Re: [PATCH] Re: [RFC PATCH] namespaces: fix leak on fork() failure

[Mike Galbraith <efault@gmx.de>]

On Fri, 2012-05-04 at 00:55 -0700, Eric W. Biederman wrote: 
> Mike Galbraith <efault@gmx.de> writes:
> 
> > Namespaces have something in common with cgroups.  synchronize_rcu()
> > makes them somewhat less than wonderful for dynamic use.
> 
> Well unlike cgroups namespaces were not designed for heavy dynamic use.
> Although it appears that vsftp puts them to that kind of use so some
> of the design decisions are with revisiting.

Yeah, the testcase was distilled from vsftp, so it must be beating on
namespaces pretty hard to induce a bug report.

> > default flags = SIGCHLD
> >
> > -namespace:  flag |= CLONE_NEWPID 
> > -all:  flags |= CLONE_NEWIPC | CLONE_NEWNET | CLONE_NEWUSER
> >
> > marge:/usr/local/tmp/starvation # ./hackbench
> > Running with 10*40 (== 400) tasks.
> > Time: 2.636
> > marge:/usr/local/tmp/starvation # ./hackbench -namespace
> > Running with 10*40 (== 400) tasks.
> > Time: 11.624
> > marge:/usr/local/tmp/starvation # ./hackbench -namespace -all
> > Running with 10*40 (== 400) tasks.
> > Time: 51.474
> 
> CLONE_NEWUSER?  I presume you have applied my latest user namespace
> patches?  Otherwise you are running completely half baked code.

I was testing in mainline.  While fiddling with the testcase and leakage
monitor script, I decided to see what happens with all namespace flags.
The others didn't cause any leakage, but did make things slow down.

> hackbench?  Which kernel are you running.  Hackbench in some kernels is
> really good at triggering cache ping-pong effects with pids, and creds.
> So I'm not certain what to say there.  In the latest kernels things
> should be better with unix domain sockets as long as you don't actually
> ask to pass your creds but hackbench is still a pretty ridiculous
> benchmark.  Oversharing is always going to be bad for performance.

Hackbench was just to show the price of hefty namespace usage.

> > You can create trash quickly, but you have to haul it away.
> 
> Well synchronize_rcu is much better in that respect than call_rcu, which
> let's the trash build up but is never carried away.
> 
> The core design assumption with namespaces is that they will be used
> much more than they will be created/destroyed, and as long as there are
> progress guarantees in place I don't have a problem with that.   At the
> same time if there are easy things we can do to make things go faster
> I am in favor of that notion.
> 
> Still especially in the case of hackbench I think it is worth asking the
> question how much of the slow down is due to cache ping-pong due to
> oversharing.

Dunno, and doubt I'll have time to tinker with it more.  Darn bugzilla
thing keeps knocking on my mailbox with interesting bugs in places I
know _diddly spit_ about.. like namespaces.

-Mike