`migratekit` convergance

[mnaser]

Hi there! :)

We've noticed that you're using the migratekit packages, are you open to the possibility of perhaps us exposing some of our APIs as public ones that you can import so that we can maintain something a little less vendor-y on your side and have benefits for the entire ecosystem?

Thanks

[stgraber]

Yeah, we'd definitely love being able to import public packages rather than having to vendor the whole thing (and then forget to update it as is common in such cases) :)

I'll let @masnax answer with more details on what pieces we've been using and whether there's any improvement we could benefit from as part of moving those pieces to being publicly exported.

[masnax]

Yeah, in general this seems like a great idea :)

That being said, we have diverged in some small ways and big ways from what is currently upstream, hopefully the public API can be sufficiently abstracted to allow for our workarounds, or eliminate the need for them.

Here's a shortlist of our changes, with some example commits, though you're likely to see more details with a proper diff:

• Removed the OpenStack bits as we are copying disks from VMware to Incus.
• Removed sirupsen/logrus dependency throughout the package.
• (https://github.com/FuturFusion/migration-manager/blob/main/internal/migratekit/nbdkit/server.go#L19-L31) Log output of nbdcopy and nbdkit commands' stdout and stderr.
• (03b3ef6) Read VMware disklib SDK from libdir=/opt/vmware-vix-disklib-distrib instead of LD_LIBRARY_PATH. I believe this was a version compatibility issue with nbdkit.
• (472de4f) Removed usage of O_EXCL during copy to target as it was causing issues on our end.
• (9b699f3) Include the tally of copied disks in the progress output, if copying more than one from source to target.
• (c354c9b) Permit full copy if unable to find any change ID
• (06821e3) Pass through progress and relevant state information via an API callback.
• (https://github.com/FuturFusion/migration-manager/blob/main/internal/migratekit/vmware/change_id.go#L54-L115) Added a helper to correctly determine the disk name for the various disk types supported by VMware. Some disks are also not eligible for migration due to being a shared disk or not supporting snapshots, so those are detected and skipped. The snapshot suffix is also trimmed by traversing up a disk's parents, and returning the root.

The biggest changes are in vmware_nbdkit.go where some Incus-specific coupling has been added as we tried to improve support of copying multiple disks from a source VM.

These includes usage of the IsSupportedDisk helper mentioned in the last bullet point above, and also ensuring that any one disk from the source is copied only to its equivalent disk on the target via a mapping of source disk names to target disk /dev/disk/by-id paths. We can't keep the source's disk name so instead we use a socket connection from the VM to the Incus daemon to determine which /dev/disk/by-id path corresponds to which source disk name via additional metadata stored on the Incus host. This also lets us keep track of the root disk in case it is not at index 0. (ad8de23).

That's probably too much detail, but I suppose in the abstract case this would be a pre-supplied dictionary of disk names to /dev/disk/by-id paths.

[mnaser]

@masnax I think what I can do is find sometime to pick up a package and try to integrate simpler changes and this can be a bit of a bit-by-bit thing?

For what it's worth, if you're okay with me cherry-picking your work in, and I can move it into pkg. Small things like moving to slog is something we'd want to do anyways, it just wasn't around back then :)

[masnax]

@mnaser that sounds reasonable to me. There's some fast-and-loose changes in there so feel free to clean up what needs cleaning up :)