Bug#848945: #848945 Empty rpool leaves ZFS systems unbootable

[Etienne Dechamps]

Control: found -1 2.04-18

I just hit this bug. It rendered my system unbootable because my root
ZFS pool has the new zstd compression feature enabled, and it looks
like grub-probe choked on that.

There are two things that I found highly frustrating here:

- While my root (/) is on ZFS, my /boot is not, precisely to try to
avoid issues like these. In this setup GRUB doesn't need to be able to
read the zpool - all that matters is the name of the pool, which can
be trivially determined without the need for any kind of subtle
low-level probing. It seems like as long as the
bootloader/kernel/initrd itself is not stored on ZFS, GRUB should Just
Work regardless of the low-level zpool details (such as zpool
features). That is not presently the case because /etc/grub.d/10_linux
insists on using grub-probe and grub-probe insists on using its own
custom low-level ZFS read code just to figure out the name of the
zpool. That's just... silly.

- The error handling here is appalling. In my case update-grub
(grub-mkconfig) *did not even display a single error message*. The
"unknown filesystem" error from grub-probe is being swallowed somehow,
and /etc/grub.d/10_linux doesn't check that the returned `rpool` looks
valid. update-grub will happily generate an obviously garbage,
unbootable config with a syntactically invalid root=ZFS= kernel
parameter and everything will look fine at first... until you try to
reboot! At the very least an error message should be displayed;
ideally update-grub should refuse to generate a new config altogether,
to avoid replacing a working config with a broken one.

By the way, here are steps to reproduce to make grub-probe choke on a
ZFS pool with zstd enabled:

# dd if=/dev/zero of=/tmp/pool bs=1 count=1 seek=$((128*1024*1024))
# zpool create testpool /tmp/pool
# zfs set compression=zstd testpool
# /usr/sbin/grub-probe --device /tmp/pool --target=fs_label
/usr/sbin/grub-probe: error: unknown filesystem.

If the "zfs set compression" step is omitted, grub-probe works. Tested
with zfs 2.0.3-8.