From: Zhang Qiao zhangqiao22@huawei.com
hulk inclusion category: feature bugzilla: https://gitee.com/openeuler/kernel/issues/I7WWMX CVE: NA
--------------------------------
In cloud-native hybrid deployment scenarios, online tasks must preempt offline tasks in a timely and offline tasks can't affect the QoS of online tasks, so we introduce the idea of qos level to scheduler, which now is supported with different scheduler policies. The qos scheduler will change the policy of correlative tasks when the qos level of a task group is modified with cpu.qos_level cpu cgroup file. In this way we are able to satisfy different needs of tasks in different qos levels.
The value of qos_level can be 0 or -1, default value is 0. If qos_level is 0, the group is an online group. otherwise it is an offline group.
Signed-off-by: Zhang Qiao zhangqiao22@huawei.com --- init/Kconfig | 12 +++++ kernel/sched/core.c | 116 +++++++++++++++++++++++++++++++++++++++++++ kernel/sched/sched.h | 4 ++ 3 files changed, 132 insertions(+)
diff --git a/init/Kconfig b/init/Kconfig index 6d35728b94b2..be94ba25ecca 100644 --- a/init/Kconfig +++ b/init/Kconfig @@ -985,6 +985,18 @@ menuconfig CGROUP_SCHED tasks.
if CGROUP_SCHED +config QOS_SCHED + bool "Qos task scheduling" + depends on CGROUP_SCHED + depends on CFS_BANDWIDTH + default n + help + This option enable qos scheduler, and support co-location online + services (Latency Sensitive) and offline tasks. colocation can + effectively improve the resource utilization. + + If in doubt, say N. + config FAIR_GROUP_SCHED bool "Group scheduling for SCHED_OTHER" depends on CGROUP_SCHED diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 9e9a45a3394f..9f466f43c023 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -7730,6 +7730,18 @@ static int __sched_setscheduler(struct task_struct *p, } change:
+#ifdef CONFIG_QOS_SCHED + /* + * If the scheduling policy of an offline task is set to a policy + * other than SCHED_IDLE, the online task preemption and cpu resource + * isolation will be invalid, so return -EINVAL in this case. + */ + if (unlikely(task_group(p)->qos_level == -1 && !idle_policy(policy))) { + task_rq_unlock(rq, p, &rf); + return -EINVAL; + } +#endif + if (user) { #ifdef CONFIG_RT_GROUP_SCHED /* @@ -10348,6 +10360,35 @@ void ia64_set_curr_task(int cpu, struct task_struct *p) /* task_group_lock serializes the addition/removal of task groups */ static DEFINE_SPINLOCK(task_group_lock);
+#ifdef CONFIG_QOS_SCHED +static inline int alloc_qos_sched_group(struct task_group *tg, + struct task_group *parent) +{ + tg->qos_level = parent->qos_level; + + return 1; +} + +static void sched_change_qos_group(struct task_struct *tsk, struct task_group *tg) +{ + struct sched_attr attr = {0}; + + /* + * No need to re-setcheduler when a task is exiting or the task + * is in an autogroup. + */ + if (!(tsk->flags & PF_EXITING) && + !task_group_is_autogroup(tg) && + (tg->qos_level == -1)) { + attr.sched_priority = 0; + attr.sched_policy = SCHED_IDLE; + attr.sched_nice = PRIO_TO_NICE(tsk->static_prio); + __setscheduler_params(tsk, &attr); + __setscheduler_prio(tsk, normal_prio(tsk)); + } +} +#endif + static inline void alloc_uclamp_sched_group(struct task_group *tg, struct task_group *parent) { @@ -10398,6 +10439,11 @@ struct task_group *sched_create_group(struct task_group *parent) if (!alloc_fair_sched_group(tg, parent)) goto err;
+#ifdef CONFIG_QOS_SCHED + if (!alloc_qos_sched_group(tg, parent)) + goto err; +#endif + if (!alloc_rt_sched_group(tg, parent)) goto err;
@@ -10484,6 +10530,10 @@ static void sched_change_group(struct task_struct *tsk, struct task_group *group { tsk->sched_task_group = group;
+#ifdef CONFIG_QOS_SCHED + sched_change_qos_group(tsk, group); +#endif + #ifdef CONFIG_FAIR_GROUP_SCHED if (tsk->sched_class->task_change_group) tsk->sched_class->task_change_group(tsk); @@ -11212,6 +11262,65 @@ static int cpu_idle_write_s64(struct cgroup_subsys_state *css, } #endif
+#ifdef CONFIG_QOS_SCHED +static int tg_change_scheduler(struct task_group *tg, void *data) +{ + int policy; + struct css_task_iter it; + struct sched_param param; + struct task_struct *tsk; + s64 qos_level = *(s64 *)data; + struct cgroup_subsys_state *css = &tg->css; + + tg->qos_level = qos_level; + if (qos_level == -1) { + policy = SCHED_IDLE; + cfs_bandwidth_usage_inc(); + } else { + policy = SCHED_NORMAL; + cfs_bandwidth_usage_dec(); + } + + param.sched_priority = 0; + css_task_iter_start(css, 0, &it); + while ((tsk = css_task_iter_next(&it))) + sched_setscheduler(tsk, policy, ¶m); + css_task_iter_end(&it); + + return 0; +} + +static int cpu_qos_write(struct cgroup_subsys_state *css, + struct cftype *cftype, s64 qos_level) +{ + struct task_group *tg = css_tg(css); + + if (!tg->se[0]) + return -EINVAL; + + if (qos_level != -1 && qos_level != 0) + return -EINVAL; + + if (tg->qos_level == qos_level) + goto done; + + if (tg->qos_level == -1 && qos_level == 0) + return -EINVAL; + + rcu_read_lock(); + walk_tg_tree_from(tg, tg_change_scheduler, tg_nop, (void *)(&qos_level)); + rcu_read_unlock(); +done: + return 0; +} + +static inline s64 cpu_qos_read(struct cgroup_subsys_state *css, + struct cftype *cft) +{ + return css_tg(css)->qos_level; +} +#endif + static struct cftype cpu_legacy_files[] = { #ifdef CONFIG_FAIR_GROUP_SCHED { @@ -11275,6 +11384,13 @@ static struct cftype cpu_legacy_files[] = { .seq_show = cpu_uclamp_max_show, .write = cpu_uclamp_max_write, }, +#endif +#ifdef CONFIG_QOS_SCHED + { + .name = "qos_level", + .read_s64 = cpu_qos_read, + .write_s64 = cpu_qos_write, + }, #endif { } /* Terminate */ }; diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h index 04846272409c..3586830dbeeb 100644 --- a/kernel/sched/sched.h +++ b/kernel/sched/sched.h @@ -403,6 +403,10 @@ struct task_group {
struct cfs_bandwidth cfs_bandwidth;
+#ifdef CONFIG_QOS_SCHED + long qos_level; +#endif + #ifdef CONFIG_UCLAMP_TASK_GROUP /* The two decimal precision [%] value requested from user-space */ unsigned int uclamp_pct[UCLAMP_CNT];