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_CGROUP_SCHED
78 #include <linux/cgroup.h>
79 #include <linux/psi.h>
82 #ifdef CONFIG_SCHED_DEBUG
83 # include <linux/static_key.h>
86 #ifdef CONFIG_PARAVIRT
87 # include <asm/paravirt.h>
88 # include <asm/paravirt_api_clock.h>
92 #include "cpudeadline.h"
94 #ifdef CONFIG_SCHED_DEBUG
95 # define SCHED_WARN_ON(x) WARN_ONCE(x, #x)
97 # define SCHED_WARN_ON(x) ({ (void)(x), 0; })
101 struct cpuidle_state;
103 /* task_struct::on_rq states: */
104 #define TASK_ON_RQ_QUEUED 1
105 #define TASK_ON_RQ_MIGRATING 2
107 extern __read_mostly int scheduler_running;
109 extern unsigned long calc_load_update;
110 extern atomic_long_t calc_load_tasks;
112 extern unsigned int sysctl_sched_child_runs_first;
114 extern void calc_global_load_tick(struct rq *this_rq);
115 extern long calc_load_fold_active(struct rq *this_rq, long adjust);
117 extern void call_trace_sched_update_nr_running(struct rq *rq, int count);
119 extern unsigned int sysctl_sched_rt_period;
120 extern int sysctl_sched_rt_runtime;
121 extern int sched_rr_timeslice;
124 * Helpers for converting nanosecond timing to jiffy resolution
126 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
129 * Increase resolution of nice-level calculations for 64-bit architectures.
130 * The extra resolution improves shares distribution and load balancing of
131 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
132 * hierarchies, especially on larger systems. This is not a user-visible change
133 * and does not change the user-interface for setting shares/weights.
135 * We increase resolution only if we have enough bits to allow this increased
136 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
137 * are pretty high and the returns do not justify the increased costs.
139 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to
140 * increase coverage and consistency always enable it on 64-bit platforms.
143 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT)
144 # define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT)
145 # define scale_load_down(w) \
147 unsigned long __w = (w); \
149 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \
153 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT)
154 # define scale_load(w) (w)
155 # define scale_load_down(w) (w)
159 * Task weight (visible to users) and its load (invisible to users) have
160 * independent resolution, but they should be well calibrated. We use
161 * scale_load() and scale_load_down(w) to convert between them. The
162 * following must be true:
164 * scale_load(sched_prio_to_weight[NICE_TO_PRIO(0)-MAX_RT_PRIO]) == NICE_0_LOAD
167 #define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT)
170 * Single value that decides SCHED_DEADLINE internal math precision.
171 * 10 -> just above 1us
172 * 9 -> just above 0.5us
177 * Single value that denotes runtime == period, ie unlimited time.
179 #define RUNTIME_INF ((u64)~0ULL)
181 static inline int idle_policy(int policy)
183 return policy == SCHED_IDLE;
185 static inline int fair_policy(int policy)
187 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
190 static inline int rt_policy(int policy)
192 return policy == SCHED_FIFO || policy == SCHED_RR;
195 static inline int dl_policy(int policy)
197 return policy == SCHED_DEADLINE;
199 static inline bool valid_policy(int policy)
201 return idle_policy(policy) || fair_policy(policy) ||
202 rt_policy(policy) || dl_policy(policy);
205 static inline int task_has_idle_policy(struct task_struct *p)
207 return idle_policy(p->policy);
210 static inline int task_has_rt_policy(struct task_struct *p)
212 return rt_policy(p->policy);
215 static inline int task_has_dl_policy(struct task_struct *p)
217 return dl_policy(p->policy);
220 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
222 static inline void update_avg(u64 *avg, u64 sample)
224 s64 diff = sample - *avg;
229 * Shifting a value by an exponent greater *or equal* to the size of said value
230 * is UB; cap at size-1.
232 #define shr_bound(val, shift) \
233 (val >> min_t(typeof(shift), shift, BITS_PER_TYPE(typeof(val)) - 1))
236 * !! For sched_setattr_nocheck() (kernel) only !!
238 * This is actually gross. :(
240 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE
241 * tasks, but still be able to sleep. We need this on platforms that cannot
242 * atomically change clock frequency. Remove once fast switching will be
243 * available on such platforms.
245 * SUGOV stands for SchedUtil GOVernor.
247 #define SCHED_FLAG_SUGOV 0x10000000
249 #define SCHED_DL_FLAGS (SCHED_FLAG_RECLAIM | SCHED_FLAG_DL_OVERRUN | SCHED_FLAG_SUGOV)
251 static inline bool dl_entity_is_special(const struct sched_dl_entity *dl_se)
253 #ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
254 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV);
261 * Tells if entity @a should preempt entity @b.
263 static inline bool dl_entity_preempt(const struct sched_dl_entity *a,
264 const struct sched_dl_entity *b)
266 return dl_entity_is_special(a) ||
267 dl_time_before(a->deadline, b->deadline);
271 * This is the priority-queue data structure of the RT scheduling class:
273 struct rt_prio_array {
274 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
275 struct list_head queue[MAX_RT_PRIO];
278 struct rt_bandwidth {
279 /* nests inside the rq lock: */
280 raw_spinlock_t rt_runtime_lock;
283 struct hrtimer rt_period_timer;
284 unsigned int rt_period_active;
287 void __dl_clear_params(struct task_struct *p);
289 static inline int dl_bandwidth_enabled(void)
291 return sysctl_sched_rt_runtime >= 0;
295 * To keep the bandwidth of -deadline tasks under control
296 * we need some place where:
297 * - store the maximum -deadline bandwidth of each cpu;
298 * - cache the fraction of bandwidth that is currently allocated in
301 * This is all done in the data structure below. It is similar to the
302 * one used for RT-throttling (rt_bandwidth), with the main difference
303 * that, since here we are only interested in admission control, we
304 * do not decrease any runtime while the group "executes", neither we
305 * need a timer to replenish it.
307 * With respect to SMP, bandwidth is given on a per root domain basis,
309 * - bw (< 100%) is the deadline bandwidth of each CPU;
310 * - total_bw is the currently allocated bandwidth in each root domain;
318 extern void init_dl_bw(struct dl_bw *dl_b);
319 extern int sched_dl_global_validate(void);
320 extern void sched_dl_do_global(void);
321 extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr);
322 extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr);
323 extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr);
324 extern bool __checkparam_dl(const struct sched_attr *attr);
325 extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr);
326 extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial);
327 extern int dl_bw_check_overflow(int cpu);
329 #ifdef CONFIG_CGROUP_SCHED
334 extern struct list_head task_groups;
336 struct cfs_bandwidth {
337 #ifdef CONFIG_CFS_BANDWIDTH
344 s64 hierarchical_quota;
349 struct hrtimer period_timer;
350 struct hrtimer slack_timer;
351 struct list_head throttled_cfs_rq;
362 /* Task group related information */
364 struct cgroup_subsys_state css;
366 #ifdef CONFIG_FAIR_GROUP_SCHED
367 /* schedulable entities of this group on each CPU */
368 struct sched_entity **se;
369 /* runqueue "owned" by this group on each CPU */
370 struct cfs_rq **cfs_rq;
371 unsigned long shares;
373 /* A positive value indicates that this is a SCHED_IDLE group. */
378 * load_avg can be heavily contended at clock tick time, so put
379 * it in its own cacheline separated from the fields above which
380 * will also be accessed at each tick.
382 atomic_long_t load_avg ____cacheline_aligned;
386 #ifdef CONFIG_RT_GROUP_SCHED
387 struct sched_rt_entity **rt_se;
388 struct rt_rq **rt_rq;
390 struct rt_bandwidth rt_bandwidth;
394 struct list_head list;
396 struct task_group *parent;
397 struct list_head siblings;
398 struct list_head children;
400 #ifdef CONFIG_SCHED_AUTOGROUP
401 struct autogroup *autogroup;
404 struct cfs_bandwidth cfs_bandwidth;
406 #ifdef CONFIG_UCLAMP_TASK_GROUP
407 /* The two decimal precision [%] value requested from user-space */
408 unsigned int uclamp_pct[UCLAMP_CNT];
409 /* Clamp values requested for a task group */
410 struct uclamp_se uclamp_req[UCLAMP_CNT];
411 /* Effective clamp values used for a task group */
412 struct uclamp_se uclamp[UCLAMP_CNT];
417 #ifdef CONFIG_FAIR_GROUP_SCHED
418 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
421 * A weight of 0 or 1 can cause arithmetics problems.
422 * A weight of a cfs_rq is the sum of weights of which entities
423 * are queued on this cfs_rq, so a weight of a entity should not be
424 * too large, so as the shares value of a task group.
425 * (The default weight is 1024 - so there's no practical
426 * limitation from this.)
428 #define MIN_SHARES (1UL << 1)
429 #define MAX_SHARES (1UL << 18)
432 typedef int (*tg_visitor)(struct task_group *, void *);
434 extern int walk_tg_tree_from(struct task_group *from,
435 tg_visitor down, tg_visitor up, void *data);
438 * Iterate the full tree, calling @down when first entering a node and @up when
439 * leaving it for the final time.
441 * Caller must hold rcu_lock or sufficient equivalent.
443 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
445 return walk_tg_tree_from(&root_task_group, down, up, data);
448 extern int tg_nop(struct task_group *tg, void *data);
450 extern void free_fair_sched_group(struct task_group *tg);
451 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
452 extern void online_fair_sched_group(struct task_group *tg);
453 extern void unregister_fair_sched_group(struct task_group *tg);
454 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
455 struct sched_entity *se, int cpu,
456 struct sched_entity *parent);
457 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
459 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
460 extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
461 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
463 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
464 struct sched_rt_entity *rt_se, int cpu,
465 struct sched_rt_entity *parent);
466 extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us);
467 extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us);
468 extern long sched_group_rt_runtime(struct task_group *tg);
469 extern long sched_group_rt_period(struct task_group *tg);
470 extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
472 extern struct task_group *sched_create_group(struct task_group *parent);
473 extern void sched_online_group(struct task_group *tg,
474 struct task_group *parent);
475 extern void sched_destroy_group(struct task_group *tg);
476 extern void sched_release_group(struct task_group *tg);
478 extern void sched_move_task(struct task_struct *tsk);
480 #ifdef CONFIG_FAIR_GROUP_SCHED
481 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
483 extern int sched_group_set_idle(struct task_group *tg, long idle);
486 extern void set_task_rq_fair(struct sched_entity *se,
487 struct cfs_rq *prev, struct cfs_rq *next);
488 #else /* !CONFIG_SMP */
489 static inline void set_task_rq_fair(struct sched_entity *se,
490 struct cfs_rq *prev, struct cfs_rq *next) { }
491 #endif /* CONFIG_SMP */
492 #endif /* CONFIG_FAIR_GROUP_SCHED */
494 #else /* CONFIG_CGROUP_SCHED */
496 struct cfs_bandwidth { };
498 #endif /* CONFIG_CGROUP_SCHED */
500 extern void unregister_rt_sched_group(struct task_group *tg);
501 extern void free_rt_sched_group(struct task_group *tg);
502 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
505 * u64_u32_load/u64_u32_store
507 * Use a copy of a u64 value to protect against data race. This is only
508 * applicable for 32-bits architectures.
511 # define u64_u32_load_copy(var, copy) var
512 # define u64_u32_store_copy(var, copy, val) (var = val)
514 # define u64_u32_load_copy(var, copy) \
516 u64 __val, __val_copy; \
520 * paired with u64_u32_store_copy(), ordering access \
525 } while (__val != __val_copy); \
528 # define u64_u32_store_copy(var, copy, val) \
530 typeof(val) __val = (val); \
533 * paired with u64_u32_load_copy(), ordering access to var and \
540 # define u64_u32_load(var) u64_u32_load_copy(var, var##_copy)
541 # define u64_u32_store(var, val) u64_u32_store_copy(var, var##_copy, val)
543 /* CFS-related fields in a runqueue */
545 struct load_weight load;
546 unsigned int nr_running;
547 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */
548 unsigned int idle_nr_running; /* SCHED_IDLE */
549 unsigned int idle_h_nr_running; /* SCHED_IDLE */
553 #ifdef CONFIG_SCHED_CORE
554 unsigned int forceidle_seq;
559 u64 min_vruntime_copy;
562 struct rb_root_cached tasks_timeline;
565 * 'curr' points to currently running entity on this cfs_rq.
566 * It is set to NULL otherwise (i.e when none are currently running).
568 struct sched_entity *curr;
569 struct sched_entity *next;
570 struct sched_entity *last;
571 struct sched_entity *skip;
573 #ifdef CONFIG_SCHED_DEBUG
574 unsigned int nr_spread_over;
581 struct sched_avg avg;
583 u64 last_update_time_copy;
586 raw_spinlock_t lock ____cacheline_aligned;
588 unsigned long load_avg;
589 unsigned long util_avg;
590 unsigned long runnable_avg;
593 #ifdef CONFIG_FAIR_GROUP_SCHED
594 unsigned long tg_load_avg_contrib;
596 long prop_runnable_sum;
599 * h_load = weight * f(tg)
601 * Where f(tg) is the recursive weight fraction assigned to
604 unsigned long h_load;
605 u64 last_h_load_update;
606 struct sched_entity *h_load_next;
607 #endif /* CONFIG_FAIR_GROUP_SCHED */
608 #endif /* CONFIG_SMP */
610 #ifdef CONFIG_FAIR_GROUP_SCHED
611 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */
614 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
615 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
616 * (like users, containers etc.)
618 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU.
619 * This list is used during load balance.
622 struct list_head leaf_cfs_rq_list;
623 struct task_group *tg; /* group that "owns" this runqueue */
625 /* Locally cached copy of our task_group's idle value */
628 #ifdef CONFIG_CFS_BANDWIDTH
630 s64 runtime_remaining;
632 u64 throttled_pelt_idle;
634 u64 throttled_pelt_idle_copy;
637 u64 throttled_clock_pelt;
638 u64 throttled_clock_pelt_time;
641 struct list_head throttled_list;
643 struct list_head throttled_csd_list;
645 #endif /* CONFIG_CFS_BANDWIDTH */
646 #endif /* CONFIG_FAIR_GROUP_SCHED */
649 static inline int rt_bandwidth_enabled(void)
651 return sysctl_sched_rt_runtime >= 0;
654 /* RT IPI pull logic requires IRQ_WORK */
655 #if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP)
656 # define HAVE_RT_PUSH_IPI
659 /* Real-Time classes' related field in a runqueue: */
661 struct rt_prio_array active;
662 unsigned int rt_nr_running;
663 unsigned int rr_nr_running;
664 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
666 int curr; /* highest queued rt task prio */
668 int next; /* next highest */
673 unsigned int rt_nr_migratory;
674 unsigned int rt_nr_total;
676 struct plist_head pushable_tasks;
678 #endif /* CONFIG_SMP */
684 /* Nests inside the rq lock: */
685 raw_spinlock_t rt_runtime_lock;
687 #ifdef CONFIG_RT_GROUP_SCHED
688 unsigned int rt_nr_boosted;
691 struct task_group *tg;
695 static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq)
697 return rt_rq->rt_queued && rt_rq->rt_nr_running;
700 /* Deadline class' related fields in a runqueue */
702 /* runqueue is an rbtree, ordered by deadline */
703 struct rb_root_cached root;
705 unsigned int dl_nr_running;
709 * Deadline values of the currently executing and the
710 * earliest ready task on this rq. Caching these facilitates
711 * the decision whether or not a ready but not running task
712 * should migrate somewhere else.
719 unsigned int dl_nr_migratory;
723 * Tasks on this rq that can be pushed away. They are kept in
724 * an rb-tree, ordered by tasks' deadlines, with caching
725 * of the leftmost (earliest deadline) element.
727 struct rb_root_cached pushable_dl_tasks_root;
732 * "Active utilization" for this runqueue: increased when a
733 * task wakes up (becomes TASK_RUNNING) and decreased when a
739 * Utilization of the tasks "assigned" to this runqueue (including
740 * the tasks that are in runqueue and the tasks that executed on this
741 * CPU and blocked). Increased when a task moves to this runqueue, and
742 * decreased when the task moves away (migrates, changes scheduling
743 * policy, or terminates).
744 * This is needed to compute the "inactive utilization" for the
745 * runqueue (inactive utilization = this_bw - running_bw).
751 * Maximum available bandwidth for reclaiming by SCHED_FLAG_RECLAIM
752 * tasks of this rq. Used in calculation of reclaimable bandwidth(GRUB).
757 * Inverse of the fraction of CPU utilization that can be reclaimed
758 * by the GRUB algorithm.
763 #ifdef CONFIG_FAIR_GROUP_SCHED
764 /* An entity is a task if it doesn't "own" a runqueue */
765 #define entity_is_task(se) (!se->my_q)
767 static inline void se_update_runnable(struct sched_entity *se)
769 if (!entity_is_task(se))
770 se->runnable_weight = se->my_q->h_nr_running;
773 static inline long se_runnable(struct sched_entity *se)
775 if (entity_is_task(se))
778 return se->runnable_weight;
782 #define entity_is_task(se) 1
784 static inline void se_update_runnable(struct sched_entity *se) {}
786 static inline long se_runnable(struct sched_entity *se)
794 * XXX we want to get rid of these helpers and use the full load resolution.
796 static inline long se_weight(struct sched_entity *se)
798 return scale_load_down(se->load.weight);
802 static inline bool sched_asym_prefer(int a, int b)
804 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b);
808 struct em_perf_domain *em_pd;
809 struct perf_domain *next;
813 /* Scheduling group status flags */
814 #define SG_OVERLOAD 0x1 /* More than one runnable task on a CPU. */
815 #define SG_OVERUTILIZED 0x2 /* One or more CPUs are over-utilized. */
818 * We add the notion of a root-domain which will be used to define per-domain
819 * variables. Each exclusive cpuset essentially defines an island domain by
820 * fully partitioning the member CPUs from any other cpuset. Whenever a new
821 * exclusive cpuset is created, we also create and attach a new root-domain
830 cpumask_var_t online;
833 * Indicate pullable load on at least one CPU, e.g:
834 * - More than one runnable task
835 * - Running task is misfit
839 /* Indicate one or more cpus over-utilized (tipping point) */
843 * The bit corresponding to a CPU gets set here if such CPU has more
844 * than one runnable -deadline task (as it is below for RT tasks).
846 cpumask_var_t dlo_mask;
852 * Indicate whether a root_domain's dl_bw has been checked or
853 * updated. It's monotonously increasing value.
855 * Also, some corner cases, like 'wrap around' is dangerous, but given
856 * that u64 is 'big enough'. So that shouldn't be a concern.
860 #ifdef HAVE_RT_PUSH_IPI
862 * For IPI pull requests, loop across the rto_mask.
864 struct irq_work rto_push_work;
865 raw_spinlock_t rto_lock;
866 /* These are only updated and read within rto_lock */
869 /* These atomics are updated outside of a lock */
870 atomic_t rto_loop_next;
871 atomic_t rto_loop_start;
874 * The "RT overload" flag: it gets set if a CPU has more than
875 * one runnable RT task.
877 cpumask_var_t rto_mask;
878 struct cpupri cpupri;
880 unsigned long max_cpu_capacity;
883 * NULL-terminated list of performance domains intersecting with the
884 * CPUs of the rd. Protected by RCU.
886 struct perf_domain __rcu *pd;
889 extern void init_defrootdomain(void);
890 extern int sched_init_domains(const struct cpumask *cpu_map);
891 extern void rq_attach_root(struct rq *rq, struct root_domain *rd);
892 extern void sched_get_rd(struct root_domain *rd);
893 extern void sched_put_rd(struct root_domain *rd);
895 #ifdef HAVE_RT_PUSH_IPI
896 extern void rto_push_irq_work_func(struct irq_work *work);
898 #endif /* CONFIG_SMP */
900 #ifdef CONFIG_UCLAMP_TASK
902 * struct uclamp_bucket - Utilization clamp bucket
903 * @value: utilization clamp value for tasks on this clamp bucket
904 * @tasks: number of RUNNABLE tasks on this clamp bucket
906 * Keep track of how many tasks are RUNNABLE for a given utilization
909 struct uclamp_bucket {
910 unsigned long value : bits_per(SCHED_CAPACITY_SCALE);
911 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE);
915 * struct uclamp_rq - rq's utilization clamp
916 * @value: currently active clamp values for a rq
917 * @bucket: utilization clamp buckets affecting a rq
919 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values.
920 * A clamp value is affecting a rq when there is at least one task RUNNABLE
921 * (or actually running) with that value.
923 * There are up to UCLAMP_CNT possible different clamp values, currently there
924 * are only two: minimum utilization and maximum utilization.
926 * All utilization clamping values are MAX aggregated, since:
927 * - for util_min: we want to run the CPU at least at the max of the minimum
928 * utilization required by its currently RUNNABLE tasks.
929 * - for util_max: we want to allow the CPU to run up to the max of the
930 * maximum utilization allowed by its currently RUNNABLE tasks.
932 * Since on each system we expect only a limited number of different
933 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track
934 * the metrics required to compute all the per-rq utilization clamp values.
938 struct uclamp_bucket bucket[UCLAMP_BUCKETS];
941 DECLARE_STATIC_KEY_FALSE(sched_uclamp_used);
942 #endif /* CONFIG_UCLAMP_TASK */
945 struct balance_callback {
946 struct balance_callback *next;
947 void (*func)(struct rq *rq);
951 * This is the main, per-CPU runqueue data structure.
953 * Locking rule: those places that want to lock multiple runqueues
954 * (such as the load balancing or the thread migration code), lock
955 * acquire operations must be ordered by ascending &runqueue.
959 raw_spinlock_t __lock;
962 * nr_running and cpu_load should be in the same cacheline because
963 * remote CPUs use both these fields when doing load calculation.
965 unsigned int nr_running;
966 #ifdef CONFIG_NUMA_BALANCING
967 unsigned int nr_numa_running;
968 unsigned int nr_preferred_running;
969 unsigned int numa_migrate_on;
971 #ifdef CONFIG_NO_HZ_COMMON
973 unsigned long last_blocked_load_update_tick;
974 unsigned int has_blocked_load;
975 call_single_data_t nohz_csd;
976 #endif /* CONFIG_SMP */
977 unsigned int nohz_tick_stopped;
979 #endif /* CONFIG_NO_HZ_COMMON */
982 unsigned int ttwu_pending;
986 #ifdef CONFIG_UCLAMP_TASK
987 /* Utilization clamp values based on CPU's RUNNABLE tasks */
988 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned;
989 unsigned int uclamp_flags;
990 #define UCLAMP_FLAG_IDLE 0x01
997 #ifdef CONFIG_FAIR_GROUP_SCHED
998 /* list of leaf cfs_rq on this CPU: */
999 struct list_head leaf_cfs_rq_list;
1000 struct list_head *tmp_alone_branch;
1001 #endif /* CONFIG_FAIR_GROUP_SCHED */
1004 * This is part of a global counter where only the total sum
1005 * over all CPUs matters. A task can increase this counter on
1006 * one CPU and if it got migrated afterwards it may decrease
1007 * it on another CPU. Always updated under the runqueue lock:
1009 unsigned int nr_uninterruptible;
1011 struct task_struct __rcu *curr;
1012 struct task_struct *idle;
1013 struct task_struct *stop;
1014 unsigned long next_balance;
1015 struct mm_struct *prev_mm;
1017 unsigned int clock_update_flags;
1019 /* Ensure that all clocks are in the same cache line */
1020 u64 clock_task ____cacheline_aligned;
1022 unsigned long lost_idle_time;
1023 u64 clock_pelt_idle;
1025 #ifndef CONFIG_64BIT
1026 u64 clock_pelt_idle_copy;
1027 u64 clock_idle_copy;
1032 #ifdef CONFIG_SCHED_DEBUG
1033 u64 last_seen_need_resched_ns;
1034 int ticks_without_resched;
1037 #ifdef CONFIG_MEMBARRIER
1038 int membarrier_state;
1042 struct root_domain *rd;
1043 struct sched_domain __rcu *sd;
1045 unsigned long cpu_capacity;
1046 unsigned long cpu_capacity_orig;
1048 struct balance_callback *balance_callback;
1050 unsigned char nohz_idle_balance;
1051 unsigned char idle_balance;
1053 unsigned long misfit_task_load;
1055 /* For active balancing */
1058 struct cpu_stop_work active_balance_work;
1060 /* CPU of this runqueue: */
1064 struct list_head cfs_tasks;
1066 struct sched_avg avg_rt;
1067 struct sched_avg avg_dl;
1068 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
1069 struct sched_avg avg_irq;
1071 #ifdef CONFIG_SCHED_THERMAL_PRESSURE
1072 struct sched_avg avg_thermal;
1077 unsigned long wake_stamp;
1080 /* This is used to determine avg_idle's max value */
1081 u64 max_idle_balance_cost;
1083 #ifdef CONFIG_HOTPLUG_CPU
1084 struct rcuwait hotplug_wait;
1086 #endif /* CONFIG_SMP */
1088 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1091 #ifdef CONFIG_PARAVIRT
1092 u64 prev_steal_time;
1094 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1095 u64 prev_steal_time_rq;
1098 /* calc_load related fields */
1099 unsigned long calc_load_update;
1100 long calc_load_active;
1102 #ifdef CONFIG_SCHED_HRTICK
1104 call_single_data_t hrtick_csd;
1106 struct hrtimer hrtick_timer;
1107 ktime_t hrtick_time;
1110 #ifdef CONFIG_SCHEDSTATS
1112 struct sched_info rq_sched_info;
1113 unsigned long long rq_cpu_time;
1114 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
1116 /* sys_sched_yield() stats */
1117 unsigned int yld_count;
1119 /* schedule() stats */
1120 unsigned int sched_count;
1121 unsigned int sched_goidle;
1123 /* try_to_wake_up() stats */
1124 unsigned int ttwu_count;
1125 unsigned int ttwu_local;
1128 #ifdef CONFIG_CPU_IDLE
1129 /* Must be inspected within a rcu lock section */
1130 struct cpuidle_state *idle_state;
1134 unsigned int nr_pinned;
1136 unsigned int push_busy;
1137 struct cpu_stop_work push_work;
1139 #ifdef CONFIG_SCHED_CORE
1142 struct task_struct *core_pick;
1143 unsigned int core_enabled;
1144 unsigned int core_sched_seq;
1145 struct rb_root core_tree;
1147 /* shared state -- careful with sched_core_cpu_deactivate() */
1148 unsigned int core_task_seq;
1149 unsigned int core_pick_seq;
1150 unsigned long core_cookie;
1151 unsigned int core_forceidle_count;
1152 unsigned int core_forceidle_seq;
1153 unsigned int core_forceidle_occupation;
1154 u64 core_forceidle_start;
1157 /* Scratch cpumask to be temporarily used under rq_lock */
1158 cpumask_var_t scratch_mask;
1160 #if defined(CONFIG_CFS_BANDWIDTH) && defined(CONFIG_SMP)
1161 call_single_data_t cfsb_csd;
1162 struct list_head cfsb_csd_list;
1166 #ifdef CONFIG_FAIR_GROUP_SCHED
1168 /* CPU runqueue to which this cfs_rq is attached */
1169 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1176 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1178 return container_of(cfs_rq, struct rq, cfs);
1182 static inline int cpu_of(struct rq *rq)
1191 #define MDF_PUSH 0x01
1193 static inline bool is_migration_disabled(struct task_struct *p)
1196 return p->migration_disabled;
1202 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1204 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
1205 #define this_rq() this_cpu_ptr(&runqueues)
1206 #define task_rq(p) cpu_rq(task_cpu(p))
1207 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
1208 #define raw_rq() raw_cpu_ptr(&runqueues)
1211 #ifdef CONFIG_SCHED_CORE
1212 static inline struct cpumask *sched_group_span(struct sched_group *sg);
1214 DECLARE_STATIC_KEY_FALSE(__sched_core_enabled);
1216 static inline bool sched_core_enabled(struct rq *rq)
1218 return static_branch_unlikely(&__sched_core_enabled) && rq->core_enabled;
1221 static inline bool sched_core_disabled(void)
1223 return !static_branch_unlikely(&__sched_core_enabled);
1227 * Be careful with this function; not for general use. The return value isn't
1228 * stable unless you actually hold a relevant rq->__lock.
1230 static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1232 if (sched_core_enabled(rq))
1233 return &rq->core->__lock;
1238 static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1240 if (rq->core_enabled)
1241 return &rq->core->__lock;
1246 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
1250 * Helpers to check if the CPU's core cookie matches with the task's cookie
1251 * when core scheduling is enabled.
1252 * A special case is that the task's cookie always matches with CPU's core
1253 * cookie if the CPU is in an idle core.
1255 static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1257 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1258 if (!sched_core_enabled(rq))
1261 return rq->core->core_cookie == p->core_cookie;
1264 static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1266 bool idle_core = true;
1269 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1270 if (!sched_core_enabled(rq))
1273 for_each_cpu(cpu, cpu_smt_mask(cpu_of(rq))) {
1274 if (!available_idle_cpu(cpu)) {
1281 * A CPU in an idle core is always the best choice for tasks with
1284 return idle_core || rq->core->core_cookie == p->core_cookie;
1287 static inline bool sched_group_cookie_match(struct rq *rq,
1288 struct task_struct *p,
1289 struct sched_group *group)
1293 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1294 if (!sched_core_enabled(rq))
1297 for_each_cpu_and(cpu, sched_group_span(group), p->cpus_ptr) {
1298 if (sched_core_cookie_match(cpu_rq(cpu), p))
1304 static inline bool sched_core_enqueued(struct task_struct *p)
1306 return !RB_EMPTY_NODE(&p->core_node);
1309 extern void sched_core_enqueue(struct rq *rq, struct task_struct *p);
1310 extern void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags);
1312 extern void sched_core_get(void);
1313 extern void sched_core_put(void);
1315 #else /* !CONFIG_SCHED_CORE */
1317 static inline bool sched_core_enabled(struct rq *rq)
1322 static inline bool sched_core_disabled(void)
1327 static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1332 static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1337 static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1342 static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1347 static inline bool sched_group_cookie_match(struct rq *rq,
1348 struct task_struct *p,
1349 struct sched_group *group)
1353 #endif /* CONFIG_SCHED_CORE */
1355 static inline void lockdep_assert_rq_held(struct rq *rq)
1357 lockdep_assert_held(__rq_lockp(rq));
1360 extern void raw_spin_rq_lock_nested(struct rq *rq, int subclass);
1361 extern bool raw_spin_rq_trylock(struct rq *rq);
1362 extern void raw_spin_rq_unlock(struct rq *rq);
1364 static inline void raw_spin_rq_lock(struct rq *rq)
1366 raw_spin_rq_lock_nested(rq, 0);
1369 static inline void raw_spin_rq_lock_irq(struct rq *rq)
1371 local_irq_disable();
1372 raw_spin_rq_lock(rq);
1375 static inline void raw_spin_rq_unlock_irq(struct rq *rq)
1377 raw_spin_rq_unlock(rq);
1381 static inline unsigned long _raw_spin_rq_lock_irqsave(struct rq *rq)
1383 unsigned long flags;
1384 local_irq_save(flags);
1385 raw_spin_rq_lock(rq);
1389 static inline void raw_spin_rq_unlock_irqrestore(struct rq *rq, unsigned long flags)
1391 raw_spin_rq_unlock(rq);
1392 local_irq_restore(flags);
1395 #define raw_spin_rq_lock_irqsave(rq, flags) \
1397 flags = _raw_spin_rq_lock_irqsave(rq); \
1400 #ifdef CONFIG_SCHED_SMT
1401 extern void __update_idle_core(struct rq *rq);
1403 static inline void update_idle_core(struct rq *rq)
1405 if (static_branch_unlikely(&sched_smt_present))
1406 __update_idle_core(rq);
1410 static inline void update_idle_core(struct rq *rq) { }
1413 #ifdef CONFIG_FAIR_GROUP_SCHED
1414 static inline struct task_struct *task_of(struct sched_entity *se)
1416 SCHED_WARN_ON(!entity_is_task(se));
1417 return container_of(se, struct task_struct, se);
1420 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
1422 return p->se.cfs_rq;
1425 /* runqueue on which this entity is (to be) queued */
1426 static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1431 /* runqueue "owned" by this group */
1432 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1439 #define task_of(_se) container_of(_se, struct task_struct, se)
1441 static inline struct cfs_rq *task_cfs_rq(const struct task_struct *p)
1443 return &task_rq(p)->cfs;
1446 static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1448 const struct task_struct *p = task_of(se);
1449 struct rq *rq = task_rq(p);
1454 /* runqueue "owned" by this group */
1455 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1461 extern void update_rq_clock(struct rq *rq);
1464 * rq::clock_update_flags bits
1466 * %RQCF_REQ_SKIP - will request skipping of clock update on the next
1467 * call to __schedule(). This is an optimisation to avoid
1468 * neighbouring rq clock updates.
1470 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is
1471 * in effect and calls to update_rq_clock() are being ignored.
1473 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been
1474 * made to update_rq_clock() since the last time rq::lock was pinned.
1476 * If inside of __schedule(), clock_update_flags will have been
1477 * shifted left (a left shift is a cheap operation for the fast path
1478 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use,
1480 * if (rq-clock_update_flags >= RQCF_UPDATED)
1482 * to check if %RQCF_UPDATED is set. It'll never be shifted more than
1483 * one position though, because the next rq_unpin_lock() will shift it
1486 #define RQCF_REQ_SKIP 0x01
1487 #define RQCF_ACT_SKIP 0x02
1488 #define RQCF_UPDATED 0x04
1490 static inline void assert_clock_updated(struct rq *rq)
1493 * The only reason for not seeing a clock update since the
1494 * last rq_pin_lock() is if we're currently skipping updates.
1496 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
1499 static inline u64 rq_clock(struct rq *rq)
1501 lockdep_assert_rq_held(rq);
1502 assert_clock_updated(rq);
1507 static inline u64 rq_clock_task(struct rq *rq)
1509 lockdep_assert_rq_held(rq);
1510 assert_clock_updated(rq);
1512 return rq->clock_task;
1516 * By default the decay is the default pelt decay period.
1517 * The decay shift can change the decay period in
1519 * Decay shift Decay period(ms)
1526 extern int sched_thermal_decay_shift;
1528 static inline u64 rq_clock_thermal(struct rq *rq)
1530 return rq_clock_task(rq) >> sched_thermal_decay_shift;
1533 static inline void rq_clock_skip_update(struct rq *rq)
1535 lockdep_assert_rq_held(rq);
1536 rq->clock_update_flags |= RQCF_REQ_SKIP;
1540 * See rt task throttling, which is the only time a skip
1541 * request is canceled.
1543 static inline void rq_clock_cancel_skipupdate(struct rq *rq)
1545 lockdep_assert_rq_held(rq);
1546 rq->clock_update_flags &= ~RQCF_REQ_SKIP;
1550 * During cpu offlining and rq wide unthrottling, we can trigger
1551 * an update_rq_clock() for several cfs and rt runqueues (Typically
1552 * when using list_for_each_entry_*)
1553 * rq_clock_start_loop_update() can be called after updating the clock
1554 * once and before iterating over the list to prevent multiple update.
1555 * After the iterative traversal, we need to call rq_clock_stop_loop_update()
1556 * to clear RQCF_ACT_SKIP of rq->clock_update_flags.
1558 static inline void rq_clock_start_loop_update(struct rq *rq)
1560 lockdep_assert_rq_held(rq);
1561 SCHED_WARN_ON(rq->clock_update_flags & RQCF_ACT_SKIP);
1562 rq->clock_update_flags |= RQCF_ACT_SKIP;
1565 static inline void rq_clock_stop_loop_update(struct rq *rq)
1567 lockdep_assert_rq_held(rq);
1568 rq->clock_update_flags &= ~RQCF_ACT_SKIP;
1572 unsigned long flags;
1573 struct pin_cookie cookie;
1574 #ifdef CONFIG_SCHED_DEBUG
1576 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the
1577 * current pin context is stashed here in case it needs to be
1578 * restored in rq_repin_lock().
1580 unsigned int clock_update_flags;
1584 extern struct balance_callback balance_push_callback;
1587 * Lockdep annotation that avoids accidental unlocks; it's like a
1588 * sticky/continuous lockdep_assert_held().
1590 * This avoids code that has access to 'struct rq *rq' (basically everything in
1591 * the scheduler) from accidentally unlocking the rq if they do not also have a
1592 * copy of the (on-stack) 'struct rq_flags rf'.
1594 * Also see Documentation/locking/lockdep-design.rst.
1596 static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
1598 rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
1600 #ifdef CONFIG_SCHED_DEBUG
1601 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
1602 rf->clock_update_flags = 0;
1604 SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
1609 static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
1611 #ifdef CONFIG_SCHED_DEBUG
1612 if (rq->clock_update_flags > RQCF_ACT_SKIP)
1613 rf->clock_update_flags = RQCF_UPDATED;
1616 lockdep_unpin_lock(__rq_lockp(rq), rf->cookie);
1619 static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf)
1621 lockdep_repin_lock(__rq_lockp(rq), rf->cookie);
1623 #ifdef CONFIG_SCHED_DEBUG
1625 * Restore the value we stashed in @rf for this pin context.
1627 rq->clock_update_flags |= rf->clock_update_flags;
1631 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1632 __acquires(rq->lock);
1634 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1635 __acquires(p->pi_lock)
1636 __acquires(rq->lock);
1638 static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf)
1639 __releases(rq->lock)
1641 rq_unpin_lock(rq, rf);
1642 raw_spin_rq_unlock(rq);
1646 task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1647 __releases(rq->lock)
1648 __releases(p->pi_lock)
1650 rq_unpin_lock(rq, rf);
1651 raw_spin_rq_unlock(rq);
1652 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
1656 rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
1657 __acquires(rq->lock)
1659 raw_spin_rq_lock_irqsave(rq, rf->flags);
1660 rq_pin_lock(rq, rf);
1664 rq_lock_irq(struct rq *rq, struct rq_flags *rf)
1665 __acquires(rq->lock)
1667 raw_spin_rq_lock_irq(rq);
1668 rq_pin_lock(rq, rf);
1672 rq_lock(struct rq *rq, struct rq_flags *rf)
1673 __acquires(rq->lock)
1675 raw_spin_rq_lock(rq);
1676 rq_pin_lock(rq, rf);
1680 rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
1681 __releases(rq->lock)
1683 rq_unpin_lock(rq, rf);
1684 raw_spin_rq_unlock_irqrestore(rq, rf->flags);
1688 rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
1689 __releases(rq->lock)
1691 rq_unpin_lock(rq, rf);
1692 raw_spin_rq_unlock_irq(rq);
1696 rq_unlock(struct rq *rq, struct rq_flags *rf)
1697 __releases(rq->lock)
1699 rq_unpin_lock(rq, rf);
1700 raw_spin_rq_unlock(rq);
1703 static inline struct rq *
1704 this_rq_lock_irq(struct rq_flags *rf)
1705 __acquires(rq->lock)
1709 local_irq_disable();
1716 enum numa_topology_type {
1721 extern enum numa_topology_type sched_numa_topology_type;
1722 extern int sched_max_numa_distance;
1723 extern bool find_numa_distance(int distance);
1724 extern void sched_init_numa(int offline_node);
1725 extern void sched_update_numa(int cpu, bool online);
1726 extern void sched_domains_numa_masks_set(unsigned int cpu);
1727 extern void sched_domains_numa_masks_clear(unsigned int cpu);
1728 extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu);
1730 static inline void sched_init_numa(int offline_node) { }
1731 static inline void sched_update_numa(int cpu, bool online) { }
1732 static inline void sched_domains_numa_masks_set(unsigned int cpu) { }
1733 static inline void sched_domains_numa_masks_clear(unsigned int cpu) { }
1734 static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1740 #ifdef CONFIG_NUMA_BALANCING
1741 /* The regions in numa_faults array from task_struct */
1742 enum numa_faults_stats {
1748 extern void sched_setnuma(struct task_struct *p, int node);
1749 extern int migrate_task_to(struct task_struct *p, int cpu);
1750 extern int migrate_swap(struct task_struct *p, struct task_struct *t,
1752 extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p);
1755 init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1758 #endif /* CONFIG_NUMA_BALANCING */
1763 queue_balance_callback(struct rq *rq,
1764 struct balance_callback *head,
1765 void (*func)(struct rq *rq))
1767 lockdep_assert_rq_held(rq);
1770 * Don't (re)queue an already queued item; nor queue anything when
1771 * balance_push() is active, see the comment with
1772 * balance_push_callback.
1774 if (unlikely(head->next || rq->balance_callback == &balance_push_callback))
1778 head->next = rq->balance_callback;
1779 rq->balance_callback = head;
1782 #define rcu_dereference_check_sched_domain(p) \
1783 rcu_dereference_check((p), \
1784 lockdep_is_held(&sched_domains_mutex))
1787 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1788 * See destroy_sched_domains: call_rcu for details.
1790 * The domain tree of any CPU may only be accessed from within
1791 * preempt-disabled sections.
1793 #define for_each_domain(cpu, __sd) \
1794 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
1795 __sd; __sd = __sd->parent)
1797 /* A mask of all the SD flags that have the SDF_SHARED_CHILD metaflag */
1798 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_SHARED_CHILD)) |
1799 static const unsigned int SD_SHARED_CHILD_MASK =
1800 #include <linux/sched/sd_flags.h>
1805 * highest_flag_domain - Return highest sched_domain containing flag.
1806 * @cpu: The CPU whose highest level of sched domain is to
1808 * @flag: The flag to check for the highest sched_domain
1809 * for the given CPU.
1811 * Returns the highest sched_domain of a CPU which contains @flag. If @flag has
1812 * the SDF_SHARED_CHILD metaflag, all the children domains also have @flag.
1814 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
1816 struct sched_domain *sd, *hsd = NULL;
1818 for_each_domain(cpu, sd) {
1819 if (sd->flags & flag) {
1825 * Stop the search if @flag is known to be shared at lower
1826 * levels. It will not be found further up.
1828 if (flag & SD_SHARED_CHILD_MASK)
1835 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
1837 struct sched_domain *sd;
1839 for_each_domain(cpu, sd) {
1840 if (sd->flags & flag)
1847 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
1848 DECLARE_PER_CPU(int, sd_llc_size);
1849 DECLARE_PER_CPU(int, sd_llc_id);
1850 DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
1851 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
1852 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
1853 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
1854 extern struct static_key_false sched_asym_cpucapacity;
1856 static __always_inline bool sched_asym_cpucap_active(void)
1858 return static_branch_unlikely(&sched_asym_cpucapacity);
1861 struct sched_group_capacity {
1864 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity
1867 unsigned long capacity;
1868 unsigned long min_capacity; /* Min per-CPU capacity in group */
1869 unsigned long max_capacity; /* Max per-CPU capacity in group */
1870 unsigned long next_update;
1871 int imbalance; /* XXX unrelated to capacity but shared group state */
1873 #ifdef CONFIG_SCHED_DEBUG
1877 unsigned long cpumask[]; /* Balance mask */
1880 struct sched_group {
1881 struct sched_group *next; /* Must be a circular list */
1884 unsigned int group_weight;
1885 struct sched_group_capacity *sgc;
1886 int asym_prefer_cpu; /* CPU of highest priority in group */
1890 * The CPUs this group covers.
1892 * NOTE: this field is variable length. (Allocated dynamically
1893 * by attaching extra space to the end of the structure,
1894 * depending on how many CPUs the kernel has booted up with)
1896 unsigned long cpumask[];
1899 static inline struct cpumask *sched_group_span(struct sched_group *sg)
1901 return to_cpumask(sg->cpumask);
1905 * See build_balance_mask().
1907 static inline struct cpumask *group_balance_mask(struct sched_group *sg)
1909 return to_cpumask(sg->sgc->cpumask);
1912 extern int group_balance_cpu(struct sched_group *sg);
1914 #ifdef CONFIG_SCHED_DEBUG
1915 void update_sched_domain_debugfs(void);
1916 void dirty_sched_domain_sysctl(int cpu);
1918 static inline void update_sched_domain_debugfs(void)
1921 static inline void dirty_sched_domain_sysctl(int cpu)
1926 extern int sched_update_scaling(void);
1928 static inline const struct cpumask *task_user_cpus(struct task_struct *p)
1930 if (!p->user_cpus_ptr)
1931 return cpu_possible_mask; /* &init_task.cpus_mask */
1932 return p->user_cpus_ptr;
1934 #endif /* CONFIG_SMP */
1938 #if defined(CONFIG_SCHED_CORE) && defined(CONFIG_SCHEDSTATS)
1940 extern void __sched_core_account_forceidle(struct rq *rq);
1942 static inline void sched_core_account_forceidle(struct rq *rq)
1944 if (schedstat_enabled())
1945 __sched_core_account_forceidle(rq);
1948 extern void __sched_core_tick(struct rq *rq);
1950 static inline void sched_core_tick(struct rq *rq)
1952 if (sched_core_enabled(rq) && schedstat_enabled())
1953 __sched_core_tick(rq);
1958 static inline void sched_core_account_forceidle(struct rq *rq) {}
1960 static inline void sched_core_tick(struct rq *rq) {}
1962 #endif /* CONFIG_SCHED_CORE && CONFIG_SCHEDSTATS */
1964 #ifdef CONFIG_CGROUP_SCHED
1967 * Return the group to which this tasks belongs.
1969 * We cannot use task_css() and friends because the cgroup subsystem
1970 * changes that value before the cgroup_subsys::attach() method is called,
1971 * therefore we cannot pin it and might observe the wrong value.
1973 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
1974 * core changes this before calling sched_move_task().
1976 * Instead we use a 'copy' which is updated from sched_move_task() while
1977 * holding both task_struct::pi_lock and rq::lock.
1979 static inline struct task_group *task_group(struct task_struct *p)
1981 return p->sched_task_group;
1984 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
1985 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
1987 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
1988 struct task_group *tg = task_group(p);
1991 #ifdef CONFIG_FAIR_GROUP_SCHED
1992 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
1993 p->se.cfs_rq = tg->cfs_rq[cpu];
1994 p->se.parent = tg->se[cpu];
1995 p->se.depth = tg->se[cpu] ? tg->se[cpu]->depth + 1 : 0;
1998 #ifdef CONFIG_RT_GROUP_SCHED
1999 p->rt.rt_rq = tg->rt_rq[cpu];
2000 p->rt.parent = tg->rt_se[cpu];
2004 #else /* CONFIG_CGROUP_SCHED */
2006 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
2007 static inline struct task_group *task_group(struct task_struct *p)
2012 #endif /* CONFIG_CGROUP_SCHED */
2014 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
2016 set_task_rq(p, cpu);
2019 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2020 * successfully executed on another CPU. We must ensure that updates of
2021 * per-task data have been completed by this moment.
2024 WRITE_ONCE(task_thread_info(p)->cpu, cpu);
2030 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
2032 #ifdef CONFIG_SCHED_DEBUG
2033 # define const_debug __read_mostly
2035 # define const_debug const
2038 #define SCHED_FEAT(name, enabled) \
2039 __SCHED_FEAT_##name ,
2042 #include "features.h"
2048 #ifdef CONFIG_SCHED_DEBUG
2051 * To support run-time toggling of sched features, all the translation units
2052 * (but core.c) reference the sysctl_sched_features defined in core.c.
2054 extern const_debug unsigned int sysctl_sched_features;
2056 #ifdef CONFIG_JUMP_LABEL
2057 #define SCHED_FEAT(name, enabled) \
2058 static __always_inline bool static_branch_##name(struct static_key *key) \
2060 return static_key_##enabled(key); \
2063 #include "features.h"
2066 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
2067 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
2069 #else /* !CONFIG_JUMP_LABEL */
2071 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2073 #endif /* CONFIG_JUMP_LABEL */
2075 #else /* !SCHED_DEBUG */
2078 * Each translation unit has its own copy of sysctl_sched_features to allow
2079 * constants propagation at compile time and compiler optimization based on
2082 #define SCHED_FEAT(name, enabled) \
2083 (1UL << __SCHED_FEAT_##name) * enabled |
2084 static const_debug __maybe_unused unsigned int sysctl_sched_features =
2085 #include "features.h"
2089 #define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2091 #endif /* SCHED_DEBUG */
2093 extern struct static_key_false sched_numa_balancing;
2094 extern struct static_key_false sched_schedstats;
2096 static inline u64 global_rt_period(void)
2098 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
2101 static inline u64 global_rt_runtime(void)
2103 if (sysctl_sched_rt_runtime < 0)
2106 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
2109 static inline int task_current(struct rq *rq, struct task_struct *p)
2111 return rq->curr == p;
2114 static inline int task_on_cpu(struct rq *rq, struct task_struct *p)
2119 return task_current(rq, p);
2123 static inline int task_on_rq_queued(struct task_struct *p)
2125 return p->on_rq == TASK_ON_RQ_QUEUED;
2128 static inline int task_on_rq_migrating(struct task_struct *p)
2130 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING;
2133 /* Wake flags. The first three directly map to some SD flag value */
2134 #define WF_EXEC 0x02 /* Wakeup after exec; maps to SD_BALANCE_EXEC */
2135 #define WF_FORK 0x04 /* Wakeup after fork; maps to SD_BALANCE_FORK */
2136 #define WF_TTWU 0x08 /* Wakeup; maps to SD_BALANCE_WAKE */
2138 #define WF_SYNC 0x10 /* Waker goes to sleep after wakeup */
2139 #define WF_MIGRATED 0x20 /* Internal use, task got migrated */
2140 #define WF_CURRENT_CPU 0x40 /* Prefer to move the wakee to the current CPU. */
2143 static_assert(WF_EXEC == SD_BALANCE_EXEC);
2144 static_assert(WF_FORK == SD_BALANCE_FORK);
2145 static_assert(WF_TTWU == SD_BALANCE_WAKE);
2149 * To aid in avoiding the subversion of "niceness" due to uneven distribution
2150 * of tasks with abnormal "nice" values across CPUs the contribution that
2151 * each task makes to its run queue's load is weighted according to its
2152 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
2153 * scaled version of the new time slice allocation that they receive on time
2157 #define WEIGHT_IDLEPRIO 3
2158 #define WMULT_IDLEPRIO 1431655765
2160 extern const int sched_prio_to_weight[40];
2161 extern const u32 sched_prio_to_wmult[40];
2164 * {de,en}queue flags:
2166 * DEQUEUE_SLEEP - task is no longer runnable
2167 * ENQUEUE_WAKEUP - task just became runnable
2169 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
2170 * are in a known state which allows modification. Such pairs
2171 * should preserve as much state as possible.
2173 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
2176 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
2177 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
2178 * ENQUEUE_MIGRATED - the task was migrated during wakeup
2182 #define DEQUEUE_SLEEP 0x01
2183 #define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */
2184 #define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */
2185 #define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */
2187 #define ENQUEUE_WAKEUP 0x01
2188 #define ENQUEUE_RESTORE 0x02
2189 #define ENQUEUE_MOVE 0x04
2190 #define ENQUEUE_NOCLOCK 0x08
2192 #define ENQUEUE_HEAD 0x10
2193 #define ENQUEUE_REPLENISH 0x20
2195 #define ENQUEUE_MIGRATED 0x40
2197 #define ENQUEUE_MIGRATED 0x00
2200 #define RETRY_TASK ((void *)-1UL)
2202 struct affinity_context {
2203 const struct cpumask *new_mask;
2204 struct cpumask *user_mask;
2208 struct sched_class {
2210 #ifdef CONFIG_UCLAMP_TASK
2214 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
2215 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
2216 void (*yield_task) (struct rq *rq);
2217 bool (*yield_to_task)(struct rq *rq, struct task_struct *p);
2219 void (*check_preempt_curr)(struct rq *rq, struct task_struct *p, int flags);
2221 struct task_struct *(*pick_next_task)(struct rq *rq);
2223 void (*put_prev_task)(struct rq *rq, struct task_struct *p);
2224 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first);
2227 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2228 int (*select_task_rq)(struct task_struct *p, int task_cpu, int flags);
2230 struct task_struct * (*pick_task)(struct rq *rq);
2232 void (*migrate_task_rq)(struct task_struct *p, int new_cpu);
2234 void (*task_woken)(struct rq *this_rq, struct task_struct *task);
2236 void (*set_cpus_allowed)(struct task_struct *p, struct affinity_context *ctx);
2238 void (*rq_online)(struct rq *rq);
2239 void (*rq_offline)(struct rq *rq);
2241 struct rq *(*find_lock_rq)(struct task_struct *p, struct rq *rq);
2244 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued);
2245 void (*task_fork)(struct task_struct *p);
2246 void (*task_dead)(struct task_struct *p);
2249 * The switched_from() call is allowed to drop rq->lock, therefore we
2250 * cannot assume the switched_from/switched_to pair is serialized by
2251 * rq->lock. They are however serialized by p->pi_lock.
2253 void (*switched_from)(struct rq *this_rq, struct task_struct *task);
2254 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
2255 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
2258 unsigned int (*get_rr_interval)(struct rq *rq,
2259 struct task_struct *task);
2261 void (*update_curr)(struct rq *rq);
2263 #ifdef CONFIG_FAIR_GROUP_SCHED
2264 void (*task_change_group)(struct task_struct *p);
2267 #ifdef CONFIG_SCHED_CORE
2268 int (*task_is_throttled)(struct task_struct *p, int cpu);
2272 static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
2274 WARN_ON_ONCE(rq->curr != prev);
2275 prev->sched_class->put_prev_task(rq, prev);
2278 static inline void set_next_task(struct rq *rq, struct task_struct *next)
2280 next->sched_class->set_next_task(rq, next, false);
2285 * Helper to define a sched_class instance; each one is placed in a separate
2286 * section which is ordered by the linker script:
2288 * include/asm-generic/vmlinux.lds.h
2290 * *CAREFUL* they are laid out in *REVERSE* order!!!
2292 * Also enforce alignment on the instance, not the type, to guarantee layout.
2294 #define DEFINE_SCHED_CLASS(name) \
2295 const struct sched_class name##_sched_class \
2296 __aligned(__alignof__(struct sched_class)) \
2297 __section("__" #name "_sched_class")
2299 /* Defined in include/asm-generic/vmlinux.lds.h */
2300 extern struct sched_class __sched_class_highest[];
2301 extern struct sched_class __sched_class_lowest[];
2303 #define for_class_range(class, _from, _to) \
2304 for (class = (_from); class < (_to); class++)
2306 #define for_each_class(class) \
2307 for_class_range(class, __sched_class_highest, __sched_class_lowest)
2309 #define sched_class_above(_a, _b) ((_a) < (_b))
2311 extern const struct sched_class stop_sched_class;
2312 extern const struct sched_class dl_sched_class;
2313 extern const struct sched_class rt_sched_class;
2314 extern const struct sched_class fair_sched_class;
2315 extern const struct sched_class idle_sched_class;
2317 static inline bool sched_stop_runnable(struct rq *rq)
2319 return rq->stop && task_on_rq_queued(rq->stop);
2322 static inline bool sched_dl_runnable(struct rq *rq)
2324 return rq->dl.dl_nr_running > 0;
2327 static inline bool sched_rt_runnable(struct rq *rq)
2329 return rq->rt.rt_queued > 0;
2332 static inline bool sched_fair_runnable(struct rq *rq)
2334 return rq->cfs.nr_running > 0;
2337 extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2338 extern struct task_struct *pick_next_task_idle(struct rq *rq);
2340 #define SCA_CHECK 0x01
2341 #define SCA_MIGRATE_DISABLE 0x02
2342 #define SCA_MIGRATE_ENABLE 0x04
2343 #define SCA_USER 0x08
2347 extern void update_group_capacity(struct sched_domain *sd, int cpu);
2349 extern void trigger_load_balance(struct rq *rq);
2351 extern void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx);
2353 static inline struct task_struct *get_push_task(struct rq *rq)
2355 struct task_struct *p = rq->curr;
2357 lockdep_assert_rq_held(rq);
2362 if (p->nr_cpus_allowed == 1)
2365 if (p->migration_disabled)
2368 rq->push_busy = true;
2369 return get_task_struct(p);
2372 extern int push_cpu_stop(void *arg);
2376 #ifdef CONFIG_CPU_IDLE
2377 static inline void idle_set_state(struct rq *rq,
2378 struct cpuidle_state *idle_state)
2380 rq->idle_state = idle_state;
2383 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2385 SCHED_WARN_ON(!rcu_read_lock_held());
2387 return rq->idle_state;
2390 static inline void idle_set_state(struct rq *rq,
2391 struct cpuidle_state *idle_state)
2395 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2401 extern void schedule_idle(void);
2403 extern void sysrq_sched_debug_show(void);
2404 extern void sched_init_granularity(void);
2405 extern void update_max_interval(void);
2407 extern void init_sched_dl_class(void);
2408 extern void init_sched_rt_class(void);
2409 extern void init_sched_fair_class(void);
2411 extern void reweight_task(struct task_struct *p, int prio);
2413 extern void resched_curr(struct rq *rq);
2414 extern void resched_cpu(int cpu);
2416 extern struct rt_bandwidth def_rt_bandwidth;
2417 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
2418 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
2420 extern void init_dl_task_timer(struct sched_dl_entity *dl_se);
2421 extern void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se);
2424 #define BW_UNIT (1 << BW_SHIFT)
2425 #define RATIO_SHIFT 8
2426 #define MAX_BW_BITS (64 - BW_SHIFT)
2427 #define MAX_BW ((1ULL << MAX_BW_BITS) - 1)
2428 unsigned long to_ratio(u64 period, u64 runtime);
2430 extern void init_entity_runnable_average(struct sched_entity *se);
2431 extern void post_init_entity_util_avg(struct task_struct *p);
2433 #ifdef CONFIG_NO_HZ_FULL
2434 extern bool sched_can_stop_tick(struct rq *rq);
2435 extern int __init sched_tick_offload_init(void);
2438 * Tick may be needed by tasks in the runqueue depending on their policy and
2439 * requirements. If tick is needed, lets send the target an IPI to kick it out of
2440 * nohz mode if necessary.
2442 static inline void sched_update_tick_dependency(struct rq *rq)
2444 int cpu = cpu_of(rq);
2446 if (!tick_nohz_full_cpu(cpu))
2449 if (sched_can_stop_tick(rq))
2450 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
2452 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
2455 static inline int sched_tick_offload_init(void) { return 0; }
2456 static inline void sched_update_tick_dependency(struct rq *rq) { }
2459 static inline void add_nr_running(struct rq *rq, unsigned count)
2461 unsigned prev_nr = rq->nr_running;
2463 rq->nr_running = prev_nr + count;
2464 if (trace_sched_update_nr_running_tp_enabled()) {
2465 call_trace_sched_update_nr_running(rq, count);
2469 if (prev_nr < 2 && rq->nr_running >= 2) {
2470 if (!READ_ONCE(rq->rd->overload))
2471 WRITE_ONCE(rq->rd->overload, 1);
2475 sched_update_tick_dependency(rq);
2478 static inline void sub_nr_running(struct rq *rq, unsigned count)
2480 rq->nr_running -= count;
2481 if (trace_sched_update_nr_running_tp_enabled()) {
2482 call_trace_sched_update_nr_running(rq, -count);
2485 /* Check if we still need preemption */
2486 sched_update_tick_dependency(rq);
2489 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
2490 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
2492 extern void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
2494 #ifdef CONFIG_PREEMPT_RT
2495 #define SCHED_NR_MIGRATE_BREAK 8
2497 #define SCHED_NR_MIGRATE_BREAK 32
2500 extern const_debug unsigned int sysctl_sched_nr_migrate;
2501 extern const_debug unsigned int sysctl_sched_migration_cost;
2503 #ifdef CONFIG_SCHED_DEBUG
2504 extern unsigned int sysctl_sched_latency;
2505 extern unsigned int sysctl_sched_min_granularity;
2506 extern unsigned int sysctl_sched_idle_min_granularity;
2507 extern unsigned int sysctl_sched_wakeup_granularity;
2508 extern int sysctl_resched_latency_warn_ms;
2509 extern int sysctl_resched_latency_warn_once;
2511 extern unsigned int sysctl_sched_tunable_scaling;
2513 extern unsigned int sysctl_numa_balancing_scan_delay;
2514 extern unsigned int sysctl_numa_balancing_scan_period_min;
2515 extern unsigned int sysctl_numa_balancing_scan_period_max;
2516 extern unsigned int sysctl_numa_balancing_scan_size;
2517 extern unsigned int sysctl_numa_balancing_hot_threshold;
2520 #ifdef CONFIG_SCHED_HRTICK
2524 * - enabled by features
2525 * - hrtimer is actually high res
2527 static inline int hrtick_enabled(struct rq *rq)
2529 if (!cpu_active(cpu_of(rq)))
2531 return hrtimer_is_hres_active(&rq->hrtick_timer);
2534 static inline int hrtick_enabled_fair(struct rq *rq)
2536 if (!sched_feat(HRTICK))
2538 return hrtick_enabled(rq);
2541 static inline int hrtick_enabled_dl(struct rq *rq)
2543 if (!sched_feat(HRTICK_DL))
2545 return hrtick_enabled(rq);
2548 void hrtick_start(struct rq *rq, u64 delay);
2552 static inline int hrtick_enabled_fair(struct rq *rq)
2557 static inline int hrtick_enabled_dl(struct rq *rq)
2562 static inline int hrtick_enabled(struct rq *rq)
2567 #endif /* CONFIG_SCHED_HRTICK */
2569 #ifndef arch_scale_freq_tick
2570 static __always_inline
2571 void arch_scale_freq_tick(void)
2576 #ifndef arch_scale_freq_capacity
2578 * arch_scale_freq_capacity - get the frequency scale factor of a given CPU.
2579 * @cpu: the CPU in question.
2581 * Return: the frequency scale factor normalized against SCHED_CAPACITY_SCALE, i.e.
2584 * ------ * SCHED_CAPACITY_SCALE
2587 static __always_inline
2588 unsigned long arch_scale_freq_capacity(int cpu)
2590 return SCHED_CAPACITY_SCALE;
2594 #ifdef CONFIG_SCHED_DEBUG
2596 * In double_lock_balance()/double_rq_lock(), we use raw_spin_rq_lock() to
2597 * acquire rq lock instead of rq_lock(). So at the end of these two functions
2598 * we need to call double_rq_clock_clear_update() to clear RQCF_UPDATED of
2599 * rq->clock_update_flags to avoid the WARN_DOUBLE_CLOCK warning.
2601 static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2)
2603 rq1->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2604 /* rq1 == rq2 for !CONFIG_SMP, so just clear RQCF_UPDATED once. */
2606 rq2->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2610 static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2) {}
2615 static inline bool rq_order_less(struct rq *rq1, struct rq *rq2)
2617 #ifdef CONFIG_SCHED_CORE
2619 * In order to not have {0,2},{1,3} turn into into an AB-BA,
2620 * order by core-id first and cpu-id second.
2624 * double_rq_lock(0,3); will take core-0, core-1 lock
2625 * double_rq_lock(1,2); will take core-1, core-0 lock
2627 * when only cpu-id is considered.
2629 if (rq1->core->cpu < rq2->core->cpu)
2631 if (rq1->core->cpu > rq2->core->cpu)
2635 * __sched_core_flip() relies on SMT having cpu-id lock order.
2638 return rq1->cpu < rq2->cpu;
2641 extern void double_rq_lock(struct rq *rq1, struct rq *rq2);
2643 #ifdef CONFIG_PREEMPTION
2646 * fair double_lock_balance: Safely acquires both rq->locks in a fair
2647 * way at the expense of forcing extra atomic operations in all
2648 * invocations. This assures that the double_lock is acquired using the
2649 * same underlying policy as the spinlock_t on this architecture, which
2650 * reduces latency compared to the unfair variant below. However, it
2651 * also adds more overhead and therefore may reduce throughput.
2653 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2654 __releases(this_rq->lock)
2655 __acquires(busiest->lock)
2656 __acquires(this_rq->lock)
2658 raw_spin_rq_unlock(this_rq);
2659 double_rq_lock(this_rq, busiest);
2666 * Unfair double_lock_balance: Optimizes throughput at the expense of
2667 * latency by eliminating extra atomic operations when the locks are
2668 * already in proper order on entry. This favors lower CPU-ids and will
2669 * grant the double lock to lower CPUs over higher ids under contention,
2670 * regardless of entry order into the function.
2672 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2673 __releases(this_rq->lock)
2674 __acquires(busiest->lock)
2675 __acquires(this_rq->lock)
2677 if (__rq_lockp(this_rq) == __rq_lockp(busiest) ||
2678 likely(raw_spin_rq_trylock(busiest))) {
2679 double_rq_clock_clear_update(this_rq, busiest);
2683 if (rq_order_less(this_rq, busiest)) {
2684 raw_spin_rq_lock_nested(busiest, SINGLE_DEPTH_NESTING);
2685 double_rq_clock_clear_update(this_rq, busiest);
2689 raw_spin_rq_unlock(this_rq);
2690 double_rq_lock(this_rq, busiest);
2695 #endif /* CONFIG_PREEMPTION */
2698 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2700 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2702 lockdep_assert_irqs_disabled();
2704 return _double_lock_balance(this_rq, busiest);
2707 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2708 __releases(busiest->lock)
2710 if (__rq_lockp(this_rq) != __rq_lockp(busiest))
2711 raw_spin_rq_unlock(busiest);
2712 lock_set_subclass(&__rq_lockp(this_rq)->dep_map, 0, _RET_IP_);
2715 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
2721 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2724 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
2730 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2733 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2739 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2743 * double_rq_unlock - safely unlock two runqueues
2745 * Note this does not restore interrupts like task_rq_unlock,
2746 * you need to do so manually after calling.
2748 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2749 __releases(rq1->lock)
2750 __releases(rq2->lock)
2752 if (__rq_lockp(rq1) != __rq_lockp(rq2))
2753 raw_spin_rq_unlock(rq2);
2755 __release(rq2->lock);
2756 raw_spin_rq_unlock(rq1);
2759 extern void set_rq_online (struct rq *rq);
2760 extern void set_rq_offline(struct rq *rq);
2761 extern bool sched_smp_initialized;
2763 #else /* CONFIG_SMP */
2766 * double_rq_lock - safely lock two runqueues
2768 * Note this does not disable interrupts like task_rq_lock,
2769 * you need to do so manually before calling.
2771 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
2772 __acquires(rq1->lock)
2773 __acquires(rq2->lock)
2775 WARN_ON_ONCE(!irqs_disabled());
2776 WARN_ON_ONCE(rq1 != rq2);
2777 raw_spin_rq_lock(rq1);
2778 __acquire(rq2->lock); /* Fake it out ;) */
2779 double_rq_clock_clear_update(rq1, rq2);
2783 * double_rq_unlock - safely unlock two runqueues
2785 * Note this does not restore interrupts like task_rq_unlock,
2786 * you need to do so manually after calling.
2788 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2789 __releases(rq1->lock)
2790 __releases(rq2->lock)
2792 WARN_ON_ONCE(rq1 != rq2);
2793 raw_spin_rq_unlock(rq1);
2794 __release(rq2->lock);
2799 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
2800 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
2802 #ifdef CONFIG_SCHED_DEBUG
2803 extern bool sched_debug_verbose;
2805 extern void print_cfs_stats(struct seq_file *m, int cpu);
2806 extern void print_rt_stats(struct seq_file *m, int cpu);
2807 extern void print_dl_stats(struct seq_file *m, int cpu);
2808 extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
2809 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2810 extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
2812 extern void resched_latency_warn(int cpu, u64 latency);
2813 #ifdef CONFIG_NUMA_BALANCING
2815 show_numa_stats(struct task_struct *p, struct seq_file *m);
2817 print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
2818 unsigned long tpf, unsigned long gsf, unsigned long gpf);
2819 #endif /* CONFIG_NUMA_BALANCING */
2821 static inline void resched_latency_warn(int cpu, u64 latency) {}
2822 #endif /* CONFIG_SCHED_DEBUG */
2824 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
2825 extern void init_rt_rq(struct rt_rq *rt_rq);
2826 extern void init_dl_rq(struct dl_rq *dl_rq);
2828 extern void cfs_bandwidth_usage_inc(void);
2829 extern void cfs_bandwidth_usage_dec(void);
2831 #ifdef CONFIG_NO_HZ_COMMON
2832 #define NOHZ_BALANCE_KICK_BIT 0
2833 #define NOHZ_STATS_KICK_BIT 1
2834 #define NOHZ_NEWILB_KICK_BIT 2
2835 #define NOHZ_NEXT_KICK_BIT 3
2837 /* Run rebalance_domains() */
2838 #define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT)
2839 /* Update blocked load */
2840 #define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT)
2841 /* Update blocked load when entering idle */
2842 #define NOHZ_NEWILB_KICK BIT(NOHZ_NEWILB_KICK_BIT)
2843 /* Update nohz.next_balance */
2844 #define NOHZ_NEXT_KICK BIT(NOHZ_NEXT_KICK_BIT)
2846 #define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK | NOHZ_NEXT_KICK)
2848 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
2850 extern void nohz_balance_exit_idle(struct rq *rq);
2852 static inline void nohz_balance_exit_idle(struct rq *rq) { }
2855 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
2856 extern void nohz_run_idle_balance(int cpu);
2858 static inline void nohz_run_idle_balance(int cpu) { }
2861 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2866 struct u64_stats_sync sync;
2869 DECLARE_PER_CPU(struct irqtime, cpu_irqtime);
2872 * Returns the irqtime minus the softirq time computed by ksoftirqd.
2873 * Otherwise ksoftirqd's sum_exec_runtime is subtracted its own runtime
2874 * and never move forward.
2876 static inline u64 irq_time_read(int cpu)
2878 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu);
2883 seq = __u64_stats_fetch_begin(&irqtime->sync);
2884 total = irqtime->total;
2885 } while (__u64_stats_fetch_retry(&irqtime->sync, seq));
2889 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2891 #ifdef CONFIG_CPU_FREQ
2892 DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data);
2895 * cpufreq_update_util - Take a note about CPU utilization changes.
2896 * @rq: Runqueue to carry out the update for.
2897 * @flags: Update reason flags.
2899 * This function is called by the scheduler on the CPU whose utilization is
2902 * It can only be called from RCU-sched read-side critical sections.
2904 * The way cpufreq is currently arranged requires it to evaluate the CPU
2905 * performance state (frequency/voltage) on a regular basis to prevent it from
2906 * being stuck in a completely inadequate performance level for too long.
2907 * That is not guaranteed to happen if the updates are only triggered from CFS
2908 * and DL, though, because they may not be coming in if only RT tasks are
2909 * active all the time (or there are RT tasks only).
2911 * As a workaround for that issue, this function is called periodically by the
2912 * RT sched class to trigger extra cpufreq updates to prevent it from stalling,
2913 * but that really is a band-aid. Going forward it should be replaced with
2914 * solutions targeted more specifically at RT tasks.
2916 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags)
2918 struct update_util_data *data;
2920 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data,
2923 data->func(data, rq_clock(rq), flags);
2926 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
2927 #endif /* CONFIG_CPU_FREQ */
2929 #ifdef arch_scale_freq_capacity
2930 # ifndef arch_scale_freq_invariant
2931 # define arch_scale_freq_invariant() true
2934 # define arch_scale_freq_invariant() false
2938 static inline unsigned long capacity_orig_of(int cpu)
2940 return cpu_rq(cpu)->cpu_capacity_orig;
2944 * enum cpu_util_type - CPU utilization type
2945 * @FREQUENCY_UTIL: Utilization used to select frequency
2946 * @ENERGY_UTIL: Utilization used during energy calculation
2948 * The utilization signals of all scheduling classes (CFS/RT/DL) and IRQ time
2949 * need to be aggregated differently depending on the usage made of them. This
2950 * enum is used within effective_cpu_util() to differentiate the types of
2951 * utilization expected by the callers, and adjust the aggregation accordingly.
2953 enum cpu_util_type {
2958 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
2959 enum cpu_util_type type,
2960 struct task_struct *p);
2963 * Verify the fitness of task @p to run on @cpu taking into account the
2964 * CPU original capacity and the runtime/deadline ratio of the task.
2966 * The function will return true if the original capacity of @cpu is
2967 * greater than or equal to task's deadline density right shifted by
2968 * (BW_SHIFT - SCHED_CAPACITY_SHIFT) and false otherwise.
2970 static inline bool dl_task_fits_capacity(struct task_struct *p, int cpu)
2972 unsigned long cap = arch_scale_cpu_capacity(cpu);
2974 return cap >= p->dl.dl_density >> (BW_SHIFT - SCHED_CAPACITY_SHIFT);
2977 static inline unsigned long cpu_bw_dl(struct rq *rq)
2979 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT;
2982 static inline unsigned long cpu_util_dl(struct rq *rq)
2984 return READ_ONCE(rq->avg_dl.util_avg);
2988 extern unsigned long cpu_util_cfs(int cpu);
2989 extern unsigned long cpu_util_cfs_boost(int cpu);
2991 static inline unsigned long cpu_util_rt(struct rq *rq)
2993 return READ_ONCE(rq->avg_rt.util_avg);
2997 #ifdef CONFIG_UCLAMP_TASK
2998 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id);
3000 static inline unsigned long uclamp_rq_get(struct rq *rq,
3001 enum uclamp_id clamp_id)
3003 return READ_ONCE(rq->uclamp[clamp_id].value);
3006 static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3009 WRITE_ONCE(rq->uclamp[clamp_id].value, value);
3012 static inline bool uclamp_rq_is_idle(struct rq *rq)
3014 return rq->uclamp_flags & UCLAMP_FLAG_IDLE;
3018 * uclamp_rq_util_with - clamp @util with @rq and @p effective uclamp values.
3019 * @rq: The rq to clamp against. Must not be NULL.
3020 * @util: The util value to clamp.
3021 * @p: The task to clamp against. Can be NULL if you want to clamp
3024 * Clamps the passed @util to the max(@rq, @p) effective uclamp values.
3026 * If sched_uclamp_used static key is disabled, then just return the util
3027 * without any clamping since uclamp aggregation at the rq level in the fast
3028 * path is disabled, rendering this operation a NOP.
3030 * Use uclamp_eff_value() if you don't care about uclamp values at rq level. It
3031 * will return the correct effective uclamp value of the task even if the
3032 * static key is disabled.
3034 static __always_inline
3035 unsigned long uclamp_rq_util_with(struct rq *rq, unsigned long util,
3036 struct task_struct *p)
3038 unsigned long min_util = 0;
3039 unsigned long max_util = 0;
3041 if (!static_branch_likely(&sched_uclamp_used))
3045 min_util = uclamp_eff_value(p, UCLAMP_MIN);
3046 max_util = uclamp_eff_value(p, UCLAMP_MAX);
3049 * Ignore last runnable task's max clamp, as this task will
3050 * reset it. Similarly, no need to read the rq's min clamp.
3052 if (uclamp_rq_is_idle(rq))
3056 min_util = max_t(unsigned long, min_util, uclamp_rq_get(rq, UCLAMP_MIN));
3057 max_util = max_t(unsigned long, max_util, uclamp_rq_get(rq, UCLAMP_MAX));
3060 * Since CPU's {min,max}_util clamps are MAX aggregated considering
3061 * RUNNABLE tasks with _different_ clamps, we can end up with an
3062 * inversion. Fix it now when the clamps are applied.
3064 if (unlikely(min_util >= max_util))
3067 return clamp(util, min_util, max_util);
3070 /* Is the rq being capped/throttled by uclamp_max? */
3071 static inline bool uclamp_rq_is_capped(struct rq *rq)
3073 unsigned long rq_util;
3074 unsigned long max_util;
3076 if (!static_branch_likely(&sched_uclamp_used))
3079 rq_util = cpu_util_cfs(cpu_of(rq)) + cpu_util_rt(rq);
3080 max_util = READ_ONCE(rq->uclamp[UCLAMP_MAX].value);
3082 return max_util != SCHED_CAPACITY_SCALE && rq_util >= max_util;
3086 * When uclamp is compiled in, the aggregation at rq level is 'turned off'
3087 * by default in the fast path and only gets turned on once userspace performs
3088 * an operation that requires it.
3090 * Returns true if userspace opted-in to use uclamp and aggregation at rq level
3093 static inline bool uclamp_is_used(void)
3095 return static_branch_likely(&sched_uclamp_used);
3097 #else /* CONFIG_UCLAMP_TASK */
3098 static inline unsigned long uclamp_eff_value(struct task_struct *p,
3099 enum uclamp_id clamp_id)
3101 if (clamp_id == UCLAMP_MIN)
3104 return SCHED_CAPACITY_SCALE;
3108 unsigned long uclamp_rq_util_with(struct rq *rq, unsigned long util,
3109 struct task_struct *p)
3114 static inline bool uclamp_rq_is_capped(struct rq *rq) { return false; }
3116 static inline bool uclamp_is_used(void)
3121 static inline unsigned long uclamp_rq_get(struct rq *rq,
3122 enum uclamp_id clamp_id)
3124 if (clamp_id == UCLAMP_MIN)
3127 return SCHED_CAPACITY_SCALE;
3130 static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3135 static inline bool uclamp_rq_is_idle(struct rq *rq)
3139 #endif /* CONFIG_UCLAMP_TASK */
3141 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
3142 static inline unsigned long cpu_util_irq(struct rq *rq)
3144 return rq->avg_irq.util_avg;
3148 unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3150 util *= (max - irq);
3157 static inline unsigned long cpu_util_irq(struct rq *rq)
3163 unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3169 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
3171 #define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus)))
3173 DECLARE_STATIC_KEY_FALSE(sched_energy_present);
3175 static inline bool sched_energy_enabled(void)
3177 return static_branch_unlikely(&sched_energy_present);
3180 #else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */
3182 #define perf_domain_span(pd) NULL
3183 static inline bool sched_energy_enabled(void) { return false; }
3185 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
3187 #ifdef CONFIG_MEMBARRIER
3189 * The scheduler provides memory barriers required by membarrier between:
3190 * - prior user-space memory accesses and store to rq->membarrier_state,
3191 * - store to rq->membarrier_state and following user-space memory accesses.
3192 * In the same way it provides those guarantees around store to rq->curr.
3194 static inline void membarrier_switch_mm(struct rq *rq,
3195 struct mm_struct *prev_mm,
3196 struct mm_struct *next_mm)
3198 int membarrier_state;
3200 if (prev_mm == next_mm)
3203 membarrier_state = atomic_read(&next_mm->membarrier_state);
3204 if (READ_ONCE(rq->membarrier_state) == membarrier_state)
3207 WRITE_ONCE(rq->membarrier_state, membarrier_state);
3210 static inline void membarrier_switch_mm(struct rq *rq,
3211 struct mm_struct *prev_mm,
3212 struct mm_struct *next_mm)
3218 static inline bool is_per_cpu_kthread(struct task_struct *p)
3220 if (!(p->flags & PF_KTHREAD))
3223 if (p->nr_cpus_allowed != 1)
3230 extern void swake_up_all_locked(struct swait_queue_head *q);
3231 extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);
3233 extern int try_to_wake_up(struct task_struct *tsk, unsigned int state, int wake_flags);
3235 #ifdef CONFIG_PREEMPT_DYNAMIC
3236 extern int preempt_dynamic_mode;
3237 extern int sched_dynamic_mode(const char *str);
3238 extern void sched_dynamic_update(int mode);
3241 static inline void update_current_exec_runtime(struct task_struct *curr,
3242 u64 now, u64 delta_exec)
3244 curr->se.sum_exec_runtime += delta_exec;
3245 account_group_exec_runtime(curr, delta_exec);
3247 curr->se.exec_start = now;
3248 cgroup_account_cputime(curr, delta_exec);
3251 #ifdef CONFIG_SCHED_MM_CID
3253 #define SCHED_MM_CID_PERIOD_NS (100ULL * 1000000) /* 100ms */
3254 #define MM_CID_SCAN_DELAY 100 /* 100ms */
3256 extern raw_spinlock_t cid_lock;
3257 extern int use_cid_lock;
3259 extern void sched_mm_cid_migrate_from(struct task_struct *t);
3260 extern void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t);
3261 extern void task_tick_mm_cid(struct rq *rq, struct task_struct *curr);
3262 extern void init_sched_mm_cid(struct task_struct *t);
3264 static inline void __mm_cid_put(struct mm_struct *mm, int cid)
3268 cpumask_clear_cpu(cid, mm_cidmask(mm));
3272 * The per-mm/cpu cid can have the MM_CID_LAZY_PUT flag set or transition to
3273 * the MM_CID_UNSET state without holding the rq lock, but the rq lock needs to
3274 * be held to transition to other states.
3276 * State transitions synchronized with cmpxchg or try_cmpxchg need to be
3277 * consistent across cpus, which prevents use of this_cpu_cmpxchg.
3279 static inline void mm_cid_put_lazy(struct task_struct *t)
3281 struct mm_struct *mm = t->mm;
3282 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3285 lockdep_assert_irqs_disabled();
3286 cid = __this_cpu_read(pcpu_cid->cid);
3287 if (!mm_cid_is_lazy_put(cid) ||
3288 !try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3290 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3293 static inline int mm_cid_pcpu_unset(struct mm_struct *mm)
3295 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3298 lockdep_assert_irqs_disabled();
3299 cid = __this_cpu_read(pcpu_cid->cid);
3301 if (mm_cid_is_unset(cid))
3302 return MM_CID_UNSET;
3304 * Attempt transition from valid or lazy-put to unset.
3306 res = cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, cid, MM_CID_UNSET);
3314 static inline void mm_cid_put(struct mm_struct *mm)
3318 lockdep_assert_irqs_disabled();
3319 cid = mm_cid_pcpu_unset(mm);
3320 if (cid == MM_CID_UNSET)
3322 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3325 static inline int __mm_cid_try_get(struct mm_struct *mm)
3327 struct cpumask *cpumask;
3330 cpumask = mm_cidmask(mm);
3332 * Retry finding first zero bit if the mask is temporarily
3333 * filled. This only happens during concurrent remote-clear
3334 * which owns a cid without holding a rq lock.
3337 cid = cpumask_first_zero(cpumask);
3338 if (cid < nr_cpu_ids)
3342 if (cpumask_test_and_set_cpu(cid, cpumask))
3348 * Save a snapshot of the current runqueue time of this cpu
3349 * with the per-cpu cid value, allowing to estimate how recently it was used.
3351 static inline void mm_cid_snapshot_time(struct rq *rq, struct mm_struct *mm)
3353 struct mm_cid *pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(rq));
3355 lockdep_assert_rq_held(rq);
3356 WRITE_ONCE(pcpu_cid->time, rq->clock);
3359 static inline int __mm_cid_get(struct rq *rq, struct mm_struct *mm)
3364 * All allocations (even those using the cid_lock) are lock-free. If
3365 * use_cid_lock is set, hold the cid_lock to perform cid allocation to
3366 * guarantee forward progress.
3368 if (!READ_ONCE(use_cid_lock)) {
3369 cid = __mm_cid_try_get(mm);
3372 raw_spin_lock(&cid_lock);
3374 raw_spin_lock(&cid_lock);
3375 cid = __mm_cid_try_get(mm);
3381 * cid concurrently allocated. Retry while forcing following
3382 * allocations to use the cid_lock to ensure forward progress.
3384 WRITE_ONCE(use_cid_lock, 1);
3386 * Set use_cid_lock before allocation. Only care about program order
3387 * because this is only required for forward progress.
3391 * Retry until it succeeds. It is guaranteed to eventually succeed once
3392 * all newcoming allocations observe the use_cid_lock flag set.
3395 cid = __mm_cid_try_get(mm);
3399 * Allocate before clearing use_cid_lock. Only care about
3400 * program order because this is for forward progress.
3403 WRITE_ONCE(use_cid_lock, 0);
3405 raw_spin_unlock(&cid_lock);
3407 mm_cid_snapshot_time(rq, mm);
3411 static inline int mm_cid_get(struct rq *rq, struct mm_struct *mm)
3413 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3414 struct cpumask *cpumask;
3417 lockdep_assert_rq_held(rq);
3418 cpumask = mm_cidmask(mm);
3419 cid = __this_cpu_read(pcpu_cid->cid);
3420 if (mm_cid_is_valid(cid)) {
3421 mm_cid_snapshot_time(rq, mm);
3424 if (mm_cid_is_lazy_put(cid)) {
3425 if (try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3426 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3428 cid = __mm_cid_get(rq, mm);
3429 __this_cpu_write(pcpu_cid->cid, cid);
3433 static inline void switch_mm_cid(struct rq *rq,
3434 struct task_struct *prev,
3435 struct task_struct *next)
3438 * Provide a memory barrier between rq->curr store and load of
3439 * {prev,next}->mm->pcpu_cid[cpu] on rq->curr->mm transition.
3441 * Should be adapted if context_switch() is modified.
3443 if (!next->mm) { // to kernel
3445 * user -> kernel transition does not guarantee a barrier, but
3446 * we can use the fact that it performs an atomic operation in
3449 if (prev->mm) // from user
3450 smp_mb__after_mmgrab();
3452 * kernel -> kernel transition does not change rq->curr->mm
3453 * state. It stays NULL.
3457 * kernel -> user transition does not provide a barrier
3458 * between rq->curr store and load of {prev,next}->mm->pcpu_cid[cpu].
3461 if (!prev->mm) // from kernel
3464 * user -> user transition guarantees a memory barrier through
3465 * switch_mm() when current->mm changes. If current->mm is
3466 * unchanged, no barrier is needed.
3469 if (prev->mm_cid_active) {
3470 mm_cid_snapshot_time(rq, prev->mm);
3471 mm_cid_put_lazy(prev);
3474 if (next->mm_cid_active)
3475 next->last_mm_cid = next->mm_cid = mm_cid_get(rq, next->mm);
3479 static inline void switch_mm_cid(struct rq *rq, struct task_struct *prev, struct task_struct *next) { }
3480 static inline void sched_mm_cid_migrate_from(struct task_struct *t) { }
3481 static inline void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t) { }
3482 static inline void task_tick_mm_cid(struct rq *rq, struct task_struct *curr) { }
3483 static inline void init_sched_mm_cid(struct task_struct *t) { }
3486 #endif /* _KERNEL_SCHED_SCHED_H */