Policy of explicit disabling of package features

Policy of explicit disabling of package features

[Tanu Kaskinen]

Hi,

I'm using the pulseaudio packaging as the starting point to learn about 
OE packaging. I'm trying to understand why everything is done as it is 
done, and my current focus is on pulseaudio's EXTRA_OECONF:

EXTRA_OECONF = "\
                 --disable-hal-compat \
                 --disable-orc \
                 --enable-tcpwrap=no \
                 --with-access-group=audio \
                 --disable-openssl \
                 --disable-xen \
                 --with-database=simple \
                 --without-fftw \
                 --without-zsh-completion-dir \
                 --with-udev-rules-dir=`pkg-config --variable=udevdir 
udev`/rules.d \
                 ac_cv_header_valgrind_memcheck_h=no \
"

I'm trying to understand the reason for each of those options. One 
question that I had was that why are some features explicitly disabled, 
when they should get anyway disabled automatically due to missing 
dependencies. I found out from the git logs that at least some of them 
are explicitly disabled, because the configure script may randomly find 
some dependencies installed even if they are not listed in DEPENDS, due 
to other packages pulling them in.

This then raises the question that what the policy of explicit vs. 
implicit disabling is, and also what the policy *should* be. It seems 
that the current policy is that things are explicitly disabled only 
after someone has run into problems with the non-determinism. Is that 
correct?

It seems to me that a better policy would be to always explicitly 
disable *all* features that have external dependencies that aren't 
listed in DEPENDS. That kind of policy should reduce these 
non-determinism issues. Full compliance with such policy may not be 
feasible to achieve, since it requires great care to check for new 
optional dependencies whenever updating to new upstream versions, but I 
hope that at least "pre-emptive" patches that add explicit feature 
disabling will be accepted even before anyone has run into actual problems.

-- 
Tanu

Re: Policy of explicit disabling of package features

[Richard Purdie]

On Fri, 2015-03-20 at 10:07 +0200, Tanu Kaskinen wrote:
> It seems to me that a better policy would be to always explicitly 
> disable *all* features that have external dependencies that aren't 
> listed in DEPENDS. That kind of policy should reduce these 
> non-determinism issues. Full compliance with such policy may not be 
> feasible to achieve, since it requires great care to check for new 
> optional dependencies whenever updating to new upstream versions, but I 
> hope that at least "pre-emptive" patches that add explicit feature 
> disabling will be accepted even before anyone has run into actual problems.

This is the policy, if its not in DEPENDS, it shouldn't be using it.

As you say, adherence to that policy can be tricky, its something we try
and improve over time. Where we do find issue we fix them as you've seen
from the patches. We do have better detection of these issues than we've
ever had before.

Cheers,

Richard

Re: Policy of explicit disabling of package features

[Tanu Kaskinen]

On Fri, 2015-03-20 at 09:30 +0000, Richard Purdie wrote:
> On Fri, 2015-03-20 at 10:07 +0200, Tanu Kaskinen wrote:
> > It seems to me that a better policy would be to always explicitly 
> > disable *all* features that have external dependencies that aren't 
> > listed in DEPENDS. That kind of policy should reduce these 
> > non-determinism issues. Full compliance with such policy may not be 
> > feasible to achieve, since it requires great care to check for new 
> > optional dependencies whenever updating to new upstream versions, but I 
> > hope that at least "pre-emptive" patches that add explicit feature 
> > disabling will be accepted even before anyone has run into actual problems.
> 
> This is the policy, if its not in DEPENDS, it shouldn't be using it.

Ok, excellent! I'll make a patch that adds some --disable-foo options to
pulseaudio.inc that are currently missing.

-- 
Tanu