From f353b9eb23586e55b99a6bfe7da9563be5fcca29 Mon Sep 17 00:00:00 2001 From: Piotr Gorski Date: Sat, 12 Aug 2023 21:05:20 +0200 Subject: [PATCH] bore-eevdf Signed-off-by: Piotr Gorski --- include/linux/sched.h | 29 ++++++ init/Kconfig | 20 ++++ kernel/sched/core.c | 122 +++++++++++++++++++++++ kernel/sched/debug.c | 4 + kernel/sched/fair.c | 219 +++++++++++++++++++++++++++++++++++++++--- kernel/sched/sched.h | 1 + 6 files changed, 384 insertions(+), 11 deletions(-) diff --git a/include/linux/sched.h b/include/linux/sched.h index c940c4dc8..984931de0 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -545,6 +545,24 @@ struct sched_statistics { #endif /* CONFIG_SCHEDSTATS */ } ____cacheline_aligned; +#ifdef CONFIG_SCHED_BORE +typedef union { + u16 u16; + s16 s16; + u8 u8[2]; + s8 s8[2]; +} x16; + +typedef union { + u32 u32; + s32 s32; + u16 u16[2]; + s16 s16[2]; + u8 u8[4]; + s8 s8[4]; +} x32; +#endif // CONFIG_SCHED_BORE + struct sched_entity { /* For load-balancing: */ struct load_weight load; @@ -559,6 +577,12 @@ struct sched_entity { u64 sum_exec_runtime; u64 prev_sum_exec_runtime; u64 vruntime; +#ifdef CONFIG_SCHED_BORE + u64 burst_time; + u16 prev_burst_penalty; + u16 curr_burst_penalty; + u16 burst_penalty; +#endif // CONFIG_SCHED_BORE s64 vlag; u64 slice; @@ -990,6 +1014,11 @@ struct task_struct { struct list_head children; struct list_head sibling; struct task_struct *group_leader; +#ifdef CONFIG_SCHED_BORE + u16 child_burst_cache; + u16 child_burst_count_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 71755cc8e..c697be79e 100644 --- a/init/Kconfig +++ b/init/Kconfig @@ -1277,6 +1277,26 @@ 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 aff81e124..a4eba9e47 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -4491,6 +4491,117 @@ int wake_up_state(struct task_struct *p, unsigned int state) return try_to_wake_up(p, state, 0); } +#ifdef CONFIG_SCHED_BORE +extern unsigned int sched_burst_cache_lifetime; +extern unsigned int sched_bore; +extern unsigned int sched_burst_fork_atavistic; + +void __init sched_init_bore(void) { + init_task.child_burst_cache = 0; + init_task.child_burst_count_cache = 0; + init_task.child_burst_last_cached = 0; + 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; +} + +void inline sched_fork_bore(struct task_struct *p) { + p->child_burst_cache = 0; + p->child_burst_count_cache = 0; + p->child_burst_last_cached = 0; + p->se.burst_time = 0; + p->se.curr_burst_penalty = 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->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) { + u16 avg = 0; + if (cnt) avg = DIV_ROUND_CLOSEST(sum, cnt); + p->child_burst_cache = max(avg, p->se.burst_penalty); + p->child_burst_count_cache = cnt; + p->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) { + 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) { + cnt++; + sum += dec->se.burst_penalty; + } else { + if (child_burst_cache_expired(dec, now)) + update_child_burst_cache_atavistic(dec, now, depth - 1, &cnt, &sum); + else { + cnt += dec->child_burst_count_cache; + sum += (u32)dec->child_burst_cache * dec->child_burst_count_cache; + } + } + } + + __update_child_burst_cache(p, cnt, sum, now); + *acnt += cnt; + *asum += sum; +} + +static void fork_burst_penalty(struct task_struct *p) { + struct sched_entity *se = &p->se; + struct task_struct *anc = p->real_parent; + u64 now = ktime_get_ns(); + u32 cnt = 0; + u32 sum = 0; + + read_lock(&tasklist_lock); + + if (likely(sched_bore) && likely(sched_burst_fork_atavistic)) { + 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, sched_burst_fork_atavistic - 1, &cnt, &sum); + } else + if (child_burst_cache_expired(anc, now)) + update_child_burst_cache(anc, now); + + read_unlock(&tasklist_lock); + + se->burst_penalty = se->prev_burst_penalty = + max(se->prev_burst_penalty, anc->child_burst_cache); +} +#endif // CONFIG_SCHED_BORE + /* * Perform scheduler related setup for a newly forked process p. * p is forked by current. @@ -4507,6 +4618,9 @@ 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; INIT_LIST_HEAD(&p->se.group_node); @@ -4828,6 +4942,9 @@ 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 + fork_burst_penalty(p); +#endif // CONFIG_SCHED_BORE uclamp_post_fork(p); } @@ -9954,6 +10071,11 @@ 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 3.1.2 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 e7e83181f..ff41a524c 100644 --- a/kernel/sched/debug.c +++ b/kernel/sched/debug.c @@ -348,6 +348,7 @@ static __init int sched_init_debug(void) #endif debugfs_create_u32("base_slice_ns", 0644, debugfs_sched, &sysctl_sched_base_slice); + debugfs_create_u32("wakeup_granularity_ns", 0644, debugfs_sched, &sysctl_sched_wakeup_granularity); debugfs_create_u32("latency_warn_ms", 0644, debugfs_sched, &sysctl_resched_latency_warn_ms); debugfs_create_u32("latency_warn_once", 0644, debugfs_sched, &sysctl_resched_latency_warn_once); @@ -594,6 +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", ((x16*)&p->se.burst_penalty)->u8[1]); +#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 461409c0e..1293fe037 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 +69,17 @@ * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus * - * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) + * (default SCHED_TUNABLESCALING_NONE = *1) */ -unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG; +unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE; /* * Minimal preemption granularity for CPU-bound tasks: * - * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) + * (default: 3 msec * 1, units: nanoseconds) */ -unsigned int sysctl_sched_base_slice = 750000ULL; -static unsigned int normalized_sysctl_sched_base_slice = 750000ULL; +unsigned int sysctl_sched_base_slice = 3000000ULL; +static unsigned int normalized_sysctl_sched_base_slice = 3000000ULL; /* * After fork, child runs first. If set to 0 (default) then @@ -84,8 +87,85 @@ static unsigned int normalized_sysctl_sched_base_slice = 750000ULL; */ unsigned int sysctl_sched_child_runs_first __read_mostly; +/* + * SCHED_OTHER wake-up granularity. + * + * This option delays the preemption effects of decoupled workloads + * and reduces their over-scheduling. Synchronous workloads will still + * have immediate wakeup/sleep latencies. + * + * (default: 1.6 msec * 1, units: nanoseconds) + */ +unsigned int sysctl_sched_wakeup_granularity = 1600000UL; +static unsigned int normalized_sysctl_sched_wakeup_granularity = 1600000UL; + 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 = 60000000; +unsigned int __read_mostly sched_burst_penalty_offset = 22; +unsigned int __read_mostly sched_burst_penalty_scale = 1366; +unsigned int __read_mostly sched_burst_smoothness_up = 1; +unsigned int __read_mostly sched_burst_smoothness_down = 0; +unsigned int __read_mostly sched_burst_fork_atavistic = 2; +static int three = 3; +static int sixty_four = 64; +static int maxval_12_bits = 4095; + +#define MAX_BURST_PENALTY ((40U << 8) - 1) + +static inline u32 log2plus1_u64_u32f8(u64 v) { + x32 result; + int msb = fls64(v); + int excess_bits = msb - 9; + result.u8[0] = (0 <= excess_bits)? v >> excess_bits: v << -excess_bits; + result.u8[1] = msb; + return result.u32; +} + +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 >> 10; + + return min(MAX_BURST_PENALTY, scaled_penalty); +} + +static 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 u64 penalty_scale(u64 delta, struct sched_entity *se, bool half) { + u32 score = ((x16*)&se->burst_penalty)->u8[1]; + if (half) score >>= 1; + return mul_u64_u32_shr(delta, sched_prio_to_wmult[score], 22); +} + +static inline u32 binary_smooth(u32 new, u32 old) { + int increment = new - old; + return (0 <= increment)? + old + ( increment >> sched_burst_smoothness_up): + old - (-increment >> sched_burst_smoothness_down); +} + +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; +} + +static inline void vruntime_backstep(s64 *vdiff, struct sched_entity *se) { + u64 delta_exec = se->sum_exec_runtime - se->prev_sum_exec_runtime; + *vdiff += delta_exec - penalty_scale(delta_exec, se, false); +} +#endif // CONFIG_SCHED_BORE + int sched_thermal_decay_shift; static int __init setup_sched_thermal_decay_shift(char *str) { @@ -145,6 +225,69 @@ 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_fork_atavistic", + .data = &sched_burst_fork_atavistic, + .maxlen = sizeof(unsigned int), + .mode = 0644, + .proc_handler = &proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &three, + }, + { + .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_down", + .data = &sched_burst_smoothness_down, + .maxlen = sizeof(unsigned int), + .mode = 0644, + .proc_handler = &proc_dointvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &three, + }, + { + .procname = "sched_burst_smoothness_up", + .data = &sched_burst_smoothness_up, + .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, @@ -238,6 +381,7 @@ static void update_sysctl(void) #define SET_SYSCTL(name) \ (sysctl_##name = (factor) * normalized_sysctl_##name) SET_SYSCTL(sched_base_slice); + SET_SYSCTL(sched_wakeup_granularity); #undef SET_SYSCTL } @@ -308,11 +452,20 @@ static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight /* * delta /= w */ +#ifdef CONFIG_SCHED_BORE +#define calc_delta_fair_half(delta, se) __calc_delta_fair(delta, se, true) +#define calc_delta_fair(delta, se) __calc_delta_fair(delta, se, false) +static inline u64 __calc_delta_fair(u64 delta, struct sched_entity *se, bool half) +#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 (likely(sched_bore)) delta = penalty_scale(delta, se, half); +#endif // CONFIG_SCHED_BORE return delta; } @@ -944,6 +1097,7 @@ int sched_update_scaling(void) #define WRT_SYSCTL(name) \ (normalized_sysctl_##name = sysctl_##name / (factor)) WRT_SYSCTL(sched_base_slice); + WRT_SYSCTL(sched_wakeup_granularity); #undef WRT_SYSCTL return 0; @@ -1121,7 +1275,11 @@ 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); @@ -4919,7 +5077,7 @@ static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} static void place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) { - u64 vslice = calc_delta_fair(se->slice, se); + u64 vslice = calc_delta_fair_half(se->slice, se); u64 vruntime = avg_vruntime(cfs_rq); s64 lag = 0; @@ -5187,6 +5345,9 @@ set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) se->prev_sum_exec_runtime = se->sum_exec_runtime; } +static int +wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); + /* * Pick the next process, keeping these things in mind, in this order: * 1) keep things fair between processes/task groups @@ -5197,14 +5358,16 @@ set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) static struct sched_entity * pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) { + struct sched_entity *candidate = pick_eevdf(cfs_rq); /* * Enabling NEXT_BUDDY will affect latency but not fairness. */ if (sched_feat(NEXT_BUDDY) && - cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next)) + cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next) && + wakeup_preempt_entity(cfs_rq->next, candidate) < 1) return cfs_rq->next; - return pick_eevdf(cfs_rq); + return candidate; } static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); @@ -6452,6 +6615,30 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) hrtick_update(rq); } +static unsigned long wakeup_gran(struct sched_entity *se) +{ + unsigned long gran = sysctl_sched_wakeup_granularity; + return calc_delta_fair(gran, se); +} + +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 (likely(sched_bore)) vruntime_backstep(&vdiff, curr); +#endif // CONFIG_SCHED_BORE + + if (vdiff <= 0) + return -1; + + gran = wakeup_gran(se); + if (vdiff > gran) + return 1; + + return 0; +} + static void set_next_buddy(struct sched_entity *se); /* @@ -6470,6 +6657,9 @@ 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); @@ -7980,7 +8170,7 @@ static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_ /* * XXX pick_eevdf(cfs_rq) != se ? */ - if (pick_eevdf(cfs_rq) == pse) + if ((pick_eevdf(cfs_rq) == pse) && (wakeup_preempt_entity(se, pse) == 1)) goto preempt; return; @@ -8197,8 +8387,12 @@ 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); +#endif // CONFIG_SCHED_BORE return; + } clear_buddies(cfs_rq, se); @@ -8207,6 +8401,9 @@ 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); +#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/sched.h b/kernel/sched/sched.h index 67cd7e1fd..04d065015 100644 --- a/kernel/sched/sched.h +++ b/kernel/sched/sched.h @@ -2506,6 +2506,7 @@ extern const_debug unsigned int sysctl_sched_nr_migrate; extern const_debug unsigned int sysctl_sched_migration_cost; extern unsigned int sysctl_sched_base_slice; +extern unsigned int sysctl_sched_wakeup_granularity; #ifdef CONFIG_SCHED_DEBUG extern int sysctl_resched_latency_warn_ms; -- 2.42.0.rc0.25.ga82fb66fed