In shrink_slab(), it directly goes to 'out' label only when it
can't get the lock of shrinker_rwsew. In this case, it doesn't
do the real work of shrinking slab, so we don't need trigger a
reschedule by cond_resched().
Signed-off-by: Xianting Tian <[email protected]>
---
mm/vmscan.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 466fc3144..676e97b28 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -687,8 +687,9 @@ static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
}
up_read(&shrinker_rwsem);
-out:
+
cond_resched();
+out:
return freed;
}
--
2.17.1
Thanks, my understanding is,
In shrink_slab(), do_shrink_slab() will do the real reclaim work, which will occupy current cpu and consume more cpu time, so we need to trigger a reschedule after reclaim.
But if it jumps to 'out' label, that means we don't do the reclaim work at this time, it won't cause other thread getting starvation, so we don't need to call cond_resched() in this case.
Is it right?
-----Original Message-----
From: Michal Hocko [mailto:[email protected]]
Sent: Friday, October 16, 2020 8:08 PM
To: tianxianting (RD) <[email protected]>
Cc: [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH] mm: vmscan: avoid a unnecessary reschedule in shrink_slab()
On Fri 16-10-20 11:39:52, Xianting Tian wrote:
> In shrink_slab(), it directly goes to 'out' label only when it can't
> get the lock of shrinker_rwsew. In this case, it doesn't do the real
> work of shrinking slab, so we don't need trigger a reschedule by
> cond_resched().
Your changelog doesn't explain why this is not needed or undesirable. Do you see any actual problem?
The point of this code is to provide a deterministic scheduling point regardless of the shrinker_rwsew.
>
> Signed-off-by: Xianting Tian <[email protected]>
> ---
> mm/vmscan.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/vmscan.c b/mm/vmscan.c index 466fc3144..676e97b28
> 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -687,8 +687,9 @@ static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
> }
>
> up_read(&shrinker_rwsem);
> -out:
> +
> cond_resched();
> +out:
> return freed;
> }
>
> --
> 2.17.1
>
--
Michal Hocko
SUSE Labs
On Fri 16-10-20 12:48:23, Tianxianting wrote:
> Thanks, my understanding is,
> In shrink_slab(), do_shrink_slab() will do the real reclaim work, which will occupy current cpu and consume more cpu time, so we need to trigger a reschedule after reclaim.
> But if it jumps to 'out' label, that means we don't do the reclaim work at this time, it won't cause other thread getting starvation, so we don't need to call cond_resched() in this case.
> Is it right?
You are almost right. But consider situation when the lock is taken for
quite some time. do_shrink_slab cannot make any forward progress and
effectivelly busy loop. Unless the caller does cond_resched it might
cause soft lockups.
Anyway let me try to ask again. Why does would this be any problem that
deserves a fix?
>
> -----Original Message-----
> From: Michal Hocko [mailto:[email protected]]
> Sent: Friday, October 16, 2020 8:08 PM
> To: tianxianting (RD) <[email protected]>
> Cc: [email protected]; [email protected]; [email protected]
> Subject: Re: [PATCH] mm: vmscan: avoid a unnecessary reschedule in shrink_slab()
>
> On Fri 16-10-20 11:39:52, Xianting Tian wrote:
> > In shrink_slab(), it directly goes to 'out' label only when it can't
> > get the lock of shrinker_rwsew. In this case, it doesn't do the real
> > work of shrinking slab, so we don't need trigger a reschedule by
> > cond_resched().
>
> Your changelog doesn't explain why this is not needed or undesirable. Do you see any actual problem?
>
> The point of this code is to provide a deterministic scheduling point regardless of the shrinker_rwsew.
>
> >
> > Signed-off-by: Xianting Tian <[email protected]>
> > ---
> > mm/vmscan.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/mm/vmscan.c b/mm/vmscan.c index 466fc3144..676e97b28
> > 100644
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -687,8 +687,9 @@ static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
> > }
> >
> > up_read(&shrinker_rwsem);
> > -out:
> > +
> > cond_resched();
> > +out:
> > return freed;
> > }
> >
> > --
> > 2.17.1
> >
>
> --
> Michal Hocko
> SUSE Labs
--
Michal Hocko
SUSE Labs
On Fri 16-10-20 13:20:41, Tianxianting wrote:
> Thanks
> I understood what you said :)
> But whether it is proper to check reschedule in every loop when lock is taken?
I do not see any actual problem TBH. cond_resched is mostly to increase
interactivity for non preemptible kernel. It can reduce throughput but
this is a memory reclaim path and I do not expect this to contribute to
any moderate hot paths. Direct reclaim doesn't really count as a hot
path.
--
Michal Hocko
SUSE Labs
On Fri 16-10-20 11:39:52, Xianting Tian wrote:
> In shrink_slab(), it directly goes to 'out' label only when it
> can't get the lock of shrinker_rwsew. In this case, it doesn't
> do the real work of shrinking slab, so we don't need trigger a
> reschedule by cond_resched().
Your changelog doesn't explain why this is not needed or undesirable. Do
you see any actual problem?
The point of this code is to provide a deterministic scheduling point
regardless of the shrinker_rwsew.
>
> Signed-off-by: Xianting Tian <[email protected]>
> ---
> mm/vmscan.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index 466fc3144..676e97b28 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -687,8 +687,9 @@ static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
> }
>
> up_read(&shrinker_rwsem);
> -out:
> +
> cond_resched();
> +out:
> return freed;
> }
>
> --
> 2.17.1
>
--
Michal Hocko
SUSE Labs
Thanks
I understood what you said :)
But whether it is proper to check reschedule in every loop when lock is taken?
By the way, I did not met a issue for this , I just learn this code and come up with one possible optimization based my understanding.
-----Original Message-----
From: Michal Hocko [mailto:[email protected]]
Sent: Friday, October 16, 2020 9:02 PM
To: tianxianting (RD) <[email protected]>
Cc: [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH] mm: vmscan: avoid a unnecessary reschedule in shrink_slab()
On Fri 16-10-20 12:48:23, Tianxianting wrote:
> Thanks, my understanding is,
> In shrink_slab(), do_shrink_slab() will do the real reclaim work, which will occupy current cpu and consume more cpu time, so we need to trigger a reschedule after reclaim.
> But if it jumps to 'out' label, that means we don't do the reclaim work at this time, it won't cause other thread getting starvation, so we don't need to call cond_resched() in this case.
> Is it right?
You are almost right. But consider situation when the lock is taken for quite some time. do_shrink_slab cannot make any forward progress and effectivelly busy loop. Unless the caller does cond_resched it might cause soft lockups.
Anyway let me try to ask again. Why does would this be any problem that deserves a fix?
>
> -----Original Message-----
> From: Michal Hocko [mailto:[email protected]]
> Sent: Friday, October 16, 2020 8:08 PM
> To: tianxianting (RD) <[email protected]>
> Cc: [email protected]; [email protected];
> [email protected]
> Subject: Re: [PATCH] mm: vmscan: avoid a unnecessary reschedule in
> shrink_slab()
>
> On Fri 16-10-20 11:39:52, Xianting Tian wrote:
> > In shrink_slab(), it directly goes to 'out' label only when it can't
> > get the lock of shrinker_rwsew. In this case, it doesn't do the real
> > work of shrinking slab, so we don't need trigger a reschedule by
> > cond_resched().
>
> Your changelog doesn't explain why this is not needed or undesirable. Do you see any actual problem?
>
> The point of this code is to provide a deterministic scheduling point regardless of the shrinker_rwsew.
>
> >
> > Signed-off-by: Xianting Tian <[email protected]>
> > ---
> > mm/vmscan.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/mm/vmscan.c b/mm/vmscan.c index 466fc3144..676e97b28
> > 100644
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -687,8 +687,9 @@ static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
> > }
> >
> > up_read(&shrinker_rwsem);
> > -out:
> > +
> > cond_resched();
> > +out:
> > return freed;
> > }
> >
> > --
> > 2.17.1
> >
>
> --
> Michal Hocko
> SUSE Labs
--
Michal Hocko
SUSE Labs
Thanks for the explain.
I got it.
-----Original Message-----
From: Michal Hocko [mailto:[email protected]]
Sent: Friday, October 16, 2020 9:45 PM
To: tianxianting (RD) <[email protected]>
Cc: [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH] mm: vmscan: avoid a unnecessary reschedule in shrink_slab()
On Fri 16-10-20 13:20:41, Tianxianting wrote:
> Thanks
> I understood what you said :)
> But whether it is proper to check reschedule in every loop when lock is taken?
I do not see any actual problem TBH. cond_resched is mostly to increase interactivity for non preemptible kernel. It can reduce throughput but this is a memory reclaim path and I do not expect this to contribute to any moderate hot paths. Direct reclaim doesn't really count as a hot path.
--
Michal Hocko
SUSE Labs