From e016cce088886f56617becc8fcc598a0114e4faa Mon Sep 17 00:00:00 2001 From: Peter Jung Date: Sat, 11 Mar 2023 18:44:19 +0100 Subject: [PATCH] bore-eevdf Signed-off-by: Peter Jung --- include/linux/sched.h | 5 ++ init/Kconfig | 20 +++++++ kernel/sched/core.c | 29 ++++++++++ kernel/sched/debug.c | 3 + kernel/sched/fair.c | 124 +++++++++++++++++++++++++++++++++++++++++- 5 files changed, 180 insertions(+), 1 deletion(-) diff --git a/include/linux/sched.h b/include/linux/sched.h index 764df627c243..f912da35db34 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -558,6 +558,11 @@ struct sched_entity { u64 sum_exec_runtime; u64 vruntime; u64 prev_sum_exec_runtime; +#ifdef CONFIG_SCHED_BORE + u64 prev_burst_time; + u64 burst_time; + u8 burst_score; +#endif // CONFIG_SCHED_BORE s64 lag; u64 slice; diff --git a/init/Kconfig b/init/Kconfig index 748a9491ca12..d10f1e6257cd 100644 --- a/init/Kconfig +++ b/init/Kconfig @@ -1318,6 +1318,26 @@ config CHECKPOINT_RESTORE If unsure, say N here. +config SCHED_BORE + bool "Burst-Oriented Response Enhancer" + default y + help + In Desktop and Mobile computing, one might prefer interactive + tasks to keep responsive no matter what they run in the background. + + Enabling this kernel feature modifies the scheduler to discriminate + tasks by their burst time (runtime since it last went sleeping or + yielding state) and prioritize those that run less bursty. + Such tasks usually include window compositor, widgets backend, + terminal emulator, video playback, games and so on. + With a little impact to scheduling fairness, it may improve + responsiveness especially under heavy background workload. + + You can turn it off by setting the sysctl kernel.sched_bore = 0. + Enabling this feature implies NO_GENTLE_FAIR_SLEEPERS by default. + + If unsure say Y here. + config SCHED_AUTOGROUP bool "Automatic process group scheduling" select CGROUPS diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 9db5f9ec9022..1f1e1f586407 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -4418,6 +4418,21 @@ int wake_up_state(struct task_struct *p, unsigned int state) return try_to_wake_up(p, state, 0); } +#ifdef CONFIG_SCHED_BORE +static inline void sched_fork_update_prev_burst(struct task_struct *p) +{ + struct task_struct *sib; + u32 cnt = 0; + u64 sum = 0, avg = 0; + list_for_each_entry(sib, &p->sibling, sibling) { + cnt++; + sum += sib->se.prev_burst_time >> 8; + } + if (cnt) avg = div_u64(sum, cnt) << 8; + if (p->se.prev_burst_time < avg) p->se.prev_burst_time = avg; +} +#endif // CONFIG_SCHED_BORE + /* * Perform scheduler related setup for a newly forked process p. * p is forked by current. @@ -4434,6 +4449,9 @@ static void __sched_fork(unsigned long clone_flags, struct task_struct *p) p->se.prev_sum_exec_runtime = 0; p->se.nr_migrations = 0; p->se.vruntime = 0; +#ifdef CONFIG_SCHED_BORE + p->se.burst_time = 0; +#endif // CONFIG_SCHED_BORE p->se.dur_avg = 0; p->se.prev_sleep_sum_runtime = 0; p->se.lag = 0; @@ -4664,6 +4682,10 @@ late_initcall(sched_core_sysctl_init); int sched_fork(unsigned long clone_flags, struct task_struct *p) { __sched_fork(clone_flags, p); +#ifdef CONFIG_SCHED_BORE + sched_fork_update_prev_burst(p); + p->se.burst_time = 0; +#endif // CONFIG_SCHED_BORE /* * We mark the process as NEW here. This guarantees that * nobody will actually run it, and a signal or other external @@ -9153,6 +9175,9 @@ void __init init_idle(struct task_struct *idle, int cpu) idle->__state = TASK_RUNNING; idle->se.exec_start = sched_clock(); +#ifdef CONFIG_SCHED_BORE + idle->se.prev_burst_time = 0; +#endif //CONFIG_SCHED_BORE /* * PF_KTHREAD should already be set at this point; regardless, make it * look like a proper per-CPU kthread. @@ -9820,6 +9845,10 @@ void __init sched_init(void) BUG_ON(&dl_sched_class != &stop_sched_class + 1); #endif +#ifdef CONFIG_SCHED_BORE + printk(KERN_INFO "BORE (Burst-Oriented Response Enhancer) CPU Scheduler modification for 1.7-eevdf2 by Masahito Suzuki"); +#endif // CONFIG_SCHED_BORE + wait_bit_init(); #ifdef CONFIG_FAIR_GROUP_SCHED diff --git a/kernel/sched/debug.c b/kernel/sched/debug.c index fe9edfa43f65..3672df7c1f6a 100644 --- a/kernel/sched/debug.c +++ b/kernel/sched/debug.c @@ -551,6 +551,9 @@ print_task(struct seq_file *m, struct rq *rq, struct task_struct *p) SPLIT_NS(schedstat_val_or_zero(p->stats.sum_sleep_runtime)), SPLIT_NS(schedstat_val_or_zero(p->stats.sum_block_runtime))); +#ifdef CONFIG_SCHED_BORE + SEQ_printf(m, " %2d", p->se.burst_score); +#endif #ifdef CONFIG_NUMA_BALANCING SEQ_printf(m, " %d %d", task_node(p), task_numa_group_id(p)); #endif diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index c40b775452bc..1e4ca5419a11 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -19,6 +19,9 @@ * * Adaptive scheduling granularity, math enhancements by Peter Zijlstra * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra + * + * Burst-Oriented Response Enhancer (BORE) CPU Scheduler + * Copyright (C) 2021 Masahito Suzuki */ #include #include @@ -141,6 +144,16 @@ static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; const_debug unsigned int sysctl_sched_migration_cost = 500000UL; +#ifdef CONFIG_SCHED_BORE +unsigned int __read_mostly sched_bore = 1; +unsigned int __read_mostly sched_burst_penalty_scale = 1280; +unsigned int __read_mostly sched_burst_granularity = 6; +unsigned int __read_mostly sched_burst_smoothness = 2; +static int three = 3; +static int sixty_four = 64; +static int maxval_12_bits = 4095; +#endif // CONFIG_SCHED_BORE + int sched_thermal_decay_shift; static int __init setup_sched_thermal_decay_shift(char *str) { @@ -204,6 +217,44 @@ static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536; #ifdef CONFIG_SYSCTL static struct ctl_table sched_fair_sysctls[] = { +#ifdef CONFIG_SCHED_BORE + { + .procname = "sched_bore", + .data = &sched_bore, + .maxlen = sizeof(unsigned int), + .mode = 0644, + .proc_handler = &proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = SYSCTL_ONE, + }, + { + .procname = "sched_burst_penalty_scale", + .data = &sched_burst_penalty_scale, + .maxlen = sizeof(unsigned int), + .mode = 0644, + .proc_handler = &proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &maxval_12_bits, + }, + { + .procname = "sched_burst_granularity", + .data = &sched_burst_granularity, + .maxlen = sizeof(unsigned int), + .mode = 0644, + .proc_handler = &proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &sixty_four, + }, + { + .procname = "sched_burst_smoothness", + .data = &sched_burst_smoothness, + .maxlen = sizeof(unsigned int), + .mode = 0644, + .proc_handler = &proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &three, + }, +#endif // CONFIG_SCHED_BORE { .procname = "sched_child_runs_first", .data = &sysctl_sched_child_runs_first, @@ -1182,6 +1233,39 @@ static void update_tg_load_avg(struct cfs_rq *cfs_rq) } #endif /* CONFIG_SMP */ +#ifdef CONFIG_SCHED_BORE +static inline void update_burst_score(struct sched_entity *se) { + u64 burst_time; + s32 bits; + u32 intgr, fdigs, dec10; + + burst_time = max(se->burst_time, se->prev_burst_time); + bits = fls64(burst_time); + intgr = max((u32)bits, sched_burst_granularity) - sched_burst_granularity; + fdigs = max(bits - 1, (s32)sched_burst_granularity); + dec10 = (intgr << 10) | (burst_time << (64 - fdigs) >> 54); + se->burst_score = min((u32)39, dec10 * sched_burst_penalty_scale >> 20); +} + +static u64 burst_scale(u64 delta, struct sched_entity *se) { + return mul_u64_u32_shr(delta, sched_prio_to_wmult[se->burst_score], 22); +} + +static u64 calc_delta_fair_bscale(u64 delta, struct sched_entity *se) { + return burst_scale(calc_delta_fair(delta, se), se); +} + +static inline u64 binary_smooth(u64 old, u64 new, unsigned int smoothness) { + return (new + old * ((1 << smoothness) - 1)) >> smoothness; +} + +static inline void reset_burst(struct sched_entity *se) { + se->prev_burst_time = binary_smooth( + se->prev_burst_time, se->burst_time, sched_burst_smoothness); + se->burst_time = 0; +} +#endif // CONFIG_SCHED_BORE + /* * Update the current task's runtime statistics. */ @@ -1211,6 +1295,13 @@ static void update_curr(struct cfs_rq *cfs_rq) curr->sum_exec_runtime += delta_exec; schedstat_add(cfs_rq->exec_clock, delta_exec); +#ifdef CONFIG_SCHED_BORE + curr->burst_time += delta_exec; + update_burst_score(curr); + if (sched_bore) + curr->vruntime += calc_delta_fair_bscale(delta_exec, curr); + else +#endif // CONFIG_SCHED_BORE curr->vruntime += calc_delta_fair(delta_exec, curr); /* * XXX: strictly: vd_i += N*r_i/w_i such that: vd_i > ve_i @@ -5283,6 +5374,11 @@ set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) se->prev_sum_exec_runtime = se->sum_exec_runtime; } +#ifdef CONFIG_SCHED_BORE +static int +wakeup_preempt_entity_bscale(struct sched_entity *curr, + struct sched_entity *se, bool do_scale); +#endif // CONFIG_SCHED_BORE static int wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); @@ -5330,7 +5426,13 @@ pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) se = second; } - if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) { +#ifdef CONFIG_SCHED_BORE + if (cfs_rq->next && wakeup_preempt_entity_bscale( + cfs_rq->next, left, sched_bore & 2) < 1) +#else // CONFIG_SCHED_BORE + if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) +#endif // CONFIG_SCHED_BORE + { /* * Someone really wants this to run. If it's not unfair, run it. */ @@ -6615,6 +6717,9 @@ static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) util_est_dequeue(&rq->cfs, p); for_each_sched_entity(se) { +#ifdef CONFIG_SCHED_BORE + if (task_sleep) reset_burst(se); +#endif // CONFIG_SCHED_BORE cfs_rq = cfs_rq_of(se); dequeue_entity(cfs_rq, se, flags); @@ -8070,7 +8175,12 @@ static unsigned long wakeup_gran(struct sched_entity *se) * */ static int +#ifdef CONFIG_SCHED_BORE +wakeup_preempt_entity_bscale(struct sched_entity *curr, + struct sched_entity *se, bool do_scale) +#else // CONFIG_SCHED_BORE wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) +#endif // CONFIG_SCHED_BORE { s64 gran, vdiff = curr->vruntime - se->vruntime; @@ -8078,11 +8188,20 @@ wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) return -1; gran = wakeup_gran(se); +#ifdef CONFIG_SCHED_BORE + if (do_scale) gran = burst_scale(gran, se); +#endif // CONFIG_SCHED_BORE if (vdiff > gran) return 1; return 0; } +#ifdef CONFIG_SCHED_BORE +static int wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) +{ + return wakeup_preempt_entity_bscale(curr, se, false); +} +#endif // CONFIG_SCHED_BORE static void set_last_buddy(struct sched_entity *se) { @@ -8430,6 +8549,9 @@ static void yield_task_fair(struct rq *rq) struct task_struct *curr = rq->curr; struct cfs_rq *cfs_rq = task_cfs_rq(curr); struct sched_entity *se = &curr->se; +#ifdef CONFIG_SCHED_BORE + reset_burst(se); +#endif // CONFIG_SCHED_BORE /* * Are we the only task in the tree? -- 2.40.0.rc2