Re: [patch] PID namespace design bug, workaround

[Ingo Molnar <mingo@elte.hu>]

* Pavel Emelyanov <xemul@openvz.org> wrote:

> The "fix" I mention is just returning -EINVAL in case user orders 
> CLONE_NEWPIDS and compiling out all the namespace cloning code. This 
> is just a more elegant way to get rid of pid namespaces rather than 
> Ingo proposed.

unfortunately i have to NACK that approach. We never allowed broken 
user-space visible APIs into the kernel like that because it just gives 
a vector for that breakage to become de-facto used and forced upon the 
core kernel. Even if they can be .config turned off. That's just a lame 
excuse that delays the fixing of it. We may mark features that have a 
good expectation to be fixed as CONFIG_EXPERIMENTAL, and we may mark 
drivers that nobody maintains anymore as CONFIG_BROKEN, but we dont 
introduce new core syscall features with CONFIG_BROKEN! We never did and 
i hope we never will.

The _only_ way to force the fixing of such type of breakages is to not 
offer them _at all_. Really, you are proposing a major new extension to 
lots of important core Linux APIs so please try to solve this problem 
cleanly, it's really severe. Right now as things stand this containers 
sub-feature is "a little bit pregnant". This is one of the few cases 
where we really _must_ say no.

	Ingo