Re: [RFC PATCH] Filesystem Journal Notifications

[Jamie Lokier <jamie@shareable.org>]

Abhijit Paithankar wrote:
> > > One way to do it is to fsync every few operations. However, fsync is
> > > blocking and affects performance.
> > >
> > > The other (more efficient) way is to have the filesystem notify the
> > > application when a transaction/change is written to disk.
> > 
> > Is this more efficient than aio_fsync, and if so, why?
> > 
> > For file writes, aio_fsync seems like a cleaner interface, and if
> > that's not fast enough, it could be made faster - perhaps using code
> > form this patch.  (An aio_fsync_ranges would be even better).
> 
> aio_fsync does not provide any relationship between the metadata
> operations and journal commit. There is no mechanism to track
> operations which made it to the journal and the ones that did not.

Oh, do you mean that aio_fsync tells you all operations have reached
the journal up to the time of aio_fsync, but your notifications say
_which_ operations reach it one at a time, so you can wait on specific
ones without waiting for them all?

If the latter, I agree that what you're doing is good :-)

-- Jamie