JUMP_LABEL
2101#define SCHED_FEAT(name, enabled) \
2102static __always_inline bool static_branch_##name(struct static_key *key) \
2103{ \
2104 return static_key_##enabled(key); \
2105}
2106
2107#include "features.h"
2108#undef SCHED_FEAT
2109
2110extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
2111#define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
2112
2113#else /* !CONFIG_JUMP_LABEL */
2114
2115#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2116
2117#endif /* CONFIG_JUMP_LABEL */
2118
2119#else /* !SCHED_DEBUG */
2120
2121/*
2122 * Each translation unit has its own copy of sysctl_sched_features to allow
2123 * constants propagation at compile time and compiler optimization based on
2124 * features default.
2125 */
2126#define SCHED_FEAT(name, enabled) \
2127 (1UL << __SCHED_FEAT_##name) * enabled |
2128static const_debug __maybe_unused unsigned int sysctl_sched_features =
2129#include "features.h"
2130 0;
2131#undef SCHED_FEAT
2132
2133#define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2134
2135#endif /* SCHED_DEBUG */
2136
2137extern struct static_key_false sched_numa_balancing;
2138extern struct static_key_false sched_schedstats;
2139
2140static inline u64 global_rt_period(void)
2141{
2142 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
2143}
2144
2145static inline u64 global_rt_runtime(void)
2146{
2147 if (sysctl_sched_rt_runtime < 0)
2148 return RUNTIME_INF;
2149
2150 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
2151}
2152
2153static inline int task_current(struct rq *rq, struct task_struct *p)
2154{
2155 return rq->curr == p;
2156}
2157
2158static inline int task_on_cpu(struct rq *rq, struct task_struct *p)
2159{
2160#ifdef CONFIG_SMP
2161 return p->on_cpu;
2162#else
2163 return task_current(rq, p);
2164#endif
2165}
2166
2167static inline int task_on_rq_queued(struct task_struct *p)
2168{
2169 return p->on_rq == TASK_ON_RQ_QUEUED;
2170}
2171
2172static inline int task_on_rq_migrating(struct task_struct *p)
2173{
2174 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING;
2175}
2176
2177/* Wake flags. The first three directly map to some SD flag value */
2178#define WF_EXEC 0x02 /* Wakeup after exec; maps to SD_BALANCE_EXEC */
2179#define WF_FORK 0x04 /* Wakeup after fork; maps to SD_BALANCE_FORK */
2180#define WF_TTWU 0x08 /* Wakeup; maps to SD_BALANCE_WAKE */
2181
2182#define WF_SYNC 0x10 /* Waker goes to sleep after wakeup */
2183#define WF_MIGRATED 0x20 /* Internal use, task got migrated */
2184#define WF_CURRENT_CPU 0x40 /* Prefer to move the wakee to the current CPU. */
2185
2186#ifdef CONFIG_SMP
2187static_assert(WF_EXEC == SD_BALANCE_EXEC);
2188static_assert(WF_FORK == SD_BALANCE_FORK);
2189static_assert(WF_TTWU == SD_BALANCE_WAKE);
2190#endif
2191
2192/*
2193 * To aid in avoiding the subversion of "niceness" due to uneven distribution
2194 * of tasks with abnormal "nice" values across CPUs the contribution that
2195 * each task makes to its run queue's load is weighted according to its
2196 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
2197 * scaled version of the new time slice allocation that they receive on time
2198 * slice expiry etc.
2199 */
2200
2201#define WEIGHT_IDLEPRIO 3
2202#define WMULT_IDLEPRIO 1431655765
2203
2204extern const int sched_prio_to_weight[40];
2205extern const u32 sched_prio_to_wmult[40];
2206
2207/*
2208 * {de,en}queue flags:
2209 *
2210 * DEQUEUE_SLEEP - task is no longer runnable
2211 * ENQUEUE_WAKEUP - task just became runnable
2212 *
2213 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
2214 * are in a known state which allows modification. Such pairs
2215 * should preserve as much state as possible.
2216 *
2217 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
2218 * in the runqueue.
2219 *
2220 * NOCLOCK - skip the update_rq_clock() (avoids double updates)
2221 *
2222 * MIGRATION - p->on_rq == TASK_ON_RQ_MIGRATING (used for DEADLINE)
2223 *
2224 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
2225 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
2226 * ENQUEUE_MIGRATED - the task was migrated during wakeup
2227 *
2228 */
2229
2230#define DEQUEUE_SLEEP 0x01
2231#define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */
2232#define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */
2233#define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */
2234#define DEQUEUE_MIGRATING 0x100 /* Matches ENQUEUE_MIGRATING */
2235
2236#define ENQUEUE_WAKEUP 0x01
2237#define ENQUEUE_RESTORE 0x02
2238#define ENQUEUE_MOVE 0x04
2239#define ENQUEUE_NOCLOCK 0x08
2240
2241#define ENQUEUE_HEAD 0x10
2242#define ENQUEUE_REPLENISH 0x20
2243#ifdef CONFIG_SMP
2244#define ENQUEUE_MIGRATED 0x40
2245#else
2246#define ENQUEUE_MIGRATED 0x00
2247#endif
2248#define ENQUEUE_INITIAL 0x80
2249#define ENQUEUE_MIGRATING 0x100
2250
2251#define RETRY_TASK ((void *)-1UL)
2252
2253struct affinity_context {
2254 const struct cpumask *new_mask;
2255 struct cpumask *user_mask;
2256 unsigned int flags;
2257};
2258
2259extern s64 update_curr_common(struct rq *rq);
2260
2261struct sched_class {
2262
2263#ifdef CONFIG_UCLAMP_TASK
2264 int uclamp_enabled;
2265#endif
2266
2267 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
2268 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
2269 void (*yield_task) (struct rq *rq);
2270 bool (*yield_to_task)(struct rq *rq, struct task_struct *p);
2271
2272 void (*wakeup_preempt)(struct rq *rq, struct task_struct *p, int flags);
2273
2274 struct task_struct *(*pick_next_task)(struct rq *rq);
2275
2276 void (*put_prev_task)(struct rq *rq, struct task_struct *p);
2277 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first);
2278
2279#ifdef CONFIG_SMP
2280 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2281 int (*select_task_rq)(struct task_struct *p, int task_cpu, int flags);
2282
2283 struct task_struct * (*pick_task)(struct rq *rq);
2284
2285 void (*migrate_task_rq)(struct task_struct *p, int new_cpu);
2286
2287 void (*task_woken)(struct rq *this_rq, struct task_struct *task);
2288
2289 void (*set_cpus_allowed)(struct task_struct *p, struct affinity_context *ctx);
2290
2291 void (*rq_online)(struct rq *rq);
2292 void (*rq_offline)(struct rq *rq);
2293
2294 struct rq *(*find_lock_rq)(struct task_struct *p, struct rq *rq);
2295#endif
2296
2297 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued);
2298 void (*task_fork)(struct task_struct *p);
2299 void (*task_dead)(struct task_struct *p);
2300
2301 /*
2302 * The switched_from() call is allowed to drop rq->lock, therefore we
2303 * cannot assume the switched_from/switched_to pair is serialized by
2304 * rq->lock. They are however serialized by p->pi_lock.
2305 */
2306 void (*switched_from)(struct rq *this_rq, struct task_struct *task);
2307 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
2308 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
2309 int oldprio);
2310
2311 unsigned int (*get_rr_interval)(struct rq *rq,
2312 struct task_struct *task);
2313
2314 void (*update_curr)(struct rq *rq);
2315
2316#ifdef CONFIG_FAIR_GROUP_SCHED
2317 void (*task_change_group)(struct task_struct *p);
2318#endif
2319
2320#ifdef CONFIG_SCHED_CORE
2321 int (*task_is_throttled)(struct task_struct *p, int cpu);
2322#endif
2323};
2324
2325static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
2326{
2327 WARN_ON_ONCE(rq->curr != prev);
2328 prev->sched_class->put_prev_task(rq, prev);
2329}
2330
2331static inline void set_next_task(struct rq *rq, struct task_struct *next)
2332{
2333 next->sched_class->set_next_task(rq, next, false);
2334}
2335
2336
2337/*
2338 * Helper to define a sched_class instance; each one is placed in a separate
2339 * section which is ordered by the linker script:
2340 *
2341 * include/asm-generic/vmlinux.lds.h
2342 *
2343 * *CAREFUL* they are laid out in *REVERSE* order!!!
2344 *
2345 * Also enforce alignment on the instance, not the type, to guarantee layout.
2346 */
2347#define DEFINE_SCHED_CLASS(name) \
2348const struct sched_class name##_sched_class \
2349 __aligned(__alignof__(struct sched_class)) \
2350 __section("__" #name "_sched_class")
2351
2352/* Defined in include/asm-generic/vmlinux.lds.h */
2353extern struct sched_class __sched_class_highest[];
2354extern struct sched_class __sched_class_lowest[];
2355
2356#define for_class_range(class, _from, _to) \
2357 for (class = (_from); class < (_to); class++)
2358
2359#define for_each_class(class) \
2360 for_class_range(class, __sched_class_highest, __sched_class_lowest)
2361
2362#define sched_class_above(_a, _b) ((_a) < (_b))
2363
2364extern const struct sched_class stop_sched_class;
2365extern const struct sched_class dl_sched_class;
2366extern const struct sched_class rt_sched_class;
2367extern const struct sched_class fair_sched_class;
2368extern const struct sched_class idle_sched_class;
2369
2370static inline bool sched_stop_runnable(struct rq *rq)
2371{
2372 return rq->stop && task_on_rq_queued(rq->stop);
2373}
2374
2375static inline bool sched_dl_runnable(struct rq *rq)
2376{
2377 return rq->dl.dl_nr_running > 0;
2378}
2379
2380static inline bool sched_rt_runnable(struct rq *rq)
2381{
2382 return rq->rt.rt_queued > 0;
2383}
2384
2385static inline bool sched_fair_runnable(struct rq *rq)
2386{
2387 return rq->cfs.nr_running > 0;
2388}
2389
2390extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2391extern struct task_struct *pick_next_task_idle(struct rq *rq);
2392
2393#define SCA_CHECK 0x01
2394#define SCA_MIGRATE_DISABLE 0x02
2395#define SCA_MIGRATE_ENABLE 0x04
2396#define SCA_USER 0x08
2397
2398#ifdef CONFIG_SMP
2399
2400extern void update_group_capacity(struct sched_domain *sd, int cpu);
2401
2402extern void trigger_load_balance(struct rq *rq);
2403
2404extern void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx);
2405
2406static inline struct task_struct *get_push_task(struct rq *rq)
2407{
2408 struct task_struct *p = rq->curr;
2409
2410 lockdep_assert_rq_held(rq);
2411
2412 if (rq->push_busy)
2413 return NULL;
2414
2415 if (p->nr_cpus_allowed == 1)
2416 return NULL;
2417
2418 if (p->migration_disabled)
2419 return NULL;
2420
2421 rq->push_busy = true;
2422 return get_task_struct(p);
2423}
2424
2425extern int push_cpu_stop(void *arg);
2426
2427#endif
2428
2429#ifdef CONFIG_CPU_IDLE
2430static inline void idle_set_state(struct rq *rq,
2431 struct cpuidle_state *idle_state)
2432{
2433 rq->idle_state = idle_state;
2434}
2435
2436static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2437{
2438 SCHED_WARN_ON(!rcu_read_lock_held());
2439
2440 return rq->idle_state;
2441}
2442#else
2443static inline void idle_set_state(struct rq *rq,
2444 struct cpuidle_state *idle_state)
2445{
2446}
2447
2448static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2449{
2450 return NULL;
2451}
2452#endif
2453
2454extern void schedule_idle(void);
2455asmlinkage void schedule_user(void);
2456
2457extern void sysrq_sched_debug_show(void);
2458extern void sched_init_granularity(void);
2459extern void update_max_interval(void);
2460
2461extern void init_sched_dl_class(void);
2462extern void init_sched_rt_class(void);
2463extern void init_sched_fair_class(void);
2464
2465extern void reweight_task(struct task_struct *p, int prio);
2466
2467extern void resched_curr(struct rq *rq);
2468extern void resched_cpu(int cpu);
2469
2470extern struct rt_bandwidth def_rt_bandwidth;
2471extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
2472extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
2473
2474extern void init_dl_entity(struct sched_dl_entity *dl_se);
2475
2476#define BW_SHIFT 20
2477#define BW_UNIT (1 << BW_SHIFT)
2478#define RATIO_SHIFT 8
2479#define MAX_BW_BITS (64 - BW_SHIFT)
2480#define MAX_BW ((1ULL << MAX_BW_BITS) - 1)
2481unsigned long to_ratio(u64 period, u64 runtime);
2482
2483extern void init_entity_runnable_average(struct sched_entity *se);
2484extern void post_init_entity_util_avg(struct task_struct *p);
2485
2486#ifdef CONFIG_NO_HZ_FULL
2487extern bool sched_can_stop_tick(struct rq *rq);
2488extern int __init sched_tick_offload_init(void);
2489
2490/*
2491 * Tick may be needed by tasks in the runqueue depending on their policy and
2492 * requirements. If tick is needed, lets send the target an IPI to kick it out of
2493 * nohz mode if necessary.
2494 */
2495static inline void sched_update_tick_dependency(struct rq *rq)
2496{
2497 int cpu = cpu_of(rq);
2498
2499 if (!tick_nohz_full_cpu(cpu))
2500 return;
2501
2502 if (sched_can_stop_tick(rq))
2503 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
2504 else
2505 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
2506}
2507#else
2508static inline int sched_tick_offload_init(void) { return 0; }
2509static inline void sched_update_tick_dependency(struct rq *rq) { }
2510#endif
2511
2512static inline void add_nr_running(struct rq *rq, unsigned count)
2513{
2514 unsigned prev_nr = rq->nr_running;
2515
2516 rq->nr_running = prev_nr + count;
2517 if (trace_sched_update_nr_running_tp_enabled()) {
2518 call_trace_sched_update_nr_running(rq, count);
2519 }
2520
2521#ifdef CONFIG_SMP
2522 if (prev_nr < 2 && rq->nr_running >= 2) {
2523 if (!READ_ONCE(rq->rd->overload))
2524 WRITE_ONCE(rq->rd->overload, 1);
2525 }
2526#endif
2527
2528 sched_update_tick_dependency(rq);
2529}
2530
2531static inline void sub_nr_running(struct rq *rq, unsigned count)
2532{
2533 rq->nr_running -= count;
2534 if (trace_sched_update_nr_running_tp_enabled()) {
2535 call_trace_sched_update_nr_running(rq, -count);
2536 }
2537
2538 /* Check if we still need preemption */
2539 sched_update_tick_dependency(rq);
2540}
2541
2542extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
2543extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
2544
2545extern void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags);
2546
2547#ifdef CONFIG_PREEMPT_RT
2548#define SCHED_NR_MIGRATE_BREAK 8
2549#else
2550#define SCHED_NR_MIGRATE_BREAK 32
2551#endif
2552
2553extern const_debug unsigned int sysctl_sched_nr_migrate;
2554extern const_debug unsigned int sysctl_sched_migration_cost;
2555
2556extern unsigned int sysctl_sched_base_slice;
2557
2558#ifdef CONFIG_SCHED_DEBUG
2559extern int sysctl_resched_latency_warn_ms;
2560extern int sysctl_resched_latency_warn_once;
2561
2562extern unsigned int sysctl_sched_tunable_scaling;
2563
2564extern unsigned int sysctl_numa_balancing_scan_delay;
2565extern unsigned int sysctl_numa_balancing_scan_period_min;
2566extern unsigned int sysctl_numa_balancing_scan_period_max;
2567extern unsigned int sysctl_numa_balancing_scan_size;
2568extern unsigned int sysctl_numa_balancing_hot_threshold;
2569#endif
2570
2571#ifdef CONFIG_SCHED_HRTICK
2572
2573/*
2574 * Use hrtick when:
2575 * - enabled by features
2576 * - hrtimer is actually high res
2577 */
2578static inline int hrtick_enabled(struct rq *rq)
2579{
2580 if (!cpu_active(cpu_of(rq)))
2581 return 0;
2582 return hrtimer_is_hres_active(&rq->hrtick_timer);
2583}
2584
2585static inline int hrtick_enabled_fair(struct rq *rq)
2586{
2587 if (!sched_feat(HRTICK))
2588 return 0;
2589 return hrtick_enabled(rq);
2590}
2591
2592static inline int hrtick_enabled_dl(struct rq *rq)
2593{
2594 if (!sched_feat(HRTICK_DL))
2595 return 0;
2596 return hrtick_enabled(rq);
2597}
2598
2599void hrtick_start(struct rq *rq, u64 delay);
2600
2601#else
2602
2603static inline int hrtick_enabled_fair(struct rq *rq)
2604{
2605 return 0;
2606}
2607
2608static inline int hrtick_enabled_dl(struct rq *rq)
2609{
2610 return 0;
2611}
2612
2613static inline int hrtick_enabled(struct rq *rq)
2614{
2615 return 0;
2616}
2617
2618#endif /* CONFIG_SCHED_HRTICK */
2619
2620#ifndef arch_scale_freq_tick
2621static __always_inline
2622void arch_scale_freq_tick(void)
2623{
2624}
2625#endif
2626
2627#ifndef arch_scale_freq_capacity
2628/**
2629 * arch_scale_freq_capacity - get the frequency scale factor of a given CPU.
2630 * @cpu: the CPU in question.
2631 *
2632 * Return: the frequency scale factor normalized against SCHED_CAPACITY_SCALE, i.e.
2633 *
2634 * f_curr
2635 * ------ * SCHED_CAPACITY_SCALE
2636 * f_max
2637 */
2638static __always_inline
2639unsigned long arch_scale_freq_capacity(int cpu)
2640{
2641 return SCHED_CAPACITY_SCALE;
2642}
2643#endif
2644
2645#ifdef CONFIG_SCHED_DEBUG
2646/*
2647 * In double_lock_balance()/double_rq_lock(), we use raw_spin_rq_lock() to
2648 * acquire rq lock instead of rq_lock(). So at the end of these two functions
2649 * we need to call double_rq_clock_clear_update() to clear RQCF_UPDATED of
2650 * rq->clock_update_flags to avoid the WARN_DOUBLE_CLOCK warning.
2651 */
2652static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2)
2653{
2654 rq1->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2655 /* rq1 == rq2 for !CONFIG_SMP, so just clear RQCF_UPDATED once. */
2656#ifdef CONFIG_SMP
2657 rq2->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2658#endif
2659}
2660#else
2661static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2) {}
2662#endif
2663
2664#define DEFINE_LOCK_GUARD_2(name, type, _lock, _unlock, ...) \
2665__DEFINE_UNLOCK_GUARD(name, type, _unlock, type *lock2; __VA_ARGS__) \
2666static inline class_##name##_t class_##name##_constructor(type *lock, type *lock2) \
2667{ class_##name##_t _t = { .lock = lock, .lock2 = lock2 }, *_T = &_t; \
2668 _lock; return _t; }
2669
2670#ifdef CONFIG_SMP
2671
2672static inline bool rq_order_less(struct rq *rq1, struct rq *rq2)
2673{
2674#ifdef CONFIG_SCHED_CORE
2675 /*
2676 * In order to not have {0,2},{1,3} turn into into an AB-BA,
2677 * order by core-id first and cpu-id second.
2678 *
2679 * Notably:
2680 *
2681 * double_rq_lock(0,3); will take core-0, core-1 lock
2682 * double_rq_lock(1,2); will take core-1, core-0 lock
2683 *
2684 * when only cpu-id is considered.
2685 */
2686 if (rq1->core->cpu < rq2->core->cpu)
2687 return true;
2688 if (rq1->core->cpu > rq2->core->cpu)
2689 return false;
2690
2691 /*
2692 * __sched_core_flip() relies on SMT having cpu-id lock order.
2693 */
2694#endif
2695 return rq1->cpu < rq2->cpu;
2696}
2697
2698extern void double_rq_lock(struct rq *rq1, struct rq *rq2);
2699
2700#ifdef CONFIG_PREEMPTION
2701
2702/*
2703 * fair double_lock_balance: Safely acquires both rq->locks in a fair
2704 * way at the expense of forcing extra atomic operations in all
2705 * invocations. This assures that the double_lock is acquired using the
2706 * same underlying policy as the spinlock_t on this architecture, which
2707 * reduces latency compared to the unfair variant below. However, it
2708 * also adds more overhead and therefore may reduce throughput.
2709 */
2710static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2711 __releases(this_rq->lock)
2712 __acquires(busiest->lock)
2713 __acquires(this_rq->lock)
2714{
2715 raw_spin_rq_unlock(this_rq);
2716 double_rq_lock(this_rq, busiest);
2717
2718 return 1;
2719}
2720
2721#else
2722/*
2723 * Unfair double_lock_balance: Optimizes throughput at the expense of
2724 * latency by eliminating extra atomic operations when the locks are
2725 * already in proper order on entry. This favors lower CPU-ids and will
2726 * grant the double lock to lower CPUs over higher ids under contention,
2727 * regardless of entry order into the function.
2728 */
2729static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2730 __releases(this_rq->lock)
2731 __acquires(busiest->lock)
2732 __acquires(this_rq->lock)
2733{
2734 if (__rq_lockp(this_rq) == __rq_lockp(busiest) ||
2735 likely(raw_spin_rq_trylock(busiest))) {
2736 double_rq_clock_clear_update(this_rq, busiest);
2737 return 0;
2738 }
2739
2740 if (rq_order_less(this_rq, busiest)) {
2741 raw_spin_rq_lock_nested(busiest, SINGLE_DEPTH_NESTING);
2742 double_rq_clock_clear_update(this_rq, busiest);
2743 return 0;
2744 }
2745
2746 raw_spin_rq_unlock(this_rq);
2747 double_rq_lock(this_rq, busiest);
2748
2749 return 1;
2750}
2751
2752#endif /* CONFIG_PREEMPTION */
2753
2754/*
2755 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2756 */
2757static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2758{
2759 lockdep_assert_irqs_disabled();
2760
2761 return _double_lock_balance(this_rq, busiest);
2762}
2763
2764static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2765 __releases(busiest->lock)
2766{
2767 if (__rq_lockp(this_rq) != __rq_lockp(busiest))
2768 raw_spin_rq_unlock(busiest);
2769 lock_set_subclass(&__rq_lockp(this_rq)->dep_map, 0, _RET_IP_);
2770}
2771
2772static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
2773{
2774 if (l1 > l2)
2775 swap(l1, l2);
2776
2777 spin_lock(l1);
2778 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2779}
2780
2781static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
2782{
2783 if (l1 > l2)
2784 swap(l1, l2);
2785
2786 spin_lock_irq(l1);
2787 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2788}
2789
2790static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2791{
2792 if (l1 > l2)
2793 swap(l1, l2);
2794
2795 raw_spin_lock(l1);
2796 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2797}
2798
2799static inline void double_raw_unlock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2800{
2801 raw_spin_unlock(l1);
2802 raw_spin_unlock(l2);
2803}
2804
2805DEFINE_LOCK_GUARD_2(double_raw_spinlock, raw_spinlock_t,
2806 double_raw_lock(_T->lock, _T->lock2),
2807 double_raw_unlock(_T->lock, _T->lock2))
2808
2809/*
2810 * double_rq_unlock - safely unlock two runqueues
2811 *
2812 * Note this does not restore interrupts like task_rq_unlock,
2813 * you need to do so manually after calling.
2814 */
2815static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2816 __releases(rq1->lock)
2817 __releases(rq2->lock)
2818{
2819 if (__rq_lockp(rq1) != __rq_lockp(rq2))
2820 raw_spin_rq_unlock(rq2);
2821 else
2822 __release(rq2->lock);
2823 raw_spin_rq_unlock(rq1);
2824}
2825
2826extern void set_rq_online (struct rq *rq);
2827extern void set_rq_offline(struct rq *rq);
2828extern bool sched_smp_initialized;
2829
2830#else /* CONFIG_SMP */
2831
2832/*
2833 * double_rq_lock - safely lock two runqueues
2834 *
2835 * Note this does not disable interrupts like task_rq_lock,
2836 * you need to do so manually before calling.
2837 */
2838static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
2839 __acquires(rq1->lock)
2840 __acquires(rq2->lock)
2841{
2842 WARN_ON_ONCE(!irqs_disabled());
2843 WARN_ON_ONCE(rq1 != rq2);
2844 raw_spin_rq_lock(rq1);
2845 __acquire(rq2->lock); /* Fake it out ;) */
2846 double_rq_clock_clear_update(rq1, rq2);
2847}
2848
2849/*
2850 * double_rq_unlock - safely unlock two runqueues
2851 *
2852 * Note this does not restore interrupts like task_rq_unlock,
2853 * you need to do so manually after calling.
2854 */
2855static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2856 __releases(rq1->lock)
2857 __releases(rq2->lock)
2858{
2859 WARN_ON_ONCE(rq1 != rq2);
2860 raw_spin_rq_unlock(rq1);
2861 __release(rq2->lock);
2862}
2863
2864#endif
2865
2866DEFINE_LOCK_GUARD_2(double_rq_lock, struct rq,
2867 double_rq_lock(_T->lock, _T->lock2),
2868 double_rq_unlock(_T->lock, _T->lock2))
2869
2870extern struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq);
2871extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
2872extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
2873
2874#ifdef CONFIG_SCHED_DEBUG
2875extern bool sched_debug_verbose;
2876
2877extern void print_cfs_stats(struct seq_file *m, int cpu);
2878extern void print_rt_stats(struct seq_file *m, int cpu);
2879extern void print_dl_stats(struct seq_file *m, int cpu);
2880extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
2881extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2882extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
2883
2884extern void resched_latency_warn(int cpu, u64 latency);
2885#ifdef CONFIG_NUMA_BALANCING
2886extern void
2887show_numa_stats(struct task_struct *p, struct seq_file *m);
2888extern void
2889print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
2890 unsigned long tpf, unsigned long gsf, unsigned long gpf);
2891#endif /* CONFIG_NUMA_BALANCING */
2892#else
2893static inline void resched_latency_warn(int cpu, u64 latency) {}
2894#endif /* CONFIG_SCHED_DEBUG */
2895
2896extern void init_cfs_rq(struct cfs_rq *cfs_rq);
2897extern void init_rt_rq(struct rt_rq *rt_rq);
2898extern void init_dl_rq(struct dl_rq *dl_rq);
2899
2900extern void cfs_bandwidth_usage_inc(void);
2901extern void cfs_bandwidth_usage_dec(void);
2902
2903#ifdef CONFIG_NO_HZ_COMMON
2904#define NOHZ_BALANCE_KICK_BIT 0
2905#define NOHZ_STATS_KICK_BIT 1
2906#define NOHZ_NEWILB_KICK_BIT 2
2907#define NOHZ_NEXT_KICK_BIT 3
2908
2909/* Run rebalance_domains() */
2910#define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT)
2911/* Update blocked load */
2912#define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT)
2913/* Update blocked load when entering idle */
2914#define NOHZ_NEWILB_KICK BIT(NOHZ_NEWILB_KICK_BIT)
2915/* Update nohz.next_balance */
2916#define NOHZ_NEXT_KICK BIT(NOHZ_NEXT_KICK_BIT)
2917
2918#define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK | NOHZ_NEXT_KICK)
2919
2920#define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
2921
2922extern void nohz_balance_exit_idle(struct rq *rq);
2923#else
2924static inline void nohz_balance_exit_idle(struct rq *rq) { }
2925#endif
2926
2927#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
2928extern void nohz_run_idle_balance(int cpu);
2929#else
2930static inline void nohz_run_idle_balance(int cpu) { }
2931#endif
2932
2933#ifdef CONFIG_IRQ_TIME_ACCOUNTING
2934struct irqtime {
2935 u64 total;
2936 u64 tick_delta;
2937 u64 irq_start_time;
2938 struct u64_stats_sync sync;
2939};
2940
2941DECLARE_PER_CPU(struct irqtime, cpu_irqtime);
2942
2943/*
2944 * Returns the irqtime minus the softirq time computed by ksoftirqd.
2945 * Otherwise ksoftirqd's sum_exec_runtime is subtracted its own runtime
2946 * and never move forward.
2947 */
2948static inline u64 irq_time_read(int cpu)
2949{
2950 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu);
2951 unsigned int seq;
2952 u64 total;
2953
2954 do {
2955 seq = __u64_stats_fetch_begin(&irqtime->sync);
2956 total = irqtime->total;
2957 } while (__u64_stats_fetch_retry(&irqtime->sync, seq));
2958
2959 return total;
2960}
2961#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2962
2963#ifdef CONFIG_CPU_FREQ
2964DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data);
2965
2966/**
2967 * cpufreq_update_util - Take a note about CPU utilization changes.
2968 * @rq: Runqueue to carry out the update for.
2969 * @flags: Update reason flags.
2970 *
2971 * This function is called by the scheduler on the CPU whose utilization is
2972 * being updated.
2973 *
2974 * It can only be called from RCU-sched read-side critical sections.
2975 *
2976 * The way cpufreq is currently arranged requires it to evaluate the CPU
2977 * performance state (frequency/voltage) on a regular basis to prevent it from
2978 * being stuck in a completely inadequate performance level for too long.
2979 * That is not guaranteed to happen if the updates are only triggered from CFS
2980 * and DL, though, because they may not be coming in if only RT tasks are
2981 * active all the time (or there are RT tasks only).
2982 *
2983 * As a workaround for that issue, this function is called periodically by the
2984 * RT sched class to trigger extra cpufreq updates to prevent it from stalling,
2985 * but that really is a band-aid. Going forward it should be replaced with
2986 * solutions targeted more specifically at RT tasks.
2987 */
2988static inline void cpufreq_update_util(struct rq *rq, unsigned int flags)
2989{
2990 struct update_util_data *data;
2991
2992 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data,
2993 cpu_of(rq)));
2994 if (data)
2995 data->func(data, rq_clock(rq), flags);
2996}
2997#else
2998static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
2999#endif /* CONFIG_CPU_FREQ */
3000
3001#ifdef arch_scale_freq_capacity
3002# ifndef arch_scale_freq_invariant
3003# define arch_scale_freq_invariant() true
3004# endif
3005#else
3006# define arch_scale_freq_invariant() false
3007#endif
3008
3009#ifdef CONFIG_SMP
3010unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
3011 unsigned long *min,
3012 unsigned long *max);
3013
3014unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
3015 unsigned long min,
3016 unsigned long max);
3017
3018
3019/*
3020 * Verify the fitness of task @p to run on @cpu taking into account the
3021 * CPU original capacity and the runtime/deadline ratio of the task.
3022 *
3023 * The function will return true if the original capacity of @cpu is
3024 * greater than or equal to task's deadline density right shifted by
3025 * (BW_SHIFT - SCHED_CAPACITY_SHIFT) and false otherwise.
3026 */
3027static inline bool dl_task_fits_capacity(struct task_struct *p, int cpu)
3028{
3029 unsigned long cap = arch_scale_cpu_capacity(cpu);
3030
3031 return cap >= p->dl.dl_density >> (BW_SHIFT - SCHED_CAPACITY_SHIFT);
3032}
3033
3034static inline unsigned long cpu_bw_dl(struct rq *rq)
3035{
3036 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT;
3037}
3038
3039static inline unsigned long cpu_util_dl(struct rq *rq)
3040{
3041 return READ_ONCE(rq->avg_dl.util_avg);
3042}
3043
3044
3045extern unsigned long cpu_util_cfs(int cpu);
3046extern unsigned long cpu_util_cfs_boost(int cpu);
3047
3048static inline unsigned long cpu_util_rt(struct rq *rq)
3049{
3050 return READ_ONCE(rq->avg_rt.util_avg);
3051}
3052#endif
3053
3054#ifdef CONFIG_UCLAMP_TASK
3055unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id);
3056
3057static inline unsigned long uclamp_rq_get(struct rq *rq,
3058 enum uclamp_id clamp_id)
3059{
3060 return READ_ONCE(rq->uclamp[clamp_id].value);
3061}
3062
3063static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3064 unsigned int value)
3065{
3066 WRITE_ONCE(rq->uclamp[clamp_id].value, value);
3067}
3068
3069static inline bool uclamp_rq_is_idle(struct rq *rq)
3070{
3071 return rq->uclamp_flags & UCLAMP_FLAG_IDLE;
3072}
3073
3074/* Is the rq being capped/throttled by uclamp_max? */
3075static inline bool uclamp_rq_is_capped(struct rq *rq)
3076{
3077 unsigned long rq_util;
3078 unsigned long max_util;
3079
3080 if (!static_branch_likely(&sched_uclamp_used))
3081 return false;
3082
3083 rq_util = cpu_util_cfs(cpu_of(rq)) + cpu_util_rt(rq);
3084 max_util = READ_ONCE(rq->uclamp[UCLAMP_MAX].value);
3085
3086 return max_util != SCHED_CAPACITY_SCALE && rq_util >= max_util;
3087}
3088
3089/*
3090 * When uclamp is compiled in, the aggregation at rq level is 'turned off'
3091 * by default in the fast path and only gets turned on once userspace performs
3092 * an operation that requires it.
3093 *
3094 * Returns true if userspace opted-in to use uclamp and aggregation at rq level
3095 * hence is active.
3096 */
3097static inline bool uclamp_is_used(void)
3098{
3099 return static_branch_likely(&sched_uclamp_used);
3100}
3101#else /* CONFIG_UCLAMP_TASK */
3102static inline unsigned long uclamp_eff_value(struct task_struct *p,
3103 enum uclamp_id clamp_id)
3104{
3105 if (clamp_id == UCLAMP_MIN)
3106 return 0;
3107
3108 return SCHED_CAPACITY_SCALE;
3109}
3110
3111static inline bool uclamp_rq_is_capped(struct rq *rq) { return false; }
3112
3113static inline bool uclamp_is_used(void)
3114{
3115 return false;
3116}
3117
3118static inline unsigned long uclamp_rq_get(struct rq *rq,
3119 enum uclamp_id clamp_id)
3120{
3121 if (clamp_id == UCLAMP_MIN)
3122 return 0;
3123
3124 return SCHED_CAPACITY_SCALE;
3125}
3126
3127static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3128 unsigned int value)
3129{
3130}
3131
3132static inline bool uclamp_rq_is_idle(struct rq *rq)
3133{
3134 return false;
3135}
3136#endif /* CONFIG_UCLAMP_TASK */
3137
3138#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
3139static inline unsigned long cpu_util_irq(struct rq *rq)
3140{
3141 return READ_ONCE(rq->avg_irq.util_avg);
3142}
3143
3144static inline
3145unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3146{
3147 util *= (max - irq);
3148 util /= max;
3149
3150 return util;
3151
3152}
3153#else
3154static inline unsigned long cpu_util_irq(struct rq *rq)
3155{
3156 return 0;
3157}
3158
3159static inline
3160unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3161{
3162 return util;
3163}
3164#endif
3165
3166#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
3167
3168#define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus)))
3169
3170DECLARE_STATIC_KEY_FALSE(sched_energy_present);
3171
3172static inline bool sched_energy_enabled(void)
3173{
3174 return static_branch_unlikely(&sched_energy_present);
3175}
3176
3177extern struct cpufreq_governor schedutil_gov;
3178
3179#else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */
3180
3181#define perf_domain_span(pd) NULL
3182static inline bool sched_energy_enabled(void) { return false; }
3183
3184#endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
3185
3186#ifdef CONFIG_MEMBARRIER
3187/*
3188 * The scheduler provides memory barriers required by membarrier between:
3189 * - prior user-space memory accesses and store to rq->membarrier_state,
3190 * - store to rq->membarrier_state and following user-space memory accesses.
3191 * In the same way it provides those guarantees around store to rq->curr.
3192 */
3193static inline void membarrier_switch_mm(struct rq *rq,
3194 struct mm_struct *prev_mm,
3195 struct mm_struct *next_mm)
3196{
3197 int membarrier_state;
3198
3199 if (prev_mm == next_mm)
3200 return;
3201
3202 membarrier_state = atomic_read(&next_mm->membarrier_state);
3203 if (READ_ONCE(rq->membarrier_state) == membarrier_state)
3204 return;
3205
3206 WRITE_ONCE(rq->membarrier_state, membarrier_state);
3207}
3208#else
3209static inline void membarrier_switch_mm(struct rq *rq,
3210 struct mm_struct *prev_mm,
3211 struct mm_struct *next_mm)
3212{
3213}
3214#endif
3215
3216#ifdef CONFIG_SMP
3217static inline bool is_per_cpu_kthread(struct task_struct *p)
3218{
3219 if (!(p->flags & PF_KTHREAD))
3220 return false;
3221
3222 if (p->nr_cpus_allowed != 1)
3223 return false;
3224
3225 return true;
3226}
3227#endif
3228
3229extern void swake_up_all_locked(struct swait_queue_head *q);
3230extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);
3231
3232extern int try_to_wake_up(struct task_struct *tsk, unsigned int state, int wake_flags);
3233
3234#ifdef CONFIG_PREEMPT_DYNAMIC
3235extern int preempt_dynamic_mode;
3236extern int sched_dynamic_mode(const char *str);
3237extern void sched_dynamic_update(int mode);
3238#endif
3239
3240#ifdef CONFIG_SCHED_MM_CID
3241
3242#define SCHED_MM_CID_PERIOD_NS (100ULL * 1000000) /* 100ms */
3243#define MM_CID_SCAN_DELAY 100 /* 100ms */
3244
3245extern raw_spinlock_t cid_lock;
3246extern int use_cid_lock;
3247
3248extern void sched_mm_cid_migrate_from(struct task_struct *t);
3249extern void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t);
3250extern void task_tick_mm_cid(struct rq *rq, struct task_struct *curr);
3251extern void init_sched_mm_cid(struct task_struct *t);
3252
3253static inline void __mm_cid_put(struct mm_struct *mm, int cid)
3254{
3255 if (cid < 0)
3256 return;
3257 cpumask_clear_cpu(cid, mm_cidmask(mm));
3258}
3259
3260/*
3261 * The per-mm/cpu cid can have the MM_CID_LAZY_PUT flag set or transition to
3262 * the MM_CID_UNSET state without holding the rq lock, but the rq lock needs to
3263 * be held to transition to other states.
3264 *
3265 * State transitions synchronized with cmpxchg or try_cmpxchg need to be
3266 * consistent across cpus, which prevents use of this_cpu_cmpxchg.
3267 */
3268static inline void mm_cid_put_lazy(struct task_struct *t)
3269{
3270 struct mm_struct *mm = t->mm;
3271 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3272 int cid;
3273
3274 lockdep_assert_irqs_disabled();
3275 cid = __this_cpu_read(pcpu_cid->cid);
3276 if (!mm_cid_is_lazy_put(cid) ||
3277 !try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3278 return;
3279 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3280}
3281
3282static inline int mm_cid_pcpu_unset(struct mm_struct *mm)
3283{
3284 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3285 int cid, res;
3286
3287 lockdep_assert_irqs_disabled();
3288 cid = __this_cpu_read(pcpu_cid->cid);
3289 for (;;) {
3290 if (mm_cid_is_unset(cid))
3291 return MM_CID_UNSET;
3292 /*
3293 * Attempt transition from valid or lazy-put to unset.
3294 */
3295 res = cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, cid, MM_CID_UNSET);
3296 if (res == cid)
3297 break;
3298 cid = res;
3299 }
3300 return cid;
3301}
3302
3303static inline void mm_cid_put(struct mm_struct *mm)
3304{
3305 int cid;
3306
3307 lockdep_assert_irqs_disabled();
3308 cid = mm_cid_pcpu_unset(mm);
3309 if (cid == MM_CID_UNSET)
3310 return;
3311 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3312}
3313
3314static inline int __mm_cid_try_get(struct mm_struct *mm)
3315{
3316 struct cpumask *cpumask;
3317 int cid;
3318
3319 cpumask = mm_cidmask(mm);
3320 /*
3321 * Retry finding first zero bit if the mask is temporarily
3322 * filled. This only happens during concurrent remote-clear
3323 * which owns a cid without holding a rq lock.
3324 */
3325 for (;;) {
3326 cid = cpumask_first_zero(cpumask);
3327 if (cid < nr_cpu_ids)
3328 break;
3329 cpu_relax();
3330 }
3331 if (cpumask_test_and_set_cpu(cid, cpumask))
3332 return -1;
3333 return cid;
3334}
3335
3336/*
3337 * Save a snapshot of the current runqueue time of this cpu
3338 * with the per-cpu cid value, allowing to estimate how recently it was used.
3339 */
3340static inline void mm_cid_snapshot_time(struct rq *rq, struct mm_struct *mm)
3341{
3342 struct mm_cid *pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(rq));
3343
3344 lockdep_assert_rq_held(rq);
3345 WRITE_ONCE(pcpu_cid->time, rq->clock);
3346}
3347
3348static inline int __mm_cid_get(struct rq *rq, struct mm_struct *mm)
3349{
3350 int cid;
3351
3352 /*
3353 * All allocations (even those using the cid_lock) are lock-free. If
3354 * use_cid_lock is set, hold the cid_lock to perform cid allocation to
3355 * guarantee forward progress.
3356 */
3357 if (!READ_ONCE(use_cid_lock)) {
3358 cid = __mm_cid_try_get(mm);
3359 if (cid >= 0)
3360 goto end;
3361 raw_spin_lock(&cid_lock);
3362 } else {
3363 raw_spin_lock(&cid_lock);
3364 cid = __mm_cid_try_get(mm);
3365 if (cid >= 0)
3366 goto unlock;
3367 }
3368
3369 /*
3370 * cid concurrently allocated. Retry while forcing following
3371 * allocations to use the cid_lock to ensure forward progress.
3372 */
3373 WRITE_ONCE(use_cid_lock, 1);
3374 /*
3375 * Set use_cid_lock before allocation. Only care about program order
3376 * because this is only required for forward progress.
3377 */
3378 barrier();
3379 /*
3380 * Retry until it succeeds. It is guaranteed to eventually succeed once
3381 * all newcoming allocations observe the use_cid_lock flag set.
3382 */
3383 do {
3384 cid = __mm_cid_try_get(mm);
3385 cpu_relax();
3386 } while (cid < 0);
3387 /*
3388 * Allocate before clearing use_cid_lock. Only care about
3389 * program order because this is for forward progress.
3390 */
3391 barrier();
3392 WRITE_ONCE(use_cid_lock, 0);
3393unlock:
3394 raw_spin_unlock(&cid_lock);
3395end:
3396 mm_cid_snapshot_time(rq, mm);
3397 return cid;
3398}
3399
3400static inline int mm_cid_get(struct rq *rq, struct mm_struct *mm)
3401{
3402 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3403 struct cpumask *cpumask;
3404 int cid;
3405
3406 lockdep_assert_rq_held(rq);
3407 cpumask = mm_cidmask(mm);
3408 cid = __this_cpu_read(pcpu_cid->cid);
3409 if (mm_cid_is_valid(cid)) {
3410 mm_cid_snapshot_time(rq, mm);
3411 return cid;
3412 }
3413 if (mm_cid_is_lazy_put(cid)) {
3414 if (try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3415 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3416 }
3417 cid = __mm_cid_get(rq, mm);
3418 __this_cpu_write(pcpu_cid->cid, cid);
3419 return cid;
3420}
3421
3422static inline void switch_mm_cid(struct rq *rq,
3423 struct task_struct *prev,
3424 struct task_struct *next)
3425{
3426 /*
3427 * Provide a memory barrier between rq->curr store and load of
3428 * {prev,next}->mm->pcpu_cid[cpu] on rq->curr->mm transition.
3429 *
3430 * Should be adapted if context_switch() is modified.
3431 */
3432 if (!next->mm) { // to kernel
3433 /*
3434 * user -> kernel transition does not guarantee a barrier, but
3435 * we can use the fact that it performs an atomic operation in
3436 * mmgrab().
3437 */
3438 if (prev->mm) // from user
3439 smp_mb__after_mmgrab();
3440 /*
3441 * kernel -> kernel transition does not change rq->curr->mm
3442 * state. It stays NULL.
3443 */
3444 } else { // to user
3445 /*
3446 * kernel -> user transition does not provide a barrier
3447 * between rq->curr store and load of {prev,next}->mm->pcpu_cid[cpu].
3448 * Provide it here.
3449 */
3450 if (!prev->mm) { // from kernel
3451 smp_mb();
3452 } else { // from user
3453 /*
3454 * user->user transition relies on an implicit
3455 * memory barrier in switch_mm() when
3456 * current->mm changes. If the architecture
3457 * switch_mm() does not have an implicit memory
3458 * barrier, it is emitted here. If current->mm
3459 * is unchanged, no barrier is needed.
3460 */
3461 smp_mb__after_switch_mm();
3462 }
3463 }
3464 if (prev->mm_cid_active) {
3465 mm_cid_snapshot_time(rq, prev->mm);
3466 mm_cid_put_lazy(prev);
3467 prev->mm_cid = -1;
3468 }
3469 if (next->mm_cid_active)
3470 next->last_mm_cid = next->mm_cid = mm_cid_get(rq, next->mm);
3471}
3472
3473#else
3474static inline void switch_mm_cid(struct rq *rq, struct task_struct *prev, struct task_struct *next) { }
3475static inline void sched_mm_cid_migrate_from(struct task_struct *t) { }
3476static inline void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t) { }
3477static inline void task_tick_mm_cid(struct rq *rq, struct task_struct *curr) { }
3478static inline void init_sched_mm_cid(struct task_struct *t) { }
3479#endif
3480
3481extern u64 avg_vruntime(struct cfs_rq *cfs_rq);
3482extern int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se);
3483
3484#endif /* _KERNEL_SCHED_SCHED_H */