On Sat, Dec 14, 2013 at 08:51:17AM +1100, Dave Chinner wrote:
> Yes, but they are still atomic from a user and crash recovery
> point of view....
I can't see how we can guarantee an atomic update for them, both
in the case of an I/O error and an actual system crash.
> Well, I think it's a bit different to the directory block case - the
> directory blocks are filesystem metadata, while xattrs contain user
> data. Hence if we log user xattrs a user can consume all of the log
> bandwidth writing xattrs and degrade the metadata modification
> performance of the rest of the filesystem.
We're getting close to do that with namespace modifications with
all your scalability work :)
I think that's a point to consider, but not really black and white. It
just makes it a bit easier to consume log bandwith, and increases the
need to have some form of per-user quotas for this sort of operations.
> So, IMO, the first question we need to answer is whether the current
> behaviour is actually a problem for anyone....
I've not heard of real life problems, but an interfaces that has very
nice behavior for the common case, but a much less optimal for a corner
cases is bound to cause trouble.