Hacking modversions

[Mike Hearn]

On Tue, 01 Mar 2005 16:49:49 -0500, Dave Jones wrote:
> Congratulations. You win todays understatement of the day award.
> You have no idea how maddening doing this is for RHEL.
> If I had to do it for Fedora too, I think I'd be institutionalised by now.
> 
> To do this properly is an incredible amount of work. It's one of the reasons
> that our enterprise kernels stick at one version for their lifetimes,
> as to bend every change to fit the abi of your release kernel is just
> so time-consuming (and sometimes its just not possible to fix some stuff
> without breaking ABI).

Well, I didn't say "Fedora should ship a stable kernel interface". I said,
changes which do not require a recompile should not force one to occur.

That's very different. To use Arjans example, I don't care about functions
no longer accepting NULL pointers or whatever, because that isn't
something a recompile would fix. I also don't care (for this proposal) if
the ABI *does* change. 

What I was talking about is when I need to recompile a module, but there's
not really any reason for it. It's just modversions being modversions.
 
> For Fedora, I can't see this happening tbh.  For RHEL we get to be
> really picky about which upstream cset's we pull in. Conservatism is the
> name of the game there. Fedora users tend to want the latest and
> greatest bits and pieces, and pulling in individual csets is just going
> to be too much overhead given each point release upstream is currently
> churning out around 4000 csets.  To maintain any kind of illusion of the
> kabi you propose, you'd need to do this to throw out the bits that break
> kabi so drastically.

I'm not suggesting throwing bits out, rather:

- Being more conservative about what changes are pushed into
  fedora-updates, and just putting them off for FC4. In particular I was
  kind of surprised to see such huge changes in security patches!

- Using that conservatism to say "This security/driver fix clearly does
  not break something that a recompile would fix, so _this time_ let's not
  force a module rebuild".

That's quite different to actually maintaining a stable kernel ABI, which
isn't what I was asking for. 

Yes, that means that some new features may not get in until the next
Fedora release. So be it. Online updates should be about bugfixes and
security patches IMHO, not about adding or removing features once the OS
is deployed.

thanks -mike