svn commit: r248352 - in stable/9: etc share/mk

[John Baldwin]

On Tuesday, April 02, 2013 1:59:03 pm Brooks Davis wrote:
> On Wed, Mar 20, 2013 at 09:18:08AM -0400, John Baldwin wrote:
> > On Tuesday, March 19, 2013 4:06:31 pm Brooks Davis wrote:
> > > On Tue, Mar 19, 2013 at 09:49:47PM +0400, Dmitry Morozovsky wrote:
> > > > On Tue, 19 Mar 2013, Brooks Davis wrote:
> > > > 
> > > > > > >   Replace all known uses of ln in the build process with appropriate
> > > > > > >   install -l invocations via new INSTALL_LINK and INSTALL_SYMLINK
> > > > > > >   variables.
> > > > > > 
> > > > > > It seems this merge breaks ``make distribution'' and hence mergemaster if your 
> > > > > > base system is not updated yet (for example, while updating jail):
> > > > > 
> > > > > Sorry for the delay in responding.  I missed this yesterday.
> > > > > 
> > > > > It works for me on a older 9.0-STABLE system where the base install
> > > > > doesn't support -l.  Did you build world or run "make toolchain" in that
> > > > > source tree to build the bootstrap copy of install?
> > > > 
> > > > Yes, this is after full ``make buildworld buildkernel'' process.
> > > 
> > > I've found the problem thanks to misc/177055.  It is that mergemaster
> > > (and etcupdate) set MAKEOBJDIRPREFIX to something in their
> > > temporary directory and thus deprive themselves of bootstrap tools.
> > > Unfortunately, I don't see a trivial fix so I've backed this out for
> > > now and will work on this in HEAD.
> > 
> > Hummmm.  In the case of etcupdate you can use 'etcupdate -B'.  That is actually safe
> > to do in the common case where you've just updated /usr/src and built the corresponding
> > world in /usr/obj.  It should possibly even by the default for etcupdate if a DESTDIR
> > is not specified.
> 
> Finally getting back to this...
> 
> etcupdate -B would correct the immediate problem for etcupdate.  I do
> think that making it the default if the tree exists makes sense.  It
> won't be more broken than a cross installworld is.

Hmmm, checking for the obj tree is a bit hackish.  I'd rather it were more
deterministic.  I think I'd like to make it just default to -B and require
a new -b flag to build a new tree, but perhaps have it check for a tree
and error out if it doesn't exist and you don't give it -b?

> I did a quick test when I first found this issue and it would be easy to
> reuse the existing MAKEOBJDIRPREFIX in mergemaster as well.
> 
> I think we'll want to update UPDATING to recommend that the
> mergemaster -p stage (and the equivalent for etcupdate) be run using the
> version in the source tree, not the installed one.  I do wonder if it
> would make sense for them to attempt to find and invoke that version so
> simplify bootstrapping.

Currently etcupdate doesn't implement something like -p.  I need to add that as
I would prefer to use its conflict resolution for adding users.  (That would
also let it serve as a full replacement for mergemaster for those who prefer it.)

-- 
John Baldwin