Re: [PATCH 0/2] finx argv_split() vs sysctl race

[Andi Kleen <andi@firstfloor.org>]

On Sat, Mar 16, 2013 at 10:23:51PM +0100, Oleg Nesterov wrote:
> On 03/16, Andi Kleen wrote:
> >
> > > Perhaps rcu can be better, although a global rwsem looks simpler,
> > > I dunno.
> >
> > It's a general problem with lots of sysctls.
> > >
> > > But argv_split() or its usage should be changed anyway, and GFP_KERNEL
> > > won't work under rcu_read_lock().
> >
> > rcu strings has a helper function to copy the string for sleepy cases.
> 
> Then you need to pre-allocate, take rcu_read_lock(), copy, and check
> that it actually fits the pre-allocated buffer. Not sure why the simple
> rwsem is worse.

The reason I did it originally like that was that some of the sysctls weren't
as "slow path" as power off. And for anything that is even moderately
often used a global lock is going to hurt eventually. The "read" in the
sem also doesn't help because it's still a hot cache line. 
I agree if it the goal was only to fix poweroff RCU is somewhat
overkill and a global lock would be fine.

> But I won't argue in any case
> 
> > > To me 1/2 looks as a simplification anyway, but I won't argue if we
> > > decide to add rcu/locking and avoid this patch.
> >
> > Ok I'll revisit.
> 
> OK, but do you agree with 1/2?

It doesn't solve the race alone because when the 0 byte can move it's
not safe to run kstrndup() in parallel. Ok given the n and that it 
force terminates it could only lead to some junk at the end.

But it seems like a useful small optimization, although I don't know
if it's used in any non slow paths.

I assume you audited all callers that they comprehend that they need
to free differently now.

-Andi