[RFC, PATCH 0/5] fsfreeze: fix sb vs bdev freeze/thaw b0rkage

[Dave Chinner <david@fromorbit.com>]

The following series is for to address bugs in the emergency thawing code, as
well as mismatcheѕ with freezing at the block layer and the superblock that
break the freeze/thaw nesting order.

The first two patches fix the emergency thaw infinite loop reported by Jeff
Merkey and the deadlock on sb->s_umount that the infinite loop hid. These may
be stable kernel candidates.

The remainder of the patches address the bdev/sb mismatch and the fact that sb
level freezing does not nest correctly. For all the places that the bdev
interfaces are used, we need a superblock anyway so we may as well make
freeze/thaw work only at the sb level. As such, this series moves all the
nesting code to the sb from the bdev level and removes the
freeze_bdev/thaw_bdev interfaces completely. It also converts the emergency
thaw to work at the superblock level such that it will now thaw manually frozen
filesystems.

A *big* outstanding problem still exists - freezing takes an active reference
to the superblock, so unmounting an frozen filesystem has some nasty and
unexpected side effects. The existing code results in an unmountable block
device:

# mount /dev/vda /mnt/test
# xfs_freeze -f /mnt/test
# umount /mnt/test
# grep test /proc/mounts
# mkfs.xfs -f -l size=128m /dev/vda
mkfs.xfs: /dev/vda contains a mounted filesystem
Usage: mkfs.xfs
....
# mount /dev/vda /mnt/test
mount: /dev/vda already mounted or /mnt/test busy
#

At this point I can't get access to /dev/vda and needs a reboot to
get it and /mnt/test back.

This patch series results in the block device being mountable, but
remains frozen across unmount/mount:

# mount /dev/vda /mnt/test
# xfs_freeze -f /mnt/test
# umount /mnt/test
# grep test /proc/mounts
# mkfs.xfs -f -l size=128m /dev/vda
mkfs.xfs: /dev/vda contains a mounted filesystem
Usage: mkfs.xfs
....
# mount /dev/vda /mnt/test
# touch /mnt/test/foo &
[1] 2647
#
# xfs_freeze -u /mnt/test
[1]+  Done                    sudo touch /mnt/test/foo
# umount /mnt/test
# mkfs.xfs -f -l size=128m /dev/vda
meta-data=/dev/vda               isize=256    agcount=4, agsize=262144 blks
         =                       sectsz=512   attr=2
data     =                       bsize=4096   blocks=1048576, imaxpct=25
         =                       sunit=0      swidth=0 blks
naming   =version 2              bsize=4096   ascii-ci=0
log      =internal log           bsize=4096   blocks=32768, version=2
         =                       sectsz=512   sunit=0 blks, lazy-count=1
realtime =none                   extsz=4096   blocks=0, rtextents=0
#

This behaviour is only marginally better than the existing behaviour
(at least you can release the references). However, I don't really
like either option - we used to disallow umount on a frozen
filesystems to avoid this problem.

So What is really supposed to happen when we unmount a frozen
superblock? Should unmount return EBUSY?  Should it be automatically
thawed so it doesn't affect block device behaviour after unmount?
Something else?

Cheers,

Dave.