On Sun, 2021-03-28 at 18:33 -0400, Mark H Weaver wrote:
> Earlier, I wrote:
> > One thing to be very careful about is to only use 'gtk-doc/stable',
> > 'dblatex/stable', and 'imagemagick/stable' in native-inputs, and
> > moreover to make sure that no references to these */stable packages
> > remain in any package outputs.
> > 
> > Of course, if any package retains references to its 'native-inputs',
> > that's always a bug, but I wouldn't be surprised if such bugs exist in
> > Guix.  Such bugs might be relatively harmless now (except when
> > cross-compiling), but they could become a security bug if a package
> > retains a reference to 'imagemagick/stable'.

It just occurred to me: could we automatically add all native-inputs
to #:disallowed-references when cross-compiling?  This shouldn't break
any packages, except possibly when cross-compiling.

Or stronger, add all native-inputs to #:disallowed-references (unless
they are also in inputs or propagated-inputs), even when compiling
natively?

Problems include:
* I guess a world rebuild, as the derivations would be different.
* In some places we have the following pattern:

    (native-inputs
     `(("autoconf" ,autoconf)
       ("automake" ,automake)
       ("pkg-config" ,pkg-config)
       ,@(if (%current-target-system)
             `(("guile" ,guile-3.0))   ;for 'guild compile' and 'guile-3.0.pc'
             '())))
    (inputs
     `(("guile" ,guile-3.0)
       ("lzlib" ,lzlib)))
    (synopsis "Guile bindings to lzlib")

  The (if (%current-target-system) ...) would need to be made unconditional.
* I guess an option to disable this behaviour might be useful.

> It occurs to me that we will need some tooling to ensure that no
> references to these buggy "*/stable" packages end up in package outputs
> that users actually use.  Otherwise, it is likely that sooner or later,
> a runtime reference to one of these buggy packages will sneak in to our
> systems.
> 
> An initial idea is that these "*/stable" packages could have a package
> property (perhaps named something like 'build-time-only') that indicates
> that references to its outputs should not occur within the outputs of
> any other package that does not have that same property.

Would this be (a) something enforced by the build process (using
#:disallowed-references or #:allowed-references), or (b) a linter?

> We'd also need to somehow ensure that users don't install these
> 'build-time-only' packages directly, at least not without an additional
> option (e.g. --force-unsafe-build-time-only) to override it.

What about a developer running "guix environment eom"?  IIUC, this would
make the developer vulnerable (at least, once I've gotten around replacing
the 'gtk-doc' input with 'gtk-doc/stable'), so it might make sense to
replace /stable -> unstable packages here.

However, now the developer ends up with a different set of packages than
wil be seen in the build environment ...

> Additionally, it might be good to issue warnings if 'build-time-only'
> packages are not hidden,

This seems good to me.  This should prevent
"guix install imagemagick@bad-version".

>  or if they are found within the 'inputs' or
> 'propagated-inputs' fields of any package that's not also
> 'build-time-only'.  Both of these last two checks have loopholes,
> however, so they are not reliable indicators.

But these (automatic "guix lint") checks could still catch many
problems in practice before they are committed!  

> Thoughts?  Other proposals?

Is this something you will be writing "guix lint" checkers (or other
checkers) for yourself?

Greetings,
Maxime.