[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20200713125543.GJ10769@hirez.programming.kicks-ass.net>
Date: Mon, 13 Jul 2020 14:55:43 +0200
From: Peter Zijlstra <peterz@...radead.org>
To: Valentin Schneider <valentin.schneider@....com>
Cc: linux-kernel@...r.kernel.org, mingo@...nel.org,
vincent.guittot@...aro.org, dietmar.eggemann@....com,
morten.rasmussen@....com
Subject: Re: [PATCH v3 7/7] sched/topology: Use prebuilt SD flag degeneration
mask
On Wed, Jul 01, 2020 at 08:06:55PM +0100, Valentin Schneider wrote:
> Leverage SD_DEGENERATE_GROUPS_MASK in sd_degenerate() and
> sd_degenerate_parent().
>
> Note that this changes sd_degenerate() somewhat: I'm using the negation of
> SD_DEGENERATE_GROUPS_MASK as the mask of flags not requiring groups, which
> is equivalent to:
>
> SD_WAKE_AFFINE | SD_SERIALIZE | SD_NUMA
>
> whereas the current mask for that is simply
>
> SD_WAKE_AFFINE
>
> I played with a few toy NUMA topologies on QEMU and couldn't cause a
> different degeneration than what mainline does currently. If that is deemed
> too risky, we can go back to using SD_WAKE_AFFINE explicitly.
Arguably SD_SERIALIZE needs groups, note how we're only having that
effective for machines with at least 2 nodes. It's a bit shit how we end
up there, but IIRC that's what it ends up as.
SD_NUMA is descriptive, and not marking a group as degenerates because
it has SD_NUMA seems a bit silly. But then, it would be the top domain
and would survive anyway?
Powered by blists - more mailing lists