Support ostree-format: oci in image.yaml

[cgwalters]

Part of implementing coreos/fedora-coreos-tracker#812

A whole lot of the story of coreos-assembler is threaded
with the tension between ostree and disk images. They
have fundamentally different tradeoffs. And now I'm trying
to add container images to the mix.

The idea of capturing an ostree repo in archive mode as a tarball
is a cosa invention. We don't actually ship anything that way.

The proposal in the above linked issue is to "productize" support
for shipping ostree-in-container, because containers are just
slightly fancy tarballs.

This patch adds support for:
echo 'ostree-format: oci' >> image.yaml
in the config git.

When enabled, the images/ostree is replaced with an oci-archive
format of an "ostree-in-container", which we might shorten to
ostcontainer or so. The code is updated to call out to
rpm-ostree's latest (really ostree-rs-ext's latest) code
to perform the export and import.

We're not making it the default yet, but I'd like to potentially
e.g. switch the FCOS next stream or so.

The next step after this lands is to add separate code in the
pipeline to push the image to a registry.
There's also a lot of deduplication/rationalization to
come later around cosa upload-oscontainer etc.

[cgwalters]

Looks like a gangplank race:

[2021-06-07T16:29:10.150Z] [logs:build/coreos-ci-coreos-assembler-9726974c-51cc-4dc6-be83-01254de07c7d-1] time="2021-06-07T16:29:07Z" level=error msg="failed to write the file" bucket=source error="Server not initialized, please try again." host=59f290ce3013 object=test
[2021-06-07T16:29:10.150Z] [logs:build/coreos-ci-coreos-assembler-9726974c-51cc-4dc6-be83-01254de07c7d-1]     remotes_test.go:53: failed to write content: Server not initialized, please try again.
[2021-06-07T16:29:10.150Z] [logs:build/coreos-ci-coreos-assembler-9726974c-51cc-4dc6-be83-01254de07c7d-1] --- FAIL: TestRemote (1.17s)

[cgwalters]

@darkmuggle when you get a sec ⬆️

[jlebon]

Some minor comments, but LGTM overall!

We're not making it the default yet, but I'd like to potentially
e.g. switch the FCOS next stream or so.

Maybe rawhide first would make more. (edit: ...sense :) )

[jlebon]

Should we chown -R the whole repo back here? (See e.g. runcompose_tree).

[cgwalters]

Urg, good catch. Actually I'd only tested this whole thing in the privileged path. I added some (as yet untested) code to handle unprivileged.

The more I think about this though, the more I feel like the whole tmp/repo thing was a mistake. Basically, if we're just going to eat the write amplification of making a big tarball (and qcow2) as a general rule, then keeping a tmp/repo is mostly pointless. All we need is the commit object to do things like rpmdb diffs.

And so e.g. we should then switch create_disk.sh to run e.g. rpm-ostree ex-container import instead. Almost everything else operating on tmp/repo just wants the commit object.

[jlebon]

Hmm, I think all the places which only need the commit object already do that optimization, no? If not, then yeah we shouldn't unpack for that. Everything else though does need the repo on disk to e.g. copy files out/inspect/build new trees and it seems like we'd be wasting an important optimization if we just nuked the repo instead of leaving it around after a build.

[jlebon]

For cosa buildfast at least, I could imagine teaching ostree commit to take e.g. --base path/to/my.ociarchive in the future.

[cgwalters]

it seems like we'd be wasting an important optimization if we just nuked the repo instead of leaving it around after a build.

The problem we hit is that for the disk image we hit issues with qemu reading xattrs over 9p which meant we can't directly read cache/repo-build. That might be fixed by virtiofs.

I'm not proposing deleting cache/repo-build but...leaving it as a cache just for rpm-ostree, not for anything else.