svn commit: r350550 - head/share/mk

[Rodney W. Grimes]

> On 07/08/2019 11:00, John Baldwin wrote:
> > On 8/6/19 9:56 AM, Glen Barber wrote:
> >> On Sat, Aug 03, 2019 at 01:06:18AM +0000, John Baldwin wrote:
> >>> Author: jhb
> >>> Date: Sat Aug  3 01:06:17 2019
> >>> New Revision: 350550
> >>> URL: https://svnweb.freebsd.org/changeset/base/350550
> >>>
> >>> Log:
> >>>    Flip REPRODUCIBLE_BUILD back to off by default in head.
> >>>    
> >>>    Having the full uname output can be useful on head even with
> >>>    unmodified trees or trees that newvers.sh fails to recognize as
> >>>    modified.
> >>>    
> >>>    Reviewed by:	emaste
> >>>    Differential Revision:	https://reviews.freebsd.org/D20895
> >>>
> >> I would like to request this commit be reverted.  While the original
> >> commit message to enable this knob stated the commit would be reverted
> >> after stable/12 branched, I have seen no public complaints about
> >> enabling REPRODUCIBLE_BUILD by default (and quite honestly, do not see
> >> the benefit of disabling it by default -- why wouldn't we want
> >> reproducibility?).
> >>
> >> To me, this feels like a step backwards, with no tangible benefit.
> >> Note, newvers.sh does properly detect a modified tree if it can find
> >> the VCS metadata directory (i.e., .git, .svn) -- I know this because
> >> I personally helped with it.
> >>
> >> In my opinion, those that want the non-reproducible metadata included in
> >> output from 'uname -a' should set WITHOUT_REPRODUCIBLE_BUILDS in their
> >> src.conf.  Turning off a sane default for the benefit of what I suspect
> >> is likely a short list of use cases feels like a step in the wrong
> >> direction.
> > My arguments for flipping this in head (and head only) are that the data
> > provided in uname -a when this is disabled is useful for development, and
> > that in head we do tailor settings towards development (e.g. GENERIC in
> > head vs GENERIC in stable).
> >
> > The logic to handle modified trees has an inherent assumption that I think
> > is false, at least for my workflow and I suspect many others.  I do builds
> > and tests of kernels on separate machines (VMs or bare metal) from where I
> > use VCS to manage sources so that a kernel crash doesn't toast my source
> > tree.  The trees are then shared to the build/test machines via NFS.  As
> > a result, the build/test machines are not always able to detect that the
> > tree is modified either because a subset of the checkout is exported via
> > NFS, or the VCS tool isn't installed on the build/test machines because
> > they are generally barebones systems with only a base installed.  This
> > does mean that flipping the knob off doesn't provide all of the same info,
> > but it does provide the path, and the path matters because 'kgdb -n last'
> > uses it, and because if you use separate directories for separate projects
> > (e.g. git worktrees), then the path tells you which test kernel you booted.
> > (It is not uncommon for me to have several test projects in flight on a
> > single test machine for different branches.)
> >
> > In the original discussion on arch, we collectively recognized that
> > developer builds vs release builds were different and needed different
> > defaults.  The compromise reached at that time was to depend on the VCS
> > to detect developer builds to choose the policy.  What I have found is that
> > in practice for at least my workflow that doesn't actually work.  I posit
> > that the majority of kernels built from head are developer builds, not
> > releases, and that the default should cater to that.  You could also always
> > patch release.sh to set WITH_REPRODUCIBLE_BUILD in the environment which I
> > think would give a more accurate sense of when builds are releases or not.
> >
> > However, I will yield to whatever the consensus is.
> 
> +1 keeping metadata in head.

I am conflicted on this one, and I think there is a reasonable argument
on both sides, but from what I have read here this appears to be mostly
the kernel that is at issue, loss of the meta data from newvers.sh in
the kernel is infact a PITA, even on stable or production release
systems.

I propose a compromise, add 2 knobs:
WITHOUT_REPRODUCIBLE_KERNEL	(aka get your metadata in uname)
WITH_REPRODUCIBLE_USERLAND	(aka reproducible userland)

WITH{,OUT}_REPRODUCIBLE_BUILD overrides both, for backwards compat,
and neither should be defined by default.

Somehow set WITH_REPRODUCIBLE_KERNEL for builds of GENERIC
for releases/snapshots, but do not ship the system with it
set (I can here a growl from Glen on this)  Thus we build
a reproducible kernel and ship it with the system but if
the user builds a kernel it gets meta data to indicate it
is no longer a stock kernel.

FYI, upon finding I could not figure out what kernel I was running
after installing 12.0 release I turnd off REPRODUCIBLE on my kernel
build VM for 12.0.  I do leave it on if I am building userland.

Thoughts?

-- 
Rod Grimes                                                 rgrimes at freebsd.org