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 * Helpers for converting nanosecond timing to jiffy resolution
117 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
120 * Increase resolution of nice-level calculations for 64-bit architectures.
121 * The extra resolution improves shares distribution and load balancing of
122 * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
123 * hierarchies, especially on larger systems. This is not a user-visible change
124 * and does not change the user-interface for setting shares/weights.
126 * We increase resolution only if we have enough bits to allow this increased
127 * resolution (i.e. 64-bit). The costs for increasing resolution when 32-bit
128 * are pretty high and the returns do not justify the increased costs.
130 * Really only required when CONFIG_FAIR_GROUP_SCHED=y is also set, but to
131 * increase coverage and consistency always enable it on 64-bit platforms.
134 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT + SCHED_FIXEDPOINT_SHIFT)
135 # define scale_load(w) ((w) << SCHED_FIXEDPOINT_SHIFT)
136 # define scale_load_down(w) \
138 unsigned long __w = (w); \
140 __w = max(2UL, __w >> SCHED_FIXEDPOINT_SHIFT); \
144 # define NICE_0_LOAD_SHIFT (SCHED_FIXEDPOINT_SHIFT)
145 # define scale_load(w) (w)
146 # define scale_load_down(w) (w)
150 * Task weight (visible to users) and its load (invisible to users) have
151 * independent resolution, but they should be well calibrated. We use
152 * scale_load() and scale_load_down(w) to convert between them. The
153 * following must be true:
155 * scale_load(sched_prio_to_weight[NICE_TO_PRIO(0)-MAX_RT_PRIO]) == NICE_0_LOAD
158 #define NICE_0_LOAD (1L << NICE_0_LOAD_SHIFT)
161 * Single value that decides SCHED_DEADLINE internal math precision.
162 * 10 -> just above 1us
163 * 9 -> just above 0.5us
168 * Single value that denotes runtime == period, ie unlimited time.
170 #define RUNTIME_INF ((u64)~0ULL)
172 static inline int idle_policy(int policy)
174 return policy == SCHED_IDLE;
176 static inline int fair_policy(int policy)
178 return policy == SCHED_NORMAL || policy == SCHED_BATCH;
181 static inline int rt_policy(int policy)
183 return policy == SCHED_FIFO || policy == SCHED_RR;
186 static inline int dl_policy(int policy)
188 return policy == SCHED_DEADLINE;
190 static inline bool valid_policy(int policy)
192 return idle_policy(policy) || fair_policy(policy) ||
193 rt_policy(policy) || dl_policy(policy);
196 static inline int task_has_idle_policy(struct task_struct *p)
198 return idle_policy(p->policy);
201 static inline int task_has_rt_policy(struct task_struct *p)
203 return rt_policy(p->policy);
206 static inline int task_has_dl_policy(struct task_struct *p)
208 return dl_policy(p->policy);
211 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
213 static inline void update_avg(u64 *avg, u64 sample)
215 s64 diff = sample - *avg;
220 * Shifting a value by an exponent greater *or equal* to the size of said value
221 * is UB; cap at size-1.
223 #define shr_bound(val, shift) \
224 (val >> min_t(typeof(shift), shift, BITS_PER_TYPE(typeof(val)) - 1))
227 * !! For sched_setattr_nocheck() (kernel) only !!
229 * This is actually gross. :(
231 * It is used to make schedutil kworker(s) higher priority than SCHED_DEADLINE
232 * tasks, but still be able to sleep. We need this on platforms that cannot
233 * atomically change clock frequency. Remove once fast switching will be
234 * available on such platforms.
236 * SUGOV stands for SchedUtil GOVernor.
238 #define SCHED_FLAG_SUGOV 0x10000000
240 #define SCHED_DL_FLAGS (SCHED_FLAG_RECLAIM | SCHED_FLAG_DL_OVERRUN | SCHED_FLAG_SUGOV)
242 static inline bool dl_entity_is_special(const struct sched_dl_entity *dl_se)
244 #ifdef CONFIG_CPU_FREQ_GOV_SCHEDUTIL
245 return unlikely(dl_se->flags & SCHED_FLAG_SUGOV);
252 * Tells if entity @a should preempt entity @b.
254 static inline bool dl_entity_preempt(const struct sched_dl_entity *a,
255 const struct sched_dl_entity *b)
257 return dl_entity_is_special(a) ||
258 dl_time_before(a->deadline, b->deadline);
262 * This is the priority-queue data structure of the RT scheduling class:
264 struct rt_prio_array {
265 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
266 struct list_head queue[MAX_RT_PRIO];
269 struct rt_bandwidth {
270 /* nests inside the rq lock: */
271 raw_spinlock_t rt_runtime_lock;
274 struct hrtimer rt_period_timer;
275 unsigned int rt_period_active;
278 static inline int dl_bandwidth_enabled(void)
280 return sysctl_sched_rt_runtime >= 0;
284 * To keep the bandwidth of -deadline tasks under control
285 * we need some place where:
286 * - store the maximum -deadline bandwidth of each cpu;
287 * - cache the fraction of bandwidth that is currently allocated in
290 * This is all done in the data structure below. It is similar to the
291 * one used for RT-throttling (rt_bandwidth), with the main difference
292 * that, since here we are only interested in admission control, we
293 * do not decrease any runtime while the group "executes", neither we
294 * need a timer to replenish it.
296 * With respect to SMP, bandwidth is given on a per root domain basis,
298 * - bw (< 100%) is the deadline bandwidth of each CPU;
299 * - total_bw is the currently allocated bandwidth in each root domain;
307 extern void init_dl_bw(struct dl_bw *dl_b);
308 extern int sched_dl_global_validate(void);
309 extern void sched_dl_do_global(void);
310 extern int sched_dl_overflow(struct task_struct *p, int policy, const struct sched_attr *attr);
311 extern void __setparam_dl(struct task_struct *p, const struct sched_attr *attr);
312 extern void __getparam_dl(struct task_struct *p, struct sched_attr *attr);
313 extern bool __checkparam_dl(const struct sched_attr *attr);
314 extern bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr);
315 extern int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur, const struct cpumask *trial);
316 extern int dl_bw_check_overflow(int cpu);
319 * SCHED_DEADLINE supports servers (nested scheduling) with the following
322 * dl_se::rq -- runqueue we belong to.
324 * dl_se::server_has_tasks() -- used on bandwidth enforcement; we 'stop' the
325 * server when it runs out of tasks to run.
327 * dl_se::server_pick() -- nested pick_next_task(); we yield the period if this
330 * dl_server_update() -- called from update_curr_common(), propagates runtime
334 * dl_server_stop() -- start/stop the server when it has (no) tasks.
336 * dl_server_init() -- initializes the server.
338 extern void dl_server_update(struct sched_dl_entity *dl_se, s64 delta_exec);
339 extern void dl_server_start(struct sched_dl_entity *dl_se);
340 extern void dl_server_stop(struct sched_dl_entity *dl_se);
341 extern void dl_server_init(struct sched_dl_entity *dl_se, struct rq *rq,
342 dl_server_has_tasks_f has_tasks,
343 dl_server_pick_f pick);
345 #ifdef CONFIG_CGROUP_SCHED
350 extern struct list_head task_groups;
352 struct cfs_bandwidth {
353 #ifdef CONFIG_CFS_BANDWIDTH
360 s64 hierarchical_quota;
365 struct hrtimer period_timer;
366 struct hrtimer slack_timer;
367 struct list_head throttled_cfs_rq;
378 /* Task group related information */
380 struct cgroup_subsys_state css;
382 #ifdef CONFIG_FAIR_GROUP_SCHED
383 /* schedulable entities of this group on each CPU */
384 struct sched_entity **se;
385 /* runqueue "owned" by this group on each CPU */
386 struct cfs_rq **cfs_rq;
387 unsigned long shares;
389 /* A positive value indicates that this is a SCHED_IDLE group. */
394 * load_avg can be heavily contended at clock tick time, so put
395 * it in its own cacheline separated from the fields above which
396 * will also be accessed at each tick.
398 atomic_long_t load_avg ____cacheline_aligned;
402 #ifdef CONFIG_RT_GROUP_SCHED
403 struct sched_rt_entity **rt_se;
404 struct rt_rq **rt_rq;
406 struct rt_bandwidth rt_bandwidth;
410 struct list_head list;
412 struct task_group *parent;
413 struct list_head siblings;
414 struct list_head children;
416 #ifdef CONFIG_SCHED_AUTOGROUP
417 struct autogroup *autogroup;
420 struct cfs_bandwidth cfs_bandwidth;
422 #ifdef CONFIG_UCLAMP_TASK_GROUP
423 /* The two decimal precision [%] value requested from user-space */
424 unsigned int uclamp_pct[UCLAMP_CNT];
425 /* Clamp values requested for a task group */
426 struct uclamp_se uclamp_req[UCLAMP_CNT];
427 /* Effective clamp values used for a task group */
428 struct uclamp_se uclamp[UCLAMP_CNT];
433 #ifdef CONFIG_FAIR_GROUP_SCHED
434 #define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
437 * A weight of 0 or 1 can cause arithmetics problems.
438 * A weight of a cfs_rq is the sum of weights of which entities
439 * are queued on this cfs_rq, so a weight of a entity should not be
440 * too large, so as the shares value of a task group.
441 * (The default weight is 1024 - so there's no practical
442 * limitation from this.)
444 #define MIN_SHARES (1UL << 1)
445 #define MAX_SHARES (1UL << 18)
448 typedef int (*tg_visitor)(struct task_group *, void *);
450 extern int walk_tg_tree_from(struct task_group *from,
451 tg_visitor down, tg_visitor up, void *data);
454 * Iterate the full tree, calling @down when first entering a node and @up when
455 * leaving it for the final time.
457 * Caller must hold rcu_lock or sufficient equivalent.
459 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
461 return walk_tg_tree_from(&root_task_group, down, up, data);
464 extern int tg_nop(struct task_group *tg, void *data);
466 #ifdef CONFIG_FAIR_GROUP_SCHED
467 extern void free_fair_sched_group(struct task_group *tg);
468 extern int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent);
469 extern void online_fair_sched_group(struct task_group *tg);
470 extern void unregister_fair_sched_group(struct task_group *tg);
472 static inline void free_fair_sched_group(struct task_group *tg) { }
473 static inline int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
477 static inline void online_fair_sched_group(struct task_group *tg) { }
478 static inline void unregister_fair_sched_group(struct task_group *tg) { }
481 extern void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
482 struct sched_entity *se, int cpu,
483 struct sched_entity *parent);
484 extern void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent);
486 extern void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b);
487 extern void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b);
488 extern void unthrottle_cfs_rq(struct cfs_rq *cfs_rq);
489 extern bool cfs_task_bw_constrained(struct task_struct *p);
491 extern void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
492 struct sched_rt_entity *rt_se, int cpu,
493 struct sched_rt_entity *parent);
494 extern int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us);
495 extern int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us);
496 extern long sched_group_rt_runtime(struct task_group *tg);
497 extern long sched_group_rt_period(struct task_group *tg);
498 extern int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk);
500 extern struct task_group *sched_create_group(struct task_group *parent);
501 extern void sched_online_group(struct task_group *tg,
502 struct task_group *parent);
503 extern void sched_destroy_group(struct task_group *tg);
504 extern void sched_release_group(struct task_group *tg);
506 extern void sched_move_task(struct task_struct *tsk);
508 #ifdef CONFIG_FAIR_GROUP_SCHED
509 extern int sched_group_set_shares(struct task_group *tg, unsigned long shares);
511 extern int sched_group_set_idle(struct task_group *tg, long idle);
514 extern void set_task_rq_fair(struct sched_entity *se,
515 struct cfs_rq *prev, struct cfs_rq *next);
516 #else /* !CONFIG_SMP */
517 static inline void set_task_rq_fair(struct sched_entity *se,
518 struct cfs_rq *prev, struct cfs_rq *next) { }
519 #endif /* CONFIG_SMP */
520 #endif /* CONFIG_FAIR_GROUP_SCHED */
522 #else /* CONFIG_CGROUP_SCHED */
524 struct cfs_bandwidth { };
525 static inline bool cfs_task_bw_constrained(struct task_struct *p) { return false; }
527 #endif /* CONFIG_CGROUP_SCHED */
529 extern void unregister_rt_sched_group(struct task_group *tg);
530 extern void free_rt_sched_group(struct task_group *tg);
531 extern int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent);
534 * u64_u32_load/u64_u32_store
536 * Use a copy of a u64 value to protect against data race. This is only
537 * applicable for 32-bits architectures.
540 # define u64_u32_load_copy(var, copy) var
541 # define u64_u32_store_copy(var, copy, val) (var = val)
543 # define u64_u32_load_copy(var, copy) \
545 u64 __val, __val_copy; \
549 * paired with u64_u32_store_copy(), ordering access \
554 } while (__val != __val_copy); \
557 # define u64_u32_store_copy(var, copy, val) \
559 typeof(val) __val = (val); \
562 * paired with u64_u32_load_copy(), ordering access to var and \
569 # define u64_u32_load(var) u64_u32_load_copy(var, var##_copy)
570 # define u64_u32_store(var, val) u64_u32_store_copy(var, var##_copy, val)
572 /* CFS-related fields in a runqueue */
574 struct load_weight load;
575 unsigned int nr_running;
576 unsigned int h_nr_running; /* SCHED_{NORMAL,BATCH,IDLE} */
577 unsigned int idle_nr_running; /* SCHED_IDLE */
578 unsigned int idle_h_nr_running; /* SCHED_IDLE */
585 #ifdef CONFIG_SCHED_CORE
586 unsigned int forceidle_seq;
591 u64 min_vruntime_copy;
594 struct rb_root_cached tasks_timeline;
597 * 'curr' points to currently running entity on this cfs_rq.
598 * It is set to NULL otherwise (i.e when none are currently running).
600 struct sched_entity *curr;
601 struct sched_entity *next;
603 #ifdef CONFIG_SCHED_DEBUG
604 unsigned int nr_spread_over;
611 struct sched_avg avg;
613 u64 last_update_time_copy;
616 raw_spinlock_t lock ____cacheline_aligned;
618 unsigned long load_avg;
619 unsigned long util_avg;
620 unsigned long runnable_avg;
623 #ifdef CONFIG_FAIR_GROUP_SCHED
624 u64 last_update_tg_load_avg;
625 unsigned long tg_load_avg_contrib;
627 long prop_runnable_sum;
630 * h_load = weight * f(tg)
632 * Where f(tg) is the recursive weight fraction assigned to
635 unsigned long h_load;
636 u64 last_h_load_update;
637 struct sched_entity *h_load_next;
638 #endif /* CONFIG_FAIR_GROUP_SCHED */
639 #endif /* CONFIG_SMP */
641 #ifdef CONFIG_FAIR_GROUP_SCHED
642 struct rq *rq; /* CPU runqueue to which this cfs_rq is attached */
645 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
646 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
647 * (like users, containers etc.)
649 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a CPU.
650 * This list is used during load balance.
653 struct list_head leaf_cfs_rq_list;
654 struct task_group *tg; /* group that "owns" this runqueue */
656 /* Locally cached copy of our task_group's idle value */
659 #ifdef CONFIG_CFS_BANDWIDTH
661 s64 runtime_remaining;
663 u64 throttled_pelt_idle;
665 u64 throttled_pelt_idle_copy;
668 u64 throttled_clock_pelt;
669 u64 throttled_clock_pelt_time;
670 u64 throttled_clock_self;
671 u64 throttled_clock_self_time;
674 struct list_head throttled_list;
675 struct list_head throttled_csd_list;
676 #endif /* CONFIG_CFS_BANDWIDTH */
677 #endif /* CONFIG_FAIR_GROUP_SCHED */
680 static inline int rt_bandwidth_enabled(void)
682 return sysctl_sched_rt_runtime >= 0;
685 /* RT IPI pull logic requires IRQ_WORK */
686 #if defined(CONFIG_IRQ_WORK) && defined(CONFIG_SMP)
687 # define HAVE_RT_PUSH_IPI
690 /* Real-Time classes' related field in a runqueue: */
692 struct rt_prio_array active;
693 unsigned int rt_nr_running;
694 unsigned int rr_nr_running;
695 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
697 int curr; /* highest queued rt task prio */
699 int next; /* next highest */
705 struct plist_head pushable_tasks;
707 #endif /* CONFIG_SMP */
713 /* Nests inside the rq lock: */
714 raw_spinlock_t rt_runtime_lock;
716 #ifdef CONFIG_RT_GROUP_SCHED
717 unsigned int rt_nr_boosted;
720 struct task_group *tg;
724 static inline bool rt_rq_is_runnable(struct rt_rq *rt_rq)
726 return rt_rq->rt_queued && rt_rq->rt_nr_running;
729 /* Deadline class' related fields in a runqueue */
731 /* runqueue is an rbtree, ordered by deadline */
732 struct rb_root_cached root;
734 unsigned int dl_nr_running;
738 * Deadline values of the currently executing and the
739 * earliest ready task on this rq. Caching these facilitates
740 * the decision whether or not a ready but not running task
741 * should migrate somewhere else.
751 * Tasks on this rq that can be pushed away. They are kept in
752 * an rb-tree, ordered by tasks' deadlines, with caching
753 * of the leftmost (earliest deadline) element.
755 struct rb_root_cached pushable_dl_tasks_root;
760 * "Active utilization" for this runqueue: increased when a
761 * task wakes up (becomes TASK_RUNNING) and decreased when a
767 * Utilization of the tasks "assigned" to this runqueue (including
768 * the tasks that are in runqueue and the tasks that executed on this
769 * CPU and blocked). Increased when a task moves to this runqueue, and
770 * decreased when the task moves away (migrates, changes scheduling
771 * policy, or terminates).
772 * This is needed to compute the "inactive utilization" for the
773 * runqueue (inactive utilization = this_bw - running_bw).
779 * Maximum available bandwidth for reclaiming by SCHED_FLAG_RECLAIM
780 * tasks of this rq. Used in calculation of reclaimable bandwidth(GRUB).
785 * Inverse of the fraction of CPU utilization that can be reclaimed
786 * by the GRUB algorithm.
791 #ifdef CONFIG_FAIR_GROUP_SCHED
792 /* An entity is a task if it doesn't "own" a runqueue */
793 #define entity_is_task(se) (!se->my_q)
795 static inline void se_update_runnable(struct sched_entity *se)
797 if (!entity_is_task(se))
798 se->runnable_weight = se->my_q->h_nr_running;
801 static inline long se_runnable(struct sched_entity *se)
803 if (entity_is_task(se))
806 return se->runnable_weight;
810 #define entity_is_task(se) 1
812 static inline void se_update_runnable(struct sched_entity *se) {}
814 static inline long se_runnable(struct sched_entity *se)
822 * XXX we want to get rid of these helpers and use the full load resolution.
824 static inline long se_weight(struct sched_entity *se)
826 return scale_load_down(se->load.weight);
830 static inline bool sched_asym_prefer(int a, int b)
832 return arch_asym_cpu_priority(a) > arch_asym_cpu_priority(b);
836 struct em_perf_domain *em_pd;
837 struct perf_domain *next;
841 /* Scheduling group status flags */
842 #define SG_OVERLOAD 0x1 /* More than one runnable task on a CPU. */
843 #define SG_OVERUTILIZED 0x2 /* One or more CPUs are over-utilized. */
846 * We add the notion of a root-domain which will be used to define per-domain
847 * variables. Each exclusive cpuset essentially defines an island domain by
848 * fully partitioning the member CPUs from any other cpuset. Whenever a new
849 * exclusive cpuset is created, we also create and attach a new root-domain
858 cpumask_var_t online;
861 * Indicate pullable load on at least one CPU, e.g:
862 * - More than one runnable task
863 * - Running task is misfit
867 /* Indicate one or more cpus over-utilized (tipping point) */
871 * The bit corresponding to a CPU gets set here if such CPU has more
872 * than one runnable -deadline task (as it is below for RT tasks).
874 cpumask_var_t dlo_mask;
880 * Indicate whether a root_domain's dl_bw has been checked or
881 * updated. It's monotonously increasing value.
883 * Also, some corner cases, like 'wrap around' is dangerous, but given
884 * that u64 is 'big enough'. So that shouldn't be a concern.
888 #ifdef HAVE_RT_PUSH_IPI
890 * For IPI pull requests, loop across the rto_mask.
892 struct irq_work rto_push_work;
893 raw_spinlock_t rto_lock;
894 /* These are only updated and read within rto_lock */
897 /* These atomics are updated outside of a lock */
898 atomic_t rto_loop_next;
899 atomic_t rto_loop_start;
902 * The "RT overload" flag: it gets set if a CPU has more than
903 * one runnable RT task.
905 cpumask_var_t rto_mask;
906 struct cpupri cpupri;
908 unsigned long max_cpu_capacity;
911 * NULL-terminated list of performance domains intersecting with the
912 * CPUs of the rd. Protected by RCU.
914 struct perf_domain __rcu *pd;
917 extern void init_defrootdomain(void);
918 extern int sched_init_domains(const struct cpumask *cpu_map);
919 extern void rq_attach_root(struct rq *rq, struct root_domain *rd);
920 extern void sched_get_rd(struct root_domain *rd);
921 extern void sched_put_rd(struct root_domain *rd);
923 #ifdef HAVE_RT_PUSH_IPI
924 extern void rto_push_irq_work_func(struct irq_work *work);
926 #endif /* CONFIG_SMP */
928 #ifdef CONFIG_UCLAMP_TASK
930 * struct uclamp_bucket - Utilization clamp bucket
931 * @value: utilization clamp value for tasks on this clamp bucket
932 * @tasks: number of RUNNABLE tasks on this clamp bucket
934 * Keep track of how many tasks are RUNNABLE for a given utilization
937 struct uclamp_bucket {
938 unsigned long value : bits_per(SCHED_CAPACITY_SCALE);
939 unsigned long tasks : BITS_PER_LONG - bits_per(SCHED_CAPACITY_SCALE);
943 * struct uclamp_rq - rq's utilization clamp
944 * @value: currently active clamp values for a rq
945 * @bucket: utilization clamp buckets affecting a rq
947 * Keep track of RUNNABLE tasks on a rq to aggregate their clamp values.
948 * A clamp value is affecting a rq when there is at least one task RUNNABLE
949 * (or actually running) with that value.
951 * There are up to UCLAMP_CNT possible different clamp values, currently there
952 * are only two: minimum utilization and maximum utilization.
954 * All utilization clamping values are MAX aggregated, since:
955 * - for util_min: we want to run the CPU at least at the max of the minimum
956 * utilization required by its currently RUNNABLE tasks.
957 * - for util_max: we want to allow the CPU to run up to the max of the
958 * maximum utilization allowed by its currently RUNNABLE tasks.
960 * Since on each system we expect only a limited number of different
961 * utilization clamp values (UCLAMP_BUCKETS), use a simple array to track
962 * the metrics required to compute all the per-rq utilization clamp values.
966 struct uclamp_bucket bucket[UCLAMP_BUCKETS];
969 DECLARE_STATIC_KEY_FALSE(sched_uclamp_used);
970 #endif /* CONFIG_UCLAMP_TASK */
973 struct balance_callback {
974 struct balance_callback *next;
975 void (*func)(struct rq *rq);
979 * This is the main, per-CPU runqueue data structure.
981 * Locking rule: those places that want to lock multiple runqueues
982 * (such as the load balancing or the thread migration code), lock
983 * acquire operations must be ordered by ascending &runqueue.
987 raw_spinlock_t __lock;
989 unsigned int nr_running;
990 #ifdef CONFIG_NUMA_BALANCING
991 unsigned int nr_numa_running;
992 unsigned int nr_preferred_running;
993 unsigned int numa_migrate_on;
995 #ifdef CONFIG_NO_HZ_COMMON
997 unsigned long last_blocked_load_update_tick;
998 unsigned int has_blocked_load;
999 call_single_data_t nohz_csd;
1000 #endif /* CONFIG_SMP */
1001 unsigned int nohz_tick_stopped;
1002 atomic_t nohz_flags;
1003 #endif /* CONFIG_NO_HZ_COMMON */
1006 unsigned int ttwu_pending;
1010 #ifdef CONFIG_UCLAMP_TASK
1011 /* Utilization clamp values based on CPU's RUNNABLE tasks */
1012 struct uclamp_rq uclamp[UCLAMP_CNT] ____cacheline_aligned;
1013 unsigned int uclamp_flags;
1014 #define UCLAMP_FLAG_IDLE 0x01
1021 #ifdef CONFIG_FAIR_GROUP_SCHED
1022 /* list of leaf cfs_rq on this CPU: */
1023 struct list_head leaf_cfs_rq_list;
1024 struct list_head *tmp_alone_branch;
1025 #endif /* CONFIG_FAIR_GROUP_SCHED */
1028 * This is part of a global counter where only the total sum
1029 * over all CPUs matters. A task can increase this counter on
1030 * one CPU and if it got migrated afterwards it may decrease
1031 * it on another CPU. Always updated under the runqueue lock:
1033 unsigned int nr_uninterruptible;
1035 struct task_struct __rcu *curr;
1036 struct task_struct *idle;
1037 struct task_struct *stop;
1038 unsigned long next_balance;
1039 struct mm_struct *prev_mm;
1041 unsigned int clock_update_flags;
1043 /* Ensure that all clocks are in the same cache line */
1044 u64 clock_task ____cacheline_aligned;
1046 unsigned long lost_idle_time;
1047 u64 clock_pelt_idle;
1049 #ifndef CONFIG_64BIT
1050 u64 clock_pelt_idle_copy;
1051 u64 clock_idle_copy;
1056 #ifdef CONFIG_SCHED_DEBUG
1057 u64 last_seen_need_resched_ns;
1058 int ticks_without_resched;
1061 #ifdef CONFIG_MEMBARRIER
1062 int membarrier_state;
1066 struct root_domain *rd;
1067 struct sched_domain __rcu *sd;
1069 unsigned long cpu_capacity;
1071 struct balance_callback *balance_callback;
1073 unsigned char nohz_idle_balance;
1074 unsigned char idle_balance;
1076 unsigned long misfit_task_load;
1078 /* For active balancing */
1081 struct cpu_stop_work active_balance_work;
1083 /* CPU of this runqueue: */
1087 struct list_head cfs_tasks;
1089 struct sched_avg avg_rt;
1090 struct sched_avg avg_dl;
1091 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
1092 struct sched_avg avg_irq;
1094 #ifdef CONFIG_SCHED_THERMAL_PRESSURE
1095 struct sched_avg avg_thermal;
1100 /* This is used to determine avg_idle's max value */
1101 u64 max_idle_balance_cost;
1103 #ifdef CONFIG_HOTPLUG_CPU
1104 struct rcuwait hotplug_wait;
1106 #endif /* CONFIG_SMP */
1108 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1111 #ifdef CONFIG_PARAVIRT
1112 u64 prev_steal_time;
1114 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1115 u64 prev_steal_time_rq;
1118 /* calc_load related fields */
1119 unsigned long calc_load_update;
1120 long calc_load_active;
1122 #ifdef CONFIG_SCHED_HRTICK
1124 call_single_data_t hrtick_csd;
1126 struct hrtimer hrtick_timer;
1127 ktime_t hrtick_time;
1130 #ifdef CONFIG_SCHEDSTATS
1132 struct sched_info rq_sched_info;
1133 unsigned long long rq_cpu_time;
1134 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
1136 /* sys_sched_yield() stats */
1137 unsigned int yld_count;
1139 /* schedule() stats */
1140 unsigned int sched_count;
1141 unsigned int sched_goidle;
1143 /* try_to_wake_up() stats */
1144 unsigned int ttwu_count;
1145 unsigned int ttwu_local;
1148 #ifdef CONFIG_CPU_IDLE
1149 /* Must be inspected within a rcu lock section */
1150 struct cpuidle_state *idle_state;
1154 unsigned int nr_pinned;
1156 unsigned int push_busy;
1157 struct cpu_stop_work push_work;
1159 #ifdef CONFIG_SCHED_CORE
1162 struct task_struct *core_pick;
1163 unsigned int core_enabled;
1164 unsigned int core_sched_seq;
1165 struct rb_root core_tree;
1167 /* shared state -- careful with sched_core_cpu_deactivate() */
1168 unsigned int core_task_seq;
1169 unsigned int core_pick_seq;
1170 unsigned long core_cookie;
1171 unsigned int core_forceidle_count;
1172 unsigned int core_forceidle_seq;
1173 unsigned int core_forceidle_occupation;
1174 u64 core_forceidle_start;
1177 /* Scratch cpumask to be temporarily used under rq_lock */
1178 cpumask_var_t scratch_mask;
1180 #if defined(CONFIG_CFS_BANDWIDTH) && defined(CONFIG_SMP)
1181 call_single_data_t cfsb_csd;
1182 struct list_head cfsb_csd_list;
1186 #ifdef CONFIG_FAIR_GROUP_SCHED
1188 /* CPU runqueue to which this cfs_rq is attached */
1189 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1196 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
1198 return container_of(cfs_rq, struct rq, cfs);
1202 static inline int cpu_of(struct rq *rq)
1211 #define MDF_PUSH 0x01
1213 static inline bool is_migration_disabled(struct task_struct *p)
1216 return p->migration_disabled;
1222 DECLARE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1224 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
1225 #define this_rq() this_cpu_ptr(&runqueues)
1226 #define task_rq(p) cpu_rq(task_cpu(p))
1227 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
1228 #define raw_rq() raw_cpu_ptr(&runqueues)
1231 #ifdef CONFIG_SCHED_CORE
1232 static inline struct cpumask *sched_group_span(struct sched_group *sg);
1234 DECLARE_STATIC_KEY_FALSE(__sched_core_enabled);
1236 static inline bool sched_core_enabled(struct rq *rq)
1238 return static_branch_unlikely(&__sched_core_enabled) && rq->core_enabled;
1241 static inline bool sched_core_disabled(void)
1243 return !static_branch_unlikely(&__sched_core_enabled);
1247 * Be careful with this function; not for general use. The return value isn't
1248 * stable unless you actually hold a relevant rq->__lock.
1250 static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1252 if (sched_core_enabled(rq))
1253 return &rq->core->__lock;
1258 static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1260 if (rq->core_enabled)
1261 return &rq->core->__lock;
1266 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
1268 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
1271 * Helpers to check if the CPU's core cookie matches with the task's cookie
1272 * when core scheduling is enabled.
1273 * A special case is that the task's cookie always matches with CPU's core
1274 * cookie if the CPU is in an idle core.
1276 static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1278 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1279 if (!sched_core_enabled(rq))
1282 return rq->core->core_cookie == p->core_cookie;
1285 static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1287 bool idle_core = true;
1290 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1291 if (!sched_core_enabled(rq))
1294 for_each_cpu(cpu, cpu_smt_mask(cpu_of(rq))) {
1295 if (!available_idle_cpu(cpu)) {
1302 * A CPU in an idle core is always the best choice for tasks with
1305 return idle_core || rq->core->core_cookie == p->core_cookie;
1308 static inline bool sched_group_cookie_match(struct rq *rq,
1309 struct task_struct *p,
1310 struct sched_group *group)
1314 /* Ignore cookie match if core scheduler is not enabled on the CPU. */
1315 if (!sched_core_enabled(rq))
1318 for_each_cpu_and(cpu, sched_group_span(group), p->cpus_ptr) {
1319 if (sched_core_cookie_match(cpu_rq(cpu), p))
1325 static inline bool sched_core_enqueued(struct task_struct *p)
1327 return !RB_EMPTY_NODE(&p->core_node);
1330 extern void sched_core_enqueue(struct rq *rq, struct task_struct *p);
1331 extern void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags);
1333 extern void sched_core_get(void);
1334 extern void sched_core_put(void);
1336 #else /* !CONFIG_SCHED_CORE */
1338 static inline bool sched_core_enabled(struct rq *rq)
1343 static inline bool sched_core_disabled(void)
1348 static inline raw_spinlock_t *rq_lockp(struct rq *rq)
1353 static inline raw_spinlock_t *__rq_lockp(struct rq *rq)
1358 static inline bool sched_cpu_cookie_match(struct rq *rq, struct task_struct *p)
1363 static inline bool sched_core_cookie_match(struct rq *rq, struct task_struct *p)
1368 static inline bool sched_group_cookie_match(struct rq *rq,
1369 struct task_struct *p,
1370 struct sched_group *group)
1374 #endif /* CONFIG_SCHED_CORE */
1376 static inline void lockdep_assert_rq_held(struct rq *rq)
1378 lockdep_assert_held(__rq_lockp(rq));
1381 extern void raw_spin_rq_lock_nested(struct rq *rq, int subclass);
1382 extern bool raw_spin_rq_trylock(struct rq *rq);
1383 extern void raw_spin_rq_unlock(struct rq *rq);
1385 static inline void raw_spin_rq_lock(struct rq *rq)
1387 raw_spin_rq_lock_nested(rq, 0);
1390 static inline void raw_spin_rq_lock_irq(struct rq *rq)
1392 local_irq_disable();
1393 raw_spin_rq_lock(rq);
1396 static inline void raw_spin_rq_unlock_irq(struct rq *rq)
1398 raw_spin_rq_unlock(rq);
1402 static inline unsigned long _raw_spin_rq_lock_irqsave(struct rq *rq)
1404 unsigned long flags;
1405 local_irq_save(flags);
1406 raw_spin_rq_lock(rq);
1410 static inline void raw_spin_rq_unlock_irqrestore(struct rq *rq, unsigned long flags)
1412 raw_spin_rq_unlock(rq);
1413 local_irq_restore(flags);
1416 #define raw_spin_rq_lock_irqsave(rq, flags) \
1418 flags = _raw_spin_rq_lock_irqsave(rq); \
1421 #ifdef CONFIG_SCHED_SMT
1422 extern void __update_idle_core(struct rq *rq);
1424 static inline void update_idle_core(struct rq *rq)
1426 if (static_branch_unlikely(&sched_smt_present))
1427 __update_idle_core(rq);
1431 static inline void update_idle_core(struct rq *rq) { }
1434 #ifdef CONFIG_FAIR_GROUP_SCHED
1435 static inline struct task_struct *task_of(struct sched_entity *se)
1437 SCHED_WARN_ON(!entity_is_task(se));
1438 return container_of(se, struct task_struct, se);
1441 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
1443 return p->se.cfs_rq;
1446 /* runqueue on which this entity is (to be) queued */
1447 static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1452 /* runqueue "owned" by this group */
1453 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1460 #define task_of(_se) container_of(_se, struct task_struct, se)
1462 static inline struct cfs_rq *task_cfs_rq(const struct task_struct *p)
1464 return &task_rq(p)->cfs;
1467 static inline struct cfs_rq *cfs_rq_of(const struct sched_entity *se)
1469 const struct task_struct *p = task_of(se);
1470 struct rq *rq = task_rq(p);
1475 /* runqueue "owned" by this group */
1476 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
1482 extern void update_rq_clock(struct rq *rq);
1485 * rq::clock_update_flags bits
1487 * %RQCF_REQ_SKIP - will request skipping of clock update on the next
1488 * call to __schedule(). This is an optimisation to avoid
1489 * neighbouring rq clock updates.
1491 * %RQCF_ACT_SKIP - is set from inside of __schedule() when skipping is
1492 * in effect and calls to update_rq_clock() are being ignored.
1494 * %RQCF_UPDATED - is a debug flag that indicates whether a call has been
1495 * made to update_rq_clock() since the last time rq::lock was pinned.
1497 * If inside of __schedule(), clock_update_flags will have been
1498 * shifted left (a left shift is a cheap operation for the fast path
1499 * to promote %RQCF_REQ_SKIP to %RQCF_ACT_SKIP), so you must use,
1501 * if (rq-clock_update_flags >= RQCF_UPDATED)
1503 * to check if %RQCF_UPDATED is set. It'll never be shifted more than
1504 * one position though, because the next rq_unpin_lock() will shift it
1507 #define RQCF_REQ_SKIP 0x01
1508 #define RQCF_ACT_SKIP 0x02
1509 #define RQCF_UPDATED 0x04
1511 static inline void assert_clock_updated(struct rq *rq)
1514 * The only reason for not seeing a clock update since the
1515 * last rq_pin_lock() is if we're currently skipping updates.
1517 SCHED_WARN_ON(rq->clock_update_flags < RQCF_ACT_SKIP);
1520 static inline u64 rq_clock(struct rq *rq)
1522 lockdep_assert_rq_held(rq);
1523 assert_clock_updated(rq);
1528 static inline u64 rq_clock_task(struct rq *rq)
1530 lockdep_assert_rq_held(rq);
1531 assert_clock_updated(rq);
1533 return rq->clock_task;
1537 * By default the decay is the default pelt decay period.
1538 * The decay shift can change the decay period in
1540 * Decay shift Decay period(ms)
1547 extern int sched_thermal_decay_shift;
1549 static inline u64 rq_clock_thermal(struct rq *rq)
1551 return rq_clock_task(rq) >> sched_thermal_decay_shift;
1554 static inline void rq_clock_skip_update(struct rq *rq)
1556 lockdep_assert_rq_held(rq);
1557 rq->clock_update_flags |= RQCF_REQ_SKIP;
1561 * See rt task throttling, which is the only time a skip
1562 * request is canceled.
1564 static inline void rq_clock_cancel_skipupdate(struct rq *rq)
1566 lockdep_assert_rq_held(rq);
1567 rq->clock_update_flags &= ~RQCF_REQ_SKIP;
1571 * During cpu offlining and rq wide unthrottling, we can trigger
1572 * an update_rq_clock() for several cfs and rt runqueues (Typically
1573 * when using list_for_each_entry_*)
1574 * rq_clock_start_loop_update() can be called after updating the clock
1575 * once and before iterating over the list to prevent multiple update.
1576 * After the iterative traversal, we need to call rq_clock_stop_loop_update()
1577 * to clear RQCF_ACT_SKIP of rq->clock_update_flags.
1579 static inline void rq_clock_start_loop_update(struct rq *rq)
1581 lockdep_assert_rq_held(rq);
1582 SCHED_WARN_ON(rq->clock_update_flags & RQCF_ACT_SKIP);
1583 rq->clock_update_flags |= RQCF_ACT_SKIP;
1586 static inline void rq_clock_stop_loop_update(struct rq *rq)
1588 lockdep_assert_rq_held(rq);
1589 rq->clock_update_flags &= ~RQCF_ACT_SKIP;
1593 unsigned long flags;
1594 struct pin_cookie cookie;
1595 #ifdef CONFIG_SCHED_DEBUG
1597 * A copy of (rq::clock_update_flags & RQCF_UPDATED) for the
1598 * current pin context is stashed here in case it needs to be
1599 * restored in rq_repin_lock().
1601 unsigned int clock_update_flags;
1605 extern struct balance_callback balance_push_callback;
1608 * Lockdep annotation that avoids accidental unlocks; it's like a
1609 * sticky/continuous lockdep_assert_held().
1611 * This avoids code that has access to 'struct rq *rq' (basically everything in
1612 * the scheduler) from accidentally unlocking the rq if they do not also have a
1613 * copy of the (on-stack) 'struct rq_flags rf'.
1615 * Also see Documentation/locking/lockdep-design.rst.
1617 static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
1619 rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
1621 #ifdef CONFIG_SCHED_DEBUG
1622 rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
1623 rf->clock_update_flags = 0;
1625 SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
1630 static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
1632 #ifdef CONFIG_SCHED_DEBUG
1633 if (rq->clock_update_flags > RQCF_ACT_SKIP)
1634 rf->clock_update_flags = RQCF_UPDATED;
1637 lockdep_unpin_lock(__rq_lockp(rq), rf->cookie);
1640 static inline void rq_repin_lock(struct rq *rq, struct rq_flags *rf)
1642 lockdep_repin_lock(__rq_lockp(rq), rf->cookie);
1644 #ifdef CONFIG_SCHED_DEBUG
1646 * Restore the value we stashed in @rf for this pin context.
1648 rq->clock_update_flags |= rf->clock_update_flags;
1652 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1653 __acquires(rq->lock);
1655 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
1656 __acquires(p->pi_lock)
1657 __acquires(rq->lock);
1659 static inline void __task_rq_unlock(struct rq *rq, struct rq_flags *rf)
1660 __releases(rq->lock)
1662 rq_unpin_lock(rq, rf);
1663 raw_spin_rq_unlock(rq);
1667 task_rq_unlock(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
1668 __releases(rq->lock)
1669 __releases(p->pi_lock)
1671 rq_unpin_lock(rq, rf);
1672 raw_spin_rq_unlock(rq);
1673 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
1676 DEFINE_LOCK_GUARD_1(task_rq_lock, struct task_struct,
1677 _T->rq = task_rq_lock(_T->lock, &_T->rf),
1678 task_rq_unlock(_T->rq, _T->lock, &_T->rf),
1679 struct rq *rq; struct rq_flags rf)
1682 rq_lock_irqsave(struct rq *rq, struct rq_flags *rf)
1683 __acquires(rq->lock)
1685 raw_spin_rq_lock_irqsave(rq, rf->flags);
1686 rq_pin_lock(rq, rf);
1690 rq_lock_irq(struct rq *rq, struct rq_flags *rf)
1691 __acquires(rq->lock)
1693 raw_spin_rq_lock_irq(rq);
1694 rq_pin_lock(rq, rf);
1698 rq_lock(struct rq *rq, struct rq_flags *rf)
1699 __acquires(rq->lock)
1701 raw_spin_rq_lock(rq);
1702 rq_pin_lock(rq, rf);
1706 rq_unlock_irqrestore(struct rq *rq, struct rq_flags *rf)
1707 __releases(rq->lock)
1709 rq_unpin_lock(rq, rf);
1710 raw_spin_rq_unlock_irqrestore(rq, rf->flags);
1714 rq_unlock_irq(struct rq *rq, struct rq_flags *rf)
1715 __releases(rq->lock)
1717 rq_unpin_lock(rq, rf);
1718 raw_spin_rq_unlock_irq(rq);
1722 rq_unlock(struct rq *rq, struct rq_flags *rf)
1723 __releases(rq->lock)
1725 rq_unpin_lock(rq, rf);
1726 raw_spin_rq_unlock(rq);
1729 DEFINE_LOCK_GUARD_1(rq_lock, struct rq,
1730 rq_lock(_T->lock, &_T->rf),
1731 rq_unlock(_T->lock, &_T->rf),
1734 DEFINE_LOCK_GUARD_1(rq_lock_irq, struct rq,
1735 rq_lock_irq(_T->lock, &_T->rf),
1736 rq_unlock_irq(_T->lock, &_T->rf),
1739 DEFINE_LOCK_GUARD_1(rq_lock_irqsave, struct rq,
1740 rq_lock_irqsave(_T->lock, &_T->rf),
1741 rq_unlock_irqrestore(_T->lock, &_T->rf),
1744 static inline struct rq *
1745 this_rq_lock_irq(struct rq_flags *rf)
1746 __acquires(rq->lock)
1750 local_irq_disable();
1757 enum numa_topology_type {
1762 extern enum numa_topology_type sched_numa_topology_type;
1763 extern int sched_max_numa_distance;
1764 extern bool find_numa_distance(int distance);
1765 extern void sched_init_numa(int offline_node);
1766 extern void sched_update_numa(int cpu, bool online);
1767 extern void sched_domains_numa_masks_set(unsigned int cpu);
1768 extern void sched_domains_numa_masks_clear(unsigned int cpu);
1769 extern int sched_numa_find_closest(const struct cpumask *cpus, int cpu);
1771 static inline void sched_init_numa(int offline_node) { }
1772 static inline void sched_update_numa(int cpu, bool online) { }
1773 static inline void sched_domains_numa_masks_set(unsigned int cpu) { }
1774 static inline void sched_domains_numa_masks_clear(unsigned int cpu) { }
1775 static inline int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1781 #ifdef CONFIG_NUMA_BALANCING
1782 /* The regions in numa_faults array from task_struct */
1783 enum numa_faults_stats {
1789 extern void sched_setnuma(struct task_struct *p, int node);
1790 extern int migrate_task_to(struct task_struct *p, int cpu);
1791 extern int migrate_swap(struct task_struct *p, struct task_struct *t,
1793 extern void init_numa_balancing(unsigned long clone_flags, struct task_struct *p);
1796 init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1799 #endif /* CONFIG_NUMA_BALANCING */
1804 queue_balance_callback(struct rq *rq,
1805 struct balance_callback *head,
1806 void (*func)(struct rq *rq))
1808 lockdep_assert_rq_held(rq);
1811 * Don't (re)queue an already queued item; nor queue anything when
1812 * balance_push() is active, see the comment with
1813 * balance_push_callback.
1815 if (unlikely(head->next || rq->balance_callback == &balance_push_callback))
1819 head->next = rq->balance_callback;
1820 rq->balance_callback = head;
1823 #define rcu_dereference_check_sched_domain(p) \
1824 rcu_dereference_check((p), \
1825 lockdep_is_held(&sched_domains_mutex))
1828 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1829 * See destroy_sched_domains: call_rcu for details.
1831 * The domain tree of any CPU may only be accessed from within
1832 * preempt-disabled sections.
1834 #define for_each_domain(cpu, __sd) \
1835 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
1836 __sd; __sd = __sd->parent)
1838 /* A mask of all the SD flags that have the SDF_SHARED_CHILD metaflag */
1839 #define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_SHARED_CHILD)) |
1840 static const unsigned int SD_SHARED_CHILD_MASK =
1841 #include <linux/sched/sd_flags.h>
1846 * highest_flag_domain - Return highest sched_domain containing flag.
1847 * @cpu: The CPU whose highest level of sched domain is to
1849 * @flag: The flag to check for the highest sched_domain
1850 * for the given CPU.
1852 * Returns the highest sched_domain of a CPU which contains @flag. If @flag has
1853 * the SDF_SHARED_CHILD metaflag, all the children domains also have @flag.
1855 static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
1857 struct sched_domain *sd, *hsd = NULL;
1859 for_each_domain(cpu, sd) {
1860 if (sd->flags & flag) {
1866 * Stop the search if @flag is known to be shared at lower
1867 * levels. It will not be found further up.
1869 if (flag & SD_SHARED_CHILD_MASK)
1876 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
1878 struct sched_domain *sd;
1880 for_each_domain(cpu, sd) {
1881 if (sd->flags & flag)
1888 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_llc);
1889 DECLARE_PER_CPU(int, sd_llc_size);
1890 DECLARE_PER_CPU(int, sd_llc_id);
1891 DECLARE_PER_CPU(int, sd_share_id);
1892 DECLARE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
1893 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_numa);
1894 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
1895 DECLARE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
1896 extern struct static_key_false sched_asym_cpucapacity;
1897 extern struct static_key_false sched_cluster_active;
1899 static __always_inline bool sched_asym_cpucap_active(void)
1901 return static_branch_unlikely(&sched_asym_cpucapacity);
1904 struct sched_group_capacity {
1907 * CPU capacity of this group, SCHED_CAPACITY_SCALE being max capacity
1910 unsigned long capacity;
1911 unsigned long min_capacity; /* Min per-CPU capacity in group */
1912 unsigned long max_capacity; /* Max per-CPU capacity in group */
1913 unsigned long next_update;
1914 int imbalance; /* XXX unrelated to capacity but shared group state */
1916 #ifdef CONFIG_SCHED_DEBUG
1920 unsigned long cpumask[]; /* Balance mask */
1923 struct sched_group {
1924 struct sched_group *next; /* Must be a circular list */
1927 unsigned int group_weight;
1929 struct sched_group_capacity *sgc;
1930 int asym_prefer_cpu; /* CPU of highest priority in group */
1934 * The CPUs this group covers.
1936 * NOTE: this field is variable length. (Allocated dynamically
1937 * by attaching extra space to the end of the structure,
1938 * depending on how many CPUs the kernel has booted up with)
1940 unsigned long cpumask[];
1943 static inline struct cpumask *sched_group_span(struct sched_group *sg)
1945 return to_cpumask(sg->cpumask);
1949 * See build_balance_mask().
1951 static inline struct cpumask *group_balance_mask(struct sched_group *sg)
1953 return to_cpumask(sg->sgc->cpumask);
1956 extern int group_balance_cpu(struct sched_group *sg);
1958 #ifdef CONFIG_SCHED_DEBUG
1959 void update_sched_domain_debugfs(void);
1960 void dirty_sched_domain_sysctl(int cpu);
1962 static inline void update_sched_domain_debugfs(void)
1965 static inline void dirty_sched_domain_sysctl(int cpu)
1970 extern int sched_update_scaling(void);
1972 static inline const struct cpumask *task_user_cpus(struct task_struct *p)
1974 if (!p->user_cpus_ptr)
1975 return cpu_possible_mask; /* &init_task.cpus_mask */
1976 return p->user_cpus_ptr;
1978 #endif /* CONFIG_SMP */
1982 #if defined(CONFIG_SCHED_CORE) && defined(CONFIG_SCHEDSTATS)
1984 extern void __sched_core_account_forceidle(struct rq *rq);
1986 static inline void sched_core_account_forceidle(struct rq *rq)
1988 if (schedstat_enabled())
1989 __sched_core_account_forceidle(rq);
1992 extern void __sched_core_tick(struct rq *rq);
1994 static inline void sched_core_tick(struct rq *rq)
1996 if (sched_core_enabled(rq) && schedstat_enabled())
1997 __sched_core_tick(rq);
2002 static inline void sched_core_account_forceidle(struct rq *rq) {}
2004 static inline void sched_core_tick(struct rq *rq) {}
2006 #endif /* CONFIG_SCHED_CORE && CONFIG_SCHEDSTATS */
2008 #ifdef CONFIG_CGROUP_SCHED
2011 * Return the group to which this tasks belongs.
2013 * We cannot use task_css() and friends because the cgroup subsystem
2014 * changes that value before the cgroup_subsys::attach() method is called,
2015 * therefore we cannot pin it and might observe the wrong value.
2017 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
2018 * core changes this before calling sched_move_task().
2020 * Instead we use a 'copy' which is updated from sched_move_task() while
2021 * holding both task_struct::pi_lock and rq::lock.
2023 static inline struct task_group *task_group(struct task_struct *p)
2025 return p->sched_task_group;
2028 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
2029 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
2031 #if defined(CONFIG_FAIR_GROUP_SCHED) || defined(CONFIG_RT_GROUP_SCHED)
2032 struct task_group *tg = task_group(p);
2035 #ifdef CONFIG_FAIR_GROUP_SCHED
2036 set_task_rq_fair(&p->se, p->se.cfs_rq, tg->cfs_rq[cpu]);
2037 p->se.cfs_rq = tg->cfs_rq[cpu];
2038 p->se.parent = tg->se[cpu];
2039 p->se.depth = tg->se[cpu] ? tg->se[cpu]->depth + 1 : 0;
2042 #ifdef CONFIG_RT_GROUP_SCHED
2043 p->rt.rt_rq = tg->rt_rq[cpu];
2044 p->rt.parent = tg->rt_se[cpu];
2048 #else /* CONFIG_CGROUP_SCHED */
2050 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
2051 static inline struct task_group *task_group(struct task_struct *p)
2056 #endif /* CONFIG_CGROUP_SCHED */
2058 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
2060 set_task_rq(p, cpu);
2063 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2064 * successfully executed on another CPU. We must ensure that updates of
2065 * per-task data have been completed by this moment.
2068 WRITE_ONCE(task_thread_info(p)->cpu, cpu);
2074 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
2076 #ifdef CONFIG_SCHED_DEBUG
2077 # define const_debug __read_mostly
2079 # define const_debug const
2082 #define SCHED_FEAT(name, enabled) \
2083 __SCHED_FEAT_##name ,
2086 #include "features.h"
2092 #ifdef CONFIG_SCHED_DEBUG
2095 * To support run-time toggling of sched features, all the translation units
2096 * (but core.c) reference the sysctl_sched_features defined in core.c.
2098 extern const_debug unsigned int sysctl_sched_features;
2100 #ifdef CONFIG_JUMP_LABEL
2101 #define SCHED_FEAT(name, enabled) \
2102 static __always_inline bool static_branch_##name(struct static_key *key) \
2104 return static_key_##enabled(key); \
2107 #include "features.h"
2110 extern struct static_key sched_feat_keys[__SCHED_FEAT_NR];
2111 #define sched_feat(x) (static_branch_##x(&sched_feat_keys[__SCHED_FEAT_##x]))
2113 #else /* !CONFIG_JUMP_LABEL */
2115 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2117 #endif /* CONFIG_JUMP_LABEL */
2119 #else /* !SCHED_DEBUG */
2122 * Each translation unit has its own copy of sysctl_sched_features to allow
2123 * constants propagation at compile time and compiler optimization based on
2126 #define SCHED_FEAT(name, enabled) \
2127 (1UL << __SCHED_FEAT_##name) * enabled |
2128 static const_debug __maybe_unused unsigned int sysctl_sched_features =
2129 #include "features.h"
2133 #define sched_feat(x) !!(sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
2135 #endif /* SCHED_DEBUG */
2137 extern struct static_key_false sched_numa_balancing;
2138 extern struct static_key_false sched_schedstats;
2140 static inline u64 global_rt_period(void)
2142 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
2145 static inline u64 global_rt_runtime(void)
2147 if (sysctl_sched_rt_runtime < 0)
2150 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
2153 static inline int task_current(struct rq *rq, struct task_struct *p)
2155 return rq->curr == p;
2158 static inline int task_on_cpu(struct rq *rq, struct task_struct *p)
2163 return task_current(rq, p);
2167 static inline int task_on_rq_queued(struct task_struct *p)
2169 return p->on_rq == TASK_ON_RQ_QUEUED;
2172 static inline int task_on_rq_migrating(struct task_struct *p)
2174 return READ_ONCE(p->on_rq) == TASK_ON_RQ_MIGRATING;
2177 /* Wake flags. The first three directly map to some SD flag value */
2178 #define WF_EXEC 0x02 /* Wakeup after exec; maps to SD_BALANCE_EXEC */
2179 #define WF_FORK 0x04 /* Wakeup after fork; maps to SD_BALANCE_FORK */
2180 #define WF_TTWU 0x08 /* Wakeup; maps to SD_BALANCE_WAKE */
2182 #define WF_SYNC 0x10 /* Waker goes to sleep after wakeup */
2183 #define WF_MIGRATED 0x20 /* Internal use, task got migrated */
2184 #define WF_CURRENT_CPU 0x40 /* Prefer to move the wakee to the current CPU. */
2187 static_assert(WF_EXEC == SD_BALANCE_EXEC);
2188 static_assert(WF_FORK == SD_BALANCE_FORK);
2189 static_assert(WF_TTWU == SD_BALANCE_WAKE);
2193 * To aid in avoiding the subversion of "niceness" due to uneven distribution
2194 * of tasks with abnormal "nice" values across CPUs the contribution that
2195 * each task makes to its run queue's load is weighted according to its
2196 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
2197 * scaled version of the new time slice allocation that they receive on time
2201 #define WEIGHT_IDLEPRIO 3
2202 #define WMULT_IDLEPRIO 1431655765
2204 extern const int sched_prio_to_weight[40];
2205 extern const u32 sched_prio_to_wmult[40];
2208 * {de,en}queue flags:
2210 * DEQUEUE_SLEEP - task is no longer runnable
2211 * ENQUEUE_WAKEUP - task just became runnable
2213 * SAVE/RESTORE - an otherwise spurious dequeue/enqueue, done to ensure tasks
2214 * are in a known state which allows modification. Such pairs
2215 * should preserve as much state as possible.
2217 * MOVE - paired with SAVE/RESTORE, explicitly does not preserve the location
2220 * NOCLOCK - skip the update_rq_clock() (avoids double updates)
2222 * MIGRATION - p->on_rq == TASK_ON_RQ_MIGRATING (used for DEADLINE)
2224 * ENQUEUE_HEAD - place at front of runqueue (tail if not specified)
2225 * ENQUEUE_REPLENISH - CBS (replenish runtime and postpone deadline)
2226 * ENQUEUE_MIGRATED - the task was migrated during wakeup
2230 #define DEQUEUE_SLEEP 0x01
2231 #define DEQUEUE_SAVE 0x02 /* Matches ENQUEUE_RESTORE */
2232 #define DEQUEUE_MOVE 0x04 /* Matches ENQUEUE_MOVE */
2233 #define DEQUEUE_NOCLOCK 0x08 /* Matches ENQUEUE_NOCLOCK */
2234 #define DEQUEUE_MIGRATING 0x100 /* Matches ENQUEUE_MIGRATING */
2236 #define ENQUEUE_WAKEUP 0x01
2237 #define ENQUEUE_RESTORE 0x02
2238 #define ENQUEUE_MOVE 0x04
2239 #define ENQUEUE_NOCLOCK 0x08
2241 #define ENQUEUE_HEAD 0x10
2242 #define ENQUEUE_REPLENISH 0x20
2244 #define ENQUEUE_MIGRATED 0x40
2246 #define ENQUEUE_MIGRATED 0x00
2248 #define ENQUEUE_INITIAL 0x80
2249 #define ENQUEUE_MIGRATING 0x100
2251 #define RETRY_TASK ((void *)-1UL)
2253 struct affinity_context {
2254 const struct cpumask *new_mask;
2255 struct cpumask *user_mask;
2259 extern s64 update_curr_common(struct rq *rq);
2261 struct sched_class {
2263 #ifdef CONFIG_UCLAMP_TASK
2267 void (*enqueue_task) (struct rq *rq, struct task_struct *p, int flags);
2268 void (*dequeue_task) (struct rq *rq, struct task_struct *p, int flags);
2269 void (*yield_task) (struct rq *rq);
2270 bool (*yield_to_task)(struct rq *rq, struct task_struct *p);
2272 void (*wakeup_preempt)(struct rq *rq, struct task_struct *p, int flags);
2274 struct task_struct *(*pick_next_task)(struct rq *rq);
2276 void (*put_prev_task)(struct rq *rq, struct task_struct *p);
2277 void (*set_next_task)(struct rq *rq, struct task_struct *p, bool first);
2280 int (*balance)(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2281 int (*select_task_rq)(struct task_struct *p, int task_cpu, int flags);
2283 struct task_struct * (*pick_task)(struct rq *rq);
2285 void (*migrate_task_rq)(struct task_struct *p, int new_cpu);
2287 void (*task_woken)(struct rq *this_rq, struct task_struct *task);
2289 void (*set_cpus_allowed)(struct task_struct *p, struct affinity_context *ctx);
2291 void (*rq_online)(struct rq *rq);
2292 void (*rq_offline)(struct rq *rq);
2294 struct rq *(*find_lock_rq)(struct task_struct *p, struct rq *rq);
2297 void (*task_tick)(struct rq *rq, struct task_struct *p, int queued);
2298 void (*task_fork)(struct task_struct *p);
2299 void (*task_dead)(struct task_struct *p);
2302 * The switched_from() call is allowed to drop rq->lock, therefore we
2303 * cannot assume the switched_from/switched_to pair is serialized by
2304 * rq->lock. They are however serialized by p->pi_lock.
2306 void (*switched_from)(struct rq *this_rq, struct task_struct *task);
2307 void (*switched_to) (struct rq *this_rq, struct task_struct *task);
2308 void (*prio_changed) (struct rq *this_rq, struct task_struct *task,
2311 unsigned int (*get_rr_interval)(struct rq *rq,
2312 struct task_struct *task);
2314 void (*update_curr)(struct rq *rq);
2316 #ifdef CONFIG_FAIR_GROUP_SCHED
2317 void (*task_change_group)(struct task_struct *p);
2320 #ifdef CONFIG_SCHED_CORE
2321 int (*task_is_throttled)(struct task_struct *p, int cpu);
2325 static inline void put_prev_task(struct rq *rq, struct task_struct *prev)
2327 WARN_ON_ONCE(rq->curr != prev);
2328 prev->sched_class->put_prev_task(rq, prev);
2331 static inline void set_next_task(struct rq *rq, struct task_struct *next)
2333 next->sched_class->set_next_task(rq, next, false);
2338 * Helper to define a sched_class instance; each one is placed in a separate
2339 * section which is ordered by the linker script:
2341 * include/asm-generic/vmlinux.lds.h
2343 * *CAREFUL* they are laid out in *REVERSE* order!!!
2345 * Also enforce alignment on the instance, not the type, to guarantee layout.
2347 #define DEFINE_SCHED_CLASS(name) \
2348 const struct sched_class name##_sched_class \
2349 __aligned(__alignof__(struct sched_class)) \
2350 __section("__" #name "_sched_class")
2352 /* Defined in include/asm-generic/vmlinux.lds.h */
2353 extern struct sched_class __sched_class_highest[];
2354 extern struct sched_class __sched_class_lowest[];
2356 #define for_class_range(class, _from, _to) \
2357 for (class = (_from); class < (_to); class++)
2359 #define for_each_class(class) \
2360 for_class_range(class, __sched_class_highest, __sched_class_lowest)
2362 #define sched_class_above(_a, _b) ((_a) < (_b))
2364 extern const struct sched_class stop_sched_class;
2365 extern const struct sched_class dl_sched_class;
2366 extern const struct sched_class rt_sched_class;
2367 extern const struct sched_class fair_sched_class;
2368 extern const struct sched_class idle_sched_class;
2370 static inline bool sched_stop_runnable(struct rq *rq)
2372 return rq->stop && task_on_rq_queued(rq->stop);
2375 static inline bool sched_dl_runnable(struct rq *rq)
2377 return rq->dl.dl_nr_running > 0;
2380 static inline bool sched_rt_runnable(struct rq *rq)
2382 return rq->rt.rt_queued > 0;
2385 static inline bool sched_fair_runnable(struct rq *rq)
2387 return rq->cfs.nr_running > 0;
2390 extern struct task_struct *pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf);
2391 extern struct task_struct *pick_next_task_idle(struct rq *rq);
2393 #define SCA_CHECK 0x01
2394 #define SCA_MIGRATE_DISABLE 0x02
2395 #define SCA_MIGRATE_ENABLE 0x04
2396 #define SCA_USER 0x08
2400 extern void update_group_capacity(struct sched_domain *sd, int cpu);
2402 extern void trigger_load_balance(struct rq *rq);
2404 extern void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx);
2406 static inline struct task_struct *get_push_task(struct rq *rq)
2408 struct task_struct *p = rq->curr;
2410 lockdep_assert_rq_held(rq);
2415 if (p->nr_cpus_allowed == 1)
2418 if (p->migration_disabled)
2421 rq->push_busy = true;
2422 return get_task_struct(p);
2425 extern int push_cpu_stop(void *arg);
2429 #ifdef CONFIG_CPU_IDLE
2430 static inline void idle_set_state(struct rq *rq,
2431 struct cpuidle_state *idle_state)
2433 rq->idle_state = idle_state;
2436 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2438 SCHED_WARN_ON(!rcu_read_lock_held());
2440 return rq->idle_state;
2443 static inline void idle_set_state(struct rq *rq,
2444 struct cpuidle_state *idle_state)
2448 static inline struct cpuidle_state *idle_get_state(struct rq *rq)
2454 extern void schedule_idle(void);
2455 asmlinkage void schedule_user(void);
2457 extern void sysrq_sched_debug_show(void);
2458 extern void sched_init_granularity(void);
2459 extern void update_max_interval(void);
2461 extern void init_sched_dl_class(void);
2462 extern void init_sched_rt_class(void);
2463 extern void init_sched_fair_class(void);
2465 extern void reweight_task(struct task_struct *p, int prio);
2467 extern void resched_curr(struct rq *rq);
2468 extern void resched_cpu(int cpu);
2470 extern struct rt_bandwidth def_rt_bandwidth;
2471 extern void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime);
2472 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
2474 extern void init_dl_entity(struct sched_dl_entity *dl_se);
2477 #define BW_UNIT (1 << BW_SHIFT)
2478 #define RATIO_SHIFT 8
2479 #define MAX_BW_BITS (64 - BW_SHIFT)
2480 #define MAX_BW ((1ULL << MAX_BW_BITS) - 1)
2481 unsigned long to_ratio(u64 period, u64 runtime);
2483 extern void init_entity_runnable_average(struct sched_entity *se);
2484 extern void post_init_entity_util_avg(struct task_struct *p);
2486 #ifdef CONFIG_NO_HZ_FULL
2487 extern bool sched_can_stop_tick(struct rq *rq);
2488 extern int __init sched_tick_offload_init(void);
2491 * Tick may be needed by tasks in the runqueue depending on their policy and
2492 * requirements. If tick is needed, lets send the target an IPI to kick it out of
2493 * nohz mode if necessary.
2495 static inline void sched_update_tick_dependency(struct rq *rq)
2497 int cpu = cpu_of(rq);
2499 if (!tick_nohz_full_cpu(cpu))
2502 if (sched_can_stop_tick(rq))
2503 tick_nohz_dep_clear_cpu(cpu, TICK_DEP_BIT_SCHED);
2505 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
2508 static inline int sched_tick_offload_init(void) { return 0; }
2509 static inline void sched_update_tick_dependency(struct rq *rq) { }
2512 static inline void add_nr_running(struct rq *rq, unsigned count)
2514 unsigned prev_nr = rq->nr_running;
2516 rq->nr_running = prev_nr + count;
2517 if (trace_sched_update_nr_running_tp_enabled()) {
2518 call_trace_sched_update_nr_running(rq, count);
2522 if (prev_nr < 2 && rq->nr_running >= 2) {
2523 if (!READ_ONCE(rq->rd->overload))
2524 WRITE_ONCE(rq->rd->overload, 1);
2528 sched_update_tick_dependency(rq);
2531 static inline void sub_nr_running(struct rq *rq, unsigned count)
2533 rq->nr_running -= count;
2534 if (trace_sched_update_nr_running_tp_enabled()) {
2535 call_trace_sched_update_nr_running(rq, -count);
2538 /* Check if we still need preemption */
2539 sched_update_tick_dependency(rq);
2542 extern void activate_task(struct rq *rq, struct task_struct *p, int flags);
2543 extern void deactivate_task(struct rq *rq, struct task_struct *p, int flags);
2545 extern void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags);
2547 #ifdef CONFIG_PREEMPT_RT
2548 #define SCHED_NR_MIGRATE_BREAK 8
2550 #define SCHED_NR_MIGRATE_BREAK 32
2553 extern const_debug unsigned int sysctl_sched_nr_migrate;
2554 extern const_debug unsigned int sysctl_sched_migration_cost;
2556 extern unsigned int sysctl_sched_base_slice;
2558 #ifdef CONFIG_SCHED_DEBUG
2559 extern int sysctl_resched_latency_warn_ms;
2560 extern int sysctl_resched_latency_warn_once;
2562 extern unsigned int sysctl_sched_tunable_scaling;
2564 extern unsigned int sysctl_numa_balancing_scan_delay;
2565 extern unsigned int sysctl_numa_balancing_scan_period_min;
2566 extern unsigned int sysctl_numa_balancing_scan_period_max;
2567 extern unsigned int sysctl_numa_balancing_scan_size;
2568 extern unsigned int sysctl_numa_balancing_hot_threshold;
2571 #ifdef CONFIG_SCHED_HRTICK
2575 * - enabled by features
2576 * - hrtimer is actually high res
2578 static inline int hrtick_enabled(struct rq *rq)
2580 if (!cpu_active(cpu_of(rq)))
2582 return hrtimer_is_hres_active(&rq->hrtick_timer);
2585 static inline int hrtick_enabled_fair(struct rq *rq)
2587 if (!sched_feat(HRTICK))
2589 return hrtick_enabled(rq);
2592 static inline int hrtick_enabled_dl(struct rq *rq)
2594 if (!sched_feat(HRTICK_DL))
2596 return hrtick_enabled(rq);
2599 void hrtick_start(struct rq *rq, u64 delay);
2603 static inline int hrtick_enabled_fair(struct rq *rq)
2608 static inline int hrtick_enabled_dl(struct rq *rq)
2613 static inline int hrtick_enabled(struct rq *rq)
2618 #endif /* CONFIG_SCHED_HRTICK */
2620 #ifndef arch_scale_freq_tick
2621 static __always_inline
2622 void arch_scale_freq_tick(void)
2627 #ifndef arch_scale_freq_capacity
2629 * arch_scale_freq_capacity - get the frequency scale factor of a given CPU.
2630 * @cpu: the CPU in question.
2632 * Return: the frequency scale factor normalized against SCHED_CAPACITY_SCALE, i.e.
2635 * ------ * SCHED_CAPACITY_SCALE
2638 static __always_inline
2639 unsigned long arch_scale_freq_capacity(int cpu)
2641 return SCHED_CAPACITY_SCALE;
2645 #ifdef CONFIG_SCHED_DEBUG
2647 * In double_lock_balance()/double_rq_lock(), we use raw_spin_rq_lock() to
2648 * acquire rq lock instead of rq_lock(). So at the end of these two functions
2649 * we need to call double_rq_clock_clear_update() to clear RQCF_UPDATED of
2650 * rq->clock_update_flags to avoid the WARN_DOUBLE_CLOCK warning.
2652 static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2)
2654 rq1->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2655 /* rq1 == rq2 for !CONFIG_SMP, so just clear RQCF_UPDATED once. */
2657 rq2->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
2661 static inline void double_rq_clock_clear_update(struct rq *rq1, struct rq *rq2) {}
2664 #define DEFINE_LOCK_GUARD_2(name, type, _lock, _unlock, ...) \
2665 __DEFINE_UNLOCK_GUARD(name, type, _unlock, type *lock2; __VA_ARGS__) \
2666 static inline class_##name##_t class_##name##_constructor(type *lock, type *lock2) \
2667 { class_##name##_t _t = { .lock = lock, .lock2 = lock2 }, *_T = &_t; \
2672 static inline bool rq_order_less(struct rq *rq1, struct rq *rq2)
2674 #ifdef CONFIG_SCHED_CORE
2676 * In order to not have {0,2},{1,3} turn into into an AB-BA,
2677 * order by core-id first and cpu-id second.
2681 * double_rq_lock(0,3); will take core-0, core-1 lock
2682 * double_rq_lock(1,2); will take core-1, core-0 lock
2684 * when only cpu-id is considered.
2686 if (rq1->core->cpu < rq2->core->cpu)
2688 if (rq1->core->cpu > rq2->core->cpu)
2692 * __sched_core_flip() relies on SMT having cpu-id lock order.
2695 return rq1->cpu < rq2->cpu;
2698 extern void double_rq_lock(struct rq *rq1, struct rq *rq2);
2700 #ifdef CONFIG_PREEMPTION
2703 * fair double_lock_balance: Safely acquires both rq->locks in a fair
2704 * way at the expense of forcing extra atomic operations in all
2705 * invocations. This assures that the double_lock is acquired using the
2706 * same underlying policy as the spinlock_t on this architecture, which
2707 * reduces latency compared to the unfair variant below. However, it
2708 * also adds more overhead and therefore may reduce throughput.
2710 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2711 __releases(this_rq->lock)
2712 __acquires(busiest->lock)
2713 __acquires(this_rq->lock)
2715 raw_spin_rq_unlock(this_rq);
2716 double_rq_lock(this_rq, busiest);
2723 * Unfair double_lock_balance: Optimizes throughput at the expense of
2724 * latency by eliminating extra atomic operations when the locks are
2725 * already in proper order on entry. This favors lower CPU-ids and will
2726 * grant the double lock to lower CPUs over higher ids under contention,
2727 * regardless of entry order into the function.
2729 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
2730 __releases(this_rq->lock)
2731 __acquires(busiest->lock)
2732 __acquires(this_rq->lock)
2734 if (__rq_lockp(this_rq) == __rq_lockp(busiest) ||
2735 likely(raw_spin_rq_trylock(busiest))) {
2736 double_rq_clock_clear_update(this_rq, busiest);
2740 if (rq_order_less(this_rq, busiest)) {
2741 raw_spin_rq_lock_nested(busiest, SINGLE_DEPTH_NESTING);
2742 double_rq_clock_clear_update(this_rq, busiest);
2746 raw_spin_rq_unlock(this_rq);
2747 double_rq_lock(this_rq, busiest);
2752 #endif /* CONFIG_PREEMPTION */
2755 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2757 static inline int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2759 lockdep_assert_irqs_disabled();
2761 return _double_lock_balance(this_rq, busiest);
2764 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2765 __releases(busiest->lock)
2767 if (__rq_lockp(this_rq) != __rq_lockp(busiest))
2768 raw_spin_rq_unlock(busiest);
2769 lock_set_subclass(&__rq_lockp(this_rq)->dep_map, 0, _RET_IP_);
2772 static inline void double_lock(spinlock_t *l1, spinlock_t *l2)
2778 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2781 static inline void double_lock_irq(spinlock_t *l1, spinlock_t *l2)
2787 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2790 static inline void double_raw_lock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2796 raw_spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
2799 static inline void double_raw_unlock(raw_spinlock_t *l1, raw_spinlock_t *l2)
2801 raw_spin_unlock(l1);
2802 raw_spin_unlock(l2);
2805 DEFINE_LOCK_GUARD_2(double_raw_spinlock, raw_spinlock_t,
2806 double_raw_lock(_T->lock, _T->lock2),
2807 double_raw_unlock(_T->lock, _T->lock2))
2810 * double_rq_unlock - safely unlock two runqueues
2812 * Note this does not restore interrupts like task_rq_unlock,
2813 * you need to do so manually after calling.
2815 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2816 __releases(rq1->lock)
2817 __releases(rq2->lock)
2819 if (__rq_lockp(rq1) != __rq_lockp(rq2))
2820 raw_spin_rq_unlock(rq2);
2822 __release(rq2->lock);
2823 raw_spin_rq_unlock(rq1);
2826 extern void set_rq_online (struct rq *rq);
2827 extern void set_rq_offline(struct rq *rq);
2828 extern bool sched_smp_initialized;
2830 #else /* CONFIG_SMP */
2833 * double_rq_lock - safely lock two runqueues
2835 * Note this does not disable interrupts like task_rq_lock,
2836 * you need to do so manually before calling.
2838 static inline void double_rq_lock(struct rq *rq1, struct rq *rq2)
2839 __acquires(rq1->lock)
2840 __acquires(rq2->lock)
2842 WARN_ON_ONCE(!irqs_disabled());
2843 WARN_ON_ONCE(rq1 != rq2);
2844 raw_spin_rq_lock(rq1);
2845 __acquire(rq2->lock); /* Fake it out ;) */
2846 double_rq_clock_clear_update(rq1, rq2);
2850 * double_rq_unlock - safely unlock two runqueues
2852 * Note this does not restore interrupts like task_rq_unlock,
2853 * you need to do so manually after calling.
2855 static inline void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2856 __releases(rq1->lock)
2857 __releases(rq2->lock)
2859 WARN_ON_ONCE(rq1 != rq2);
2860 raw_spin_rq_unlock(rq1);
2861 __release(rq2->lock);
2866 DEFINE_LOCK_GUARD_2(double_rq_lock, struct rq,
2867 double_rq_lock(_T->lock, _T->lock2),
2868 double_rq_unlock(_T->lock, _T->lock2))
2870 extern struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq);
2871 extern struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq);
2872 extern struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq);
2874 #ifdef CONFIG_SCHED_DEBUG
2875 extern bool sched_debug_verbose;
2877 extern void print_cfs_stats(struct seq_file *m, int cpu);
2878 extern void print_rt_stats(struct seq_file *m, int cpu);
2879 extern void print_dl_stats(struct seq_file *m, int cpu);
2880 extern void print_cfs_rq(struct seq_file *m, int cpu, struct cfs_rq *cfs_rq);
2881 extern void print_rt_rq(struct seq_file *m, int cpu, struct rt_rq *rt_rq);
2882 extern void print_dl_rq(struct seq_file *m, int cpu, struct dl_rq *dl_rq);
2884 extern void resched_latency_warn(int cpu, u64 latency);
2885 #ifdef CONFIG_NUMA_BALANCING
2887 show_numa_stats(struct task_struct *p, struct seq_file *m);
2889 print_numa_stats(struct seq_file *m, int node, unsigned long tsf,
2890 unsigned long tpf, unsigned long gsf, unsigned long gpf);
2891 #endif /* CONFIG_NUMA_BALANCING */
2893 static inline void resched_latency_warn(int cpu, u64 latency) {}
2894 #endif /* CONFIG_SCHED_DEBUG */
2896 extern void init_cfs_rq(struct cfs_rq *cfs_rq);
2897 extern void init_rt_rq(struct rt_rq *rt_rq);
2898 extern void init_dl_rq(struct dl_rq *dl_rq);
2900 extern void cfs_bandwidth_usage_inc(void);
2901 extern void cfs_bandwidth_usage_dec(void);
2903 #ifdef CONFIG_NO_HZ_COMMON
2904 #define NOHZ_BALANCE_KICK_BIT 0
2905 #define NOHZ_STATS_KICK_BIT 1
2906 #define NOHZ_NEWILB_KICK_BIT 2
2907 #define NOHZ_NEXT_KICK_BIT 3
2909 /* Run rebalance_domains() */
2910 #define NOHZ_BALANCE_KICK BIT(NOHZ_BALANCE_KICK_BIT)
2911 /* Update blocked load */
2912 #define NOHZ_STATS_KICK BIT(NOHZ_STATS_KICK_BIT)
2913 /* Update blocked load when entering idle */
2914 #define NOHZ_NEWILB_KICK BIT(NOHZ_NEWILB_KICK_BIT)
2915 /* Update nohz.next_balance */
2916 #define NOHZ_NEXT_KICK BIT(NOHZ_NEXT_KICK_BIT)
2918 #define NOHZ_KICK_MASK (NOHZ_BALANCE_KICK | NOHZ_STATS_KICK | NOHZ_NEXT_KICK)
2920 #define nohz_flags(cpu) (&cpu_rq(cpu)->nohz_flags)
2922 extern void nohz_balance_exit_idle(struct rq *rq);
2924 static inline void nohz_balance_exit_idle(struct rq *rq) { }
2927 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
2928 extern void nohz_run_idle_balance(int cpu);
2930 static inline void nohz_run_idle_balance(int cpu) { }
2933 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2938 struct u64_stats_sync sync;
2941 DECLARE_PER_CPU(struct irqtime, cpu_irqtime);
2944 * Returns the irqtime minus the softirq time computed by ksoftirqd.
2945 * Otherwise ksoftirqd's sum_exec_runtime is subtracted its own runtime
2946 * and never move forward.
2948 static inline u64 irq_time_read(int cpu)
2950 struct irqtime *irqtime = &per_cpu(cpu_irqtime, cpu);
2955 seq = __u64_stats_fetch_begin(&irqtime->sync);
2956 total = irqtime->total;
2957 } while (__u64_stats_fetch_retry(&irqtime->sync, seq));
2961 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2963 #ifdef CONFIG_CPU_FREQ
2964 DECLARE_PER_CPU(struct update_util_data __rcu *, cpufreq_update_util_data);
2967 * cpufreq_update_util - Take a note about CPU utilization changes.
2968 * @rq: Runqueue to carry out the update for.
2969 * @flags: Update reason flags.
2971 * This function is called by the scheduler on the CPU whose utilization is
2974 * It can only be called from RCU-sched read-side critical sections.
2976 * The way cpufreq is currently arranged requires it to evaluate the CPU
2977 * performance state (frequency/voltage) on a regular basis to prevent it from
2978 * being stuck in a completely inadequate performance level for too long.
2979 * That is not guaranteed to happen if the updates are only triggered from CFS
2980 * and DL, though, because they may not be coming in if only RT tasks are
2981 * active all the time (or there are RT tasks only).
2983 * As a workaround for that issue, this function is called periodically by the
2984 * RT sched class to trigger extra cpufreq updates to prevent it from stalling,
2985 * but that really is a band-aid. Going forward it should be replaced with
2986 * solutions targeted more specifically at RT tasks.
2988 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags)
2990 struct update_util_data *data;
2992 data = rcu_dereference_sched(*per_cpu_ptr(&cpufreq_update_util_data,
2995 data->func(data, rq_clock(rq), flags);
2998 static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
2999 #endif /* CONFIG_CPU_FREQ */
3001 #ifdef arch_scale_freq_capacity
3002 # ifndef arch_scale_freq_invariant
3003 # define arch_scale_freq_invariant() true
3006 # define arch_scale_freq_invariant() false
3010 unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
3012 unsigned long *max);
3014 unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
3020 * Verify the fitness of task @p to run on @cpu taking into account the
3021 * CPU original capacity and the runtime/deadline ratio of the task.
3023 * The function will return true if the original capacity of @cpu is
3024 * greater than or equal to task's deadline density right shifted by
3025 * (BW_SHIFT - SCHED_CAPACITY_SHIFT) and false otherwise.
3027 static inline bool dl_task_fits_capacity(struct task_struct *p, int cpu)
3029 unsigned long cap = arch_scale_cpu_capacity(cpu);
3031 return cap >= p->dl.dl_density >> (BW_SHIFT - SCHED_CAPACITY_SHIFT);
3034 static inline unsigned long cpu_bw_dl(struct rq *rq)
3036 return (rq->dl.running_bw * SCHED_CAPACITY_SCALE) >> BW_SHIFT;
3039 static inline unsigned long cpu_util_dl(struct rq *rq)
3041 return READ_ONCE(rq->avg_dl.util_avg);
3045 extern unsigned long cpu_util_cfs(int cpu);
3046 extern unsigned long cpu_util_cfs_boost(int cpu);
3048 static inline unsigned long cpu_util_rt(struct rq *rq)
3050 return READ_ONCE(rq->avg_rt.util_avg);
3054 #ifdef CONFIG_UCLAMP_TASK
3055 unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id);
3057 static inline unsigned long uclamp_rq_get(struct rq *rq,
3058 enum uclamp_id clamp_id)
3060 return READ_ONCE(rq->uclamp[clamp_id].value);
3063 static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3066 WRITE_ONCE(rq->uclamp[clamp_id].value, value);
3069 static inline bool uclamp_rq_is_idle(struct rq *rq)
3071 return rq->uclamp_flags & UCLAMP_FLAG_IDLE;
3074 /* Is the rq being capped/throttled by uclamp_max? */
3075 static inline bool uclamp_rq_is_capped(struct rq *rq)
3077 unsigned long rq_util;
3078 unsigned long max_util;
3080 if (!static_branch_likely(&sched_uclamp_used))
3083 rq_util = cpu_util_cfs(cpu_of(rq)) + cpu_util_rt(rq);
3084 max_util = READ_ONCE(rq->uclamp[UCLAMP_MAX].value);
3086 return max_util != SCHED_CAPACITY_SCALE && rq_util >= max_util;
3090 * When uclamp is compiled in, the aggregation at rq level is 'turned off'
3091 * by default in the fast path and only gets turned on once userspace performs
3092 * an operation that requires it.
3094 * Returns true if userspace opted-in to use uclamp and aggregation at rq level
3097 static inline bool uclamp_is_used(void)
3099 return static_branch_likely(&sched_uclamp_used);
3101 #else /* CONFIG_UCLAMP_TASK */
3102 static inline unsigned long uclamp_eff_value(struct task_struct *p,
3103 enum uclamp_id clamp_id)
3105 if (clamp_id == UCLAMP_MIN)
3108 return SCHED_CAPACITY_SCALE;
3111 static inline bool uclamp_rq_is_capped(struct rq *rq) { return false; }
3113 static inline bool uclamp_is_used(void)
3118 static inline unsigned long uclamp_rq_get(struct rq *rq,
3119 enum uclamp_id clamp_id)
3121 if (clamp_id == UCLAMP_MIN)
3124 return SCHED_CAPACITY_SCALE;
3127 static inline void uclamp_rq_set(struct rq *rq, enum uclamp_id clamp_id,
3132 static inline bool uclamp_rq_is_idle(struct rq *rq)
3136 #endif /* CONFIG_UCLAMP_TASK */
3138 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
3139 static inline unsigned long cpu_util_irq(struct rq *rq)
3141 return rq->avg_irq.util_avg;
3145 unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3147 util *= (max - irq);
3154 static inline unsigned long cpu_util_irq(struct rq *rq)
3160 unsigned long scale_irq_capacity(unsigned long util, unsigned long irq, unsigned long max)
3166 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
3168 #define perf_domain_span(pd) (to_cpumask(((pd)->em_pd->cpus)))
3170 DECLARE_STATIC_KEY_FALSE(sched_energy_present);
3172 static inline bool sched_energy_enabled(void)
3174 return static_branch_unlikely(&sched_energy_present);
3177 extern struct cpufreq_governor schedutil_gov;
3179 #else /* ! (CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL) */
3181 #define perf_domain_span(pd) NULL
3182 static inline bool sched_energy_enabled(void) { return false; }
3184 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL */
3186 #ifdef CONFIG_MEMBARRIER
3188 * The scheduler provides memory barriers required by membarrier between:
3189 * - prior user-space memory accesses and store to rq->membarrier_state,
3190 * - store to rq->membarrier_state and following user-space memory accesses.
3191 * In the same way it provides those guarantees around store to rq->curr.
3193 static inline void membarrier_switch_mm(struct rq *rq,
3194 struct mm_struct *prev_mm,
3195 struct mm_struct *next_mm)
3197 int membarrier_state;
3199 if (prev_mm == next_mm)
3202 membarrier_state = atomic_read(&next_mm->membarrier_state);
3203 if (READ_ONCE(rq->membarrier_state) == membarrier_state)
3206 WRITE_ONCE(rq->membarrier_state, membarrier_state);
3209 static inline void membarrier_switch_mm(struct rq *rq,
3210 struct mm_struct *prev_mm,
3211 struct mm_struct *next_mm)
3217 static inline bool is_per_cpu_kthread(struct task_struct *p)
3219 if (!(p->flags & PF_KTHREAD))
3222 if (p->nr_cpus_allowed != 1)
3229 extern void swake_up_all_locked(struct swait_queue_head *q);
3230 extern void __prepare_to_swait(struct swait_queue_head *q, struct swait_queue *wait);
3232 extern int try_to_wake_up(struct task_struct *tsk, unsigned int state, int wake_flags);
3234 #ifdef CONFIG_PREEMPT_DYNAMIC
3235 extern int preempt_dynamic_mode;
3236 extern int sched_dynamic_mode(const char *str);
3237 extern void sched_dynamic_update(int mode);
3240 #ifdef CONFIG_SCHED_MM_CID
3242 #define SCHED_MM_CID_PERIOD_NS (100ULL * 1000000) /* 100ms */
3243 #define MM_CID_SCAN_DELAY 100 /* 100ms */
3245 extern raw_spinlock_t cid_lock;
3246 extern int use_cid_lock;
3248 extern void sched_mm_cid_migrate_from(struct task_struct *t);
3249 extern void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t);
3250 extern void task_tick_mm_cid(struct rq *rq, struct task_struct *curr);
3251 extern void init_sched_mm_cid(struct task_struct *t);
3253 static inline void __mm_cid_put(struct mm_struct *mm, int cid)
3257 cpumask_clear_cpu(cid, mm_cidmask(mm));
3261 * The per-mm/cpu cid can have the MM_CID_LAZY_PUT flag set or transition to
3262 * the MM_CID_UNSET state without holding the rq lock, but the rq lock needs to
3263 * be held to transition to other states.
3265 * State transitions synchronized with cmpxchg or try_cmpxchg need to be
3266 * consistent across cpus, which prevents use of this_cpu_cmpxchg.
3268 static inline void mm_cid_put_lazy(struct task_struct *t)
3270 struct mm_struct *mm = t->mm;
3271 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3274 lockdep_assert_irqs_disabled();
3275 cid = __this_cpu_read(pcpu_cid->cid);
3276 if (!mm_cid_is_lazy_put(cid) ||
3277 !try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3279 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3282 static inline int mm_cid_pcpu_unset(struct mm_struct *mm)
3284 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3287 lockdep_assert_irqs_disabled();
3288 cid = __this_cpu_read(pcpu_cid->cid);
3290 if (mm_cid_is_unset(cid))
3291 return MM_CID_UNSET;
3293 * Attempt transition from valid or lazy-put to unset.
3295 res = cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, cid, MM_CID_UNSET);
3303 static inline void mm_cid_put(struct mm_struct *mm)
3307 lockdep_assert_irqs_disabled();
3308 cid = mm_cid_pcpu_unset(mm);
3309 if (cid == MM_CID_UNSET)
3311 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3314 static inline int __mm_cid_try_get(struct mm_struct *mm)
3316 struct cpumask *cpumask;
3319 cpumask = mm_cidmask(mm);
3321 * Retry finding first zero bit if the mask is temporarily
3322 * filled. This only happens during concurrent remote-clear
3323 * which owns a cid without holding a rq lock.
3326 cid = cpumask_first_zero(cpumask);
3327 if (cid < nr_cpu_ids)
3331 if (cpumask_test_and_set_cpu(cid, cpumask))
3337 * Save a snapshot of the current runqueue time of this cpu
3338 * with the per-cpu cid value, allowing to estimate how recently it was used.
3340 static inline void mm_cid_snapshot_time(struct rq *rq, struct mm_struct *mm)
3342 struct mm_cid *pcpu_cid = per_cpu_ptr(mm->pcpu_cid, cpu_of(rq));
3344 lockdep_assert_rq_held(rq);
3345 WRITE_ONCE(pcpu_cid->time, rq->clock);
3348 static inline int __mm_cid_get(struct rq *rq, struct mm_struct *mm)
3353 * All allocations (even those using the cid_lock) are lock-free. If
3354 * use_cid_lock is set, hold the cid_lock to perform cid allocation to
3355 * guarantee forward progress.
3357 if (!READ_ONCE(use_cid_lock)) {
3358 cid = __mm_cid_try_get(mm);
3361 raw_spin_lock(&cid_lock);
3363 raw_spin_lock(&cid_lock);
3364 cid = __mm_cid_try_get(mm);
3370 * cid concurrently allocated. Retry while forcing following
3371 * allocations to use the cid_lock to ensure forward progress.
3373 WRITE_ONCE(use_cid_lock, 1);
3375 * Set use_cid_lock before allocation. Only care about program order
3376 * because this is only required for forward progress.
3380 * Retry until it succeeds. It is guaranteed to eventually succeed once
3381 * all newcoming allocations observe the use_cid_lock flag set.
3384 cid = __mm_cid_try_get(mm);
3388 * Allocate before clearing use_cid_lock. Only care about
3389 * program order because this is for forward progress.
3392 WRITE_ONCE(use_cid_lock, 0);
3394 raw_spin_unlock(&cid_lock);
3396 mm_cid_snapshot_time(rq, mm);
3400 static inline int mm_cid_get(struct rq *rq, struct mm_struct *mm)
3402 struct mm_cid __percpu *pcpu_cid = mm->pcpu_cid;
3403 struct cpumask *cpumask;
3406 lockdep_assert_rq_held(rq);
3407 cpumask = mm_cidmask(mm);
3408 cid = __this_cpu_read(pcpu_cid->cid);
3409 if (mm_cid_is_valid(cid)) {
3410 mm_cid_snapshot_time(rq, mm);
3413 if (mm_cid_is_lazy_put(cid)) {
3414 if (try_cmpxchg(&this_cpu_ptr(pcpu_cid)->cid, &cid, MM_CID_UNSET))
3415 __mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
3417 cid = __mm_cid_get(rq, mm);
3418 __this_cpu_write(pcpu_cid->cid, cid);
3422 static inline void switch_mm_cid(struct rq *rq,
3423 struct task_struct *prev,
3424 struct task_struct *next)
3427 * Provide a memory barrier between rq->curr store and load of
3428 * {prev,next}->mm->pcpu_cid[cpu] on rq->curr->mm transition.
3430 * Should be adapted if context_switch() is modified.
3432 if (!next->mm) { // to kernel
3434 * user -> kernel transition does not guarantee a barrier, but
3435 * we can use the fact that it performs an atomic operation in
3438 if (prev->mm) // from user
3439 smp_mb__after_mmgrab();
3441 * kernel -> kernel transition does not change rq->curr->mm
3442 * state. It stays NULL.
3446 * kernel -> user transition does not provide a barrier
3447 * between rq->curr store and load of {prev,next}->mm->pcpu_cid[cpu].
3450 if (!prev->mm) { // from kernel
3452 } else { // from user
3454 * user->user transition relies on an implicit
3455 * memory barrier in switch_mm() when
3456 * current->mm changes. If the architecture
3457 * switch_mm() does not have an implicit memory
3458 * barrier, it is emitted here. If current->mm
3459 * is unchanged, no barrier is needed.
3461 smp_mb__after_switch_mm();
3464 if (prev->mm_cid_active) {
3465 mm_cid_snapshot_time(rq, prev->mm);
3466 mm_cid_put_lazy(prev);
3469 if (next->mm_cid_active)
3470 next->last_mm_cid = next->mm_cid = mm_cid_get(rq, next->mm);
3474 static inline void switch_mm_cid(struct rq *rq, struct task_struct *prev, struct task_struct *next) { }
3475 static inline void sched_mm_cid_migrate_from(struct task_struct *t) { }
3476 static inline void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t) { }
3477 static inline void task_tick_mm_cid(struct rq *rq, struct task_struct *curr) { }
3478 static inline void init_sched_mm_cid(struct task_struct *t) { }
3481 extern u64 avg_vruntime(struct cfs_rq *cfs_rq);
3482 extern int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se);
3484 #endif /* _KERNEL_SCHED_SCHED_H */