Re: [PATCH] sched: Add a new version sysctl to control child runs first

[Peter Zijlstra]

On Mon, Sep 13, 2021 at 11:37:31AM +0000, CGEL wrote:
> On Mon, Sep 13, 2021 at 10:13:54AM +0200, Peter Zijlstra wrote:
> > On Sun, Sep 12, 2021 at 04:12:23AM +0000, cgel.zte@xxxxxxxxx wrote:
> > > From: Yang Yang <yang.yang29@xxxxxxxxxx>
> > > 
> > > The old version sysctl has some problems. First, it allows set value
> > > bigger than 1, which is unnecessary. Second, it didn't follow the
> > > rule of capabilities. Thirdly, it didn't use static key. This new
> > > version fixes all the problems.
> > 
> > Does any of that actually matter?
> 
> For the first problem, I think the reason why sysctl_schedstats() only
> accepts 0 or 1, is suitbale for sysctl_child_runs_first(). Since
> task_fork_fair() only need sysctl_sched_child_runs_first to be
> zero or non-zero.

This could potentially break people that already write a larger value in
it -- by accident or otherwise.

> For the second problem, I remember there is a rule: try to
> administration system through capilities but not depends on
> root identity. Just like sysctl_schedstats() or other
> sysctl_xx().

It seems entirely daft to me; those files are already 644, if root opens
the file and passes it along, it gets to keep the pieces.

> For the thirdly problem, sysctl_child_runs_first maynot changes
> often, but may accessed often, like static_key delayacct_key
> controlled by sysctl_delayacct().

Can you actually show it makes a performance difference in a fork
micro-bench? Given the amount of gunk fork() already does, I don't think
it'll matter one way or the other, and in that case, simpler is better.