Sched_Boost小结
之前遇到一个耗电问题,最后发现是/proc/sys/kernel/sched_boost节点设置异常,一直处于boost状态。导致所有场景功耗上升。
现在总结一下sched_boost的相关知识。
Sched_Boost
sched_boost主要是通过影响Task placement的方式,来进行boost。它属于QTI EAS中的一部分。
默认task placement policy
计算每个cpu的负载,并将task分配到负载最轻的cpu上。如果有多个cpu的负载相同(一般是都处于idle),那么就会把task分配到系统中capacity最大的cpu上。
设置sched_boost
通过设置节点:/proc/sys/kernel/sched_boost 或者内核调用sched_set_boost()函数,可以进行sched_boost,并且在分配任务时,忽略对energy的消耗。
boost一旦设置之后,就必须显示写0来关闭。同时也支持个应用同时调用设置,设置会选择boost等级最高的生效; 而当所有应用都都关闭boost时,boost才会真正失效。
boost等级
sched_boost一共有4个等级,除了0代表关闭boost以外,其他3个等级灵活地控制功耗和性能的不同倾向程度。
在通过节点设置,会调用sched_boost_handler
{ .procname = "sched_boost", .data = &sysctl_sched_boost, .maxlen = sizeof(unsigned int), .mode = 0644, .proc_handler = sched_boost_handler, .extra1 = &neg_three, .extra2 = &three, },
经过verify之后,调用_sched_set_boost来设置boost。
int sched_boost_handler(struct ctl_table *table, int write, void __user *buffer, size_t *lenp, loff_t *ppos) { int ret; unsigned int *data = (unsigned int *)table->data; mutex_lock(&boost_mutex); ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); if (ret || !write) goto done; if (verify_boost_params(*data)) _sched_set_boost(*data); else ret = -EINVAL; done: mutex_unlock(&boost_mutex); return ret;
而通过内核调用的方式,同样最后也是调用_sched_set_boost来设置boost。
int sched_set_boost(int type) { int ret = 0; mutex_lock(&boost_mutex); if (verify_boost_params(type)) _sched_set_boost(type); else ret = -EINVAL; mutex_unlock(&boost_mutex); return ret; }
接下来,我们看关键的设置函数_sched_set_boost:
static void _sched_set_boost(int type) { if (type == 0) //通过type参数判断是否enable/disable boost sched_boost_disable_all(); //(1)disable all boost else if (type > 0) sched_boost_enable(type); //(2) enable boost else sched_boost_disable(-type); //(3) disable boost /* * sysctl_sched_boost holds the boost request from * user space which could be different from the * effectively enabled boost. Update the effective * boost here. */ sched_boost_type = sched_effective_boost(); sysctl_sched_boost = sched_boost_type; set_boost_policy(sysctl_sched_boost); //(4) 设置boost policy trace_sched_set_boost(sysctl_sched_boost); }
首先看一下sched_boost的4个用于控制配置的结构体:
其中refcount来记录设置的次数。enter函数表示切换到该boost配置的动作;exit则是退出该boost配置的动作。
static struct sched_boost_data sched_boosts[] = { [NO_BOOST] = { .refcount = 0, .enter = sched_no_boost_nop, .exit = sched_no_boost_nop, }, [FULL_THROTTLE_BOOST] = { .refcount = 0, .enter = sched_full_throttle_boost_enter, .exit = sched_full_throttle_boost_exit, }, [CONSERVATIVE_BOOST] = { .refcount = 0, .enter = sched_conservative_boost_enter, .exit = sched_conservative_boost_exit, }, [RESTRAINED_BOOST] = { .refcount = 0, .enter = sched_restrained_boost_enter, .exit = sched_restrained_boost_exit, }, };
(1)disable all boost
调用除no boost外,所有boost配置的exit函数并且将他们的refcount清0。
#define SCHED_BOOST_START FULL_THROTTLE_BOOST #define SCHED_BOOST_END (RESTRAINED_BOOST + 1 static void sched_boost_disable_all(void) { int i; for (i = SCHED_BOOST_START; i < SCHED_BOOST_END; i++) { if (sched_boosts[i].refcount > 0) { sched_boosts[i].exit(); sched_boosts[i].refcount = 0; } } }
(2) enable boost
refcount记录调用次数+;
由于sched+boost支持多应用同时调用的,所以在设置boost之前,要先检查当前有效的boost配置。
优先级是No boost > Full Throttle > Conservative > Restrained。
static void sched_boost_enable(int type) { struct sched_boost_data *sb = &sched_boosts[type]; int next_boost, prev_boost = sched_boost_type; sb->refcount++; //refcount记录次数+1 if (sb->refcount != 1) return; /* * This boost enable request did not come before. * Take this new request and find the next boost * by aggregating all the enabled boosts. If there * is a change, disable the previous boost and enable * the next boost. */ next_boost = sched_effective_boost(); //设置boost之前,检查当前有效的boost配置 if (next_boost == prev_boost) return; sched_boosts[prev_boost].exit(); //调用之前配置的exit,退出之前的boost sched_boosts[next_boost].enter(); //调用现在配置的enter,进入当前boost状态
通过检查refcount,来确认当前有效的boost。
static int sched_effective_boost(void) { int i; /* * The boosts are sorted in descending order by * priority. */ for (i = SCHED_BOOST_START; i < SCHED_BOOST_END; i++) { if (sched_boosts[i].refcount >= 1) return i; } return NO_BOOST; }
(3)disable boost
同样假如是disable boost的话,就会相应的对refcount--,并且调用当前boost类型的exit函数来退出boost。
因为sched_boost支持多种boost同时开启,并按优先级设置。所以当disable一种boost时,最后通过检查当前有效的boost来进入余下优先级高的boost模式。
static void sched_boost_disable(int type) { struct sched_boost_data *sb = &sched_boosts[type]; int next_boost; if (sb->refcount <= 0) return; sb->refcount--; if (sb->refcount) return; /* * This boost's refcount becomes zero, so it must * be disabled. Disable it first and then apply * the next boost. */ sb->exit(); next_boost = sched_effective_boost(); sched_boosts[next_boost].enter(); }
(4)设置boost policy
在最后一步中,设置policy来体现task是否需要进行up migrate。
如下是sched_boost不同等级对应的up migrate迁移策略。
Full throttle和Conservative:SCHED_BOOST_ON_BIG---在进行task placement时,仅考虑capacity最大的cpu core
无:SCHED_BOOST_ON_ALL---在进行task placement时,仅不考虑capacity最小的cpu core
No Boost和Restrained:SCHED_BOOST_NONE---正常EAS
/* * Scheduler boost type and boost policy might at first seem unrelated, * however, there exists a connection between them that will allow us * to use them interchangeably during placement decisions. We'll explain * the connection here in one possible way so that the implications are * clear when looking at placement policies. * * When policy = SCHED_BOOST_NONE, type is either none or RESTRAINED * When policy = SCHED_BOOST_ON_ALL or SCHED_BOOST_ON_BIG, type can * neither be none nor RESTRAINED. */ static void set_boost_policy(int type) { if (type == NO_BOOST || type == RESTRAINED_BOOST) { //conservative和full throttle模式才会进行向上迁移 boost_policy = SCHED_BOOST_NONE; return; } if (boost_policy_dt) { boost_policy = boost_policy_dt; return; } if (min_possible_efficiency != max_possible_efficiency) { //左边是cpu中efficiency最小值,右边为最大值。big.LITTLE架构应该恒成立 boost_policy = SCHED_BOOST_ON_BIG; return; } boost_policy = SCHED_BOOST_ON_ALL; }
接下来详细分析3种boost设置的原理:
Full Throttle
full throttle(全速)模式下的sched boost,主要有如下2个动作:
(1)core control
(2)freq aggregation
static void sched_full_throttle_boost_enter(void) { core_ctl_set_boost(true); //(1)core control walt_enable_frequency_aggregation(true); //(2)freq aggregation }
(1)core control:isoloate/unisoloate cpu cores;enable boost时,开所有cpu core
int core_ctl_set_boost(bool boost) { unsigned int index = 0; struct cluster_data *cluster; unsigned long flags; int ret = 0; bool boost_state_changed = false; if (unlikely(!initialized)) return 0; spin_lock_irqsave(&state_lock, flags); for_each_cluster(cluster, index) { //修改并记录每个cluster的boost状态 if (boost) { boost_state_changed = !cluster->boost; ++cluster->boost; } else { if (!cluster->boost) { ret = -EINVAL; break; } else { --cluster->boost; boost_state_changed = !cluster->boost; } } } spin_unlock_irqrestore(&state_lock, flags); if (boost_state_changed) { index = 0; for_each_cluster(cluster, index) //针对每个cluster,apply boost设置 apply_need(cluster); } trace_core_ctl_set_boost(cluster->boost, ret); return ret; } EXPORT_SYMBOL(core_ctl_set_boost);
static void apply_need(struct cluster_data *cluster) { if (eval_need(cluster)) //判断是否需要 wake_up_core_ctl_thread(cluster); //唤醒cluster的core control thread }
具体如何判断的:
enable boost时:判断是否需要unisolate cpu,
disable boost时:判断need_cpus < active_cpus是否成立。
并且与上一次更新的间隔时间满足 > delay time。
static bool eval_need(struct cluster_data *cluster) { unsigned long flags; struct cpu_data *c; unsigned int need_cpus = 0, last_need, thres_idx; int ret = 0; bool need_flag = false; unsigned int new_need; s64 now, elapsed; if (unlikely(!cluster->inited)) return 0; spin_lock_irqsave(&state_lock, flags); if (cluster->boost || !cluster->enable) { need_cpus = cluster->max_cpus; //当enable boost时,设置need_cpus为所有cpu } else { cluster->active_cpus = get_active_cpu_count(cluster); //当disable boost时,首先获取active的cpu thres_idx = cluster->active_cpus ? cluster->active_cpus - 1 : 0; list_for_each_entry(c, &cluster->lru, sib) { bool old_is_busy = c->is_busy; if (c->busy >= cluster->busy_up_thres[thres_idx] || sched_cpu_high_irqload(c->cpu)) c->is_busy = true; else if (c->busy < cluster->busy_down_thres[thres_idx]) c->is_busy = false; trace_core_ctl_set_busy(c->cpu, c->busy, old_is_busy, c->is_busy); need_cpus += c->is_busy; } need_cpus = apply_task_need(cluster, need_cpus); //根据task需要,计算need_cpus } new_need = apply_limits(cluster, need_cpus); //限制need_cpus范围:cluster->min_cpus <= need_cpus <= clusterr->max_cpus need_flag = adjustment_possible(cluster, new_need); //(*)enable boost时:判断是否需要unisolate cpu; disable boost时:判断need_cpus < active_cpus是否成立 last_need = cluster->need_cpus; now = ktime_to_ms(ktime_get()); if (new_need > cluster->active_cpus) { ret = 1; //enable boost } else { /* * When there is no change in need and there are no more * active CPUs than currently needed, just update the * need time stamp and return. //当需要的cpu没有变化时,只需要更新时间戳,然后return */ if (new_need == last_need && new_need == cluster->active_cpus) { cluster->need_ts = now; spin_unlock_irqrestore(&state_lock, flags); return 0; } elapsed = now - cluster->need_ts; ret = elapsed >= cluster->offline_delay_ms; //修改need_cpus的时间要大于delay时间,才认为有必要进行更改 } if (ret) { cluster->need_ts = now; //更新时间戳,need_cpus cluster->need_cpus = new_need; } trace_core_ctl_eval_need(cluster->first_cpu, last_need, new_need, ret && need_flag); spin_unlock_irqrestore(&state_lock, flags); return ret && need_flag; }
满足更新要求的条件后,就会唤醒core control thread
static void wake_up_core_ctl_thread(struct cluster_data *cluster) { unsigned long flags; spin_lock_irqsave(&cluster->pending_lock, flags); cluster->pending = true; spin_unlock_irqrestore(&cluster->pending_lock, flags); wake_up_process(cluster->core_ctl_thread); }
其中会有一个检测pending的防止重入的操作。假如pending标志已经改写,那么就会将当前进程移出rq。
static int __ref try_core_ctl(void *data) { struct cluster_data *cluster = data; unsigned long flags; while (1) { set_current_state(TASK_INTERRUPTIBLE); //先退出RUNNING状态,设为TASK_INTERRUPTIBLE,后面再调用schedule()。会判断当前task已不处于TASK_RUNNING状态,随后会进行dequeue,并调度其他进程到rq->curr。(典型的出rq操作)
spin_lock_irqsave(&cluster->pending_lock, flags); if (!cluster->pending) { //检测pending,如果已经core control完成,则直接进行schedule(),并准备退出该thread spin_unlock_irqrestore(&cluster->pending_lock, flags); schedule(); if (kthread_should_stop()) break; spin_lock_irqsave(&cluster->pending_lock, flags); } set_current_state(TASK_RUNNING); cluster->pending = false; spin_unlock_irqrestore(&cluster->pending_lock, flags); do_core_ctl(cluster); //do work } return 0; }
static void __ref do_core_ctl(struct cluster_data *cluster) { unsigned int need; need = apply_limits(cluster, cluster->need_cpus); //再次check need_cpus是否合法 if (adjustment_possible(cluster, need)) { //再次check是否需要更改 pr_debug("Trying to adjust group %u from %u to %u\n", cluster->first_cpu, cluster->active_cpus, need); if (cluster->active_cpus > need) //根据need进行cpu un/isolate try_to_isolate(cluster, need); else if (cluster->active_cpus < need) try_to_unisolate(cluster, need); } }
(2)freq aggregation:提升cpu freq
设置flag:
static inline void walt_enable_frequency_aggregation(bool enable) { sched_freq_aggr_en = enable; }
后续在以下这一个地方会影响freq选择:
- load freq polcy时,会根据是否开启freq_aggregation的flag而影响load的计算。
开启flag之后,会使用aggr_grp_load替代rq->grp_time.prev_runnable_sum
aggr_grp_load是当前cluster中所有cpu core的rq->grp_time.prev_runnable_sum的总和:sum。
那么开启之后,就会使load变大,从而提升cpu freq或者进行
static inline u64 freq_policy_load(struct rq *rq) { 。。。 if (sched_freq_aggr_en) load = rq->prev_runnable_sum + aggr_grp_load; else load = rq->prev_runnable_sum + rq->grp_time.prev_runnable_sum; 。。。 }
Conservative
static void sched_conservative_boost_enter(void) { update_cgroup_boost_settings(); //(1)更新cgroup boost设置 sched_task_filter_util = sysctl_sched_min_task_util_for_boost; //(2)task util调节 }
(1)遍历group,group一般有以下这些:backgroun,foreground,top-app,rt。
将所有sched_boost_enabled设置为false,除了写了no_override的group。而从init.target.rc中可以看到top-app和foregroung写了该flag(如下),所以最后就会有
write /dev/stune/foreground/schedtune.sched_boost_no_override 1 write /dev/stune/top-app/schedtune.sched_boost_no_override 1
void update_cgroup_boost_settings(void) { int i; for (i = 0; i < BOOSTGROUPS_COUNT; i++) { if (!allocated_group[i]) break; if (allocated_group[i]->sched_boost_no_override) continue; allocated_group[i]->sched_boost_enabled = false; } }
(2)修改min_task_util的门限
/* 1ms default for 20ms window size scaled to 1024 */ unsigned int sysctl_sched_min_task_util_for_boost = 51; //conservative 保守 /* 0.68ms default for 20ms window size scaled to 1024 */ unsigned int sysctl_sched_min_task_util_for_colocation = 35; //normal
目前仅发现在如下2处会用到这个值:
1、获取sched_boost_enabled仍然为true的task。过滤task_util较小的task,关闭boost。
下面这个函数是返回task boost policy的,也就是placement boost。函数主要逻辑:
task所在group中打开了sched_boost_endabled,并且sched_boost设置非0: 假如sched_boost=1(full throttle),则policy是SCHED_BOOST_ON_BIG 假如sched_boost=2(Conservative),则还需要判断task util是否超过sched_task_filter_util。超过了,policy=SCHED_BOOST_ON_BIG;没超过,policy=SCHED_BOOST_NONE
根据上面看到只有top-app和foreground这2个group是打开了sched_boost_enabled的。
所以,这里就是将top-app和foreground中的task_util较大的task挑选出来,enable了boost;而其他的task,都关闭了boost。
static inline enum sched_boost_policy task_boost_policy(struct task_struct *p) { enum sched_boost_policy policy = task_sched_boost(p) ? sched_boost_policy() : SCHED_BOOST_NONE; if (policy == SCHED_BOOST_ON_BIG) { /* * Filter out tasks less than min task util threshold * under conservative boost. */ if (sched_boost() == CONSERVATIVE_BOOST && task_util(p) <= sched_task_filter_util) //修改了这个门限 policy = SCHED_BOOST_NONE; } return policy; }
2、在更新walt负载时
更新了unfilter的计数,增大到51,相当于:/* 1ms default for 20ms window size scaled to 1024 */(原先默认为:35,即0.68ms)。也就是当demand_scaled大于这个门限后,会设置一个nr_windows的缓冲时间。缓冲时间每一次update_history就会减1,直到为0。
static void update_history(struct rq *rq, struct task_struct *p, u32 runtime, int samples, int event) { 。。。 if (demand_scaled > sched_task_filter_util) p->unfilter = sysctl_sched_task_unfilter_nr_windows; else if (p->unfilter) p->unfilter = p->unfilter - 1; 。。。 }
p->unfilter这个参数又会在如下两个地方用于判断:
首先是在fair.c中,但是由于现在是CONSERVATIVE_BOOST,所以,只会return false。这条暂不分析下去。
static inline bool task_skip_min_cpu(struct task_struct *p) { return sched_boost() != CONSERVATIVE_BOOST && get_rtg_status(p) && p->unfilter; }
另外一个地方是如下walt.h中,由于p->unfilter非0,那么就会判断当前cpu是否为小核,如果是小核,那么return true。
static inline bool walt_should_kick_upmigrate(struct task_struct *p, int cpu) { struct related_thread_group *rtg = p->grp; if (is_suh_max() && rtg && rtg->id == DEFAULT_CGROUP_COLOC_ID && rtg->skip_min && p->unfilter) return is_min_capacity_cpu(cpu); return false; }
返回值会影响如下函数的返回值,也是return flase。
所以,会告诉系统当前的task load与当前cpu的capacity不匹配,需要进行迁移。
此外,在进行迁移时,也是调用该函数,用于排除目标cpu为小核。
综上,就是将task从小核迁移到大核。
static inline bool task_fits_max(struct task_struct *p, int cpu) { unsigned long capacity = capacity_orig_of(cpu); unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val; unsigned long task_boost = per_task_boost(p); if (capacity == max_capacity) return true; if (is_min_capacity_cpu(cpu)) { if (task_boost_policy(p) == SCHED_BOOST_ON_BIG || task_boost > 0 || schedtune_task_boost(p) > 0 || walt_should_kick_upmigrate(p, cpu)) //条件为true return false; } else { /* mid cap cpu */ if (task_boost > TASK_BOOST_ON_MID) return false; } return task_fits_capacity(p, capacity, cpu); }
Restrained
该模式下,从代码看,仅打开了freq aggregation。与上面全速模式类似,不再赘述。
static void sched_restrained_boost_enter(void) { walt_enable_frequency_aggregation(true); }
总结各个boost的效果
Full throttle:
1、通过core control,将所有cpu都进行unisolation
2、通过freq聚合,将load计算放大。从而触发提升freq,或者迁移等
3、通过设置boost policy= SCHED_BOOST_ON_BIG,迁移挑选target cpu时,只会选择大核
最终效果应该尽可能把任务都放在大核运行(除了cpuset中有限制)
Conservative:
1、通过更新group boost配置,仅让top-app和foreground组进行task placement boost
2、提高min_task_util的门限,让进行up migrate的条件更苛刻。只有load较大(>1ms)的task,会进行up migrate。
2、同上,更改min_task_util门限后,会提醒系统task与cpu是misfit,需要进行迁移。
3、通过设置boost policy= SCHED_BOOST_ON_BIG,迁移挑选target cpu时,只会选择大核
最终效果:top-app和foreground的一些task会迁移到大核运行
Restrained:
1、通过freq聚合,将load计算放大。从而触发提升freq,或者迁移等
load放大后,仍遵循基本EAS。提升freq或者迁移,视情况而定。
注意关于early_datection部分简述参考之前文章:https://www.cnblogs.com/lingjiajun/p/12317090.html 中CPU freq调节章节。不过代码中貌似没有看到会影响的地方。由于EAS部分本人还没有仔细学习分析代码,可能还有疏漏或者错误,欢迎交流。