fix(backup): mount the zfs snapshot itself so a container can read it

The zfs snapshot was read through <subject>/.zfs/snapshot/<name>. The
kernel automounts that directory through a usermode helper that runs in
the host's mount namespace, so a containerised baudolo never sees the
mount and finds no snapshot. The snapshot is now mounted explicitly at
<subject>/.baudolo-<tag>, beside where btrfs carves its own, and removed
in order: umount, rmdir, zfs destroy. A failed mount destroys the
snapshot at once instead of leaking one per run.

The e2e suite could not reach this code at all: zpool create on the /img
file failed with "no such pool or dataset", because the kernel opens
vdevs from a taskq thread that resolves paths in the host's root, where
the container's /img does not exist. The pool now sits on a loop device,
whose /dev/loopN node the kernel creates on the host as well, and it is
destroyed on exit once this script created it, since a pool is
kernel-wide.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-10-18 00:22:42 +02:00
parent 389b418f42
commit 0735e5fa7b
3 changed files with 90 additions and 15 deletions

View File

@@ -36,9 +36,14 @@ ZFS_READY = "{ [ -c /dev/zfs ] || modprobe zfs 2>/dev/null; }; [ -c /dev/zfs ]"
def mount_script(fstype: str) -> str:
"""Build a filesystem on a loop device and carve out the snapshot subject."""
if fstype == "zfs":
# Exception: a loop device, not the /img file, because the kernel opens
# vdevs from a taskq thread that resolves paths in the host's root.
# The pool is kernel-wide, so it is destroyed on exit, but only once
# this script created it.
return (
f"{PACKAGES} && {ZFS_READY} && truncate -s 400M /img "
"&& zpool create -m none baudolo /img "
f"{PACKAGES} && {ZFS_READY} && truncate -s 400M /img && {ATTACH} "
'&& zpool create -m none baudolo "$LOOP" '
"&& trap 'zpool destroy -f baudolo; losetup -d \"$LOOP\"' EXIT "
"&& zfs create -o mountpoint=/subject/docker baudolo/docker"
)
return (