Re: [Devel] [PATCH] pidns: Place under CONFIG_EXPERIMENTAL (take 2)

[Adrian Bunk <bunk@kernel.org>]

On Fri, Oct 26, 2007 at 07:31:04PM -0600, Eric W. Biederman wrote:
> Adrian Bunk <bunk@kernel.org> writes:
> 
> > CONFIG_EXPERIMENTAL is a weak hint that some code might not (yet) be in 
> > a perfect state, but it does not have any semantics regarding
> > userspace ABIs.
> 
> Code that might not (yet) be in a perfect state sums it up pretty
> well.  There is not plan or expectation to change magic numbers or
> things like that but the behavior of the code may change as bug and
> such are fixed.
> 
> > A dependency on BROKEN seems more appropriate.
> 
> Since you can't select that it seems a little strong.
> 
> ...
> 
> One of the things we talked about at the kernel summit is how
> almost inevitably when new user space interfaces are introduced
> there are problems.  Someone over looks something, something
> gets changed to get through the review something like that.  There was
> discussion but no consensus on how do introduce something like that
> but still allow our selves the ability to fix it.  Keeping the
> code under CONFIG_EXPERIMENTAL is the best suggest I have seen
> so far.  Even if it is slightly expanding the definition of
> CONFIG_EXPERIMENTAL.
> 
> Every place the kernel uses pids is a huge scope.  It is very
> easy to miss something with a scope that wide.  So the engineer
> in me says chances of us missing something are pretty huge.
> Especially since I know we have bugs in -rc1.
> 
> If it turns out that making multiple pid namespaces work is
> hopeless we can always change the dependency to BROKEN later.

No, we can't after 2.6.24 got released.

Let me make an example:
- looking at the timelines, e.g. Ubuntu 8.04 LTS is likely to
  ship with 2.6.24
- this experimental feature might be enabled there
- this Ubuntu release with this kernel will be supported and used
  for five years

> As for ABI and behavioral characteristics currently that is
> largely well defined.  You are supposed to get the exact
> same thing as you would on the system if you only had a
> single pid namespace.  The places where we have questionable
> semantics is in the intersections between namespaces.  
> 
> That is not an area I am willing to stand up and say we got
> it perfect the first time, I'm going to support our behavior
> quirks forever if I can find soon enough.  Very few applications
> will care, and the differences might really matter at some point.
> 
> So does any one have any better suggestions on how to deal
> with features that are enough work you aren't going to get them
> perfect the first time.  You need the code merged or else you
> can not complete the feature (too many dependencies through out the
> code).  You want early adopters to start playing with the feature
> so you can get feed back and you can test to make certain everything
> is going ok.  You want to retain the ability to fix implementation
> details even if those details are user visible, for a time until
> things seem as good as they can reasonably get.
> 
> Roughly that sounds like CONFIG_EXPERIMENTAL to me.  But I would
> be happy to hear if someone has a better idea.

There is a difference between "complete the feature" and "early adopters 
to start playing with the feature" on the one side, and making something 
available in a released kernel on the other side.

For development and playing with it it can depend on BROKEN (perhaps 
with the dependency removed through the first -rc kernels), but as soon 
as it's available in a -final kernel the ABI is fixed.

> Eric

cu
Adrian

-- 

       "Is there not promise of rain?" Ling Tan asked suddenly out
        of the darkness. There had been need of rain for many days.
       "Only a promise," Lao Er said.
                                       Pearl S. Buck - Dragon Seed