bug#55043: Some packages depend on nss-certs, some bundle it.

[Liliana Marie Prikler <liliana.prikler@ist.tugraz.at>]

Am Mittwoch, dem 20.04.2022 um 17:22 +0200 schrieb Maxime Devos:
> X-Debbugs-CC: Hartmut Goebel <h.goebel@crazy-compilers.com>
> 
> Hi,
> 
> There are some packages bundling CA certificates:
> 
>  * nss-certs / le-certs (this one is not a problem)
>  * python-certifi
>  * perl-mozilla-ca
>  * rust-webpki-roots
>  * erlang-certifi (not yet, see
> <https://issues.guix.gnu.org/54796#3>)
>  * go-github-com-certifi-gocertifi
> 
> Worse, these packages have many dependencies!
> 
> $ guix refresh -l nss-certs le-certs python-certifi perl-mozilla-ca
> rust-webpki-roots 
> Het bouwen van het volgende 534 pakketten zorgt ervoor dat 1575
> afhankelijke pakketten opnieuw worden gebouwd: ...
This is only if you update all of them at once.  An update to any one
of them might go to staging instead, which for the record I don't see
merged too often in recent times.  Also, with the exception of our
snowflake static linkers Rust and Go, we should be able to graft them
on master.

> Why is this a problem?
> 
>  * I don't think that anybody is actually looking into keeping
>    python-certifi / perl-mozilla-ca / rust-webpki-roots / ...
>    up to date.  Security problems!
>  * Even so, this seems a waste of time to me, why not just use
>    $SSL_CERT_DIR / $SSL_CERT_FILE instead?
Point taken, I think these might just be different programming language
implementation that essentially store the same certificate in a known
location and then refer to it.  Which Guix can already do by specifying
SSL_CERT_DIR or SSL_CERT_FILE in a wrapper script.
>  * Lots of rebuilds to update things.
>  * (relatively minir) Allowing overriding the certificates trusted
> with
>    $SSL_CERT_DIR / $SSL_CERT_FILE would be nice.
I don't think SSL_CERT_DIR and SSL_CERT_FILE customization is useful in
all apps however.  I think there is a legitimate concern, that users
could be tricked into downloading a malicious certificate chain and
then running their instant messaging app with that, enabling a man in
the middle attack.  Again, Guix could solve this by hardcoding
SSL_CERT_*, but...

> Also relevant to the third point: some packages depend on nss-certs.
We'd be increasing this number if instead of language-specific certifi
packages, we consolidated them into nss-certs.  On the other hand, we
would be reducing the number of dependents in libraries that don't need
to hardcode certs for security reasons, so perhaps life will become
easier at some point.

Also, w.r.t. updating nss-certs, since we do have grafts, security
updates should not be a concern in my opinion.  Making sure that all
our security-critical instant messaging applications use our trusted
bundle rather than their own very curated set is of higher priority in
my opinion.

> I've heard an argument in favour of just using the certifi packages
> instead of using our own certificates:
> 
> > (from Hartmut Goebel, at <https://issues.guix.gnu.org/54796#52>)
> > Neither python-certifi nor gocertifi build on nss-cert. Addind some
> > update mechanism into the Guix package is not a good idea IMO: This
> > would make “erlang-certif@2.9.0“ contain different certificates
> > than the release 2.9.0, making debugging a hell.
> 
> ... but I don't follow, it's just a different set of certificates,
> could you elaborate? 
From a library/app developer's point of view, you can't rely on the
system certificate store if your system's vendor starts with M, ends
with t and has 7 letters in between.  I'm not sure how missing erlang-
certif would impact building any of its packages, but perhaps they
found a way to break builds if some certificate is missing.

> Proposal:
> 
>  * eventually remove python-certifi, perl-mozilla-ca, ... because
>    nobody appears to be keeping them up-to-date and for security it
>    is important for them to be up to date.
Sure.  In the meantime, however, I'd suggest mocking them by having
them simply return nss-certs or whatever else we have for a trusted
certificate bundle.

>  * likewise, forbid new packages from being included as-is if they
>    depend on a certifi package or nss-certs.
See above.

>  * Look into removing the certifi packages from the inputs of
>    packages, submitting patches to upstream for using $SSL_CERT_... /
>    /etc/ssl/certs ... as appropriate.
That I agree with.


Cheers