Extend wiki page table of distributions

[jmid]

We should extend the table in the wiki page https://github.com/ocaml/opam-repository/wiki/Depexts-os-distribution---os-family-values

First off for consistency with the second table, I suggest the first table should add an os column with the following entries:

• "linux"
• ...
• "linux"
• "macos"
• "macos"
• "freebsd"
• "openbsd"
• "netbsd"

Second, it would be useful to add new distribution lines:

• a NixOS line: os = linux | os-distribution = nixos | os-family = nixos from fix conf-oniguruma depext on nixos #28921 to help explain and expand the NixOS support we already have
• a Rocky line: os = linux | os-distribution = rocky | os-family = rhel from conf-openblas: fix installation on Rocky9 #28925 to document a seemingly popular RHEL variant
• a Mint line: os = linux | os-distribution = linuxmint | os-family = ubuntu to document a representative Ubuntu-family variant that I've helped add support for myself

I don't have write permission to make these changes myself unfortunately 🤷

(If I am allowed to dream, it would also be great to have these tested in the CI...)

[edwintorok]

Might be useful to treat all RHEL clones the same way, CentOS, Rocky, but there are others as well (I'm aware of AlmaLinux as well). Most of these tend to be binary compatible (or close to it), and they should at least be consistent at the source/build level.
Although CentOS nowadays is actually ahead of RHEL, and Rocky/AlmaLinux are the RHEL clones that CentOS used to be.

That is why I suggested that it'd be good if opam used pkg-config instead of relying on specific distro names. I think at least for RPM-like distributions, you can also install package by specifying their pkgconfig name using the provides/require mechanisms:
E.g.

rpm -q --provides openblas-devel
...
pkgconfig(openblas) = 0.3.26

sudo dnf install 'pkgconfig(openblas)'
...
Package openblas-devel-0.3.26-2.el9.x86_64 is already installed.
...

Unfortunately I'm not aware of an equivalent for Debian based distributions, so we have to rely on distribution name matching and package names/flags instead.
So even if we have to do distro name matching to figure out package names, for the actual compile flags/paths we should be able to rely on pkg-config/pkgconf (except for the odd package that doesn't have a .pc file).

Meanwhile using os-family instead of os-distribution might increase the chance that an opam package would work on more distributions?

[jmid]

OK, cool with pkgconfig-based installation! 😎

Regarding os-family I was wondering about this recommendation in the wiki-page:

> In the case of CentOS/Fedora/Oraclelinux, using os-family can get really painful with the addition of the EPEL repository so just using os-distribution in this case is usually easier.

I suspect this was added by @kit-ty-kate (polite ping) and that the recommendation is hard earned.
Kate: could you perhaps elaborate a bit?

[edwintorok]

See here for a discussion from a Haskell perspective on depending on pkg-config package names vs actual library names
https://discourse.haskell.org/t/whats-the-benefit-of-pkgconfig-depends/8670/12. There are some inconsistencies there as well (e.g. sometimes the pkg-config file is added by the distro, so not all distros would have them), so either way probably needs testing on all supported distros.

In the case of CentOS/Fedora/Oraclelinux, using os-family can get really painful with the addition of the EPEL repository so just using os-distribution in this case is usually easier.

Maybe we need a way to distinguish RHEL family from Fedora. Fedora shouldn't be considered part of the RHEL family (well first of all due EPEL as mentioned), but the wiki says that fedora has os-family = fedora. So is that a bug that has since been fixed in opam, or are there still subtle breakages when using os-family = rhel?

[kit-ty-kate]

Kate: could you perhaps elaborate a bit?

The original thought comes from ocaml/opam#4440. The context is that i was trying to find a standard way to write depexts fields so when you take into account the required EPEL repository for centos for many packages (at least at the time, i don't know about now) you would end up having one line for centos, one for other rpm-based distributions and one for fedora.

I think it should be safe to use os-family = "fedora" at the very least. For the rest of the RHEL variants i don't know.

So is that a bug that has since been fixed in opam, or are there still subtle breakages when using os-family = rhel?

I'm not sure what you are saying. Do you mean that you expect fedora to have os-family = rhel? On Linux, os-family is defined using the /etc/os-release file provided by linux distributions. If is set to the first value of ID_LIKE, or ID if the former is undefined. It seems to be a choice by Fedora to be viewed as a "base" distribution and not a RHEL derivative.

[edwintorok]

Kate: could you perhaps elaborate a bit?

The original thought comes from ocaml/opam#4440. The context is that i was trying to find a standard way to write depexts fields so when you take into account the required EPEL repository for centos for many packages (at least at the time, i don't know about now) you would end up having one line for centos, one for other rpm-based distributions and one for fedora.

Yes, that is quite tedious, but consistently using os-family for all of these would at least avoids problems with supporting new RHEL clones?
I'm hoping that actually my pkg-config idea would work for all these distros (if the pkg-config package names and their availability is consistent across all these distros, at least for the small subset of packages that the opam repository needs).
I'll have to do some experiments though in containers and report back.
If it works then the situation would become simpler, you'd need:

• 1 line for all RPM based distros (fedora/rhel/suse would all be the same)
• an additional line for RHEL based distros when the pakage is not available in the main repo (additional epel-repository dependency)
• 1 line for all debian-based
• 1 line for BSD
• 1 line for Mac
• ...

Still not very uniform, but at least the situation won't look so bad when it comes to build flags, which could use pkg-config consistently for all (I don't know about Windows, and whether pkgconf is available there).

Another possibility would be for the file to specify just the pkgconfig name, and some tool to look up the mapping (can be done by apt-file on Debian based distros I think) and generate the opam file with all the names.

I think it should be safe to use os-family = "fedora" at the very least. For the rest of the RHEL variants i don't know.

So is that a bug that has since been fixed in opam, or are there still subtle breakages when using os-family = rhel?

I'm not sure what you are saying. Do you mean that you expect fedora to have os-family = rhel?

No, it is fine as it is, my comment was a reply to this, which listed Fedora together with RHEL and mentioned EPEL.
Fedora is already a separate family from RHEL, so that avoids the EPEL issue already (you only need to add EPEL to os-family = rhel).

On Linux, os-family is defined using the /etc/os-release file provided by linux distributions. If is set to the first value of ID_LIKE, or ID if the former is undefined. It seems to be a choice by Fedora to be viewed as a "base" distribution and not a RHEL derivative.

I think that is correct.