Lines Matching +full:differential +full:- +full:pair
1 // SPDX-License-Identifier: GPL-2.0
44 #include <linux/memory-tiers.h>
62 * The initial- and re-scaling of tunables is configurable
66 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
67 * SCHED_TUNABLESCALING_LOG - scaled logarithmically, *1+ilog(ncpus)
68 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
75 * Minimal preemption granularity for CPU-bound tasks:
97 return -cpu; in arch_asym_cpu_priority()
118 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
169 lw->weight += inc; in update_load_add()
170 lw->inv_weight = 0; in update_load_add()
175 lw->weight -= dec; in update_load_sub()
176 lw->inv_weight = 0; in update_load_sub()
181 lw->weight = w; in update_load_set()
182 lw->inv_weight = 0; in update_load_set()
189 * so pick a second-best guess by going with the log2 of the
237 if (likely(lw->inv_weight)) in __update_inv_weight()
240 w = scale_load_down(lw->weight); in __update_inv_weight()
243 lw->inv_weight = 1; in __update_inv_weight()
245 lw->inv_weight = WMULT_CONST; in __update_inv_weight()
247 lw->inv_weight = WMULT_CONST / w; in __update_inv_weight()
253 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
273 shift -= fs; in __calc_delta()
277 fact = mul_u32_u32(fact, lw->inv_weight); in __calc_delta()
282 shift -= fs; in __calc_delta()
294 if (unlikely(se->load.weight != NICE_0_LOAD)) in calc_delta_fair()
295 delta = __calc_delta(delta, NICE_0_LOAD, &se->load); in calc_delta_fair()
310 for (; se; se = se->parent)
317 if (cfs_rq->on_list) in list_add_leaf_cfs_rq()
318 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; in list_add_leaf_cfs_rq()
320 cfs_rq->on_list = 1; in list_add_leaf_cfs_rq()
325 * enqueued. The fact that we always enqueue bottom-up in list_add_leaf_cfs_rq()
331 if (cfs_rq->tg->parent && in list_add_leaf_cfs_rq()
332 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { in list_add_leaf_cfs_rq()
339 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, in list_add_leaf_cfs_rq()
340 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); in list_add_leaf_cfs_rq()
346 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; in list_add_leaf_cfs_rq()
350 if (!cfs_rq->tg->parent) { in list_add_leaf_cfs_rq()
355 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, in list_add_leaf_cfs_rq()
356 &rq->leaf_cfs_rq_list); in list_add_leaf_cfs_rq()
361 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; in list_add_leaf_cfs_rq()
371 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); in list_add_leaf_cfs_rq()
376 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; in list_add_leaf_cfs_rq()
382 if (cfs_rq->on_list) { in list_del_leaf_cfs_rq()
389 * to the prev element but it will point to rq->leaf_cfs_rq_list in list_del_leaf_cfs_rq()
392 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) in list_del_leaf_cfs_rq()
393 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; in list_del_leaf_cfs_rq()
395 list_del_rcu(&cfs_rq->leaf_cfs_rq_list); in list_del_leaf_cfs_rq()
396 cfs_rq->on_list = 0; in list_del_leaf_cfs_rq()
402 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); in assert_list_leaf_cfs_rq()
407 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
414 if (se->cfs_rq == pse->cfs_rq) in is_same_group()
415 return se->cfs_rq; in is_same_group()
422 return se->parent; in parent_entity()
438 se_depth = (*se)->depth; in find_matching_se()
439 pse_depth = (*pse)->depth; in find_matching_se()
442 se_depth--; in find_matching_se()
447 pse_depth--; in find_matching_se()
459 return tg->idle > 0; in tg_is_idle()
464 return cfs_rq->idle > 0; in cfs_rq_is_idle()
493 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
531 s64 delta = (s64)(vruntime - max_vruntime); in max_vruntime()
540 s64 delta = (s64)(vruntime - min_vruntime); in min_vruntime()
554 return (s64)(a->deadline - b->deadline) < 0; in entity_before()
559 return (s64)(se->vruntime - cfs_rq->min_vruntime); in entity_key()
566 * Compute virtual time from the per-task service numbers:
574 * lag_i = S - s_i = w_i * (V - v_i)
580 * \Sum w_i * (V - v_i) = 0
581 * \Sum w_i * V - w_i * v_i = 0
584 * se->vruntime):
587 * V = -------------- = --------------
594 * virtual time has non-contiguous motion equivalent to:
596 * V +-= lag_i / W
603 * Substitute: v_i == (v_i - v0) + v0
605 * \Sum ((v_i - v0) + v0) * w_i \Sum (v_i - v0) * w_i
606 * V = ---------------------------- = --------------------- + v0
611 * v0 := cfs_rq->min_vruntime
612 * \Sum (v_i - v0) * w_i := cfs_rq->avg_vruntime
613 * \Sum w_i := cfs_rq->avg_load
616 * the per-task service, these deltas: (v_i - v), will be in the order of the
626 unsigned long weight = scale_load_down(se->load.weight); in avg_vruntime_add()
629 cfs_rq->avg_vruntime += key * weight; in avg_vruntime_add()
630 cfs_rq->avg_load += weight; in avg_vruntime_add()
636 unsigned long weight = scale_load_down(se->load.weight); in avg_vruntime_sub()
639 cfs_rq->avg_vruntime -= key * weight; in avg_vruntime_sub()
640 cfs_rq->avg_load -= weight; in avg_vruntime_sub()
647 * v' = v + d ==> avg_vruntime' = avg_runtime - d*avg_load in avg_vruntime_update()
649 cfs_rq->avg_vruntime -= cfs_rq->avg_load * delta; in avg_vruntime_update()
658 struct sched_entity *curr = cfs_rq->curr; in avg_vruntime()
659 s64 avg = cfs_rq->avg_vruntime; in avg_vruntime()
660 long load = cfs_rq->avg_load; in avg_vruntime()
662 if (curr && curr->on_rq) { in avg_vruntime()
663 unsigned long weight = scale_load_down(curr->load.weight); in avg_vruntime()
672 avg -= (load - 1); in avg_vruntime()
676 return cfs_rq->min_vruntime + avg; in avg_vruntime()
680 * lag_i = S - s_i = w_i * (V - v_i)
683 * is possible -- by addition/removal/reweight to the tree -- to move V around
691 * -r_max < lag < max(r_max, q)
699 SCHED_WARN_ON(!se->on_rq); in update_entity_lag()
701 vlag = avg_vruntime(cfs_rq) - se->vruntime; in update_entity_lag()
702 limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se); in update_entity_lag()
704 se->vlag = clamp(vlag, -limit, limit); in update_entity_lag()
711 * lag_i = S - s_i = w_i*(V - v_i)
713 * lag_i >= 0 -> V >= v_i
715 * \Sum (v_i - v)*w_i
716 * V = ------------------ + v
719 * lag_i >= 0 -> \Sum (v_i - v)*w_i >= (v_i - v)*(\Sum w_i)
721 * Note: using 'avg_vruntime() > se->vruntime' is inaccurate due
726 struct sched_entity *curr = cfs_rq->curr; in vruntime_eligible()
727 s64 avg = cfs_rq->avg_vruntime; in vruntime_eligible()
728 long load = cfs_rq->avg_load; in vruntime_eligible()
730 if (curr && curr->on_rq) { in vruntime_eligible()
731 unsigned long weight = scale_load_down(curr->load.weight); in vruntime_eligible()
737 return avg >= (s64)(vruntime - cfs_rq->min_vruntime) * load; in vruntime_eligible()
742 return vruntime_eligible(cfs_rq, se->vruntime); in entity_eligible()
747 u64 min_vruntime = cfs_rq->min_vruntime; in __update_min_vruntime()
751 s64 delta = (s64)(vruntime - min_vruntime); in __update_min_vruntime()
762 struct sched_entity *curr = cfs_rq->curr; in update_min_vruntime()
763 u64 vruntime = cfs_rq->min_vruntime; in update_min_vruntime()
766 if (curr->on_rq) in update_min_vruntime()
767 vruntime = curr->vruntime; in update_min_vruntime()
774 vruntime = se->min_vruntime; in update_min_vruntime()
776 vruntime = min_vruntime(vruntime, se->min_vruntime); in update_min_vruntime()
780 cfs_rq->min_vruntime = __update_min_vruntime(cfs_rq, vruntime); in update_min_vruntime()
786 struct sched_entity *curr = cfs_rq->curr; in cfs_rq_min_slice()
789 if (curr && curr->on_rq) in cfs_rq_min_slice()
790 min_slice = curr->slice; in cfs_rq_min_slice()
793 min_slice = min(min_slice, root->min_slice); in cfs_rq_min_slice()
803 #define vruntime_gt(field, lse, rse) ({ (s64)((lse)->field - (rse)->field) > 0; })
810 se->min_vruntime = rse->min_vruntime; in __min_vruntime_update()
818 if (rse->min_slice < se->min_slice) in __min_slice_update()
819 se->min_slice = rse->min_slice; in __min_slice_update()
824 * se->min_vruntime = min(se->vruntime, {left,right}->min_vruntime)
828 u64 old_min_vruntime = se->min_vruntime; in min_vruntime_update()
829 u64 old_min_slice = se->min_slice; in min_vruntime_update()
830 struct rb_node *node = &se->run_node; in min_vruntime_update()
832 se->min_vruntime = se->vruntime; in min_vruntime_update()
833 __min_vruntime_update(se, node->rb_right); in min_vruntime_update()
834 __min_vruntime_update(se, node->rb_left); in min_vruntime_update()
836 se->min_slice = se->slice; in min_vruntime_update()
837 __min_slice_update(se, node->rb_right); in min_vruntime_update()
838 __min_slice_update(se, node->rb_left); in min_vruntime_update()
840 return se->min_vruntime == old_min_vruntime && in min_vruntime_update()
841 se->min_slice == old_min_slice; in min_vruntime_update()
848 * Enqueue an entity into the rb-tree:
853 se->min_vruntime = se->vruntime; in __enqueue_entity()
854 se->min_slice = se->slice; in __enqueue_entity()
855 rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline, in __enqueue_entity()
861 rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline, in __dequeue_entity()
868 struct rb_node *root = cfs_rq->tasks_timeline.rb_root.rb_node; in __pick_root_entity()
878 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); in __pick_first_entity()
897 * We can do this in O(log n) time due to an augmented RB-tree. The
901 * se->min_vruntime = min(se->vruntime, se->{left,right}->min_vruntime)
907 struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node; in pick_eevdf()
909 struct sched_entity *curr = cfs_rq->curr; in pick_eevdf()
916 if (cfs_rq->nr_queued == 1) in pick_eevdf()
917 return curr && curr->on_rq ? curr : se; in pick_eevdf()
919 if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr))) in pick_eevdf()
923 * Once selected, run a task until it either becomes non-eligible or in pick_eevdf()
926 if (sched_feat(RUN_TO_PARITY) && curr && curr->vlag == curr->deadline) in pick_eevdf()
937 struct rb_node *left = node->rb_left; in pick_eevdf()
944 __node_2_se(left)->min_vruntime)) { in pick_eevdf()
961 node = node->rb_right; in pick_eevdf()
973 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); in __pick_last_entity()
1007 if ((s64)(se->vruntime - se->deadline) < 0) in update_deadline()
1015 if (!se->custom_slice) in update_deadline()
1016 se->slice = sysctl_sched_base_slice; in update_deadline()
1021 se->deadline = se->vruntime + calc_delta_fair(se->slice, se); in update_deadline()
1039 struct sched_avg *sa = &se->avg; in init_entity_runnable_average()
1050 sa->load_avg = scale_load_down(se->load.weight); in init_entity_runnable_average()
1059 * util_avg = cfs_rq->avg.util_avg / (cfs_rq->avg.load_avg + 1)
1069 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
1084 struct sched_entity *se = &p->se; in post_init_entity_util_avg()
1086 struct sched_avg *sa = &se->avg; in post_init_entity_util_avg()
1088 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; in post_init_entity_util_avg()
1090 if (p->sched_class != &fair_sched_class) { in post_init_entity_util_avg()
1101 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); in post_init_entity_util_avg()
1106 if (cfs_rq->avg.util_avg != 0) { in post_init_entity_util_avg()
1107 sa->util_avg = cfs_rq->avg.util_avg * se_weight(se); in post_init_entity_util_avg()
1108 sa->util_avg /= (cfs_rq->avg.load_avg + 1); in post_init_entity_util_avg()
1110 if (sa->util_avg > cap) in post_init_entity_util_avg()
1111 sa->util_avg = cap; in post_init_entity_util_avg()
1113 sa->util_avg = cap; in post_init_entity_util_avg()
1117 sa->runnable_avg = sa->util_avg; in post_init_entity_util_avg()
1137 delta_exec = now - curr->exec_start; in update_curr_se()
1141 curr->exec_start = now; in update_curr_se()
1142 curr->sum_exec_runtime += delta_exec; in update_curr_se()
1148 __schedstat_set(stats->exec_max, in update_curr_se()
1149 max(delta_exec, stats->exec_max)); in update_curr_se()
1167 if (curr->vlag == curr->deadline) in did_preempt_short()
1179 if (pse->slice >= se->slice) in do_preempt_short()
1199 struct task_struct *donor = rq->donor; in update_curr_common()
1202 delta_exec = update_curr_se(rq, &donor->se); in update_curr_common()
1214 struct sched_entity *curr = cfs_rq->curr; in update_curr()
1226 curr->vruntime += calc_delta_fair(delta_exec, curr); in update_curr()
1239 * - If the task is running on behalf of fair_server, we need in update_curr()
1241 * - Fair task that runs outside of fair_server should account in update_curr()
1245 if (dl_server_active(&rq->fair_server)) in update_curr()
1246 dl_server_update(&rq->fair_server, delta_exec); in update_curr()
1251 if (cfs_rq->nr_queued == 1) in update_curr()
1262 update_curr(cfs_rq_of(&rq->donor->se)); in update_curr_fair()
1295 * maybe already in the runqueue, the se->statistics.wait_start in update_stats_wait_end_fair()
1299 if (unlikely(!schedstat_val(stats->wait_start))) in update_stats_wait_end_fair()
1326 * Task is being enqueued - update stats:
1338 if (se != cfs_rq->curr) in update_stats_enqueue_fair()
1356 if (se != cfs_rq->curr) in update_stats_dequeue_fair()
1364 state = READ_ONCE(tsk->__state); in update_stats_dequeue_fair()
1366 __schedstat_set(tsk->stats.sleep_start, in update_stats_dequeue_fair()
1369 __schedstat_set(tsk->stats.block_start, in update_stats_dequeue_fair()
1375 * We are picking a new current task - update its stats:
1383 se->exec_start = rq_clock_task(rq_of(cfs_rq)); in update_stats_curr_start()
1424 * Allow a small imbalance based on a simple pair of communicating in adjust_numa_imbalance()
1475 * ->numa_group (see struct task_struct for locking rules).
1479 return rcu_dereference_check(p->numa_group, p == current || in deref_task_numa_group()
1480 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu))); in deref_task_numa_group()
1485 return rcu_dereference_protected(p->numa_group, p == current); in deref_curr_numa_group()
1497 * Calculations based on RSS as non-present and empty pages are skipped in task_nr_scan_windows()
1501 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); in task_nr_scan_windows()
1502 rss = get_mm_rss(p->mm); in task_nr_scan_windows()
1535 ng = rcu_dereference(p->numa_group); in task_scan_start()
1540 period *= refcount_read(&ng->refcount); in task_scan_start()
1565 period *= refcount_read(&ng->refcount); in task_scan_max()
1577 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); in account_numa_enqueue()
1578 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); in account_numa_enqueue()
1583 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE); in account_numa_dequeue()
1584 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); in account_numa_dequeue()
1602 ng = rcu_dereference(p->numa_group); in task_numa_group_id()
1604 gid = ng->gid; in task_numa_group_id()
1623 if (!p->numa_faults) in task_faults()
1626 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + in task_faults()
1627 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; in task_faults()
1637 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + in group_faults()
1638 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; in group_faults()
1643 return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] + in group_faults_cpu()
1644 group->faults[task_faults_idx(NUMA_CPU, nid, 1)]; in group_faults_cpu()
1653 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; in group_faults_priv()
1665 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; in group_faults_shared()
1673 * considered part of a numa group's pseudo-interleaving set. Migrations
1680 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; in numa_is_active_node()
1739 faults *= (max_dist - dist); in score_nearby_nodes()
1740 faults /= (max_dist - LOCAL_DISTANCE); in score_nearby_nodes()
1760 if (!p->numa_faults) in task_weight()
1763 total_faults = p->total_numa_faults; in task_weight()
1783 total_faults = ng->total_faults; in group_weight()
1819 pgdat->node_present_pages >> 4); in pgdat_free_space_enough()
1820 for (z = pgdat->nr_zones - 1; z >= 0; z--) { in pgdat_free_space_enough()
1821 struct zone *zone = pgdat->node_zones + z; in pgdat_free_space_enough()
1841 * hint page fault latency = hint page fault time - scan time
1853 return (time - last_time) & PAGE_ACCESS_TIME_MASK; in numa_hint_fault_latency()
1870 start = pgdat->nbp_rl_start; in numa_promotion_rate_limit()
1871 if (now - start > MSEC_PER_SEC && in numa_promotion_rate_limit()
1872 cmpxchg(&pgdat->nbp_rl_start, start, now) == start) in numa_promotion_rate_limit()
1873 pgdat->nbp_rl_nr_cand = nr_cand; in numa_promotion_rate_limit()
1874 if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit) in numa_promotion_rate_limit()
1890 start = pgdat->nbp_th_start; in numa_promotion_adjust_threshold()
1891 if (now - start > th_period && in numa_promotion_adjust_threshold()
1892 cmpxchg(&pgdat->nbp_th_start, start, now) == start) { in numa_promotion_adjust_threshold()
1896 diff_cand = nr_cand - pgdat->nbp_th_nr_cand; in numa_promotion_adjust_threshold()
1898 th = pgdat->nbp_threshold ? : ref_th; in numa_promotion_adjust_threshold()
1900 th = max(th - unit_th, unit_th); in numa_promotion_adjust_threshold()
1903 pgdat->nbp_th_nr_cand = nr_cand; in numa_promotion_adjust_threshold()
1904 pgdat->nbp_threshold = th; in numa_promotion_adjust_threshold()
1933 pgdat->nbp_threshold = 0; in should_numa_migrate_memory()
1939 (20 - PAGE_SHIFT); in should_numa_migrate_memory()
1942 th = pgdat->nbp_threshold ? : def_th; in should_numa_migrate_memory()
1951 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); in should_numa_migrate_memory()
1961 * two full passes of the "multi-stage node selection" test that is in should_numa_migrate_memory()
1964 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && in should_numa_migrate_memory()
1969 * Multi-stage node selection is used in conjunction with a periodic in should_numa_migrate_memory()
1970 * migration fault to build a temporal task<->page relation. By using in should_numa_migrate_memory()
1971 * a two-stage filter we remove short/unlikely relations. in should_numa_migrate_memory()
1975 * page (n_t) (in a given time-span) to a probability. in should_numa_migrate_memory()
1983 * act on an unlikely task<->page relation. in should_numa_migrate_memory()
1993 /* A shared fault, but p->numa_group has not been set up yet. */ in should_numa_migrate_memory()
2010 * --------------- * - > --------------- in should_numa_migrate_memory()
2072 if ((ns->nr_running > ns->weight) && in numa_classify()
2073 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) || in numa_classify()
2074 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100)))) in numa_classify()
2077 if ((ns->nr_running < ns->weight) || in numa_classify()
2078 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) && in numa_classify()
2079 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100)))) in numa_classify()
2120 int cpu, idle_core = -1; in update_numa_stats()
2123 ns->idle_cpu = -1; in update_numa_stats()
2129 ns->load += cpu_load(rq); in update_numa_stats()
2130 ns->runnable += cpu_runnable(rq); in update_numa_stats()
2131 ns->util += cpu_util_cfs(cpu); in update_numa_stats()
2132 ns->nr_running += rq->cfs.h_nr_runnable; in update_numa_stats()
2133 ns->compute_capacity += capacity_of(cpu); in update_numa_stats()
2135 if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) { in update_numa_stats()
2136 if (READ_ONCE(rq->numa_migrate_on) || in update_numa_stats()
2137 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) in update_numa_stats()
2140 if (ns->idle_cpu == -1) in update_numa_stats()
2141 ns->idle_cpu = cpu; in update_numa_stats()
2148 ns->weight = cpumask_weight(cpumask_of_node(nid)); in update_numa_stats()
2150 ns->node_type = numa_classify(env->imbalance_pct, ns); in update_numa_stats()
2153 ns->idle_cpu = idle_core; in update_numa_stats()
2159 struct rq *rq = cpu_rq(env->dst_cpu); in task_numa_assign()
2161 /* Check if run-queue part of active NUMA balance. */ in task_numa_assign()
2162 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) { in task_numa_assign()
2164 int start = env->dst_cpu; in task_numa_assign()
2167 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) { in task_numa_assign()
2168 if (cpu == env->best_cpu || !idle_cpu(cpu) || in task_numa_assign()
2169 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) { in task_numa_assign()
2173 env->dst_cpu = cpu; in task_numa_assign()
2174 rq = cpu_rq(env->dst_cpu); in task_numa_assign()
2175 if (!xchg(&rq->numa_migrate_on, 1)) in task_numa_assign()
2185 * Clear previous best_cpu/rq numa-migrate flag, since task now in task_numa_assign()
2188 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) { in task_numa_assign()
2189 rq = cpu_rq(env->best_cpu); in task_numa_assign()
2190 WRITE_ONCE(rq->numa_migrate_on, 0); in task_numa_assign()
2193 if (env->best_task) in task_numa_assign()
2194 put_task_struct(env->best_task); in task_numa_assign()
2198 env->best_task = p; in task_numa_assign()
2199 env->best_imp = imp; in task_numa_assign()
2200 env->best_cpu = env->dst_cpu; in task_numa_assign()
2214 * ------------ vs --------- in load_too_imbalanced()
2217 src_capacity = env->src_stats.compute_capacity; in load_too_imbalanced()
2218 dst_capacity = env->dst_stats.compute_capacity; in load_too_imbalanced()
2220 imb = abs(dst_load * src_capacity - src_load * dst_capacity); in load_too_imbalanced()
2222 orig_src_load = env->src_stats.load; in load_too_imbalanced()
2223 orig_dst_load = env->dst_stats.load; in load_too_imbalanced()
2225 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); in load_too_imbalanced()
2247 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); in task_numa_compare()
2248 struct rq *dst_rq = cpu_rq(env->dst_cpu); in task_numa_compare()
2252 int dist = env->dist; in task_numa_compare()
2257 if (READ_ONCE(dst_rq->numa_migrate_on)) in task_numa_compare()
2261 cur = rcu_dereference(dst_rq->curr); in task_numa_compare()
2262 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) in task_numa_compare()
2267 * end try selecting ourselves (current == env->p) as a swap candidate. in task_numa_compare()
2269 if (cur == env->p) { in task_numa_compare()
2275 if (maymove && moveimp >= env->best_imp) in task_numa_compare()
2282 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) in task_numa_compare()
2289 if (env->best_task && in task_numa_compare()
2290 env->best_task->numa_preferred_nid == env->src_nid && in task_numa_compare()
2291 cur->numa_preferred_nid != env->src_nid) { in task_numa_compare()
2296 * "imp" is the fault differential for the source task between the in task_numa_compare()
2297 * source and destination node. Calculate the total differential for in task_numa_compare()
2305 cur_ng = rcu_dereference(cur->numa_group); in task_numa_compare()
2313 if (env->dst_stats.node_type == node_has_spare) in task_numa_compare()
2316 imp = taskimp + task_weight(cur, env->src_nid, dist) - in task_numa_compare()
2317 task_weight(cur, env->dst_nid, dist); in task_numa_compare()
2323 imp -= imp / 16; in task_numa_compare()
2330 imp += group_weight(cur, env->src_nid, dist) - in task_numa_compare()
2331 group_weight(cur, env->dst_nid, dist); in task_numa_compare()
2333 imp += task_weight(cur, env->src_nid, dist) - in task_numa_compare()
2334 task_weight(cur, env->dst_nid, dist); in task_numa_compare()
2338 if (cur->numa_preferred_nid == env->dst_nid) in task_numa_compare()
2339 imp -= imp / 16; in task_numa_compare()
2347 if (cur->numa_preferred_nid == env->src_nid) in task_numa_compare()
2350 if (maymove && moveimp > imp && moveimp > env->best_imp) { in task_numa_compare()
2360 if (env->best_task && cur->numa_preferred_nid == env->src_nid && in task_numa_compare()
2361 env->best_task->numa_preferred_nid != env->src_nid) { in task_numa_compare()
2371 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) in task_numa_compare()
2377 load = task_h_load(env->p) - task_h_load(cur); in task_numa_compare()
2381 dst_load = env->dst_stats.load + load; in task_numa_compare()
2382 src_load = env->src_stats.load - load; in task_numa_compare()
2390 int cpu = env->dst_stats.idle_cpu; in task_numa_compare()
2394 cpu = env->dst_cpu; in task_numa_compare()
2400 if (!idle_cpu(cpu) && env->best_cpu >= 0 && in task_numa_compare()
2401 idle_cpu(env->best_cpu)) { in task_numa_compare()
2402 cpu = env->best_cpu; in task_numa_compare()
2405 env->dst_cpu = cpu; in task_numa_compare()
2415 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu)) in task_numa_compare()
2422 if (!maymove && env->best_task && in task_numa_compare()
2423 env->best_task->numa_preferred_nid == env->src_nid) { in task_numa_compare()
2442 if (env->dst_stats.node_type == node_has_spare) { in task_numa_find_cpu()
2452 src_running = env->src_stats.nr_running - 1; in task_numa_find_cpu()
2453 dst_running = env->dst_stats.nr_running + 1; in task_numa_find_cpu()
2454 imbalance = max(0, dst_running - src_running); in task_numa_find_cpu()
2456 env->imb_numa_nr); in task_numa_find_cpu()
2461 if (env->dst_stats.idle_cpu >= 0) { in task_numa_find_cpu()
2462 env->dst_cpu = env->dst_stats.idle_cpu; in task_numa_find_cpu()
2470 * If the improvement from just moving env->p direction is better in task_numa_find_cpu()
2473 load = task_h_load(env->p); in task_numa_find_cpu()
2474 dst_load = env->dst_stats.load + load; in task_numa_find_cpu()
2475 src_load = env->src_stats.load - load; in task_numa_find_cpu()
2479 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { in task_numa_find_cpu()
2481 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) in task_numa_find_cpu()
2484 env->dst_cpu = cpu; in task_numa_find_cpu()
2502 .best_cpu = -1, in task_numa_migrate()
2516 * random movement of tasks -- counter the numa conditions we're trying in task_numa_migrate()
2522 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; in task_numa_migrate()
2523 env.imb_numa_nr = sd->imb_numa_nr; in task_numa_migrate()
2535 return -EINVAL; in task_numa_migrate()
2538 env.dst_nid = p->numa_preferred_nid; in task_numa_migrate()
2543 taskimp = task_weight(p, env.dst_nid, dist) - taskweight; in task_numa_migrate()
2544 groupimp = group_weight(p, env.dst_nid, dist) - groupweight; in task_numa_migrate()
2552 * - there is no space available on the preferred_nid in task_numa_migrate()
2553 * - the task is part of a numa_group that is interleaved across in task_numa_migrate()
2558 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { in task_numa_migrate()
2560 if (nid == env.src_nid || nid == p->numa_preferred_nid) in task_numa_migrate()
2571 taskimp = task_weight(p, nid, dist) - taskweight; in task_numa_migrate()
2572 groupimp = group_weight(p, nid, dist) - groupweight; in task_numa_migrate()
2592 if (env.best_cpu == -1) in task_numa_migrate()
2597 if (nid != p->numa_preferred_nid) in task_numa_migrate()
2602 if (env.best_cpu == -1) { in task_numa_migrate()
2603 trace_sched_stick_numa(p, env.src_cpu, NULL, -1); in task_numa_migrate()
2604 return -EAGAIN; in task_numa_migrate()
2610 WRITE_ONCE(best_rq->numa_migrate_on, 0); in task_numa_migrate()
2617 WRITE_ONCE(best_rq->numa_migrate_on, 0); in task_numa_migrate()
2631 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) in numa_migrate_preferred()
2635 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); in numa_migrate_preferred()
2636 p->numa_migrate_retry = jiffies + interval; in numa_migrate_preferred()
2639 if (task_node(p) == p->numa_preferred_nid) in numa_migrate_preferred()
2669 numa_group->max_faults_cpu = max_faults; in numa_group_count_active_nodes()
2670 numa_group->active_nodes = active_nodes; in numa_group_count_active_nodes()
2696 unsigned long remote = p->numa_faults_locality[0]; in update_task_scan_period()
2697 unsigned long local = p->numa_faults_locality[1]; in update_task_scan_period()
2706 if (local + shared == 0 || p->numa_faults_locality[2]) { in update_task_scan_period()
2707 p->numa_scan_period = min(p->numa_scan_period_max, in update_task_scan_period()
2708 p->numa_scan_period << 1); in update_task_scan_period()
2710 p->mm->numa_next_scan = jiffies + in update_task_scan_period()
2711 msecs_to_jiffies(p->numa_scan_period); in update_task_scan_period()
2722 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); in update_task_scan_period()
2731 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; in update_task_scan_period()
2741 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; in update_task_scan_period()
2747 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, in update_task_scan_period()
2752 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; in update_task_scan_period()
2755 p->numa_scan_period = clamp(p->numa_scan_period + diff, in update_task_scan_period()
2757 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); in update_task_scan_period()
2764 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2771 now = p->se.exec_start; in numa_get_avg_runtime()
2772 runtime = p->se.sum_exec_runtime; in numa_get_avg_runtime()
2774 if (p->last_task_numa_placement) { in numa_get_avg_runtime()
2775 delta = runtime - p->last_sum_exec_runtime; in numa_get_avg_runtime()
2776 *period = now - p->last_task_numa_placement; in numa_get_avg_runtime()
2782 delta = p->se.avg.load_sum; in numa_get_avg_runtime()
2786 p->last_sum_exec_runtime = runtime; in numa_get_avg_runtime()
2787 p->last_task_numa_placement = now; in numa_get_avg_runtime()
2837 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { in preferred_group_nid()
2891 * The p->mm->numa_scan_seq field gets updated without in task_numa_placement()
2895 seq = READ_ONCE(p->mm->numa_scan_seq); in task_numa_placement()
2896 if (p->numa_scan_seq == seq) in task_numa_placement()
2898 p->numa_scan_seq = seq; in task_numa_placement()
2899 p->numa_scan_period_max = task_scan_max(p); in task_numa_placement()
2901 total_faults = p->numa_faults_locality[0] + in task_numa_placement()
2902 p->numa_faults_locality[1]; in task_numa_placement()
2908 group_lock = &ng->lock; in task_numa_placement()
2928 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; in task_numa_placement()
2929 fault_types[priv] += p->numa_faults[membuf_idx]; in task_numa_placement()
2930 p->numa_faults[membuf_idx] = 0; in task_numa_placement()
2936 * little over-all impact on throughput, and thus their in task_numa_placement()
2940 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / in task_numa_placement()
2942 f_diff = f_weight - p->numa_faults[cpu_idx] / 2; in task_numa_placement()
2943 p->numa_faults[cpubuf_idx] = 0; in task_numa_placement()
2945 p->numa_faults[mem_idx] += diff; in task_numa_placement()
2946 p->numa_faults[cpu_idx] += f_diff; in task_numa_placement()
2947 faults += p->numa_faults[mem_idx]; in task_numa_placement()
2948 p->total_numa_faults += diff; in task_numa_placement()
2957 ng->faults[mem_idx] += diff; in task_numa_placement()
2958 ng->faults[cpu_idx] += f_diff; in task_numa_placement()
2959 ng->total_faults += diff; in task_numa_placement()
2960 group_faults += ng->faults[mem_idx]; in task_numa_placement()
2975 /* Cannot migrate task to CPU-less node */ in task_numa_placement()
2986 if (max_nid != p->numa_preferred_nid) in task_numa_placement()
2995 return refcount_inc_not_zero(&grp->refcount); in get_numa_group()
3000 if (refcount_dec_and_test(&grp->refcount)) in put_numa_group()
3022 refcount_set(&grp->refcount, 1); in task_numa_group()
3023 grp->active_nodes = 1; in task_numa_group()
3024 grp->max_faults_cpu = 0; in task_numa_group()
3025 spin_lock_init(&grp->lock); in task_numa_group()
3026 grp->gid = p->pid; in task_numa_group()
3029 grp->faults[i] = p->numa_faults[i]; in task_numa_group()
3031 grp->total_faults = p->total_numa_faults; in task_numa_group()
3033 grp->nr_tasks++; in task_numa_group()
3034 rcu_assign_pointer(p->numa_group, grp); in task_numa_group()
3038 tsk = READ_ONCE(cpu_rq(cpu)->curr); in task_numa_group()
3043 grp = rcu_dereference(tsk->numa_group); in task_numa_group()
3055 if (my_grp->nr_tasks > grp->nr_tasks) in task_numa_group()
3059 * Tie-break on the grp address. in task_numa_group()
3061 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) in task_numa_group()
3065 if (tsk->mm == current->mm) in task_numa_group()
3084 double_lock_irq(&my_grp->lock, &grp->lock); in task_numa_group()
3087 my_grp->faults[i] -= p->numa_faults[i]; in task_numa_group()
3088 grp->faults[i] += p->numa_faults[i]; in task_numa_group()
3090 my_grp->total_faults -= p->total_numa_faults; in task_numa_group()
3091 grp->total_faults += p->total_numa_faults; in task_numa_group()
3093 my_grp->nr_tasks--; in task_numa_group()
3094 grp->nr_tasks++; in task_numa_group()
3096 spin_unlock(&my_grp->lock); in task_numa_group()
3097 spin_unlock_irq(&grp->lock); in task_numa_group()
3099 rcu_assign_pointer(p->numa_group, grp); in task_numa_group()
3114 * reset the data back to default state without freeing ->numa_faults.
3119 struct numa_group *grp = rcu_dereference_raw(p->numa_group); in task_numa_free()
3120 unsigned long *numa_faults = p->numa_faults; in task_numa_free()
3128 spin_lock_irqsave(&grp->lock, flags); in task_numa_free()
3130 grp->faults[i] -= p->numa_faults[i]; in task_numa_free()
3131 grp->total_faults -= p->total_numa_faults; in task_numa_free()
3133 grp->nr_tasks--; in task_numa_free()
3134 spin_unlock_irqrestore(&grp->lock, flags); in task_numa_free()
3135 RCU_INIT_POINTER(p->numa_group, NULL); in task_numa_free()
3140 p->numa_faults = NULL; in task_numa_free()
3143 p->total_numa_faults = 0; in task_numa_free()
3165 if (!p->mm) in task_numa_fault()
3177 /* Allocate buffer to track faults on a per-node basis */ in task_numa_fault()
3178 if (unlikely(!p->numa_faults)) { in task_numa_fault()
3179 int size = sizeof(*p->numa_faults) * in task_numa_fault()
3182 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); in task_numa_fault()
3183 if (!p->numa_faults) in task_numa_fault()
3186 p->total_numa_faults = 0; in task_numa_fault()
3187 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); in task_numa_fault()
3194 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { in task_numa_fault()
3209 if (!priv && !local && ng && ng->active_nodes > 1 && in task_numa_fault()
3218 if (time_after(jiffies, p->numa_migrate_retry)) { in task_numa_fault()
3224 p->numa_pages_migrated += pages; in task_numa_fault()
3226 p->numa_faults_locality[2] += pages; in task_numa_fault()
3228 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; in task_numa_fault()
3229 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; in task_numa_fault()
3230 p->numa_faults_locality[local] += pages; in task_numa_fault()
3237 * p->mm->numa_scan_seq is written to without exclusive access in reset_ptenuma_scan()
3243 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); in reset_ptenuma_scan()
3244 p->mm->numa_scan_offset = 0; in reset_ptenuma_scan()
3256 if ((READ_ONCE(current->mm->numa_scan_seq) - vma->numab_state->start_scan_seq) < 2) in vma_is_accessed()
3259 pids = vma->numab_state->pids_active[0] | vma->numab_state->pids_active[1]; in vma_is_accessed()
3260 if (test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids)) in vma_is_accessed()
3265 * some VMAs may never be scanned in multi-threaded applications: in vma_is_accessed()
3267 if (mm->numa_scan_offset > vma->vm_start) { in vma_is_accessed()
3277 if (READ_ONCE(mm->numa_scan_seq) > in vma_is_accessed()
3278 (vma->numab_state->prev_scan_seq + get_nr_threads(current))) in vma_is_accessed()
3294 struct mm_struct *mm = p->mm; in task_numa_work()
3295 u64 runtime = p->se.sum_exec_runtime; in task_numa_work()
3306 work->next = work; in task_numa_work()
3310 * NOTE: make sure not to dereference p->mm before this check, in task_numa_work()
3312 * without p->mm even though we still had it when we enqueued this in task_numa_work()
3315 if (p->flags & PF_EXITING) in task_numa_work()
3318 if (!mm->numa_next_scan) { in task_numa_work()
3319 mm->numa_next_scan = now + in task_numa_work()
3326 migrate = mm->numa_next_scan; in task_numa_work()
3330 if (p->numa_scan_period == 0) { in task_numa_work()
3331 p->numa_scan_period_max = task_scan_max(p); in task_numa_work()
3332 p->numa_scan_period = task_scan_start(p); in task_numa_work()
3335 next_scan = now + msecs_to_jiffies(p->numa_scan_period); in task_numa_work()
3336 if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan)) in task_numa_work()
3343 p->node_stamp += 2 * TICK_NSEC; in task_numa_work()
3346 pages <<= 20 - PAGE_SHIFT; /* MB in pages */ in task_numa_work()
3363 start = mm->numa_scan_offset; in task_numa_work()
3375 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { in task_numa_work()
3383 * hinting faults in read-only file-backed mappings or the vDSO in task_numa_work()
3386 if (!vma->vm_mm || in task_numa_work()
3387 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) { in task_numa_work()
3401 /* Initialise new per-VMA NUMAB state. */ in task_numa_work()
3402 if (!vma->numab_state) { in task_numa_work()
3409 if (cmpxchg(&vma->numab_state, NULL, ptr)) { in task_numa_work()
3414 vma->numab_state->start_scan_seq = mm->numa_scan_seq; in task_numa_work()
3416 vma->numab_state->next_scan = now + in task_numa_work()
3420 vma->numab_state->pids_active_reset = vma->numab_state->next_scan + in task_numa_work()
3428 vma->numab_state->prev_scan_seq = mm->numa_scan_seq - 1; in task_numa_work()
3435 if (mm->numa_scan_seq && time_before(jiffies, in task_numa_work()
3436 vma->numab_state->next_scan)) { in task_numa_work()
3442 if (mm->numa_scan_seq && in task_numa_work()
3443 time_after(jiffies, vma->numab_state->pids_active_reset)) { in task_numa_work()
3444 vma->numab_state->pids_active_reset = vma->numab_state->pids_active_reset + in task_numa_work()
3446 vma->numab_state->pids_active[0] = READ_ONCE(vma->numab_state->pids_active[1]); in task_numa_work()
3447 vma->numab_state->pids_active[1] = 0; in task_numa_work()
3451 if (vma->numab_state->prev_scan_seq == mm->numa_scan_seq) { in task_numa_work()
3452 mm->numa_scan_offset = vma->vm_end; in task_numa_work()
3468 start = max(start, vma->vm_start); in task_numa_work()
3470 end = min(end, vma->vm_end); in task_numa_work()
3476 * is not already PTE-numa. If the VMA contains in task_numa_work()
3482 pages -= (end - start) >> PAGE_SHIFT; in task_numa_work()
3483 virtpages -= (end - start) >> PAGE_SHIFT; in task_numa_work()
3490 } while (end != vma->vm_end); in task_numa_work()
3493 vma->numab_state->prev_scan_seq = mm->numa_scan_seq; in task_numa_work()
3521 mm->numa_scan_offset = start; in task_numa_work()
3532 if (unlikely(p->se.sum_exec_runtime != runtime)) { in task_numa_work()
3533 u64 diff = p->se.sum_exec_runtime - runtime; in task_numa_work()
3534 p->node_stamp += 32 * diff; in task_numa_work()
3541 struct mm_struct *mm = p->mm; in init_numa_balancing()
3544 mm_users = atomic_read(&mm->mm_users); in init_numa_balancing()
3546 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); in init_numa_balancing()
3547 mm->numa_scan_seq = 0; in init_numa_balancing()
3550 p->node_stamp = 0; in init_numa_balancing()
3551 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; in init_numa_balancing()
3552 p->numa_scan_period = sysctl_numa_balancing_scan_delay; in init_numa_balancing()
3553 p->numa_migrate_retry = 0; in init_numa_balancing()
3555 p->numa_work.next = &p->numa_work; in init_numa_balancing()
3556 p->numa_faults = NULL; in init_numa_balancing()
3557 p->numa_pages_migrated = 0; in init_numa_balancing()
3558 p->total_numa_faults = 0; in init_numa_balancing()
3559 RCU_INIT_POINTER(p->numa_group, NULL); in init_numa_balancing()
3560 p->last_task_numa_placement = 0; in init_numa_balancing()
3561 p->last_sum_exec_runtime = 0; in init_numa_balancing()
3563 init_task_work(&p->numa_work, task_numa_work); in init_numa_balancing()
3567 p->numa_preferred_nid = NUMA_NO_NODE; in init_numa_balancing()
3579 current->numa_scan_period * mm_users * NSEC_PER_MSEC); in init_numa_balancing()
3581 p->node_stamp = delay; in init_numa_balancing()
3590 struct callback_head *work = &curr->numa_work; in task_tick_numa()
3596 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work) in task_tick_numa()
3605 now = curr->se.sum_exec_runtime; in task_tick_numa()
3606 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; in task_tick_numa()
3608 if (now > curr->node_stamp + period) { in task_tick_numa()
3609 if (!curr->node_stamp) in task_tick_numa()
3610 curr->numa_scan_period = task_scan_start(curr); in task_tick_numa()
3611 curr->node_stamp += period; in task_tick_numa()
3613 if (!time_before(jiffies, curr->mm->numa_next_scan)) in task_tick_numa()
3626 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) in update_scan_period()
3635 * is pulled cross-node due to wakeups or load balancing. in update_scan_period()
3637 if (p->numa_scan_seq) { in update_scan_period()
3643 if (dst_nid == p->numa_preferred_nid || in update_scan_period()
3644 (p->numa_preferred_nid != NUMA_NO_NODE && in update_scan_period()
3645 src_nid != p->numa_preferred_nid)) in update_scan_period()
3649 p->numa_scan_period = task_scan_start(p); in update_scan_period()
3674 update_load_add(&cfs_rq->load, se->load.weight); in account_entity_enqueue()
3680 list_add(&se->group_node, &rq->cfs_tasks); in account_entity_enqueue()
3683 cfs_rq->nr_queued++; in account_entity_enqueue()
3689 update_load_sub(&cfs_rq->load, se->load.weight); in account_entity_dequeue()
3693 list_del_init(&se->group_node); in account_entity_dequeue()
3696 cfs_rq->nr_queued--; in account_entity_dequeue()
3702 * Explicitly do a load-store to ensure the intermediate value never hits
3722 * Explicitly do a load-store to ensure the intermediate value never hits
3730 res = var - val; \
3739 * A variant of sub_positive(), which does not use explicit load-store
3744 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3751 cfs_rq->avg.load_avg += se->avg.load_avg; in enqueue_load_avg()
3752 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; in enqueue_load_avg()
3758 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); in dequeue_load_avg()
3759 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); in dequeue_load_avg()
3761 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum, in dequeue_load_avg()
3762 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER); in dequeue_load_avg()
3776 bool curr = cfs_rq->curr == se; in reweight_entity()
3778 if (se->on_rq) { in reweight_entity()
3782 se->deadline -= se->vruntime; in reweight_entity()
3783 se->rel_deadline = 1; in reweight_entity()
3786 update_load_sub(&cfs_rq->load, se->load.weight); in reweight_entity()
3791 * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i), in reweight_entity()
3792 * we need to scale se->vlag when w_i changes. in reweight_entity()
3794 se->vlag = div_s64(se->vlag * se->load.weight, weight); in reweight_entity()
3795 if (se->rel_deadline) in reweight_entity()
3796 se->deadline = div_s64(se->deadline * se->load.weight, weight); in reweight_entity()
3798 update_load_set(&se->load, weight); in reweight_entity()
3802 u32 divider = get_pelt_divider(&se->avg); in reweight_entity()
3804 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); in reweight_entity()
3809 if (se->on_rq) { in reweight_entity()
3810 update_load_add(&cfs_rq->load, se->load.weight); in reweight_entity()
3817 * whether the rq-wide min_vruntime needs updated too. Since in reweight_entity()
3819 * than up-to-date one, we do the update at the end of the in reweight_entity()
3829 struct sched_entity *se = &p->se; in reweight_task_fair()
3831 struct load_weight *load = &se->load; in reweight_task_fair()
3833 reweight_entity(cfs_rq, se, lw->weight); in reweight_task_fair()
3834 load->inv_weight = lw->inv_weight; in reweight_task_fair()
3848 * tg->weight * grq->load.weight
3849 * ge->load.weight = ----------------------------- (1)
3850 * \Sum grq->load.weight
3858 * grq->load.weight -> grq->avg.load_avg (2)
3862 * tg->weight * grq->avg.load_avg
3863 * ge->load.weight = ------------------------------ (3)
3864 * tg->load_avg
3866 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3870 * The problem with it is that because the average is slow -- it was designed
3871 * to be exactly that of course -- this leads to transients in boundary
3873 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3878 * tg->weight * grq->load.weight
3879 * ge->load.weight = ----------------------------- = tg->weight (4)
3880 * grp->load.weight
3887 * ge->load.weight =
3889 * tg->weight * grq->load.weight
3890 * --------------------------------------------------- (5)
3891 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3893 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3894 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3897 * tg->weight * grq->load.weight
3898 * ge->load.weight = ----------------------------- (6)
3903 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3904 * max(grq->load.weight, grq->avg.load_avg)
3908 * overestimates the ge->load.weight and therefore:
3910 * \Sum ge->load.weight >= tg->weight
3917 struct task_group *tg = cfs_rq->tg; in calc_group_shares()
3919 tg_shares = READ_ONCE(tg->shares); in calc_group_shares()
3921 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); in calc_group_shares()
3923 tg_weight = atomic_long_read(&tg->load_avg); in calc_group_shares()
3926 tg_weight -= cfs_rq->tg_load_avg_contrib; in calc_group_shares()
3934 * MIN_SHARES has to be unscaled here to support per-CPU partitioning in calc_group_shares()
3935 * of a group with small tg->shares value. It is a floor value which is in calc_group_shares()
3939 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 in calc_group_shares()
3940 * on an 8-core system with 8 tasks each runnable on one CPU shares has in calc_group_shares()
3962 if (!gcfs_rq || !gcfs_rq->load.weight) in update_cfs_group()
3969 shares = READ_ONCE(gcfs_rq->tg->shares); in update_cfs_group()
3973 if (unlikely(se->load.weight != shares)) in update_cfs_group()
3987 if (&rq->cfs == cfs_rq) { in cfs_rq_util_change()
3997 * As is, the util number is not freq-invariant (we'd have to in cfs_rq_util_change()
4009 if (sa->load_sum) in load_avg_is_decayed()
4012 if (sa->util_sum) in load_avg_is_decayed()
4015 if (sa->runnable_sum) in load_avg_is_decayed()
4023 SCHED_WARN_ON(sa->load_avg || in load_avg_is_decayed()
4024 sa->util_avg || in load_avg_is_decayed()
4025 sa->runnable_avg); in load_avg_is_decayed()
4032 return u64_u32_load_copy(cfs_rq->avg.last_update_time, in cfs_rq_last_update_time()
4033 cfs_rq->last_update_time_copy); in cfs_rq_last_update_time()
4039 * bottom-up, we only have to test whether the cfs_rq before us on the list
4050 if (cfs_rq->on_list) { in child_cfs_rq_on_list()
4051 prev = cfs_rq->leaf_cfs_rq_list.prev; in child_cfs_rq_on_list()
4053 prev = rq->tmp_alone_branch; in child_cfs_rq_on_list()
4056 if (prev == &rq->leaf_cfs_rq_list) in child_cfs_rq_on_list()
4061 return (prev_cfs_rq->tg->parent == cfs_rq->tg); in child_cfs_rq_on_list()
4066 if (cfs_rq->load.weight) in cfs_rq_is_decayed()
4069 if (!load_avg_is_decayed(&cfs_rq->avg)) in cfs_rq_is_decayed()
4079 * update_tg_load_avg - update the tg's load avg
4082 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
4083 * However, because tg->load_avg is a global value there are performance
4087 * differential update where we store the last value we propagated. This in
4088 * turn allows skipping updates if the differential is 'small'.
4100 if (cfs_rq->tg == &root_task_group) in update_tg_load_avg()
4108 * For migration heavy workloads, access to tg->load_avg can be in update_tg_load_avg()
4112 if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC) in update_tg_load_avg()
4115 delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; in update_tg_load_avg()
4116 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { in update_tg_load_avg()
4117 atomic_long_add(delta, &cfs_rq->tg->load_avg); in update_tg_load_avg()
4118 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; in update_tg_load_avg()
4119 cfs_rq->last_update_tg_load_avg = now; in update_tg_load_avg()
4131 if (cfs_rq->tg == &root_task_group) in clear_tg_load_avg()
4135 delta = 0 - cfs_rq->tg_load_avg_contrib; in clear_tg_load_avg()
4136 atomic_long_add(delta, &cfs_rq->tg->load_avg); in clear_tg_load_avg()
4137 cfs_rq->tg_load_avg_contrib = 0; in clear_tg_load_avg()
4138 cfs_rq->last_update_tg_load_avg = now; in clear_tg_load_avg()
4157 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; in clear_tg_offline_cfs_rqs()
4168 * caller only guarantees p->pi_lock is held; no other assumptions,
4169 * including the state of rq->lock, should be made.
4183 * getting what current time is, so simply throw away the out-of-date in set_task_rq_fair()
4187 if (!(se->avg.last_update_time && prev)) in set_task_rq_fair()
4194 se->avg.last_update_time = n_last_update_time; in set_task_rq_fair()
4202 * ge->avg == grq->avg (1)
4213 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
4218 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
4222 * ge->avg.runnable_avg == grq->avg.runnable_avg
4226 * ge->load.weight * grq->avg.load_avg
4227 * ge->avg.load_avg = ----------------------------------- (4)
4228 * grq->load.weight
4241 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4252 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4259 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4267 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg; in update_tg_cfs_util()
4275 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. in update_tg_cfs_util()
4278 divider = get_pelt_divider(&cfs_rq->avg); in update_tg_cfs_util()
4282 se->avg.util_avg = gcfs_rq->avg.util_avg; in update_tg_cfs_util()
4283 new_sum = se->avg.util_avg * divider; in update_tg_cfs_util()
4284 delta_sum = (long)new_sum - (long)se->avg.util_sum; in update_tg_cfs_util()
4285 se->avg.util_sum = new_sum; in update_tg_cfs_util()
4288 add_positive(&cfs_rq->avg.util_avg, delta_avg); in update_tg_cfs_util()
4289 add_positive(&cfs_rq->avg.util_sum, delta_sum); in update_tg_cfs_util()
4292 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum, in update_tg_cfs_util()
4293 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER); in update_tg_cfs_util()
4299 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg; in update_tg_cfs_runnable()
4307 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. in update_tg_cfs_runnable()
4310 divider = get_pelt_divider(&cfs_rq->avg); in update_tg_cfs_runnable()
4313 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg; in update_tg_cfs_runnable()
4314 new_sum = se->avg.runnable_avg * divider; in update_tg_cfs_runnable()
4315 delta_sum = (long)new_sum - (long)se->avg.runnable_sum; in update_tg_cfs_runnable()
4316 se->avg.runnable_sum = new_sum; in update_tg_cfs_runnable()
4319 add_positive(&cfs_rq->avg.runnable_avg, delta_avg); in update_tg_cfs_runnable()
4320 add_positive(&cfs_rq->avg.runnable_sum, delta_sum); in update_tg_cfs_runnable()
4322 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum, in update_tg_cfs_runnable()
4323 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER); in update_tg_cfs_runnable()
4329 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; in update_tg_cfs_load()
4338 gcfs_rq->prop_runnable_sum = 0; in update_tg_cfs_load()
4341 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. in update_tg_cfs_load()
4344 divider = get_pelt_divider(&cfs_rq->avg); in update_tg_cfs_load()
4351 runnable_sum += se->avg.load_sum; in update_tg_cfs_load()
4358 if (scale_load_down(gcfs_rq->load.weight)) { in update_tg_cfs_load()
4359 load_sum = div_u64(gcfs_rq->avg.load_sum, in update_tg_cfs_load()
4360 scale_load_down(gcfs_rq->load.weight)); in update_tg_cfs_load()
4364 runnable_sum = min(se->avg.load_sum, load_sum); in update_tg_cfs_load()
4373 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; in update_tg_cfs_load()
4379 delta_avg = load_avg - se->avg.load_avg; in update_tg_cfs_load()
4383 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum; in update_tg_cfs_load()
4385 se->avg.load_sum = runnable_sum; in update_tg_cfs_load()
4386 se->avg.load_avg = load_avg; in update_tg_cfs_load()
4387 add_positive(&cfs_rq->avg.load_avg, delta_avg); in update_tg_cfs_load()
4388 add_positive(&cfs_rq->avg.load_sum, delta_sum); in update_tg_cfs_load()
4390 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum, in update_tg_cfs_load()
4391 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER); in update_tg_cfs_load()
4396 cfs_rq->propagate = 1; in add_tg_cfs_propagate()
4397 cfs_rq->prop_runnable_sum += runnable_sum; in add_tg_cfs_propagate()
4409 if (!gcfs_rq->propagate) in propagate_entity_load_avg()
4412 gcfs_rq->propagate = 0; in propagate_entity_load_avg()
4416 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); in propagate_entity_load_avg()
4440 if (se->avg.load_avg || se->avg.util_avg) in skip_blocked_update()
4447 if (gcfs_rq->propagate) in skip_blocked_update()
4481 if (load_avg_is_decayed(&se->avg)) in migrate_se_pelt_lag()
4488 is_idle = is_idle_task(rcu_dereference(rq->curr)); in migrate_se_pelt_lag()
4505 * - cfs->throttled_clock_pelt_time@cfs_rq_idle in migrate_se_pelt_lag()
4508 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle in migrate_se_pelt_lag()
4511 * = sched_clock_cpu() - rq_clock()@rq_idle in migrate_se_pelt_lag()
4515 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time + in migrate_se_pelt_lag()
4516 * sched_clock_cpu() - rq_clock()@rq_idle in migrate_se_pelt_lag()
4518 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle in migrate_se_pelt_lag()
4519 * rq_clock()@rq_idle is rq->clock_idle in migrate_se_pelt_lag()
4520 * cfs->throttled_clock_pelt_time@cfs_rq_idle in migrate_se_pelt_lag()
4521 * is cfs_rq->throttled_pelt_idle in migrate_se_pelt_lag()
4525 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle); in migrate_se_pelt_lag()
4530 now = u64_u32_load(rq->clock_pelt_idle); in migrate_se_pelt_lag()
4540 now -= throttled; in migrate_se_pelt_lag()
4543 * cfs_rq->avg.last_update_time is more recent than our in migrate_se_pelt_lag()
4548 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle); in migrate_se_pelt_lag()
4557 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4564 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4568 * Since both these conditions indicate a changed cfs_rq->avg.load we should
4575 struct sched_avg *sa = &cfs_rq->avg; in update_cfs_rq_load_avg()
4578 if (cfs_rq->removed.nr) { in update_cfs_rq_load_avg()
4580 u32 divider = get_pelt_divider(&cfs_rq->avg); in update_cfs_rq_load_avg()
4582 raw_spin_lock(&cfs_rq->removed.lock); in update_cfs_rq_load_avg()
4583 swap(cfs_rq->removed.util_avg, removed_util); in update_cfs_rq_load_avg()
4584 swap(cfs_rq->removed.load_avg, removed_load); in update_cfs_rq_load_avg()
4585 swap(cfs_rq->removed.runnable_avg, removed_runnable); in update_cfs_rq_load_avg()
4586 cfs_rq->removed.nr = 0; in update_cfs_rq_load_avg()
4587 raw_spin_unlock(&cfs_rq->removed.lock); in update_cfs_rq_load_avg()
4590 sub_positive(&sa->load_avg, r); in update_cfs_rq_load_avg()
4591 sub_positive(&sa->load_sum, r * divider); in update_cfs_rq_load_avg()
4592 /* See sa->util_sum below */ in update_cfs_rq_load_avg()
4593 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER); in update_cfs_rq_load_avg()
4596 sub_positive(&sa->util_avg, r); in update_cfs_rq_load_avg()
4597 sub_positive(&sa->util_sum, r * divider); in update_cfs_rq_load_avg()
4599 * Because of rounding, se->util_sum might ends up being +1 more than in update_cfs_rq_load_avg()
4600 * cfs->util_sum. Although this is not a problem by itself, detaching in update_cfs_rq_load_avg()
4602 * util_avg (~1ms) can make cfs->util_sum becoming null whereas in update_cfs_rq_load_avg()
4609 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER); in update_cfs_rq_load_avg()
4612 sub_positive(&sa->runnable_avg, r); in update_cfs_rq_load_avg()
4613 sub_positive(&sa->runnable_sum, r * divider); in update_cfs_rq_load_avg()
4614 /* See sa->util_sum above */ in update_cfs_rq_load_avg()
4615 sa->runnable_sum = max_t(u32, sa->runnable_sum, in update_cfs_rq_load_avg()
4616 sa->runnable_avg * PELT_MIN_DIVIDER); in update_cfs_rq_load_avg()
4623 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT); in update_cfs_rq_load_avg()
4629 u64_u32_store_copy(sa->last_update_time, in update_cfs_rq_load_avg()
4630 cfs_rq->last_update_time_copy, in update_cfs_rq_load_avg()
4631 sa->last_update_time); in update_cfs_rq_load_avg()
4636 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4641 * cfs_rq->avg.last_update_time being current.
4646 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. in attach_entity_load_avg()
4649 u32 divider = get_pelt_divider(&cfs_rq->avg); in attach_entity_load_avg()
4658 se->avg.last_update_time = cfs_rq->avg.last_update_time; in attach_entity_load_avg()
4659 se->avg.period_contrib = cfs_rq->avg.period_contrib; in attach_entity_load_avg()
4667 se->avg.util_sum = se->avg.util_avg * divider; in attach_entity_load_avg()
4669 se->avg.runnable_sum = se->avg.runnable_avg * divider; in attach_entity_load_avg()
4671 se->avg.load_sum = se->avg.load_avg * divider; in attach_entity_load_avg()
4672 if (se_weight(se) < se->avg.load_sum) in attach_entity_load_avg()
4673 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se)); in attach_entity_load_avg()
4675 se->avg.load_sum = 1; in attach_entity_load_avg()
4678 cfs_rq->avg.util_avg += se->avg.util_avg; in attach_entity_load_avg()
4679 cfs_rq->avg.util_sum += se->avg.util_sum; in attach_entity_load_avg()
4680 cfs_rq->avg.runnable_avg += se->avg.runnable_avg; in attach_entity_load_avg()
4681 cfs_rq->avg.runnable_sum += se->avg.runnable_sum; in attach_entity_load_avg()
4683 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); in attach_entity_load_avg()
4691 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4696 * cfs_rq->avg.last_update_time being current.
4701 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); in detach_entity_load_avg()
4702 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); in detach_entity_load_avg()
4704 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum, in detach_entity_load_avg()
4705 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER); in detach_entity_load_avg()
4707 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg); in detach_entity_load_avg()
4708 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum); in detach_entity_load_avg()
4710 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum, in detach_entity_load_avg()
4711 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER); in detach_entity_load_avg()
4713 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); in detach_entity_load_avg()
4738 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) in update_load_avg()
4744 if (!se->avg.last_update_time && (flags & DO_ATTACH)) { in update_load_avg()
4794 * tasks cannot exit without having gone through wake_up_new_task() -> in remove_entity_load_avg()
4801 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); in remove_entity_load_avg()
4802 ++cfs_rq->removed.nr; in remove_entity_load_avg()
4803 cfs_rq->removed.util_avg += se->avg.util_avg; in remove_entity_load_avg()
4804 cfs_rq->removed.load_avg += se->avg.load_avg; in remove_entity_load_avg()
4805 cfs_rq->removed.runnable_avg += se->avg.runnable_avg; in remove_entity_load_avg()
4806 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); in remove_entity_load_avg()
4811 return cfs_rq->avg.runnable_avg; in cfs_rq_runnable_avg()
4816 return cfs_rq->avg.load_avg; in cfs_rq_load_avg()
4823 return READ_ONCE(p->se.avg.util_avg); in task_util()
4828 return READ_ONCE(p->se.avg.runnable_avg); in task_runnable()
4833 return READ_ONCE(p->se.avg.util_est) & ~UTIL_AVG_UNCHANGED; in _task_util_est()
4850 enqueued = cfs_rq->avg.util_est; in util_est_enqueue()
4852 WRITE_ONCE(cfs_rq->avg.util_est, enqueued); in util_est_enqueue()
4866 enqueued = cfs_rq->avg.util_est; in util_est_dequeue()
4867 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p)); in util_est_dequeue()
4868 WRITE_ONCE(cfs_rq->avg.util_est, enqueued); in util_est_dequeue()
4892 ewma = READ_ONCE(p->se.avg.util_est); in util_est_update()
4917 last_ewma_diff = ewma - dequeued; in util_est_update()
4943 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) in util_est_update()
4944 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) in util_est_update()
4945 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) in util_est_update()
4946 * = w * ( -last_ewma_diff ) + ewma(t-1) in util_est_update()
4947 * = w * (-last_ewma_diff + ewma(t-1) / w) in util_est_update()
4953 ewma -= last_ewma_diff; in util_est_update()
4957 WRITE_ONCE(p->se.avg.util_est, ewma); in util_est_update()
4959 trace_sched_util_est_se_tp(&p->se); in util_est_update()
4966 capacity -= max(hw_load_avg(cpu_rq(cpu)), cpufreq_get_pressure(cpu)); in get_actual_cpu_capacity()
5024 * +---------------------------------------- in util_fits_cpu()
5062 * +---------------------------------------- in util_fits_cpu()
5085 return -1; in util_fits_cpu()
5113 if (!p || (p->nr_cpus_allowed == 1) || in update_misfit_status()
5114 (arch_scale_cpu_capacity(cpu) == p->max_allowed_capacity) || in update_misfit_status()
5117 rq->misfit_task_load = 0; in update_misfit_status()
5125 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1); in update_misfit_status()
5132 return !cfs_rq->nr_queued; in cfs_rq_is_decayed()
5172 struct sched_entity *se = &p->se; in __setparam_fair()
5174 p->static_prio = NICE_TO_PRIO(attr->sched_nice); in __setparam_fair()
5175 if (attr->sched_runtime) { in __setparam_fair()
5176 se->custom_slice = 1; in __setparam_fair()
5177 se->slice = clamp_t(u64, attr->sched_runtime, in __setparam_fair()
5181 se->custom_slice = 0; in __setparam_fair()
5182 se->slice = sysctl_sched_base_slice; in __setparam_fair()
5192 if (!se->custom_slice) in place_entity()
5193 se->slice = sysctl_sched_base_slice; in place_entity()
5194 vslice = calc_delta_fair(se->slice, se); in place_entity()
5204 if (sched_feat(PLACE_LAG) && cfs_rq->nr_queued && se->vlag) { in place_entity()
5205 struct sched_entity *curr = cfs_rq->curr; in place_entity()
5208 lag = se->vlag; in place_entity()
5218 * lag_i = S - s_i = w_i * (V - v_i) in place_entity()
5223 * vl_i = V - v_i <=> v_i = V - vl_i in place_entity()
5235 * = (W*V + w_i*(V - vl_i)) / (W + w_i) in place_entity()
5236 * = (W*V + w_i*V - w_i*vl_i) / (W + w_i) in place_entity()
5237 * = (V*(W + w_i) - w_i*l) / (W + w_i) in place_entity()
5238 * = V - w_i*vl_i / (W + w_i) in place_entity()
5242 * vl'_i = V' - v_i in place_entity()
5243 * = V - w_i*vl_i / (W + w_i) - (V - vl_i) in place_entity()
5244 * = vl_i - w_i*vl_i / (W + w_i) in place_entity()
5254 * vl'_i = vl_i - w_i*vl_i / (W + w_i) in place_entity()
5255 * = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i) in place_entity()
5257 * (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i in place_entity()
5262 load = cfs_rq->avg_load; in place_entity()
5263 if (curr && curr->on_rq) in place_entity()
5264 load += scale_load_down(curr->load.weight); in place_entity()
5266 lag *= load + scale_load_down(se->load.weight); in place_entity()
5272 se->vruntime = vruntime - lag; in place_entity()
5274 if (se->rel_deadline) { in place_entity()
5275 se->deadline += se->vruntime; in place_entity()
5276 se->rel_deadline = 0; in place_entity()
5291 se->deadline = se->vruntime + vslice; in place_entity()
5303 bool curr = cfs_rq->curr == se; in enqueue_entity()
5316 * - Update loads to have both entity and cfs_rq synced with now. in enqueue_entity()
5317 * - For group_entity, update its runnable_weight to reflect the new in enqueue_entity()
5319 * - For group_entity, update its weight to reflect the new share of in enqueue_entity()
5321 * - Add its new weight to cfs_rq->load.weight in enqueue_entity()
5327 * but update_cfs_group() here will re-adjust the weight and have to in enqueue_entity()
5333 * XXX now that the entity has been re-weighted, and it's lag adjusted, in enqueue_entity()
5343 se->exec_start = 0; in enqueue_entity()
5349 se->on_rq = 1; in enqueue_entity()
5351 if (cfs_rq->nr_queued == 1) { in enqueue_entity()
5359 if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock) in enqueue_entity()
5360 cfs_rq->throttled_clock = rq_clock(rq); in enqueue_entity()
5361 if (!cfs_rq->throttled_clock_self) in enqueue_entity()
5362 cfs_rq->throttled_clock_self = rq_clock(rq); in enqueue_entity()
5372 if (cfs_rq->next != se) in __clear_buddies_next()
5375 cfs_rq->next = NULL; in __clear_buddies_next()
5381 if (cfs_rq->next == se) in clear_buddies()
5389 se->sched_delayed = 1; in set_delayed()
5402 cfs_rq->h_nr_runnable--; in set_delayed()
5410 se->sched_delayed = 0; in clear_delayed()
5424 cfs_rq->h_nr_runnable++; in clear_delayed()
5433 if (sched_feat(DELAY_ZERO) && se->vlag > 0) in finish_delayed_dequeue_entity()
5434 se->vlag = 0; in finish_delayed_dequeue_entity()
5447 SCHED_WARN_ON(!se->sched_delayed); in dequeue_entity()
5457 SCHED_WARN_ON(delay && se->sched_delayed); in dequeue_entity()
5472 * - Update loads to have both entity and cfs_rq synced with now. in dequeue_entity()
5473 * - For group_entity, update its runnable_weight to reflect the new in dequeue_entity()
5475 * - Subtract its previous weight from cfs_rq->load.weight. in dequeue_entity()
5476 * - For group entity, update its weight to reflect the new share in dequeue_entity()
5486 se->deadline -= se->vruntime; in dequeue_entity()
5487 se->rel_deadline = 1; in dequeue_entity()
5490 if (se != cfs_rq->curr) in dequeue_entity()
5492 se->on_rq = 0; in dequeue_entity()
5504 * further than we started -- i.e. we'll be penalized. in dequeue_entity()
5512 if (cfs_rq->nr_queued == 0) in dequeue_entity()
5524 if (se->on_rq) { in set_next_entity()
5537 se->vlag = se->deadline; in set_next_entity()
5541 SCHED_WARN_ON(cfs_rq->curr); in set_next_entity()
5542 cfs_rq->curr = se; in set_next_entity()
5547 * when there are only lesser-weight tasks around): in set_next_entity()
5550 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { in set_next_entity()
5554 __schedstat_set(stats->slice_max, in set_next_entity()
5555 max((u64)stats->slice_max, in set_next_entity()
5556 se->sum_exec_runtime - se->prev_sum_exec_runtime)); in set_next_entity()
5559 se->prev_sum_exec_runtime = se->sum_exec_runtime; in set_next_entity()
5577 * Picking the ->next buddy will affect latency but not fairness. in pick_next_entity()
5580 cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next)) { in pick_next_entity()
5581 /* ->next will never be delayed */ in pick_next_entity()
5582 SCHED_WARN_ON(cfs_rq->next->sched_delayed); in pick_next_entity()
5583 return cfs_rq->next; in pick_next_entity()
5587 if (se->sched_delayed) { in pick_next_entity()
5605 if (prev->on_rq) in put_prev_entity()
5611 if (prev->on_rq) { in put_prev_entity()
5618 SCHED_WARN_ON(cfs_rq->curr != prev); in put_prev_entity()
5619 cfs_rq->curr = NULL; in put_prev_entity()
5626 * Update run-time statistics of the 'current'. in entity_tick()
5698 * directly instead of rq->clock to avoid adding additional synchronization
5699 * around rq->lock.
5701 * requires cfs_b->lock
5707 if (unlikely(cfs_b->quota == RUNTIME_INF)) in __refill_cfs_bandwidth_runtime()
5710 cfs_b->runtime += cfs_b->quota; in __refill_cfs_bandwidth_runtime()
5711 runtime = cfs_b->runtime_snap - cfs_b->runtime; in __refill_cfs_bandwidth_runtime()
5713 cfs_b->burst_time += runtime; in __refill_cfs_bandwidth_runtime()
5714 cfs_b->nr_burst++; in __refill_cfs_bandwidth_runtime()
5717 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst); in __refill_cfs_bandwidth_runtime()
5718 cfs_b->runtime_snap = cfs_b->runtime; in __refill_cfs_bandwidth_runtime()
5723 return &tg->cfs_bandwidth; in tg_cfs_bandwidth()
5732 lockdep_assert_held(&cfs_b->lock); in __assign_cfs_rq_runtime()
5735 min_amount = target_runtime - cfs_rq->runtime_remaining; in __assign_cfs_rq_runtime()
5737 if (cfs_b->quota == RUNTIME_INF) in __assign_cfs_rq_runtime()
5742 if (cfs_b->runtime > 0) { in __assign_cfs_rq_runtime()
5743 amount = min(cfs_b->runtime, min_amount); in __assign_cfs_rq_runtime()
5744 cfs_b->runtime -= amount; in __assign_cfs_rq_runtime()
5745 cfs_b->idle = 0; in __assign_cfs_rq_runtime()
5749 cfs_rq->runtime_remaining += amount; in __assign_cfs_rq_runtime()
5751 return cfs_rq->runtime_remaining > 0; in __assign_cfs_rq_runtime()
5757 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); in assign_cfs_rq_runtime()
5760 raw_spin_lock(&cfs_b->lock); in assign_cfs_rq_runtime()
5762 raw_spin_unlock(&cfs_b->lock); in assign_cfs_rq_runtime()
5770 cfs_rq->runtime_remaining -= delta_exec; in __account_cfs_rq_runtime()
5772 if (likely(cfs_rq->runtime_remaining > 0)) in __account_cfs_rq_runtime()
5775 if (cfs_rq->throttled) in __account_cfs_rq_runtime()
5781 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) in __account_cfs_rq_runtime()
5788 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) in account_cfs_rq_runtime()
5796 return cfs_bandwidth_used() && cfs_rq->throttled; in cfs_rq_throttled()
5802 return cfs_bandwidth_used() && cfs_rq->throttle_count; in throttled_hierarchy()
5808 * load-balance operations.
5815 src_cfs_rq = tg->cfs_rq[src_cpu]; in throttled_lb_pair()
5816 dest_cfs_rq = tg->cfs_rq[dest_cpu]; in throttled_lb_pair()
5825 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; in tg_unthrottle_up()
5827 cfs_rq->throttle_count--; in tg_unthrottle_up()
5828 if (!cfs_rq->throttle_count) { in tg_unthrottle_up()
5829 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) - in tg_unthrottle_up()
5830 cfs_rq->throttled_clock_pelt; in tg_unthrottle_up()
5836 if (cfs_rq->throttled_clock_self) { in tg_unthrottle_up()
5837 u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self; in tg_unthrottle_up()
5839 cfs_rq->throttled_clock_self = 0; in tg_unthrottle_up()
5844 cfs_rq->throttled_clock_self_time += delta; in tg_unthrottle_up()
5854 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; in tg_throttle_down()
5857 if (!cfs_rq->throttle_count) { in tg_throttle_down()
5858 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq); in tg_throttle_down()
5861 SCHED_WARN_ON(cfs_rq->throttled_clock_self); in tg_throttle_down()
5862 if (cfs_rq->nr_queued) in tg_throttle_down()
5863 cfs_rq->throttled_clock_self = rq_clock(rq); in tg_throttle_down()
5865 cfs_rq->throttle_count++; in tg_throttle_down()
5873 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); in throttle_cfs_rq()
5876 long rq_h_nr_queued = rq->cfs.h_nr_queued; in throttle_cfs_rq()
5878 raw_spin_lock(&cfs_b->lock); in throttle_cfs_rq()
5891 list_add_tail_rcu(&cfs_rq->throttled_list, in throttle_cfs_rq()
5892 &cfs_b->throttled_cfs_rq); in throttle_cfs_rq()
5894 raw_spin_unlock(&cfs_b->lock); in throttle_cfs_rq()
5899 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; in throttle_cfs_rq()
5903 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); in throttle_cfs_rq()
5906 queued_delta = cfs_rq->h_nr_queued; in throttle_cfs_rq()
5907 runnable_delta = cfs_rq->h_nr_runnable; in throttle_cfs_rq()
5908 idle_delta = cfs_rq->h_nr_idle; in throttle_cfs_rq()
5913 /* throttled entity or throttle-on-deactivate */ in throttle_cfs_rq()
5914 if (!se->on_rq) in throttle_cfs_rq()
5923 if (se->sched_delayed) in throttle_cfs_rq()
5928 idle_delta = cfs_rq->h_nr_queued; in throttle_cfs_rq()
5930 qcfs_rq->h_nr_queued -= queued_delta; in throttle_cfs_rq()
5931 qcfs_rq->h_nr_runnable -= runnable_delta; in throttle_cfs_rq()
5932 qcfs_rq->h_nr_idle -= idle_delta; in throttle_cfs_rq()
5934 if (qcfs_rq->load.weight) { in throttle_cfs_rq()
5935 /* Avoid re-evaluating load for this entity: */ in throttle_cfs_rq()
5943 /* throttled entity or throttle-on-deactivate */ in throttle_cfs_rq()
5944 if (!se->on_rq) in throttle_cfs_rq()
5951 idle_delta = cfs_rq->h_nr_queued; in throttle_cfs_rq()
5953 qcfs_rq->h_nr_queued -= queued_delta; in throttle_cfs_rq()
5954 qcfs_rq->h_nr_runnable -= runnable_delta; in throttle_cfs_rq()
5955 qcfs_rq->h_nr_idle -= idle_delta; in throttle_cfs_rq()
5962 if (rq_h_nr_queued && !rq->cfs.h_nr_queued) in throttle_cfs_rq()
5963 dl_server_stop(&rq->fair_server); in throttle_cfs_rq()
5967 * throttled-list. rq->lock protects completion. in throttle_cfs_rq()
5969 cfs_rq->throttled = 1; in throttle_cfs_rq()
5970 SCHED_WARN_ON(cfs_rq->throttled_clock); in throttle_cfs_rq()
5971 if (cfs_rq->nr_queued) in throttle_cfs_rq()
5972 cfs_rq->throttled_clock = rq_clock(rq); in throttle_cfs_rq()
5979 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); in unthrottle_cfs_rq()
5982 long rq_h_nr_queued = rq->cfs.h_nr_queued; in unthrottle_cfs_rq()
5984 se = cfs_rq->tg->se[cpu_of(rq)]; in unthrottle_cfs_rq()
5986 cfs_rq->throttled = 0; in unthrottle_cfs_rq()
5990 raw_spin_lock(&cfs_b->lock); in unthrottle_cfs_rq()
5991 if (cfs_rq->throttled_clock) { in unthrottle_cfs_rq()
5992 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; in unthrottle_cfs_rq()
5993 cfs_rq->throttled_clock = 0; in unthrottle_cfs_rq()
5995 list_del_rcu(&cfs_rq->throttled_list); in unthrottle_cfs_rq()
5996 raw_spin_unlock(&cfs_b->lock); in unthrottle_cfs_rq()
5999 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); in unthrottle_cfs_rq()
6001 if (!cfs_rq->load.weight) { in unthrottle_cfs_rq()
6002 if (!cfs_rq->on_list) in unthrottle_cfs_rq()
6015 queued_delta = cfs_rq->h_nr_queued; in unthrottle_cfs_rq()
6016 runnable_delta = cfs_rq->h_nr_runnable; in unthrottle_cfs_rq()
6017 idle_delta = cfs_rq->h_nr_idle; in unthrottle_cfs_rq()
6022 if (se->sched_delayed) { in unthrottle_cfs_rq()
6026 } else if (se->on_rq) in unthrottle_cfs_rq()
6031 idle_delta = cfs_rq->h_nr_queued; in unthrottle_cfs_rq()
6033 qcfs_rq->h_nr_queued += queued_delta; in unthrottle_cfs_rq()
6034 qcfs_rq->h_nr_runnable += runnable_delta; in unthrottle_cfs_rq()
6035 qcfs_rq->h_nr_idle += idle_delta; in unthrottle_cfs_rq()
6049 idle_delta = cfs_rq->h_nr_queued; in unthrottle_cfs_rq()
6051 qcfs_rq->h_nr_queued += queued_delta; in unthrottle_cfs_rq()
6052 qcfs_rq->h_nr_runnable += runnable_delta; in unthrottle_cfs_rq()
6053 qcfs_rq->h_nr_idle += idle_delta; in unthrottle_cfs_rq()
6060 /* Start the fair server if un-throttling resulted in new runnable tasks */ in unthrottle_cfs_rq()
6061 if (!rq_h_nr_queued && rq->cfs.h_nr_queued) in unthrottle_cfs_rq()
6062 dl_server_start(&rq->fair_server); in unthrottle_cfs_rq()
6071 if (rq->curr == rq->idle && rq->cfs.nr_queued) in unthrottle_cfs_rq()
6095 * fact that we pair with sched_free_group_rcu(), so that we cannot in __cfsb_csd_unthrottle()
6101 list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list, in __cfsb_csd_unthrottle()
6103 list_del_init(&cursor->throttled_csd_list); in __cfsb_csd_unthrottle()
6126 if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list))) in __unthrottle_cfs_rq_async()
6129 first = list_empty(&rq->cfsb_csd_list); in __unthrottle_cfs_rq_async()
6130 list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list); in __unthrottle_cfs_rq_async()
6132 smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd); in __unthrottle_cfs_rq_async()
6146 cfs_rq->runtime_remaining <= 0)) in unthrottle_cfs_rq_async()
6163 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, in distribute_cfs_runtime()
6177 if (!list_empty(&cfs_rq->throttled_csd_list)) in distribute_cfs_runtime()
6181 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0); in distribute_cfs_runtime()
6183 raw_spin_lock(&cfs_b->lock); in distribute_cfs_runtime()
6184 runtime = -cfs_rq->runtime_remaining + 1; in distribute_cfs_runtime()
6185 if (runtime > cfs_b->runtime) in distribute_cfs_runtime()
6186 runtime = cfs_b->runtime; in distribute_cfs_runtime()
6187 cfs_b->runtime -= runtime; in distribute_cfs_runtime()
6188 remaining = cfs_b->runtime; in distribute_cfs_runtime()
6189 raw_spin_unlock(&cfs_b->lock); in distribute_cfs_runtime()
6191 cfs_rq->runtime_remaining += runtime; in distribute_cfs_runtime()
6194 if (cfs_rq->runtime_remaining > 0) { in distribute_cfs_runtime()
6203 list_add_tail(&cfs_rq->throttled_csd_list, in distribute_cfs_runtime()
6220 list_del_init(&cfs_rq->throttled_csd_list); in distribute_cfs_runtime()
6237 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
6245 if (cfs_b->quota == RUNTIME_INF) in do_sched_cfs_period_timer()
6248 throttled = !list_empty(&cfs_b->throttled_cfs_rq); in do_sched_cfs_period_timer()
6249 cfs_b->nr_periods += overrun; in do_sched_cfs_period_timer()
6251 /* Refill extra burst quota even if cfs_b->idle */ in do_sched_cfs_period_timer()
6258 if (cfs_b->idle && !throttled) in do_sched_cfs_period_timer()
6263 cfs_b->idle = 1; in do_sched_cfs_period_timer()
6268 cfs_b->nr_throttled += overrun; in do_sched_cfs_period_timer()
6271 * This check is repeated as we release cfs_b->lock while we unthrottle. in do_sched_cfs_period_timer()
6273 while (throttled && cfs_b->runtime > 0) { in do_sched_cfs_period_timer()
6274 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); in do_sched_cfs_period_timer()
6275 /* we can't nest cfs_b->lock while distributing bandwidth */ in do_sched_cfs_period_timer()
6277 raw_spin_lock_irqsave(&cfs_b->lock, flags); in do_sched_cfs_period_timer()
6286 cfs_b->idle = 0; in do_sched_cfs_period_timer()
6304 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
6310 struct hrtimer *refresh_timer = &cfs_b->period_timer; in runtime_refresh_within()
6313 /* if the call-back is running a quota refresh is already occurring */ in runtime_refresh_within()
6334 if (cfs_b->slack_started) in start_cfs_slack_bandwidth()
6336 cfs_b->slack_started = true; in start_cfs_slack_bandwidth()
6338 hrtimer_start(&cfs_b->slack_timer, in start_cfs_slack_bandwidth()
6346 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); in __return_cfs_rq_runtime()
6347 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; in __return_cfs_rq_runtime()
6352 raw_spin_lock(&cfs_b->lock); in __return_cfs_rq_runtime()
6353 if (cfs_b->quota != RUNTIME_INF) { in __return_cfs_rq_runtime()
6354 cfs_b->runtime += slack_runtime; in __return_cfs_rq_runtime()
6356 /* we are under rq->lock, defer unthrottling using a timer */ in __return_cfs_rq_runtime()
6357 if (cfs_b->runtime > sched_cfs_bandwidth_slice() && in __return_cfs_rq_runtime()
6358 !list_empty(&cfs_b->throttled_cfs_rq)) in __return_cfs_rq_runtime()
6361 raw_spin_unlock(&cfs_b->lock); in __return_cfs_rq_runtime()
6364 cfs_rq->runtime_remaining -= slack_runtime; in __return_cfs_rq_runtime()
6372 if (!cfs_rq->runtime_enabled || cfs_rq->nr_queued) in return_cfs_rq_runtime()
6380 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6388 raw_spin_lock_irqsave(&cfs_b->lock, flags); in do_sched_cfs_slack_timer()
6389 cfs_b->slack_started = false; in do_sched_cfs_slack_timer()
6392 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); in do_sched_cfs_slack_timer()
6396 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) in do_sched_cfs_slack_timer()
6397 runtime = cfs_b->runtime; in do_sched_cfs_slack_timer()
6399 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); in do_sched_cfs_slack_timer()
6410 * runtime as update_curr() throttling can not trigger until it's on-rq.
6417 /* an active group must be handled by the update_curr()->put() path */ in check_enqueue_throttle()
6418 if (!cfs_rq->runtime_enabled || cfs_rq->curr) in check_enqueue_throttle()
6427 if (cfs_rq->runtime_remaining <= 0) in check_enqueue_throttle()
6438 if (!tg->parent) in sync_throttle()
6441 cfs_rq = tg->cfs_rq[cpu]; in sync_throttle()
6442 pcfs_rq = tg->parent->cfs_rq[cpu]; in sync_throttle()
6444 cfs_rq->throttle_count = pcfs_rq->throttle_count; in sync_throttle()
6445 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu)); in sync_throttle()
6454 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) in check_cfs_rq_runtime()
6488 raw_spin_lock_irqsave(&cfs_b->lock, flags); in sched_cfs_period_timer()
6490 overrun = hrtimer_forward_now(timer, cfs_b->period); in sched_cfs_period_timer()
6497 u64 new, old = ktime_to_ns(cfs_b->period); in sched_cfs_period_timer()
6506 cfs_b->period = ns_to_ktime(new); in sched_cfs_period_timer()
6507 cfs_b->quota *= 2; in sched_cfs_period_timer()
6508 cfs_b->burst *= 2; in sched_cfs_period_timer()
6514 div_u64(cfs_b->quota, NSEC_PER_USEC)); in sched_cfs_period_timer()
6520 div_u64(cfs_b->quota, NSEC_PER_USEC)); in sched_cfs_period_timer()
6528 cfs_b->period_active = 0; in sched_cfs_period_timer()
6529 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); in sched_cfs_period_timer()
6536 raw_spin_lock_init(&cfs_b->lock); in init_cfs_bandwidth()
6537 cfs_b->runtime = 0; in init_cfs_bandwidth()
6538 cfs_b->quota = RUNTIME_INF; in init_cfs_bandwidth()
6539 cfs_b->period = ns_to_ktime(default_cfs_period()); in init_cfs_bandwidth()
6540 cfs_b->burst = 0; in init_cfs_bandwidth()
6541 cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF; in init_cfs_bandwidth()
6543 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); in init_cfs_bandwidth()
6544 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); in init_cfs_bandwidth()
6545 cfs_b->period_timer.function = sched_cfs_period_timer; in init_cfs_bandwidth()
6548 hrtimer_set_expires(&cfs_b->period_timer, in init_cfs_bandwidth()
6549 get_random_u32_below(cfs_b->period)); in init_cfs_bandwidth()
6550 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); in init_cfs_bandwidth()
6551 cfs_b->slack_timer.function = sched_cfs_slack_timer; in init_cfs_bandwidth()
6552 cfs_b->slack_started = false; in init_cfs_bandwidth()
6557 cfs_rq->runtime_enabled = 0; in init_cfs_rq_runtime()
6558 INIT_LIST_HEAD(&cfs_rq->throttled_list); in init_cfs_rq_runtime()
6559 INIT_LIST_HEAD(&cfs_rq->throttled_csd_list); in init_cfs_rq_runtime()
6564 lockdep_assert_held(&cfs_b->lock); in start_cfs_bandwidth()
6566 if (cfs_b->period_active) in start_cfs_bandwidth()
6569 cfs_b->period_active = 1; in start_cfs_bandwidth()
6570 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); in start_cfs_bandwidth()
6571 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); in start_cfs_bandwidth()
6579 if (!cfs_b->throttled_cfs_rq.next) in destroy_cfs_bandwidth()
6582 hrtimer_cancel(&cfs_b->period_timer); in destroy_cfs_bandwidth()
6583 hrtimer_cancel(&cfs_b->slack_timer); in destroy_cfs_bandwidth()
6600 if (list_empty(&rq->cfsb_csd_list)) in destroy_cfs_bandwidth()
6626 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; in update_runtime_enabled()
6627 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; in update_runtime_enabled()
6629 raw_spin_lock(&cfs_b->lock); in update_runtime_enabled()
6630 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; in update_runtime_enabled()
6631 raw_spin_unlock(&cfs_b->lock); in update_runtime_enabled()
6656 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; in unthrottle_offline_cfs_rqs()
6658 if (!cfs_rq->runtime_enabled) in unthrottle_offline_cfs_rqs()
6665 cfs_rq->runtime_enabled = 0; in unthrottle_offline_cfs_rqs()
6674 cfs_rq->runtime_remaining = 1; in unthrottle_offline_cfs_rqs()
6689 if (cfs_rq->runtime_enabled || in cfs_task_bw_constrained()
6690 tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF) in cfs_task_bw_constrained()
6708 if (rq->nr_running != 1) in sched_fair_update_stop_tick()
6777 struct sched_entity *se = &p->se; in hrtick_start_fair()
6781 if (rq->cfs.h_nr_queued > 1) { in hrtick_start_fair()
6782 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; in hrtick_start_fair()
6783 u64 slice = se->slice; in hrtick_start_fair()
6784 s64 delta = slice - ran; in hrtick_start_fair()
6802 struct task_struct *donor = rq->donor; in hrtick_update()
6804 if (!hrtick_enabled_fair(rq) || donor->sched_class != &fair_sched_class) in hrtick_update()
6840 return !sched_energy_enabled() || READ_ONCE(rd->overutilized); in is_rd_overutilized()
6848 WRITE_ONCE(rd->overutilized, flag); in set_rd_overutilized()
6859 if (!is_rd_overutilized(rq->rd) && cpu_overutilized(rq->cpu)) in check_update_overutilized_status()
6860 set_rd_overutilized(rq->rd, 1); in check_update_overutilized_status()
6869 return unlikely(rq->nr_running == rq->cfs.h_nr_idle && in sched_idle_rq()
6870 rq->nr_running); in sched_idle_rq()
6886 * se->sched_delayed should imply: se->on_rq == 1. in requeue_delayed_entity()
6890 SCHED_WARN_ON(!se->sched_delayed); in requeue_delayed_entity()
6891 SCHED_WARN_ON(!se->on_rq); in requeue_delayed_entity()
6895 if (se->vlag > 0) { in requeue_delayed_entity()
6896 cfs_rq->nr_queued--; in requeue_delayed_entity()
6897 if (se != cfs_rq->curr) in requeue_delayed_entity()
6899 se->vlag = 0; in requeue_delayed_entity()
6901 if (se != cfs_rq->curr) in requeue_delayed_entity()
6903 cfs_rq->nr_queued++; in requeue_delayed_entity()
6920 struct sched_entity *se = &p->se; in enqueue_task_fair()
6924 int rq_h_nr_queued = rq->cfs.h_nr_queued; in enqueue_task_fair()
6933 if (!(p->se.sched_delayed && (task_on_rq_migrating(p) || (flags & ENQUEUE_RESTORE)))) in enqueue_task_fair()
6934 util_est_enqueue(&rq->cfs, p); in enqueue_task_fair()
6946 if (p->in_iowait) in enqueue_task_fair()
6949 if (task_new && se->sched_delayed) in enqueue_task_fair()
6953 if (se->on_rq) { in enqueue_task_fair()
6954 if (se->sched_delayed) in enqueue_task_fair()
6963 * its entities in the desired time-frame. in enqueue_task_fair()
6966 se->slice = slice; in enqueue_task_fair()
6967 se->custom_slice = 1; in enqueue_task_fair()
6972 cfs_rq->h_nr_runnable += h_nr_runnable; in enqueue_task_fair()
6973 cfs_rq->h_nr_queued++; in enqueue_task_fair()
6974 cfs_rq->h_nr_idle += h_nr_idle; in enqueue_task_fair()
6993 se->slice = slice; in enqueue_task_fair()
6996 cfs_rq->h_nr_runnable += h_nr_runnable; in enqueue_task_fair()
6997 cfs_rq->h_nr_queued++; in enqueue_task_fair()
6998 cfs_rq->h_nr_idle += h_nr_idle; in enqueue_task_fair()
7008 if (!rq_h_nr_queued && rq->cfs.h_nr_queued) { in enqueue_task_fair()
7010 if (!rq->nr_running) in enqueue_task_fair()
7011 dl_server_update_idle_time(rq, rq->curr); in enqueue_task_fair()
7012 dl_server_start(&rq->fair_server); in enqueue_task_fair()
7045 * failing half-way through and resume the dequeue later.
7048 * -1 - dequeue delayed
7049 * 0 - dequeue throttled
7050 * 1 - dequeue complete
7055 int rq_h_nr_queued = rq->cfs.h_nr_queued; in dequeue_entities()
7069 if (task_sleep || task_delayed || !se->sched_delayed) in dequeue_entities()
7080 if (p && &p->se == se) in dequeue_entities()
7081 return -1; in dequeue_entities()
7086 cfs_rq->h_nr_runnable -= h_nr_runnable; in dequeue_entities()
7087 cfs_rq->h_nr_queued -= h_nr_queued; in dequeue_entities()
7088 cfs_rq->h_nr_idle -= h_nr_idle; in dequeue_entities()
7098 if (cfs_rq->load.weight) { in dequeue_entities()
7101 /* Avoid re-evaluating load for this entity: */ in dequeue_entities()
7122 se->slice = slice; in dequeue_entities()
7125 cfs_rq->h_nr_runnable -= h_nr_runnable; in dequeue_entities()
7126 cfs_rq->h_nr_queued -= h_nr_queued; in dequeue_entities()
7127 cfs_rq->h_nr_idle -= h_nr_idle; in dequeue_entities()
7139 if (rq_h_nr_queued && !rq->cfs.h_nr_queued) in dequeue_entities()
7140 dl_server_stop(&rq->fair_server); in dequeue_entities()
7144 rq->next_balance = jiffies; in dequeue_entities()
7148 SCHED_WARN_ON(p->on_rq != 1); in dequeue_entities()
7150 /* Fix-up what dequeue_task_fair() skipped */ in dequeue_entities()
7154 * Fix-up what block_task() skipped. in dequeue_entities()
7171 if (!(p->se.sched_delayed && (task_on_rq_migrating(p) || (flags & DEQUEUE_SAVE)))) in dequeue_task_fair()
7172 util_est_dequeue(&rq->cfs, p); in dequeue_task_fair()
7174 util_est_update(&rq->cfs, p, flags & DEQUEUE_SLEEP); in dequeue_task_fair()
7175 if (dequeue_entities(rq, &p->se, flags) < 0) in dequeue_task_fair()
7208 return cfs_rq_load_avg(&rq->cfs); in cpu_load()
7212 * cpu_load_without - compute CPU load without any contributions from *p
7230 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) in cpu_load_without()
7233 cfs_rq = &rq->cfs; in cpu_load_without()
7234 load = READ_ONCE(cfs_rq->avg.load_avg); in cpu_load_without()
7244 return cfs_rq_runnable_avg(&rq->cfs); in cpu_runnable()
7253 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) in cpu_runnable_without()
7256 cfs_rq = &rq->cfs; in cpu_runnable_without()
7257 runnable = READ_ONCE(cfs_rq->avg.runnable_avg); in cpu_runnable_without()
7260 lsub_positive(&runnable, p->se.avg.runnable_avg); in cpu_runnable_without()
7267 return cpu_rq(cpu)->cpu_capacity; in capacity_of()
7276 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { in record_wakee()
7277 current->wakee_flips >>= 1; in record_wakee()
7278 current->wakee_flip_decay_ts = jiffies; in record_wakee()
7281 if (current->last_wakee != p) { in record_wakee()
7282 current->last_wakee = p; in record_wakee()
7283 current->wakee_flips++; in record_wakee()
7288 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
7298 * non-monogamous, with partner count exceeding socket size.
7306 unsigned int master = current->wakee_flips; in wake_wide()
7307 unsigned int slave = p->wakee_flips; in wake_wide()
7322 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
7323 * cache-affine and is (or will be) idle.
7325 * wake_affine_weight() - considers the weight to reflect the average
7347 if (sync && cpu_rq(this_cpu)->nr_running == 1) in wake_affine_idle()
7371 this_eff_load -= current_load; in wake_affine_weight()
7382 prev_eff_load -= task_load; in wake_affine_weight()
7384 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; in wake_affine_weight()
7410 schedstat_inc(p->stats.nr_wakeups_affine_attempts); in wake_affine()
7414 schedstat_inc(sd->ttwu_move_affine); in wake_affine()
7415 schedstat_inc(p->stats.nr_wakeups_affine); in wake_affine()
7423 * sched_balance_find_dst_group_cpu - find the idlest CPU among the CPUs in the group.
7432 int shallowest_idle_cpu = -1; in sched_balance_find_dst_group_cpu()
7436 if (group->group_weight == 1) in sched_balance_find_dst_group_cpu()
7440 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { in sched_balance_find_dst_group_cpu()
7451 if (idle && idle->exit_latency < min_exit_latency) { in sched_balance_find_dst_group_cpu()
7457 min_exit_latency = idle->exit_latency; in sched_balance_find_dst_group_cpu()
7458 latest_idle_timestamp = rq->idle_stamp; in sched_balance_find_dst_group_cpu()
7460 } else if ((!idle || idle->exit_latency == min_exit_latency) && in sched_balance_find_dst_group_cpu()
7461 rq->idle_stamp > latest_idle_timestamp) { in sched_balance_find_dst_group_cpu()
7467 latest_idle_timestamp = rq->idle_stamp; in sched_balance_find_dst_group_cpu()
7470 } else if (shallowest_idle_cpu == -1) { in sched_balance_find_dst_group_cpu()
7479 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; in sched_balance_find_dst_group_cpu()
7487 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) in sched_balance_find_dst_cpu()
7495 sync_entity_load_avg(&p->se); in sched_balance_find_dst_cpu()
7502 if (!(sd->flags & sd_flag)) { in sched_balance_find_dst_cpu()
7503 sd = sd->child; in sched_balance_find_dst_cpu()
7509 sd = sd->child; in sched_balance_find_dst_cpu()
7516 sd = sd->child; in sched_balance_find_dst_cpu()
7522 weight = sd->span_weight; in sched_balance_find_dst_cpu()
7525 if (weight <= tmp->span_weight) in sched_balance_find_dst_cpu()
7527 if (tmp->flags & sd_flag) in sched_balance_find_dst_cpu()
7541 return -1; in __select_idle_cpu()
7554 WRITE_ONCE(sds->has_idle_cores, val); in set_idle_cores()
7563 return READ_ONCE(sds->has_idle_cores); in test_idle_cores()
7570 * information in sd_llc_shared->has_idle_cores.
7600 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7610 if (*idle_cpu == -1) { in select_idle_core()
7619 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, cpus)) in select_idle_core()
7627 return -1; in select_idle_core()
7637 for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) { in select_idle_smt()
7650 return -1; in select_idle_smt()
7671 return -1; in select_idle_smt()
7678 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7679 * average idle time for this rq (as found in rq->avg_idle).
7684 int i, cpu, idle_cpu = -1, nr = INT_MAX; in select_idle_cpu()
7687 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); in select_idle_cpu()
7692 /* because !--nr is the condition to stop scan */ in select_idle_cpu()
7693 nr = READ_ONCE(sd_share->nr_idle_scan) + 1; in select_idle_cpu()
7696 return -1; in select_idle_cpu()
7701 struct sched_group *sg = sd->groups; in select_idle_cpu()
7703 if (sg->flags & SD_CLUSTER) { in select_idle_cpu()
7713 if (--nr <= 0) in select_idle_cpu()
7714 return -1; in select_idle_cpu()
7731 if (--nr <= 0) in select_idle_cpu()
7732 return -1; in select_idle_cpu()
7755 int cpu, best_cpu = -1; in select_idle_capacity()
7759 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); in select_idle_capacity()
7784 * First, select CPU which fits better (-1 being better than 0). in select_idle_capacity()
7821 int i, recent_used_cpu, prev_aff = -1; in select_idle_sibling()
7828 sync_entity_load_avg(&p->se); in select_idle_sibling()
7835 * per-cpu select_rq_mask usage in select_idle_sibling()
7858 * Allow a per-cpu kthread to stack with the wakee if the in select_idle_sibling()
7861 * per-cpu kthread that is now complete and the wakeup is in select_idle_sibling()
7868 this_rq()->nr_running <= 1 && in select_idle_sibling()
7874 recent_used_cpu = p->recent_used_cpu; in select_idle_sibling()
7875 p->recent_used_cpu = prev; in select_idle_sibling()
7880 cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) && in select_idle_sibling()
7888 recent_used_cpu = -1; in select_idle_sibling()
7944 * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7947 * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7954 * recent utilization of currently non-runnable tasks on that CPU.
7962 * previously-executed tasks, which helps better deduce how busy a CPU will
7963 * be when a long-sleeping task wakes up. The contribution to CPU utilization
7977 * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7980 * after task migrations (scheduler-driven DVFS).
7987 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; in cpu_util()
7988 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg); in cpu_util()
7992 runnable = READ_ONCE(cfs_rq->avg.runnable_avg); in cpu_util()
7997 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its in cpu_util()
8010 util_est = READ_ONCE(cfs_rq->avg.util_est); in cpu_util()
8013 * During wake-up @p isn't enqueued yet and doesn't contribute in cpu_util()
8014 * to any cpu_rq(cpu)->cfs.avg.util_est. in cpu_util()
8018 * During exec (@dst_cpu = -1) @p is enqueued and does in cpu_util()
8019 * contribute to cpu_rq(cpu)->cfs.util_est. in cpu_util()
8028 * p->on_rq = TASK_ON_RQ_MIGRATING; in cpu_util()
8029 * -------------------------------- A in cpu_util()
8033 * -------------------------------- B in cpu_util()
8051 return cpu_util(cpu, NULL, -1, 0); in cpu_util_cfs()
8056 return cpu_util(cpu, NULL, -1, 1); in cpu_util_cfs_boost()
8075 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) in cpu_util_without()
8078 return cpu_util(cpu, p, -1, 0); in cpu_util_without()
8093 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
8094 * which excludes things like IRQ and steal-time. These latter are then accrued
8112 * because of inaccuracies in how we track these -- see in effective_cpu_util()
8127 * - the computed DL bandwidth needed with the IRQ pressure which in effective_cpu_util()
8129 * - The minimum performance requirement for CFS and/or RT. in effective_cpu_util()
8137 if (!uclamp_is_used() && rt_rq_is_runnable(&rq->rt)) in effective_cpu_util()
8165 * max - irq in effective_cpu_util()
8166 * U' = irq + --------- * U in effective_cpu_util()
8181 * energy_env - Utilization landscape for energy estimation.
8187 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
8213 eenv->task_busy_time = busy_time; in eenv_task_busy_time()
8227 * - A stable PD utilization, no matter which CPU of that PD we want to place
8230 * - A fair comparison between CPUs as the task contribution (task_util())
8235 * exceed @eenv->pd_cap.
8245 unsigned long util = cpu_util(cpu, p, -1, 0); in eenv_pd_busy_time()
8250 eenv->pd_busy_time = min(eenv->pd_cap, busy_time); in eenv_pd_busy_time()
8257 * Returns the maximum utilization among @eenv->cpus. This utilization can't
8258 * exceed @eenv->cpu_cap.
8299 return min(max_util, eenv->cpu_cap); in eenv_pd_max_util()
8312 unsigned long busy_time = eenv->pd_busy_time; in compute_energy()
8316 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time); in compute_energy()
8318 energy = em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap); in compute_energy()
8326 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
8330 * out which of the CPU candidates is the most energy-efficient.
8347 * cluster-packing, and spreading inside a cluster. That should at least be
8354 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
8358 * to be energy-inefficient in some use-cases. The alternative would be to
8361 * other use-cases too. So, until someone finds a better way to solve this,
8362 * let's keep things simple by re-using the existing slow path.
8370 struct root_domain *rd = this_rq()->rd; in find_energy_efficient_cpu()
8371 int cpu, best_energy_cpu, target = -1; in find_energy_efficient_cpu()
8372 int prev_fits = -1, best_fits = -1; in find_energy_efficient_cpu()
8380 pd = rcu_dereference(rd->pd); in find_energy_efficient_cpu()
8385 * Energy-aware wake-up happens on the lowest sched_domain starting in find_energy_efficient_cpu()
8390 sd = sd->parent; in find_energy_efficient_cpu()
8396 sync_entity_load_avg(&p->se); in find_energy_efficient_cpu()
8402 for (; pd; pd = pd->next) { in find_energy_efficient_cpu()
8405 long prev_spare_cap = -1, max_spare_cap = -1; in find_energy_efficient_cpu()
8408 int max_spare_cap_cpu = -1; in find_energy_efficient_cpu()
8409 int fits, max_fits = -1; in find_energy_efficient_cpu()
8431 if (!cpumask_test_cpu(cpu, p->cpus_ptr)) in find_energy_efficient_cpu()
8450 * max-aggregated uclamp_{min, max}. in find_energy_efficient_cpu()
8487 base_energy = compute_energy(&eenv, pd, cpus, p, -1); in find_energy_efficient_cpu()
8490 if (prev_spare_cap > -1) { in find_energy_efficient_cpu()
8496 prev_delta -= base_energy; in find_energy_efficient_cpu()
8520 cur_delta -= base_energy; in find_energy_efficient_cpu()
8564 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); in select_task_rq_fair()
8573 * required for stable ->cpus_allowed in select_task_rq_fair()
8575 lockdep_assert_held(&p->pi_lock); in select_task_rq_fair()
8580 cpumask_test_cpu(cpu, p->cpus_ptr)) in select_task_rq_fair()
8583 if (!is_rd_overutilized(this_rq()->rd)) { in select_task_rq_fair()
8590 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr); in select_task_rq_fair()
8599 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && in select_task_rq_fair()
8613 if (tmp->flags & sd_flag) in select_task_rq_fair()
8634 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
8638 struct sched_entity *se = &p->se; in migrate_task_rq_fair()
8647 * leading to an inflation after wake-up on the new rq. in migrate_task_rq_fair()
8657 se->avg.last_update_time = 0; in migrate_task_rq_fair()
8664 struct sched_entity *se = &p->se; in task_dead_fair()
8666 if (se->sched_delayed) { in task_dead_fair()
8671 if (se->sched_delayed) { in task_dead_fair()
8696 if (!cpumask_intersects(p->cpus_ptr, cpumask)) in set_task_max_allowed_capacity()
8699 p->max_allowed_capacity = entry->capacity; in set_task_max_allowed_capacity()
8726 if (SCHED_WARN_ON(!se->on_rq)) in set_next_buddy()
8730 cfs_rq_of(se)->next = se; in set_next_buddy()
8739 struct task_struct *donor = rq->donor; in check_preempt_wakeup_fair()
8740 struct sched_entity *se = &donor->se, *pse = &p->se; in check_preempt_wakeup_fair()
8751 * next-buddy nomination below. in check_preempt_wakeup_fair()
8756 if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK) && !pse->sched_delayed) { in check_preempt_wakeup_fair()
8764 * Note: this also catches the edge-case of curr being in a throttled in check_preempt_wakeup_fair()
8770 if (test_tsk_need_resched(rq->curr)) in check_preempt_wakeup_fair()
8783 * Preempt an idle entity in favor of a non-idle entity (and don't preempt in check_preempt_wakeup_fair()
8794 if (unlikely(!normal_policy(p->policy))) in check_preempt_wakeup_fair()
8806 if (do_preempt_short(cfs_rq, pse, se) && se->vlag == se->deadline) in check_preempt_wakeup_fair()
8807 se->vlag = se->deadline + 1; in check_preempt_wakeup_fair()
8827 cfs_rq = &rq->cfs; in pick_task_fair()
8828 if (!cfs_rq->nr_queued) in pick_task_fair()
8833 if (cfs_rq->curr && cfs_rq->curr->on_rq) in pick_task_fair()
8862 se = &p->se; in pick_next_task_fair()
8865 if (prev->sched_class != &fair_sched_class) in pick_next_task_fair()
8882 struct sched_entity *pse = &prev->se; in pick_next_task_fair()
8886 int se_depth = se->depth; in pick_next_task_fair()
8887 int pse_depth = pse->depth; in pick_next_task_fair()
8919 * Because sched_balance_newidle() releases (and re-acquires) rq->lock, it is in pick_next_task_fair()
8921 * must re-start the pick_next_entity() loop. in pick_next_task_fair()
8945 return !!dl_se->rq->cfs.nr_queued; in fair_server_has_tasks()
8950 return pick_task_fair(dl_se->rq); in fair_server_pick_task()
8955 struct sched_dl_entity *dl_se = &rq->fair_server; in fair_server_init()
8967 struct sched_entity *se = &prev->se; in put_prev_task_fair()
8981 struct task_struct *curr = rq->curr; in yield_task_fair()
8983 struct sched_entity *se = &curr->se; in yield_task_fair()
8988 if (unlikely(rq->nr_running == 1)) in yield_task_fair()
8995 * Update run-time statistics of the 'current'. in yield_task_fair()
9005 se->deadline += calc_delta_fair(se->slice, se); in yield_task_fair()
9010 struct sched_entity *se = &p->se; in yield_to_task_fair()
9013 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) in yield_to_task_fair()
9026 * Fair scheduling class load-balancing methods.
9030 * The purpose of load-balancing is to achieve the same basic fairness the
9031 * per-CPU scheduler provides, namely provide a proportional amount of compute
9036 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
9041 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
9047 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
9056 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
9063 * - infeasible weights;
9064 * - local vs global optima in the discrete case. ]
9074 * of load-balance at each level inversely proportional to the number of CPUs in
9080 * \Sum { --- * --- * 2^i } = O(n) (5)
9082 * `- size of each group
9083 * | | `- number of CPUs doing load-balance
9084 * | `- freq
9085 * `- sum over all levels
9127 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
9134 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
9216 /* The set of CPUs under consideration for load-balancing */
9231 * Is this task likely cache-hot:
9237 lockdep_assert_rq_held(env->src_rq); in task_hot()
9239 if (p->sched_class != &fair_sched_class) in task_hot()
9246 if (env->sd->flags & SD_SHARE_CPUCAPACITY) in task_hot()
9252 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && in task_hot()
9253 (&p->se == cfs_rq_of(&p->se)->next)) in task_hot()
9256 if (sysctl_sched_migration_cost == -1) in task_hot()
9263 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p)) in task_hot()
9269 delta = rq_clock_task(env->src_rq) - p->se.exec_start; in task_hot()
9282 struct numa_group *numa_group = rcu_dereference(p->numa_group); in migrate_degrades_locality()
9289 if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) in migrate_degrades_locality()
9292 src_nid = cpu_to_node(env->src_cpu); in migrate_degrades_locality()
9293 dst_nid = cpu_to_node(env->dst_cpu); in migrate_degrades_locality()
9299 if (src_nid == p->numa_preferred_nid) { in migrate_degrades_locality()
9300 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) in migrate_degrades_locality()
9307 if (dst_nid == p->numa_preferred_nid) in migrate_degrades_locality()
9308 return -1; in migrate_degrades_locality()
9311 if (env->idle == CPU_IDLE) in migrate_degrades_locality()
9323 return src_weight - dst_weight; in migrate_degrades_locality()
9338 * dst_cfs_rq->nr_queued is greater than 1, if the task
9347 dst_cfs_rq = task_group(p)->cfs_rq[dest_cpu]; in task_is_ineligible_on_dst_cpu()
9349 dst_cfs_rq = &cpu_rq(dest_cpu)->cfs; in task_is_ineligible_on_dst_cpu()
9351 if (sched_feat(PLACE_LAG) && dst_cfs_rq->nr_queued && in task_is_ineligible_on_dst_cpu()
9352 !entity_eligible(task_cfs_rq(p), &p->se)) in task_is_ineligible_on_dst_cpu()
9359 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
9366 lockdep_assert_rq_held(env->src_rq); in can_migrate_task()
9367 if (p->sched_task_hot) in can_migrate_task()
9368 p->sched_task_hot = 0; in can_migrate_task()
9376 * 5) are cache-hot on their current CPU. in can_migrate_task()
9378 if ((p->se.sched_delayed) && (env->migration_type != migrate_load)) in can_migrate_task()
9381 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) in can_migrate_task()
9386 * For ineligible tasks we soft-limit them and only allow in can_migrate_task()
9387 * them to migrate when nr_balance_failed is non-zero to in can_migrate_task()
9388 * avoid load-balancing trying very hard to balance the load. in can_migrate_task()
9390 if (!env->sd->nr_balance_failed && in can_migrate_task()
9391 task_is_ineligible_on_dst_cpu(p, env->dst_cpu)) in can_migrate_task()
9398 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { in can_migrate_task()
9401 schedstat_inc(p->stats.nr_failed_migrations_affine); in can_migrate_task()
9403 env->flags |= LBF_SOME_PINNED; in can_migrate_task()
9411 * - for NEWLY_IDLE in can_migrate_task()
9412 * - if we have already computed one in current iteration in can_migrate_task()
9413 * - if it's an active balance in can_migrate_task()
9415 if (env->idle == CPU_NEWLY_IDLE || in can_migrate_task()
9416 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB)) in can_migrate_task()
9419 /* Prevent to re-select dst_cpu via env's CPUs: */ in can_migrate_task()
9420 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { in can_migrate_task()
9421 if (cpumask_test_cpu(cpu, p->cpus_ptr)) { in can_migrate_task()
9422 env->flags |= LBF_DST_PINNED; in can_migrate_task()
9423 env->new_dst_cpu = cpu; in can_migrate_task()
9432 env->flags &= ~LBF_ALL_PINNED; in can_migrate_task()
9434 if (task_on_cpu(env->src_rq, p)) { in can_migrate_task()
9435 schedstat_inc(p->stats.nr_failed_migrations_running); in can_migrate_task()
9446 if (env->flags & LBF_ACTIVE_LB) in can_migrate_task()
9455 if (!hot || env->sd->nr_balance_failed > env->sd->cache_nice_tries) { in can_migrate_task()
9457 p->sched_task_hot = 1; in can_migrate_task()
9461 schedstat_inc(p->stats.nr_failed_migrations_hot); in can_migrate_task()
9466 * detach_task() -- detach the task for the migration specified in env
9470 lockdep_assert_rq_held(env->src_rq); in detach_task()
9472 if (p->sched_task_hot) { in detach_task()
9473 p->sched_task_hot = 0; in detach_task()
9474 schedstat_inc(env->sd->lb_hot_gained[env->idle]); in detach_task()
9475 schedstat_inc(p->stats.nr_forced_migrations); in detach_task()
9478 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); in detach_task()
9479 set_task_cpu(p, env->dst_cpu); in detach_task()
9483 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
9492 lockdep_assert_rq_held(env->src_rq); in detach_one_task()
9495 &env->src_rq->cfs_tasks, se.group_node) { in detach_one_task()
9503 * lb_gained[env->idle] is updated (other is detach_tasks) in detach_one_task()
9507 schedstat_inc(env->sd->lb_gained[env->idle]); in detach_one_task()
9514 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
9521 struct list_head *tasks = &env->src_rq->cfs_tasks; in detach_tasks()
9526 lockdep_assert_rq_held(env->src_rq); in detach_tasks()
9532 if (env->src_rq->nr_running <= 1) { in detach_tasks()
9533 env->flags &= ~LBF_ALL_PINNED; in detach_tasks()
9537 if (env->imbalance <= 0) in detach_tasks()
9545 if (env->idle && env->src_rq->nr_running <= 1) in detach_tasks()
9548 env->loop++; in detach_tasks()
9550 if (env->loop > env->loop_max) in detach_tasks()
9554 if (env->loop > env->loop_break) { in detach_tasks()
9555 env->loop_break += SCHED_NR_MIGRATE_BREAK; in detach_tasks()
9556 env->flags |= LBF_NEED_BREAK; in detach_tasks()
9565 switch (env->migration_type) { in detach_tasks()
9570 * value. Make sure that env->imbalance decreases in detach_tasks()
9577 load < 16 && !env->sd->nr_balance_failed) in detach_tasks()
9586 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance) in detach_tasks()
9589 env->imbalance -= load; in detach_tasks()
9595 if (shr_bound(util, env->sd->nr_balance_failed) > env->imbalance) in detach_tasks()
9598 env->imbalance -= util; in detach_tasks()
9602 env->imbalance--; in detach_tasks()
9607 if (task_fits_cpu(p, env->src_cpu)) in detach_tasks()
9610 env->imbalance = 0; in detach_tasks()
9615 list_add(&p->se.group_node, &env->tasks); in detach_tasks()
9625 if (env->idle == CPU_NEWLY_IDLE) in detach_tasks()
9633 if (env->imbalance <= 0) in detach_tasks()
9638 if (p->sched_task_hot) in detach_tasks()
9639 schedstat_inc(p->stats.nr_failed_migrations_hot); in detach_tasks()
9641 list_move(&p->se.group_node, tasks); in detach_tasks()
9649 schedstat_add(env->sd->lb_gained[env->idle], detached); in detach_tasks()
9655 * attach_task() -- attach the task detached by detach_task() to its new rq.
9667 * attach_one_task() -- attaches the task returned from detach_one_task() to
9681 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
9686 struct list_head *tasks = &env->tasks; in attach_tasks()
9690 rq_lock(env->dst_rq, &rf); in attach_tasks()
9691 update_rq_clock(env->dst_rq); in attach_tasks()
9695 list_del_init(&p->se.group_node); in attach_tasks()
9697 attach_task(env->dst_rq, p); in attach_tasks()
9700 rq_unlock(env->dst_rq, &rf); in attach_tasks()
9706 if (cfs_rq->avg.load_avg) in cfs_rq_has_blocked()
9709 if (cfs_rq->avg.util_avg) in cfs_rq_has_blocked()
9734 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies); in update_blocked_load_tick()
9740 rq->has_blocked_load = 0; in update_blocked_load_status()
9783 if (cfs_rq->nr_queued == 0) in __update_blocked_fair()
9786 if (cfs_rq == &rq->cfs) in __update_blocked_fair()
9791 se = cfs_rq->tg->se[cpu]; in __update_blocked_fair()
9812 * This needs to be done in a top-down fashion because the load of a child
9818 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; in update_cfs_rq_h_load()
9822 if (cfs_rq->last_h_load_update == now) in update_cfs_rq_h_load()
9825 WRITE_ONCE(cfs_rq->h_load_next, NULL); in update_cfs_rq_h_load()
9828 WRITE_ONCE(cfs_rq->h_load_next, se); in update_cfs_rq_h_load()
9829 if (cfs_rq->last_h_load_update == now) in update_cfs_rq_h_load()
9834 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); in update_cfs_rq_h_load()
9835 cfs_rq->last_h_load_update = now; in update_cfs_rq_h_load()
9838 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { in update_cfs_rq_h_load()
9839 load = cfs_rq->h_load; in update_cfs_rq_h_load()
9840 load = div64_ul(load * se->avg.load_avg, in update_cfs_rq_h_load()
9843 cfs_rq->h_load = load; in update_cfs_rq_h_load()
9844 cfs_rq->last_h_load_update = now; in update_cfs_rq_h_load()
9853 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, in task_h_load()
9859 struct cfs_rq *cfs_rq = &rq->cfs; in __update_blocked_fair()
9871 return p->se.avg.load_avg; in task_h_load()
9897 * sg_lb_stats - stats of a sched_group required for load-balancing:
9920 * sd_lb_stats - stats of a sched_domain required for load-balancing:
9977 free = max - used; in scale_rt_capacity()
9985 struct sched_group *sdg = sd->groups; in update_cpu_capacity()
9990 cpu_rq(cpu)->cpu_capacity = capacity; in update_cpu_capacity()
9993 sdg->sgc->capacity = capacity; in update_cpu_capacity()
9994 sdg->sgc->min_capacity = capacity; in update_cpu_capacity()
9995 sdg->sgc->max_capacity = capacity; in update_cpu_capacity()
10000 struct sched_domain *child = sd->child; in update_group_capacity()
10001 struct sched_group *group, *sdg = sd->groups; in update_group_capacity()
10005 interval = msecs_to_jiffies(sd->balance_interval); in update_group_capacity()
10007 sdg->sgc->next_update = jiffies + interval; in update_group_capacity()
10018 if (child->flags & SD_OVERLAP) { in update_group_capacity()
10037 group = child->groups; in update_group_capacity()
10039 struct sched_group_capacity *sgc = group->sgc; in update_group_capacity()
10041 capacity += sgc->capacity; in update_group_capacity()
10042 min_capacity = min(sgc->min_capacity, min_capacity); in update_group_capacity()
10043 max_capacity = max(sgc->max_capacity, max_capacity); in update_group_capacity()
10044 group = group->next; in update_group_capacity()
10045 } while (group != child->groups); in update_group_capacity()
10048 sdg->sgc->capacity = capacity; in update_group_capacity()
10049 sdg->sgc->min_capacity = min_capacity; in update_group_capacity()
10050 sdg->sgc->max_capacity = max_capacity; in update_group_capacity()
10061 return ((rq->cpu_capacity * sd->imbalance_pct) < in check_cpu_capacity()
10068 return rq->misfit_task_load; in check_misfit_status()
10073 * groups is inadequate due to ->cpus_ptr constraints.
10082 * If we were to balance group-wise we'd place two tasks in the first group and
10102 return group->sgc->imbalance; in sg_imbalanced()
10120 if (sgs->sum_nr_running < sgs->group_weight) in group_has_capacity()
10123 if ((sgs->group_capacity * imbalance_pct) < in group_has_capacity()
10124 (sgs->group_runnable * 100)) in group_has_capacity()
10127 if ((sgs->group_capacity * 100) > in group_has_capacity()
10128 (sgs->group_util * imbalance_pct)) in group_has_capacity()
10145 if (sgs->sum_nr_running <= sgs->group_weight) in group_is_overloaded()
10148 if ((sgs->group_capacity * 100) < in group_is_overloaded()
10149 (sgs->group_util * imbalance_pct)) in group_is_overloaded()
10152 if ((sgs->group_capacity * imbalance_pct) < in group_is_overloaded()
10153 (sgs->group_runnable * 100)) in group_is_overloaded()
10170 if (sgs->group_asym_packing) in group_classify()
10173 if (sgs->group_smt_balance) in group_classify()
10176 if (sgs->group_misfit_task_load) in group_classify()
10186 * sched_use_asym_prio - Check whether asym_packing priority must be used
10198 if (!(sd->flags & SD_ASYM_PACKING)) in sched_use_asym_prio()
10204 return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu); in sched_use_asym_prio()
10218 * sched_group_asym - Check if the destination CPU can do asym_packing balance
10220 * @sgs: Load-balancing statistics of the candidate busiest group
10236 if ((group->flags & SD_SHARE_CPUCAPACITY) && in sched_group_asym()
10237 (sgs->group_weight - sgs->idle_cpus != 1)) in sched_group_asym()
10240 return sched_asym(env->sd, env->dst_cpu, group->asym_prefer_cpu); in sched_group_asym()
10250 return (sg1->flags & SD_SHARE_CPUCAPACITY) != in smt_vs_nonsmt_groups()
10251 (sg2->flags & SD_SHARE_CPUCAPACITY); in smt_vs_nonsmt_groups()
10257 if (!env->idle) in smt_balance()
10266 if (group->flags & SD_SHARE_CPUCAPACITY && in smt_balance()
10267 sgs->sum_h_nr_running > 1) in smt_balance()
10281 if (!env->idle || !busiest->sum_nr_running) in sibling_imbalance()
10284 ncores_busiest = sds->busiest->cores; in sibling_imbalance()
10285 ncores_local = sds->local->cores; in sibling_imbalance()
10288 imbalance = busiest->sum_nr_running; in sibling_imbalance()
10289 lsub_positive(&imbalance, local->sum_nr_running); in sibling_imbalance()
10294 imbalance = ncores_local * busiest->sum_nr_running; in sibling_imbalance()
10295 lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running); in sibling_imbalance()
10301 if (imbalance <= 1 && local->sum_nr_running == 0 && in sibling_imbalance()
10302 busiest->sum_nr_running > 1) in sibling_imbalance()
10315 if (rq->cfs.h_nr_runnable != 1) in sched_reduced_capacity()
10322 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
10324 * @sds: Load-balancing data with statistics of the local group.
10337 int i, nr_running, local_group, sd_flags = env->sd->flags; in update_sg_lb_stats()
10338 bool balancing_at_rd = !env->sd->parent; in update_sg_lb_stats()
10342 local_group = group == sds->local; in update_sg_lb_stats()
10344 for_each_cpu_and(i, sched_group_span(group), env->cpus) { in update_sg_lb_stats()
10348 sgs->group_load += load; in update_sg_lb_stats()
10349 sgs->group_util += cpu_util_cfs(i); in update_sg_lb_stats()
10350 sgs->group_runnable += cpu_runnable(rq); in update_sg_lb_stats()
10351 sgs->sum_h_nr_running += rq->cfs.h_nr_runnable; in update_sg_lb_stats()
10353 nr_running = rq->nr_running; in update_sg_lb_stats()
10354 sgs->sum_nr_running += nr_running; in update_sg_lb_stats()
10363 sgs->idle_cpus++; in update_sg_lb_stats()
10375 sgs->nr_numa_running += rq->nr_numa_running; in update_sg_lb_stats()
10376 sgs->nr_preferred_running += rq->nr_preferred_running; in update_sg_lb_stats()
10384 if (sgs->group_misfit_task_load < rq->misfit_task_load) { in update_sg_lb_stats()
10385 sgs->group_misfit_task_load = rq->misfit_task_load; in update_sg_lb_stats()
10388 } else if (env->idle && sched_reduced_capacity(rq, env->sd)) { in update_sg_lb_stats()
10390 if (sgs->group_misfit_task_load < load) in update_sg_lb_stats()
10391 sgs->group_misfit_task_load = load; in update_sg_lb_stats()
10395 sgs->group_capacity = group->sgc->capacity; in update_sg_lb_stats()
10397 sgs->group_weight = group->group_weight; in update_sg_lb_stats()
10400 if (!local_group && env->idle && sgs->sum_h_nr_running && in update_sg_lb_stats()
10402 sgs->group_asym_packing = 1; in update_sg_lb_stats()
10406 sgs->group_smt_balance = 1; in update_sg_lb_stats()
10408 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs); in update_sg_lb_stats()
10411 if (sgs->group_type == group_overloaded) in update_sg_lb_stats()
10412 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / in update_sg_lb_stats()
10413 sgs->group_capacity; in update_sg_lb_stats()
10417 * update_sd_pick_busiest - return 1 on busiest group
10434 struct sg_lb_stats *busiest = &sds->busiest_stat; in update_sd_pick_busiest()
10437 if (!sgs->sum_h_nr_running) in update_sd_pick_busiest()
10446 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && in update_sd_pick_busiest()
10447 (sgs->group_type == group_misfit_task) && in update_sd_pick_busiest()
10448 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) || in update_sd_pick_busiest()
10449 sds->local_stat.group_type != group_has_spare)) in update_sd_pick_busiest()
10452 if (sgs->group_type > busiest->group_type) in update_sd_pick_busiest()
10455 if (sgs->group_type < busiest->group_type) in update_sd_pick_busiest()
10463 switch (sgs->group_type) { in update_sd_pick_busiest()
10466 return sgs->avg_load > busiest->avg_load; in update_sd_pick_busiest()
10477 return sched_asym_prefer(sds->busiest->asym_prefer_cpu, sg->asym_prefer_cpu); in update_sd_pick_busiest()
10484 return sgs->group_misfit_task_load > busiest->group_misfit_task_load; in update_sd_pick_busiest()
10491 if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0) in update_sd_pick_busiest()
10509 if (sgs->avg_load < busiest->avg_load) in update_sd_pick_busiest()
10512 if (sgs->avg_load == busiest->avg_load) { in update_sd_pick_busiest()
10514 * SMT sched groups need more help than non-SMT groups. in update_sd_pick_busiest()
10517 if (sds->busiest->flags & SD_SHARE_CPUCAPACITY) in update_sd_pick_busiest()
10529 if (smt_vs_nonsmt_groups(sds->busiest, sg)) { in update_sd_pick_busiest()
10530 if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1) in update_sd_pick_busiest()
10544 if (sgs->idle_cpus > busiest->idle_cpus) in update_sd_pick_busiest()
10546 else if ((sgs->idle_cpus == busiest->idle_cpus) && in update_sd_pick_busiest()
10547 (sgs->sum_nr_running <= busiest->sum_nr_running)) in update_sd_pick_busiest()
10555 * per-CPU capacity. Migrating tasks to less capable CPUs may harm in update_sd_pick_busiest()
10559 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && in update_sd_pick_busiest()
10560 (sgs->group_type <= group_fully_busy) && in update_sd_pick_busiest()
10561 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu)))) in update_sd_pick_busiest()
10570 if (sgs->sum_h_nr_running > sgs->nr_numa_running) in fbq_classify_group()
10572 if (sgs->sum_h_nr_running > sgs->nr_preferred_running) in fbq_classify_group()
10579 if (rq->nr_running > rq->nr_numa_running) in fbq_classify_rq()
10581 if (rq->nr_running > rq->nr_preferred_running) in fbq_classify_rq()
10601 * task_running_on_cpu - return 1 if @p is running on @cpu.
10607 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) in task_running_on_cpu()
10617 * idle_cpu_without - would a given CPU be idle without p ?
10627 if (rq->curr != rq->idle && rq->curr != p) in idle_cpu_without()
10631 * rq->nr_running can't be used but an updated version without the in idle_cpu_without()
10636 if (rq->ttwu_pending) in idle_cpu_without()
10643 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
10659 if (sd->flags & SD_ASYM_CPUCAPACITY) in update_sg_wakeup_stats()
10660 sgs->group_misfit_task_load = 1; in update_sg_wakeup_stats()
10666 sgs->group_load += cpu_load_without(rq, p); in update_sg_wakeup_stats()
10667 sgs->group_util += cpu_util_without(i, p); in update_sg_wakeup_stats()
10668 sgs->group_runnable += cpu_runnable_without(rq, p); in update_sg_wakeup_stats()
10670 sgs->sum_h_nr_running += rq->cfs.h_nr_runnable - local; in update_sg_wakeup_stats()
10672 nr_running = rq->nr_running - local; in update_sg_wakeup_stats()
10673 sgs->sum_nr_running += nr_running; in update_sg_wakeup_stats()
10679 sgs->idle_cpus++; in update_sg_wakeup_stats()
10682 if (sd->flags & SD_ASYM_CPUCAPACITY && in update_sg_wakeup_stats()
10683 sgs->group_misfit_task_load && in update_sg_wakeup_stats()
10685 sgs->group_misfit_task_load = 0; in update_sg_wakeup_stats()
10689 sgs->group_capacity = group->sgc->capacity; in update_sg_wakeup_stats()
10691 sgs->group_weight = group->group_weight; in update_sg_wakeup_stats()
10693 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs); in update_sg_wakeup_stats()
10699 if (sgs->group_type == group_fully_busy || in update_sg_wakeup_stats()
10700 sgs->group_type == group_overloaded) in update_sg_wakeup_stats()
10701 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / in update_sg_wakeup_stats()
10702 sgs->group_capacity; in update_sg_wakeup_stats()
10710 if (sgs->group_type < idlest_sgs->group_type) in update_pick_idlest()
10713 if (sgs->group_type > idlest_sgs->group_type) in update_pick_idlest()
10721 switch (sgs->group_type) { in update_pick_idlest()
10725 if (idlest_sgs->avg_load <= sgs->avg_load) in update_pick_idlest()
10737 if (idlest->sgc->max_capacity >= group->sgc->max_capacity) in update_pick_idlest()
10743 if (idlest_sgs->idle_cpus > sgs->idle_cpus) in update_pick_idlest()
10747 if (idlest_sgs->idle_cpus == sgs->idle_cpus && in update_pick_idlest()
10748 idlest_sgs->group_util <= sgs->group_util) in update_pick_idlest()
10766 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups; in sched_balance_find_dst_group()
10780 p->cpus_ptr)) in sched_balance_find_dst_group()
10804 } while (group = group->next, group != sd->groups); in sched_balance_find_dst_group()
10835 (sd->imbalance_pct-100) / 100; in sched_balance_find_dst_group()
10842 * cross-domain, add imbalance to the load on the remote node in sched_balance_find_dst_group()
10846 if ((sd->flags & SD_NUMA) && in sched_balance_find_dst_group()
10857 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load) in sched_balance_find_dst_group()
10869 if (local->sgc->max_capacity >= idlest->sgc->max_capacity) in sched_balance_find_dst_group()
10875 if (sd->flags & SD_NUMA) { in sched_balance_find_dst_group()
10876 int imb_numa_nr = sd->imb_numa_nr; in sched_balance_find_dst_group()
10883 if (cpu_to_node(this_cpu) == p->numa_preferred_nid) in sched_balance_find_dst_group()
10887 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid) in sched_balance_find_dst_group()
10899 if (p->nr_cpus_allowed != NR_CPUS) { in sched_balance_find_dst_group()
10902 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr); in sched_balance_find_dst_group()
10903 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr); in sched_balance_find_dst_group()
10906 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus); in sched_balance_find_dst_group()
10943 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE) in update_idle_cpu_scan()
10946 llc_weight = per_cpu(sd_llc_size, env->dst_cpu); in update_idle_cpu_scan()
10947 if (env->sd->span_weight != llc_weight) in update_idle_cpu_scan()
10950 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu)); in update_idle_cpu_scan()
10960 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1] in update_idle_cpu_scan()
10978 * y = SCHED_CAPACITY_SCALE - in update_idle_cpu_scan()
10987 pct = env->sd->imbalance_pct; in update_idle_cpu_scan()
10991 y = SCHED_CAPACITY_SCALE - tmp; in update_idle_cpu_scan()
10996 if ((int)y != sd_share->nr_idle_scan) in update_idle_cpu_scan()
10997 WRITE_ONCE(sd_share->nr_idle_scan, (int)y); in update_idle_cpu_scan()
11001 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
11008 struct sched_group *sg = env->sd->groups; in update_sd_lb_stats()
11009 struct sg_lb_stats *local = &sds->local_stat; in update_sd_lb_stats()
11018 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); in update_sd_lb_stats()
11020 sds->local = sg; in update_sd_lb_stats()
11023 if (env->idle != CPU_NEWLY_IDLE || in update_sd_lb_stats()
11024 time_after_eq(jiffies, sg->sgc->next_update)) in update_sd_lb_stats()
11025 update_group_capacity(env->sd, env->dst_cpu); in update_sd_lb_stats()
11031 sds->busiest = sg; in update_sd_lb_stats()
11032 sds->busiest_stat = *sgs; in update_sd_lb_stats()
11036 sds->total_load += sgs->group_load; in update_sd_lb_stats()
11037 sds->total_capacity += sgs->group_capacity; in update_sd_lb_stats()
11039 sum_util += sgs->group_util; in update_sd_lb_stats()
11040 sg = sg->next; in update_sd_lb_stats()
11041 } while (sg != env->sd->groups); in update_sd_lb_stats()
11048 if (sds->busiest) in update_sd_lb_stats()
11049 sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING); in update_sd_lb_stats()
11052 if (env->sd->flags & SD_NUMA) in update_sd_lb_stats()
11053 env->fbq_type = fbq_classify_group(&sds->busiest_stat); in update_sd_lb_stats()
11055 if (!env->sd->parent) { in update_sd_lb_stats()
11057 set_rd_overloaded(env->dst_rq->rd, sg_overloaded); in update_sd_lb_stats()
11059 /* Update over-utilization (tipping point, U >= 0) indicator */ in update_sd_lb_stats()
11060 set_rd_overutilized(env->dst_rq->rd, sg_overutilized); in update_sd_lb_stats()
11062 set_rd_overutilized(env->dst_rq->rd, sg_overutilized); in update_sd_lb_stats()
11069 * calculate_imbalance - Calculate the amount of imbalance present within the
11078 local = &sds->local_stat; in calculate_imbalance()
11079 busiest = &sds->busiest_stat; in calculate_imbalance()
11081 if (busiest->group_type == group_misfit_task) { in calculate_imbalance()
11082 if (env->sd->flags & SD_ASYM_CPUCAPACITY) { in calculate_imbalance()
11084 env->migration_type = migrate_misfit; in calculate_imbalance()
11085 env->imbalance = 1; in calculate_imbalance()
11091 env->migration_type = migrate_load; in calculate_imbalance()
11092 env->imbalance = busiest->group_misfit_task_load; in calculate_imbalance()
11097 if (busiest->group_type == group_asym_packing) { in calculate_imbalance()
11102 env->migration_type = migrate_task; in calculate_imbalance()
11103 env->imbalance = busiest->sum_h_nr_running; in calculate_imbalance()
11107 if (busiest->group_type == group_smt_balance) { in calculate_imbalance()
11109 env->migration_type = migrate_task; in calculate_imbalance()
11110 env->imbalance = 1; in calculate_imbalance()
11114 if (busiest->group_type == group_imbalanced) { in calculate_imbalance()
11116 * In the group_imb case we cannot rely on group-wide averages in calculate_imbalance()
11117 * to ensure CPU-load equilibrium, try to move any task to fix in calculate_imbalance()
11121 env->migration_type = migrate_task; in calculate_imbalance()
11122 env->imbalance = 1; in calculate_imbalance()
11130 if (local->group_type == group_has_spare) { in calculate_imbalance()
11131 if ((busiest->group_type > group_fully_busy) && in calculate_imbalance()
11132 !(env->sd->flags & SD_SHARE_LLC)) { in calculate_imbalance()
11141 env->migration_type = migrate_util; in calculate_imbalance()
11142 env->imbalance = max(local->group_capacity, local->group_util) - in calculate_imbalance()
11143 local->group_util; in calculate_imbalance()
11152 if (env->idle && env->imbalance == 0) { in calculate_imbalance()
11153 env->migration_type = migrate_task; in calculate_imbalance()
11154 env->imbalance = 1; in calculate_imbalance()
11160 if (busiest->group_weight == 1 || sds->prefer_sibling) { in calculate_imbalance()
11165 env->migration_type = migrate_task; in calculate_imbalance()
11166 env->imbalance = sibling_imbalance(env, sds, busiest, local); in calculate_imbalance()
11173 env->migration_type = migrate_task; in calculate_imbalance()
11174 env->imbalance = max_t(long, 0, in calculate_imbalance()
11175 (local->idle_cpus - busiest->idle_cpus)); in calculate_imbalance()
11180 if (env->sd->flags & SD_NUMA) { in calculate_imbalance()
11181 env->imbalance = adjust_numa_imbalance(env->imbalance, in calculate_imbalance()
11182 local->sum_nr_running + 1, in calculate_imbalance()
11183 env->sd->imb_numa_nr); in calculate_imbalance()
11188 env->imbalance >>= 1; in calculate_imbalance()
11197 if (local->group_type < group_overloaded) { in calculate_imbalance()
11203 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) / in calculate_imbalance()
11204 local->group_capacity; in calculate_imbalance()
11210 if (local->avg_load >= busiest->avg_load) { in calculate_imbalance()
11211 env->imbalance = 0; in calculate_imbalance()
11215 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) / in calculate_imbalance()
11216 sds->total_capacity; in calculate_imbalance()
11222 if (local->avg_load >= sds->avg_load) { in calculate_imbalance()
11223 env->imbalance = 0; in calculate_imbalance()
11237 env->migration_type = migrate_load; in calculate_imbalance()
11238 env->imbalance = min( in calculate_imbalance()
11239 (busiest->avg_load - sds->avg_load) * busiest->group_capacity, in calculate_imbalance()
11240 (sds->avg_load - local->avg_load) * local->group_capacity in calculate_imbalance()
11267 * sched_balance_find_src_group - Returns the busiest group within the sched_domain
11274 * Return: - The busiest group if imbalance exists.
11296 if (busiest->group_type == group_misfit_task) in sched_balance_find_src_group()
11299 if (!is_rd_overutilized(env->dst_rq->rd) && in sched_balance_find_src_group()
11300 rcu_dereference(env->dst_rq->rd->pd)) in sched_balance_find_src_group()
11304 if (busiest->group_type == group_asym_packing) in sched_balance_find_src_group()
11312 if (busiest->group_type == group_imbalanced) in sched_balance_find_src_group()
11320 if (local->group_type > busiest->group_type) in sched_balance_find_src_group()
11327 if (local->group_type == group_overloaded) { in sched_balance_find_src_group()
11332 if (local->avg_load >= busiest->avg_load) in sched_balance_find_src_group()
11343 if (local->avg_load >= sds.avg_load) in sched_balance_find_src_group()
11350 if (100 * busiest->avg_load <= in sched_balance_find_src_group()
11351 env->sd->imbalance_pct * local->avg_load) in sched_balance_find_src_group()
11359 if (sds.prefer_sibling && local->group_type == group_has_spare && in sched_balance_find_src_group()
11363 if (busiest->group_type != group_overloaded) { in sched_balance_find_src_group()
11364 if (!env->idle) { in sched_balance_find_src_group()
11373 if (busiest->group_type == group_smt_balance && in sched_balance_find_src_group()
11379 if (busiest->group_weight > 1 && in sched_balance_find_src_group()
11380 local->idle_cpus <= (busiest->idle_cpus + 1)) { in sched_balance_find_src_group()
11393 if (busiest->sum_h_nr_running == 1) { in sched_balance_find_src_group()
11404 return env->imbalance ? sds.busiest : NULL; in sched_balance_find_src_group()
11407 env->imbalance = 0; in sched_balance_find_src_group()
11412 * sched_balance_find_src_rq - find the busiest runqueue among the CPUs in the group.
11422 for_each_cpu_and(i, sched_group_span(group), env->cpus) { in sched_balance_find_src_rq()
11432 * - regular: there are !numa tasks in sched_balance_find_src_rq()
11433 * - remote: there are numa tasks that run on the 'wrong' node in sched_balance_find_src_rq()
11434 * - all: there is no distinction in sched_balance_find_src_rq()
11449 if (rt > env->fbq_type) in sched_balance_find_src_rq()
11452 nr_running = rq->cfs.h_nr_runnable; in sched_balance_find_src_rq()
11460 * eventually lead to active_balancing high->low capacity. in sched_balance_find_src_rq()
11461 * Higher per-CPU capacity is considered better than balancing in sched_balance_find_src_rq()
11464 if (env->sd->flags & SD_ASYM_CPUCAPACITY && in sched_balance_find_src_rq()
11465 !capacity_greater(capacity_of(env->dst_cpu), capacity) && in sched_balance_find_src_rq()
11476 if (sched_asym(env->sd, i, env->dst_cpu) && nr_running == 1) in sched_balance_find_src_rq()
11479 switch (env->migration_type) { in sched_balance_find_src_rq()
11487 if (nr_running == 1 && load > env->imbalance && in sched_balance_find_src_rq()
11488 !check_cpu_capacity(rq, env->sd)) in sched_balance_find_src_rq()
11540 if (rq->misfit_task_load > busiest_load) { in sched_balance_find_src_rq()
11541 busiest_load = rq->misfit_task_load; in sched_balance_find_src_rq()
11572 return env->idle && sched_use_asym_prio(env->sd, env->dst_cpu) && in asym_active_balance()
11573 (sched_asym_prefer(env->dst_cpu, env->src_cpu) || in asym_active_balance()
11574 !sched_use_asym_prio(env->sd, env->src_cpu)); in asym_active_balance()
11580 struct sched_domain *sd = env->sd; in imbalanced_active_balance()
11587 if ((env->migration_type == migrate_task) && in imbalanced_active_balance()
11588 (sd->nr_balance_failed > sd->cache_nice_tries+2)) in imbalanced_active_balance()
11596 struct sched_domain *sd = env->sd; in need_active_balance()
11610 if (env->idle && in need_active_balance()
11611 (env->src_rq->cfs.h_nr_runnable == 1)) { in need_active_balance()
11612 if ((check_cpu_capacity(env->src_rq, sd)) && in need_active_balance()
11613 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) in need_active_balance()
11617 if (env->migration_type == migrate_misfit) in need_active_balance()
11628 struct sched_group *sg = env->sd->groups; in should_we_balance()
11629 int cpu, idle_smt = -1; in should_we_balance()
11635 if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) in should_we_balance()
11645 if (env->idle == CPU_NEWLY_IDLE) { in should_we_balance()
11646 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending) in should_we_balance()
11653 for_each_cpu_and(cpu, swb_cpus, env->cpus) { in should_we_balance()
11662 if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) { in should_we_balance()
11663 if (idle_smt == -1) in should_we_balance()
11677 * Are we the first idle core in a non-SMT domain or higher, in should_we_balance()
11680 return cpu == env->dst_cpu; in should_we_balance()
11684 if (idle_smt != -1) in should_we_balance()
11685 return idle_smt == env->dst_cpu; in should_we_balance()
11688 return group_balance_cpu(sg) == env->dst_cpu; in should_we_balance()
11697 switch (env->migration_type) { in update_lb_imbalance_stat()
11699 __schedstat_add(sd->lb_imbalance_load[idle], env->imbalance); in update_lb_imbalance_stat()
11702 __schedstat_add(sd->lb_imbalance_util[idle], env->imbalance); in update_lb_imbalance_stat()
11705 __schedstat_add(sd->lb_imbalance_task[idle], env->imbalance); in update_lb_imbalance_stat()
11708 __schedstat_add(sd->lb_imbalance_misfit[idle], env->imbalance); in update_lb_imbalance_stat()
11722 struct sched_domain *sd_parent = sd->parent; in sched_balance_rq()
11731 .dst_grpmask = group_balance_mask(sd->groups), in sched_balance_rq()
11741 schedstat_inc(sd->lb_count[idle]); in sched_balance_rq()
11751 schedstat_inc(sd->lb_nobusyg[idle]); in sched_balance_rq()
11757 schedstat_inc(sd->lb_nobusyq[idle]); in sched_balance_rq()
11765 env.src_cpu = busiest->cpu; in sched_balance_rq()
11771 if (busiest->nr_running > 1) { in sched_balance_rq()
11774 * an imbalance but busiest->nr_running <= 1, the group is in sched_balance_rq()
11778 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); in sched_balance_rq()
11785 * cur_ld_moved - load moved in current iteration in sched_balance_rq()
11786 * ld_moved - cumulative load moved across iterations in sched_balance_rq()
11793 * unlock busiest->lock, and we are able to be sure in sched_balance_rq()
11822 * nohz-idle), we now have balance_cpu in a position to move in sched_balance_rq()
11833 /* Prevent to re-select dst_cpu via env's CPUs */ in sched_balance_rq()
11853 int *group_imbalance = &sd_parent->groups->sgc->imbalance; in sched_balance_rq()
11880 schedstat_inc(sd->lb_failed[idle]); in sched_balance_rq()
11892 sd->nr_balance_failed++; in sched_balance_rq()
11904 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { in sched_balance_rq()
11913 * ->active_balance synchronizes accesses to in sched_balance_rq()
11914 * ->active_balance_work. Once set, it's cleared in sched_balance_rq()
11917 if (!busiest->active_balance) { in sched_balance_rq()
11918 busiest->active_balance = 1; in sched_balance_rq()
11919 busiest->push_cpu = this_cpu; in sched_balance_rq()
11928 &busiest->active_balance_work); in sched_balance_rq()
11933 sd->nr_balance_failed = 0; in sched_balance_rq()
11938 sd->balance_interval = sd->min_interval; in sched_balance_rq()
11950 int *group_imbalance = &sd_parent->groups->sgc->imbalance; in sched_balance_rq()
11962 schedstat_inc(sd->lb_balanced[idle]); in sched_balance_rq()
11964 sd->nr_balance_failed = 0; in sched_balance_rq()
11985 sd->balance_interval < MAX_PINNED_INTERVAL) || in sched_balance_rq()
11986 sd->balance_interval < sd->max_interval) in sched_balance_rq()
11987 sd->balance_interval *= 2; in sched_balance_rq()
11995 unsigned long interval = sd->balance_interval; in get_sd_balance_interval()
11998 interval *= sd->busy_factor; in get_sd_balance_interval()
12009 interval -= 1; in get_sd_balance_interval()
12023 next = sd->last_balance + interval; in update_next_balance()
12039 int target_cpu = busiest_rq->push_cpu; in active_load_balance_cpu_stop()
12047 * Between queueing the stop-work and running it is a hole in which in active_load_balance_cpu_stop()
12056 !busiest_rq->active_balance)) in active_load_balance_cpu_stop()
12060 if (busiest_rq->nr_running <= 1) in active_load_balance_cpu_stop()
12066 * Bjorn Helgaas on a 128-CPU setup. in active_load_balance_cpu_stop()
12082 .src_cpu = busiest_rq->cpu, in active_load_balance_cpu_stop()
12088 schedstat_inc(sd->alb_count); in active_load_balance_cpu_stop()
12093 schedstat_inc(sd->alb_pushed); in active_load_balance_cpu_stop()
12095 sd->nr_balance_failed = 0; in active_load_balance_cpu_stop()
12097 schedstat_inc(sd->alb_failed); in active_load_balance_cpu_stop()
12102 busiest_rq->active_balance = 0; in active_load_balance_cpu_stop()
12114 * This flag serializes load-balancing passes over large domains
12115 * (above the NODE topology level) - only one load-balancing instance
12119 * - Note that load-balancing passes triggered while another one
12120 * is executing are skipped and not re-tried.
12122 * - Also note that this does not serialize rebalance_domains()
12123 * execution, as non-SD_SERIALIZE domains will still be
12124 * load-balanced in parallel.
12130 * This trades load-balance latency on larger machines for less cross talk.
12139 if (cost > sd->max_newidle_lb_cost) { in update_newidle_cost()
12144 sd->max_newidle_lb_cost = cost; in update_newidle_cost()
12145 sd->last_decay_max_lb_cost = jiffies; in update_newidle_cost()
12146 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) { in update_newidle_cost()
12152 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256; in update_newidle_cost()
12153 sd->last_decay_max_lb_cost = jiffies; in update_newidle_cost()
12170 int cpu = rq->cpu; in sched_balance_domains()
12187 max_cost += sd->max_newidle_lb_cost; in sched_balance_domains()
12202 need_serialize = sd->flags & SD_SERIALIZE; in sched_balance_domains()
12208 if (time_after_eq(jiffies, sd->last_balance + interval)) { in sched_balance_domains()
12212 * env->dst_cpu, so we can't know our idle in sched_balance_domains()
12218 sd->last_balance = jiffies; in sched_balance_domains()
12224 if (time_after(next_balance, sd->last_balance + interval)) { in sched_balance_domains()
12225 next_balance = sd->last_balance + interval; in sched_balance_domains()
12231 * Ensure the rq-wide value also decays but keep it at a in sched_balance_domains()
12232 * reasonable floor to avoid funnies with rq->avg_idle. in sched_balance_domains()
12234 rq->max_idle_balance_cost = in sched_balance_domains()
12245 rq->next_balance = next_balance; in sched_balance_domains()
12251 return unlikely(!rcu_dereference_sched(rq->sd)); in on_null_domain()
12258 * - When one of the busy CPUs notices that there may be an idle rebalancing
12278 return -1; in find_new_ilb()
12282 * Kick a CPU to do the NOHZ balancing, if it is time for it, via a cross-CPU
12323 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd); in kick_ilb()
12335 int nr_busy, i, cpu = rq->cpu; in nohz_balancer_kick()
12338 if (unlikely(rq->idle_balance)) in nohz_balancer_kick()
12361 if (rq->nr_running >= 2) { in nohz_balancer_kick()
12368 sd = rcu_dereference(rq->sd); in nohz_balancer_kick()
12374 if (rq->cfs.h_nr_runnable >= 1 && check_cpu_capacity(rq, sd)) { in nohz_balancer_kick()
12423 * increase the overall cache utilization), we need a less-loaded LLC in nohz_balancer_kick()
12427 * the others are - so just get a NOHZ balance going if it looks in nohz_balancer_kick()
12430 nr_busy = atomic_read(&sds->nr_busy_cpus); in nohz_balancer_kick()
12453 if (!sd || !sd->nohz_idle) in set_cpu_sd_state_busy()
12455 sd->nohz_idle = 0; in set_cpu_sd_state_busy()
12457 atomic_inc(&sd->shared->nr_busy_cpus); in set_cpu_sd_state_busy()
12466 if (likely(!rq->nohz_tick_stopped)) in nohz_balance_exit_idle()
12469 rq->nohz_tick_stopped = 0; in nohz_balance_exit_idle()
12470 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); in nohz_balance_exit_idle()
12473 set_cpu_sd_state_busy(rq->cpu); in nohz_balance_exit_idle()
12483 if (!sd || sd->nohz_idle) in set_cpu_sd_state_idle()
12485 sd->nohz_idle = 1; in set_cpu_sd_state_idle()
12487 atomic_dec(&sd->shared->nr_busy_cpus); in set_cpu_sd_state_idle()
12507 * Can be set safely without rq->lock held in nohz_balance_enter_idle()
12509 * rq->lock is held during the check and the clear in nohz_balance_enter_idle()
12511 rq->has_blocked_load = 1; in nohz_balance_enter_idle()
12519 if (rq->nohz_tick_stopped) in nohz_balance_enter_idle()
12526 rq->nohz_tick_stopped = 1; in nohz_balance_enter_idle()
12551 unsigned int cpu = rq->cpu; in update_nohz_stats()
12553 if (!rq->has_blocked_load) in update_nohz_stats()
12559 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick))) in update_nohz_stats()
12564 return rq->has_blocked_load; in update_nohz_stats()
12579 int this_cpu = this_rq->cpu; in _nohz_idle_balance()
12636 if (time_after_eq(jiffies, rq->next_balance)) { in _nohz_idle_balance()
12647 if (time_after(next_balance, rq->next_balance)) { in _nohz_idle_balance()
12648 next_balance = rq->next_balance; in _nohz_idle_balance()
12677 unsigned int flags = this_rq->nohz_idle_balance; in nohz_idle_balance()
12682 this_rq->nohz_idle_balance = 0; in nohz_idle_balance()
12723 int this_cpu = this_rq->cpu; in nohz_newidle_balance()
12726 if (this_rq->avg_idle < sysctl_sched_migration_cost) in nohz_newidle_balance()
12757 * < 0 - we released the lock and there are !fair tasks present
12758 * 0 - failed, no new tasks
12759 * > 0 - success, new (fair) tasks present
12764 int this_cpu = this_rq->cpu; in sched_balance_newidle()
12776 if (this_rq->ttwu_pending) in sched_balance_newidle()
12784 this_rq->idle_stamp = rq_clock(this_rq); in sched_balance_newidle()
12794 * for load-balance and preemption/IRQs are still disabled avoiding in sched_balance_newidle()
12796 * re-start the picking loop. in sched_balance_newidle()
12801 sd = rcu_dereference_check_sched_domain(this_rq->sd); in sched_balance_newidle()
12803 if (!get_rd_overloaded(this_rq->rd) || in sched_balance_newidle()
12804 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) { in sched_balance_newidle()
12825 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) in sched_balance_newidle()
12828 if (sd->flags & SD_BALANCE_NEWIDLE) { in sched_balance_newidle()
12835 domain_cost = t1 - t0; in sched_balance_newidle()
12853 if (curr_cost > this_rq->max_idle_balance_cost) in sched_balance_newidle()
12854 this_rq->max_idle_balance_cost = curr_cost; in sched_balance_newidle()
12861 if (this_rq->cfs.h_nr_queued && !pulled_task) in sched_balance_newidle()
12865 if (this_rq->nr_running != this_rq->cfs.h_nr_queued) in sched_balance_newidle()
12866 pulled_task = -1; in sched_balance_newidle()
12870 if (time_after(this_rq->next_balance, next_balance)) in sched_balance_newidle()
12871 this_rq->next_balance = next_balance; in sched_balance_newidle()
12874 this_rq->idle_stamp = 0; in sched_balance_newidle()
12886 * - directly from the local sched_tick() for periodic load balancing
12888 * - indirectly from a remote sched_tick() for NOHZ idle balancing
12889 * through the SMP cross-call nohz_csd_func()
12894 enum cpu_idle_type idle = this_rq->idle_balance; in sched_balance_softirq()
12907 sched_balance_update_blocked_averages(this_rq->cpu); in sched_balance_softirq()
12923 if (time_after_eq(jiffies, rq->next_balance)) in sched_balance_trigger()
12953 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime; in __entity_slice_used()
12954 u64 slice = se->slice; in __entity_slice_used()
12976 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check in task_tick_core()
12979 if (rq->core->core_forceidle_count && rq->cfs.nr_queued == 1 && in task_tick_core()
12980 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE)) in task_tick_core()
12985 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
12994 if (cfs_rq->forceidle_seq == fi_seq) in se_fi_update()
12996 cfs_rq->forceidle_seq = fi_seq; in se_fi_update()
12999 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime; in se_fi_update()
13005 struct sched_entity *se = &p->se; in task_vruntime_update()
13007 if (p->sched_class != &fair_sched_class) in task_vruntime_update()
13010 se_fi_update(se, rq->core->core_forceidle_seq, in_fi); in task_vruntime_update()
13017 const struct sched_entity *sea = &a->se; in cfs_prio_less()
13018 const struct sched_entity *seb = &b->se; in cfs_prio_less()
13023 SCHED_WARN_ON(task_rq(b)->core != rq->core); in cfs_prio_less()
13030 while (sea->cfs_rq->tg != seb->cfs_rq->tg) { in cfs_prio_less()
13031 int sea_depth = sea->depth; in cfs_prio_less()
13032 int seb_depth = seb->depth; in cfs_prio_less()
13040 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi); in cfs_prio_less()
13041 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi); in cfs_prio_less()
13043 cfs_rqa = sea->cfs_rq; in cfs_prio_less()
13044 cfs_rqb = seb->cfs_rq; in cfs_prio_less()
13046 cfs_rqa = &task_rq(a)->cfs; in cfs_prio_less()
13047 cfs_rqb = &task_rq(b)->cfs; in cfs_prio_less()
13055 delta = (s64)(sea->vruntime - seb->vruntime) + in cfs_prio_less()
13056 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi); in cfs_prio_less()
13066 cfs_rq = task_group(p)->cfs_rq[cpu]; in task_is_throttled_fair()
13068 cfs_rq = &cpu_rq(cpu)->cfs; in task_is_throttled_fair()
13087 struct sched_entity *se = &curr->se; in task_tick_fair()
13105 * - child not yet on the tasklist
13106 * - preemption disabled
13123 if (rq->cfs.nr_queued == 1) in prio_changed_fair()
13132 if (p->prio > oldprio) in prio_changed_fair()
13154 se = se->parent; in propagate_entity_cfs_rq()
13179 * - A forked task which hasn't been woken up by wake_up_new_task(). in detach_entity_cfs_rq()
13180 * - A task which has been woken up by try_to_wake_up() but is in detach_entity_cfs_rq()
13183 if (!se->avg.last_update_time) in detach_entity_cfs_rq()
13207 struct sched_entity *se = &p->se; in detach_task_cfs_rq()
13214 struct sched_entity *se = &p->se; in attach_task_cfs_rq()
13226 SCHED_WARN_ON(p->se.sched_delayed); in switched_to_fair()
13247 struct sched_entity *se = &p->se; in __set_next_task_fair()
13255 list_move(&se->group_node, &rq->cfs_tasks); in __set_next_task_fair()
13261 SCHED_WARN_ON(se->sched_delayed); in __set_next_task_fair()
13273 * This routine is mostly called to set cfs_rq->curr field when a task
13278 struct sched_entity *se = &p->se; in set_next_task_fair()
13293 cfs_rq->tasks_timeline = RB_ROOT_CACHED; in init_cfs_rq()
13294 cfs_rq->min_vruntime = (u64)(-(1LL << 20)); in init_cfs_rq()
13296 raw_spin_lock_init(&cfs_rq->removed.lock); in init_cfs_rq()
13307 if (READ_ONCE(p->__state) == TASK_NEW) in task_change_group_fair()
13313 /* Tell se's cfs_rq has been changed -- migrated */ in task_change_group_fair()
13314 p->se.avg.last_update_time = 0; in task_change_group_fair()
13325 if (tg->cfs_rq) in free_fair_sched_group()
13326 kfree(tg->cfs_rq[i]); in free_fair_sched_group()
13327 if (tg->se) in free_fair_sched_group()
13328 kfree(tg->se[i]); in free_fair_sched_group()
13331 kfree(tg->cfs_rq); in free_fair_sched_group()
13332 kfree(tg->se); in free_fair_sched_group()
13341 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); in alloc_fair_sched_group()
13342 if (!tg->cfs_rq) in alloc_fair_sched_group()
13344 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); in alloc_fair_sched_group()
13345 if (!tg->se) in alloc_fair_sched_group()
13348 tg->shares = NICE_0_LOAD; in alloc_fair_sched_group()
13364 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); in alloc_fair_sched_group()
13385 se = tg->se[i]; in online_fair_sched_group()
13401 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; in unregister_fair_sched_group()
13402 struct sched_entity *se = tg->se[cpu]; in unregister_fair_sched_group()
13406 if (se->sched_delayed) { in unregister_fair_sched_group()
13408 if (se->sched_delayed) { in unregister_fair_sched_group()
13419 * check on_list without danger of it being re-added. in unregister_fair_sched_group()
13421 if (cfs_rq->on_list) { in unregister_fair_sched_group()
13434 cfs_rq->tg = tg; in init_tg_cfs_entry()
13435 cfs_rq->rq = rq; in init_tg_cfs_entry()
13438 tg->cfs_rq[cpu] = cfs_rq; in init_tg_cfs_entry()
13439 tg->se[cpu] = se; in init_tg_cfs_entry()
13446 se->cfs_rq = &rq->cfs; in init_tg_cfs_entry()
13447 se->depth = 0; in init_tg_cfs_entry()
13449 se->cfs_rq = parent->my_q; in init_tg_cfs_entry()
13450 se->depth = parent->depth + 1; in init_tg_cfs_entry()
13453 se->my_q = cfs_rq; in init_tg_cfs_entry()
13455 update_load_set(&se->load, NICE_0_LOAD); in init_tg_cfs_entry()
13456 se->parent = parent; in init_tg_cfs_entry()
13470 if (!tg->se[0]) in __sched_group_set_shares()
13471 return -EINVAL; in __sched_group_set_shares()
13475 if (tg->shares == shares) in __sched_group_set_shares()
13478 tg->shares = shares; in __sched_group_set_shares()
13481 struct sched_entity *se = tg->se[i]; in __sched_group_set_shares()
13503 ret = -EINVAL; in sched_group_set_shares()
13516 return -EINVAL; in sched_group_set_idle()
13519 return -EINVAL; in sched_group_set_idle()
13523 if (tg->idle == idle) { in sched_group_set_idle()
13528 tg->idle = idle; in sched_group_set_idle()
13532 struct sched_entity *se = tg->se[i]; in sched_group_set_idle()
13533 struct cfs_rq *grp_cfs_rq = tg->cfs_rq[i]; in sched_group_set_idle()
13540 grp_cfs_rq->idle = idle; in sched_group_set_idle()
13544 idle_task_delta = grp_cfs_rq->h_nr_queued - in sched_group_set_idle()
13545 grp_cfs_rq->h_nr_idle; in sched_group_set_idle()
13547 idle_task_delta *= -1; in sched_group_set_idle()
13552 if (!se->on_rq) in sched_group_set_idle()
13555 cfs_rq->h_nr_idle += idle_task_delta; in sched_group_set_idle()
13581 struct sched_entity *se = &task->se; in get_rr_interval_fair()
13588 if (rq->cfs.load.weight) in get_rr_interval_fair()
13589 rr_interval = NS_TO_JIFFIES(se->slice); in get_rr_interval_fair()
13667 ng = rcu_dereference(p->numa_group); in show_numa_stats()
13669 if (p->numa_faults) { in show_numa_stats()
13670 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; in show_numa_stats()
13671 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; in show_numa_stats()
13674 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], in show_numa_stats()
13675 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; in show_numa_stats()
13696 INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i)); in init_sched_fair_class()
13697 INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list); in init_sched_fair_class()