Lines Matching +full:overrun +full:- +full:throttle +full:- +full:ms
1 // SPDX-License-Identifier: GPL-2.0
26 * Targeted preemption latency for CPU-bound tasks:
29 * 'timeslice length' - timeslices in CFS are of variable length
30 * and have no persistent notion like in traditional, time-slice
34 * run vmstat and monitor the context-switches (cs) field)
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * The initial- and re-scaling of tunables is configurable
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
55 * Minimal preemption granularity for CPU-bound tasks:
74 * SCHED_OTHER wake-up granularity.
77 * and reduces their over-scheduling. Synchronous workloads will still
106 return -cpu; in arch_asym_cpu_priority()
127 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
141 lw->weight += inc; in update_load_add()
142 lw->inv_weight = 0; in update_load_add()
147 lw->weight -= dec; in update_load_sub()
148 lw->inv_weight = 0; in update_load_sub()
153 lw->weight = w; in update_load_set()
154 lw->inv_weight = 0; in update_load_set()
161 * so pick a second-best guess by going with the log2 of the
211 if (likely(lw->inv_weight)) in __update_inv_weight()
214 w = scale_load_down(lw->weight); in __update_inv_weight()
217 lw->inv_weight = 1; in __update_inv_weight()
219 lw->inv_weight = WMULT_CONST; in __update_inv_weight()
221 lw->inv_weight = WMULT_CONST / w; in __update_inv_weight()
227 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
247 shift -= fs; in __calc_delta()
251 fact = mul_u32_u32(fact, lw->inv_weight); in __calc_delta()
256 shift -= fs; in __calc_delta()
274 for (; se; se = se->parent)
281 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg)) in cfs_rq_tg_path()
282 autogroup_path(cfs_rq->tg, path, len); in cfs_rq_tg_path()
283 else if (cfs_rq && cfs_rq->tg->css.cgroup) in cfs_rq_tg_path()
284 cgroup_path(cfs_rq->tg->css.cgroup, path, len); in cfs_rq_tg_path()
294 if (cfs_rq->on_list) in list_add_leaf_cfs_rq()
295 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; in list_add_leaf_cfs_rq()
297 cfs_rq->on_list = 1; in list_add_leaf_cfs_rq()
302 * enqueued. The fact that we always enqueue bottom-up in list_add_leaf_cfs_rq()
308 if (cfs_rq->tg->parent && in list_add_leaf_cfs_rq()
309 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { in list_add_leaf_cfs_rq()
316 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, in list_add_leaf_cfs_rq()
317 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); in list_add_leaf_cfs_rq()
323 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; in list_add_leaf_cfs_rq()
327 if (!cfs_rq->tg->parent) { in list_add_leaf_cfs_rq()
332 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, in list_add_leaf_cfs_rq()
333 &rq->leaf_cfs_rq_list); in list_add_leaf_cfs_rq()
338 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; in list_add_leaf_cfs_rq()
348 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); in list_add_leaf_cfs_rq()
353 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; in list_add_leaf_cfs_rq()
359 if (cfs_rq->on_list) { in list_del_leaf_cfs_rq()
366 * to the prev element but it will point to rq->leaf_cfs_rq_list in list_del_leaf_cfs_rq()
369 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) in list_del_leaf_cfs_rq()
370 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; in list_del_leaf_cfs_rq()
372 list_del_rcu(&cfs_rq->leaf_cfs_rq_list); in list_del_leaf_cfs_rq()
373 cfs_rq->on_list = 0; in list_del_leaf_cfs_rq()
379 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); in assert_list_leaf_cfs_rq()
384 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
391 if (se->cfs_rq == pse->cfs_rq) in is_same_group()
392 return se->cfs_rq; in is_same_group()
399 return se->parent; in parent_entity()
415 se_depth = (*se)->depth; in find_matching_se()
416 pse_depth = (*pse)->depth; in find_matching_se()
419 se_depth--; in find_matching_se()
424 pse_depth--; in find_matching_se()
436 return tg->idle > 0; in tg_is_idle()
441 return cfs_rq->idle > 0; in cfs_rq_is_idle()
476 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
514 s64 delta = (s64)(vruntime - max_vruntime); in max_vruntime()
523 s64 delta = (s64)(vruntime - min_vruntime); in min_vruntime()
533 return (s64)(a->vruntime - b->vruntime) < 0; in entity_before()
541 struct sched_entity *curr = cfs_rq->curr; in update_min_vruntime()
542 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); in update_min_vruntime()
544 u64 vruntime = cfs_rq->min_vruntime; in update_min_vruntime()
547 if (curr->on_rq) in update_min_vruntime()
548 vruntime = curr->vruntime; in update_min_vruntime()
553 if (leftmost) { /* non-empty tree */ in update_min_vruntime()
557 vruntime = se->vruntime; in update_min_vruntime()
559 vruntime = min_vruntime(vruntime, se->vruntime); in update_min_vruntime()
563 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); in update_min_vruntime()
566 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; in update_min_vruntime()
576 * Enqueue an entity into the rb-tree:
580 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less); in __enqueue_entity()
585 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); in __dequeue_entity()
590 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); in __pick_first_entity()
600 struct rb_node *next = rb_next(&se->run_node); in __pick_next_entity()
611 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); in __pick_last_entity()
646 if (unlikely(se->load.weight != NICE_0_LOAD)) in calc_delta_fair()
647 delta = __calc_delta(delta, NICE_0_LOAD, &se->load); in calc_delta_fair()
669 * We calculate the wall-time slice from the period by taking a part
676 unsigned int nr_running = cfs_rq->nr_running; in sched_slice()
680 nr_running = rq_of(cfs_rq)->cfs.h_nr_running; in sched_slice()
682 slice = __sched_period(nr_running + !se->on_rq); in sched_slice()
689 load = &cfs_rq->load; in sched_slice()
691 if (unlikely(!se->on_rq)) { in sched_slice()
692 lw = cfs_rq->load; in sched_slice()
694 update_load_add(&lw, se->load.weight); in sched_slice()
697 slice = __calc_delta(slice, se->load.weight, load); in sched_slice()
707 * We calculate the vruntime slice of a to-be-inserted task.
726 struct sched_avg *sa = &se->avg; in init_entity_runnable_average()
737 sa->load_avg = scale_load_down(se->load.weight); in init_entity_runnable_average()
748 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
757 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
772 struct sched_entity *se = &p->se; in post_init_entity_util_avg()
774 struct sched_avg *sa = &se->avg; in post_init_entity_util_avg()
776 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; in post_init_entity_util_avg()
779 if (cfs_rq->avg.util_avg != 0) { in post_init_entity_util_avg()
780 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; in post_init_entity_util_avg()
781 sa->util_avg /= (cfs_rq->avg.load_avg + 1); in post_init_entity_util_avg()
783 if (sa->util_avg > cap) in post_init_entity_util_avg()
784 sa->util_avg = cap; in post_init_entity_util_avg()
786 sa->util_avg = cap; in post_init_entity_util_avg()
790 sa->runnable_avg = sa->util_avg; in post_init_entity_util_avg()
792 if (p->sched_class != &fair_sched_class) { in post_init_entity_util_avg()
803 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); in post_init_entity_util_avg()
827 struct sched_entity *curr = cfs_rq->curr; in update_curr()
834 delta_exec = now - curr->exec_start; in update_curr()
838 curr->exec_start = now; in update_curr()
840 schedstat_set(curr->statistics.exec_max, in update_curr()
841 max(delta_exec, curr->statistics.exec_max)); in update_curr()
843 curr->sum_exec_runtime += delta_exec; in update_curr()
844 schedstat_add(cfs_rq->exec_clock, delta_exec); in update_curr()
846 curr->vruntime += calc_delta_fair(delta_exec, curr); in update_curr()
852 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); in update_curr()
862 update_curr(cfs_rq_of(&rq->curr->se)); in update_curr_fair()
874 prev_wait_start = schedstat_val(se->statistics.wait_start); in update_stats_wait_start()
878 wait_start -= prev_wait_start; in update_stats_wait_start()
880 __schedstat_set(se->statistics.wait_start, wait_start); in update_stats_wait_start()
894 * maybe already in the runqueue, the se->statistics.wait_start in update_stats_wait_end()
898 if (unlikely(!schedstat_val(se->statistics.wait_start))) in update_stats_wait_end()
901 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); in update_stats_wait_end()
911 __schedstat_set(se->statistics.wait_start, delta); in update_stats_wait_end()
917 __schedstat_set(se->statistics.wait_max, in update_stats_wait_end()
918 max(schedstat_val(se->statistics.wait_max), delta)); in update_stats_wait_end()
919 __schedstat_inc(se->statistics.wait_count); in update_stats_wait_end()
920 __schedstat_add(se->statistics.wait_sum, delta); in update_stats_wait_end()
921 __schedstat_set(se->statistics.wait_start, 0); in update_stats_wait_end()
933 sleep_start = schedstat_val(se->statistics.sleep_start); in update_stats_enqueue_sleeper()
934 block_start = schedstat_val(se->statistics.block_start); in update_stats_enqueue_sleeper()
940 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; in update_stats_enqueue_sleeper()
945 if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) in update_stats_enqueue_sleeper()
946 __schedstat_set(se->statistics.sleep_max, delta); in update_stats_enqueue_sleeper()
948 __schedstat_set(se->statistics.sleep_start, 0); in update_stats_enqueue_sleeper()
949 __schedstat_add(se->statistics.sum_sleep_runtime, delta); in update_stats_enqueue_sleeper()
957 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; in update_stats_enqueue_sleeper()
962 if (unlikely(delta > schedstat_val(se->statistics.block_max))) in update_stats_enqueue_sleeper()
963 __schedstat_set(se->statistics.block_max, delta); in update_stats_enqueue_sleeper()
965 __schedstat_set(se->statistics.block_start, 0); in update_stats_enqueue_sleeper()
966 __schedstat_add(se->statistics.sum_sleep_runtime, delta); in update_stats_enqueue_sleeper()
969 if (tsk->in_iowait) { in update_stats_enqueue_sleeper()
970 __schedstat_add(se->statistics.iowait_sum, delta); in update_stats_enqueue_sleeper()
971 __schedstat_inc(se->statistics.iowait_count); in update_stats_enqueue_sleeper()
979 * 20 to get a milliseconds-range estimation of the in update_stats_enqueue_sleeper()
993 * Task is being enqueued - update stats:
1005 if (se != cfs_rq->curr) in update_stats_enqueue()
1023 if (se != cfs_rq->curr) in update_stats_dequeue()
1031 state = READ_ONCE(tsk->__state); in update_stats_dequeue()
1033 __schedstat_set(se->statistics.sleep_start, in update_stats_dequeue()
1036 __schedstat_set(se->statistics.block_start, in update_stats_dequeue()
1042 * We are picking a new current task - update its stats:
1050 se->exec_start = rq_clock_task(rq_of(cfs_rq)); in update_stats_curr_start()
1059 * Approximate time to scan a full NUMA task in ms. The task scan period is
1069 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1094 * ->numa_group (see struct task_struct for locking rules).
1098 return rcu_dereference_check(p->numa_group, p == current || in deref_task_numa_group()
1099 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu))); in deref_task_numa_group()
1104 return rcu_dereference_protected(p->numa_group, p == current); in deref_curr_numa_group()
1116 * Calculations based on RSS as non-present and empty pages are skipped in task_nr_scan_windows()
1120 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); in task_nr_scan_windows()
1121 rss = get_mm_rss(p->mm); in task_nr_scan_windows()
1154 ng = rcu_dereference(p->numa_group); in task_scan_start()
1159 period *= refcount_read(&ng->refcount); in task_scan_start()
1184 period *= refcount_read(&ng->refcount); in task_scan_max()
1196 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); in account_numa_enqueue()
1197 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); in account_numa_enqueue()
1202 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE); in account_numa_dequeue()
1203 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); in account_numa_dequeue()
1221 ng = rcu_dereference(p->numa_group); in task_numa_group_id()
1223 gid = ng->gid; in task_numa_group_id()
1242 if (!p->numa_faults) in task_faults()
1245 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + in task_faults()
1246 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; in task_faults()
1256 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + in group_faults()
1257 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; in group_faults()
1262 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + in group_faults_cpu()
1263 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; in group_faults_cpu()
1272 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; in group_faults_priv()
1284 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; in group_faults_shared()
1292 * considered part of a numa group's pseudo-interleaving set. Migrations
1299 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; in numa_is_active_node()
1357 faults *= (sched_max_numa_distance - dist); in score_nearby_nodes()
1358 faults /= (sched_max_numa_distance - LOCAL_DISTANCE); in score_nearby_nodes()
1378 if (!p->numa_faults) in task_weight()
1381 total_faults = p->total_numa_faults; in task_weight()
1401 total_faults = ng->total_faults; in group_weight()
1419 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); in should_numa_migrate_memory()
1425 * two full passes of the "multi-stage node selection" test that is in should_numa_migrate_memory()
1428 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && in should_numa_migrate_memory()
1433 * Multi-stage node selection is used in conjunction with a periodic in should_numa_migrate_memory()
1434 * migration fault to build a temporal task<->page relation. By using in should_numa_migrate_memory()
1435 * a two-stage filter we remove short/unlikely relations. in should_numa_migrate_memory()
1439 * page (n_t) (in a given time-span) to a probability. in should_numa_migrate_memory()
1447 * act on an unlikely task<->page relation. in should_numa_migrate_memory()
1457 /* A shared fault, but p->numa_group has not been set up yet. */ in should_numa_migrate_memory()
1474 * --------------- * - > --------------- in should_numa_migrate_memory()
1555 if ((ns->nr_running > ns->weight) && in numa_classify()
1556 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) || in numa_classify()
1557 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100)))) in numa_classify()
1560 if ((ns->nr_running < ns->weight) || in numa_classify()
1561 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) && in numa_classify()
1562 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100)))) in numa_classify()
1603 int cpu, idle_core = -1; in update_numa_stats()
1606 ns->idle_cpu = -1; in update_numa_stats()
1612 ns->load += cpu_load(rq); in update_numa_stats()
1613 ns->runnable += cpu_runnable(rq); in update_numa_stats()
1614 ns->util += cpu_util(cpu); in update_numa_stats()
1615 ns->nr_running += rq->cfs.h_nr_running; in update_numa_stats()
1616 ns->compute_capacity += capacity_of(cpu); in update_numa_stats()
1618 if (find_idle && !rq->nr_running && idle_cpu(cpu)) { in update_numa_stats()
1619 if (READ_ONCE(rq->numa_migrate_on) || in update_numa_stats()
1620 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) in update_numa_stats()
1623 if (ns->idle_cpu == -1) in update_numa_stats()
1624 ns->idle_cpu = cpu; in update_numa_stats()
1631 ns->weight = cpumask_weight(cpumask_of_node(nid)); in update_numa_stats()
1633 ns->node_type = numa_classify(env->imbalance_pct, ns); in update_numa_stats()
1636 ns->idle_cpu = idle_core; in update_numa_stats()
1642 struct rq *rq = cpu_rq(env->dst_cpu); in task_numa_assign()
1644 /* Check if run-queue part of active NUMA balance. */ in task_numa_assign()
1645 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) { in task_numa_assign()
1647 int start = env->dst_cpu; in task_numa_assign()
1650 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) { in task_numa_assign()
1651 if (cpu == env->best_cpu || !idle_cpu(cpu) || in task_numa_assign()
1652 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) { in task_numa_assign()
1656 env->dst_cpu = cpu; in task_numa_assign()
1657 rq = cpu_rq(env->dst_cpu); in task_numa_assign()
1658 if (!xchg(&rq->numa_migrate_on, 1)) in task_numa_assign()
1668 * Clear previous best_cpu/rq numa-migrate flag, since task now in task_numa_assign()
1671 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) { in task_numa_assign()
1672 rq = cpu_rq(env->best_cpu); in task_numa_assign()
1673 WRITE_ONCE(rq->numa_migrate_on, 0); in task_numa_assign()
1676 if (env->best_task) in task_numa_assign()
1677 put_task_struct(env->best_task); in task_numa_assign()
1681 env->best_task = p; in task_numa_assign()
1682 env->best_imp = imp; in task_numa_assign()
1683 env->best_cpu = env->dst_cpu; in task_numa_assign()
1697 * ------------ vs --------- in load_too_imbalanced()
1700 src_capacity = env->src_stats.compute_capacity; in load_too_imbalanced()
1701 dst_capacity = env->dst_stats.compute_capacity; in load_too_imbalanced()
1703 imb = abs(dst_load * src_capacity - src_load * dst_capacity); in load_too_imbalanced()
1705 orig_src_load = env->src_stats.load; in load_too_imbalanced()
1706 orig_dst_load = env->dst_stats.load; in load_too_imbalanced()
1708 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); in load_too_imbalanced()
1730 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); in task_numa_compare()
1731 struct rq *dst_rq = cpu_rq(env->dst_cpu); in task_numa_compare()
1735 int dist = env->dist; in task_numa_compare()
1740 if (READ_ONCE(dst_rq->numa_migrate_on)) in task_numa_compare()
1744 cur = rcu_dereference(dst_rq->curr); in task_numa_compare()
1745 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) in task_numa_compare()
1750 * end try selecting ourselves (current == env->p) as a swap candidate. in task_numa_compare()
1752 if (cur == env->p) { in task_numa_compare()
1758 if (maymove && moveimp >= env->best_imp) in task_numa_compare()
1765 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) in task_numa_compare()
1772 if (env->best_task && in task_numa_compare()
1773 env->best_task->numa_preferred_nid == env->src_nid && in task_numa_compare()
1774 cur->numa_preferred_nid != env->src_nid) { in task_numa_compare()
1788 cur_ng = rcu_dereference(cur->numa_group); in task_numa_compare()
1790 imp = taskimp + task_weight(cur, env->src_nid, dist) - in task_numa_compare()
1791 task_weight(cur, env->dst_nid, dist); in task_numa_compare()
1797 imp -= imp / 16; in task_numa_compare()
1804 imp += group_weight(cur, env->src_nid, dist) - in task_numa_compare()
1805 group_weight(cur, env->dst_nid, dist); in task_numa_compare()
1807 imp += task_weight(cur, env->src_nid, dist) - in task_numa_compare()
1808 task_weight(cur, env->dst_nid, dist); in task_numa_compare()
1812 if (cur->numa_preferred_nid == env->dst_nid) in task_numa_compare()
1813 imp -= imp / 16; in task_numa_compare()
1821 if (cur->numa_preferred_nid == env->src_nid) in task_numa_compare()
1824 if (maymove && moveimp > imp && moveimp > env->best_imp) { in task_numa_compare()
1834 if (env->best_task && cur->numa_preferred_nid == env->src_nid && in task_numa_compare()
1835 env->best_task->numa_preferred_nid != env->src_nid) { in task_numa_compare()
1845 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) in task_numa_compare()
1851 load = task_h_load(env->p) - task_h_load(cur); in task_numa_compare()
1855 dst_load = env->dst_stats.load + load; in task_numa_compare()
1856 src_load = env->src_stats.load - load; in task_numa_compare()
1864 int cpu = env->dst_stats.idle_cpu; in task_numa_compare()
1868 cpu = env->dst_cpu; in task_numa_compare()
1874 if (!idle_cpu(cpu) && env->best_cpu >= 0 && in task_numa_compare()
1875 idle_cpu(env->best_cpu)) { in task_numa_compare()
1876 cpu = env->best_cpu; in task_numa_compare()
1879 env->dst_cpu = cpu; in task_numa_compare()
1889 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu)) in task_numa_compare()
1896 if (!maymove && env->best_task && in task_numa_compare()
1897 env->best_task->numa_preferred_nid == env->src_nid) { in task_numa_compare()
1916 if (env->dst_stats.node_type == node_has_spare) { in task_numa_find_cpu()
1926 src_running = env->src_stats.nr_running - 1; in task_numa_find_cpu()
1927 dst_running = env->dst_stats.nr_running + 1; in task_numa_find_cpu()
1928 imbalance = max(0, dst_running - src_running); in task_numa_find_cpu()
1930 env->dst_stats.weight); in task_numa_find_cpu()
1935 if (env->dst_stats.idle_cpu >= 0) { in task_numa_find_cpu()
1936 env->dst_cpu = env->dst_stats.idle_cpu; in task_numa_find_cpu()
1944 * If the improvement from just moving env->p direction is better in task_numa_find_cpu()
1947 load = task_h_load(env->p); in task_numa_find_cpu()
1948 dst_load = env->dst_stats.load + load; in task_numa_find_cpu()
1949 src_load = env->src_stats.load - load; in task_numa_find_cpu()
1953 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { in task_numa_find_cpu()
1955 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) in task_numa_find_cpu()
1958 env->dst_cpu = cpu; in task_numa_find_cpu()
1976 .best_cpu = -1, in task_numa_migrate()
1990 * random movement of tasks -- counter the numa conditions we're trying in task_numa_migrate()
1996 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; in task_numa_migrate()
2007 return -EINVAL; in task_numa_migrate()
2010 env.dst_nid = p->numa_preferred_nid; in task_numa_migrate()
2015 taskimp = task_weight(p, env.dst_nid, dist) - taskweight; in task_numa_migrate()
2016 groupimp = group_weight(p, env.dst_nid, dist) - groupweight; in task_numa_migrate()
2024 * - there is no space available on the preferred_nid in task_numa_migrate()
2025 * - the task is part of a numa_group that is interleaved across in task_numa_migrate()
2030 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { in task_numa_migrate()
2032 if (nid == env.src_nid || nid == p->numa_preferred_nid) in task_numa_migrate()
2043 taskimp = task_weight(p, nid, dist) - taskweight; in task_numa_migrate()
2044 groupimp = group_weight(p, nid, dist) - groupweight; in task_numa_migrate()
2064 if (env.best_cpu == -1) in task_numa_migrate()
2069 if (nid != p->numa_preferred_nid) in task_numa_migrate()
2074 if (env.best_cpu == -1) { in task_numa_migrate()
2075 trace_sched_stick_numa(p, env.src_cpu, NULL, -1); in task_numa_migrate()
2076 return -EAGAIN; in task_numa_migrate()
2082 WRITE_ONCE(best_rq->numa_migrate_on, 0); in task_numa_migrate()
2089 WRITE_ONCE(best_rq->numa_migrate_on, 0); in task_numa_migrate()
2103 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) in numa_migrate_preferred()
2107 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); in numa_migrate_preferred()
2108 p->numa_migrate_retry = jiffies + interval; in numa_migrate_preferred()
2111 if (task_node(p) == p->numa_preferred_nid) in numa_migrate_preferred()
2141 numa_group->max_faults_cpu = max_faults; in numa_group_count_active_nodes()
2142 numa_group->active_nodes = active_nodes; in numa_group_count_active_nodes()
2168 unsigned long remote = p->numa_faults_locality[0]; in update_task_scan_period()
2169 unsigned long local = p->numa_faults_locality[1]; in update_task_scan_period()
2178 if (local + shared == 0 || p->numa_faults_locality[2]) { in update_task_scan_period()
2179 p->numa_scan_period = min(p->numa_scan_period_max, in update_task_scan_period()
2180 p->numa_scan_period << 1); in update_task_scan_period()
2182 p->mm->numa_next_scan = jiffies + in update_task_scan_period()
2183 msecs_to_jiffies(p->numa_scan_period); in update_task_scan_period()
2194 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); in update_task_scan_period()
2203 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; in update_task_scan_period()
2213 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; in update_task_scan_period()
2219 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, in update_task_scan_period()
2224 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; in update_task_scan_period()
2227 p->numa_scan_period = clamp(p->numa_scan_period + diff, in update_task_scan_period()
2229 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); in update_task_scan_period()
2235 * decays those on a 32ms period, which is orders of magnitude off
2236 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2243 now = p->se.exec_start; in numa_get_avg_runtime()
2244 runtime = p->se.sum_exec_runtime; in numa_get_avg_runtime()
2246 if (p->last_task_numa_placement) { in numa_get_avg_runtime()
2247 delta = runtime - p->last_sum_exec_runtime; in numa_get_avg_runtime()
2248 *period = now - p->last_task_numa_placement; in numa_get_avg_runtime()
2254 delta = p->se.avg.load_sum; in numa_get_avg_runtime()
2258 p->last_sum_exec_runtime = runtime; in numa_get_avg_runtime()
2259 p->last_task_numa_placement = now; in numa_get_avg_runtime()
2309 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { in preferred_group_nid()
2363 * The p->mm->numa_scan_seq field gets updated without in task_numa_placement()
2367 seq = READ_ONCE(p->mm->numa_scan_seq); in task_numa_placement()
2368 if (p->numa_scan_seq == seq) in task_numa_placement()
2370 p->numa_scan_seq = seq; in task_numa_placement()
2371 p->numa_scan_period_max = task_scan_max(p); in task_numa_placement()
2373 total_faults = p->numa_faults_locality[0] + in task_numa_placement()
2374 p->numa_faults_locality[1]; in task_numa_placement()
2380 group_lock = &ng->lock; in task_numa_placement()
2400 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; in task_numa_placement()
2401 fault_types[priv] += p->numa_faults[membuf_idx]; in task_numa_placement()
2402 p->numa_faults[membuf_idx] = 0; in task_numa_placement()
2408 * little over-all impact on throughput, and thus their in task_numa_placement()
2412 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / in task_numa_placement()
2414 f_diff = f_weight - p->numa_faults[cpu_idx] / 2; in task_numa_placement()
2415 p->numa_faults[cpubuf_idx] = 0; in task_numa_placement()
2417 p->numa_faults[mem_idx] += diff; in task_numa_placement()
2418 p->numa_faults[cpu_idx] += f_diff; in task_numa_placement()
2419 faults += p->numa_faults[mem_idx]; in task_numa_placement()
2420 p->total_numa_faults += diff; in task_numa_placement()
2429 ng->faults[mem_idx] += diff; in task_numa_placement()
2430 ng->faults_cpu[mem_idx] += f_diff; in task_numa_placement()
2431 ng->total_faults += diff; in task_numa_placement()
2432 group_faults += ng->faults[mem_idx]; in task_numa_placement()
2455 if (max_nid != p->numa_preferred_nid) in task_numa_placement()
2464 return refcount_inc_not_zero(&grp->refcount); in get_numa_group()
2469 if (refcount_dec_and_test(&grp->refcount)) in put_numa_group()
2490 refcount_set(&grp->refcount, 1); in task_numa_group()
2491 grp->active_nodes = 1; in task_numa_group()
2492 grp->max_faults_cpu = 0; in task_numa_group()
2493 spin_lock_init(&grp->lock); in task_numa_group()
2494 grp->gid = p->pid; in task_numa_group()
2496 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * in task_numa_group()
2500 grp->faults[i] = p->numa_faults[i]; in task_numa_group()
2502 grp->total_faults = p->total_numa_faults; in task_numa_group()
2504 grp->nr_tasks++; in task_numa_group()
2505 rcu_assign_pointer(p->numa_group, grp); in task_numa_group()
2509 tsk = READ_ONCE(cpu_rq(cpu)->curr); in task_numa_group()
2514 grp = rcu_dereference(tsk->numa_group); in task_numa_group()
2526 if (my_grp->nr_tasks > grp->nr_tasks) in task_numa_group()
2530 * Tie-break on the grp address. in task_numa_group()
2532 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) in task_numa_group()
2536 if (tsk->mm == current->mm) in task_numa_group()
2555 double_lock_irq(&my_grp->lock, &grp->lock); in task_numa_group()
2558 my_grp->faults[i] -= p->numa_faults[i]; in task_numa_group()
2559 grp->faults[i] += p->numa_faults[i]; in task_numa_group()
2561 my_grp->total_faults -= p->total_numa_faults; in task_numa_group()
2562 grp->total_faults += p->total_numa_faults; in task_numa_group()
2564 my_grp->nr_tasks--; in task_numa_group()
2565 grp->nr_tasks++; in task_numa_group()
2567 spin_unlock(&my_grp->lock); in task_numa_group()
2568 spin_unlock_irq(&grp->lock); in task_numa_group()
2570 rcu_assign_pointer(p->numa_group, grp); in task_numa_group()
2585 * reset the data back to default state without freeing ->numa_faults.
2590 struct numa_group *grp = rcu_dereference_raw(p->numa_group); in task_numa_free()
2591 unsigned long *numa_faults = p->numa_faults; in task_numa_free()
2599 spin_lock_irqsave(&grp->lock, flags); in task_numa_free()
2601 grp->faults[i] -= p->numa_faults[i]; in task_numa_free()
2602 grp->total_faults -= p->total_numa_faults; in task_numa_free()
2604 grp->nr_tasks--; in task_numa_free()
2605 spin_unlock_irqrestore(&grp->lock, flags); in task_numa_free()
2606 RCU_INIT_POINTER(p->numa_group, NULL); in task_numa_free()
2611 p->numa_faults = NULL; in task_numa_free()
2614 p->total_numa_faults = 0; in task_numa_free()
2636 if (!p->mm) in task_numa_fault()
2639 /* Allocate buffer to track faults on a per-node basis */ in task_numa_fault()
2640 if (unlikely(!p->numa_faults)) { in task_numa_fault()
2641 int size = sizeof(*p->numa_faults) * in task_numa_fault()
2644 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); in task_numa_fault()
2645 if (!p->numa_faults) in task_numa_fault()
2648 p->total_numa_faults = 0; in task_numa_fault()
2649 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); in task_numa_fault()
2656 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { in task_numa_fault()
2671 if (!priv && !local && ng && ng->active_nodes > 1 && in task_numa_fault()
2680 if (time_after(jiffies, p->numa_migrate_retry)) { in task_numa_fault()
2686 p->numa_pages_migrated += pages; in task_numa_fault()
2688 p->numa_faults_locality[2] += pages; in task_numa_fault()
2690 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; in task_numa_fault()
2691 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; in task_numa_fault()
2692 p->numa_faults_locality[local] += pages; in task_numa_fault()
2699 * p->mm->numa_scan_seq is written to without exclusive access in reset_ptenuma_scan()
2705 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); in reset_ptenuma_scan()
2706 p->mm->numa_scan_offset = 0; in reset_ptenuma_scan()
2717 struct mm_struct *mm = p->mm; in task_numa_work()
2718 u64 runtime = p->se.sum_exec_runtime; in task_numa_work()
2726 work->next = work; in task_numa_work()
2730 * NOTE: make sure not to dereference p->mm before this check, in task_numa_work()
2732 * without p->mm even though we still had it when we enqueued this in task_numa_work()
2735 if (p->flags & PF_EXITING) in task_numa_work()
2738 if (!mm->numa_next_scan) { in task_numa_work()
2739 mm->numa_next_scan = now + in task_numa_work()
2746 migrate = mm->numa_next_scan; in task_numa_work()
2750 if (p->numa_scan_period == 0) { in task_numa_work()
2751 p->numa_scan_period_max = task_scan_max(p); in task_numa_work()
2752 p->numa_scan_period = task_scan_start(p); in task_numa_work()
2755 next_scan = now + msecs_to_jiffies(p->numa_scan_period); in task_numa_work()
2756 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) in task_numa_work()
2763 p->node_stamp += 2 * TICK_NSEC; in task_numa_work()
2765 start = mm->numa_scan_offset; in task_numa_work()
2767 pages <<= 20 - PAGE_SHIFT; /* MB in pages */ in task_numa_work()
2779 vma = mm->mmap; in task_numa_work()
2781 for (; vma; vma = vma->vm_next) { in task_numa_work()
2783 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { in task_numa_work()
2790 * hinting faults in read-only file-backed mappings or the vdso in task_numa_work()
2793 if (!vma->vm_mm || in task_numa_work()
2794 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) in task_numa_work()
2805 start = max(start, vma->vm_start); in task_numa_work()
2807 end = min(end, vma->vm_end); in task_numa_work()
2813 * is not already pte-numa. If the VMA contains in task_numa_work()
2819 pages -= (end - start) >> PAGE_SHIFT; in task_numa_work()
2820 virtpages -= (end - start) >> PAGE_SHIFT; in task_numa_work()
2827 } while (end != vma->vm_end); in task_numa_work()
2838 mm->numa_scan_offset = start; in task_numa_work()
2849 if (unlikely(p->se.sum_exec_runtime != runtime)) { in task_numa_work()
2850 u64 diff = p->se.sum_exec_runtime - runtime; in task_numa_work()
2851 p->node_stamp += 32 * diff; in task_numa_work()
2858 struct mm_struct *mm = p->mm; in init_numa_balancing()
2861 mm_users = atomic_read(&mm->mm_users); in init_numa_balancing()
2863 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); in init_numa_balancing()
2864 mm->numa_scan_seq = 0; in init_numa_balancing()
2867 p->node_stamp = 0; in init_numa_balancing()
2868 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; in init_numa_balancing()
2869 p->numa_scan_period = sysctl_numa_balancing_scan_delay; in init_numa_balancing()
2871 p->numa_work.next = &p->numa_work; in init_numa_balancing()
2872 p->numa_faults = NULL; in init_numa_balancing()
2873 RCU_INIT_POINTER(p->numa_group, NULL); in init_numa_balancing()
2874 p->last_task_numa_placement = 0; in init_numa_balancing()
2875 p->last_sum_exec_runtime = 0; in init_numa_balancing()
2877 init_task_work(&p->numa_work, task_numa_work); in init_numa_balancing()
2881 p->numa_preferred_nid = NUMA_NO_NODE; in init_numa_balancing()
2893 current->numa_scan_period * mm_users * NSEC_PER_MSEC); in init_numa_balancing()
2895 p->node_stamp = delay; in init_numa_balancing()
2904 struct callback_head *work = &curr->numa_work; in task_tick_numa()
2910 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work) in task_tick_numa()
2919 now = curr->se.sum_exec_runtime; in task_tick_numa()
2920 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; in task_tick_numa()
2922 if (now > curr->node_stamp + period) { in task_tick_numa()
2923 if (!curr->node_stamp) in task_tick_numa()
2924 curr->numa_scan_period = task_scan_start(curr); in task_tick_numa()
2925 curr->node_stamp += period; in task_tick_numa()
2927 if (!time_before(jiffies, curr->mm->numa_next_scan)) in task_tick_numa()
2940 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) in update_scan_period()
2949 * is pulled cross-node due to wakeups or load balancing. in update_scan_period()
2951 if (p->numa_scan_seq) { in update_scan_period()
2957 if (dst_nid == p->numa_preferred_nid || in update_scan_period()
2958 (p->numa_preferred_nid != NUMA_NO_NODE && in update_scan_period()
2959 src_nid != p->numa_preferred_nid)) in update_scan_period()
2963 p->numa_scan_period = task_scan_start(p); in update_scan_period()
2988 update_load_add(&cfs_rq->load, se->load.weight); in account_entity_enqueue()
2994 list_add(&se->group_node, &rq->cfs_tasks); in account_entity_enqueue()
2997 cfs_rq->nr_running++; in account_entity_enqueue()
3003 update_load_sub(&cfs_rq->load, se->load.weight); in account_entity_dequeue()
3007 list_del_init(&se->group_node); in account_entity_dequeue()
3010 cfs_rq->nr_running--; in account_entity_dequeue()
3016 * Explicitly do a load-store to ensure the intermediate value never hits
3036 * Explicitly do a load-store to ensure the intermediate value never hits
3044 res = var - val; \
3053 * A variant of sub_positive(), which does not use explicit load-store
3058 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3065 cfs_rq->avg.load_avg += se->avg.load_avg; in enqueue_load_avg()
3066 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; in enqueue_load_avg()
3072 u32 divider = get_pelt_divider(&se->avg); in dequeue_load_avg()
3073 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); in dequeue_load_avg()
3074 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * divider; in dequeue_load_avg()
3086 if (se->on_rq) { in reweight_entity()
3088 if (cfs_rq->curr == se) in reweight_entity()
3090 update_load_sub(&cfs_rq->load, se->load.weight); in reweight_entity()
3094 update_load_set(&se->load, weight); in reweight_entity()
3098 u32 divider = get_pelt_divider(&se->avg); in reweight_entity()
3100 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); in reweight_entity()
3105 if (se->on_rq) in reweight_entity()
3106 update_load_add(&cfs_rq->load, se->load.weight); in reweight_entity()
3112 struct sched_entity *se = &p->se; in reweight_task()
3114 struct load_weight *load = &se->load; in reweight_task()
3118 load->inv_weight = sched_prio_to_wmult[prio]; in reweight_task()
3130 * tg->weight * grq->load.weight
3131 * ge->load.weight = ----------------------------- (1)
3132 * \Sum grq->load.weight
3140 * grq->load.weight -> grq->avg.load_avg (2)
3144 * tg->weight * grq->avg.load_avg
3145 * ge->load.weight = ------------------------------ (3)
3146 * tg->load_avg
3148 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3152 * The problem with it is that because the average is slow -- it was designed
3153 * to be exactly that of course -- this leads to transients in boundary
3155 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3160 * tg->weight * grq->load.weight
3161 * ge->load.weight = ----------------------------- = tg->weight (4)
3162 * grp->load.weight
3169 * ge->load.weight =
3171 * tg->weight * grq->load.weight
3172 * --------------------------------------------------- (5)
3173 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3175 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3176 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3179 * tg->weight * grq->load.weight
3180 * ge->load.weight = ----------------------------- (6)
3185 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3186 * max(grq->load.weight, grq->avg.load_avg)
3190 * overestimates the ge->load.weight and therefore:
3192 * \Sum ge->load.weight >= tg->weight
3199 struct task_group *tg = cfs_rq->tg; in calc_group_shares()
3201 tg_shares = READ_ONCE(tg->shares); in calc_group_shares()
3203 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); in calc_group_shares()
3205 tg_weight = atomic_long_read(&tg->load_avg); in calc_group_shares()
3208 tg_weight -= cfs_rq->tg_load_avg_contrib; in calc_group_shares()
3216 * MIN_SHARES has to be unscaled here to support per-CPU partitioning in calc_group_shares()
3217 * of a group with small tg->shares value. It is a floor value which is in calc_group_shares()
3221 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 in calc_group_shares()
3222 * on an 8-core system with 8 tasks each runnable on one CPU shares has in calc_group_shares()
3249 shares = READ_ONCE(gcfs_rq->tg->shares); in update_cfs_group()
3251 if (likely(se->load.weight == shares)) in update_cfs_group()
3270 if (&rq->cfs == cfs_rq) { in cfs_rq_util_change()
3280 * As is, the util number is not freq-invariant (we'd have to in cfs_rq_util_change()
3294 * bottom-up, we only have to test whether the cfs_rq before us on the list
3304 if (cfs_rq->on_list) { in child_cfs_rq_on_list()
3305 prev = cfs_rq->leaf_cfs_rq_list.prev; in child_cfs_rq_on_list()
3309 prev = rq->tmp_alone_branch; in child_cfs_rq_on_list()
3314 return (prev_cfs_rq->tg->parent == cfs_rq->tg); in child_cfs_rq_on_list()
3319 if (cfs_rq->load.weight) in cfs_rq_is_decayed()
3322 if (cfs_rq->avg.load_sum) in cfs_rq_is_decayed()
3325 if (cfs_rq->avg.util_sum) in cfs_rq_is_decayed()
3328 if (cfs_rq->avg.runnable_sum) in cfs_rq_is_decayed()
3339 SCHED_WARN_ON(cfs_rq->avg.load_avg || in cfs_rq_is_decayed()
3340 cfs_rq->avg.util_avg || in cfs_rq_is_decayed()
3341 cfs_rq->avg.runnable_avg); in cfs_rq_is_decayed()
3347 * update_tg_load_avg - update the tg's load avg
3350 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3351 * However, because tg->load_avg is a global value there are performance
3362 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; in update_tg_load_avg()
3367 if (cfs_rq->tg == &root_task_group) in update_tg_load_avg()
3370 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { in update_tg_load_avg()
3371 atomic_long_add(delta, &cfs_rq->tg->load_avg); in update_tg_load_avg()
3372 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; in update_tg_load_avg()
3378 * caller only guarantees p->pi_lock is held; no other assumptions,
3379 * including the state of rq->lock, should be made.
3393 * getting what current time is, so simply throw away the out-of-date in set_task_rq_fair()
3397 if (!(se->avg.last_update_time && prev)) in set_task_rq_fair()
3406 p_last_update_time_copy = prev->load_last_update_time_copy; in set_task_rq_fair()
3407 n_last_update_time_copy = next->load_last_update_time_copy; in set_task_rq_fair()
3411 p_last_update_time = prev->avg.last_update_time; in set_task_rq_fair()
3412 n_last_update_time = next->avg.last_update_time; in set_task_rq_fair()
3418 p_last_update_time = prev->avg.last_update_time; in set_task_rq_fair()
3419 n_last_update_time = next->avg.last_update_time; in set_task_rq_fair()
3422 se->avg.last_update_time = n_last_update_time; in set_task_rq_fair()
3431 * ge->avg == grq->avg (1)
3442 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3447 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3451 * ge->avg.runnable_avg == grq->avg.runnable_avg
3455 * ge->load.weight * grq->avg.load_avg
3456 * ge->avg.load_avg = ----------------------------------- (4)
3457 * grq->load.weight
3470 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3481 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3488 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3497 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg; in update_tg_cfs_util()
3505 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. in update_tg_cfs_util()
3508 divider = get_pelt_divider(&cfs_rq->avg); in update_tg_cfs_util()
3511 se->avg.util_avg = gcfs_rq->avg.util_avg; in update_tg_cfs_util()
3512 se->avg.util_sum = se->avg.util_avg * divider; in update_tg_cfs_util()
3515 add_positive(&cfs_rq->avg.util_avg, delta); in update_tg_cfs_util()
3516 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider; in update_tg_cfs_util()
3522 long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg; in update_tg_cfs_runnable()
3530 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. in update_tg_cfs_runnable()
3533 divider = get_pelt_divider(&cfs_rq->avg); in update_tg_cfs_runnable()
3536 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg; in update_tg_cfs_runnable()
3537 se->avg.runnable_sum = se->avg.runnable_avg * divider; in update_tg_cfs_runnable()
3540 add_positive(&cfs_rq->avg.runnable_avg, delta); in update_tg_cfs_runnable()
3541 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider; in update_tg_cfs_runnable()
3547 long delta, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; in update_tg_cfs_load()
3555 gcfs_rq->prop_runnable_sum = 0; in update_tg_cfs_load()
3558 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. in update_tg_cfs_load()
3561 divider = get_pelt_divider(&cfs_rq->avg); in update_tg_cfs_load()
3568 runnable_sum += se->avg.load_sum; in update_tg_cfs_load()
3575 if (scale_load_down(gcfs_rq->load.weight)) { in update_tg_cfs_load()
3576 load_sum = div_s64(gcfs_rq->avg.load_sum, in update_tg_cfs_load()
3577 scale_load_down(gcfs_rq->load.weight)); in update_tg_cfs_load()
3581 runnable_sum = min(se->avg.load_sum, load_sum); in update_tg_cfs_load()
3590 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; in update_tg_cfs_load()
3596 se->avg.load_sum = runnable_sum; in update_tg_cfs_load()
3598 delta = load_avg - se->avg.load_avg; in update_tg_cfs_load()
3602 se->avg.load_avg = load_avg; in update_tg_cfs_load()
3604 add_positive(&cfs_rq->avg.load_avg, delta); in update_tg_cfs_load()
3605 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * divider; in update_tg_cfs_load()
3610 cfs_rq->propagate = 1; in add_tg_cfs_propagate()
3611 cfs_rq->prop_runnable_sum += runnable_sum; in add_tg_cfs_propagate()
3623 if (!gcfs_rq->propagate) in propagate_entity_load_avg()
3626 gcfs_rq->propagate = 0; in propagate_entity_load_avg()
3630 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); in propagate_entity_load_avg()
3654 if (se->avg.load_avg || se->avg.util_avg) in skip_blocked_update()
3661 if (gcfs_rq->propagate) in skip_blocked_update()
3686 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3694 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3698 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3705 struct sched_avg *sa = &cfs_rq->avg; in update_cfs_rq_load_avg()
3708 if (cfs_rq->removed.nr) { in update_cfs_rq_load_avg()
3710 u32 divider = get_pelt_divider(&cfs_rq->avg); in update_cfs_rq_load_avg()
3712 raw_spin_lock(&cfs_rq->removed.lock); in update_cfs_rq_load_avg()
3713 swap(cfs_rq->removed.util_avg, removed_util); in update_cfs_rq_load_avg()
3714 swap(cfs_rq->removed.load_avg, removed_load); in update_cfs_rq_load_avg()
3715 swap(cfs_rq->removed.runnable_avg, removed_runnable); in update_cfs_rq_load_avg()
3716 cfs_rq->removed.nr = 0; in update_cfs_rq_load_avg()
3717 raw_spin_unlock(&cfs_rq->removed.lock); in update_cfs_rq_load_avg()
3720 sub_positive(&sa->load_avg, r); in update_cfs_rq_load_avg()
3721 sa->load_sum = sa->load_avg * divider; in update_cfs_rq_load_avg()
3724 sub_positive(&sa->util_avg, r); in update_cfs_rq_load_avg()
3725 sa->util_sum = sa->util_avg * divider; in update_cfs_rq_load_avg()
3728 sub_positive(&sa->runnable_avg, r); in update_cfs_rq_load_avg()
3729 sa->runnable_sum = sa->runnable_avg * divider; in update_cfs_rq_load_avg()
3736 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT); in update_cfs_rq_load_avg()
3745 cfs_rq->load_last_update_time_copy = sa->last_update_time; in update_cfs_rq_load_avg()
3752 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3757 * cfs_rq->avg.last_update_time being current.
3762 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. in attach_entity_load_avg()
3765 u32 divider = get_pelt_divider(&cfs_rq->avg); in attach_entity_load_avg()
3774 se->avg.last_update_time = cfs_rq->avg.last_update_time; in attach_entity_load_avg()
3775 se->avg.period_contrib = cfs_rq->avg.period_contrib; in attach_entity_load_avg()
3783 se->avg.util_sum = se->avg.util_avg * divider; in attach_entity_load_avg()
3785 se->avg.runnable_sum = se->avg.runnable_avg * divider; in attach_entity_load_avg()
3787 se->avg.load_sum = divider; in attach_entity_load_avg()
3789 se->avg.load_sum = in attach_entity_load_avg()
3790 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se)); in attach_entity_load_avg()
3794 cfs_rq->avg.util_avg += se->avg.util_avg; in attach_entity_load_avg()
3795 cfs_rq->avg.util_sum += se->avg.util_sum; in attach_entity_load_avg()
3796 cfs_rq->avg.runnable_avg += se->avg.runnable_avg; in attach_entity_load_avg()
3797 cfs_rq->avg.runnable_sum += se->avg.runnable_sum; in attach_entity_load_avg()
3799 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); in attach_entity_load_avg()
3807 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3812 * cfs_rq->avg.last_update_time being current.
3817 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. in detach_entity_load_avg()
3820 u32 divider = get_pelt_divider(&cfs_rq->avg); in detach_entity_load_avg()
3823 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); in detach_entity_load_avg()
3824 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider; in detach_entity_load_avg()
3825 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg); in detach_entity_load_avg()
3826 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider; in detach_entity_load_avg()
3828 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); in detach_entity_load_avg()
3852 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) in update_load_avg()
3858 if (!se->avg.last_update_time && (flags & DO_ATTACH)) { in update_load_avg()
3885 last_update_time_copy = cfs_rq->load_last_update_time_copy; in cfs_rq_last_update_time()
3887 last_update_time = cfs_rq->avg.last_update_time; in cfs_rq_last_update_time()
3895 return cfs_rq->avg.last_update_time; in cfs_rq_last_update_time()
3922 * tasks cannot exit without having gone through wake_up_new_task() -> in remove_entity_load_avg()
3929 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); in remove_entity_load_avg()
3930 ++cfs_rq->removed.nr; in remove_entity_load_avg()
3931 cfs_rq->removed.util_avg += se->avg.util_avg; in remove_entity_load_avg()
3932 cfs_rq->removed.load_avg += se->avg.load_avg; in remove_entity_load_avg()
3933 cfs_rq->removed.runnable_avg += se->avg.runnable_avg; in remove_entity_load_avg()
3934 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); in remove_entity_load_avg()
3939 return cfs_rq->avg.runnable_avg; in cfs_rq_runnable_avg()
3944 return cfs_rq->avg.load_avg; in cfs_rq_load_avg()
3951 return READ_ONCE(p->se.avg.util_avg); in task_util()
3956 struct util_est ue = READ_ONCE(p->se.avg.util_est); in _task_util_est()
3989 enqueued = cfs_rq->avg.util_est.enqueued; in util_est_enqueue()
3991 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); in util_est_enqueue()
4005 enqueued = cfs_rq->avg.util_est.enqueued; in util_est_dequeue()
4006 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p)); in util_est_dequeue()
4007 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); in util_est_dequeue()
4018 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4024 return ((unsigned int)(value + margin - 1) < (2 * margin - 1)); in within_margin()
4048 ue = p->se.avg.util_est; in util_est_update()
4070 last_ewma_diff = ue.enqueued - ue.ewma; in util_est_update()
4071 last_enqueued_diff -= ue.enqueued; in util_est_update()
4094 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) in util_est_update()
4095 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) in util_est_update()
4096 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) in util_est_update()
4097 * = w * ( last_ewma_diff ) + ewma(t-1) in util_est_update()
4098 * = w * (last_ewma_diff + ewma(t-1) / w) in util_est_update()
4108 WRITE_ONCE(p->se.avg.util_est, ue); in util_est_update()
4110 trace_sched_util_est_se_tp(&p->se); in util_est_update()
4123 if (!p || p->nr_cpus_allowed == 1) { in update_misfit_status()
4124 rq->misfit_task_load = 0; in update_misfit_status()
4129 rq->misfit_task_load = 0; in update_misfit_status()
4137 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1); in update_misfit_status()
4184 s64 d = se->vruntime - cfs_rq->min_vruntime; in check_spread()
4187 d = -d; in check_spread()
4190 schedstat_inc(cfs_rq->nr_spread_over); in check_spread()
4197 u64 vruntime = cfs_rq->min_vruntime; in place_entity()
4219 vruntime -= thresh; in place_entity()
4223 se->vruntime = max_vruntime(se->vruntime, vruntime); in place_entity()
4256 * vruntime -= min_vruntime
4264 * min_vruntime are up-to-date.
4268 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4269 * vruntime -= min_vruntime
4276 * this way we don't have the most up-to-date min_vruntime on the originating
4277 * CPU and an up-to-date min_vruntime on the destination CPU.
4284 bool curr = cfs_rq->curr == se; in enqueue_entity()
4291 se->vruntime += cfs_rq->min_vruntime; in enqueue_entity()
4302 se->vruntime += cfs_rq->min_vruntime; in enqueue_entity()
4306 * - Update loads to have both entity and cfs_rq synced with now. in enqueue_entity()
4307 * - Add its load to cfs_rq->runnable_avg in enqueue_entity()
4308 * - For group_entity, update its weight to reflect the new share of in enqueue_entity()
4310 * - Add its new weight to cfs_rq->load.weight in enqueue_entity()
4325 se->on_rq = 1; in enqueue_entity()
4329 * because of a parent been throttled but cfs->nr_running > 1. Try to in enqueue_entity()
4332 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used()) in enqueue_entity()
4335 if (cfs_rq->nr_running == 1) in enqueue_entity()
4343 if (cfs_rq->last != se) in __clear_buddies_last()
4346 cfs_rq->last = NULL; in __clear_buddies_last()
4354 if (cfs_rq->next != se) in __clear_buddies_next()
4357 cfs_rq->next = NULL; in __clear_buddies_next()
4365 if (cfs_rq->skip != se) in __clear_buddies_skip()
4368 cfs_rq->skip = NULL; in __clear_buddies_skip()
4374 if (cfs_rq->last == se) in clear_buddies()
4377 if (cfs_rq->next == se) in clear_buddies()
4380 if (cfs_rq->skip == se) in clear_buddies()
4390 * Update run-time statistics of the 'current'. in dequeue_entity()
4396 * - Update loads to have both entity and cfs_rq synced with now. in dequeue_entity()
4397 * - Subtract its load from the cfs_rq->runnable_avg. in dequeue_entity()
4398 * - Subtract its previous weight from cfs_rq->load.weight. in dequeue_entity()
4399 * - For group entity, update its weight to reflect the new share in dequeue_entity()
4409 if (se != cfs_rq->curr) in dequeue_entity()
4411 se->on_rq = 0; in dequeue_entity()
4421 se->vruntime -= cfs_rq->min_vruntime; in dequeue_entity()
4432 * further than we started -- ie. we'll be penalized. in dequeue_entity()
4449 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; in check_preempt_tick()
4454 * re-elected due to buddy favours. in check_preempt_tick()
4469 delta = curr->vruntime - se->vruntime; in check_preempt_tick()
4484 if (se->on_rq) { in set_next_entity()
4496 cfs_rq->curr = se; in set_next_entity()
4501 * when there are only lesser-weight tasks around): in set_next_entity()
4504 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { in set_next_entity()
4505 schedstat_set(se->statistics.slice_max, in set_next_entity()
4506 max((u64)schedstat_val(se->statistics.slice_max), in set_next_entity()
4507 se->sum_exec_runtime - se->prev_sum_exec_runtime)); in set_next_entity()
4510 se->prev_sum_exec_runtime = se->sum_exec_runtime; in set_next_entity()
4542 if (cfs_rq->skip && cfs_rq->skip == se) { in pick_next_entity()
4557 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) { in pick_next_entity()
4561 se = cfs_rq->next; in pick_next_entity()
4562 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) { in pick_next_entity()
4566 se = cfs_rq->last; in pick_next_entity()
4580 if (prev->on_rq) in put_prev_entity()
4583 /* throttle cfs_rqs exceeding runtime */ in put_prev_entity()
4588 if (prev->on_rq) { in put_prev_entity()
4595 cfs_rq->curr = NULL; in put_prev_entity()
4602 * Update run-time statistics of the 'current'. in entity_tick()
4625 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) in entity_tick()
4629 if (cfs_rq->nr_running > 1) in entity_tick()
4683 * directly instead of rq->clock to avoid adding additional synchronization
4684 * around rq->lock.
4686 * requires cfs_b->lock
4690 if (unlikely(cfs_b->quota == RUNTIME_INF)) in __refill_cfs_bandwidth_runtime()
4693 cfs_b->runtime += cfs_b->quota; in __refill_cfs_bandwidth_runtime()
4694 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst); in __refill_cfs_bandwidth_runtime()
4699 return &tg->cfs_bandwidth; in tg_cfs_bandwidth()
4708 lockdep_assert_held(&cfs_b->lock); in __assign_cfs_rq_runtime()
4711 min_amount = target_runtime - cfs_rq->runtime_remaining; in __assign_cfs_rq_runtime()
4713 if (cfs_b->quota == RUNTIME_INF) in __assign_cfs_rq_runtime()
4718 if (cfs_b->runtime > 0) { in __assign_cfs_rq_runtime()
4719 amount = min(cfs_b->runtime, min_amount); in __assign_cfs_rq_runtime()
4720 cfs_b->runtime -= amount; in __assign_cfs_rq_runtime()
4721 cfs_b->idle = 0; in __assign_cfs_rq_runtime()
4725 cfs_rq->runtime_remaining += amount; in __assign_cfs_rq_runtime()
4727 return cfs_rq->runtime_remaining > 0; in __assign_cfs_rq_runtime()
4733 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); in assign_cfs_rq_runtime()
4736 raw_spin_lock(&cfs_b->lock); in assign_cfs_rq_runtime()
4738 raw_spin_unlock(&cfs_b->lock); in assign_cfs_rq_runtime()
4746 cfs_rq->runtime_remaining -= delta_exec; in __account_cfs_rq_runtime()
4748 if (likely(cfs_rq->runtime_remaining > 0)) in __account_cfs_rq_runtime()
4751 if (cfs_rq->throttled) in __account_cfs_rq_runtime()
4757 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) in __account_cfs_rq_runtime()
4764 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) in account_cfs_rq_runtime()
4772 return cfs_bandwidth_used() && cfs_rq->throttled; in cfs_rq_throttled()
4778 return cfs_bandwidth_used() && cfs_rq->throttle_count; in throttled_hierarchy()
4784 * load-balance operations.
4791 src_cfs_rq = tg->cfs_rq[src_cpu]; in throttled_lb_pair()
4792 dest_cfs_rq = tg->cfs_rq[dest_cpu]; in throttled_lb_pair()
4801 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; in tg_unthrottle_up()
4803 cfs_rq->throttle_count--; in tg_unthrottle_up()
4804 if (!cfs_rq->throttle_count) { in tg_unthrottle_up()
4805 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - in tg_unthrottle_up()
4806 cfs_rq->throttled_clock_task; in tg_unthrottle_up()
4809 if (!cfs_rq_is_decayed(cfs_rq) || cfs_rq->nr_running) in tg_unthrottle_up()
4819 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; in tg_throttle_down()
4822 if (!cfs_rq->throttle_count) { in tg_throttle_down()
4823 cfs_rq->throttled_clock_task = rq_clock_task(rq); in tg_throttle_down()
4826 cfs_rq->throttle_count++; in tg_throttle_down()
4834 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); in throttle_cfs_rq()
4838 raw_spin_lock(&cfs_b->lock); in throttle_cfs_rq()
4845 * subsequent check_cfs_rq_runtime calls agree not to throttle in throttle_cfs_rq()
4851 list_add_tail_rcu(&cfs_rq->throttled_list, in throttle_cfs_rq()
4852 &cfs_b->throttled_cfs_rq); in throttle_cfs_rq()
4854 raw_spin_unlock(&cfs_b->lock); in throttle_cfs_rq()
4857 return false; /* Throttle no longer required. */ in throttle_cfs_rq()
4859 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; in throttle_cfs_rq()
4863 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); in throttle_cfs_rq()
4866 task_delta = cfs_rq->h_nr_running; in throttle_cfs_rq()
4867 idle_task_delta = cfs_rq->idle_h_nr_running; in throttle_cfs_rq()
4870 /* throttled entity or throttle-on-deactivate */ in throttle_cfs_rq()
4871 if (!se->on_rq) in throttle_cfs_rq()
4877 idle_task_delta = cfs_rq->h_nr_running; in throttle_cfs_rq()
4879 qcfs_rq->h_nr_running -= task_delta; in throttle_cfs_rq()
4880 qcfs_rq->idle_h_nr_running -= idle_task_delta; in throttle_cfs_rq()
4882 if (qcfs_rq->load.weight) { in throttle_cfs_rq()
4883 /* Avoid re-evaluating load for this entity: */ in throttle_cfs_rq()
4891 /* throttled entity or throttle-on-deactivate */ in throttle_cfs_rq()
4892 if (!se->on_rq) in throttle_cfs_rq()
4899 idle_task_delta = cfs_rq->h_nr_running; in throttle_cfs_rq()
4901 qcfs_rq->h_nr_running -= task_delta; in throttle_cfs_rq()
4902 qcfs_rq->idle_h_nr_running -= idle_task_delta; in throttle_cfs_rq()
4911 * throttled-list. rq->lock protects completion. in throttle_cfs_rq()
4913 cfs_rq->throttled = 1; in throttle_cfs_rq()
4914 cfs_rq->throttled_clock = rq_clock(rq); in throttle_cfs_rq()
4921 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); in unthrottle_cfs_rq()
4925 se = cfs_rq->tg->se[cpu_of(rq)]; in unthrottle_cfs_rq()
4927 cfs_rq->throttled = 0; in unthrottle_cfs_rq()
4931 raw_spin_lock(&cfs_b->lock); in unthrottle_cfs_rq()
4932 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; in unthrottle_cfs_rq()
4933 list_del_rcu(&cfs_rq->throttled_list); in unthrottle_cfs_rq()
4934 raw_spin_unlock(&cfs_b->lock); in unthrottle_cfs_rq()
4936 /* update hierarchical throttle state */ in unthrottle_cfs_rq()
4937 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); in unthrottle_cfs_rq()
4940 if (!cfs_rq->load.weight) { in unthrottle_cfs_rq()
4941 if (cfs_rq->on_list) in unthrottle_cfs_rq()
4946 task_delta = cfs_rq->h_nr_running; in unthrottle_cfs_rq()
4947 idle_task_delta = cfs_rq->idle_h_nr_running; in unthrottle_cfs_rq()
4951 if (se->on_rq) in unthrottle_cfs_rq()
4956 idle_task_delta = cfs_rq->h_nr_running; in unthrottle_cfs_rq()
4958 qcfs_rq->h_nr_running += task_delta; in unthrottle_cfs_rq()
4959 qcfs_rq->idle_h_nr_running += idle_task_delta; in unthrottle_cfs_rq()
4973 idle_task_delta = cfs_rq->h_nr_running; in unthrottle_cfs_rq()
4975 qcfs_rq->h_nr_running += task_delta; in unthrottle_cfs_rq()
4976 qcfs_rq->idle_h_nr_running += idle_task_delta; in unthrottle_cfs_rq()
5009 if (rq->curr == rq->idle && rq->cfs.nr_running) in unthrottle_cfs_rq()
5019 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, in distribute_cfs_runtime()
5029 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0); in distribute_cfs_runtime()
5031 raw_spin_lock(&cfs_b->lock); in distribute_cfs_runtime()
5032 runtime = -cfs_rq->runtime_remaining + 1; in distribute_cfs_runtime()
5033 if (runtime > cfs_b->runtime) in distribute_cfs_runtime()
5034 runtime = cfs_b->runtime; in distribute_cfs_runtime()
5035 cfs_b->runtime -= runtime; in distribute_cfs_runtime()
5036 remaining = cfs_b->runtime; in distribute_cfs_runtime()
5037 raw_spin_unlock(&cfs_b->lock); in distribute_cfs_runtime()
5039 cfs_rq->runtime_remaining += runtime; in distribute_cfs_runtime()
5042 if (cfs_rq->runtime_remaining > 0) in distribute_cfs_runtime()
5057 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5060 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) in do_sched_cfs_period_timer() argument
5065 if (cfs_b->quota == RUNTIME_INF) in do_sched_cfs_period_timer()
5068 throttled = !list_empty(&cfs_b->throttled_cfs_rq); in do_sched_cfs_period_timer()
5069 cfs_b->nr_periods += overrun; in do_sched_cfs_period_timer()
5071 /* Refill extra burst quota even if cfs_b->idle */ in do_sched_cfs_period_timer()
5078 if (cfs_b->idle && !throttled) in do_sched_cfs_period_timer()
5083 cfs_b->idle = 1; in do_sched_cfs_period_timer()
5088 cfs_b->nr_throttled += overrun; in do_sched_cfs_period_timer()
5091 * This check is repeated as we release cfs_b->lock while we unthrottle. in do_sched_cfs_period_timer()
5093 while (throttled && cfs_b->runtime > 0) { in do_sched_cfs_period_timer()
5094 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); in do_sched_cfs_period_timer()
5095 /* we can't nest cfs_b->lock while distributing bandwidth */ in do_sched_cfs_period_timer()
5097 raw_spin_lock_irqsave(&cfs_b->lock, flags); in do_sched_cfs_period_timer()
5099 throttled = !list_empty(&cfs_b->throttled_cfs_rq); in do_sched_cfs_period_timer()
5108 cfs_b->idle = 0; in do_sched_cfs_period_timer()
5126 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5132 struct hrtimer *refresh_timer = &cfs_b->period_timer; in runtime_refresh_within()
5135 /* if the call-back is running a quota refresh is already occurring */ in runtime_refresh_within()
5156 if (cfs_b->slack_started) in start_cfs_slack_bandwidth()
5158 cfs_b->slack_started = true; in start_cfs_slack_bandwidth()
5160 hrtimer_start(&cfs_b->slack_timer, in start_cfs_slack_bandwidth()
5168 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); in __return_cfs_rq_runtime()
5169 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; in __return_cfs_rq_runtime()
5174 raw_spin_lock(&cfs_b->lock); in __return_cfs_rq_runtime()
5175 if (cfs_b->quota != RUNTIME_INF) { in __return_cfs_rq_runtime()
5176 cfs_b->runtime += slack_runtime; in __return_cfs_rq_runtime()
5178 /* we are under rq->lock, defer unthrottling using a timer */ in __return_cfs_rq_runtime()
5179 if (cfs_b->runtime > sched_cfs_bandwidth_slice() && in __return_cfs_rq_runtime()
5180 !list_empty(&cfs_b->throttled_cfs_rq)) in __return_cfs_rq_runtime()
5183 raw_spin_unlock(&cfs_b->lock); in __return_cfs_rq_runtime()
5186 cfs_rq->runtime_remaining -= slack_runtime; in __return_cfs_rq_runtime()
5194 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) in return_cfs_rq_runtime()
5202 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5210 raw_spin_lock_irqsave(&cfs_b->lock, flags); in do_sched_cfs_slack_timer()
5211 cfs_b->slack_started = false; in do_sched_cfs_slack_timer()
5214 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); in do_sched_cfs_slack_timer()
5218 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) in do_sched_cfs_slack_timer()
5219 runtime = cfs_b->runtime; in do_sched_cfs_slack_timer()
5221 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); in do_sched_cfs_slack_timer()
5232 * runtime as update_curr() throttling can not trigger until it's on-rq.
5239 /* an active group must be handled by the update_curr()->put() path */ in check_enqueue_throttle()
5240 if (!cfs_rq->runtime_enabled || cfs_rq->curr) in check_enqueue_throttle()
5249 if (cfs_rq->runtime_remaining <= 0) in check_enqueue_throttle()
5260 if (!tg->parent) in sync_throttle()
5263 cfs_rq = tg->cfs_rq[cpu]; in sync_throttle()
5264 pcfs_rq = tg->parent->cfs_rq[cpu]; in sync_throttle()
5266 cfs_rq->throttle_count = pcfs_rq->throttle_count; in sync_throttle()
5267 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); in sync_throttle()
5270 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5276 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) in check_cfs_rq_runtime()
5306 int overrun; in sched_cfs_period_timer() local
5310 raw_spin_lock_irqsave(&cfs_b->lock, flags); in sched_cfs_period_timer()
5312 overrun = hrtimer_forward_now(timer, cfs_b->period); in sched_cfs_period_timer()
5313 if (!overrun) in sched_cfs_period_timer()
5316 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); in sched_cfs_period_timer()
5319 u64 new, old = ktime_to_ns(cfs_b->period); in sched_cfs_period_timer()
5328 cfs_b->period = ns_to_ktime(new); in sched_cfs_period_timer()
5329 cfs_b->quota *= 2; in sched_cfs_period_timer()
5330 cfs_b->burst *= 2; in sched_cfs_period_timer()
5336 div_u64(cfs_b->quota, NSEC_PER_USEC)); in sched_cfs_period_timer()
5342 div_u64(cfs_b->quota, NSEC_PER_USEC)); in sched_cfs_period_timer()
5350 cfs_b->period_active = 0; in sched_cfs_period_timer()
5351 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); in sched_cfs_period_timer()
5358 raw_spin_lock_init(&cfs_b->lock); in init_cfs_bandwidth()
5359 cfs_b->runtime = 0; in init_cfs_bandwidth()
5360 cfs_b->quota = RUNTIME_INF; in init_cfs_bandwidth()
5361 cfs_b->period = ns_to_ktime(default_cfs_period()); in init_cfs_bandwidth()
5362 cfs_b->burst = 0; in init_cfs_bandwidth()
5364 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); in init_cfs_bandwidth()
5365 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); in init_cfs_bandwidth()
5366 cfs_b->period_timer.function = sched_cfs_period_timer; in init_cfs_bandwidth()
5367 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); in init_cfs_bandwidth()
5368 cfs_b->slack_timer.function = sched_cfs_slack_timer; in init_cfs_bandwidth()
5369 cfs_b->slack_started = false; in init_cfs_bandwidth()
5374 cfs_rq->runtime_enabled = 0; in init_cfs_rq_runtime()
5375 INIT_LIST_HEAD(&cfs_rq->throttled_list); in init_cfs_rq_runtime()
5380 lockdep_assert_held(&cfs_b->lock); in start_cfs_bandwidth()
5382 if (cfs_b->period_active) in start_cfs_bandwidth()
5385 cfs_b->period_active = 1; in start_cfs_bandwidth()
5386 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); in start_cfs_bandwidth()
5387 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); in start_cfs_bandwidth()
5393 if (!cfs_b->throttled_cfs_rq.next) in destroy_cfs_bandwidth()
5396 hrtimer_cancel(&cfs_b->period_timer); in destroy_cfs_bandwidth()
5397 hrtimer_cancel(&cfs_b->slack_timer); in destroy_cfs_bandwidth()
5416 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; in update_runtime_enabled()
5417 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; in update_runtime_enabled()
5419 raw_spin_lock(&cfs_b->lock); in update_runtime_enabled()
5420 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; in update_runtime_enabled()
5421 raw_spin_unlock(&cfs_b->lock); in update_runtime_enabled()
5435 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; in unthrottle_offline_cfs_rqs()
5437 if (!cfs_rq->runtime_enabled) in unthrottle_offline_cfs_rqs()
5444 cfs_rq->runtime_remaining = 1; in unthrottle_offline_cfs_rqs()
5449 cfs_rq->runtime_enabled = 0; in unthrottle_offline_cfs_rqs()
5509 struct sched_entity *se = &p->se; in hrtick_start_fair()
5514 if (rq->cfs.h_nr_running > 1) { in hrtick_start_fair()
5516 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; in hrtick_start_fair()
5517 s64 delta = slice - ran; in hrtick_start_fair()
5535 struct task_struct *curr = rq->curr; in hrtick_update()
5537 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class) in hrtick_update()
5540 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) in hrtick_update()
5564 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) { in update_overutilized_status()
5565 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED); in update_overutilized_status()
5566 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED); in update_overutilized_status()
5576 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running && in sched_idle_rq()
5577 rq->nr_running); in sched_idle_rq()
5596 struct sched_entity *se = &p->se; in enqueue_task_fair()
5606 util_est_enqueue(&rq->cfs, p); in enqueue_task_fair()
5613 if (p->in_iowait) in enqueue_task_fair()
5617 if (se->on_rq) in enqueue_task_fair()
5622 cfs_rq->h_nr_running++; in enqueue_task_fair()
5623 cfs_rq->idle_h_nr_running += idle_h_nr_running; in enqueue_task_fair()
5642 cfs_rq->h_nr_running++; in enqueue_task_fair()
5643 cfs_rq->idle_h_nr_running += idle_h_nr_running; in enqueue_task_fair()
5711 struct sched_entity *se = &p->se; in dequeue_task_fair()
5716 util_est_dequeue(&rq->cfs, p); in dequeue_task_fair()
5722 cfs_rq->h_nr_running--; in dequeue_task_fair()
5723 cfs_rq->idle_h_nr_running -= idle_h_nr_running; in dequeue_task_fair()
5733 if (cfs_rq->load.weight) { in dequeue_task_fair()
5734 /* Avoid re-evaluating load for this entity: */ in dequeue_task_fair()
5754 cfs_rq->h_nr_running--; in dequeue_task_fair()
5755 cfs_rq->idle_h_nr_running -= idle_h_nr_running; in dequeue_task_fair()
5771 rq->next_balance = jiffies; in dequeue_task_fair()
5774 util_est_update(&rq->cfs, p, task_sleep); in dequeue_task_fair()
5798 return cfs_rq_load_avg(&rq->cfs); in cpu_load()
5802 * cpu_load_without - compute CPU load without any contributions from *p
5820 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) in cpu_load_without()
5823 cfs_rq = &rq->cfs; in cpu_load_without()
5824 load = READ_ONCE(cfs_rq->avg.load_avg); in cpu_load_without()
5834 return cfs_rq_runnable_avg(&rq->cfs); in cpu_runnable()
5843 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) in cpu_runnable_without()
5846 cfs_rq = &rq->cfs; in cpu_runnable_without()
5847 runnable = READ_ONCE(cfs_rq->avg.runnable_avg); in cpu_runnable_without()
5850 lsub_positive(&runnable, p->se.avg.runnable_avg); in cpu_runnable_without()
5857 return cpu_rq(cpu)->cpu_capacity; in capacity_of()
5866 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { in record_wakee()
5867 current->wakee_flips >>= 1; in record_wakee()
5868 current->wakee_flip_decay_ts = jiffies; in record_wakee()
5871 if (current->last_wakee != p) { in record_wakee()
5872 current->last_wakee = p; in record_wakee()
5873 current->wakee_flips++; in record_wakee()
5878 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5888 * non-monogamous, with partner count exceeding socket size.
5896 unsigned int master = current->wakee_flips; in wake_wide()
5897 unsigned int slave = p->wakee_flips; in wake_wide()
5912 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5913 * cache-affine and is (or will be) idle.
5915 * wake_affine_weight() - considers the weight to reflect the average
5937 if (sync && cpu_rq(this_cpu)->nr_running == 1) in wake_affine_idle()
5961 this_eff_load -= current_load; in wake_affine_weight()
5972 prev_eff_load -= task_load; in wake_affine_weight()
5974 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; in wake_affine_weight()
6000 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); in wake_affine()
6004 schedstat_inc(sd->ttwu_move_affine); in wake_affine()
6005 schedstat_inc(p->se.statistics.nr_wakeups_affine); in wake_affine()
6013 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6022 int shallowest_idle_cpu = -1; in find_idlest_group_cpu()
6026 if (group->group_weight == 1) in find_idlest_group_cpu()
6030 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { in find_idlest_group_cpu()
6041 if (idle && idle->exit_latency < min_exit_latency) { in find_idlest_group_cpu()
6047 min_exit_latency = idle->exit_latency; in find_idlest_group_cpu()
6048 latest_idle_timestamp = rq->idle_stamp; in find_idlest_group_cpu()
6050 } else if ((!idle || idle->exit_latency == min_exit_latency) && in find_idlest_group_cpu()
6051 rq->idle_stamp > latest_idle_timestamp) { in find_idlest_group_cpu()
6057 latest_idle_timestamp = rq->idle_stamp; in find_idlest_group_cpu()
6060 } else if (shallowest_idle_cpu == -1) { in find_idlest_group_cpu()
6069 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; in find_idlest_group_cpu()
6077 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) in find_idlest_cpu()
6085 sync_entity_load_avg(&p->se); in find_idlest_cpu()
6092 if (!(sd->flags & sd_flag)) { in find_idlest_cpu()
6093 sd = sd->child; in find_idlest_cpu()
6099 sd = sd->child; in find_idlest_cpu()
6106 sd = sd->child; in find_idlest_cpu()
6112 weight = sd->span_weight; in find_idlest_cpu()
6115 if (weight <= tmp->span_weight) in find_idlest_cpu()
6117 if (tmp->flags & sd_flag) in find_idlest_cpu()
6131 return -1; in __select_idle_cpu()
6144 WRITE_ONCE(sds->has_idle_cores, val); in set_idle_cores()
6153 return READ_ONCE(sds->has_idle_cores); in test_idle_cores()
6160 * information in sd_llc_shared->has_idle_cores.
6190 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6203 if (*idle_cpu == -1) { in select_idle_core()
6204 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) { in select_idle_core()
6212 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr)) in select_idle_core()
6220 return -1; in select_idle_core()
6231 if (!cpumask_test_cpu(cpu, p->cpus_ptr) || in select_idle_smt()
6238 return -1; in select_idle_smt()
6259 return -1; in select_idle_smt()
6266 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6267 * average idle time for this rq (as found in rq->avg_idle).
6272 int i, cpu, idle_cpu = -1, nr = INT_MAX; in select_idle_cpu()
6280 return -1; in select_idle_cpu()
6282 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); in select_idle_cpu()
6293 if (unlikely(this_rq->wake_stamp < now)) { in select_idle_cpu()
6294 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) { in select_idle_cpu()
6295 this_rq->wake_stamp++; in select_idle_cpu()
6296 this_rq->wake_avg_idle >>= 1; in select_idle_cpu()
6300 avg_idle = this_rq->wake_avg_idle; in select_idle_cpu()
6301 avg_cost = this_sd->avg_scan_cost + 1; in select_idle_cpu()
6303 span_avg = sd->span_weight * avg_idle; in select_idle_cpu()
6319 if (!--nr) in select_idle_cpu()
6320 return -1; in select_idle_cpu()
6331 time = cpu_clock(this) - time; in select_idle_cpu()
6337 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time); in select_idle_cpu()
6339 update_avg(&this_sd->avg_scan_cost, time); in select_idle_cpu()
6354 int cpu, best_cpu = -1; in select_idle_capacity()
6358 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); in select_idle_capacity()
6402 sync_entity_load_avg(&p->se); in select_idle_sibling()
6407 * per-cpu select_idle_mask usage in select_idle_sibling()
6424 * Allow a per-cpu kthread to stack with the wakee if the in select_idle_sibling()
6427 * per-cpu kthread that is now complete and the wakeup is in select_idle_sibling()
6433 this_rq()->nr_running <= 1) { in select_idle_sibling()
6438 recent_used_cpu = p->recent_used_cpu; in select_idle_sibling()
6439 p->recent_used_cpu = prev; in select_idle_sibling()
6444 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) && in select_idle_sibling()
6450 p->recent_used_cpu = prev; in select_idle_sibling()
6496 * cpu_util - Estimates the amount of capacity of a CPU used by CFS tasks.
6504 * recent utilization of currently non-runnable tasks on a CPU. It represents
6529 * migrations (scheduler-driven DVFS).
6538 cfs_rq = &cpu_rq(cpu)->cfs; in cpu_util()
6539 util = READ_ONCE(cfs_rq->avg.util_avg); in cpu_util()
6542 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued)); in cpu_util()
6566 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) in cpu_util_without()
6569 cfs_rq = &cpu_rq(cpu)->cfs; in cpu_util_without()
6570 util = READ_ONCE(cfs_rq->avg.util_avg); in cpu_util_without()
6581 * cpu_util_without = (cpu_util - task_util) = 0 in cpu_util_without()
6588 * cpu_util_without = (cpu_util - task_util) >= 0 in cpu_util_without()
6603 READ_ONCE(cfs_rq->avg.util_est.enqueued); in cpu_util_without()
6611 * p->on_rq = TASK_ON_RQ_MIGRATING; in cpu_util_without()
6612 * ---------------------------------- A in cpu_util_without()
6616 * ---------------------------------- B in cpu_util_without()
6642 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; in cpu_util_next()
6643 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg); in cpu_util_next()
6657 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued); in cpu_util_next()
6660 * During wake-up, the task isn't enqueued yet and doesn't in cpu_util_next()
6661 * appear in the cfs_rq->avg.util_est.enqueued of any rq, in cpu_util_next()
6690 _cpu_cap -= arch_scale_thermal_pressure(cpumask_first(pd_mask)); in compute_energy()
6718 cpu_util_next(cpu, p, -1) + task_util_est(p); in compute_energy()
6744 return em_cpu_energy(pd->em_pd, max_util, sum_util, _cpu_cap); in compute_energy()
6748 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6752 * out which of the CPU candidates is the most energy-efficient.
6769 * cluster-packing, and spreading inside a cluster. That should at least be
6776 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6780 * to be energy-inefficient in some use-cases. The alternative would be to
6783 * other use-cases too. So, until someone finds a better way to solve this,
6784 * let's keep things simple by re-using the existing slow path.
6789 struct root_domain *rd = cpu_rq(smp_processor_id())->rd; in find_energy_efficient_cpu()
6790 int cpu, best_energy_cpu = prev_cpu, target = -1; in find_energy_efficient_cpu()
6796 pd = rcu_dereference(rd->pd); in find_energy_efficient_cpu()
6797 if (!pd || READ_ONCE(rd->overutilized)) in find_energy_efficient_cpu()
6801 * Energy-aware wake-up happens on the lowest sched_domain starting in find_energy_efficient_cpu()
6806 sd = sd->parent; in find_energy_efficient_cpu()
6812 sync_entity_load_avg(&p->se); in find_energy_efficient_cpu()
6816 for (; pd; pd = pd->next) { in find_energy_efficient_cpu()
6820 int max_spare_cap_cpu = -1; in find_energy_efficient_cpu()
6823 if (!cpumask_test_cpu(cpu, p->cpus_ptr)) in find_energy_efficient_cpu()
6859 base_energy_pd = compute_energy(p, -1, pd); in find_energy_efficient_cpu()
6867 prev_delta -= base_energy_pd; in find_energy_efficient_cpu()
6876 cur_delta -= base_energy_pd; in find_energy_efficient_cpu()
6890 (prev_delta - best_delta) > ((prev_delta + base_energy) >> 4)) in find_energy_efficient_cpu()
6914 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); in select_task_rq_fair()
6923 * required for stable ->cpus_allowed in select_task_rq_fair()
6925 lockdep_assert_held(&p->pi_lock); in select_task_rq_fair()
6936 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr); in select_task_rq_fair()
6945 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && in select_task_rq_fair()
6954 if (tmp->flags & sd_flag) in select_task_rq_fair()
6977 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6984 * min_vruntime -- the latter is done by enqueue_entity() when placing in migrate_task_rq_fair()
6987 if (READ_ONCE(p->__state) == TASK_WAKING) { in migrate_task_rq_fair()
6988 struct sched_entity *se = &p->se; in migrate_task_rq_fair()
6996 min_vruntime_copy = cfs_rq->min_vruntime_copy; in migrate_task_rq_fair()
6998 min_vruntime = cfs_rq->min_vruntime; in migrate_task_rq_fair()
7001 min_vruntime = cfs_rq->min_vruntime; in migrate_task_rq_fair()
7004 se->vruntime -= min_vruntime; in migrate_task_rq_fair()
7007 if (p->on_rq == TASK_ON_RQ_MIGRATING) { in migrate_task_rq_fair()
7010 * rq->lock and can modify state directly. in migrate_task_rq_fair()
7013 detach_entity_cfs_rq(&p->se); in migrate_task_rq_fair()
7020 * throw away the out-of-date time. This will result in the in migrate_task_rq_fair()
7024 remove_entity_load_avg(&p->se); in migrate_task_rq_fair()
7028 p->se.avg.last_update_time = 0; in migrate_task_rq_fair()
7031 p->se.exec_start = 0; in migrate_task_rq_fair()
7038 remove_entity_load_avg(&p->se); in task_dead_fair()
7044 if (rq->nr_running) in balance_fair()
7056 * Since its curr running now, convert the gran from real-time in wakeup_gran()
7057 * to virtual-time in his units. in wakeup_gran()
7078 * |<--->|c
7080 * w(c, s1) = -1
7088 s64 gran, vdiff = curr->vruntime - se->vruntime; in wakeup_preempt_entity()
7091 return -1; in wakeup_preempt_entity()
7103 if (SCHED_WARN_ON(!se->on_rq)) in set_last_buddy()
7107 cfs_rq_of(se)->last = se; in set_last_buddy()
7114 if (SCHED_WARN_ON(!se->on_rq)) in set_next_buddy()
7118 cfs_rq_of(se)->next = se; in set_next_buddy()
7125 cfs_rq_of(se)->skip = se; in set_skip_buddy()
7133 struct task_struct *curr = rq->curr; in check_preempt_wakeup()
7134 struct sched_entity *se = &curr->se, *pse = &p->se; in check_preempt_wakeup()
7136 int scale = cfs_rq->nr_running >= sched_nr_latency; in check_preempt_wakeup()
7146 * lead to a throttle). This both saves work and prevents false in check_preempt_wakeup()
7147 * next-buddy nomination below. in check_preempt_wakeup()
7161 * Note: this also catches the edge-case of curr being in a throttled in check_preempt_wakeup()
7170 /* Idle tasks are by definition preempted by non-idle tasks. */ in check_preempt_wakeup()
7176 * Batch and idle tasks do not preempt non-idle tasks (their preemption in check_preempt_wakeup()
7179 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) in check_preempt_wakeup()
7189 * Preempt an idle group in favor of a non-idle group (and don't preempt in check_preempt_wakeup()
7215 * with schedule on the ->pre_schedule() or idle_balance() in check_preempt_wakeup()
7221 if (unlikely(!se->on_rq || curr == rq->idle)) in check_preempt_wakeup()
7235 cfs_rq = &rq->cfs; in pick_task_fair()
7236 if (!cfs_rq->nr_running) in pick_task_fair()
7240 struct sched_entity *curr = cfs_rq->curr; in pick_task_fair()
7244 if (curr->on_rq) in pick_task_fair()
7264 struct cfs_rq *cfs_rq = &rq->cfs; in pick_next_task_fair()
7274 if (!prev || prev->sched_class != &fair_sched_class) in pick_next_task_fair()
7286 struct sched_entity *curr = cfs_rq->curr; in pick_next_task_fair()
7290 * have to consider cfs_rq->curr. If it is still a runnable in pick_next_task_fair()
7295 if (curr->on_rq) in pick_next_task_fair()
7302 * throttle and dequeue its entity in the parent(s). in pick_next_task_fair()
7307 cfs_rq = &rq->cfs; in pick_next_task_fair()
7309 if (!cfs_rq->nr_running) in pick_next_task_fair()
7328 struct sched_entity *pse = &prev->se; in pick_next_task_fair()
7331 int se_depth = se->depth; in pick_next_task_fair()
7332 int pse_depth = pse->depth; in pick_next_task_fair()
7369 list_move(&p->se.group_node, &rq->cfs_tasks); in pick_next_task_fair()
7386 * Because newidle_balance() releases (and re-acquires) rq->lock, it is in pick_next_task_fair()
7388 * must re-start the pick_next_entity() loop. in pick_next_task_fair()
7415 struct sched_entity *se = &prev->se; in put_prev_task_fair()
7427 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7431 struct task_struct *curr = rq->curr; in yield_task_fair()
7433 struct sched_entity *se = &curr->se; in yield_task_fair()
7438 if (unlikely(rq->nr_running == 1)) in yield_task_fair()
7443 if (curr->policy != SCHED_BATCH) { in yield_task_fair()
7446 * Update run-time statistics of the 'current'. in yield_task_fair()
7462 struct sched_entity *se = &p->se; in yield_to_task_fair()
7465 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) in yield_to_task_fair()
7478 * Fair scheduling class load-balancing methods.
7482 * The purpose of load-balancing is to achieve the same basic fairness the
7483 * per-CPU scheduler provides, namely provide a proportional amount of compute
7488 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7493 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7499 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7508 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7515 * - infeasible weights;
7516 * - local vs global optima in the discrete case. ]
7526 * of load-balance at each level inv. proportional to the number of CPUs in
7532 * \Sum { --- * --- * 2^i } = O(n) (5)
7534 * `- size of each group
7535 * | | `- number of CPUs doing load-balance
7536 * | `- freq
7537 * `- sum over all levels
7579 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7586 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7663 /* The set of CPUs under consideration for load-balancing */
7678 * Is this task likely cache-hot:
7684 lockdep_assert_rq_held(env->src_rq); in task_hot()
7686 if (p->sched_class != &fair_sched_class) in task_hot()
7693 if (env->sd->flags & SD_SHARE_CPUCAPACITY) in task_hot()
7699 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && in task_hot()
7700 (&p->se == cfs_rq_of(&p->se)->next || in task_hot()
7701 &p->se == cfs_rq_of(&p->se)->last)) in task_hot()
7704 if (sysctl_sched_migration_cost == -1) in task_hot()
7711 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p)) in task_hot()
7717 delta = rq_clock_task(env->src_rq) - p->se.exec_start; in task_hot()
7726 * Returns -1, if task migration is not affected by locality.
7730 struct numa_group *numa_group = rcu_dereference(p->numa_group); in migrate_degrades_locality()
7735 return -1; in migrate_degrades_locality()
7737 if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) in migrate_degrades_locality()
7738 return -1; in migrate_degrades_locality()
7740 src_nid = cpu_to_node(env->src_cpu); in migrate_degrades_locality()
7741 dst_nid = cpu_to_node(env->dst_cpu); in migrate_degrades_locality()
7744 return -1; in migrate_degrades_locality()
7747 if (src_nid == p->numa_preferred_nid) { in migrate_degrades_locality()
7748 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) in migrate_degrades_locality()
7751 return -1; in migrate_degrades_locality()
7755 if (dst_nid == p->numa_preferred_nid) in migrate_degrades_locality()
7759 if (env->idle == CPU_IDLE) in migrate_degrades_locality()
7760 return -1; in migrate_degrades_locality()
7778 return -1; in migrate_degrades_locality()
7783 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7790 lockdep_assert_rq_held(env->src_rq); in can_migrate_task()
7797 * 4) are cache-hot on their current CPU. in can_migrate_task()
7799 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) in can_migrate_task()
7806 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { in can_migrate_task()
7809 schedstat_inc(p->se.statistics.nr_failed_migrations_affine); in can_migrate_task()
7811 env->flags |= LBF_SOME_PINNED; in can_migrate_task()
7819 * - for NEWLY_IDLE in can_migrate_task()
7820 * - if we have already computed one in current iteration in can_migrate_task()
7821 * - if it's an active balance in can_migrate_task()
7823 if (env->idle == CPU_NEWLY_IDLE || in can_migrate_task()
7824 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB)) in can_migrate_task()
7827 /* Prevent to re-select dst_cpu via env's CPUs: */ in can_migrate_task()
7828 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { in can_migrate_task()
7829 if (cpumask_test_cpu(cpu, p->cpus_ptr)) { in can_migrate_task()
7830 env->flags |= LBF_DST_PINNED; in can_migrate_task()
7831 env->new_dst_cpu = cpu; in can_migrate_task()
7840 env->flags &= ~LBF_ALL_PINNED; in can_migrate_task()
7842 if (task_running(env->src_rq, p)) { in can_migrate_task()
7843 schedstat_inc(p->se.statistics.nr_failed_migrations_running); in can_migrate_task()
7854 if (env->flags & LBF_ACTIVE_LB) in can_migrate_task()
7858 if (tsk_cache_hot == -1) in can_migrate_task()
7862 env->sd->nr_balance_failed > env->sd->cache_nice_tries) { in can_migrate_task()
7864 schedstat_inc(env->sd->lb_hot_gained[env->idle]); in can_migrate_task()
7865 schedstat_inc(p->se.statistics.nr_forced_migrations); in can_migrate_task()
7870 schedstat_inc(p->se.statistics.nr_failed_migrations_hot); in can_migrate_task()
7875 * detach_task() -- detach the task for the migration specified in env
7879 lockdep_assert_rq_held(env->src_rq); in detach_task()
7881 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); in detach_task()
7882 set_task_cpu(p, env->dst_cpu); in detach_task()
7886 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7895 lockdep_assert_rq_held(env->src_rq); in detach_one_task()
7898 &env->src_rq->cfs_tasks, se.group_node) { in detach_one_task()
7906 * lb_gained[env->idle] is updated (other is detach_tasks) in detach_one_task()
7910 schedstat_inc(env->sd->lb_gained[env->idle]); in detach_one_task()
7919 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7926 struct list_head *tasks = &env->src_rq->cfs_tasks; in detach_tasks()
7931 lockdep_assert_rq_held(env->src_rq); in detach_tasks()
7937 if (env->src_rq->nr_running <= 1) { in detach_tasks()
7938 env->flags &= ~LBF_ALL_PINNED; in detach_tasks()
7942 if (env->imbalance <= 0) in detach_tasks()
7950 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) in detach_tasks()
7955 env->loop++; in detach_tasks()
7957 if (env->loop > env->loop_max) in detach_tasks()
7961 if (env->loop > env->loop_break) { in detach_tasks()
7962 env->loop_break += sched_nr_migrate_break; in detach_tasks()
7963 env->flags |= LBF_NEED_BREAK; in detach_tasks()
7970 switch (env->migration_type) { in detach_tasks()
7975 * value. Make sure that env->imbalance decreases in detach_tasks()
7982 load < 16 && !env->sd->nr_balance_failed) in detach_tasks()
7991 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance) in detach_tasks()
7994 env->imbalance -= load; in detach_tasks()
8000 if (util > env->imbalance) in detach_tasks()
8003 env->imbalance -= util; in detach_tasks()
8007 env->imbalance--; in detach_tasks()
8012 if (task_fits_capacity(p, capacity_of(env->src_cpu))) in detach_tasks()
8015 env->imbalance = 0; in detach_tasks()
8020 list_add(&p->se.group_node, &env->tasks); in detach_tasks()
8030 if (env->idle == CPU_NEWLY_IDLE) in detach_tasks()
8038 if (env->imbalance <= 0) in detach_tasks()
8043 list_move(&p->se.group_node, tasks); in detach_tasks()
8051 schedstat_add(env->sd->lb_gained[env->idle], detached); in detach_tasks()
8057 * attach_task() -- attach the task detached by detach_task() to its new rq.
8069 * attach_one_task() -- attaches the task returned from detach_one_task() to
8083 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8088 struct list_head *tasks = &env->tasks; in attach_tasks()
8092 rq_lock(env->dst_rq, &rf); in attach_tasks()
8093 update_rq_clock(env->dst_rq); in attach_tasks()
8097 list_del_init(&p->se.group_node); in attach_tasks()
8099 attach_task(env->dst_rq, p); in attach_tasks()
8102 rq_unlock(env->dst_rq, &rf); in attach_tasks()
8108 if (cfs_rq->avg.load_avg) in cfs_rq_has_blocked()
8111 if (cfs_rq->avg.util_avg) in cfs_rq_has_blocked()
8119 if (READ_ONCE(rq->avg_rt.util_avg)) in others_have_blocked()
8122 if (READ_ONCE(rq->avg_dl.util_avg)) in others_have_blocked()
8129 if (READ_ONCE(rq->avg_irq.util_avg)) in others_have_blocked()
8138 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies); in update_blocked_load_tick()
8144 rq->has_blocked_load = 0; in update_blocked_load_status()
8164 curr_class = rq->curr->sched_class; in __update_blocked_others()
8197 if (cfs_rq == &rq->cfs) in __update_blocked_fair()
8202 se = cfs_rq->tg->se[cpu]; in __update_blocked_fair()
8223 * This needs to be done in a top-down fashion because the load of a child
8229 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; in update_cfs_rq_h_load()
8233 if (cfs_rq->last_h_load_update == now) in update_cfs_rq_h_load()
8236 WRITE_ONCE(cfs_rq->h_load_next, NULL); in update_cfs_rq_h_load()
8239 WRITE_ONCE(cfs_rq->h_load_next, se); in update_cfs_rq_h_load()
8240 if (cfs_rq->last_h_load_update == now) in update_cfs_rq_h_load()
8245 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); in update_cfs_rq_h_load()
8246 cfs_rq->last_h_load_update = now; in update_cfs_rq_h_load()
8249 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { in update_cfs_rq_h_load()
8250 load = cfs_rq->h_load; in update_cfs_rq_h_load()
8251 load = div64_ul(load * se->avg.load_avg, in update_cfs_rq_h_load()
8254 cfs_rq->h_load = load; in update_cfs_rq_h_load()
8255 cfs_rq->last_h_load_update = now; in update_cfs_rq_h_load()
8264 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, in task_h_load()
8270 struct cfs_rq *cfs_rq = &rq->cfs; in __update_blocked_fair()
8282 return p->se.avg.load_avg; in task_h_load()
8308 * sg_lb_stats - stats of a sched_group required for load_balancing
8330 * sd_lb_stats - Structure to store the statistics of a sched_domain
8384 used = READ_ONCE(rq->avg_rt.util_avg); in scale_rt_capacity()
8385 used += READ_ONCE(rq->avg_dl.util_avg); in scale_rt_capacity()
8391 free = max - used; in scale_rt_capacity()
8399 struct sched_group *sdg = sd->groups; in update_cpu_capacity()
8401 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu); in update_cpu_capacity()
8406 cpu_rq(cpu)->cpu_capacity = capacity; in update_cpu_capacity()
8409 sdg->sgc->capacity = capacity; in update_cpu_capacity()
8410 sdg->sgc->min_capacity = capacity; in update_cpu_capacity()
8411 sdg->sgc->max_capacity = capacity; in update_cpu_capacity()
8416 struct sched_domain *child = sd->child; in update_group_capacity()
8417 struct sched_group *group, *sdg = sd->groups; in update_group_capacity()
8421 interval = msecs_to_jiffies(sd->balance_interval); in update_group_capacity()
8423 sdg->sgc->next_update = jiffies + interval; in update_group_capacity()
8434 if (child->flags & SD_OVERLAP) { in update_group_capacity()
8453 group = child->groups; in update_group_capacity()
8455 struct sched_group_capacity *sgc = group->sgc; in update_group_capacity()
8457 capacity += sgc->capacity; in update_group_capacity()
8458 min_capacity = min(sgc->min_capacity, min_capacity); in update_group_capacity()
8459 max_capacity = max(sgc->max_capacity, max_capacity); in update_group_capacity()
8460 group = group->next; in update_group_capacity()
8461 } while (group != child->groups); in update_group_capacity()
8464 sdg->sgc->capacity = capacity; in update_group_capacity()
8465 sdg->sgc->min_capacity = min_capacity; in update_group_capacity()
8466 sdg->sgc->max_capacity = max_capacity; in update_group_capacity()
8477 return ((rq->cpu_capacity * sd->imbalance_pct) < in check_cpu_capacity()
8478 (rq->cpu_capacity_orig * 100)); in check_cpu_capacity()
8488 return rq->misfit_task_load && in check_misfit_status()
8489 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity || in check_misfit_status()
8495 * groups is inadequate due to ->cpus_ptr constraints.
8504 * If we were to balance group-wise we'd place two tasks in the first group and
8524 return group->sgc->imbalance; in sg_imbalanced()
8542 if (sgs->sum_nr_running < sgs->group_weight) in group_has_capacity()
8545 if ((sgs->group_capacity * imbalance_pct) < in group_has_capacity()
8546 (sgs->group_runnable * 100)) in group_has_capacity()
8549 if ((sgs->group_capacity * 100) > in group_has_capacity()
8550 (sgs->group_util * imbalance_pct)) in group_has_capacity()
8567 if (sgs->sum_nr_running <= sgs->group_weight) in group_is_overloaded()
8570 if ((sgs->group_capacity * 100) < in group_is_overloaded()
8571 (sgs->group_util * imbalance_pct)) in group_is_overloaded()
8574 if ((sgs->group_capacity * imbalance_pct) < in group_is_overloaded()
8575 (sgs->group_runnable * 100)) in group_is_overloaded()
8592 if (sgs->group_asym_packing) in group_classify()
8595 if (sgs->group_misfit_task_load) in group_classify()
8605 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8620 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group)); in update_sg_lb_stats()
8622 for_each_cpu_and(i, sched_group_span(group), env->cpus) { in update_sg_lb_stats()
8625 sgs->group_load += cpu_load(rq); in update_sg_lb_stats()
8626 sgs->group_util += cpu_util(i); in update_sg_lb_stats()
8627 sgs->group_runnable += cpu_runnable(rq); in update_sg_lb_stats()
8628 sgs->sum_h_nr_running += rq->cfs.h_nr_running; in update_sg_lb_stats()
8630 nr_running = rq->nr_running; in update_sg_lb_stats()
8631 sgs->sum_nr_running += nr_running; in update_sg_lb_stats()
8640 sgs->nr_numa_running += rq->nr_numa_running; in update_sg_lb_stats()
8641 sgs->nr_preferred_running += rq->nr_preferred_running; in update_sg_lb_stats()
8647 sgs->idle_cpus++; in update_sg_lb_stats()
8656 if (env->sd->flags & SD_ASYM_CPUCAPACITY && in update_sg_lb_stats()
8657 sgs->group_misfit_task_load < rq->misfit_task_load) { in update_sg_lb_stats()
8658 sgs->group_misfit_task_load = rq->misfit_task_load; in update_sg_lb_stats()
8664 if (env->sd->flags & SD_ASYM_PACKING && in update_sg_lb_stats()
8665 env->idle != CPU_NOT_IDLE && in update_sg_lb_stats()
8666 sgs->sum_h_nr_running && in update_sg_lb_stats()
8667 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) { in update_sg_lb_stats()
8668 sgs->group_asym_packing = 1; in update_sg_lb_stats()
8671 sgs->group_capacity = group->sgc->capacity; in update_sg_lb_stats()
8673 sgs->group_weight = group->group_weight; in update_sg_lb_stats()
8675 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs); in update_sg_lb_stats()
8678 if (sgs->group_type == group_overloaded) in update_sg_lb_stats()
8679 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / in update_sg_lb_stats()
8680 sgs->group_capacity; in update_sg_lb_stats()
8684 * update_sd_pick_busiest - return 1 on busiest group
8701 struct sg_lb_stats *busiest = &sds->busiest_stat; in update_sd_pick_busiest()
8704 if (!sgs->sum_h_nr_running) in update_sd_pick_busiest()
8713 if (sgs->group_type == group_misfit_task && in update_sd_pick_busiest()
8714 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) || in update_sd_pick_busiest()
8715 sds->local_stat.group_type != group_has_spare)) in update_sd_pick_busiest()
8718 if (sgs->group_type > busiest->group_type) in update_sd_pick_busiest()
8721 if (sgs->group_type < busiest->group_type) in update_sd_pick_busiest()
8729 switch (sgs->group_type) { in update_sd_pick_busiest()
8732 if (sgs->avg_load <= busiest->avg_load) in update_sd_pick_busiest()
8745 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu)) in update_sd_pick_busiest()
8754 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load) in update_sd_pick_busiest()
8769 if (sgs->avg_load <= busiest->avg_load) in update_sd_pick_busiest()
8781 if (sgs->idle_cpus > busiest->idle_cpus) in update_sd_pick_busiest()
8783 else if ((sgs->idle_cpus == busiest->idle_cpus) && in update_sd_pick_busiest()
8784 (sgs->sum_nr_running <= busiest->sum_nr_running)) in update_sd_pick_busiest()
8792 * per-CPU capacity. Migrating tasks to less capable CPUs may harm in update_sd_pick_busiest()
8796 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && in update_sd_pick_busiest()
8797 (sgs->group_type <= group_fully_busy) && in update_sd_pick_busiest()
8798 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu)))) in update_sd_pick_busiest()
8807 if (sgs->sum_h_nr_running > sgs->nr_numa_running) in fbq_classify_group()
8809 if (sgs->sum_h_nr_running > sgs->nr_preferred_running) in fbq_classify_group()
8816 if (rq->nr_running > rq->nr_numa_running) in fbq_classify_rq()
8818 if (rq->nr_running > rq->nr_preferred_running) in fbq_classify_rq()
8838 * task_running_on_cpu - return 1 if @p is running on @cpu.
8844 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) in task_running_on_cpu()
8854 * idle_cpu_without - would a given CPU be idle without p ?
8864 if (rq->curr != rq->idle && rq->curr != p) in idle_cpu_without()
8868 * rq->nr_running can't be used but an updated version without the in idle_cpu_without()
8874 if (rq->ttwu_pending) in idle_cpu_without()
8882 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8901 sgs->group_load += cpu_load_without(rq, p); in update_sg_wakeup_stats()
8902 sgs->group_util += cpu_util_without(i, p); in update_sg_wakeup_stats()
8903 sgs->group_runnable += cpu_runnable_without(rq, p); in update_sg_wakeup_stats()
8905 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local; in update_sg_wakeup_stats()
8907 nr_running = rq->nr_running - local; in update_sg_wakeup_stats()
8908 sgs->sum_nr_running += nr_running; in update_sg_wakeup_stats()
8914 sgs->idle_cpus++; in update_sg_wakeup_stats()
8919 if (sd->flags & SD_ASYM_CPUCAPACITY && in update_sg_wakeup_stats()
8920 !task_fits_capacity(p, group->sgc->max_capacity)) { in update_sg_wakeup_stats()
8921 sgs->group_misfit_task_load = 1; in update_sg_wakeup_stats()
8924 sgs->group_capacity = group->sgc->capacity; in update_sg_wakeup_stats()
8926 sgs->group_weight = group->group_weight; in update_sg_wakeup_stats()
8928 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs); in update_sg_wakeup_stats()
8934 if (sgs->group_type == group_fully_busy || in update_sg_wakeup_stats()
8935 sgs->group_type == group_overloaded) in update_sg_wakeup_stats()
8936 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / in update_sg_wakeup_stats()
8937 sgs->group_capacity; in update_sg_wakeup_stats()
8945 if (sgs->group_type < idlest_sgs->group_type) in update_pick_idlest()
8948 if (sgs->group_type > idlest_sgs->group_type) in update_pick_idlest()
8956 switch (sgs->group_type) { in update_pick_idlest()
8960 if (idlest_sgs->avg_load <= sgs->avg_load) in update_pick_idlest()
8971 if (idlest->sgc->max_capacity >= group->sgc->max_capacity) in update_pick_idlest()
8977 if (idlest_sgs->idle_cpus > sgs->idle_cpus) in update_pick_idlest()
8981 if (idlest_sgs->idle_cpus == sgs->idle_cpus && in update_pick_idlest()
8982 idlest_sgs->group_util <= sgs->group_util) in update_pick_idlest()
9010 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups; in find_idlest_group()
9024 p->cpus_ptr)) in find_idlest_group()
9048 } while (group = group->next, group != sd->groups); in find_idlest_group()
9079 (sd->imbalance_pct-100) / 100; in find_idlest_group()
9086 * cross-domain, add imbalance to the load on the remote node in find_idlest_group()
9090 if ((sd->flags & SD_NUMA) && in find_idlest_group()
9101 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load) in find_idlest_group()
9112 if (local->sgc->max_capacity >= idlest->sgc->max_capacity) in find_idlest_group()
9117 if (sd->flags & SD_NUMA) { in find_idlest_group()
9124 if (cpu_to_node(this_cpu) == p->numa_preferred_nid) in find_idlest_group()
9128 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid) in find_idlest_group()
9137 if (allow_numa_imbalance(local_sgs.sum_nr_running, sd->span_weight)) in find_idlest_group()
9156 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9163 struct sched_domain *child = env->sd->child; in update_sd_lb_stats()
9164 struct sched_group *sg = env->sd->groups; in update_sd_lb_stats()
9165 struct sg_lb_stats *local = &sds->local_stat; in update_sd_lb_stats()
9173 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); in update_sd_lb_stats()
9175 sds->local = sg; in update_sd_lb_stats()
9178 if (env->idle != CPU_NEWLY_IDLE || in update_sd_lb_stats()
9179 time_after_eq(jiffies, sg->sgc->next_update)) in update_sd_lb_stats()
9180 update_group_capacity(env->sd, env->dst_cpu); in update_sd_lb_stats()
9190 sds->busiest = sg; in update_sd_lb_stats()
9191 sds->busiest_stat = *sgs; in update_sd_lb_stats()
9196 sds->total_load += sgs->group_load; in update_sd_lb_stats()
9197 sds->total_capacity += sgs->group_capacity; in update_sd_lb_stats()
9199 sg = sg->next; in update_sd_lb_stats()
9200 } while (sg != env->sd->groups); in update_sd_lb_stats()
9203 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING; in update_sd_lb_stats()
9206 if (env->sd->flags & SD_NUMA) in update_sd_lb_stats()
9207 env->fbq_type = fbq_classify_group(&sds->busiest_stat); in update_sd_lb_stats()
9209 if (!env->sd->parent) { in update_sd_lb_stats()
9210 struct root_domain *rd = env->dst_rq->rd; in update_sd_lb_stats()
9213 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD); in update_sd_lb_stats()
9215 /* Update over-utilization (tipping point, U >= 0) indicator */ in update_sd_lb_stats()
9216 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED); in update_sd_lb_stats()
9219 struct root_domain *rd = env->dst_rq->rd; in update_sd_lb_stats()
9221 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED); in update_sd_lb_stats()
9245 * calculate_imbalance - Calculate the amount of imbalance present within the
9254 local = &sds->local_stat; in calculate_imbalance()
9255 busiest = &sds->busiest_stat; in calculate_imbalance()
9257 if (busiest->group_type == group_misfit_task) { in calculate_imbalance()
9259 env->migration_type = migrate_misfit; in calculate_imbalance()
9260 env->imbalance = 1; in calculate_imbalance()
9264 if (busiest->group_type == group_asym_packing) { in calculate_imbalance()
9269 env->migration_type = migrate_task; in calculate_imbalance()
9270 env->imbalance = busiest->sum_h_nr_running; in calculate_imbalance()
9274 if (busiest->group_type == group_imbalanced) { in calculate_imbalance()
9276 * In the group_imb case we cannot rely on group-wide averages in calculate_imbalance()
9277 * to ensure CPU-load equilibrium, try to move any task to fix in calculate_imbalance()
9281 env->migration_type = migrate_task; in calculate_imbalance()
9282 env->imbalance = 1; in calculate_imbalance()
9290 if (local->group_type == group_has_spare) { in calculate_imbalance()
9291 if ((busiest->group_type > group_fully_busy) && in calculate_imbalance()
9292 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) { in calculate_imbalance()
9301 env->migration_type = migrate_util; in calculate_imbalance()
9302 env->imbalance = max(local->group_capacity, local->group_util) - in calculate_imbalance()
9303 local->group_util; in calculate_imbalance()
9312 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) { in calculate_imbalance()
9313 env->migration_type = migrate_task; in calculate_imbalance()
9314 env->imbalance = 1; in calculate_imbalance()
9320 if (busiest->group_weight == 1 || sds->prefer_sibling) { in calculate_imbalance()
9321 unsigned int nr_diff = busiest->sum_nr_running; in calculate_imbalance()
9326 env->migration_type = migrate_task; in calculate_imbalance()
9327 lsub_positive(&nr_diff, local->sum_nr_running); in calculate_imbalance()
9328 env->imbalance = nr_diff >> 1; in calculate_imbalance()
9335 env->migration_type = migrate_task; in calculate_imbalance()
9336 env->imbalance = max_t(long, 0, (local->idle_cpus - in calculate_imbalance()
9337 busiest->idle_cpus) >> 1); in calculate_imbalance()
9341 if (env->sd->flags & SD_NUMA) { in calculate_imbalance()
9342 env->imbalance = adjust_numa_imbalance(env->imbalance, in calculate_imbalance()
9343 busiest->sum_nr_running, busiest->group_weight); in calculate_imbalance()
9353 if (local->group_type < group_overloaded) { in calculate_imbalance()
9359 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) / in calculate_imbalance()
9360 local->group_capacity; in calculate_imbalance()
9362 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) / in calculate_imbalance()
9363 sds->total_capacity; in calculate_imbalance()
9368 if (local->avg_load >= busiest->avg_load) { in calculate_imbalance()
9369 env->imbalance = 0; in calculate_imbalance()
9382 env->migration_type = migrate_load; in calculate_imbalance()
9383 env->imbalance = min( in calculate_imbalance()
9384 (busiest->avg_load - sds->avg_load) * busiest->group_capacity, in calculate_imbalance()
9385 (sds->avg_load - local->avg_load) * local->group_capacity in calculate_imbalance()
9412 * find_busiest_group - Returns the busiest group within the sched_domain
9420 * Return: - The busiest group if imbalance exists.
9436 struct root_domain *rd = env->dst_rq->rd; in find_busiest_group()
9438 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)) in find_busiest_group()
9450 if (busiest->group_type == group_misfit_task) in find_busiest_group()
9454 if (busiest->group_type == group_asym_packing) in find_busiest_group()
9462 if (busiest->group_type == group_imbalanced) in find_busiest_group()
9469 if (local->group_type > busiest->group_type) in find_busiest_group()
9476 if (local->group_type == group_overloaded) { in find_busiest_group()
9481 if (local->avg_load >= busiest->avg_load) in find_busiest_group()
9492 if (local->avg_load >= sds.avg_load) in find_busiest_group()
9499 if (100 * busiest->avg_load <= in find_busiest_group()
9500 env->sd->imbalance_pct * local->avg_load) in find_busiest_group()
9505 if (sds.prefer_sibling && local->group_type == group_has_spare && in find_busiest_group()
9506 busiest->sum_nr_running > local->sum_nr_running + 1) in find_busiest_group()
9509 if (busiest->group_type != group_overloaded) { in find_busiest_group()
9510 if (env->idle == CPU_NOT_IDLE) in find_busiest_group()
9518 if (busiest->group_weight > 1 && in find_busiest_group()
9519 local->idle_cpus <= (busiest->idle_cpus + 1)) in find_busiest_group()
9531 if (busiest->sum_h_nr_running == 1) in find_busiest_group()
9541 return env->imbalance ? sds.busiest : NULL; in find_busiest_group()
9544 env->imbalance = 0; in find_busiest_group()
9549 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9559 for_each_cpu_and(i, sched_group_span(group), env->cpus) { in find_busiest_queue()
9569 * - regular: there are !numa tasks in find_busiest_queue()
9570 * - remote: there are numa tasks that run on the 'wrong' node in find_busiest_queue()
9571 * - all: there is no distinction in find_busiest_queue()
9586 if (rt > env->fbq_type) in find_busiest_queue()
9589 nr_running = rq->cfs.h_nr_running; in find_busiest_queue()
9597 * eventually lead to active_balancing high->low capacity. in find_busiest_queue()
9598 * Higher per-CPU capacity is considered better than balancing in find_busiest_queue()
9601 if (env->sd->flags & SD_ASYM_CPUCAPACITY && in find_busiest_queue()
9602 !capacity_greater(capacity_of(env->dst_cpu), capacity) && in find_busiest_queue()
9606 switch (env->migration_type) { in find_busiest_queue()
9614 if (nr_running == 1 && load > env->imbalance && in find_busiest_queue()
9615 !check_cpu_capacity(rq, env->sd)) in find_busiest_queue()
9667 if (rq->misfit_task_load > busiest_load) { in find_busiest_queue()
9668 busiest_load = rq->misfit_task_load; in find_busiest_queue()
9694 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) && in asym_active_balance()
9695 sched_asym_prefer(env->dst_cpu, env->src_cpu); in asym_active_balance()
9701 struct sched_domain *sd = env->sd; in imbalanced_active_balance()
9708 if ((env->migration_type == migrate_task) && in imbalanced_active_balance()
9709 (sd->nr_balance_failed > sd->cache_nice_tries+2)) in imbalanced_active_balance()
9717 struct sched_domain *sd = env->sd; in need_active_balance()
9731 if ((env->idle != CPU_NOT_IDLE) && in need_active_balance()
9732 (env->src_rq->cfs.h_nr_running == 1)) { in need_active_balance()
9733 if ((check_cpu_capacity(env->src_rq, sd)) && in need_active_balance()
9734 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) in need_active_balance()
9738 if (env->migration_type == migrate_misfit) in need_active_balance()
9748 struct sched_group *sg = env->sd->groups; in should_we_balance()
9755 if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) in should_we_balance()
9762 if (env->idle == CPU_NEWLY_IDLE) in should_we_balance()
9766 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { in should_we_balance()
9771 return cpu == env->dst_cpu; in should_we_balance()
9775 return group_balance_cpu(sg) == env->dst_cpu; in should_we_balance()
9787 struct sched_domain *sd_parent = sd->parent; in load_balance()
9797 .dst_grpmask = sched_group_span(sd->groups), in load_balance()
9807 schedstat_inc(sd->lb_count[idle]); in load_balance()
9817 schedstat_inc(sd->lb_nobusyg[idle]); in load_balance()
9823 schedstat_inc(sd->lb_nobusyq[idle]); in load_balance()
9829 schedstat_add(sd->lb_imbalance[idle], env.imbalance); in load_balance()
9831 env.src_cpu = busiest->cpu; in load_balance()
9837 if (busiest->nr_running > 1) { in load_balance()
9840 * an imbalance but busiest->nr_running <= 1, the group is in load_balance()
9844 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); in load_balance()
9851 * cur_ld_moved - load moved in current iteration in load_balance()
9852 * ld_moved - cumulative load moved across iterations in load_balance()
9859 * unlock busiest->lock, and we are able to be sure in load_balance()
9888 * nohz-idle), we now have balance_cpu in a position to move in load_balance()
9899 /* Prevent to re-select dst_cpu via env's CPUs */ in load_balance()
9919 int *group_imbalance = &sd_parent->groups->sgc->imbalance; in load_balance()
9946 schedstat_inc(sd->lb_failed[idle]); in load_balance()
9954 sd->nr_balance_failed++; in load_balance()
9966 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { in load_balance()
9975 * ->active_balance synchronizes accesses to in load_balance()
9976 * ->active_balance_work. Once set, it's cleared in load_balance()
9979 if (!busiest->active_balance) { in load_balance()
9980 busiest->active_balance = 1; in load_balance()
9981 busiest->push_cpu = this_cpu; in load_balance()
9989 &busiest->active_balance_work); in load_balance()
9993 sd->nr_balance_failed = 0; in load_balance()
9998 sd->balance_interval = sd->min_interval; in load_balance()
10010 int *group_imbalance = &sd_parent->groups->sgc->imbalance; in load_balance()
10022 schedstat_inc(sd->lb_balanced[idle]); in load_balance()
10024 sd->nr_balance_failed = 0; in load_balance()
10040 sd->balance_interval < MAX_PINNED_INTERVAL) || in load_balance()
10041 sd->balance_interval < sd->max_interval) in load_balance()
10042 sd->balance_interval *= 2; in load_balance()
10050 unsigned long interval = sd->balance_interval; in get_sd_balance_interval()
10053 interval *= sd->busy_factor; in get_sd_balance_interval()
10055 /* scale ms to jiffies */ in get_sd_balance_interval()
10064 interval -= 1; in get_sd_balance_interval()
10078 next = sd->last_balance + interval; in update_next_balance()
10094 int target_cpu = busiest_rq->push_cpu; in active_load_balance_cpu_stop()
10102 * Between queueing the stop-work and running it is a hole in which in active_load_balance_cpu_stop()
10111 !busiest_rq->active_balance)) in active_load_balance_cpu_stop()
10115 if (busiest_rq->nr_running <= 1) in active_load_balance_cpu_stop()
10121 * Bjorn Helgaas on a 128-CPU setup. in active_load_balance_cpu_stop()
10137 .src_cpu = busiest_rq->cpu, in active_load_balance_cpu_stop()
10143 schedstat_inc(sd->alb_count); in active_load_balance_cpu_stop()
10148 schedstat_inc(sd->alb_pushed); in active_load_balance_cpu_stop()
10150 sd->nr_balance_failed = 0; in active_load_balance_cpu_stop()
10152 schedstat_inc(sd->alb_failed); in active_load_balance_cpu_stop()
10157 busiest_rq->active_balance = 0; in active_load_balance_cpu_stop()
10172 * This trades load-balance latency on larger machines for less cross talk.
10188 int cpu = rq->cpu; in rebalance_domains()
10204 if (time_after(jiffies, sd->next_decay_max_lb_cost)) { in rebalance_domains()
10205 sd->max_newidle_lb_cost = in rebalance_domains()
10206 (sd->max_newidle_lb_cost * 253) / 256; in rebalance_domains()
10207 sd->next_decay_max_lb_cost = jiffies + HZ; in rebalance_domains()
10210 max_cost += sd->max_newidle_lb_cost; in rebalance_domains()
10225 need_serialize = sd->flags & SD_SERIALIZE; in rebalance_domains()
10231 if (time_after_eq(jiffies, sd->last_balance + interval)) { in rebalance_domains()
10235 * env->dst_cpu, so we can't know our idle in rebalance_domains()
10241 sd->last_balance = jiffies; in rebalance_domains()
10247 if (time_after(next_balance, sd->last_balance + interval)) { in rebalance_domains()
10248 next_balance = sd->last_balance + interval; in rebalance_domains()
10254 * Ensure the rq-wide value also decays but keep it at a in rebalance_domains()
10255 * reasonable floor to avoid funnies with rq->avg_idle. in rebalance_domains()
10257 rq->max_idle_balance_cost = in rebalance_domains()
10268 rq->next_balance = next_balance; in rebalance_domains()
10274 return unlikely(!rcu_dereference_sched(rq->sd)); in on_null_domain()
10280 * - When one of the busy CPUs notice that there may be an idle rebalancing
10283 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10339 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd); in kick_ilb()
10351 int nr_busy, i, cpu = rq->cpu; in nohz_balancer_kick()
10354 if (unlikely(rq->idle_balance)) in nohz_balancer_kick()
10377 if (rq->nr_running >= 2) { in nohz_balancer_kick()
10384 sd = rcu_dereference(rq->sd); in nohz_balancer_kick()
10391 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) { in nohz_balancer_kick()
10437 * increase the overall cache use), we need some less-loaded LLC in nohz_balancer_kick()
10441 * the others are - so just get a nohz balance going if it looks in nohz_balancer_kick()
10444 nr_busy = atomic_read(&sds->nr_busy_cpus); in nohz_balancer_kick()
10464 if (!sd || !sd->nohz_idle) in set_cpu_sd_state_busy()
10466 sd->nohz_idle = 0; in set_cpu_sd_state_busy()
10468 atomic_inc(&sd->shared->nr_busy_cpus); in set_cpu_sd_state_busy()
10477 if (likely(!rq->nohz_tick_stopped)) in nohz_balance_exit_idle()
10480 rq->nohz_tick_stopped = 0; in nohz_balance_exit_idle()
10481 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); in nohz_balance_exit_idle()
10484 set_cpu_sd_state_busy(rq->cpu); in nohz_balance_exit_idle()
10494 if (!sd || sd->nohz_idle) in set_cpu_sd_state_idle()
10496 sd->nohz_idle = 1; in set_cpu_sd_state_idle()
10498 atomic_dec(&sd->shared->nr_busy_cpus); in set_cpu_sd_state_idle()
10522 * Can be set safely without rq->lock held in nohz_balance_enter_idle()
10524 * rq->lock is held during the check and the clear in nohz_balance_enter_idle()
10526 rq->has_blocked_load = 1; in nohz_balance_enter_idle()
10534 if (rq->nohz_tick_stopped) in nohz_balance_enter_idle()
10541 rq->nohz_tick_stopped = 1; in nohz_balance_enter_idle()
10565 unsigned int cpu = rq->cpu; in update_nohz_stats()
10567 if (!rq->has_blocked_load) in update_nohz_stats()
10573 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick))) in update_nohz_stats()
10578 return rq->has_blocked_load; in update_nohz_stats()
10594 int this_cpu = this_rq->cpu; in _nohz_idle_balance()
10642 if (time_after_eq(jiffies, rq->next_balance)) { in _nohz_idle_balance()
10653 if (time_after(next_balance, rq->next_balance)) { in _nohz_idle_balance()
10654 next_balance = rq->next_balance; in _nohz_idle_balance()
10682 unsigned int flags = this_rq->nohz_idle_balance; in nohz_idle_balance()
10687 this_rq->nohz_idle_balance = 0; in nohz_idle_balance()
10717 int this_cpu = this_rq->cpu; in nohz_newidle_balance()
10727 if (this_rq->avg_idle < sysctl_sched_migration_cost) in nohz_newidle_balance()
10758 * < 0 - we released the lock and there are !fair tasks present
10759 * 0 - failed, no new tasks
10760 * > 0 - success, new (fair) tasks present
10765 int this_cpu = this_rq->cpu; in newidle_balance()
10776 if (this_rq->ttwu_pending) in newidle_balance()
10783 this_rq->idle_stamp = rq_clock(this_rq); in newidle_balance()
10793 * for load-balance and preemption/IRQs are still disabled avoiding in newidle_balance()
10795 * re-start the picking loop. in newidle_balance()
10799 if (this_rq->avg_idle < sysctl_sched_migration_cost || in newidle_balance()
10800 !READ_ONCE(this_rq->rd->overload)) { in newidle_balance()
10803 sd = rcu_dereference_check_sched_domain(this_rq->sd); in newidle_balance()
10819 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { in newidle_balance()
10824 if (sd->flags & SD_BALANCE_NEWIDLE) { in newidle_balance()
10831 domain_cost = sched_clock_cpu(this_cpu) - t0; in newidle_balance()
10832 if (domain_cost > sd->max_newidle_lb_cost) in newidle_balance()
10833 sd->max_newidle_lb_cost = domain_cost; in newidle_balance()
10844 if (pulled_task || this_rq->nr_running > 0 || in newidle_balance()
10845 this_rq->ttwu_pending) in newidle_balance()
10852 if (curr_cost > this_rq->max_idle_balance_cost) in newidle_balance()
10853 this_rq->max_idle_balance_cost = curr_cost; in newidle_balance()
10860 if (this_rq->cfs.h_nr_running && !pulled_task) in newidle_balance()
10864 if (this_rq->nr_running != this_rq->cfs.h_nr_running) in newidle_balance()
10865 pulled_task = -1; in newidle_balance()
10869 if (time_after(this_rq->next_balance, next_balance)) in newidle_balance()
10870 this_rq->next_balance = next_balance; in newidle_balance()
10873 this_rq->idle_stamp = 0; in newidle_balance()
10889 enum cpu_idle_type idle = this_rq->idle_balance ? in run_rebalance_domains()
10904 update_blocked_averages(this_rq->cpu); in run_rebalance_domains()
10920 if (time_after_eq(jiffies, rq->next_balance)) in trigger_load_balance()
10948 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime; in __entity_slice_used()
10970 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check in task_tick_core()
10973 if (rq->core->core_forceidle && rq->cfs.nr_running == 1 && in task_tick_core()
10974 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE)) in task_tick_core()
10979 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
10987 if (cfs_rq->forceidle_seq == fi_seq) in se_fi_update()
10989 cfs_rq->forceidle_seq = fi_seq; in se_fi_update()
10992 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime; in se_fi_update()
10998 struct sched_entity *se = &p->se; in task_vruntime_update()
11000 if (p->sched_class != &fair_sched_class) in task_vruntime_update()
11003 se_fi_update(se, rq->core->core_forceidle_seq, in_fi); in task_vruntime_update()
11009 struct sched_entity *sea = &a->se; in cfs_prio_less()
11010 struct sched_entity *seb = &b->se; in cfs_prio_less()
11015 SCHED_WARN_ON(task_rq(b)->core != rq->core); in cfs_prio_less()
11022 while (sea->cfs_rq->tg != seb->cfs_rq->tg) { in cfs_prio_less()
11023 int sea_depth = sea->depth; in cfs_prio_less()
11024 int seb_depth = seb->depth; in cfs_prio_less()
11032 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi); in cfs_prio_less()
11033 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi); in cfs_prio_less()
11035 cfs_rqa = sea->cfs_rq; in cfs_prio_less()
11036 cfs_rqb = seb->cfs_rq; in cfs_prio_less()
11038 cfs_rqa = &task_rq(a)->cfs; in cfs_prio_less()
11039 cfs_rqb = &task_rq(b)->cfs; in cfs_prio_less()
11047 delta = (s64)(sea->vruntime - seb->vruntime) + in cfs_prio_less()
11048 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi); in cfs_prio_less()
11067 struct sched_entity *se = &curr->se; in task_tick_fair()
11085 * - child not yet on the tasklist
11086 * - preemption disabled
11091 struct sched_entity *se = &p->se, *curr; in task_fork_fair()
11099 curr = cfs_rq->curr; in task_fork_fair()
11102 se->vruntime = curr->vruntime; in task_fork_fair()
11111 swap(curr->vruntime, se->vruntime); in task_fork_fair()
11115 se->vruntime -= cfs_rq->min_vruntime; in task_fork_fair()
11129 if (rq->cfs.nr_running == 1) in prio_changed_fair()
11138 if (p->prio > oldprio) in prio_changed_fair()
11146 struct sched_entity *se = &p->se; in vruntime_normalized()
11153 if (p->on_rq) in vruntime_normalized()
11160 * - A forked child which is waiting for being woken up by in vruntime_normalized()
11162 * - A task which has been woken up by try_to_wake_up() and in vruntime_normalized()
11165 if (!se->sum_exec_runtime || in vruntime_normalized()
11166 (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup)) in vruntime_normalized()
11184 se = se->parent; in propagate_entity_cfs_rq()
11220 * Since the real-depth could have been changed (only FAIR in attach_entity_cfs_rq()
11223 se->depth = se->parent ? se->parent->depth + 1 : 0; in attach_entity_cfs_rq()
11235 struct sched_entity *se = &p->se; in detach_task_cfs_rq()
11244 se->vruntime -= cfs_rq->min_vruntime; in detach_task_cfs_rq()
11252 struct sched_entity *se = &p->se; in attach_task_cfs_rq()
11258 se->vruntime += cfs_rq->min_vruntime; in attach_task_cfs_rq()
11285 * This routine is mostly called to set cfs_rq->curr field when a task
11290 struct sched_entity *se = &p->se; in set_next_task_fair()
11298 list_move(&se->group_node, &rq->cfs_tasks); in set_next_task_fair()
11313 cfs_rq->tasks_timeline = RB_ROOT_CACHED; in init_cfs_rq()
11314 cfs_rq->min_vruntime = (u64)(-(1LL << 20)); in init_cfs_rq()
11316 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; in init_cfs_rq()
11319 raw_spin_lock_init(&cfs_rq->removed.lock); in init_cfs_rq()
11326 struct sched_entity *se = &p->se; in task_set_group_fair()
11329 se->depth = se->parent ? se->parent->depth + 1 : 0; in task_set_group_fair()
11338 /* Tell se's cfs_rq has been changed -- migrated */ in task_move_group_fair()
11339 p->se.avg.last_update_time = 0; in task_move_group_fair()
11364 if (tg->cfs_rq) in free_fair_sched_group()
11365 kfree(tg->cfs_rq[i]); in free_fair_sched_group()
11366 if (tg->se) in free_fair_sched_group()
11367 kfree(tg->se[i]); in free_fair_sched_group()
11370 kfree(tg->cfs_rq); in free_fair_sched_group()
11371 kfree(tg->se); in free_fair_sched_group()
11380 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); in alloc_fair_sched_group()
11381 if (!tg->cfs_rq) in alloc_fair_sched_group()
11383 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); in alloc_fair_sched_group()
11384 if (!tg->se) in alloc_fair_sched_group()
11387 tg->shares = NICE_0_LOAD; in alloc_fair_sched_group()
11403 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); in alloc_fair_sched_group()
11424 se = tg->se[i]; in online_fair_sched_group()
11440 if (tg->se[cpu]) in unregister_fair_sched_group()
11441 remove_entity_load_avg(tg->se[cpu]); in unregister_fair_sched_group()
11445 * check on_list without danger of it being re-added. in unregister_fair_sched_group()
11447 if (!tg->cfs_rq[cpu]->on_list) in unregister_fair_sched_group()
11453 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); in unregister_fair_sched_group()
11464 cfs_rq->tg = tg; in init_tg_cfs_entry()
11465 cfs_rq->rq = rq; in init_tg_cfs_entry()
11468 tg->cfs_rq[cpu] = cfs_rq; in init_tg_cfs_entry()
11469 tg->se[cpu] = se; in init_tg_cfs_entry()
11476 se->cfs_rq = &rq->cfs; in init_tg_cfs_entry()
11477 se->depth = 0; in init_tg_cfs_entry()
11479 se->cfs_rq = parent->my_q; in init_tg_cfs_entry()
11480 se->depth = parent->depth + 1; in init_tg_cfs_entry()
11483 se->my_q = cfs_rq; in init_tg_cfs_entry()
11485 update_load_set(&se->load, NICE_0_LOAD); in init_tg_cfs_entry()
11486 se->parent = parent; in init_tg_cfs_entry()
11500 if (!tg->se[0]) in __sched_group_set_shares()
11501 return -EINVAL; in __sched_group_set_shares()
11505 if (tg->shares == shares) in __sched_group_set_shares()
11508 tg->shares = shares; in __sched_group_set_shares()
11511 struct sched_entity *se = tg->se[i]; in __sched_group_set_shares()
11533 ret = -EINVAL; in sched_group_set_shares()
11546 return -EINVAL; in sched_group_set_idle()
11549 return -EINVAL; in sched_group_set_idle()
11553 if (tg->idle == idle) { in sched_group_set_idle()
11558 tg->idle = idle; in sched_group_set_idle()
11562 struct sched_entity *se = tg->se[i]; in sched_group_set_idle()
11563 struct cfs_rq *grp_cfs_rq = tg->cfs_rq[i]; in sched_group_set_idle()
11570 grp_cfs_rq->idle = idle; in sched_group_set_idle()
11574 idle_task_delta = grp_cfs_rq->h_nr_running - in sched_group_set_idle()
11575 grp_cfs_rq->idle_h_nr_running; in sched_group_set_idle()
11577 idle_task_delta *= -1; in sched_group_set_idle()
11582 if (!se->on_rq) in sched_group_set_idle()
11585 cfs_rq->idle_h_nr_running += idle_task_delta; in sched_group_set_idle()
11624 struct sched_entity *se = &task->se; in get_rr_interval_fair()
11631 if (rq->cfs.load.weight) in get_rr_interval_fair()
11705 ng = rcu_dereference(p->numa_group); in show_numa_stats()
11707 if (p->numa_faults) { in show_numa_stats()
11708 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; in show_numa_stats()
11709 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; in show_numa_stats()
11712 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], in show_numa_stats()
11713 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; in show_numa_stats()
11743 return cfs_rq ? &cfs_rq->avg : NULL; in sched_trace_cfs_rq_avg()
11766 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1; in sched_trace_cfs_rq_cpu()
11773 return rq ? &rq->avg_rt : NULL; in sched_trace_rq_avg_rt()
11783 return rq ? &rq->avg_dl : NULL; in sched_trace_rq_avg_dl()
11793 return rq ? &rq->avg_irq : NULL; in sched_trace_rq_avg_irq()
11802 return rq ? cpu_of(rq) : -1; in sched_trace_rq_cpu()
11810 rq->cpu_capacity in sched_trace_rq_cpu_capacity()
11814 : -1; in sched_trace_rq_cpu_capacity()
11821 return rd ? rd->span : NULL; in sched_trace_rd_span()
11830 return rq ? rq->nr_running : -1; in sched_trace_rq_nr_running()