From b6a7058a13f345d5aa5426466f9104da43d47ce4 Mon Sep 17 00:00:00 2001 From: Piotr Gorski Date: Tue, 5 Dec 2023 15:49:10 +0100 Subject: [PATCH] bore-cachy Signed-off-by: Piotr Gorski --- include/linux/sched.h | 10 ++ init/Kconfig | 19 ++++ kernel/sched/core.c | 128 ++++++++++++++++++++++++ kernel/sched/debug.c | 3 + kernel/sched/fair.c | 216 +++++++++++++++++++++++++++++++++++++--- kernel/sched/features.h | 4 + 6 files changed, 366 insertions(+), 14 deletions(-) diff --git a/include/linux/sched.h b/include/linux/sched.h index 77f01ac38..01f2839ad 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -559,6 +559,16 @@ struct sched_entity { u64 sum_exec_runtime; u64 prev_sum_exec_runtime; u64 vruntime; +#ifdef CONFIG_SCHED_BORE + u64 burst_time; + u8 prev_burst_penalty; + u8 curr_burst_penalty; + u8 burst_penalty; + u8 slice_score; + u8 child_burst; + u16 child_burst_cnt; + u64 child_burst_last_cached; +#endif // CONFIG_SCHED_BORE s64 vlag; u64 slice; diff --git a/init/Kconfig b/init/Kconfig index 9dee4c100..49d343e97 100644 --- a/init/Kconfig +++ b/init/Kconfig @@ -1278,6 +1278,25 @@ 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. + + 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 a854b7183..a98cfa7ab 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -4488,6 +4488,123 @@ int wake_up_state(struct task_struct *p, unsigned int state) return try_to_wake_up(p, state, 0); } +#ifdef CONFIG_SCHED_BORE +extern bool sched_bore; +extern u8 sched_burst_fork_atavistic; +extern uint sched_burst_cache_lifetime; + +void __init sched_init_bore(void) { + init_task.se.burst_time = 0; + init_task.se.prev_burst_penalty = 0; + init_task.se.curr_burst_penalty = 0; + init_task.se.burst_penalty = 0; + init_task.se.slice_score = 0; + init_task.se.child_burst_last_cached = 0; +} + +void inline sched_fork_bore(struct task_struct *p) { + p->se.burst_time = 0; + p->se.curr_burst_penalty = 0; + p->se.slice_score = 0; + p->se.child_burst_last_cached = 0; +} + +static u32 count_child_tasks(struct task_struct *p) { + struct task_struct *child; + u32 cnt = 0; + list_for_each_entry(child, &p->children, sibling) {cnt++;} + return cnt; +} + +static inline bool child_burst_cache_expired(struct task_struct *p, u64 now) { + return (p->se.child_burst_last_cached + sched_burst_cache_lifetime < now); +} + +static void __update_child_burst_cache( + struct task_struct *p, u32 cnt, u32 sum, u64 now) { + u8 avg = 0; + if (cnt) avg = sum / cnt; + p->se.child_burst = max(avg, p->se.burst_penalty); + p->se.child_burst_cnt = cnt; + p->se.child_burst_last_cached = now; +} + +static void update_child_burst_cache(struct task_struct *p, u64 now) { + struct task_struct *child; + u32 cnt = 0; + u32 sum = 0; + + list_for_each_entry(child, &p->children, sibling) { + if (child->sched_class != &fair_sched_class) continue; + cnt++; + sum += child->se.burst_penalty; + } + + __update_child_burst_cache(p, cnt, sum, now); +} + +static void update_child_burst_cache_atavistic( + struct task_struct *p, u64 now, u32 depth, u32 *acnt, u32 *asum) { + struct task_struct *child, *dec; + u32 cnt = 0, dcnt = 0; + u32 sum = 0; + + list_for_each_entry(child, &p->children, sibling) { + dec = child; + while ((dcnt = count_child_tasks(dec)) == 1) + dec = list_first_entry(&dec->children, struct task_struct, sibling); + + if (!dcnt || !depth) { + if (dec->sched_class != &fair_sched_class) continue; + cnt++; + sum += dec->se.burst_penalty; + continue; + } + if (!child_burst_cache_expired(dec, now)) { + cnt += dec->se.child_burst_cnt; + sum += (u32)dec->se.child_burst * dec->se.child_burst_cnt; + continue; + } + update_child_burst_cache_atavistic(dec, now, depth - 1, &cnt, &sum); + } + + __update_child_burst_cache(p, cnt, sum, now); + *acnt += cnt; + *asum += sum; +} + +static void sched_post_fork_bore(struct task_struct *p) { + struct sched_entity *se = &p->se; + struct task_struct *anc; + u64 now; + u32 cnt = 0, sum = 0, depth; + u8 burst_cache; + + if (likely(sched_bore)) { + now = ktime_get_ns(); + read_lock(&tasklist_lock); + + anc = p->real_parent; + depth = sched_burst_fork_atavistic; + if (likely(depth)) { + while ((anc->real_parent != anc) && (count_child_tasks(anc) == 1)) + anc = anc->real_parent; + if (child_burst_cache_expired(anc, now)) + update_child_burst_cache_atavistic( + anc, now, depth - 1, &cnt, &sum); + } else + if (child_burst_cache_expired(anc, now)) + update_child_burst_cache(anc, now); + + burst_cache = anc->se.child_burst; + + read_unlock(&tasklist_lock); + se->prev_burst_penalty = max(se->prev_burst_penalty, burst_cache); + } + se->burst_penalty = se->prev_burst_penalty; +} +#endif // CONFIG_SCHED_BORE + /* * Perform scheduler related setup for a newly forked process p. * p is forked by current. @@ -4504,6 +4621,9 @@ 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 + sched_fork_bore(p); +#endif // CONFIG_SCHED_BORE p->se.vlag = 0; p->se.slice = sysctl_sched_base_slice; INIT_LIST_HEAD(&p->se.group_node); @@ -4823,6 +4943,9 @@ void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs) void sched_post_fork(struct task_struct *p) { +#ifdef CONFIG_SCHED_BORE + sched_post_fork_bore(p); +#endif // CONFIG_SCHED_BORE uclamp_post_fork(p); } @@ -9922,6 +10045,11 @@ void __init sched_init(void) BUG_ON(&dl_sched_class != &stop_sched_class + 1); #endif +#ifdef CONFIG_SCHED_BORE + sched_init_bore(); + printk(KERN_INFO "BORE (Burst-Oriented Response Enhancer) CPU Scheduler modification 3.5.7 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 4c3d0d9f3..e37fdfad1 100644 --- a/kernel/sched/debug.c +++ b/kernel/sched/debug.c @@ -595,6 +595,9 @@ 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.slice_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 fa9fff0f9..5e4f0ccff 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) 2021-2023 Masahito Suzuki */ #include #include @@ -66,17 +69,28 @@ * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus * - * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) + * (BORE default SCHED_TUNABLESCALING_NONE = *1 constant) + * (EEVDF default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) */ +#ifdef CONFIG_SCHED_BORE +unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE; +#else // CONFIG_SCHED_BORE unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG; +#endif // CONFIG_SCHED_BORE /* * Minimal preemption granularity for CPU-bound tasks: * - * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) + * (BORE default: 3 msec constant, units: nanoseconds) + * (EEVDF default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) */ +#ifdef CONFIG_SCHED_BORE +unsigned int sysctl_sched_base_slice = 3000000ULL; +static unsigned int normalized_sysctl_sched_base_slice = 3000000ULL; +#else // CONFIG_SCHED_BORE unsigned int sysctl_sched_base_slice = 750000ULL; static unsigned int normalized_sysctl_sched_base_slice = 750000ULL; +#endif // CONFIG_SCHED_BORE /* * After fork, child runs first. If set to 0 (default) then @@ -86,6 +100,68 @@ unsigned int sysctl_sched_child_runs_first __read_mostly; const_debug unsigned int sysctl_sched_migration_cost = 500000UL; +#ifdef CONFIG_SCHED_BORE +bool __read_mostly sched_bore = 1; +bool __read_mostly sched_burst_score_rounding = 0; +bool __read_mostly sched_burst_smoothness_long = 1; +bool __read_mostly sched_burst_smoothness_short = 0; +u8 __read_mostly sched_burst_fork_atavistic = 2; +u8 __read_mostly sched_burst_penalty_offset = 22; +uint __read_mostly sched_burst_penalty_scale = 1280; +uint __read_mostly sched_burst_cache_lifetime = 60000000; +static u8 sixty_four = 64; +static uint maxval_12_bits = 4095; + +#define MAX_BURST_PENALTY (39U <<2) + +static inline u32 log2plus1_u64_u32f8(u64 v) { + u32 msb = fls64(v); + s32 excess_bits = msb - 9; + u8 fractional = (0 <= excess_bits)? v >> excess_bits: v << -excess_bits; + return msb << 8 | fractional; +} + +static inline u32 calc_burst_penalty(u64 burst_time) { + u32 greed, tolerance, penalty, scaled_penalty; + + greed = log2plus1_u64_u32f8(burst_time); + tolerance = sched_burst_penalty_offset << 8; + penalty = max(0, (s32)greed - (s32)tolerance); + scaled_penalty = penalty * sched_burst_penalty_scale >> 16; + + return min(MAX_BURST_PENALTY, scaled_penalty); +} + +static inline void update_burst_penalty(struct sched_entity *se) { + se->curr_burst_penalty = calc_burst_penalty(se->burst_time); + se->burst_penalty = max(se->prev_burst_penalty, se->curr_burst_penalty); +} + +static inline void update_slice_score(struct sched_entity *se) { + u32 penalty = se->burst_penalty; + if (sched_burst_score_rounding) penalty += 0x2U; + se->slice_score = penalty >> 2; +} + +static inline u64 scale_slice(u64 delta, struct sched_entity *se) { + return mul_u64_u32_shr(delta, sched_prio_to_wmult[se->slice_score], 22); +} + +static inline u32 binary_smooth(u32 new, u32 old) { + int increment = new - old; + return (0 <= increment)? + old + ( increment >> (int)sched_burst_smoothness_long): + old - (-increment >> (int)sched_burst_smoothness_short); +} + +static void restart_burst(struct sched_entity *se) { + se->burst_penalty = se->prev_burst_penalty = + binary_smooth(se->curr_burst_penalty, se->prev_burst_penalty); + se->curr_burst_penalty = 0; + se->burst_time = 0; +} +#endif // CONFIG_SCHED_BORE + int sched_thermal_decay_shift; static int __init setup_sched_thermal_decay_shift(char *str) { @@ -145,6 +221,70 @@ 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(bool), + .mode = 0644, + .proc_handler = &proc_dobool, + }, + { + .procname = "sched_burst_cache_lifetime", + .data = &sched_burst_cache_lifetime, + .maxlen = sizeof(uint), + .mode = 0644, + .proc_handler = proc_douintvec, + }, + { + .procname = "sched_burst_fork_atavistic", + .data = &sched_burst_fork_atavistic, + .maxlen = sizeof(u8), + .mode = 0644, + .proc_handler = &proc_dou8vec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = SYSCTL_THREE, + }, + { + .procname = "sched_burst_penalty_offset", + .data = &sched_burst_penalty_offset, + .maxlen = sizeof(u8), + .mode = 0644, + .proc_handler = &proc_dou8vec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &sixty_four, + }, + { + .procname = "sched_burst_penalty_scale", + .data = &sched_burst_penalty_scale, + .maxlen = sizeof(uint), + .mode = 0644, + .proc_handler = &proc_douintvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &maxval_12_bits, + }, + { + .procname = "sched_burst_score_rounding", + .data = &sched_burst_score_rounding, + .maxlen = sizeof(bool), + .mode = 0644, + .proc_handler = &proc_dobool, + }, + { + .procname = "sched_burst_smoothness_long", + .data = &sched_burst_smoothness_long, + .maxlen = sizeof(bool), + .mode = 0644, + .proc_handler = &proc_dobool, + }, + { + .procname = "sched_burst_smoothness_short", + .data = &sched_burst_smoothness_short, + .maxlen = sizeof(bool), + .mode = 0644, + .proc_handler = &proc_dobool, + }, +#endif // CONFIG_SCHED_BORE { .procname = "sched_child_runs_first", .data = &sysctl_sched_child_runs_first, @@ -313,6 +453,9 @@ static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) if (unlikely(se->load.weight != NICE_0_LOAD)) delta = __calc_delta(delta, NICE_0_LOAD, &se->load); +#ifdef CONFIG_SCHED_BORE + if (likely(sched_bore)) delta = scale_slice(delta, se); +#endif // CONFIG_SCHED_BORE return delta; } @@ -668,7 +811,7 @@ void avg_vruntime_update(struct cfs_rq *cfs_rq, s64 delta) * Specifically: avg_runtime() + 0 must result in entity_eligible() := true * For this to be so, the result of this function must have a left bias. */ -u64 avg_vruntime(struct cfs_rq *cfs_rq) +static u64 avg_key(struct cfs_rq *cfs_rq) { struct sched_entity *curr = cfs_rq->curr; s64 avg = cfs_rq->avg_vruntime; @@ -688,7 +831,11 @@ u64 avg_vruntime(struct cfs_rq *cfs_rq) avg = div_s64(avg, load); } - return cfs_rq->min_vruntime + avg; + return avg; +} + +inline u64 avg_vruntime(struct cfs_rq *cfs_rq) { + return cfs_rq->min_vruntime + avg_key(cfs_rq); } /* @@ -709,13 +856,8 @@ u64 avg_vruntime(struct cfs_rq *cfs_rq) */ static void update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se) { - s64 lag, limit; - SCHED_WARN_ON(!se->on_rq); - lag = avg_vruntime(cfs_rq) - se->vruntime; - - limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se); - se->vlag = clamp(lag, -limit, limit); + se->vlag = avg_vruntime(cfs_rq) - se->vruntime; } /* @@ -981,7 +1123,6 @@ static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq) return se; } -#ifdef CONFIG_SCHED_DEBUG struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) { struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); @@ -995,6 +1136,7 @@ struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) /************************************************************** * Scheduling class statistics methods: */ +#ifdef CONFIG_SCHED_DEBUG #ifdef CONFIG_SMP int sched_update_scaling(void) { @@ -1031,6 +1173,9 @@ static void update_deadline(struct cfs_rq *cfs_rq, struct sched_entity *se) /* * EEVDF: vd_i = ve_i + r_i / w_i */ +#ifdef CONFIG_SCHED_BORE + update_slice_score(se); +#endif // CONFIG_SCHED_BORE se->deadline = se->vruntime + calc_delta_fair(se->slice, se); /* @@ -1173,7 +1318,11 @@ static void update_curr(struct cfs_rq *cfs_rq) curr->sum_exec_runtime += delta_exec; schedstat_add(cfs_rq->exec_clock, delta_exec); - curr->vruntime += calc_delta_fair(delta_exec, curr); +#ifdef CONFIG_SCHED_BORE + curr->burst_time += delta_exec; + update_burst_penalty(curr); +#endif // CONFIG_SCHED_BORE + curr->vruntime += max(1ULL, calc_delta_fair(delta_exec, curr)); update_deadline(cfs_rq, curr); update_min_vruntime(cfs_rq); @@ -5066,6 +5215,9 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) s64 lag = 0; se->slice = sysctl_sched_base_slice; +#ifdef CONFIG_SCHED_BORE + update_slice_score(se); +#endif // CONFIG_SCHED_BORE vslice = calc_delta_fair(se->slice, se); /* @@ -5080,7 +5232,13 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) struct sched_entity *curr = cfs_rq->curr; unsigned long load; - lag = se->vlag; + u64 slice = se->slice; +#ifdef CONFIG_SCHED_BORE + if (unlikely(!sched_bore)) +#endif // CONFIG_SCHED_BORE + slice *= 2; + s64 limit = calc_delta_fair(max_t(u64, slice, TICK_NSEC), se); + lag = clamp(se->vlag, -limit, limit); /* * If we want to place a task and preserve lag, we have to @@ -5142,6 +5300,21 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) if (WARN_ON_ONCE(!load)) load = 1; lag = div_s64(lag, load); + +#ifdef CONFIG_SCHED_BORE + if (flags & ENQUEUE_MIGRATED && likely(sched_bore)) { + s64 left_vruntime = vruntime, right_vruntime = vruntime; + struct sched_entity *first = __pick_first_entity(cfs_rq), + *last = __pick_last_entity(cfs_rq); + + if (first) left_vruntime = first->vruntime; + if (last) right_vruntime = last->vruntime; + + lag = clamp(lag, + (s64)vruntime - right_vruntime, + (s64)vruntime - left_vruntime); + } +#endif // CONFIG_SCHED_BORE } se->vruntime = vruntime - lag; @@ -6698,6 +6871,12 @@ static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) bool was_sched_idle = sched_idle_rq(rq); util_est_dequeue(&rq->cfs, p); +#ifdef CONFIG_SCHED_BORE + if (task_sleep) { + update_curr(cfs_rq_of(se)); + restart_burst(se); + } +#endif // CONFIG_SCHED_BORE for_each_sched_entity(se) { cfs_rq = cfs_rq_of(se); @@ -8429,8 +8608,13 @@ static void yield_task_fair(struct rq *rq) /* * Are we the only task in the tree? */ - if (unlikely(rq->nr_running == 1)) + if (unlikely(rq->nr_running == 1)) { +#ifdef CONFIG_SCHED_BORE + restart_burst(se); + update_slice_score(se); +#endif // CONFIG_SCHED_BORE return; + } clear_buddies(cfs_rq, se); @@ -8439,6 +8623,10 @@ static void yield_task_fair(struct rq *rq) * Update run-time statistics of the 'current'. */ update_curr(cfs_rq); +#ifdef CONFIG_SCHED_BORE + restart_burst(se); + update_slice_score(se); +#endif // CONFIG_SCHED_BORE /* * Tell update_rq_clock() that we've just updated, * so we don't do microscopic update in schedule() diff --git a/kernel/sched/features.h b/kernel/sched/features.h index f77016823..a2e09c04f 100644 --- a/kernel/sched/features.h +++ b/kernel/sched/features.h @@ -6,7 +6,11 @@ */ SCHED_FEAT(PLACE_LAG, true) SCHED_FEAT(PLACE_DEADLINE_INITIAL, true) +#ifdef CONFIG_SCHED_BORE +SCHED_FEAT(RUN_TO_PARITY, false) +#else // CONFIG_SCHED_BORE SCHED_FEAT(RUN_TO_PARITY, true) +#endif // CONFIG_SCHED_BORE /* * Prefer to schedule the task we woke last (assuming it failed -- 2.43.0.rc2