prepare-root: Move sysroot.tmp creation earlier

[cgwalters]

Main motivation is prep for composefs in
#2640
In the interest of that, we add a bool using_composefs but it's currently always false.

[alexlarsson]

Shouldn't this be using root_mountpoint, not "/sysroot". They will mostly be the same, but all the other places use root_mountpoint.

[cgwalters]

Yes. That said, I had to re-read and re-re-read this code and think through things because I realized this change overall was sufficiently large that it had removed my confidence in knowing how things work.

I added some more comments, and updated.

[alexlarsson]

Minor comment, but otherwise LGTM

[cgwalters]

I find myself wondering though if we shouldn't instead create ostree-prepare-composefs.c that is dedicated to just that. Among other things, there we can not care about the "no-initramfs" path, we can hardcode the sysroot_readonly flag, etc.

But bigger picture also, I think with composefs we should stop doing the hardlinked checkouts at all. And that also leads into the discussions around having ostree=1|2 and not checksums, etc., which also drops out the need to call this resolve_deploy_path() etc.

Also related to this, I'm thinking that instead of parsing options from the target repo, we should inject a config file into the initramfs (e.g. /usr/lib/ostree/initramfs/config). In fact I'm pretty sure that's what we must do because it chains into the SB story. And if we do that it's again different from what ostree-prepare-root.c is doing today and an argument to make a new file.

[alexlarsson]

For the config, the current PR uses a ot-composefs kernel cmdline options for similar reasons, i.e. in a secureboot system it will be part of a trusted set of info. But yes, the initrd is also a trusted place for such info.

As for ostree-prepare-composefs. What would decide to run this instead of ostree-prepare-root?

[alexlarsson]

The current PR setup is kinda incrementally adding composefs. It deploys both styles, and the code picks up whichever works. I agree that a more opinionated version may be better where we just do the composefs during the deploy if that is configured.

Don't we still need a partial checkout for the etc merging to work though?

[alexlarsson]

Hmm, also it feels like we shouldn't embedd details about how the image is deployed in the initrd, because you may want to have a commit that works in both kinds of deployment.

[cgwalters]

As for ostree-prepare-composefs. What would decide to run this instead of ostree-prepare-root?

With systemd in the initramfs, there's an immense amount of flexibilty. In the general case we can install a generator which parses an injected config file and chooses whether to activate ostree-prepare-root.service or ostree-prepare-composefs.service. Or we can just have a single binary that parses the config itself and dispatches.

I think what seems simplest though is having one binary with two source files, and very early on in ostree-prepare-root.c we detect the composefs case and just dispatch to that - explicitly erroring out if we're pid 1.

[cgwalters]

@alexlarsson mind adding an official "approve" on this PR? Branch protection here requires it.

[alexlarsson]

ack