Bug 236168

RHEL 5.1 clone - some code exists (see bug 174701),
nice to have this in RHEL5 too.

+++ This bug was initially created as a clone of Bug #174701 +++

LVM2 pvcreate should make an automatic backup of the part of
the device that it overwrites, so you can undo it.

(Could add as a generic option in the low-level write function, so other
commands could use it too.)

[split off from bugzilla 168330]

...

-- Additional comment from prockai on 2007-04-11 12:43 EST --
Created an attachment (id=152288)
lvm2-pvcreate-undo-prelim

Here comes a preliminary patch implementing basic pvcreate --undo
functionality. It has a few rough edges yet, like there is no configuration
option to disable pvcreate archive creation (easy to fix) and that
offset/length pairs are written out in host byte order (needs ntohll/htonll
which seem to be missing from libc? needs investigation).

The design is simplistic, it provides new device write primitives suffixed
_backup that take an extra file descriptor argument, which they use to write
out a copy of whatever they are going to overwrite, prepended with offset and
length of the block.

The restore function takes a file that contains several such blocks and writes
them over in their original positions, in reverse order from the file. This
allows for easy backup implementation, where a backup file is created and
opened for writing, its fd is passed to every device write operation and file
is closed. Then, restore takes the opened device and a filename of backup file.


Things like further metadata about the operation are left for higher layers, in
this case pvcreate. Current implementation just uses <archive_dir>/<uuid>.pv as
backup file and --undo on a PV looks for this file which when found is used for
restore. There is room for a more elaborate operation logfile, which may
reference the (simplistic) binary backup files. Any checksumming should be
probably done on that (operation logfile) level as well.

To make it possible for format_text to use the backup-enabled device
operations, i have added device_backup_fd to toolcontext. This may need
different addressing as well, although there is not much room for maneuvering
-- probably either way, all dev_write/dev_set calls that need backup
functionality will have to be modified in a way or other (either replace them
with versions that take toolcontext parameter, or the _backup versions).

I will followup with improved patches either tonight or tomorrow.