This is not an issue when initially creating the partition because
device mapper will just map on top of the filesystem, but it breaks
incremental updates because the check for `TYPE=` will fail, and the
subsequent attempt to run `mkswap` also fails:
mkswap: cannot open /dev/disk/by-partlabel/disk-nix-store-swap: Device or resource busy
Instead, just don't do anything, NixOS will take care of it.
Running `zfs set mountpoint=/mnt/my-ds tank/my-ds`, ZFS may try to
unmount the dataset even if the mountpoint didn't change.
To avoid the confusing error message, this command is now only run when
the mountpoint actually changes.
This is needed to prevent two layers of emulation when offering a build for a host running binfmt such as an x86 host trying to compile for aarch64 via binfmt
This is especially useful when trying to build images for embedded systems such as the pi, which have a vendor kernel that can't be booted as part of the disko image builder process
This allows for the disko scripts to be ran in more foreign environments
where we can't assume the basic utilities coreutils provides are
available or aren't 1-for-1 compatible (e.g. busybox utilities).
I.e. instead of effectively ignoring `discardPolicy` when combined with
`randomEncryption`, or adding a second option requiring the user to
repeat themselves (DRY violation), just enable both when requested via
`discardPolicy`.
Without this `discardPolicy` is a NOP when `randomEncryption` is
enabled.
This hopefully fixes the large memory consumption we have seen in the
past. Also presumably coreutils cp is more tuned for performance than
whatever nix does.
closure-info has one flaw, it's possible that this file contains stale
information when build are not reproducible and the local checksum
doesn't match what the remote build created.
The work-around here is that we disregard the hashes and re-compute with
what nar's we actually have locally instead.
