[PATCH OLK-6.6 V1 0/4] olk-6.6-backport-bugfix-for-dynamic-affinity
olk-6.6-backport-bugfix-for-dynamic-affinity He Yujie (3): sched/dynamic_affinity: fix preffered_cpu offline problem sched/dynamic_affinity: fix CI test alway select prefer_cpu problem sched/dynamic_affinity: Calculate cpu capacity in real time when realtime tasks are running on this cpu zhangwei123171 (1): sched/fair: reset idlest_cpu if not meet the conditions for preferred CPU kernel/sched/fair.c | 33 ++++++++++++++++++++++++++++++--- 1 file changed, 30 insertions(+), 3 deletions(-) -- 2.25.1
From: zhangwei123171 <zhangwei123171@jd.com> jingdong inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I8PGAS ----------------------------------------------------- We have set the preferred CPU, but when the load is too high, the conditions for the preferred CPU are not met. At this time, the idle CPU needs to be reset. This will be corrected in the subsequent fallback process. dynamic affinity should not break this logic. Fixes: 2e1dfc02d115 ("sched: Adjust wakeup cpu range according CPU util dynamicly") Signed-off-by: zhangwei123171 <zhangwei123171@jd.com> Reviewed-by: zhaoxiaoqiang11 <zhaoxiaoqiang11@jd.com> Signed-off-by: Cheng Yu <serein.chengyu@huawei.com> --- kernel/sched/fair.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index c4c3afa6e7b4..fdb352c870ea 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -9179,6 +9179,8 @@ static void set_task_select_cpus(struct task_struct *p, int *idlest_cpu, p->select_cpus = p->prefer_cpus; if (sd_flag & SD_BALANCE_WAKE) schedstat_inc(p->stats.nr_wakeups_preferred_cpus); + } else if (idlest_cpu) { + *idlest_cpu = -1; } } #endif -- 2.25.1
From: He Yujie <coka.heyujie@huawei.com> hulk inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/IB42TC -------------------------------- After the preferred_cpu goes offline, the core selection of dynamic affinity does not check whether the preferred_cpu is valid. As a result, related dynamic affinity processes are concentrated on a shared_cpu. This patch resolves this problem to checks whether the preferred_cpu is valid and compares only the usage threshold of the valid preferred_cpu. Fixes: 2e1dfc02d115 ("sched: Adjust wakeup cpu range according CPU util dynamicly") Signed-off-by: He Yujie <coka.heyujie@huawei.com> Signed-off-by: Cheng Yu <serein.chengyu@huawei.com> --- kernel/sched/fair.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index fdb352c870ea..f83094121207 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -9149,7 +9149,7 @@ static void set_task_select_cpus(struct task_struct *p, int *idlest_cpu, /* manual mode */ tg = task_group(p); - for_each_cpu(cpu, p->prefer_cpus) { + for_each_cpu_and(cpu, p->prefer_cpus, cpu_online_mask) { if (idlest_cpu && (available_idle_cpu(cpu) || sched_idle_cpu(cpu))) { *idlest_cpu = cpu; } else if (idlest_cpu) { @@ -9174,8 +9174,7 @@ static void set_task_select_cpus(struct task_struct *p, int *idlest_cpu, } rcu_read_unlock(); - if (tg_capacity > cpumask_weight(p->prefer_cpus) && - util_avg_sum * 100 <= tg_capacity * sysctl_sched_util_low_pct) { + if (util_avg_sum * 100 < tg_capacity * sysctl_sched_util_low_pct) { p->select_cpus = p->prefer_cpus; if (sd_flag & SD_BALANCE_WAKE) schedstat_inc(p->stats.nr_wakeups_preferred_cpus); -- 2.25.1
From: He Yujie <coka.heyujie@huawei.com> hulk inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/IB42TC -------------------------------- Commit aed539da59ad fix preffer_cpu offline problem, but introducing one CI problem. When one cpu usage is 100% occupied by RT tasks, CFS tasks util_avg is 0 but cpu capacity is not 0. During core selection, commit aed539da59ad delete the judgment of prefer_cpu set capacity and prefer_cpumask weight. As a result, dynamic_affinity task always select prefer_cpu as idlest_cpu. This patch resolves this problem to add the check of the valid prefer_cpu set capacity and the valid prefer_cpu weight. Fixes: 56bfe5272963 ("sched/dynamic_affinity: fix preffered_cpu offline problem") Signed-off-by: He Yujie <coka.heyujie@huawei.com> Signed-off-by: Cheng Yu <serein.chengyu@huawei.com> --- kernel/sched/fair.c | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index f83094121207..d2ad18d4b554 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -9128,6 +9128,7 @@ static void set_task_select_cpus(struct task_struct *p, int *idlest_cpu, struct task_group *tg; long spare; int cpu, mode; + int nr_cpus_valid = 0; p->select_cpus = p->cpus_ptr; if (!prefer_cpus_valid(p)) @@ -9171,10 +9172,18 @@ static void set_task_select_cpus(struct task_struct *p, int *idlest_cpu, util_avg_sum += taskgroup_cpu_util(tg, cpu); tg_capacity += capacity_of(cpu); + nr_cpus_valid++; } rcu_read_unlock(); - if (util_avg_sum * 100 < tg_capacity * sysctl_sched_util_low_pct) { + /* + * Follow cases should select cpus_ptr, checking by condition of + * tg_capacity > nr_cpus_valid: + * 1. all prefer_cpus offline; + * 2. all prefer_cpus has no cfs capaicity(tg_capacity = nr_cpus_valid * 1) + */ + if (tg_capacity > nr_cpus_valid && + util_avg_sum * 100 <= tg_capacity * sysctl_sched_util_low_pct) { p->select_cpus = p->prefer_cpus; if (sd_flag & SD_BALANCE_WAKE) schedstat_inc(p->stats.nr_wakeups_preferred_cpus); -- 2.25.1
From: He Yujie <coka.heyujie@huawei.com> hulk inclusion category: bugfix bugzilla: https://gitee.com/src-openeuler/kernel/issues/IBBVWN -------------------------------- The task_rq selection of dynamic affinity use cpu capacity to determine select_cpus range. When realtime tasks are running on the cpu all the time, cfs tasks and the thread of softirq is suppressed, and the cpu capacity is not updated in time. As a result, the select_cpus range is always selected for preferred_cpus. then cfs task will never be able to run because realtime tasks has been running. Therefore, if realtime tasks is running during the task_rq selection of dynamic affinity, the cpu capacity should be calculated to solve such a problem. Fixes: 2e1dfc02d115 ("sched: Adjust wakeup cpu range according CPU util dynamicly") Signed-off-by: He Yujie <coka.heyujie@huawei.com> Signed-off-by: Cheng Yu <serein.chengyu@huawei.com> --- kernel/sched/fair.c | 19 ++++++++++++++++++- 1 file changed, 18 insertions(+), 1 deletion(-) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index d2ad18d4b554..71661d6c5b54 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -9106,6 +9106,19 @@ static inline unsigned long taskgroup_cpu_util(struct task_group *tg, return cpu_util_cfs(cpu); } +static unsigned long scale_rt_capacity(int cpu); + +static inline unsigned long calc_cpu_capacity(int cpu) +{ + unsigned long capacity; + + capacity = scale_rt_capacity(cpu); + if (!capacity) + capacity = 1; + + return capacity; +} + /* * set_task_select_cpus: select the cpu range for task * @p: the task whose available cpu range will to set @@ -9171,8 +9184,12 @@ static void set_task_select_cpus(struct task_struct *p, int *idlest_cpu, } util_avg_sum += taskgroup_cpu_util(tg, cpu); - tg_capacity += capacity_of(cpu); nr_cpus_valid++; + + if (cpu_rq(cpu)->rt.rt_nr_running) + tg_capacity += calc_cpu_capacity(cpu); + else + tg_capacity += capacity_of(cpu); } rcu_read_unlock(); -- 2.25.1
反馈： 您发送到kernel@openeuler.org的补丁/补丁集，已成功转换为PR！ PR链接地址： https://gitee.com/openeuler/kernel/pulls/17161 邮件列表地址：https://mailweb.openeuler.org/archives/list/kernel@openeuler.org/message/GN6... FeedBack: The patch(es) which you have sent to kernel@openeuler.org mailing list has been converted to a pull request successfully! Pull request link: https://gitee.com/openeuler/kernel/pulls/17161 Mailing list address: https://mailweb.openeuler.org/archives/list/kernel@openeuler.org/message/GN6...
participants (2)
-
Cheng Yu -
patchwork bot