Recent Mk/bsd.perl.mk changes (r320679)

[Jeremy Chadwick]

On Wed, Jun 26, 2013 at 08:40:52AM +0200, Baptiste Daroussin wrote:
> On Tue, Jun 25, 2013 at 11:12:19PM -0700, Jeremy Chadwick wrote:
> > On Wed, Jun 26, 2013 at 09:42:37AM +0400, Andrej Zverev wrote:
> > > Hello, and first please accept my apologies for this situation.
> > 
> > Understood, I just hope this can get addressed/fixed sooner than later,
> > because what we have right now is reproducible breakage.  :-)
> > 
> > > > pkg_add -r perl   (this will install perl-5.14.2_3.tbz)
> > > > svn up /usr/ports
> > > > cd /usr/ports/whatever/p5-whatever
> > > > make install
> > > > pkg_delete p5-whatever
> > > 
> > > As I know we are never supported mixing of ports and packages.
> > > If you initially installed something from package and decide to use
> > > ports in this case better to rebuild all or stay with packages.
> > 
> > And here is where you are sadly mistaken.  While I cannot find any sort
> > of official statement on the "mixing of the two", the fact of the matter
> > is (and possibly you did not know this -- I'm not sure) that packages
> > are built **from** ports ("make package" does the magic).  This is
> > confirmed as well:
> > 
> > http://www.freebsd.org/ports/index.html
> > http://www.freebsd.org/doc/en/articles/linux-users/software.html
> > 
> > Please be aware when I say "package" I am talking about the .tbz balls
> > utilised by the base system's pkg_* tools (not pkg(8)) and used by the
> > OS installer and so on.  For pkg(8) I believe poudriere is used to make
> > the packages.  I don't use/can't speak about pkg(8) et al.
> 
> Wrong and totally wrong, the way the ports tree works with pkg(8) is exactly the
> same way it works with pkg_install (make package does the same).
> 
> poudriere is just some script to make sure everything is done in the right
> order, cleanly and works transparently with both pkg(8) and pkg_install.

Thanks for the clarification and the education on this point -- like I
said, I have no experience with pkg(8) and its kin.

> > For both pkg_* tools and ports, both use the same default base path
> > (/usr/local), **and** both register their bits in /var/db/pkg.
> 
> Wrong and wrong, ports knows nothing about /var/db/pkg, ports ask the package
> tool to register a package may that be pkg_install or pkg(8) this tool will
> register the package where ever it seems suitable to itself meaning /var/db/pkg
> in the case of both pkg(8) and pkg_install.

Please go look at Mk/bsd.port.mk and look up PKG_DBDIR.  Just go look
for it throughout the entire file and look at how/where it's used.  It's
used *heavily* throughout dependency checking and actual installation
(think "Registering installation of ..." messages and what happens under
the hood, if I remember right -- otherwise you have a port installed
that you cannot pkg_delete).

> > Package installation does not utilise any part of the ports/Mk/*
> > framework, but that isn't anything new -- it's been like that since at
> > least the 2.2.x days, possibly earlier.
> > 
> > So in summary, when making changes to ports/Mk/*, you really have to
> > think about the combination of the two, and make sure you don't break
> > anything when changing key pieces of those framework bits.
> 
> That is exactly what has been done in the case of the change.

But only when looking forward, not when looking at existing
installations or new installations.  I will repeat my instructions for
how to reproduce the problem:

pkg_add -r perl   (this will install perl-5.14.2_3.tbz)
svn up /usr/ports
cd /usr/ports/whatever/p5-whatever
make install
pkg_delete p5-whatever

Please go do it on a fresh FreeBSD 9.2-RELEASE system and watch what
happens -- it breaks, and that breakage results in leftover shit in
LOCALBASE.

> > > > What I'd like to know:
> > > >
> > > > - Why the major.minor.patchlevel --> major.minor path change in the
> > > > first place.  I have never, ever seen this done anywhere on any *IX
> > > > system I've used.  Where's the justification?  Was this discussed on
> > > > some perl mailing list somewhere as a "new and better way"?  It's
> > > > essentially saying "x.y.z is always going to be compatible with x.y.z+1"
> > > > which is not true (particularly with XS, as I understand it).  Where
> > > > was this discussed publicly?
> > > 
> > > http://docs.freebsd.org/cgi/getmsg.cgi?fetch=26605+0+archive/2013/freebsd-perl/20130609.freebsd-perl
> > > 
> > > I don't want to start yet another bikeshed here. Maybe link above will
> > > make some things more clear to you.
> > 
> > Thanks -- I wasn't aware of the freebsd-perl mailing list.
> > 
> > I just finished reading the entire thread.  The justification seems to
> > be "because Fedora and Debian do it" (that's the best I can find).
> > Okay, fine/great/whatever -- I guess that's a form of justification,
> > just not one I was hoping for.  I won't dwell on this at this point -- I
> > got the answer I was looking for.
> 
> No the justification is that we use to have a perl-after-upgrade script to
> workaround the fact that we used major.minor.patchlevel my bypassing the package
> tool to modify directly the content of the package database and more some files
> on the file system each time a new perl update comes (like perl 5.12.2 ->
> 5.12.4) without this script being run every single minor update of perl was
> requiring manual intervention to fix up all the installed packages whereever
> they came from ports or package etc.

That sounds like a mess, and I agree getting rid of that sounds great.
I wasn't aware of that piece.  But my stand on the breakage due to the
backwards compatibility shims not being there stands.

> The reference concerning Fedora/Debian was just when some asked if we were doing
> something exotic to say no this is a "normal way" of doing things, look others
> do and the perl build system is extecting use to do that.

On the Solaris systems I use to administrate, I remember the paths
containing major.minor.patchlevel.  I doubt I'll find anything out about
this in any of the perl documentation online though.

> > I see no actual harm in moving to major.minor, but the issue here is
> > that backwards compatibility in Mk/bsd.perl.mk for existing perl
> > installations.
> > 
> > Many people, for example, do not want to build X.org from source -- so
> > they pkg_add -r X, then build other bits/pieces related to X from
> > ports/source.  Even more people do this for OpenOffice/LibreOffice or
> > whatever it's called today ( :-) ).
> > 
> > This "combination" needs to be supported.  I know it hurts to have to
> > retain backwards compatibility, but it's necessary given how all of this
> > was designed.
> 
> This has never been supported neither been unsupported, this commit doesn't
> changed anything here.

See above, re: what the commit has broken.

> > Such backwards compatibility could be removed, as I said, once
> > sufficient time has passed, particularly once the FreeBSD versions
> > shipping with a ports tree snapshot with perl versions older than those
> > in r320679 have been EOL'd.  After that you could remove the framework
> > and cite EOL on such support.  This is normal too -- for example on
> > occasion FreeBSD [67].x people show up on -ports and complain that
> > something won't build/some part of the Mk/* framework is broken for
> > them, and EOL is the proper response to that.
> 
> Such compatibility hasn't been removed but improved with this patch:
> old behaviour update from perl-X.Y.Z -> perl-X.Y.Z+1 then the perl path was
> changed and all compatibility was broken.
> new behaviour update from perl-X.Y.Z -> perl-X.Y.Z+1: nothing to do

See above, re: only looking forward, not looking at existing support and
what happens if someone does what I said.

> > I know this probably won't hold any ground because I'm not one any
> > longer, but I was a FreeBSD ports committer myself some years ago (~5),
> > so please don't think that I'm just flying off the handle without some
> > existing familiarity with things.
> > 
> > If there is truly going to be a "split" between ports and packages
> > in this way, then someone needs to start doing SVN tagging/branches
> > for major changes like this.
> 
> This has never been the matter of the patch and noone said that their will be a
> split like this in the futur.

Please correct me if I'm wrong here -- and god do I hope I am -- but the
short version sounds to me like you're saying "nothing is wrong with the
commit, nothing is broken, there is no problem, what are you complaining
about?"  I want to make sure I understand if that's what you're saying
before I react past this point.  But I would urge you to go do what I
said, re: those 5 commands, and watch the breakage for yourself.

Of course, I guess the alternate "solution" would be to have whoever
maintains the official package sites (not for pkg, but pkg_*) update all
the packages in Latest/ so that they refer to the new path convention...

-- 
| Jeremy Chadwick                                   jdc at koitsu.org |
| UNIX Systems Administrator                http://jdc.koitsu.org/ |
| Making life hard for others since 1977.             PGP 4BD6C0CB |