We've added the variant to `DerivationOutput` to support them, but made
`DerivationOutput::path` partial to avoid actually implementing them.
With this chage, we can all collaborate on "just" removing
`DerivationOutput::path` calls to implement CA derivations.
This reduces memory consumption of
nix-instantiate \
-E 'with import <nixpkgs> {}; runCommand "foo" { src = ./blender; } "echo foo"' \
--option nar-buffer-size 10000
(where ./blender is a 1.1 GiB tree) from 1716 to 36 MiB, while still
ensuring that we don't do any write I/O for small source paths (up to
'nar-buffer-size' bytes). The downside is that large paths are now
always written to a temporary location in the store, even if they
produce an already valid store path. Thus, adding large paths might be
slower and run out of disk space. ¯\_(ツ)_/¯ Of course, you can always
restore the old behaviour by setting 'nar-buffer-size' to a very high
value.
"uid-range" provides 65536 UIDs to a build and runs the build as root
in its user namespace. "systemd-cgroup" allows the build to mount the
systemd cgroup controller (needed for running systemd-nspawn and NixOS
containers).
Also, add a configuration option "auto-allocate-uids" which is needed
to enable these features, and some experimental feature gates.
So to enable support for containers you need the following in
nix.conf:
experimental-features = auto-allocate-uids systemd-cgroup
auto-allocate-uids = true
system-features = uid-range systemd-cgroup
2^18 was overkill. The idea was to enable multiple containers to run
inside a build. However, those containers can use the same UID range -
we don't really care about perfect isolation between containers inside
a build.
Also, run builds in a cgroup namespace (ensuring /proc/self/cgroup
doesn't leak information about the outside world) and mount /sys. This
enables running systemd-nspawn and thus NixOS containers in a Nix
build.
Rather than rely on a nixbld group, we now allocate UIDs/GIDs
dynamically starting at a configurable ID (872415232 by default).
Also, we allocate 2^18 UIDs and GIDs per build, and run the build as
root in its UID namespace. (This should not be the default since it
breaks some builds. We probably should enable this conditional on a
requiredSystemFeature.) The goal is to be able to run (NixOS)
containers in a build. However, this will also require some cgroup
initialisation.
The 2^18 UIDs/GIDs is intended to provide enough ID space to run
multiple containers per build, e.g. for distributed NixOS tests.
