Chris Webb wrote:
Hi. There's a connected discussion on the sheepdog list about locking, and I
have a patch there which could complement this one quite well.
Sheepdog is a distributed, replicated block store being developed
(primarily) for Qemu. Images have a mandatory exclusive locking requirement,
enforced by the cluster manager. Without this, the replication scheme
breaks down and you can end up with inconsistent copies of the block
image.
The initial release of sheepdog took these locks in the block driver
bdrv_open() and bdrv_close() hooks. They also added a bdrv_closeall() and
ensured it was called in all the usual qemu exit paths to avoid stray locks.
(The rarer case of crashing hosts or crashing qemus will have to be handled
externally, and is 'to do'.)
The problem was that this prevented live migration, because both ends wanted
to open the image at once, even though only one would be using it at a time.
Yeah, this is a bigger problem I think. Technically speaking, when
using NFS as the backing filesystem, we really should not open the
destination end before we close the source end to keep the caches fully
coherent.
I've resisted this because I'm concerned that if we delay the opening of
the file on the destination, it could fail. That's a very late failure
and that makes me uncomfortable as just a work around for NFS.
But considering this locking situation, I think it is not a bad idea now.
Regards,
Anthony Liguori
