Re: [PATCH] direct I/O fallback sync simplification

[Jamie Lokier <jamie@shareable.org>]

Christoph Hellwig wrote:
> On Tue, Sep 29, 2009 at 10:30:02PM +0100, Jamie Lokier wrote:
> > What I'm suggesting is that there is no need to commit the data to the
> > disk, and sometimes it's an unwanted pessimisation.  So those calls
> > may be removed entirely.
> 
> I'm not going to losen up these semantics.  They might be utterly wrong,
> but it's what we have given to users for a long time.  If you want to
> losen it up send a patch with a good rational for use case that it
> really matters, and extended version of the proof below and argue for
> it to get included.  And take all the blamer later in case something
> breaks anyway.

Tbh, I'm not sure what the semantics _are_.  If I understand right,
the data is written (without a barrier), but the metadata needed to
reach the data is not written at all by that point, and (for the
filesystems we care about) this branch is used only for filling holes
and extending files - precisely the cases where lack of metadata
is most likely to occur.

So the written data is simply unreachable if a crash occurs shortly
after a write.  Later it's fine, but normal writeback would take care
of that anyway.

Do I misunderstand that?  If I got that right, I'll see about a test
to find out of the data is really inaccessible when a crash occurs
after writing.

-- Jamie