From 32c617afc05751783be3eb0f5a1d15e31dfc7919 Mon Sep 17 00:00:00 2001 From: Piotr Gorski Date: Thu, 8 Jun 2023 11:14:23 +0200 Subject: [PATCH] bore-cachy Signed-off-by: Piotr Gorski --- include/linux/sched.h | 10 +++ init/Kconfig | 20 ++++++ kernel/sched/core.c | 62 ++++++++++++++++++ kernel/sched/debug.c | 3 + kernel/sched/fair.c | 136 ++++++++++++++++++++++++++++++++++++++++ kernel/sched/features.h | 8 +++ 6 files changed, 239 insertions(+) diff --git a/include/linux/sched.h b/include/linux/sched.h index eed5d65b8..38fbebe4d 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -557,6 +557,12 @@ 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; + u64 max_burst_time; + u8 penalty_score; +#endif // CONFIG_SCHED_BORE u64 nr_migrations; @@ -985,6 +991,10 @@ struct task_struct { struct list_head children; struct list_head sibling; struct task_struct *group_leader; +#ifdef CONFIG_SCHED_BORE + u64 child_burst_cache; + u64 child_burst_last_cached; +#endif // CONFIG_SCHED_BORE /* * 'ptraced' is the list of tasks this task is using ptrace() on. diff --git a/init/Kconfig b/init/Kconfig index 0147b4a33..4ab7e154b 100644 --- a/init/Kconfig +++ b/init/Kconfig @@ -1290,6 +1290,26 @@ 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 bcb3a7e68..a0f227344 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -4484,6 +4484,57 @@ int wake_up_state(struct task_struct *p, unsigned int state) return try_to_wake_up(p, state, 0); } +#ifdef CONFIG_SCHED_BORE +#define CHILD_BURST_CUTOFF_BITS 9 +extern unsigned int sched_burst_cache_lifetime; + +void __init sched_init_bore(void) { + init_task.child_burst_cache = 0; + init_task.child_burst_last_cached = 0; + init_task.se.prev_burst_time = 0; + init_task.se.burst_time = 0; + init_task.se.max_burst_time = 0; +} + +void inline __sched_fork_bore(struct task_struct *p) { + p->child_burst_cache = 0; + p->child_burst_last_cached = 0; + p->se.burst_time = 0; +} + +static inline void update_task_child_burst_time_cache(struct task_struct *p) { + u64 sum = 0, avg_burst_time = 0; + u32 cnt = 0; + struct task_struct *child; + + read_lock(&tasklist_lock); + list_for_each_entry(child, &p->children, sibling) { + cnt++; + sum += child->se.max_burst_time >> CHILD_BURST_CUTOFF_BITS; + } + read_unlock(&tasklist_lock); + + if (cnt) avg_burst_time = div_u64(sum, cnt) << CHILD_BURST_CUTOFF_BITS; + p->child_burst_cache = max(avg_burst_time, p->se.max_burst_time); +} + +static void update_task_initial_burst_time(struct task_struct *task) { + struct sched_entity *se = &task->se; + struct task_struct *par = task->real_parent; + u64 ktime = ktime_to_ns(ktime_get()); + + if (likely(par)) { + if (par->child_burst_last_cached + sched_burst_cache_lifetime < ktime) { + par->child_burst_last_cached = ktime; + update_task_child_burst_time_cache(par); + } + se->prev_burst_time = max(se->prev_burst_time, par->child_burst_cache); + } + + se->max_burst_time = se->prev_burst_time; +} +#endif // CONFIG_SCHED_BORE + /* * Perform scheduler related setup for a newly forked process p. * p is forked by current. @@ -4500,6 +4551,9 @@ 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 INIT_LIST_HEAD(&p->se.group_node); #ifdef CONFIG_FAIR_GROUP_SCHED @@ -4726,6 +4780,9 @@ 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 + update_task_initial_burst_time(p); +#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 @@ -9922,6 +9979,11 @@ 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 2.4.0 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 066ff1c8a..4bc07d405 100644 --- a/kernel/sched/debug.c +++ b/kernel/sched/debug.c @@ -593,6 +593,9 @@ 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.penalty_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 9fe8288b1..ac29ac350 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 @@ -140,6 +143,61 @@ 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_cache_lifetime = 15000000; +unsigned int __read_mostly sched_burst_penalty_offset = 12; +unsigned int __read_mostly sched_burst_penalty_scale = 1292; +unsigned int __read_mostly sched_burst_smoothness = 1; +static int three = 3; +static int sixty_four = 64; +static int maxval_12_bits = 4095; + +#define FIXED_SHIFT 10 +#define FIXED_ONE (1 << FIXED_SHIFT) +typedef u32 fixed; + +static void update_burst_score(struct sched_entity *se) { + u64 burst_time = se->max_burst_time; + + int msb = fls64(burst_time); + fixed integer_part = msb << FIXED_SHIFT; + fixed fractional_part = burst_time << (64 - msb) << 1 >> (64 - FIXED_SHIFT); + fixed greed = integer_part | fractional_part; + + fixed tolerance = sched_burst_penalty_offset << FIXED_SHIFT; + fixed penalty = max(0, (s32)greed - (s32)tolerance); + fixed scaled_penalty = penalty * sched_burst_penalty_scale >> 10; + + u8 score = min(39U, scaled_penalty >> FIXED_SHIFT); + se->penalty_score = score; +} + +static inline u64 penalty_scale(u64 delta, struct sched_entity *se) { + return mul_u64_u32_shr(delta, sched_prio_to_wmult[se->penalty_score], 22); +} + +static inline u64 __binary_smooth(u64 new, u64 old, unsigned int smoothness) { + return (new + old * ((1 << smoothness) - 1)) >> smoothness; +} + +void restart_burst(struct sched_entity *se) { + se->max_burst_time = se->prev_burst_time = __binary_smooth( + se->burst_time, se->prev_burst_time, sched_burst_smoothness); + se->burst_time = 0; +} + +#define calc_delta_fair(delta, se) __calc_delta_fair(delta, se, true) +#define calc_delta_fair_unscaled(delta, se) __calc_delta_fair(delta, se, false) +static inline u64 +__calc_delta_fair(u64 delta, struct sched_entity *se, bool bscale); + +static s64 wakeup_preempt_backstep_delta(u64 rtime, struct sched_entity *se) { + u64 delta = calc_delta_fair_unscaled(rtime, se); + return delta - penalty_scale(delta, se); +} +#endif // CONFIG_SCHED_BORE + int sched_thermal_decay_shift; static int __init setup_sched_thermal_decay_shift(char *str) { @@ -203,6 +261,51 @@ 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_cache_lifetime", + .data = &sched_burst_cache_lifetime, + .maxlen = sizeof(unsigned int), + .mode = 0644, + .proc_handler = proc_dointvec, + }, + { + .procname = "sched_burst_penalty_offset", + .data = &sched_burst_penalty_offset, + .maxlen = sizeof(unsigned int), + .mode = 0644, + .proc_handler = &proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &sixty_four, + }, + { + .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_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, @@ -724,11 +827,19 @@ int sched_update_scaling(void) /* * delta /= w */ +#ifdef CONFIG_SCHED_BORE +static inline u64 +__calc_delta_fair(u64 delta, struct sched_entity *se, bool bscale) +#else // CONFIG_SCHED_BORE static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) +#endif // CONFIG_SCHED_BORE { if (unlikely(se->load.weight != NICE_0_LOAD)) delta = __calc_delta(delta, NICE_0_LOAD, &se->load); +#ifdef CONFIG_SCHED_BORE + if (bscale && sched_bore) delta = penalty_scale(delta, se); +#endif // CONFIG_SCHED_BORE return delta; } @@ -938,6 +1049,14 @@ 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; + curr->max_burst_time = max(curr->max_burst_time, curr->burst_time); + update_burst_score(curr); + if (sched_bore) + curr->vruntime += penalty_scale(calc_delta_fair(delta_exec, curr), curr); + else +#endif // CONFIG_SCHED_BORE curr->vruntime += calc_delta_fair(delta_exec, curr); update_min_vruntime(cfs_rq); @@ -6410,6 +6529,9 @@ 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) restart_burst(se); +#endif // CONFIG_SCHED_BORE cfs_rq = cfs_rq_of(se); dequeue_entity(cfs_rq, se, flags); @@ -7844,7 +7966,11 @@ static unsigned long wakeup_gran(struct sched_entity *se) * This is especially important for buddies when the leftmost * task is higher priority than the buddy. */ +#ifdef CONFIG_SCHED_BORE + return calc_delta_fair_unscaled(gran, se); +#else // CONFIG_SCHED_BORE return calc_delta_fair(gran, se); +#endif // CONFIG_SCHED_BORE } /* @@ -7865,6 +7991,13 @@ static int wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) { s64 gran, vdiff = curr->vruntime - se->vruntime; +#ifdef CONFIG_SCHED_BORE + if (sched_bore) { + u64 rtime = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; + vdiff += wakeup_preempt_backstep_delta(rtime, curr) + - wakeup_preempt_backstep_delta(rtime, se); + } +#endif // CONFIG_SCHED_BORE if (vdiff <= 0) return -1; @@ -8210,6 +8343,9 @@ 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 + restart_burst(se); +#endif // CONFIG_SCHED_BORE /* * Are we the only task in the tree? diff --git a/kernel/sched/features.h b/kernel/sched/features.h index 9e390eb82..696ea7081 100644 --- a/kernel/sched/features.h +++ b/kernel/sched/features.h @@ -4,7 +4,11 @@ * them to run sooner, but does not allow tons of sleepers to * rip the spread apart. */ +#ifdef CONFIG_SCHED_BORE +SCHED_FEAT(GENTLE_FAIR_SLEEPERS, false) +#else // CONFIG_SCHED_BORE SCHED_FEAT(GENTLE_FAIR_SLEEPERS, true) +#endif // CONFIG_SCHED_BORE /* * Place new tasks ahead so that they do not starve already running @@ -17,7 +21,11 @@ SCHED_FEAT(START_DEBIT, true) * wakeup-preemption), since its likely going to consume data we * touched, increases cache locality. */ +#ifdef CONFIG_SCHED_BORE +SCHED_FEAT(NEXT_BUDDY, true) +#else // CONFIG_SCHED_BORE SCHED_FEAT(NEXT_BUDDY, false) +#endif // CONFIG_SCHED_BORE /* * Prefer to schedule the task that ran last (when we did -- 2.41.0.rc2