2 * Only give sleepers 50% of their service deficit. This allows
3 * them to run sooner, but does not allow tons of sleepers to
4 * rip the spread apart.
6 SCHED_FEAT(GENTLE_FAIR_SLEEPERS, true)
9 * Place new tasks ahead so that they do not starve already running
12 SCHED_FEAT(START_DEBIT, true)
15 * Prefer to schedule the task we woke last (assuming it failed
16 * wakeup-preemption), since its likely going to consume data we
17 * touched, increases cache locality.
19 SCHED_FEAT(NEXT_BUDDY, false)
22 * Prefer to schedule the task that ran last (when we did
23 * wake-preempt) as that likely will touch the same data, increases
26 SCHED_FEAT(LAST_BUDDY, true)
29 * Consider buddies to be cache hot, decreases the likelyness of a
30 * cache buddy being migrated away, increases cache locality.
32 SCHED_FEAT(CACHE_HOT_BUDDY, true)
35 * Allow wakeup-time preemption of the current task:
37 SCHED_FEAT(WAKEUP_PREEMPTION, true)
40 * Use arch dependent cpu capacity functions
42 SCHED_FEAT(ARCH_CAPACITY, true)
44 SCHED_FEAT(HRTICK, false)
45 SCHED_FEAT(DOUBLE_TICK, false)
46 SCHED_FEAT(LB_BIAS, true)
49 * Decrement CPU capacity based on time not spent running tasks
51 SCHED_FEAT(NONTASK_CAPACITY, true)
54 * Queue remote wakeups on the target CPU and process them
55 * using the scheduler IPI. Reduces rq->lock contention/bounces.
57 SCHED_FEAT(TTWU_QUEUE, true)
59 #ifdef HAVE_RT_PUSH_IPI
61 * In order to avoid a thundering herd attack of CPUs that are
62 * lowering their priorities at the same time, and there being
63 * a single CPU that has an RT task that can migrate and is waiting
64 * to run, where the other CPUs will try to take that CPUs
65 * rq lock and possibly create a large contention, sending an
66 * IPI to that CPU and let that CPU push the RT task to where
67 * it should go may be a better scenario.
69 SCHED_FEAT(RT_PUSH_IPI, true)
72 SCHED_FEAT(FORCE_SD_OVERLAP, false)
73 SCHED_FEAT(RT_RUNTIME_SHARE, true)
74 SCHED_FEAT(LB_MIN, false)
77 * Apply the automatic NUMA scheduling policy. Enabled automatically
78 * at runtime if running on a NUMA machine. Can be controlled via
81 #ifdef CONFIG_NUMA_BALANCING
84 * NUMA will favor moving tasks towards nodes where a higher number of
85 * hinting faults are recorded during active load balancing. It will
86 * resist moving tasks towards nodes where a lower number of hinting
87 * faults have been recorded.
89 SCHED_FEAT(NUMA, true)