[PATCH] sched/fair: Fix update_min_vruntime() to get proper min_vruntime

[PATCH] sched/fair: Fix update_min_vruntime() to get proper min_vruntime

[Byungchul Park]

I am sorry for making you confused. The patch I posted one year ago,
Commit 97a7142 'sched/fair: Make update_min_vruntime() more readable',
has a bug, which you picked up.

Please add this patch to fix it, or combine these two patches into
one, or revert commit 97a7142. Whatever you want. Sorry.

Thank you,
Byungchul

-----8<-----
>From 295895be8befbab040d6054bb8186c03daabcedd Mon Sep 17 00:00:00 2001
From: Byungchul Park <byungchul.park@lge.com>
Date: Tue, 6 Sep 2016 12:22:26 +0900
Subject: [PATCH] sched/fair: Fix update_min_vruntime() to get proper
 min_vruntime

Commit 97a7142 'sched/fair: Make update_min_vruntime() more readable'
introduces a bug that cfs_rq gets a wrong min_vruntime if
!cfs_rq->rb_leftmost && cfs_rq->curr. This fixes it and makes it more
readable and simple.

Signed-off-by: Byungchul Park <byungchul.park@lge.com>
---
 kernel/sched/fair.c | 17 ++++++++---------
 1 file changed, 8 insertions(+), 9 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index a6820b3..0a5f666b 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -462,21 +462,20 @@ static inline int entity_before(struct sched_entity *a,
 
 static void update_min_vruntime(struct cfs_rq *cfs_rq)
 {
-	u64 vruntime = cfs_rq->min_vruntime;
-
-	if (cfs_rq->rb_leftmost) {
-		struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
-						   struct sched_entity,
-						   run_node);
+	u64 vruntime = U64_MAX;
 
-		vruntime = se->vruntime;
-	}
+	if (cfs_rq->rb_leftmost)
+		vruntime = rb_entry(cfs_rq->rb_leftmost,
+				    struct sched_entity,
+				    run_node)->vruntime;
 
 	if (cfs_rq->curr)
 		vruntime = min_vruntime(vruntime, cfs_rq->curr->vruntime);
 
 	/* ensure we never gain time by being placed backwards. */
-	cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
+	if (vruntime != U64_MAX)
+		cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
+
 #ifndef CONFIG_64BIT
 	smp_wmb();
 	cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
-- 
1.9.1

Re: [PATCH] sched/fair: Fix update_min_vruntime() to get proper min_vruntime

[Peter Zijlstra]

On Tue, Sep 06, 2016 at 02:05:17PM +0900, Byungchul Park wrote:
> From 295895be8befbab040d6054bb8186c03daabcedd Mon Sep 17 00:00:00 2001
> From: Byungchul Park <byungchul.park@lge.com>
> Date: Tue, 6 Sep 2016 12:22:26 +0900
> Subject: [PATCH] sched/fair: Fix update_min_vruntime() to get proper
>  min_vruntime
> 
> Commit 97a7142 'sched/fair: Make update_min_vruntime() more readable'
> introduces a bug that cfs_rq gets a wrong min_vruntime if
> !cfs_rq->rb_leftmost && cfs_rq->curr. This fixes it and makes it more
> readable and simple.

Urgh, I actually stared at that patch for quite a time and somehow
convinced myself it was good. I actually considered that scenario I
think.

/me thinks more..

Argh, I'm an idiot, you're right. By using min_vruntime() on
cfs_rq->curr we take the leftmost and fail to advance min_vruntime in
that case.

I'll ask Ingo to take the patch out.

Re: [PATCH] sched/fair: Fix update_min_vruntime() to get proper min_vruntime

[Byungchul Park]

On Tue, Sep 06, 2016 at 10:48:39AM +0200, Peter Zijlstra wrote:
> On Tue, Sep 06, 2016 at 02:05:17PM +0900, Byungchul Park wrote:
> > From 295895be8befbab040d6054bb8186c03daabcedd Mon Sep 17 00:00:00 2001
> > From: Byungchul Park <byungchul.park@lge.com>
> > Date: Tue, 6 Sep 2016 12:22:26 +0900
> > Subject: [PATCH] sched/fair: Fix update_min_vruntime() to get proper
> >  min_vruntime
> > 
> > Commit 97a7142 'sched/fair: Make update_min_vruntime() more readable'
> > introduces a bug that cfs_rq gets a wrong min_vruntime if
> > !cfs_rq->rb_leftmost && cfs_rq->curr. This fixes it and makes it more
> > readable and simple.
> 
> Urgh, I actually stared at that patch for quite a time and somehow
> convinced myself it was good. I actually considered that scenario I
> think.
> 
> /me thinks more..
> 
> Argh, I'm an idiot, you're right. By using min_vruntime() on
> cfs_rq->curr we take the leftmost and fail to advance min_vruntime in
> that case.
> 
> I'll ask Ingo to take the patch out.

Ok. But I think the combined one is more readable.
Don't you think so. It's ok even if you don't.

Thank you,
Byungchul