1 /* SPDX-License-Identifier: GPL-2.0 */
3 * Scheduler internal types and methods:
5 #ifndef _KERNEL_SCHED_SCHED_H
6 #define _KERNEL_SCHED_SCHED_H
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>
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>
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>
72 #include <trace/events/power.h>
73 #include <trace/events/sched.h>
75 #include "../workqueue_internal.h"
77 #ifdef CONFIG_PARAVIRT
78 # include <asm/paravirt.h>
79 # include <asm/paravirt_api_clock.h>
82 #include <asm/barrier.h>
85 #include "cpudeadline.h"
87 #ifdef CONFIG_SCHED_DEBUG
88 # define SCHED_WARN_ON(x) WARN_ONCE(x, #x)
90 # define SCHED_WARN_ON(x) ({ (void)(x), 0; })
96 /* task_struct::on_rq states: */
97 #define TASK_ON_RQ_QUEUED 1
98 #define TASK_ON_RQ_MIGRATING 2
100 extern __read_mostly int scheduler_running;
102 extern unsigned long calc_load_update;
103 extern atomic_long_t calc_load_tasks;
105 extern void calc_global_load_tick(struct rq *this_rq);
106 extern long calc_load_fold_active(struct rq *this_rq, long adjust);
108 extern void call_trace_sched_update_nr_running(struct rq *rq, int count);
110 extern int sysctl_sched_rt_period;
111 extern int sysctl_sched_rt_runtime;
112 extern int sched_rr_timeslice;
115 * Asymmetric CPU capacity bits
117 struct asym_cap_data {
118 struct list_head link;
120 unsigned long capacity;
121 unsigned long cpus[];
124 extern struct list_head asym_cap_list;
126 #define cpu_capacity_span(asym_data) to_cpumask((asym_data)->cpus)
129 * Helpers for converting nanosecond timing to jiffy resolution
131 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
134 * Increase resolution of nice-level calculations for 64-bit architectures.
135 * The extra resolution improves shares distribution and load balancing of
136 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
137 * hierarchies, especially on larger systems. This is not a user-visible change
138 * and does not change the user-interface for setting shares/weights.
140 * We increase resolution only if we have enough bits to allow this increased
141 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
142 * are pretty high and the returns do not justify the increased costs.
144 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to
145 * increase coverage and consistency always enable it on 64-bit platforms.
148 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT)
149 # define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT)
150 # define scale_load_down(w) \
152 unsigned long __w = (w); \
154 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \
158 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT)
159 # define scale_load(w) (w)
160 # define scale_load_down(w) (w)
164 * Task weight (visible to users) and its load (invisible to users) have
165 * independent resolution, but they should be well calibrated. We use
166 * scale_load() and scale_load_down(w) to convert between them. The
167 * following must be true:
169 * scale_load(sched_prio_to_weight[NICE_TO_PRIO(0)-MAX_RT_PRIO]) == NICE_0_LOAD
172 #define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT)
175 * Single value that decides SCHED_DEADLINE internal math precision.
176 * 10 -> just above 1us
177 * 9 -> just above 0.5us
182 * Single value that denotes runtime == period, ie unlimited time.
184 #define RUNTIME_INF ((u64)~0ULL)
186 static inline int idle_policy(int policy)
188 return policy == SCHED_IDLE;
190 static inline int fair_policy(int policy)
192 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
195 static inline int rt_policy(int policy)
197 return policy == SCHED_FIFO || policy == SCHED_RR;
200 static inline int dl_policy(int policy)
202 return policy == SCHED_DEADLINE;
204 static inline bool valid_policy(int policy)
206 return idle_policy(policy) || fair_policy(policy) ||
207 rt_policy(policy) || dl_policy(policy);
210 static inline int task_has_idle_policy(struct task_struct *p)
212 return idle_policy(p->policy);
215 static inline int task_has_rt_policy(struct task_struct *p)
217 return rt_policy(p->policy);
220 static inline int task_has_dl_policy(struct task_struct *p)
222 return dl_policy(p->policy);
225 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
227 static inline void update_avg(u64 *avg, u64 sample)
229 s64 diff = sample - *avg;
234 * Shifting a value by an exponent greater *or equal* to the size of said value
235 * is UB; cap at size-1.
237 #define shr_bound(val, shift) \
238 (val >> min_t(typeof(shift), shift, BITS_PER_TYPE(typeof(val)) - 1))
241 * !! For sched_setattr_nocheck() (kernel) only !!
243 * This is actually gross. :(
245 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE
246 * tasks, but still be able to sleep. We need this on platforms that cannot
247 * atomically change clock frequency. Remove once fast switching will be
248 * available on such platforms.
250 * SUGOV stands for SchedUtil GOVernor.
252 #define SCHED_FLAG_SUGOV 0x10000000
254 #define SCHED_DL_FLAGS (SCHED_FLAG_RECLAIM | SCHED_FLAG_DL_OVERRUN | SCHED_FLAG_SUGOV)
256 static inline bool dl_entity_is_special(const struct sched_dl_entity *dl_se)
258 #ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
259 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV);
266 * Tells if entity @a should preempt entity @b.
268 static inline bool dl_entity_preempt(const struct sched_dl_entity *a,
269 const struct sched_dl_entity *b)
271 return dl_entity_is_special(a) ||
272 dl_time_before(a->deadline, b->deadline);
276 * This is the priority-queue data structure of the RT scheduling class:
278 struct rt_prio_array {
279 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
280 struct list_head queue[MAX_RT_PRIO];
283 struct rt_bandwidth {
284 /* nests inside the rq lock: */
285 raw_spinlock_t rt_runtime_lock;
288 struct hrtimer rt_period_timer;
289 unsigned int rt_period_active;
292 static inline int dl_bandwidth_enabled(void)
294 return sysctl_sched_rt_runtime >= 0;
298 * To keep the bandwidth of -deadline tasks under control
299 * we need some place where:
300 * - store the maximum -deadline bandwidth of each cpu;
301 * - cache the fraction of bandwidth that is currently allocated in
304 * This is all done in the data structure below. It is similar to the
305 * one used for RT-throttling (rt_bandwidth), with the main difference
306 * that, since here we are only interested in admission control, we
307 * do not decrease any runtime while the group "executes", neither we
308 * need a timer to replenish it.
310 * With respect to SMP, bandwidth is given on a per root domain basis,
312 * - bw (< 100%) is the deadline bandwidth of each CPU;
313 * - total_bw is the currently allocated bandwidth in each root domain;
321 extern void init_dl_bw(struct dl_bw *dl_b);
322 extern int sched_dl_global_validate(void);
323 extern void sched_dl_do_global(void);
324 extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr);
325 extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr);
326 extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr);
327 extern bool __checkparam_dl(const struct sched_attr *attr);
328 extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr);
329 extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial);
330 extern int dl_bw_check_overflow(int cpu);
333 * SCHED_DEADLINE supports servers (nested scheduling) with the following
336 * dl_se::rq -- runqueue we belong to.
338 * dl_se::server_has_tasks() -- used on bandwidth enforcement; we 'stop' the
339 * server when it runs out of tasks to run.
341 * dl_se::server_pick() -- nested pick_next_task(); we yield the period if this
344 * dl_server_update() -- called from update_curr_common(), propagates runtime
348 * dl_server_stop() -- start/stop the server when it has (no) tasks.
350 * dl_server_init() -- initializes the server.
352 extern void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec);
353 extern void dl_server_start(struct sched_dl_entity *dl_se);
354 extern void dl_server_stop(struct sched_dl_entity *dl_se);
355 extern void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
356 dl_server_has_tasks_f has_tasks,
357 dl_server_pick_f pick);
359 #ifdef CONFIG_CGROUP_SCHED
364 extern struct list_head task_groups;
366 struct cfs_bandwidth {
367 #ifdef CONFIG_CFS_BANDWIDTH
374 s64 hierarchical_quota;
379 struct hrtimer period_timer;
380 struct hrtimer slack_timer;
381 struct list_head throttled_cfs_rq;
392 /* Task group related information */
394 struct cgroup_subsys_state css;
396 #ifdef CONFIG_FAIR_GROUP_SCHED
397 /* schedulable entities of this group on each CPU */
398 struct sched_entity **se;
399 /* runqueue "owned" by this group on each CPU */
400 struct cfs_rq **cfs_rq;
401 unsigned long shares;
403 /* A positive value indicates that this is a SCHED_IDLE group. */
408 * load_avg can be heavily contended at clock tick time, so put
409 * it in its own cacheline separated from the fields above which
410 * will also be accessed at each tick.
412 atomic_long_t load_avg ____cacheline_aligned;
416 #ifdef CONFIG_RT_GROUP_SCHED
417 struct sched_rt_entity **rt_se;
418 struct rt_rq **rt_rq;
420 struct rt_bandwidth rt_bandwidth;
424 struct list_head list;
426 struct task_group *parent;
427 struct list_head siblings;
428 struct list_head children;
430 #ifdef CONFIG_SCHED_AUTOGROUP
431 struct autogroup *autogroup;
434 struct cfs_bandwidth cfs_bandwidth;
436 #ifdef CONFIG_UCLAMP_TASK_GROUP
437 /* The two decimal precision [%] value requested from user-space */
438 unsigned int uclamp_pct[UCLAMP_CNT];
439 /* Clamp values requested for a task group */
440 struct uclamp_se uclamp_req[UCLAMP_CNT];
441 /* Effective clamp values used for a task group */
442 struct uclamp_se uclamp[UCLAMP_CNT];
447 #ifdef CONFIG_FAIR_GROUP_SCHED
448 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
451 * A weight of 0 or 1 can cause arithmetics problems.
452 * A weight of a cfs_rq is the sum of weights of which entities
453 * are queued on this cfs_rq, so a weight of a entity should not be
454 * too large, so as the shares value of a task group.
455 * (The default weight is 1024 - so there's no practical
456 * limitation from this.)
458 #define MIN_SHARES (1UL << 1)
459 #define MAX_SHARES (1UL << 18)
462 typedef int (*tg_visitor)(struct task_group *, void *);
464 extern int walk_tg_tree_from(struct task_group *from,
465 tg_visitor down, tg_visitor up, void *data);
468 * Iterate the full tree, calling @down when first entering a node and @up when
469 * leaving it for the final time.
471 * Caller must hold rcu_lock or sufficient equivalent.
473 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
475 return walk_tg_tree_from(&root_task_group, down, up, data);
478 extern int tg_nop(struct task_group *tg, void *data);
480 #ifdef CONFIG_FAIR_GROUP_SCHED
481 extern void free_fair_sched_group(struct task_group *tg);
482 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
483 extern void online_fair_sched_group(struct task_group *tg);
484 extern void unregister_fair_sched_group(struct task_group *tg);
486 static inline void free_fair_sched_group(struct task_group *tg) { }
487 static inline int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
491 static inline void online_fair_sched_group(struct task_group *tg) { }
492 static inline void unregister_fair_sched_group(struct task_group *tg) { }
495 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
496 struct sched_entity *se, int cpu,
497 struct sched_entity *parent);
498 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent);
500 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
501 extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
502 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
503 extern bool cfs_task_bw_constrained(struct task_struct *p);
505 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
506 struct sched_rt_entity *rt_se, int cpu,
507 struct sched_rt_entity *parent);
508 extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us);
509 extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us);
510 extern long sched_group_rt_runtime(struct task_group *tg);
511 extern long sched_group_rt_period(struct task_group *tg);
512 extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
514 extern struct task_group *sched_create_group(struct task_group *parent);
515 extern void sched_online_group(struct task_group *tg,
516 struct task_group *parent);
517 extern void sched_destroy_group(struct task_group *tg);
518 extern void sched_release_group(struct task_group *tg);
520 extern void sched_move_task(struct task_struct *tsk);
522 #ifdef CONFIG_FAIR_GROUP_SCHED
523 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
525 extern int sched_group_set_idle(struct task_group *tg, long idle);
528 extern void set_task_rq_fair(struct sched_entity *se,
529 struct cfs_rq *prev, struct cfs_rq *next);
530 #else /* !CONFIG_SMP */
531 static inline void set_task_rq_fair(struct sched_entity *se,
532 struct cfs_rq *prev, struct cfs_rq *next) { }
533 #endif /* CONFIG_SMP */
534 #endif /* CONFIG_FAIR_GROUP_SCHED */
536 #else /* CONFIG_CGROUP_SCHED */
538 struct cfs_bandwidth { };
539 static inline bool cfs_task_bw_constrained(struct task_struct *p) { return false; }
541 #endif /* CONFIG_CGROUP_SCHED */
543 extern void unregister_rt_sched_group(struct task_group *tg);
544 extern void free_rt_sched_group(struct task_group *tg);
545 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
548 * u64_u32_load/u64_u32_store
550 * Use a copy of a u64 value to protect against data race. This is only
551 * applicable for 32-bits architectures.
554 # define u64_u32_load_copy(var, copy) var
555 # define u64_u32_store_copy(var, copy, val) (var = val)
557 # define u64_u32_load_copy(var, copy) \
559 u64 __val, __val_copy; \
563 * paired with u64_u32_store_copy(), ordering access \
568 } while (__val != __val_copy); \
571 # define u64_u32_store_copy(var, copy, val) \
573 typeof(val) __val = (val); \
576 * paired with u64_u32_load_copy(), ordering access to var and \
583 # define u64_u32_load(var) u64_u32_load_copy(var, var##_copy)
584 # define u64_u32_store(var, val) u64_u32_store_copy(var, var##_copy, val)
586 /* CFS-related fields in a runqueue */
588 struct load_weight load;
589 unsigned int nr_running;
590 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */
591 unsigned int idle_nr_running; /* SCHED_IDLE */
592 unsigned int idle_h_nr_running; /* SCHED_IDLE */
599 #ifdef CONFIG_SCHED_CORE
600 unsigned int forceidle_seq;
605 u64 min_vruntime_copy;
608 struct rb_root_cached tasks_timeline;
611 * 'curr' points to currently running entity on this cfs_rq.
612 * It is set to NULL otherwise (i.e when none are currently running).
614 struct sched_entity *curr;
615 struct sched_entity *next;
617 #ifdef CONFIG_SCHED_DEBUG
618 unsigned int nr_spread_over;
625 struct sched_avg avg;
627 u64 last_update_time_copy;
630 raw_spinlock_t lock ____cacheline_aligned;
632 unsigned long load_avg;
633 unsigned long util_avg;
634 unsigned long runnable_avg;
637 #ifdef CONFIG_FAIR_GROUP_SCHED
638 u64 last_update_tg_load_avg;
639 unsigned long tg_load_avg_contrib;
641 long prop_runnable_sum;
644 * h_load = weight * f(tg)
646 * Where f(tg) is the recursive weight fraction assigned to
649 unsigned long h_load;
650 u64 last_h_load_update;
651 struct sched_entity *h_load_next;
652 #endif /* CONFIG_FAIR_GROUP_SCHED */
653 #endif /* CONFIG_SMP */
655 #ifdef CONFIG_FAIR_GROUP_SCHED
656 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */
659 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
660 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
661 * (like users, containers etc.)
663 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU.
664 * This list is used during load balance.
667 struct list_head leaf_cfs_rq_list;
668 struct task_group *tg; /* group that "owns" this runqueue */
670 /* Locally cached copy of our task_group's idle value */
673 #ifdef CONFIG_CFS_BANDWIDTH
675 s64 runtime_remaining;
677 u64 throttled_pelt_idle;
679 u64 throttled_pelt_idle_copy;
682 u64 throttled_clock_pelt;
683 u64 throttled_clock_pelt_time;
684 u64 throttled_clock_self;
685 u64 throttled_clock_self_time;
688 struct list_head throttled_list;
689 struct list_head throttled_csd_list;
690 #endif /* CONFIG_CFS_BANDWIDTH */
691 #endif /* CONFIG_FAIR_GROUP_SCHED */
694 static inline int rt_bandwidth_enabled(void)
696 return sysctl_sched_rt_runtime >= 0;
699 /* RT IPI pull logic requires IRQ_WORK */
700 #if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP)
701 # define HAVE_RT_PUSH_IPI
704 /* Real-Time classes' related field in a runqueue: */
706 struct rt_prio_array active;
707 unsigned int rt_nr_running;
708 unsigned int rr_nr_running;
709 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
711 int curr; /* highest queued rt task prio */
713 int next; /* next highest */
719 struct plist_head pushable_tasks;
721 #endif /* CONFIG_SMP */
727 /* Nests inside the rq lock: */
728 raw_spinlock_t rt_runtime_lock;
730 #ifdef CONFIG_RT_GROUP_SCHED
731 unsigned int rt_nr_boosted;
734 struct task_group *tg;
738 static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq)
740 return rt_rq->rt_queued && rt_rq->rt_nr_running;
743 /* Deadline class' related fields in a runqueue */
745 /* runqueue is an rbtree, ordered by deadline */
746 struct rb_root_cached root;
748 unsigned int dl_nr_running;
752 * Deadline values of the currently executing and the
753 * earliest ready task on this rq. Caching these facilitates
754 * the decision whether or not a ready but not running task
755 * should migrate somewhere else.
765 * Tasks on this rq that can be pushed away. They are kept in
766 * an rb-tree, ordered by tasks' deadlines, with caching
767 * of the leftmost (earliest deadline) element.
769 struct rb_root_cached pushable_dl_tasks_root;
774 * "Active utilization" for this runqueue: increased when a
775 * task wakes up (becomes TASK_RUNNING) and decreased when a
781 * Utilization of the tasks "assigned" to this runqueue (including
782 * the tasks that are in runqueue and the tasks that executed on this
783 * CPU and blocked). Increased when a task moves to this runqueue, and
784 * decreased when the task moves away (migrates, changes scheduling
785 * policy, or terminates).
786 * This is needed to compute the "inactive utilization" for the
787 * runqueue (inactive utilization = this_bw - running_bw).
793 * Maximum available bandwidth for reclaiming by SCHED_FLAG_RECLAIM
794 * tasks of this rq. Used in calculation of reclaimable bandwidth(GRUB).
799 * Inverse of the fraction of CPU utilization that can be reclaimed
800 * by the GRUB algorithm.
805 #ifdef CONFIG_FAIR_GROUP_SCHED
806 /* An entity is a task if it doesn't "own" a runqueue */
807 #define entity_is_task(se) (!se->my_q)
809 static inline void se_update_runnable(struct sched_entity *se)
811 if (!entity_is_task(se))
812 se->runnable_weight = se->my_q->h_nr_running;
815 static inline long se_runnable(struct sched_entity *se)
817 if (entity_is_task(se))
820 return se->runnable_weight;
824 #define entity_is_task(se) 1
826 static inline void se_update_runnable(struct sched_entity *se) {}
828 static inline long se_runnable(struct sched_entity *se)
836 * XXX we want to get rid of these helpers and use the full load resolution.
838 static inline long se_weight(struct sched_entity *se)
840 return scale_load_down(se->load.weight);
844 static inline bool sched_asym_prefer(int a, int b)
846 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b);
850 struct em_perf_domain *em_pd;
851 struct perf_domain *next;
856 * We add the notion of a root-domain which will be used to define per-domain
857 * variables. Each exclusive cpuset essentially defines an island domain by
858 * fully partitioning the member CPUs from any other cpuset. Whenever a new
859 * exclusive cpuset is created, we also create and attach a new root-domain
868 cpumask_var_t online;
871 * Indicate pullable load on at least one CPU, e.g:
872 * - More than one runnable task
873 * - Running task is misfit
877 /* Indicate one or more cpus over-utilized (tipping point) */
881 * The bit corresponding to a CPU gets set here if such CPU has more
882 * than one runnable -deadline task (as it is below for RT tasks).
884 cpumask_var_t dlo_mask;
890 * Indicate whether a root_domain's dl_bw has been checked or
891 * updated. It's monotonously increasing value.
893 * Also, some corner cases, like 'wrap around' is dangerous, but given
894 * that u64 is 'big enough'. So that shouldn't be a concern.
898 #ifdef HAVE_RT_PUSH_IPI
900 * For IPI pull requests, loop across the rto_mask.
902 struct irq_work rto_push_work;
903 raw_spinlock_t rto_lock;
904 /* These are only updated and read within rto_lock */
907 /* These atomics are updated outside of a lock */
908 atomic_t rto_loop_next;
909 atomic_t rto_loop_start;
912 * The "RT overload" flag: it gets set if a CPU has more than
913 * one runnable RT task.
915 cpumask_var_t rto_mask;
916 struct cpupri cpupri;
919 * NULL-terminated list of performance domains intersecting with the
920 * CPUs of the rd. Protected by RCU.
922 struct perf_domain __rcu *pd;
925 extern void init_defrootdomain(void);
926 extern int sched_init_domains(const struct cpumask *cpu_map);
927 extern void rq_attach_root(struct rq *rq, struct root_domain *rd);
928 extern void sched_get_rd(struct root_domain *rd);
929 extern void sched_put_rd(struct root_domain *rd);
931 static inline int get_rd_overloaded(struct root_domain *rd)
933 return READ_ONCE(rd->overloaded);
936 static inline void set_rd_overloaded(struct root_domain *rd, int status)
938 if (get_rd_overloaded(rd) != status)
939 WRITE_ONCE(rd->overloaded, status);
942 #ifdef HAVE_RT_PUSH_IPI
943 extern void rto_push_irq_work_func(struct irq_work *work);
945 #endif /* CONFIG_SMP */
947 #ifdef CONFIG_UCLAMP_TASK
949 * struct uclamp_bucket - Utilization clamp bucket
950 * @value: utilization clamp value for tasks on this clamp bucket
951 * @tasks: number of RUNNABLE tasks on this clamp bucket
953 * Keep track of how many tasks are RUNNABLE for a given utilization
956 struct uclamp_bucket {
957 unsigned long value : bits_per(SCHED_CAPACITY_SCALE);
958 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE);
962 * struct uclamp_rq - rq's utilization clamp
963 * @value: currently active clamp values for a rq
964 * @bucket: utilization clamp buckets affecting a rq
966 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values.
967 * A clamp value is affecting a rq when there is at least one task RUNNABLE
968 * (or actually running) with that value.
970 * There are up to UCLAMP_CNT possible different clamp values, currently there
971 * are only two: minimum utilization and maximum utilization.
973 * All utilization clamping values are MAX aggregated, since:
974 * - for util_min: we want to run the CPU at least at the max of the minimum
975 * utilization required by its currently RUNNABLE tasks.
976 * - for util_max: we want to allow the CPU to run up to the max of the
977 * maximum utilization allowed by its currently RUNNABLE tasks.
979 * Since on each system we expect only a limited number of different
980 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track
981 * the metrics required to compute all the per-rq utilization clamp values.
985 struct uclamp_bucket bucket[UCLAMP_BUCKETS];
988 DECLARE_STATIC_KEY_FALSE(sched_uclamp_used);
989 #endif /* CONFIG_UCLAMP_TASK */
992 struct balance_callback {
993 struct balance_callback *next;
994 void (*func)(struct rq *rq);
998 * This is the main, per-CPU runqueue data structure.
1000 * Locking rule: those places that want to lock multiple runqueues
1001 * (such as the load balancing or the thread migration code), lock
1002 * acquire operations must be ordered by ascending &runqueue.
1005 /* runqueue lock: */
1006 raw_spinlock_t __lock;
1008 unsigned int nr_running;
1009 #ifdef CONFIG_NUMA_BALANCING
1010 unsigned int nr_numa_running;
1011 unsigned int nr_preferred_running;
1012 unsigned int numa_migrate_on;
1014 #ifdef CONFIG_NO_HZ_COMMON
1016 unsigned long last_blocked_load_update_tick;
1017 unsigned int has_blocked_load;
1018 call_single_data_t nohz_csd;
1019 #endif /* CONFIG_SMP */
1020 unsigned int nohz_tick_stopped;
1021 atomic_t nohz_flags;
1022 #endif /* CONFIG_NO_HZ_COMMON */
1025 unsigned int ttwu_pending;
1029 #ifdef CONFIG_UCLAMP_TASK
1030 /* Utilization clamp values based on CPU's RUNNABLE tasks */
1031 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned;
1032 unsigned int uclamp_flags;
1033 #define UCLAMP_FLAG_IDLE 0x01
1040 #ifdef CONFIG_FAIR_GROUP_SCHED
1041 /* list of leaf cfs_rq on this CPU: */
1042 struct list_head leaf_cfs_rq_list;
1043 struct list_head *tmp_alone_branch;
1044 #endif /* CONFIG_FAIR_GROUP_SCHED */
1047 * This is part of a global counter where only the total sum
1048 * over all CPUs matters. A task can increase this counter on
1049 * one CPU and if it got migrated afterwards it may decrease
1050 * it on another CPU. Always updated under the runqueue lock:
1052 unsigned int nr_uninterruptible;
1054 struct task_struct __rcu *curr;
1055 struct task_struct *idle;
1056 struct task_struct *stop;
1057 unsigned long next_balance;
1058 struct mm_struct *prev_mm;
1060 unsigned int clock_update_flags;
1062 /* Ensure that all clocks are in the same cache line */
1063 u64 clock_task ____cacheline_aligned;
1065 unsigned long lost_idle_time;
1066 u64 clock_pelt_idle;
1068 #ifndef CONFIG_64BIT
1069 u64 clock_pelt_idle_copy;
1070 u64 clock_idle_copy;
1075 #ifdef CONFIG_SCHED_DEBUG
1076 u64 last_seen_need_resched_ns;
1077 int ticks_without_resched;
1080 #ifdef CONFIG_MEMBARRIER
1081 int membarrier_state;
1085 struct root_domain *rd;
1086 struct sched_domain __rcu *sd;
1088 unsigned long cpu_capacity;
1090 struct balance_callback *balance_callback;
1092 unsigned char nohz_idle_balance;
1093 unsigned char idle_balance;
1095 unsigned long misfit_task_load;
1097 /* For active balancing */
1100 struct cpu_stop_work active_balance_work;
1102 /* CPU of this runqueue: */
1106 struct list_head cfs_tasks;
1108 struct sched_avg avg_rt;
1109 struct sched_avg avg_dl;
1110 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
1111 struct sched_avg avg_irq;
1113 #ifdef CONFIG_SCHED_HW_PRESSURE
1114 struct sched_avg avg_hw;
1119 /* This is used to determine avg_idle's max value */
1120 u64 max_idle_balance_cost;
1122 #ifdef CONFIG_HOTPLUG_CPU
1123 struct rcuwait hotplug_wait;
1125 #endif /* CONFIG_SMP */
1127 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1130 #ifdef CONFIG_PARAVIRT
1131 u64 prev_steal_time;
1133 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1134 u64 prev_steal_time_rq;
1137 /* calc_load related fields */
1138 unsigned long calc_load_update;
1139 long calc_load_active;
1141 #ifdef CONFIG_SCHED_HRTICK
1143 call_single_data_t hrtick_csd;
1145 struct hrtimer hrtick_timer;
1146 ktime_t hrtick_time;
1149 #ifdef CONFIG_SCHEDSTATS
1151 struct sched_info rq_sched_info;
1152 unsigned long long rq_cpu_time;
1153 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
1155 /* sys_sched_yield() stats */
1156 unsigned int yld_count;
1158 /* schedule() stats */
1159 unsigned int sched_count;
1160 unsigned int sched_goidle;
1162 /* try_to_wake_up() stats */
1163 unsigned int ttwu_count;
1164 unsigned int ttwu_local;
1167 #ifdef CONFIG_CPU_IDLE
1168 /* Must be inspected within a rcu lock section */
1169 struct cpuidle_state *idle_state;
1173 unsigned int nr_pinned;
1175 unsigned int push_busy;
1176 struct cpu_stop_work push_work;
1178 #ifdef CONFIG_SCHED_CORE
1181 struct task_struct *core_pick;
1182 unsigned int core_enabled;
1183 unsigned int core_sched_seq;
1184 struct rb_root core_tree;
1186 /* shared state -- careful with sched_core_cpu_deactivate() */
1187 unsigned int core_task_seq;
1188 unsigned int core_pick_seq;
1189 unsigned long core_cookie;
1190 unsigned int core_forceidle_count;
1191 unsigned int core_forceidle_seq;
1192 unsigned int core_forceidle_occupation;
1193 u64 core_forceidle_start;
1196 /* Scratch cpumask to be temporarily used under rq_lock */
1197 cpumask_var_t scratch_mask;
1199 #if defined(CONFIG_CFS_BANDWIDTH) && defined(CONFIG_SMP)
1200 call_single_data_t cfsb_csd;
1201 struct list_head cfsb_csd_list;
1205 #ifdef CONFIG_FAIR_GROUP_SCHED
1207 /* CPU runqueue to which this cfs_rq is attached */
1208 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1215 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1217 return container_of(cfs_rq, struct rq, cfs);
1221 static inline int cpu_of(struct rq *rq)
1230 #define MDF_PUSH 0x01
1232 static inline bool is_migration_disabled(struct task_struct *p)
1235 return p->migration_disabled;
1241 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1243 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
1244 #define this_rq() this_cpu_ptr(&runqueues)
1245 #define task_rq(p) cpu_rq(task_cpu(p))
1246 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
1247 #define raw_rq() raw_cpu_ptr(&runqueues)
1250 #ifdef CONFIG_SCHED_CORE
1251 static inline struct cpumask *sched_group_span(struct sched_group *sg);
1253 DECLARE_STATIC_KEY_FALSE(__sched_core_enabled);
1255 static inline bool sched_core_enabled(struct rq *rq)
1257 return static_branch_unlikely(&__sched_core_enabled) && rq->core_enabled;
1260 static inline bool sched_core_disabled(void)
1262 return !static_branch_unlikely(&__sched_core_enabled);
1266 * Be careful with this function; not for general use. The return value isn't
1267 * stable unless you actually hold a relevant rq->__lock.
1269 static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1271 if (sched_core_enabled(rq))
1272 return &rq->core->__lock;
1277 static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1279 if (rq->core_enabled)
1280 return &rq->core->__lock;
1285 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
1287 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
1290 * Helpers to check if the CPU's core cookie matches with the task's cookie
1291 * when core scheduling is enabled.
1292 * A special case is that the task's cookie always matches with CPU's core
1293 * cookie if the CPU is in an idle core.
1295 static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1297 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1298 if (!sched_core_enabled(rq))
1301 return rq->core->core_cookie == p->core_cookie;
1304 static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1306 bool idle_core = true;
1309 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1310 if (!sched_core_enabled(rq))
1313 for_each_cpu(cpu, cpu_smt_mask(cpu_of(rq))) {
1314 if (!available_idle_cpu(cpu)) {
1321 * A CPU in an idle core is always the best choice for tasks with
1324 return idle_core || rq->core->core_cookie == p->core_cookie;
1327 static inline bool sched_group_cookie_match(struct rq *rq,
1328 struct task_struct *p,
1329 struct sched_group *group)
1333 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1334 if (!sched_core_enabled(rq))
1337 for_each_cpu_and(cpu, sched_group_span(group), p->cpus_ptr) {
1338 if (sched_core_cookie_match(cpu_rq(cpu), p))
1344 static inline bool sched_core_enqueued(struct task_struct *p)
1346 return !RB_EMPTY_NODE(&p->core_node);
1349 extern void sched_core_enqueue(struct rq *rq, struct task_struct *p);
1350 extern void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags);
1352 extern void sched_core_get(void);
1353 extern void sched_core_put(void);
1355 #else /* !CONFIG_SCHED_CORE */
1357 static inline bool sched_core_enabled(struct rq *rq)
1362 static inline bool sched_core_disabled(void)
1367 static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1372 static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1377 static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1382 static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1387 static inline bool sched_group_cookie_match(struct rq *rq,
1388 struct task_struct *p,
1389 struct sched_group *group)
1393 #endif /* CONFIG_SCHED_CORE */
1395 static inline void lockdep_assert_rq_held(struct rq *rq)
1397 lockdep_assert_held(__rq_lockp(rq));
1400 extern void raw_spin_rq_lock_nested(struct rq *rq, int subclass);
1401 extern bool raw_spin_rq_trylock(struct rq *rq);
1402 extern void raw_spin_rq_unlock(struct rq *rq);
1404 static inline void raw_spin_rq_lock(struct rq *rq)
1406 raw_spin_rq_lock_nested(rq, 0);
1409 static inline void raw_spin_rq_lock_irq(struct rq *rq)
1411 local_irq_disable();
1412 raw_spin_rq_lock(rq);
1415 static inline void raw_spin_rq_unlock_irq(struct rq *rq)
1417 raw_spin_rq_unlock(rq);
1421 static inline unsigned long _raw_spin_rq_lock_irqsave(struct rq *rq)
1423 unsigned long flags;
1424 local_irq_save(flags);
1425 raw_spin_rq_lock(rq);
1429 static inline void raw_spin_rq_unlock_irqrestore(struct rq *rq, unsigned long flags)
1431 raw_spin_rq_unlock(rq);
1432 local_irq_restore(flags);
1435 #define raw_spin_rq_lock_irqsave(rq, flags) \
1437 flags = _raw_spin_rq_lock_irqsave(rq); \
1440 #ifdef CONFIG_SCHED_SMT
1441 extern void __update_idle_core(struct rq *rq);
1443 static inline void update_idle_core(struct rq *rq)
1445 if (static_branch_unlikely(&sched_smt_present))
1446 __update_idle_core(rq);
1450 static inline void update_idle_core(struct rq *rq) { }
1453 #ifdef CONFIG_FAIR_GROUP_SCHED
1454 static inline struct task_struct *task_of(struct sched_entity *se)
1456 SCHED_WARN_ON(!entity_is_task(se));
1457 return container_of(se, struct task_struct, se);
1460 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
1462 return p->se.cfs_rq;
1465 /* runqueue on which this entity is (to be) queued */
1466 static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1471 /* runqueue "owned" by this group */
1472 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1479 #define task_of(_se) container_of(_se, struct task_struct, se)
1481 static inline struct cfs_rq *task_cfs_rq(const struct task_struct *p)
1483 return &task_rq(p)->cfs;
1486 static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1488 const struct task_struct *p = task_of(se);
1489 struct rq *rq = task_rq(p);
1494 /* runqueue "owned" by this group */
1495 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1501 extern void update_rq_clock(struct rq *rq);
1504 * rq::clock_update_flags bits
1506 * %RQCF_REQ_SKIP - will request skipping of clock update on the next
1507 * call to __schedule(). This is an optimisation to avoid
1508 * neighbouring rq clock updates.
1510 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is
1511 * in effect and calls to update_rq_clock() are being ignored.
1513 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been
1514 * made to update_rq_clock() since the last time rq::lock was pinned.
1516 * If inside of __schedule(), clock_update_flags will have been
1517 * shifted left (a left shift is a cheap operation for the fast path
1518 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use,
1520 * if (rq-clock_update_flags >= RQCF_UPDATED)
1522 * to check if %RQCF_UPDATED is set. It'll never be shifted more than
1523 * one position though, because the next rq_unpin_lock() will shift it
1526 #define RQCF_REQ_SKIP 0x01
1527 #define RQCF_ACT_SKIP 0x02
1528 #define RQCF_UPDATED 0x04
1530 static inline void assert_clock_updated(struct rq *rq)
1533 * The only reason for not seeing a clock update since the
1534 * last rq_pin_lock() is if we're currently skipping updates.
1536 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
1539 static inline u64 rq_clock(struct rq *rq)
1541 lockdep_assert_rq_held(rq);
1542 assert_clock_updated(rq);
1547 static inline u64 rq_clock_task(struct rq *rq)
1549 lockdep_assert_rq_held(rq);
1550 assert_clock_updated(rq);
1552 return rq->clock_task;
1555 static inline void rq_clock_skip_update(struct rq *rq)
1557 lockdep_assert_rq_held(rq);
1558 rq->clock_update_flags |= RQCF_REQ_SKIP;
1562 * See rt task throttling, which is the only time a skip
1563 * request is canceled.
1565 static inline void rq_clock_cancel_skipupdate(struct rq *rq)
1567 lockdep_assert_rq_held(rq);
1568 rq->clock_update_flags &= ~RQCF_REQ_SKIP;
1572 * During cpu offlining and rq wide unthrottling, we can trigger
1573 * an update_rq_clock() for several cfs and rt runqueues (Typically
1574 * when using list_for_each_entry_*)
1575 * rq_clock_start_loop_update() can be called after updating the clock
1576 * once and before iterating over the list to prevent multiple update.
1577 * After the iterative traversal, we need to call rq_clock_stop_loop_update()
1578 * to clear RQCF_ACT_SKIP of rq->clock_update_flags.
1580 static inline void rq_clock_start_loop_update(struct rq *rq)
1582 lockdep_assert_rq_held(rq);
1583 SCHED_WARN_ON(rq->clock_update_flags & RQCF_ACT_SKIP);
1584 rq->clock_update_flags |= RQCF_ACT_SKIP;
1587 static inline void rq_clock_stop_loop_update(struct rq *rq)
1589 lockdep_assert_rq_held(rq);
1590 rq->clock_update_flags &= ~RQCF_ACT_SKIP;
1594 unsigned long flags;
1595 struct pin_cookie cookie;
1596 #ifdef CONFIG_SCHED_DEBUG
1598 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the
1599 * current pin context is stashed here in case it needs to be
1600 * restored in rq_repin_lock().
1602 unsigned int clock_update_flags;
1606 extern struct balance_callback balance_push_callback;
1609 * Lockdep annotation that avoids accidental unlocks; it's like a
1610 * sticky/continuous lockdep_assert_held().
1612 * This avoids code that has access to 'struct rq *rq' (basically everything in
1613 * the scheduler) from accidentally unlocking the rq if they do not also have a
1614 * copy of the (on-stack) 'struct rq_flags rf'.
1616 * Also see Documentation/locking/lockdep-design.rst.
1618 static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
1620 rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
1622 #ifdef CONFIG_SCHED_DEBUG
1623 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
1624 rf->clock_update_flags = 0;
1626 SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
1631 static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
1633 #ifdef CONFIG_SCHED_DEBUG
1634 if (rq->clock_update_flags > RQCF_ACT_SKIP)
1635 rf->clock_update_flags = RQCF_UPDATED;
1638 lockdep_unpin_lock(__rq_lockp(rq), rf->cookie);
1641 static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf)
1643 lockdep_repin_lock(__rq_lockp(rq), rf->cookie);
1645 #ifdef CONFIG_SCHED_DEBUG
1647 * Restore the value we stashed in @rf for this pin context.
1649 rq->clock_update_flags |= rf->clock_update_flags;
1653 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1654 __acquires(rq->lock);
1656 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1657 __acquires(p->pi_lock)
1658 __acquires(rq->lock);
1660 static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf)
1661 __releases(rq->lock)
1663 rq_unpin_lock(rq, rf);
1664 raw_spin_rq_unlock(rq);
1668 task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1669 __releases(rq->lock)
1670 __releases(p->pi_lock)
1672 rq_unpin_lock(rq, rf);
1673 raw_spin_rq_unlock(rq);
1674 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
1677 DEFINE_LOCK_GUARD_1(task_rq_lock, struct task_struct,
1678 _T->rq = task_rq_lock(_T->lock, &_T->rf),
1679 task_rq_unlock(_T->rq, _T->lock, &_T->rf),
1680 struct rq *rq; struct rq_flags rf)
1683 rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
1684 __acquires(rq->lock)
1686 raw_spin_rq_lock_irqsave(rq, rf->flags);
1687 rq_pin_lock(rq, rf);
1691 rq_lock_irq(struct rq *rq, struct rq_flags *rf)
1692 __acquires(rq->lock)
1694 raw_spin_rq_lock_irq(rq);
1695 rq_pin_lock(rq, rf);
1699 rq_lock(struct rq *rq, struct rq_flags *rf)
1700 __acquires(rq->lock)
1702 raw_spin_rq_lock(rq);
1703 rq_pin_lock(rq, rf);
1707 rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
1708 __releases(rq->lock)
1710 rq_unpin_lock(rq, rf);
1711 raw_spin_rq_unlock_irqrestore(rq, rf->flags);
1715 rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
1716 __releases(rq->lock)
1718 rq_unpin_lock(rq, rf);
1719 raw_spin_rq_unlock_irq(rq);
1723 rq_unlock(struct rq *rq, struct rq_flags *rf)
1724 __releases(rq->lock)
1726 rq_unpin_lock(rq, rf);
1727 raw_spin_rq_unlock(rq);
1730 DEFINE_LOCK_GUARD_1(rq_lock, struct rq,
1731 rq_lock(_T->lock, &_T->rf),
1732 rq_unlock(_T->lock, &_T->rf),
1735 DEFINE_LOCK_GUARD_1(rq_lock_irq, struct rq,
1736 rq_lock_irq(_T->lock, &_T->rf),
1737 rq_unlock_irq(_T->lock, &_T->rf),
1740 DEFINE_LOCK_GUARD_1(rq_lock_irqsave, struct rq,
1741 rq_lock_irqsave(_T->lock, &_T->rf),
1742 rq_unlock_irqrestore(_T->lock, &_T->rf),
1745 static inline struct rq *
1746 this_rq_lock_irq(struct rq_flags *rf)
1747 __acquires(rq->lock)
1751 local_irq_disable();
1758 enum numa_topology_type {
1763 extern enum numa_topology_type sched_numa_topology_type;
1764 extern int sched_max_numa_distance;
1765 extern bool find_numa_distance(int distance);
1766 extern void sched_init_numa(int offline_node);
1767 extern void sched_update_numa(int cpu, bool online);
1768 extern void sched_domains_numa_masks_set(unsigned int cpu);
1769 extern void sched_domains_numa_masks_clear(unsigned int cpu);
1770 extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu);
1772 static inline void sched_init_numa(int offline_node) { }
1773 static inline void sched_update_numa(int cpu, bool online) { }
1774 static inline void sched_domains_numa_masks_set(unsigned int cpu) { }
1775 static inline void sched_domains_numa_masks_clear(unsigned int cpu) { }
1776 static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1782 #ifdef CONFIG_NUMA_BALANCING
1783 /* The regions in numa_faults array from task_struct */
1784 enum numa_faults_stats {
1790 extern void sched_setnuma(struct task_struct *p, int node);
1791 extern int migrate_task_to(struct task_struct *p, int cpu);
1792 extern int migrate_swap(struct task_struct *p, struct task_struct *t,
1794 extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p);
1797 init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1800 #endif /* CONFIG_NUMA_BALANCING */
1805 queue_balance_callback(struct rq *rq,
1806 struct balance_callback *head,
1807 void (*func)(struct rq *rq))
1809 lockdep_assert_rq_held(rq);
1812 * Don't (re)queue an already queued item; nor queue anything when
1813 * balance_push() is active, see the comment with
1814 * balance_push_callback.
1816 if (unlikely(head->next || rq->balance_callback == &balance_push_callback))
1820 head->next = rq->balance_callback;
1821 rq->balance_callback = head;
1824 #define rcu_dereference_check_sched_domain(p) \
1825 rcu_dereference_check((p), \
1826 lockdep_is_held(&sched_domains_mutex))
1829 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1830 * See destroy_sched_domains: call_rcu for details.
1832 * The domain tree of any CPU may only be accessed from within
1833 * preempt-disabled sections.
1835 #define for_each_domain(cpu, __sd) \
1836 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
1837 __sd; __sd = __sd->parent)
1839 /* A mask of all the SD flags that have the SDF_SHARED_CHILD metaflag */
1840 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_SHARED_CHILD)) |
1841 static const unsigned int SD_SHARED_CHILD_MASK =
1842 #include <linux/sched/sd_flags.h>
1847 * highest_flag_domain - Return highest sched_domain containing flag.
1848 * @cpu: The CPU whose highest level of sched domain is to
1850 * @flag: The flag to check for the highest sched_domain
1851 * for the given CPU.
1853 * Returns the highest sched_domain of a CPU which contains @flag. If @flag has
1854 * the SDF_SHARED_CHILD metaflag, all the children domains also have @flag.
1856 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
1858 struct sched_domain *sd, *hsd = NULL;
1860 for_each_domain(cpu, sd) {
1861 if (sd->flags & flag) {
1867 * Stop the search if @flag is known to be shared at lower
1868 * levels. It will not be found further up.
1870 if (flag & SD_SHARED_CHILD_MASK)
1877 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
1879 struct sched_domain *sd;
1881 for_each_domain(cpu, sd) {
1882 if (sd->flags & flag)
1889 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
1890 DECLARE_PER_CPU(int, sd_llc_size);
1891 DECLARE_PER_CPU(int, sd_llc_id);
1892 DECLARE_PER_CPU(int, sd_share_id);
1893 DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
1894 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
1895 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
1896 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
1897 extern struct static_key_false sched_asym_cpucapacity;
1898 extern struct static_key_false sched_cluster_active;
1900 static __always_inline bool sched_asym_cpucap_active(void)
1902 return static_branch_unlikely(&sched_asym_cpucapacity);
1905 struct sched_group_capacity {
1908 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity
1911 unsigned long capacity;
1912 unsigned long min_capacity; /* Min per-CPU capacity in group */
1913 unsigned long max_capacity; /* Max per-CPU capacity in group */
1914 unsigned long next_update;
1915 int imbalance; /* XXX unrelated to capacity but shared group state */
1917 #ifdef CONFIG_SCHED_DEBUG
1921 unsigned long cpumask[]; /* Balance mask */
1924 struct sched_group {
1925 struct sched_group *next; /* Must be a circular list */
1928 unsigned int group_weight;
1930 struct sched_group_capacity *sgc;
1931 int asym_prefer_cpu; /* CPU of highest priority in group */
1935 * The CPUs this group covers.
1937 * NOTE: this field is variable length. (Allocated dynamically
1938 * by attaching extra space to the end of the structure,
1939 * depending on how many CPUs the kernel has booted up with)
1941 unsigned long cpumask[];
1944 static inline struct cpumask *sched_group_span(struct sched_group *sg)
1946 return to_cpumask(sg->cpumask);
1950 * See build_balance_mask().
1952 static inline struct cpumask *group_balance_mask(struct sched_group *sg)
1954 return to_cpumask(sg->sgc->cpumask);
1957 extern int group_balance_cpu(struct sched_group *sg);
1959 #ifdef CONFIG_SCHED_DEBUG
1960 void update_sched_domain_debugfs(void);
1961 void dirty_sched_domain_sysctl(int cpu);
1963 static inline void update_sched_domain_debugfs(void)
1966 static inline void dirty_sched_domain_sysctl(int cpu)
1971 extern int sched_update_scaling(void);
1973 static inline const struct cpumask *task_user_cpus(struct task_struct *p)
1975 if (!p->user_cpus_ptr)
1976 return cpu_possible_mask; /* &init_task.cpus_mask */
1977 return p->user_cpus_ptr;
1979 #endif /* CONFIG_SMP */
1983 #if defined(CONFIG_SCHED_CORE) && defined(CONFIG_SCHEDSTATS)
1985 extern void __sched_core_account_forceidle(struct rq *rq);
1987 static inline void sched_core_account_forceidle(struct rq *rq)
1989 if (schedstat_enabled())
1990 __sched_core_account_forceidle(rq);
1993 extern void __sched_core_tick(struct rq *rq);
1995 static inline void sched_core_tick(struct rq *rq)
1997 if (sched_core_enabled(rq) && schedstat_enabled())
1998 __sched_core_tick(rq);
2003 static inline void sched_core_account_forceidle(struct rq *rq) {}
2005 static inline void sched_core_tick(struct rq *rq) {}
2007 #endif /* CONFIG_SCHED_CORE && CONFIG_SCHEDSTATS */
2009 #ifdef CONFIG_CGROUP_SCHED
2012 * Return the group to which this tasks belongs.
2014 * We cannot use task_css() and friends because the cgroup subsystem
2015 * changes that value before the cgroup_subsys::attach() method is called,
2016 * therefore we cannot pin it and might observe the wrong value.
2018 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
2019 * core changes this before calling sched_move_task().
2021 * Instead we use a 'copy' which is updated from sched_move_task() while
2022 * holding both task_struct::pi_lock and rq::lock.
2024 static inline struct task_group *task_group(struct task_struct *p)
2026 return p->sched_task_group;
2029 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
2030 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
2032 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
2033 struct task_group *tg = task_group(p);
2036 #ifdef CONFIG_FAIR_GROUP_SCHED
2037 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
2038 p->se.cfs_rq = tg->cfs_rq[cpu];
2039 p->se.parent = tg->se[cpu];
2040 p->se.depth = tg->se[cpu] ? tg->se[cpu]->depth + 1 : 0;
2043 #ifdef CONFIG_RT_GROUP_SCHED
2044 p->rt.rt_rq = tg->rt_rq[cpu];
2045 p->rt.parent = tg->rt_se[cpu];
2049 #else /* CONFIG_CGROUP_SCHED */
2051 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
2052 static inline struct task_group *task_group(struct task_struct *p)
2057 #endif /* CONFIG_CGROUP_SCHED */
2059 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
2061 set_task_rq(p, cpu);
2064 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2065 * successfully executed on another CPU. We must ensure that updates of
2066 * per-task data have been completed by this moment.
2069 WRITE_ONCE(task_thread_info(p)->cpu, cpu);
2075 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
2077 #ifdef CONFIG_SCHED_DEBUG
2078 # define const_debug __read_mostly
2080 # define const_debug const
2083 #define SCHED_FEAT(name, enabled) \
2084 __SCHED_FEAT_##name ,
2087 #include "features.h"
2093 #ifdef CONFIG_SCHED_DEBUG
2096 * To support run-time toggling of sched features, all the translation units
2097 * (but core.c) reference the sysctl_sched_features defined in core.c.
2099 extern const_debug unsigned int sysctl_sched_features;
2101 #ifdef CONFIG_JUMP_LABEL
2102 #define SCHED_FEAT(name, enabled) \
2103 static __always_inline bool static_branch_##name(struct static_key *key) \
2105 return static_key_##enabled(key); \
2108 #include "features.h"
2111 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
2112 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
2114 #else /* !CONFIG_JUMP_LABEL */
2116 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2118 #endif /* CONFIG_JUMP_LABEL */
2120 #else /* !SCHED_DEBUG */
2123 * Each translation unit has its own copy of sysctl_sched_features to allow
2124 * constants propagation at compile time and compiler optimization based on
2127 #define SCHED_FEAT(name, enabled) \
2128 (1UL << __SCHED_FEAT_##name) * enabled |
2129 static const_debug __maybe_unused unsigned int sysctl_sched_features =
2130 #include "features.h"
2134 #define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2136 #endif /* SCHED_DEBUG */
2138 extern struct static_key_false sched_numa_balancing;
2139 extern struct static_key_false sched_schedstats;
2141 static inline u64 global_rt_period(void)
2143 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
2146 static inline u64 global_rt_runtime(void)
2148 if (sysctl_sched_rt_runtime < 0)
2151 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
2154 static inline int task_current(struct rq *rq, struct task_struct *p)
2156 return rq->curr == p;
2159 static inline int task_on_cpu(struct rq *rq, struct task_struct *p)
2164 return task_current(rq, p);
2168 static inline int task_on_rq_queued(struct task_struct *p)
2170 return p->on_rq == TASK_ON_RQ_QUEUED;
2173 static inline int task_on_rq_migrating(struct task_struct *p)
2175 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING;
2178 /* Wake flags. The first three directly map to some SD flag value */
2179 #define WF_EXEC 0x02 /* Wakeup after exec; maps to SD_BALANCE_EXEC */
2180 #define WF_FORK 0x04 /* Wakeup after fork; maps to SD_BALANCE_FORK */
2181 #define WF_TTWU 0x08 /* Wakeup; maps to SD_BALANCE_WAKE */
2183 #define WF_SYNC 0x10 /* Waker goes to sleep after wakeup */
2184 #define WF_MIGRATED 0x20 /* Internal use, task got migrated */
2185 #define WF_CURRENT_CPU 0x40 /* Prefer to move the wakee to the current CPU. */
2188 static_assert(WF_EXEC == SD_BALANCE_EXEC);
2189 static_assert(WF_FORK == SD_BALANCE_FORK);
2190 static_assert(WF_TTWU == SD_BALANCE_WAKE);
2194 * To aid in avoiding the subversion of "niceness" due to uneven distribution
2195 * of tasks with abnormal "nice" values across CPUs the contribution that
2196 * each task makes to its run queue's load is weighted according to its
2197 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
2198 * scaled version of the new time slice allocation that they receive on time
2202 #define WEIGHT_IDLEPRIO 3
2203 #define WMULT_IDLEPRIO 1431655765
2205 extern const int sched_prio_to_weight[40];
2206 extern const u32 sched_prio_to_wmult[40];
2209 * {de,en}queue flags:
2211 * DEQUEUE_SLEEP - task is no longer runnable
2212 * ENQUEUE_WAKEUP - task just became runnable
2214 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
2215 * are in a known state which allows modification. Such pairs
2216 * should preserve as much state as possible.
2218 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
2221 * NOCLOCK - skip the update_rq_clock() (avoids double updates)
2223 * MIGRATION - p->on_rq == TASK_ON_RQ_MIGRATING (used for DEADLINE)
2225 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
2226 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
2227 * ENQUEUE_MIGRATED - the task was migrated during wakeup
2231 #define DEQUEUE_SLEEP 0x01
2232 #define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */
2233 #define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */
2234 #define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */
2235 #define DEQUEUE_MIGRATING 0x100 /* Matches ENQUEUE_MIGRATING */
2237 #define ENQUEUE_WAKEUP 0x01
2238 #define ENQUEUE_RESTORE 0x02
2239 #define ENQUEUE_MOVE 0x04
2240 #define ENQUEUE_NOCLOCK 0x08
2242 #define ENQUEUE_HEAD 0x10
2243 #define ENQUEUE_REPLENISH 0x20
2245 #define ENQUEUE_MIGRATED 0x40
2247 #define ENQUEUE_MIGRATED 0x00
2249 #define ENQUEUE_INITIAL 0x80
2250 #define ENQUEUE_MIGRATING 0x100
2252 #define RETRY_TASK ((void *)-1UL)
2254 struct affinity_context {
2255 const struct cpumask *new_mask;
2256 struct cpumask *user_mask;
2260 extern s64 update_curr_common(struct rq *rq);
2262 struct sched_class {
2264 #ifdef CONFIG_UCLAMP_TASK
2268 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
2269 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
2270 void (*yield_task) (struct rq *rq);
2271 bool (*yield_to_task)(struct rq *rq, struct task_struct *p);
2273 void (*wakeup_preempt)(struct rq *rq, struct task_struct *p, int flags);
2275 struct task_struct *(*pick_next_task)(struct rq *rq);
2277 void (*put_prev_task)(struct rq *rq, struct task_struct *p);
2278 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first);
2281 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2282 int (*select_task_rq)(struct task_struct *p, int task_cpu, int flags);
2284 struct task_struct * (*pick_task)(struct rq *rq);
2286 void (*migrate_task_rq)(struct task_struct *p, int new_cpu);
2288 void (*task_woken)(struct rq *this_rq, struct task_struct *task);
2290 void (*set_cpus_allowed)(struct task_struct *p, struct affinity_context *ctx);
2292 void (*rq_online)(struct rq *rq);
2293 void (*rq_offline)(struct rq *rq);
2295 struct rq *(*find_lock_rq)(struct task_struct *p, struct rq *rq);
2298 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued);
2299 void (*task_fork)(struct task_struct *p);
2300 void (*task_dead)(struct task_struct *p);
2303 * The switched_from() call is allowed to drop rq->lock, therefore we
2304 * cannot assume the switched_from/switched_to pair is serialized by
2305 * rq->lock. They are however serialized by p->pi_lock.
2307 void (*switched_from)(struct rq *this_rq, struct task_struct *task);
2308 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
2309 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
2312 unsigned int (*get_rr_interval)(struct rq *rq,
2313 struct task_struct *task);
2315 void (*update_curr)(struct rq *rq);
2317 #ifdef CONFIG_FAIR_GROUP_SCHED
2318 void (*task_change_group)(struct task_struct *p);
2321 #ifdef CONFIG_SCHED_CORE
2322 int (*task_is_throttled)(struct task_struct *p, int cpu);
2326 static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
2328 WARN_ON_ONCE(rq->curr != prev);
2329 prev->sched_class->put_prev_task(rq, prev);
2332 static inline void set_next_task(struct rq *rq, struct task_struct *next)
2334 next->sched_class->set_next_task(rq, next, false);
2339 * Helper to define a sched_class instance; each one is placed in a separate
2340 * section which is ordered by the linker script:
2342 * include/asm-generic/vmlinux.lds.h
2344 * *CAREFUL* they are laid out in *REVERSE* order!!!
2346 * Also enforce alignment on the instance, not the type, to guarantee layout.
2348 #define DEFINE_SCHED_CLASS(name) \
2349 const struct sched_class name##_sched_class \
2350 __aligned(__alignof__(struct sched_class)) \
2351 __section("__" #name "_sched_class")
2353 /* Defined in include/asm-generic/vmlinux.lds.h */
2354 extern struct sched_class __sched_class_highest[];
2355 extern struct sched_class __sched_class_lowest[];
2357 #define for_class_range(class, _from, _to) \
2358 for (class = (_from); class < (_to); class++)
2360 #define for_each_class(class) \
2361 for_class_range(class, __sched_class_highest, __sched_class_lowest)
2363 #define sched_class_above(_a, _b) ((_a) < (_b))
2365 extern const struct sched_class stop_sched_class;
2366 extern const struct sched_class dl_sched_class;
2367 extern const struct sched_class rt_sched_class;
2368 extern const struct sched_class fair_sched_class;
2369 extern const struct sched_class idle_sched_class;
2371 static inline bool sched_stop_runnable(struct rq *rq)
2373 return rq->stop && task_on_rq_queued(rq->stop);
2376 static inline bool sched_dl_runnable(struct rq *rq)
2378 return rq->dl.dl_nr_running > 0;
2381 static inline bool sched_rt_runnable(struct rq *rq)
2383 return rq->rt.rt_queued > 0;
2386 static inline bool sched_fair_runnable(struct rq *rq)
2388 return rq->cfs.nr_running > 0;
2391 extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2392 extern struct task_struct *pick_next_task_idle(struct rq *rq);
2394 #define SCA_CHECK 0x01
2395 #define SCA_MIGRATE_DISABLE 0x02
2396 #define SCA_MIGRATE_ENABLE 0x04
2397 #define SCA_USER 0x08
2401 extern void update_group_capacity(struct sched_domain *sd, int cpu);
2403 extern void sched_balance_trigger(struct rq *rq);
2405 extern void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx);
2407 static inline struct task_struct *get_push_task(struct rq *rq)
2409 struct task_struct *p = rq->curr;
2411 lockdep_assert_rq_held(rq);
2416 if (p->nr_cpus_allowed == 1)
2419 if (p->migration_disabled)
2422 rq->push_busy = true;
2423 return get_task_struct(p);
2426 extern int push_cpu_stop(void *arg);
2430 #ifdef CONFIG_CPU_IDLE
2431 static inline void idle_set_state(struct rq *rq,
2432 struct cpuidle_state *idle_state)
2434 rq->idle_state = idle_state;
2437 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2439 SCHED_WARN_ON(!rcu_read_lock_held());
2441 return rq->idle_state;
2444 static inline void idle_set_state(struct rq *rq,
2445 struct cpuidle_state *idle_state)
2449 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2455 extern void schedule_idle(void);
2456 asmlinkage void schedule_user(void);
2458 extern void sysrq_sched_debug_show(void);
2459 extern void sched_init_granularity(void);
2460 extern void update_max_interval(void);
2462 extern void init_sched_dl_class(void);
2463 extern void init_sched_rt_class(void);
2464 extern void init_sched_fair_class(void);
2466 extern void reweight_task(struct task_struct *p, int prio);
2468 extern void resched_curr(struct rq *rq);
2469 extern void resched_cpu(int cpu);
2471 extern struct rt_bandwidth def_rt_bandwidth;
2472 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
2473 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
2475 extern void init_dl_entity(struct sched_dl_entity *dl_se);
2478 #define BW_UNIT (1 << BW_SHIFT)
2479 #define RATIO_SHIFT 8
2480 #define MAX_BW_BITS (64 - BW_SHIFT)
2481 #define MAX_BW ((1ULL << MAX_BW_BITS) - 1)
2482 unsigned long to_ratio(u64 period, u64 runtime);
2484 extern void init_entity_runnable_average(struct sched_entity *se);
2485 extern void post_init_entity_util_avg(struct task_struct *p);
2487 #ifdef CONFIG_NO_HZ_FULL
2488 extern bool sched_can_stop_tick(struct rq *rq);
2489 extern int __init sched_tick_offload_init(void);
2492 * Tick may be needed by tasks in the runqueue depending on their policy and
2493 * requirements. If tick is needed, lets send the target an IPI to kick it out of
2494 * nohz mode if necessary.
2496 static inline void sched_update_tick_dependency(struct rq *rq)
2498 int cpu = cpu_of(rq);
2500 if (!tick_nohz_full_cpu(cpu))
2503 if (sched_can_stop_tick(rq))
2504 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
2506 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
2509 static inline int sched_tick_offload_init(void) { return 0; }
2510 static inline void sched_update_tick_dependency(struct rq *rq) { }
2513 static inline void add_nr_running(struct rq *rq, unsigned count)
2515 unsigned prev_nr = rq->nr_running;
2517 rq->nr_running = prev_nr + count;
2518 if (trace_sched_update_nr_running_tp_enabled()) {
2519 call_trace_sched_update_nr_running(rq, count);
2523 if (prev_nr < 2 && rq->nr_running >= 2)
2524 set_rd_overloaded(rq->rd, 1);
2527 sched_update_tick_dependency(rq);
2530 static inline void sub_nr_running(struct rq *rq, unsigned count)
2532 rq->nr_running -= count;
2533 if (trace_sched_update_nr_running_tp_enabled()) {
2534 call_trace_sched_update_nr_running(rq, -count);
2537 /* Check if we still need preemption */
2538 sched_update_tick_dependency(rq);
2541 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
2542 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
2544 extern void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags);
2546 #ifdef CONFIG_PREEMPT_RT
2547 #define SCHED_NR_MIGRATE_BREAK 8
2549 #define SCHED_NR_MIGRATE_BREAK 32
2552 extern const_debug unsigned int sysctl_sched_nr_migrate;
2553 extern const_debug unsigned int sysctl_sched_migration_cost;
2555 extern unsigned int sysctl_sched_base_slice;
2557 #ifdef CONFIG_SCHED_DEBUG
2558 extern int sysctl_resched_latency_warn_ms;
2559 extern int sysctl_resched_latency_warn_once;
2561 extern unsigned int sysctl_sched_tunable_scaling;
2563 extern unsigned int sysctl_numa_balancing_scan_delay;
2564 extern unsigned int sysctl_numa_balancing_scan_period_min;
2565 extern unsigned int sysctl_numa_balancing_scan_period_max;
2566 extern unsigned int sysctl_numa_balancing_scan_size;
2567 extern unsigned int sysctl_numa_balancing_hot_threshold;
2570 #ifdef CONFIG_SCHED_HRTICK
2574 * - enabled by features
2575 * - hrtimer is actually high res
2577 static inline int hrtick_enabled(struct rq *rq)
2579 if (!cpu_active(cpu_of(rq)))
2581 return hrtimer_is_hres_active(&rq->hrtick_timer);
2584 static inline int hrtick_enabled_fair(struct rq *rq)
2586 if (!sched_feat(HRTICK))
2588 return hrtick_enabled(rq);
2591 static inline int hrtick_enabled_dl(struct rq *rq)
2593 if (!sched_feat(HRTICK_DL))
2595 return hrtick_enabled(rq);
2598 void hrtick_start(struct rq *rq, u64 delay);
2602 static inline int hrtick_enabled_fair(struct rq *rq)
2607 static inline int hrtick_enabled_dl(struct rq *rq)
2612 static inline int hrtick_enabled(struct rq *rq)
2617 #endif /* CONFIG_SCHED_HRTICK */
2619 #ifndef arch_scale_freq_tick
2620 static __always_inline
2621 void arch_scale_freq_tick(void)
2626 #ifndef arch_scale_freq_capacity
2628 * arch_scale_freq_capacity - get the frequency scale factor of a given CPU.
2629 * @cpu: the CPU in question.
2631 * Return: the frequency scale factor normalized against SCHED_CAPACITY_SCALE, i.e.
2634 * ------ * SCHED_CAPACITY_SCALE
2637 static __always_inline
2638 unsigned long arch_scale_freq_capacity(int cpu)
2640 return SCHED_CAPACITY_SCALE;
2644 #ifdef CONFIG_SCHED_DEBUG
2646 * In double_lock_balance()/double_rq_lock(), we use raw_spin_rq_lock() to
2647 * acquire rq lock instead of rq_lock(). So at the end of these two functions
2648 * we need to call double_rq_clock_clear_update() to clear RQCF_UPDATED of
2649 * rq->clock_update_flags to avoid the WARN_DOUBLE_CLOCK warning.
2651 static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2)
2653 rq1->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2654 /* rq1 == rq2 for !CONFIG_SMP, so just clear RQCF_UPDATED once. */
2656 rq2->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2660 static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2) {}
2663 #define DEFINE_LOCK_GUARD_2(name, type, _lock, _unlock, ...) \
2664 __DEFINE_UNLOCK_GUARD(name, type, _unlock, type *lock2; __VA_ARGS__) \
2665 static inline class_##name##_t class_##name##_constructor(type *lock, type *lock2) \
2666 { class_##name##_t _t = { .lock = lock, .lock2 = lock2 }, *_T = &_t; \
2671 static inline bool rq_order_less(struct rq *rq1, struct rq *rq2)
2673 #ifdef CONFIG_SCHED_CORE
2675 * In order to not have {0,2},{1,3} turn into into an AB-BA,
2676 * order by core-id first and cpu-id second.
2680 * double_rq_lock(0,3); will take core-0, core-1 lock
2681 * double_rq_lock(1,2); will take core-1, core-0 lock
2683 * when only cpu-id is considered.
2685 if (rq1->core->cpu < rq2->core->cpu)
2687 if (rq1->core->cpu > rq2->core->cpu)
2691 * __sched_core_flip() relies on SMT having cpu-id lock order.
2694 return rq1->cpu < rq2->cpu;
2697 extern void double_rq_lock(struct rq *rq1, struct rq *rq2);
2699 #ifdef CONFIG_PREEMPTION
2702 * fair double_lock_balance: Safely acquires both rq->locks in a fair
2703 * way at the expense of forcing extra atomic operations in all
2704 * invocations. This assures that the double_lock is acquired using the
2705 * same underlying policy as the spinlock_t on this architecture, which
2706 * reduces latency compared to the unfair variant below. However, it
2707 * also adds more overhead and therefore may reduce throughput.
2709 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2710 __releases(this_rq->lock)
2711 __acquires(busiest->lock)
2712 __acquires(this_rq->lock)
2714 raw_spin_rq_unlock(this_rq);
2715 double_rq_lock(this_rq, busiest);
2722 * Unfair double_lock_balance: Optimizes throughput at the expense of
2723 * latency by eliminating extra atomic operations when the locks are
2724 * already in proper order on entry. This favors lower CPU-ids and will
2725 * grant the double lock to lower CPUs over higher ids under contention,
2726 * regardless of entry order into the function.
2728 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2729 __releases(this_rq->lock)
2730 __acquires(busiest->lock)
2731 __acquires(this_rq->lock)
2733 if (__rq_lockp(this_rq) == __rq_lockp(busiest) ||
2734 likely(raw_spin_rq_trylock(busiest))) {
2735 double_rq_clock_clear_update(this_rq, busiest);
2739 if (rq_order_less(this_rq, busiest)) {
2740 raw_spin_rq_lock_nested(busiest, SINGLE_DEPTH_NESTING);
2741 double_rq_clock_clear_update(this_rq, busiest);
2745 raw_spin_rq_unlock(this_rq);
2746 double_rq_lock(this_rq, busiest);
2751 #endif /* CONFIG_PREEMPTION */
2754 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2756 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2758 lockdep_assert_irqs_disabled();
2760 return _double_lock_balance(this_rq, busiest);
2763 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2764 __releases(busiest->lock)
2766 if (__rq_lockp(this_rq) != __rq_lockp(busiest))
2767 raw_spin_rq_unlock(busiest);
2768 lock_set_subclass(&__rq_lockp(this_rq)->dep_map, 0, _RET_IP_);
2771 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
2777 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2780 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
2786 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2789 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2795 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2798 static inline void double_raw_unlock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2800 raw_spin_unlock(l1);
2801 raw_spin_unlock(l2);
2804 DEFINE_LOCK_GUARD_2(double_raw_spinlock, raw_spinlock_t,
2805 double_raw_lock(_T->lock, _T->lock2),
2806 double_raw_unlock(_T->lock, _T->lock2))
2809 * double_rq_unlock - safely unlock two runqueues
2811 * Note this does not restore interrupts like task_rq_unlock,
2812 * you need to do so manually after calling.
2814 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2815 __releases(rq1->lock)
2816 __releases(rq2->lock)
2818 if (__rq_lockp(rq1) != __rq_lockp(rq2))
2819 raw_spin_rq_unlock(rq2);
2821 __release(rq2->lock);
2822 raw_spin_rq_unlock(rq1);
2825 extern void set_rq_online (struct rq *rq);
2826 extern void set_rq_offline(struct rq *rq);
2827 extern bool sched_smp_initialized;
2829 #else /* CONFIG_SMP */
2832 * double_rq_lock - safely lock two runqueues
2834 * Note this does not disable interrupts like task_rq_lock,
2835 * you need to do so manually before calling.
2837 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
2838 __acquires(rq1->lock)
2839 __acquires(rq2->lock)
2841 WARN_ON_ONCE(!irqs_disabled());
2842 WARN_ON_ONCE(rq1 != rq2);
2843 raw_spin_rq_lock(rq1);
2844 __acquire(rq2->lock); /* Fake it out ;) */
2845 double_rq_clock_clear_update(rq1, rq2);
2849 * double_rq_unlock - safely unlock two runqueues
2851 * Note this does not restore interrupts like task_rq_unlock,
2852 * you need to do so manually after calling.
2854 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2855 __releases(rq1->lock)
2856 __releases(rq2->lock)
2858 WARN_ON_ONCE(rq1 != rq2);
2859 raw_spin_rq_unlock(rq1);
2860 __release(rq2->lock);
2865 DEFINE_LOCK_GUARD_2(double_rq_lock, struct rq,
2866 double_rq_lock(_T->lock, _T->lock2),
2867 double_rq_unlock(_T->lock, _T->lock2))
2869 extern struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq);
2870 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
2871 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
2873 #ifdef CONFIG_SCHED_DEBUG
2874 extern bool sched_debug_verbose;
2876 extern void print_cfs_stats(struct seq_file *m, int cpu);
2877 extern void print_rt_stats(struct seq_file *m, int cpu);
2878 extern void print_dl_stats(struct seq_file *m, int cpu);
2879 extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
2880 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2881 extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
2883 extern void resched_latency_warn(int cpu, u64 latency);
2884 #ifdef CONFIG_NUMA_BALANCING
2886 show_numa_stats(struct task_struct *p, struct seq_file *m);
2888 print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
2889 unsigned long tpf, unsigned long gsf, unsigned long gpf);
2890 #endif /* CONFIG_NUMA_BALANCING */
2892 static inline void resched_latency_warn(int cpu, u64 latency) {}
2893 #endif /* CONFIG_SCHED_DEBUG */
2895 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
2896 extern void init_rt_rq(struct rt_rq *rt_rq);
2897 extern void init_dl_rq(struct dl_rq *dl_rq);
2899 extern void cfs_bandwidth_usage_inc(void);
2900 extern void cfs_bandwidth_usage_dec(void);
2902 #ifdef CONFIG_NO_HZ_COMMON
2903 #define NOHZ_BALANCE_KICK_BIT 0
2904 #define NOHZ_STATS_KICK_BIT 1
2905 #define NOHZ_NEWILB_KICK_BIT 2
2906 #define NOHZ_NEXT_KICK_BIT 3
2908 /* Run sched_balance_domains() */
2909 #define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT)
2910 /* Update blocked load */
2911 #define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT)
2912 /* Update blocked load when entering idle */
2913 #define NOHZ_NEWILB_KICK BIT(NOHZ_NEWILB_KICK_BIT)
2914 /* Update nohz.next_balance */
2915 #define NOHZ_NEXT_KICK BIT(NOHZ_NEXT_KICK_BIT)
2917 #define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK | NOHZ_NEXT_KICK)
2919 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
2921 extern void nohz_balance_exit_idle(struct rq *rq);
2923 static inline void nohz_balance_exit_idle(struct rq *rq) { }
2926 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
2927 extern void nohz_run_idle_balance(int cpu);
2929 static inline void nohz_run_idle_balance(int cpu) { }
2932 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2937 struct u64_stats_sync sync;
2940 DECLARE_PER_CPU(struct irqtime, cpu_irqtime);
2943 * Returns the irqtime minus the softirq time computed by ksoftirqd.
2944 * Otherwise ksoftirqd's sum_exec_runtime is subtracted its own runtime
2945 * and never move forward.
2947 static inline u64 irq_time_read(int cpu)
2949 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu);
2954 seq = __u64_stats_fetch_begin(&irqtime->sync);
2955 total = irqtime->total;
2956 } while (__u64_stats_fetch_retry(&irqtime->sync, seq));
2960 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2962 #ifdef CONFIG_CPU_FREQ
2963 DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data);
2966 * cpufreq_update_util - Take a note about CPU utilization changes.
2967 * @rq: Runqueue to carry out the update for.
2968 * @flags: Update reason flags.
2970 * This function is called by the scheduler on the CPU whose utilization is
2973 * It can only be called from RCU-sched read-side critical sections.
2975 * The way cpufreq is currently arranged requires it to evaluate the CPU
2976 * performance state (frequency/voltage) on a regular basis to prevent it from
2977 * being stuck in a completely inadequate performance level for too long.
2978 * That is not guaranteed to happen if the updates are only triggered from CFS
2979 * and DL, though, because they may not be coming in if only RT tasks are
2980 * active all the time (or there are RT tasks only).
2982 * As a workaround for that issue, this function is called periodically by the
2983 * RT sched class to trigger extra cpufreq updates to prevent it from stalling,
2984 * but that really is a band-aid. Going forward it should be replaced with
2985 * solutions targeted more specifically at RT tasks.
2987 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags)
2989 struct update_util_data *data;
2991 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data,
2994 data->func(data, rq_clock(rq), flags);
2997 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
2998 #endif /* CONFIG_CPU_FREQ */
3000 #ifdef arch_scale_freq_capacity
3001 # ifndef arch_scale_freq_invariant
3002 # define arch_scale_freq_invariant() true
3005 # define arch_scale_freq_invariant() false
3009 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
3011 unsigned long *max);
3013 unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
3019 * Verify the fitness of task @p to run on @cpu taking into account the
3020 * CPU original capacity and the runtime/deadline ratio of the task.
3022 * The function will return true if the original capacity of @cpu is
3023 * greater than or equal to task's deadline density right shifted by
3024 * (BW_SHIFT - SCHED_CAPACITY_SHIFT) and false otherwise.
3026 static inline bool dl_task_fits_capacity(struct task_struct *p, int cpu)
3028 unsigned long cap = arch_scale_cpu_capacity(cpu);
3030 return cap >= p->dl.dl_density >> (BW_SHIFT - SCHED_CAPACITY_SHIFT);
3033 static inline unsigned long cpu_bw_dl(struct rq *rq)
3035 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT;
3038 static inline unsigned long cpu_util_dl(struct rq *rq)
3040 return READ_ONCE(rq->avg_dl.util_avg);
3044 extern unsigned long cpu_util_cfs(int cpu);
3045 extern unsigned long cpu_util_cfs_boost(int cpu);
3047 static inline unsigned long cpu_util_rt(struct rq *rq)
3049 return READ_ONCE(rq->avg_rt.util_avg);
3053 #ifdef CONFIG_UCLAMP_TASK
3054 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id);
3056 static inline unsigned long uclamp_rq_get(struct rq *rq,
3057 enum uclamp_id clamp_id)
3059 return READ_ONCE(rq->uclamp[clamp_id].value);
3062 static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3065 WRITE_ONCE(rq->uclamp[clamp_id].value, value);
3068 static inline bool uclamp_rq_is_idle(struct rq *rq)
3070 return rq->uclamp_flags & UCLAMP_FLAG_IDLE;
3073 /* Is the rq being capped/throttled by uclamp_max? */
3074 static inline bool uclamp_rq_is_capped(struct rq *rq)
3076 unsigned long rq_util;
3077 unsigned long max_util;
3079 if (!static_branch_likely(&sched_uclamp_used))
3082 rq_util = cpu_util_cfs(cpu_of(rq)) + cpu_util_rt(rq);
3083 max_util = READ_ONCE(rq->uclamp[UCLAMP_MAX].value);
3085 return max_util != SCHED_CAPACITY_SCALE && rq_util >= max_util;
3089 * When uclamp is compiled in, the aggregation at rq level is 'turned off'
3090 * by default in the fast path and only gets turned on once userspace performs
3091 * an operation that requires it.
3093 * Returns true if userspace opted-in to use uclamp and aggregation at rq level
3096 static inline bool uclamp_is_used(void)
3098 return static_branch_likely(&sched_uclamp_used);
3100 #else /* CONFIG_UCLAMP_TASK */
3101 static inline unsigned long uclamp_eff_value(struct task_struct *p,
3102 enum uclamp_id clamp_id)
3104 if (clamp_id == UCLAMP_MIN)
3107 return SCHED_CAPACITY_SCALE;
3110 static inline bool uclamp_rq_is_capped(struct rq *rq) { return false; }
3112 static inline bool uclamp_is_used(void)
3117 static inline unsigned long uclamp_rq_get(struct rq *rq,
3118 enum uclamp_id clamp_id)
3120 if (clamp_id == UCLAMP_MIN)
3123 return SCHED_CAPACITY_SCALE;
3126 static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3131 static inline bool uclamp_rq_is_idle(struct rq *rq)
3135 #endif /* CONFIG_UCLAMP_TASK */
3137 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
3138 static inline unsigned long cpu_util_irq(struct rq *rq)
3140 return READ_ONCE(rq->avg_irq.util_avg);
3144 unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3146 util *= (max - irq);
3153 static inline unsigned long cpu_util_irq(struct rq *rq)
3159 unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3165 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
3167 #define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus)))
3169 DECLARE_STATIC_KEY_FALSE(sched_energy_present);
3171 static inline bool sched_energy_enabled(void)
3173 return static_branch_unlikely(&sched_energy_present);
3176 extern struct cpufreq_governor schedutil_gov;
3178 #else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */
3180 #define perf_domain_span(pd) NULL
3181 static inline bool sched_energy_enabled(void) { return false; }
3183 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
3185 #ifdef CONFIG_MEMBARRIER
3187 * The scheduler provides memory barriers required by membarrier between:
3188 * - prior user-space memory accesses and store to rq->membarrier_state,
3189 * - store to rq->membarrier_state and following user-space memory accesses.
3190 * In the same way it provides those guarantees around store to rq->curr.
3192 static inline void membarrier_switch_mm(struct rq *rq,
3193 struct mm_struct *prev_mm,
3194 struct mm_struct *next_mm)
3196 int membarrier_state;
3198 if (prev_mm == next_mm)
3201 membarrier_state = atomic_read(&next_mm->membarrier_state);
3202 if (READ_ONCE(rq->membarrier_state) == membarrier_state)
3205 WRITE_ONCE(rq->membarrier_state, membarrier_state);
3208 static inline void membarrier_switch_mm(struct rq *rq,
3209 struct mm_struct *prev_mm,
3210 struct mm_struct *next_mm)
3216 static inline bool is_per_cpu_kthread(struct task_struct *p)
3218 if (!(p->flags & PF_KTHREAD))
3221 if (p->nr_cpus_allowed != 1)
3228 extern void swake_up_all_locked(struct swait_queue_head *q);
3229 extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);
3231 extern int try_to_wake_up(struct task_struct *tsk, unsigned int state, int wake_flags);
3233 #ifdef CONFIG_PREEMPT_DYNAMIC
3234 extern int preempt_dynamic_mode;
3235 extern int sched_dynamic_mode(const char *str);
3236 extern void sched_dynamic_update(int mode);
3239 #ifdef CONFIG_SCHED_MM_CID
3241 #define SCHED_MM_CID_PERIOD_NS (100ULL * 1000000) /* 100ms */
3242 #define MM_CID_SCAN_DELAY 100 /* 100ms */
3244 extern raw_spinlock_t cid_lock;
3245 extern int use_cid_lock;
3247 extern void sched_mm_cid_migrate_from(struct task_struct *t);
3248 extern void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t);
3249 extern void task_tick_mm_cid(struct rq *rq, struct task_struct *curr);
3250 extern void init_sched_mm_cid(struct task_struct *t);
3252 static inline void __mm_cid_put(struct mm_struct *mm, int cid)
3256 cpumask_clear_cpu(cid, mm_cidmask(mm));
3260 * The per-mm/cpu cid can have the MM_CID_LAZY_PUT flag set or transition to
3261 * the MM_CID_UNSET state without holding the rq lock, but the rq lock needs to
3262 * be held to transition to other states.
3264 * State transitions synchronized with cmpxchg or try_cmpxchg need to be
3265 * consistent across cpus, which prevents use of this_cpu_cmpxchg.
3267 static inline void mm_cid_put_lazy(struct task_struct *t)
3269 struct mm_struct *mm = t->mm;
3270 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3273 lockdep_assert_irqs_disabled();
3274 cid = __this_cpu_read(pcpu_cid->cid);
3275 if (!mm_cid_is_lazy_put(cid) ||
3276 !try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3278 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3281 static inline int mm_cid_pcpu_unset(struct mm_struct *mm)
3283 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3286 lockdep_assert_irqs_disabled();
3287 cid = __this_cpu_read(pcpu_cid->cid);
3289 if (mm_cid_is_unset(cid))
3290 return MM_CID_UNSET;
3292 * Attempt transition from valid or lazy-put to unset.
3294 res = cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, cid, MM_CID_UNSET);
3302 static inline void mm_cid_put(struct mm_struct *mm)
3306 lockdep_assert_irqs_disabled();
3307 cid = mm_cid_pcpu_unset(mm);
3308 if (cid == MM_CID_UNSET)
3310 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3313 static inline int __mm_cid_try_get(struct mm_struct *mm)
3315 struct cpumask *cpumask;
3318 cpumask = mm_cidmask(mm);
3320 * Retry finding first zero bit if the mask is temporarily
3321 * filled. This only happens during concurrent remote-clear
3322 * which owns a cid without holding a rq lock.
3325 cid = cpumask_first_zero(cpumask);
3326 if (cid < nr_cpu_ids)
3330 if (cpumask_test_and_set_cpu(cid, cpumask))
3336 * Save a snapshot of the current runqueue time of this cpu
3337 * with the per-cpu cid value, allowing to estimate how recently it was used.
3339 static inline void mm_cid_snapshot_time(struct rq *rq, struct mm_struct *mm)
3341 struct mm_cid *pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(rq));
3343 lockdep_assert_rq_held(rq);
3344 WRITE_ONCE(pcpu_cid->time, rq->clock);
3347 static inline int __mm_cid_get(struct rq *rq, struct mm_struct *mm)
3352 * All allocations (even those using the cid_lock) are lock-free. If
3353 * use_cid_lock is set, hold the cid_lock to perform cid allocation to
3354 * guarantee forward progress.
3356 if (!READ_ONCE(use_cid_lock)) {
3357 cid = __mm_cid_try_get(mm);
3360 raw_spin_lock(&cid_lock);
3362 raw_spin_lock(&cid_lock);
3363 cid = __mm_cid_try_get(mm);
3369 * cid concurrently allocated. Retry while forcing following
3370 * allocations to use the cid_lock to ensure forward progress.
3372 WRITE_ONCE(use_cid_lock, 1);
3374 * Set use_cid_lock before allocation. Only care about program order
3375 * because this is only required for forward progress.
3379 * Retry until it succeeds. It is guaranteed to eventually succeed once
3380 * all newcoming allocations observe the use_cid_lock flag set.
3383 cid = __mm_cid_try_get(mm);
3387 * Allocate before clearing use_cid_lock. Only care about
3388 * program order because this is for forward progress.
3391 WRITE_ONCE(use_cid_lock, 0);
3393 raw_spin_unlock(&cid_lock);
3395 mm_cid_snapshot_time(rq, mm);
3399 static inline int mm_cid_get(struct rq *rq, struct mm_struct *mm)
3401 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3402 struct cpumask *cpumask;
3405 lockdep_assert_rq_held(rq);
3406 cpumask = mm_cidmask(mm);
3407 cid = __this_cpu_read(pcpu_cid->cid);
3408 if (mm_cid_is_valid(cid)) {
3409 mm_cid_snapshot_time(rq, mm);
3412 if (mm_cid_is_lazy_put(cid)) {
3413 if (try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3414 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3416 cid = __mm_cid_get(rq, mm);
3417 __this_cpu_write(pcpu_cid->cid, cid);
3421 static inline void switch_mm_cid(struct rq *rq,
3422 struct task_struct *prev,
3423 struct task_struct *next)
3426 * Provide a memory barrier between rq->curr store and load of
3427 * {prev,next}->mm->pcpu_cid[cpu] on rq->curr->mm transition.
3429 * Should be adapted if context_switch() is modified.
3431 if (!next->mm) { // to kernel
3433 * user -> kernel transition does not guarantee a barrier, but
3434 * we can use the fact that it performs an atomic operation in
3437 if (prev->mm) // from user
3438 smp_mb__after_mmgrab();
3440 * kernel -> kernel transition does not change rq->curr->mm
3441 * state. It stays NULL.
3445 * kernel -> user transition does not provide a barrier
3446 * between rq->curr store and load of {prev,next}->mm->pcpu_cid[cpu].
3449 if (!prev->mm) { // from kernel
3451 } else { // from user
3453 * user->user transition relies on an implicit
3454 * memory barrier in switch_mm() when
3455 * current->mm changes. If the architecture
3456 * switch_mm() does not have an implicit memory
3457 * barrier, it is emitted here. If current->mm
3458 * is unchanged, no barrier is needed.
3460 smp_mb__after_switch_mm();
3463 if (prev->mm_cid_active) {
3464 mm_cid_snapshot_time(rq, prev->mm);
3465 mm_cid_put_lazy(prev);
3468 if (next->mm_cid_active)
3469 next->last_mm_cid = next->mm_cid = mm_cid_get(rq, next->mm);
3473 static inline void switch_mm_cid(struct rq *rq, struct task_struct *prev, struct task_struct *next) { }
3474 static inline void sched_mm_cid_migrate_from(struct task_struct *t) { }
3475 static inline void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t) { }
3476 static inline void task_tick_mm_cid(struct rq *rq, struct task_struct *curr) { }
3477 static inline void init_sched_mm_cid(struct task_struct *t) { }
3480 extern u64 avg_vruntime(struct cfs_rq *cfs_rq);
3481 extern int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se);
3483 #endif /* _KERNEL_SCHED_SCHED_H */