The ollama module in its default configuration relies on systemd's
`DynamicUser=` feature for user allocation. In #305076 that allocation
was made conditional and tied to the `sandboxing` option, that was
intended to fix access to model directories outside the allocated state
directory.
However, by disabling sandboxing ollama would inadvertently run as root,
given that `User=` and `Group=` are not required to be set.
The correct way to grant access to other paths is to allocate static
user and group, and grant permissions to the destination path to that
allocation.
We therefore replace the sandboxing option user and group options, that
default to `null`, which means they default to `DynamicUser=`, but can
be replaced with a statically allocated user/group, and thereby a stable
uid/gid.
Fixes: 552eb759 ("nixos/ollama: add options to bypass sandboxing")
We want to get rid of specialFileSystems / earlyMountScript eventually and
there is no need to run this before systemd anymore now that
the wrappers themselves are set up in a systemd unit since https://github.com/NixOS/nixpkgs/pull/263203
Also this is needed to make soft-reboot work. We want to make sure
that we remount /run/wrappers with the nosuid bit removed on soft-reboot
but because @earlyMountScript@ happens in initrd, this wouldn't happen
The activation script that remounts the /etc overlay now handles other
mount points on top of /etc by bind mounting them to the new temporary
/etc overlay and then atomically revealing it.
Summary of this change:
- Simplify code.
- Stop a disk image from being cached in the binary cache.
- Make erofs Nix Store image build in an acceptable time outside of
testing environments (like `darwin.builder`).
- Do not regress on performance for tests that use many store paths in
their Nix store image.
- Slightly longer startup time for tests where not many store paths are
included in the image (these probably shouldn't use `useNixStoreImage`
anyways).
- Slightly longer startup time when inputs of VM do not change because
the Nix store image is not cached anymore.
Remove the `storeImage` built with make-disk-image.nix. This produced a
separate derivation which is then cached in the binary cache. These
types of images should be avoided because they gunk up the cache as they
change frequently. Now all Nix store images, whether read-only or
writable are based on the erofs image previously only used for read-only
images.
Additionally, simplify the way the erofs image is built by copying the
paths to include to a separate directory and build the erofs image from
there.
Before this change, the list of Nix store paths to include in the Nix
store image was converted to a complex regex that *excludes* all other
paths from a potentially large Nix store.
This previous approach suffers from two issues:
1. The regex is complex and, as admitted in the source code of the
includes-to-excludes.py script, most likely contains at least one
error. This means that it's unlikely that anyone will touch this
piece of software again.
2. When the Nix store image is built from a large Nix store (like when
you build the VM script to run outside of any testing context) this
regex becomes painfully slow. There is at least one prominent
use-case where this matters: `darwin.builder`.
Benchmarking impressions:
- Building Nix store via make-disk-image.nix takes ~25s
- Building Nix store as an erofs image takes ~4s
- Running nixosTests.qemu-vm-writable-store-image takes ~10s when
building the erofs image with the regex vs ~14s when building by
copying to a temporary directory.
- nixosTests.gitlab which had the biggest gains from the initial erofs
change takes the same time as before.
- On a host with ~140k paths in /nix/store, building the erofs image
with the regex takes 410s as opposed to 6s when copying to a temporary
directory.
