2 May 2023 |
Arian | though i do think it can be a matter of opinion https://github.com/systemd/systemd/blame/main/units/systemd-repart.service.in#L25 | 21:02:47 |
Arian | upstream runs after /sysusr is mounted (which will become /usr after switchroot) so it definitely expects the main software to already be mounted and only add additional partitions for storage not software | 21:03:29 |
Arian | I think for us that'd mean run before root but after nix.mount | 21:03:52 |
Arian | I think that makes sense for the nixos case if /nix is on separate partition | 21:05:10 |
Arian | but not sure.... | 21:05:19 |
@elvishjerricco:matrix.org | Arian: Yea I was looking into this last night, and there isn't really a clean answer without patching systemd or something. It really expects at least an OS to be installed on a disk, but in NixOS we go even further and don't even have an OS on a disk. The other thing is that we actually can't just use /sysroot ; we would have to wait for nixos-activation.service to run so that /etc is set up | 21:19:39 |
Arian | I think doing systemd-repart --empty=allow /dev/vda is the way to go | 21:20:17 |
Arian | for nixos | 21:20:24 |
Arian | that will re-partition the disk even if it doesnt have a GPT partition table | 21:20:42 |
Arian | and run that in initrd | 21:20:59 |
@elvishjerricco:matrix.org | Yea, I'm really starting to think that the only sane option is to put the config files in the initrd and do systemd-repart --empty=allow ${config.fileSystems."/".device} except that isn't quite right because we have to point at the root disk, not the root partition | 21:22:27 |
@elvishjerricco:matrix.org | Which makes me think we actually need a boot.initrd.systemd.repart.device option | 21:23:21 |
Arian | boot.initrd.systemd/repart.device | 21:23:25 |
Arian | yeh | 21:23:29 |
Arian | exactly | 21:23:30 |
@elvishjerricco:matrix.org | But that's super lame because we don't need to do that for boot.initrd.growPartition , which I was hoping to implement in systemd-initrd via repart | 21:23:58 |
@elvishjerricco:matrix.org | * But that's super lame because we don't need to do that for boot.growPartition , which I was hoping to implement in systemd-initrd via repart | 21:24:14 |
@elvishjerricco:matrix.org | but growPartition just assumes you've set fileSystems."/".device to something with a number at the end that can be dropped to get to the disk | 21:24:49 |
Arian | well we can still implement that as a compatibiliy hack? | 21:25:31 |
@elvishjerricco:matrix.org | Yea I guess we can move the number-dropping logic out of shell script and into nix code and set boot.initrd.systemd.repart.device with that | 21:26:05 |
Arian | exactly | 21:26:17 |
Arian | having nixos-install call systemd-repart would be nice too.
have nixos-install --flake github:arianvp#mymachine just work would be kinda magical
| 21:26:54 |
@elvishjerricco:matrix.org | oh that's a cool idea | 21:27:22 |
Arian | I now do this: https://github.com/arianvp/nixos-stuff/blob/master/configs/utm/install.sh | 21:27:45 |
@elvishjerricco:matrix.org | So one slightly crazier idea than parsing device file names with nix would be to go nuts with udev | 21:29:05 |
@elvishjerricco:matrix.org | i.e. set up a udev rule that waits for fileSystems."/".device and creates a /dev/repart-actual-root-node symlink that points to the parent-most device of that device | 21:30:00 |
@elvishjerricco:matrix.org | oh but that does expect the root partition to at least exist already | 21:30:24 |
@elvishjerricco:matrix.org | so nevermind | 21:30:27 |
@elvishjerricco:matrix.org | Arian: Anyway, I think the only other thing we could really do besides a boot.initrd.repart.device option would be to repurpose /sysusr for /nix/store (which is kind of morally correct anyway tbh), and and pull the system's toplevel store path from the init= parameter like we doin nixos-activation.servicein order to find repart definitions. The big advantage here is that the root device would be derived from the backing device of /sysusr/usr(which we would hack up to be derived from /sysusr/nix/store), so that we don't need an explicit device to be set. But that does mean your /nix/store` has to be on a partition on your root disk, and can't be something fancy like a 9p file system or something | 21:36:37 |
@elvishjerricco:matrix.org | * Arian: Anyway, I think the only other thing we could really do besides a boot.initrd.repart.device option would be to repurpose /sysusr for /nix/store (which is kind of morally correct anyway tbh), and and pull the system's toplevel store path from the init= parameter like we do innixos-activation.service in order to find repart definitions. The big advantage here is that the root device would be derived from the backing device of/sysusr/usr (which we would hack up to be derived from/sysusr/nix/store ), so that we don't need an explicit device to be set. But that does mean your /nix/store\ has to be on a partition on your root disk, and can't be something fancy like a 9p file system or something | 21:37:01 |