Instead of using a string to describe kernel config, use a nix
attribute set, then converted to a string.
- allows to override the config, aka convert 'yes' into 'modules' or
vice-versa
- while for now merging different configs is still crude (last spec wins),
at least there should be only one CONFIG_XYZ value compared to the current string
config where the first defined would be used and others ignored.
[initial idea by copumpkin in 2016, a major rebase to 2018 by teto]
common-config.nix has:
${kernelPlatform.kernelExtraConfig or ""}
and indeed kernelExtraConfig is in hostPlatform.platform, and not in
hostPlatform. (Ugh.).
- defined buildLinux as generic.nix instead of manual-config.nix. This
makes kernel derivations a tad more similar to your typical derivations.
- moved $buildRoot to within the source folder, this way it doesn't have to be created before the unpackPhase
and make it easier to work on kernel source without running the unpackPhase
For a while now, the only thing the 'uboot' attribute does is to tell
whether to add ubootTools to kernel/initrd builds. That can be
determined with platform.kernelTarget == "uImage" just as well.
This allows to use kernelAutoModules but still compile in any options that are set so in template config.
It's helpful for ARM and maybe other platforms where defaul configurations are useful because they compile in
modules that we and udev cannot autodetect now.
Testing showed the linux build is sensitive to /usr/include/ncursesw
unless chrooted (on non-nixos).
On a single chrooted nixos machine, -A linux is binary reproducible.
CC #2281 & @alexanderkjeldaas.
Based on access analysis with strace, I determined an essentially
minimal required set of files from the kernel source that was needed to
build all current kernel packages on 3.10, which ultimately resulted in
keeping 30M of source. Generalizing from that minimal set, which
required ad-hoc specifications of which headers outside of include/ and
arch/*/include and which files in the scripts/ directory should be kept,
to a policy of keeping all non-arch-specific headers that aren't part of
the drivers/ directory and the entire scripts/ directory added an
additional 17M, but there was nothing in the analysis that indicated
that that ad-hoc specification was at all complete so I think the extra
hit is worth the likely greater compatibility.
For reference, we now keep:
* All headers that are NOT in arch/${notTargetArch}/include or drivers/
* The scripts/ directory
* Makefile
* arch/${targetArch}/Makefile
IMO the most likely cause of future problems are the headers in
drivers/, but hopefully they won't actually be needed as they add 50M
Ideally kernel packages would only use include and
arch/${targetArch}/include, but alas this is observably not the case.
master:
* $out
* size: 234M
* references-closure: linux-headers, glibc, attr, acl, zlib, gcc,
coreutils, perl, bash
merge-kernel-builds:
* $out
* size: 152M
* references-closure: none
* $dev
* size: 57M
* references-closure: linux-headers, glibc, zlib, gcc
So even with the non-minimal set we still beat out master. Keeping the
drivers headers would make us only slightly bigger.
Signed-off-by: Shea Levy <shea@shealevy.com>
This makes the disk usage footprint of building the kernel smaller in 3
ways:
1) There is no separate kernel source derivation
2) Rather than using the entire build tree, only the output of make
modules_prepare is kept in the $dev output (plus the module symbol
versioning file generated during the build)
3) Only the subset of the source tree known to be needed for external
builds is kept in $dev
Note that while 2) is supported by official kernel documentation, I
couldn't find any source describing what we need to keep for 3). I've
started with the bare minimum (the main Makefile is called by the
Makefile generated by make modules_prepare) and we can/should add more
as needed for kernelPackages.
Signed-off-by: Shea Levy <shea@shealevy.com>
This has three major benefits:
1. We no longer have two kernel build processes to maintain
2. The build process is (IMO) cleaner and cleaves more closely to
upstream. In partuclar, we use make install to install the kernel and
development source/build trees, eliminating the guesswork about which
files to copy.
3. The derivation has multiple outputs: the kernel and modules are in
the default `out' output, while the build and source trees are in a
`dev' output. This makes it possible for the full source and build tree
to be kept (which is expected by out-of-tree modules) without bloating
the closure of the system derivation.
In addition, if a solution for how to handle queries in the presence of
imports from derivations ever makes it into nix, a framework for
querying the full configuration of the kernel in nix expressions is
already in place.
Signed-off-by: Shea Levy <shea@shealevy.com>
The function ‘mkDerivation’ now checks whether the current platform
type is included in a package's meta.platform field. If not, it
throws an exception:
$ nix-build -A linux --argstr system x86_64-darwin
error: user-thrown exception: the package ‘linux-3.10.15’ is not supported on ‘x86_64-darwin’
These packages also no longer show up in ‘nix-env -qa’ output. This
means, for instance, that the number of packages shown on
x86_64-freebsd has dropped from 9268 to 4764.
Since meta.platforms was also used to prevent Hydra from building some
packages, there now is a new attribute meta.hydraPlatforms listing the
platforms on which Hydra should build the package (which defaults to
meta.platforms).
Having N different copies of the NixOS kernel configuration is bad
because these copies tend to diverge. For instance, our 3.10 config
lacked some modules that were enabled in older configs, probably
because the 3.10 config had been copied off an earlier version of some
older kernel config.
So now there is a single kernel config in common-config.nix. It has a
few conditionals to deal with new/removed kernel options, but
otherwise it's pretty straightforward.
Also, a lot of cut&paste boilerplate between the kernel Nix
expressions is gone (such as preConfigure).
This is for consistency with terminology in stdenv (and the terms
"hostDrv" and "buildDrv" are not very intuitive, even if they're
consistent with GNU terminology).
