On 6/15/20 12:10 PM, Peter Krempa wrote:
Chaining bitmaps for checkpoints (disabling the active one and creating
a new) severely overcomplicated all operations in regards to bitmaps.
Specifically it requires us re-matching the on-disk state to the
internal metadata and in case of merging during block jobs it makes it
almost impossible to cover all corner cases.
Since the checkpoints and incremental backups were not yet enabled,
let's change the design to keep one bitmap per checkpoint. In case of
layered snapshots this will be filled in by using dirty-bitmap-populate.
Finally the main reason for this unnecessary complexity was the fear
that qemu's performance could degrade. In the end I think that
addressing the performance issue will be better done in qemu (e.g by
keeping an internal bitmap updated with changes and merging it
periodically back to the real bitmaps. QEMU writes out changes to disk
at shutdown so consistency is not a problem).
Removing the relationships between bitmaps frees us from complex
handling and also makes all the surrounding code more robust as one
broken bitmap doesn't necessarily invalidate whole chains of backups.
Signed-off-by: Peter Krempa <pkrempa@xxxxxxxxxx>
---
src/qemu/qemu_checkpoint.c | 21 +--------------------
1 file changed, 1 insertion(+), 20 deletions(-)
Definitely looks simpler. In graphics, pre-patch we had:
time --C1-----C2-----C3-----now
:b1....:
:b2....:
:b3....
where the differential from C1 is b1+b2+b3, the differential from C2 is
b2+b3, the incremental from C3 is b3.
post-patch, we have:
time --C1-----C2-----C3-----now
:b1...................
:b2............
:b3....
where the differential from C1 is b1, the differential from C2 is b2,
the incremental from C3 is b3.
Reviewed-by: Eric Blake <eblake@xxxxxxxxxx>
--
Eric Blake, Principal Software Engineer
Red Hat, Inc. +1-919-301-3226
Virtualization: qemu.org | libvirt.org
