zfs: load keys for encrypted datasets during pool import

If a user has set up their own zpools and given them to us to manage,
it's possible they've configured ZFS-native encryption. For the most
part, this works completely transparently to us. However, because we
manually do zpool-import and zpool-export during startup and shutdown of
Incus, ZFS datasets with keys will have their keys unloaded during
shutdown and then the keys are not automatically loaded on startup. This
results in containers being unable to start on startup because all IOs
are blocked indefinitely until the dataset keys are loaded manually by
the admin -- even if the admin has configured automatic key loading on
their system!

The simplest solution would be to pass -l to zfs-import (which causes
ZFS to auto-import all keys for all datasets in the pool). However, it
is slightly nicer to do a separate zfs-load-key so that we can unmount
the pool if the key import fails (zfs-import will leave the pool
imported but without keys loaded).

If the user has configured keylocation=prompt (or otherwise
misconfigured the encryption settings for their pool), the command will
fail and the pool import will fail loudly (rather than silently failing
in the form of an imported pool that is not usable as a filesystem).

Signed-off-by: Aleksa Sarai <[email protected]>

internal/server/storage/drivers/driver_zfs.go

@@ -567,6 +567,21 @@ func (d *zfs) importPool() (bool, error) {
 		return false, err
 	}
 
+	// We need to explicitly import the keys here so containers can start. This
+	// is always needed because even if the admin has set up auto-import of
+	// keys on the system, because incus manually imports and exports the pools
+	// the keys can get unloaded.
+	//
+	// We could do "zpool import -l" to request the keys during import, but by
+	// doing it separately we know that the key loading specifically failed and
+	// not some other operation. If a user has keylocation=prompt configured,
+	// this command will fail and the pool will fail to load.
+	_, err = subprocess.RunCommand("zfs", "load-key", "-r", d.config["zfs.pool_name"])
+	if err != nil {
+		_, _ = d.Unmount()
+		return false, fmt.Errorf("Failed to load keys for ZFS dataset %q: %w", d.config["zfs.pool_name"], err)
+	}
+
 	if exists {
 		return true, nil
 	}