From 7b726493e44bba968c8c0369c387b5e462d6840e Mon Sep 17 00:00:00 2001 From: Piotr Gorski Date: Mon, 16 Sep 2024 13:24:13 +0200 Subject: [PATCH] bore-cachy-ext Signed-off-by: Piotr Gorski --- include/linux/sched.h | 10 ++ init/Kconfig | 17 ++ kernel/Kconfig.hz | 16 ++ kernel/sched/core.c | 141 +++++++++++++++ kernel/sched/debug.c | 60 ++++++- kernel/sched/fair.c | 383 +++++++++++++++++++++++++++++++++++++--- kernel/sched/features.h | 16 +- kernel/sched/sched.h | 7 + 8 files changed, 623 insertions(+), 27 deletions(-) diff --git a/include/linux/sched.h b/include/linux/sched.h index 5b4f78fe3..b9e5ea2aa 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -551,6 +551,16 @@ 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 burst_score; + u8 child_burst; + u32 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 e1a88d48d..3aea8e43c 100644 --- a/init/Kconfig +++ b/init/Kconfig @@ -1327,6 +1327,23 @@ 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. + + If unsure, say Y here. + config SCHED_AUTOGROUP bool "Automatic process group scheduling" select CGROUPS diff --git a/kernel/Kconfig.hz b/kernel/Kconfig.hz index 0f78364ef..b50189ee5 100644 --- a/kernel/Kconfig.hz +++ b/kernel/Kconfig.hz @@ -79,5 +79,21 @@ config HZ default 750 if HZ_750 default 1000 if HZ_1000 +config MIN_BASE_SLICE_NS + int "Default value for min_base_slice_ns" + default 2000000 + help + The BORE Scheduler automatically calculates the optimal base + slice for the configured HZ using the following equation: + + base_slice_ns = max(min_base_slice_ns, 1000000000/HZ) + + This option sets the default lower bound limit of the base slice + to prevent the loss of task throughput due to overscheduling. + + Setting this value too high can cause the system to boot with + an unnecessarily large base slice, resulting in high scheduling + latency and poor system responsiveness. + config SCHED_HRTICK def_bool HIGH_RES_TIMERS diff --git a/kernel/sched/core.c b/kernel/sched/core.c index c792a6feb..79a913e2f 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -4336,6 +4336,136 @@ int wake_up_state(struct task_struct *p, unsigned int state) return try_to_wake_up(p, state, 0); } +#ifdef CONFIG_SCHED_BORE +extern u8 sched_burst_fork_atavistic; +extern uint sched_burst_cache_lifetime; + +static 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.burst_score = 0; + init_task.se.child_burst_last_cached = 0; +} + +inline void sched_fork_bore(struct task_struct *p) { + p->se.burst_time = 0; + p->se.curr_burst_penalty = 0; + p->se.burst_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 task_burst_inheritable(struct task_struct *p) { + return (p->sched_class == &fair_sched_class); +} + +static inline bool child_burst_cache_expired(struct task_struct *p, u64 now) { + u64 expiration_time = + p->se.child_burst_last_cached + sched_burst_cache_lifetime; + return ((s64)(expiration_time - now) < 0); +} + +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 inline void update_child_burst_direct(struct task_struct *p, u64 now) { + struct task_struct *child; + u32 cnt = 0, sum = 0; + + list_for_each_entry(child, &p->children, sibling) { + if (!task_burst_inheritable(child)) continue; + cnt++; + sum += child->se.burst_penalty; + } + + __update_child_burst_cache(p, cnt, sum, now); +} + +static inline u8 __inherit_burst_direct(struct task_struct *p, u64 now) { + struct task_struct *parent = p->real_parent; + if (child_burst_cache_expired(parent, now)) + update_child_burst_direct(parent, now); + + return parent->se.child_burst; +} + +static void update_child_burst_topological( + struct task_struct *p, u64 now, u32 depth, u32 *acnt, u32 *asum) { + struct task_struct *child, *dec; + u32 cnt = 0, dcnt = 0, 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 (!task_burst_inheritable(dec)) 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_topological(dec, now, depth - 1, &cnt, &sum); + } + + __update_child_burst_cache(p, cnt, sum, now); + *acnt += cnt; + *asum += sum; +} + +static inline u8 __inherit_burst_topological(struct task_struct *p, u64 now) { + struct task_struct *anc = p->real_parent; + u32 cnt = 0, sum = 0; + + while (anc->real_parent != anc && count_child_tasks(anc) == 1) + anc = anc->real_parent; + + if (child_burst_cache_expired(anc, now)) + update_child_burst_topological( + anc, now, sched_burst_fork_atavistic - 1, &cnt, &sum); + + return anc->se.child_burst; +} + +static inline void inherit_burst(struct task_struct *p) { + u8 burst_cache; + u64 now = ktime_get_ns(); + + read_lock(&tasklist_lock); + burst_cache = likely(sched_burst_fork_atavistic)? + __inherit_burst_topological(p, now): + __inherit_burst_direct(p, now); + read_unlock(&tasklist_lock); + + p->se.prev_burst_penalty = max(p->se.prev_burst_penalty, burst_cache); +} + +static void sched_post_fork_bore(struct task_struct *p) { + if (task_burst_inheritable(p)) + inherit_burst(p); + p->se.burst_penalty = p->se.prev_burst_penalty; +} +#endif // CONFIG_SCHED_BORE + /* * Perform scheduler related setup for a newly forked process p. * p is forked by current. @@ -4352,6 +4482,9 @@ 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); @@ -4686,6 +4819,9 @@ void sched_cancel_fork(struct task_struct *p) 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); scx_post_fork(p); } @@ -8283,6 +8419,11 @@ void __init sched_init(void) BUG_ON(!sched_class_above(&ext_sched_class, &idle_sched_class)); #endif +#ifdef CONFIG_SCHED_BORE + sched_init_bore(); + printk(KERN_INFO "BORE (Burst-Oriented Response Enhancer) CPU Scheduler modification 5.2.12 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 c057ef46c..3cab39e34 100644 --- a/kernel/sched/debug.c +++ b/kernel/sched/debug.c @@ -167,7 +167,52 @@ static const struct file_operations sched_feat_fops = { }; #ifdef CONFIG_SMP +#ifdef CONFIG_SCHED_BORE +static ssize_t sched_min_base_slice_write(struct file *filp, const char __user *ubuf, + size_t cnt, loff_t *ppos) +{ + char buf[16]; + unsigned int value; + + if (cnt > 15) + cnt = 15; + + if (copy_from_user(&buf, ubuf, cnt)) + return -EFAULT; + buf[cnt] = '\0'; + + if (kstrtouint(buf, 10, &value)) + return -EINVAL; + if (!value) + return -EINVAL; + + sysctl_sched_min_base_slice = value; + sched_update_min_base_slice(); + + *ppos += cnt; + return cnt; +} + +static int sched_min_base_slice_show(struct seq_file *m, void *v) +{ + seq_printf(m, "%d\n", sysctl_sched_min_base_slice); + return 0; +} + +static int sched_min_base_slice_open(struct inode *inode, struct file *filp) +{ + return single_open(filp, sched_min_base_slice_show, NULL); +} + +static const struct file_operations sched_min_base_slice_fops = { + .open = sched_min_base_slice_open, + .write = sched_min_base_slice_write, + .read = seq_read, + .llseek = seq_lseek, + .release = single_release, +}; +#else // !CONFIG_SCHED_BORE static ssize_t sched_scaling_write(struct file *filp, const char __user *ubuf, size_t cnt, loff_t *ppos) { @@ -213,7 +258,7 @@ static const struct file_operations sched_scaling_fops = { .llseek = seq_lseek, .release = single_release, }; - +#endif // CONFIG_SCHED_BORE #endif /* SMP */ #ifdef CONFIG_PREEMPT_DYNAMIC @@ -347,13 +392,20 @@ static __init int sched_init_debug(void) debugfs_create_file("preempt", 0644, debugfs_sched, NULL, &sched_dynamic_fops); #endif +#ifdef CONFIG_SCHED_BORE + debugfs_create_file("min_base_slice_ns", 0644, debugfs_sched, NULL, &sched_min_base_slice_fops); + debugfs_create_u32("base_slice_ns", 0400, debugfs_sched, &sysctl_sched_base_slice); +#else // !CONFIG_SCHED_BORE debugfs_create_u32("base_slice_ns", 0644, debugfs_sched, &sysctl_sched_base_slice); +#endif // CONFIG_SCHED_BORE 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); #ifdef CONFIG_SMP +#if !defined(CONFIG_SCHED_BORE) debugfs_create_file("tunable_scaling", 0644, debugfs_sched, NULL, &sched_scaling_fops); +#endif // CONFIG_SCHED_BORE debugfs_create_u32("migration_cost_ns", 0644, debugfs_sched, &sysctl_sched_migration_cost); debugfs_create_u32("nr_migrate", 0644, debugfs_sched, &sysctl_sched_nr_migrate); @@ -596,6 +648,9 @@ 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 // CONFIG_SCHED_BORE #ifdef CONFIG_NUMA_BALANCING SEQ_printf(m, " %d %d", task_node(p), task_numa_group_id(p)); #endif @@ -1069,6 +1124,9 @@ void proc_sched_show_task(struct task_struct *p, struct pid_namespace *ns, P(se.load.weight); #ifdef CONFIG_SMP +#ifdef CONFIG_SCHED_BORE + P(se.burst_score); +#endif // CONFIG_SCHED_BORE P(se.avg.load_sum); P(se.avg.runnable_sum); P(se.avg.util_sum); diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 2928026d7..91a870173 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) 2021-2024 Masahito Suzuki */ #include #include @@ -64,28 +67,176 @@ * SCHED_TUNABLESCALING_LOG - scaled logarithmically, *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: max(1 sec / HZ, min_base_slice) constant, units: nanoseconds) + * (EEVDF default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) */ -#ifdef CONFIG_CACHY -unsigned int sysctl_sched_base_slice = 350000ULL; -static unsigned int normalized_sysctl_sched_base_slice = 350000ULL; -#else +#ifdef CONFIG_SCHED_BORE +unsigned int sysctl_sched_base_slice = 1000000000ULL / HZ; +static unsigned int configured_sched_base_slice = 1000000000ULL / HZ; +unsigned int sysctl_sched_min_base_slice = CONFIG_MIN_BASE_SLICE_NS; +#else // !CONFIG_SCHED_BORE unsigned int sysctl_sched_base_slice = 750000ULL; static unsigned int normalized_sysctl_sched_base_slice = 750000ULL; -#endif +#endif // CONFIG_SCHED_BORE -#ifdef CONFIG_CACHY -const_debug unsigned int sysctl_sched_migration_cost = 300000UL; -#else const_debug unsigned int sysctl_sched_migration_cost = 500000UL; -#endif + +#ifdef CONFIG_SCHED_BORE +u8 __read_mostly sched_bore = 1; +u8 __read_mostly sched_burst_exclude_kthreads = 1; +u8 __read_mostly sched_burst_smoothness_long = 1; +u8 __read_mostly sched_burst_smoothness_short = 0; +u8 __read_mostly sched_burst_fork_atavistic = 2; +u8 __read_mostly sched_burst_parity_threshold = 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; +uint __read_mostly sched_deadline_boost_mask = ENQUEUE_INITIAL + | ENQUEUE_WAKEUP; +uint __read_mostly sched_deadline_preserve_mask = ENQUEUE_RESTORE + | ENQUEUE_MIGRATED; +static int __maybe_unused sixty_four = 64; +static int __maybe_unused maxval_u8 = 255; +static int __maybe_unused maxval_12_bits = 4095; + +#define MAX_BURST_PENALTY (39U <<2) + +static inline u32 log2plus1_u64_u32f8(u64 v) { + u32 msb = fls64(v); + u8 fractional = (v << (64 - msb) >> 55); + 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 u64 __scale_slice(u64 delta, u8 score) { + return mul_u64_u32_shr(delta, sched_prio_to_wmult[score], 22); +} + +static inline u64 __unscale_slice(u64 delta, u8 score) { + return mul_u64_u32_shr(delta, sched_prio_to_weight[score], 10); +} + +static void reweight_entity( + struct cfs_rq *cfs_rq, struct sched_entity *se, unsigned long weight); + +static void reweight_task_by_prio(struct task_struct *p, int prio) +{ + struct sched_entity *se = &p->se; + struct cfs_rq *cfs_rq = cfs_rq_of(se); + struct load_weight *load = &se->load; + unsigned long weight = scale_load(sched_prio_to_weight[prio]); + + reweight_entity(cfs_rq, se, weight); + load->inv_weight = sched_prio_to_wmult[prio]; +} + +static inline u8 effective_prio(struct task_struct *p) { + u8 prio = p->static_prio - MAX_RT_PRIO; + if (likely(sched_bore)) + prio += p->se.burst_score; + return min(39, prio); +} + +static void update_burst_score(struct sched_entity *se) { + if (!entity_is_task(se)) return; + struct task_struct *p = task_of(se); + u8 prev_prio = effective_prio(p); + + u8 burst_score = 0; + if (!((p->flags & PF_KTHREAD) && likely(sched_burst_exclude_kthreads))) + burst_score = se->burst_penalty >> 2; + se->burst_score = burst_score; + + u8 new_prio = effective_prio(p); + if (new_prio != prev_prio) + reweight_task_by_prio(p, new_prio); +} + +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); + update_burst_score(se); +} + +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; + update_burst_score(se); +} + +static void restart_burst_rescale_deadline(struct sched_entity *se) { + s64 vscaled, wremain, vremain = se->deadline - se->vruntime; + struct task_struct *p = task_of(se); + u8 prev_prio = effective_prio(p); + restart_burst(se); + u8 new_prio = effective_prio(p); + if (prev_prio > new_prio) { + wremain = __unscale_slice(abs(vremain), prev_prio); + vscaled = __scale_slice(wremain, new_prio); + if (unlikely(vremain < 0)) + vscaled = -vscaled; + se->deadline = se->vruntime + vscaled; + } +} + +static void reset_task_weights_bore(void) { + struct task_struct *task; + struct rq *rq; + struct rq_flags rf; + + write_lock_irq(&tasklist_lock); + for_each_process(task) { + rq = task_rq(task); + rq_lock_irqsave(rq, &rf); + reweight_task_by_prio(task, effective_prio(task)); + rq_unlock_irqrestore(rq, &rf); + } + write_unlock_irq(&tasklist_lock); +} + +int sched_bore_update_handler(const struct ctl_table *table, int write, + void __user *buffer, size_t *lenp, loff_t *ppos) +{ + int ret = proc_dou8vec_minmax(table, write, buffer, lenp, ppos); + if (ret || !write) + return ret; + + reset_task_weights_bore(); + + return 0; +} +#endif // CONFIG_SCHED_BORE static int __init setup_sched_thermal_decay_shift(char *str) { @@ -130,12 +281,8 @@ int __weak arch_asym_cpu_priority(int cpu) * * (default: 5 msec, units: microseconds) */ -#ifdef CONFIG_CACHY -static unsigned int sysctl_sched_cfs_bandwidth_slice = 3000UL; -#else static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; #endif -#endif #ifdef CONFIG_NUMA_BALANCING /* Restrict the NUMA promotion throughput (MB/s) for each target node. */ @@ -144,6 +291,101 @@ 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(u8), + .mode = 0644, + .proc_handler = sched_bore_update_handler, + .extra1 = SYSCTL_ZERO, + .extra2 = SYSCTL_ONE, + }, + { + .procname = "sched_burst_exclude_kthreads", + .data = &sched_burst_exclude_kthreads, + .maxlen = sizeof(u8), + .mode = 0644, + .proc_handler = proc_dou8vec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = SYSCTL_ONE, + }, + { + .procname = "sched_burst_smoothness_long", + .data = &sched_burst_smoothness_long, + .maxlen = sizeof(u8), + .mode = 0644, + .proc_handler = proc_dou8vec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = SYSCTL_ONE, + }, + { + .procname = "sched_burst_smoothness_short", + .data = &sched_burst_smoothness_short, + .maxlen = sizeof(u8), + .mode = 0644, + .proc_handler = proc_dou8vec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = SYSCTL_ONE, + }, + { + .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_parity_threshold", + .data = &sched_burst_parity_threshold, + .maxlen = sizeof(u8), + .mode = 0644, + .proc_handler = proc_dou8vec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &maxval_u8, + }, + { + .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_cache_lifetime", + .data = &sched_burst_cache_lifetime, + .maxlen = sizeof(uint), + .mode = 0644, + .proc_handler = proc_douintvec, + }, + { + .procname = "sched_deadline_boost_mask", + .data = &sched_deadline_boost_mask, + .maxlen = sizeof(uint), + .mode = 0644, + .proc_handler = proc_douintvec, + }, + { + .procname = "sched_deadline_preserve_mask", + .data = &sched_deadline_preserve_mask, + .maxlen = sizeof(uint), + .mode = 0644, + .proc_handler = proc_douintvec, + }, +#endif // CONFIG_SCHED_BORE #ifdef CONFIG_CFS_BANDWIDTH { .procname = "sched_cfs_bandwidth_slice_us", @@ -201,6 +443,13 @@ static inline void update_load_set(struct load_weight *lw, unsigned long w) * * This idea comes from the SD scheduler of Con Kolivas: */ +#ifdef CONFIG_SCHED_BORE +static void update_sysctl(void) { + sysctl_sched_base_slice = + max(sysctl_sched_min_base_slice, configured_sched_base_slice); +} +void sched_update_min_base_slice(void) { update_sysctl(); } +#else // !CONFIG_SCHED_BORE static unsigned int get_update_sysctl_factor(void) { unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); @@ -231,6 +480,7 @@ static void update_sysctl(void) SET_SYSCTL(sched_base_slice); #undef SET_SYSCTL } +#endif // CONFIG_SCHED_BORE void __init sched_init_granularity(void) { @@ -708,6 +958,9 @@ static s64 entity_lag(u64 avruntime, struct sched_entity *se) vlag = avruntime - se->vruntime; limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se); +#ifdef CONFIG_SCHED_BORE + limit >>= !!sched_bore; +#endif // CONFIG_SCHED_BORE return clamp(vlag, -limit, limit); } @@ -868,6 +1121,45 @@ struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) return __node_2_se(left); } +static inline bool pick_curr(struct cfs_rq *cfs_rq, + struct sched_entity *curr, struct sched_entity *wakee) +{ + /* + * Nothing to preserve... + */ + if (!curr || !sched_feat(RESPECT_SLICE)) + return false; + +#ifdef CONFIG_SCHED_BORE + if (likely(sched_bore) && likely(sched_burst_parity_threshold) && + sched_burst_parity_threshold < cfs_rq->nr_running) + return false; +#endif // CONFIG_SCHED_BORE + + /* + * Allow preemption at the 0-lag point -- even if not all of the slice + * is consumed. Note: placement of positive lag can push V left and render + * @curr instantly ineligible irrespective the time on-cpu. + */ + if (sched_feat(RUN_TO_PARITY) && !entity_eligible(cfs_rq, curr)) + return false; + + /* + * Don't preserve @curr when the @wakee has a shorter slice and earlier + * deadline. IOW, explicitly allow preemption. + */ + if (sched_feat(PREEMPT_SHORT) && wakee && + wakee->slice < curr->slice && + (s64)(wakee->deadline - curr->deadline) < 0) + return false; + + /* + * Preserve @curr to allow it to finish its first slice. + * See the HACK in set_next_entity(). + */ + return curr->vlag == curr->deadline; +} + /* * Earliest Eligible Virtual Deadline First * @@ -887,28 +1179,27 @@ struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) * * Which allows tree pruning through eligibility. */ -static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq) +static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq, struct sched_entity *wakee) { struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node; struct sched_entity *se = __pick_first_entity(cfs_rq); struct sched_entity *curr = cfs_rq->curr; struct sched_entity *best = NULL; + if (curr && !curr->on_rq) + curr = NULL; + /* * We can safely skip eligibility check if there is only one entity * in this cfs_rq, saving some cycles. */ if (cfs_rq->nr_running == 1) - return curr && curr->on_rq ? curr : se; - - if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr))) - curr = NULL; + return curr ?: se; /* - * Once selected, run a task until it either becomes non-eligible or - * until it gets a new slice. See the HACK in set_next_entity(). + * Preserve @curr to let it finish its slice. */ - if (sched_feat(RUN_TO_PARITY) && curr && curr->vlag == curr->deadline) + if (pick_curr(cfs_rq, curr, wakee)) return curr; /* Pick the leftmost entity if it's eligible */ @@ -967,6 +1258,7 @@ struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) * Scheduling class statistics methods: */ #ifdef CONFIG_SMP +#if !defined(CONFIG_SCHED_BORE) int sched_update_scaling(void) { unsigned int factor = get_update_sysctl_factor(); @@ -978,6 +1270,7 @@ int sched_update_scaling(void) return 0; } +#endif // CONFIG_SCHED_BORE #endif #endif @@ -1178,6 +1471,10 @@ static void update_curr(struct cfs_rq *cfs_rq) if (unlikely(delta_exec <= 0)) return; +#ifdef CONFIG_SCHED_BORE + curr->burst_time += delta_exec; + update_burst_penalty(curr); +#endif // CONFIG_SCHED_BORE curr->vruntime += calc_delta_fair(delta_exec, curr); update_deadline(cfs_rq, curr); update_min_vruntime(cfs_rq); @@ -5193,6 +5490,12 @@ 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 + if (likely(sched_bore) && + (flags & ~sched_deadline_boost_mask & sched_deadline_preserve_mask)) + vslice = se->deadline - se->vruntime; + else +#endif // CONFIG_SCHED_BORE vslice = calc_delta_fair(se->slice, se); /* @@ -5203,6 +5506,9 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) * * EEVDF: placement strategy #1 / #2 */ +#ifdef CONFIG_SCHED_BORE + if (se->vlag) +#endif // CONFIG_SCHED_BORE if (sched_feat(PLACE_LAG) && cfs_rq->nr_running) { struct sched_entity *curr = cfs_rq->curr; unsigned long load; @@ -5278,6 +5584,13 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) * on average, halfway through their slice, as such start tasks * off with half a slice to ease into the competition. */ +#ifdef CONFIG_SCHED_BORE + if (likely(sched_bore)) { + if (flags & sched_deadline_boost_mask) + vslice /= 2; + } + else +#endif // CONFIG_SCHED_BORE if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL)) vslice /= 2; @@ -5492,7 +5805,7 @@ pick_next_entity(struct cfs_rq *cfs_rq) cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next)) return cfs_rq->next; - return pick_eevdf(cfs_rq); + return pick_eevdf(cfs_rq, NULL); } static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); @@ -6860,6 +7173,14 @@ 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) { + cfs_rq = cfs_rq_of(se); + if (cfs_rq->curr == se) + update_curr(cfs_rq); + restart_burst(se); + } +#endif // CONFIG_SCHED_BORE for_each_sched_entity(se) { cfs_rq = cfs_rq_of(se); @@ -8428,7 +8749,7 @@ static void check_preempt_wakeup_fair(struct rq *rq, struct task_struct *p, int /* * XXX pick_eevdf(cfs_rq) != se ? */ - if (pick_eevdf(cfs_rq) == pse) + if (pick_eevdf(cfs_rq, pse) == pse) goto preempt; return; @@ -8646,16 +8967,25 @@ static void yield_task_fair(struct rq *rq) /* * Are we the only task in the tree? */ +#if !defined(CONFIG_SCHED_BORE) if (unlikely(rq->nr_running == 1)) return; clear_buddies(cfs_rq, se); +#endif // CONFIG_SCHED_BORE update_rq_clock(rq); /* * Update run-time statistics of the 'current'. */ update_curr(cfs_rq); +#ifdef CONFIG_SCHED_BORE + restart_burst_rescale_deadline(se); + if (unlikely(rq->nr_running == 1)) + return; + + clear_buddies(cfs_rq, se); +#endif // CONFIG_SCHED_BORE /* * Tell update_rq_clock() that we've just updated, * so we don't do microscopic update in schedule() @@ -12720,6 +13050,9 @@ static void task_fork_fair(struct task_struct *p) curr = cfs_rq->curr; if (curr) update_curr(cfs_rq); +#ifdef CONFIG_SCHED_BORE + update_burst_score(se); +#endif // CONFIG_SCHED_BORE place_entity(cfs_rq, se, ENQUEUE_INITIAL); rq_unlock(rq, &rf); } diff --git a/kernel/sched/features.h b/kernel/sched/features.h index 143f55df8..09599a67c 100644 --- a/kernel/sched/features.h +++ b/kernel/sched/features.h @@ -5,8 +5,22 @@ * sleep+wake cycles. EEVDF placement strategy #1, #2 if disabled. */ SCHED_FEAT(PLACE_LAG, true) +/* + * Give new tasks half a slice to ease into the competition. + */ SCHED_FEAT(PLACE_DEADLINE_INITIAL, true) -SCHED_FEAT(RUN_TO_PARITY, true) +/* + * Inhibit (wakeup) preemption until the current task has exhausted its slice. + */ +SCHED_FEAT(RESPECT_SLICE, true) +/* + * Relax RESPECT_SLICE to allow preemption once current has reached 0-lag. + */ +SCHED_FEAT(RUN_TO_PARITY, false) +/* + * Allow tasks with a shorter slice to disregard RESPECT_SLICE + */ +SCHED_FEAT(PREEMPT_SHORT, true) /* * Prefer to schedule the task we woke last (assuming it failed diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h index 207a04f02..c99430161 100644 --- a/kernel/sched/sched.h +++ b/kernel/sched/sched.h @@ -2063,7 +2063,11 @@ static inline void update_sched_domain_debugfs(void) { } static inline void dirty_sched_domain_sysctl(int cpu) { } #endif +#ifdef CONFIG_SCHED_BORE +extern void sched_update_min_base_slice(void); +#else // !CONFIG_SCHED_BORE extern int sched_update_scaling(void); +#endif // CONFIG_SCHED_BORE static inline const struct cpumask *task_user_cpus(struct task_struct *p) { @@ -2736,6 +2740,9 @@ 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; +#ifdef CONFIG_SCHED_BORE +extern unsigned int sysctl_sched_min_base_slice; +#endif // CONFIG_SCHED_BORE #ifdef CONFIG_SCHED_DEBUG extern int sysctl_resched_latency_warn_ms; -- 2.45.2.606.g9005149a4a