Cargo.toml: switch to a workspace

[allisonkarlitskaya]

Split into a few separate crates:

• composefs
• composefs-oci
• composefs-boot
• cfsctl
• composefs-setup-root

This is not a huge improvement in terms of compile speed, and it has some drawbacks (like 'cargo run' no longer defaulting to cfsctl) but we might want to do it anyway, for two reasons:

• I have a couple of experimental pieces of work (like a fuse filesystem implementation and an HTTP downloader) that I'd like to commit somewhere, but don't want to add to the main composefs crate, mostly for dependency reasons. Features could help there..

• I've wanted to do xtask stuff for a long time, but it requires workspaces...

This is something like the transformation required to make it work. Maybe let's start discussing it.

[allisonkarlitskaya]

I think I'm starting to sour a little on this idea already. We could be using something like Just or just straight-up make or shellscripts to get us the xtask thing and features would get us the rest...

[cgwalters]

I think splitting things up into separate crates is often a good idea anyways, there's clearly a "core composefs" and "boot stuff" at least.

I don't have a really strong opinion though.

[allisonkarlitskaya]

Seems we're both kinda stuck in the same "meh" place on this one :)

[shawn111]

@allisonkarlitskaya just add two small cargo comments

[allisonkarlitskaya]

The same error, three times, even when it looks completely unrelated to the PR, and even when it's rawhide, but when it works on main (cf #124) and I've never seen anything like it before? hmmmmm

qemu-img create -q -f qcow2 -b /home/runner/work/composefs-rs/composefs-rs/examples/bls/rawhide-bls-efi.qcow2 -F qcow2 /var/tmp/bots-run/cockpit-qczwawbd.qcow2
+ ! test -f /run/nologin && cat /proc/sys/kernel/random/boot_id
+ ! touch /a
touch: cannot touch '/a': Read-only file system
+ ls /sysroot
Traceback (most recent call last):
  File "/home/runner/work/composefs-rs/composefs-rs/examples/test/run-tests", line 62, in <module>
    main()
    ~~~~^^
  File "/home/runner/work/composefs-rs/composefs-rs/examples/test/run-tests", line 51, in main
    case(machine)
    ~~~~^^^^^^^^^
  File "/home/runner/work/composefs-rs/composefs-rs/examples/test/run-tests", line 14, in test_basic
    assert m.execute('ls /sysroot') == 'composefs\nlost+found\nstate\n'
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError
client_loop: send disconnect: Broken pipe
Error: Process completed with exit code 1.

[allisonkarlitskaya]

OK. After a lot of fighting, I got this reproduced locally. The issue is caused by the crate separation commit (works before, fails after) and the result is that /sysroot in the booted system is now a second copy of the composefs. The reason it only fails on rawhide is because that's the only case that doesn't have pre-6.15 enabled. Looking into this.

[allisonkarlitskaya]

It's getting clearer: pre-6.15 is the default flags for the composefs crate. cargo build -F has no impact on the binaries we build because they include the composefs crate with the default features. That means that rawhide is getting the pre-6.15 version of the mount code, even if it doesn't want it. That conflicts with our explicitly different handling of the mounts in composefs-setup-root, which also has the feature.

Pretty sure the fix is to declare that we take no-default-features on the composefs crate and wire it through...

[allisonkarlitskaya]

Ok, so the pre-6.15 on rawhide thing looks like a legit kernel regression. We're lucky we caught it: I added a commit here to introduce a test variant that will test it in the future (but it's currently disabled in CI).

I wrote to fsdevel: https://lore.kernel.org/linux-fsdevel/CAOYeF9WQhFDe+BGW=Dp5fK8oRy5AgZ6zokVyTj1Wp4EUiYgt4w@mail.gmail.com/

[travier]

LGTM from a quick look.