sched: add vslice
authorPeter Zijlstra <a.p.zijlstra@chello.nl>
Mon, 15 Oct 2007 15:00:10 +0000 (17:00 +0200)
committerIngo Molnar <mingo@elte.hu>
Mon, 15 Oct 2007 15:00:10 +0000 (17:00 +0200)
add vslice: the load-dependent "virtual slice" a task should
run ideally, so that the observed latency stays within the
sched_latency window.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
include/linux/sched.h
kernel/sched.c
kernel/sched_fair.c

index 2c33227..d74830c 100644 (file)
@@ -908,6 +908,7 @@ struct sched_entity {
        u64                     sum_exec_runtime;
        u64                     vruntime;
        u64                     prev_sum_exec_runtime;
+       u64                     last_min_vruntime;
 
 #ifdef CONFIG_SCHEDSTATS
        u64                     wait_start;
index 5004dff..fe1165b 100644 (file)
@@ -1615,6 +1615,7 @@ static void __sched_fork(struct task_struct *p)
        p->se.exec_start                = 0;
        p->se.sum_exec_runtime          = 0;
        p->se.prev_sum_exec_runtime     = 0;
+       p->se.last_min_vruntime         = 0;
 
 #ifdef CONFIG_SCHEDSTATS
        p->se.wait_start                = 0;
@@ -6495,6 +6496,7 @@ static inline void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
 #ifdef CONFIG_FAIR_GROUP_SCHED
        cfs_rq->rq = rq;
 #endif
+       cfs_rq->min_vruntime = (u64)(-(1LL << 20));
 }
 
 void __init sched_init(void)
index 5db7bd1..87acc5c 100644 (file)
@@ -243,6 +243,15 @@ static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
        return period;
 }
 
+static u64 __sched_vslice(unsigned long nr_running)
+{
+       u64 period = __sched_period(nr_running);
+
+       do_div(period, nr_running);
+
+       return period;
+}
+
 /*
  * Update the current task's runtime statistics. Skip current tasks that
  * are not in our scheduling class.
@@ -441,32 +450,33 @@ static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
 static void
 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
 {
-       u64 min_runtime, latency;
+       u64 vruntime;
 
-       min_runtime = cfs_rq->min_vruntime;
+       vruntime = cfs_rq->min_vruntime;
 
        if (sched_feat(USE_TREE_AVG)) {
                struct sched_entity *last = __pick_last_entity(cfs_rq);
                if (last) {
-                       min_runtime = __pick_next_entity(cfs_rq)->vruntime;
-                       min_runtime += last->vruntime;
-                       min_runtime >>= 1;
+                       vruntime += last->vruntime;
+                       vruntime >>= 1;
                }
-       } else if (sched_feat(APPROX_AVG))
-               min_runtime += sysctl_sched_latency/2;
+       } else if (sched_feat(APPROX_AVG) && cfs_rq->nr_running)
+               vruntime += __sched_vslice(cfs_rq->nr_running)/2;
 
        if (initial && sched_feat(START_DEBIT))
-               min_runtime += sched_slice(cfs_rq, se);
+               vruntime += __sched_vslice(cfs_rq->nr_running + 1);
 
        if (!initial && sched_feat(NEW_FAIR_SLEEPERS)) {
-               latency = sysctl_sched_latency;
-               if (min_runtime > latency)
-                       min_runtime -= latency;
+               s64 latency = cfs_rq->min_vruntime - se->last_min_vruntime;
+               if (latency < 0 || !cfs_rq->nr_running)
+                       latency = 0;
                else
-                       min_runtime = 0;
+                       latency = min_t(s64, latency, sysctl_sched_latency);
+               vruntime -= latency;
        }
 
-       se->vruntime = max(se->vruntime, min_runtime);
+       se->vruntime = vruntime;
+
 }
 
 static void
@@ -478,6 +488,7 @@ enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
        update_curr(cfs_rq);
 
        if (wakeup) {
+               /* se->vruntime += cfs_rq->min_vruntime; */
                place_entity(cfs_rq, se, 0);
                enqueue_sleeper(cfs_rq, se);
        }
@@ -492,8 +503,8 @@ static void
 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
 {
        update_stats_dequeue(cfs_rq, se);
-#ifdef CONFIG_SCHEDSTATS
        if (sleep) {
+#ifdef CONFIG_SCHEDSTATS
                if (entity_is_task(se)) {
                        struct task_struct *tsk = task_of(se);
 
@@ -502,8 +513,11 @@ dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
                        if (tsk->state & TASK_UNINTERRUPTIBLE)
                                se->block_start = rq_of(cfs_rq)->clock;
                }
-       }
 #endif
+               /* se->vruntime = entity_key(cfs_rq, se); */
+               se->last_min_vruntime = cfs_rq->min_vruntime;
+       }
+
        if (se != cfs_rq->curr)
                __dequeue_entity(cfs_rq, se);
        account_entity_dequeue(cfs_rq, se);