[PATCH] gnu: zfs: Split into packages specific for each of our major supported kernel versions.

Fixes: https://issues.guix.gnu.org/45401

Untested --- no idea how to repoint my guix to a temporary commit on my local file repo, would appreciate any guidance.

From 242ee858b057f1c3ca4a80c9c449f710c5b1ff96 Mon Sep 17 00:00:00 2001

supported kernel versions.

---

gnu/packages/file-systems.scm | 31 ++++++++++++++++++++++++++++++-

1 file changed, 30 insertions(+), 1 deletion(-)

--

2.29.2

On Thu, Dec 24, 2020 at 10:59:56AM +0000, raid5atemyhomework via Guix-patches via wrote:

There are primarily two ways to use Guix with a local Git repo.

First, you can follow the instructions in the manual section

Contributing, specifically the sections Building from Git and Running

Guix Before It Is Installed.

Or, you can use something like `guix pull --url=/path/to/my/repo

--commit=mycommit`.

Beyond that, does this patch create linux-libre packages with ZFS

support compiled in?

Thank you very much, will check.

No, this patch continues to use the existing technique of downloading ZFS as source and compiling it as a kernel module you have to load in, as per licensing incompatibility.

What it changes is that it compiles kernel modules for all guix-provided kernel versions that ZFS currently is rated to support, rather than compile a kernel module for whatever `(default-linux)` is, which is always the latest and greatest kernel (currently 5.10) which ZFS is currently not rated for.

As Guix goes through the trouble of maintaining a `linux-libre-4.4` package I assume at least one user somewhere is using Linux 4.4, and did not want to prevent that user from using ZFS if they decide to do so someday.

I admit the naming scheme `linux-5.9-zfs` does make it seem like the package contains an entire Linux kernel.

The alternative is to call it `zfs-for-linux-5.9` but that makes the entire package name with version `zfs-for-linux-5.9-0.8.5` which I think is worse, because the version numbers are right next to each other.

What naming scheme would you suggest?

I imagine a similar problem would occur for other kernel modules --- the internal kernel interfaces are *not* stable at all, so complex external kernel modules like ZFS need to know what kernel version it's being compiled for and include a bunch of `#ifdef`s to compensate.

So we need some decent convention for "how do we name kernel modules so that they are targeted for each kernel version".

Hullo RAIDperson!

Thanks for working to improve Guix's currently broken ZFS package!

raid5atemyhomework via Guix-patches via ???

However, I oppose this (and not because ‘rated’ sets off my

corpo-speak alarms :-). It makes your system.scm tedious to

maintain, especially if we'd do this for all modules (and why

not?).

Worse, there's no such thing as ‘a 5.10’ Linux module that loads

on ‘a 5.10’ Linux kernel. We still need an easy generic method to

build modules against their chosen kernel and configuration.

As your patch illustrates, that already exists:

It just belongs in your system.scm, not in Guix itself.

Kind regards,

T G-R

P.S.nitpick: unfortunately this breaks localisation and can't be

done:

Descriptions need to be simple self-contained strings, even if it

means duplicating entire paragraphs.

On Fri, 25 Dec 2020 05:09:44 +0100

Tobias Geerinckx-Rice via Guix-patches via <guix-patches@gnu.org> wrote:

Why is this needed? KERNEL-LOADABLE-MODULES are already automatically

adapted in this way.

See this in gnu/system.scm :

(define* (operating-system-directory-base-entries os)

"Return the basic entries of the 'system' directory of OS for use as the

value of the SYSTEM-SERVICE-TYPE service."

(let* ((locale (operating-system-locale-directory os))

(kernel (operating-system-kernel os))

(hurd (operating-system-hurd os))

(modules (operating-system-kernel-loadable-modules os))

(kernel (if hurd

kernel

(profile

(content (packages->manifest

(cons kernel

(map (lambda (module)

(if (package? module)

(package-for-kernel kernel <-----

module)

module))

modules))))

(hooks (list linux-module-database)))))

(initrd (and (not hurd) (operating-system-initrd-file os)))

(params (operating-system-boot-parameters-file os)))

`(("kernel" ,kernel)

,@(if hurd `(("hurd" ,hurd)) '())

("parameters" ,params)

,@(if initrd `(("initrd" ,initrd)) '())

("locale" ,locale)))) ;used by libc

This doesn't work, for two reasons:

* So, for example we can't just remove the "module" output and have the `.ko` installed in the default "out" output, our core problem is we need to compile a ZFS module ***and*** userspace tools.

* We can try to split the compilation to separate the ZFS module from the userspace tool, but that risks brittleness and subtle incompatibilities between a ZFS module compiled for your local kernel (which cannot be the latest and greatest kernel since the kernel does not maintain interface compatibility with kernel modules and the ZFS module is always having to catch up), and ZFS userspace tool compiled for the default latest-and-greatest kernel (even if we somehow manage to compile ***just*** the userspace tools with the latest kernel, there may be implicit assumptions about kernel constants and so on passed from the userspace tools to the kernel module, which may cause problems and birttleness).

So it's just safer to always override the `zfs` `#:linux` with the system kernel, and install it in `(packages ...)` and in `(kernel-loadable-modules ...)`.

Thanks

raid5atemyhomework

Here's another alternative:

* What the provided binaries do is:

* It looks at `/run/booted-system/kernel/manifest` and looks for a `"zfs"` package.

* It `exec`s into the corresponding file in the `/gnu/store` for that package.

This way, we can have a separate package for ZFS userspace tools, which just calls the corresponding tools in the actual ZFS that gets installed.

It's somewhat magical, though.

Thanks

raid5atemyhomework

Unfortunately, this confounds `bash-completion`. We want the bash-completion from the actual compiled ZFS, not the shim `zfs-guix-tools`. We could try digging for it too in the `kernel/manifest`, but that just adds more complexity and maintenance headaches.

It's also an extra layer of complexity and maintenance burden to create this shim wrapper that digs through your kernel manifest to look for the userland tools you need to manage ZFS in the same package that contains the actual ZFS kernel module. I think it's better to just expose a `make-zfs-package` that overrides the `#:linux` and creates a new package that you can install in `(packages ...)` **and** in `(kernel-loadable-modules ...)`.

See https://issues.guix.gnu.org/45643for my current latest patchset.

Thanks

raid5atemyhomework

Close in favor of 45692