Re: [Qemu-devel] [PATCH] raw-posix: Linearize direct I/O on Linux NFS

[Badari Pulavarty]

On Fri, 2011-04-15 at 13:09 -0500, Anthony Liguori wrote:
> On 04/15/2011 11:23 AM, Badari Pulavarty wrote:
> > On Fri, 2011-04-15 at 17:34 +0200, Christoph Hellwig wrote:
> >> On Fri, Apr 15, 2011 at 04:26:41PM +0100, Stefan Hajnoczi wrote:
> >>> On Fri, Apr 15, 2011 at 4:05 PM, Christoph Hellwig<address@hidden>  wrote:
> >>>> NAK. ?Just wait for the bloody NFS client fix to get in instead of
> >>>> adding crap like that.
> >>> That's totally fine if NFS client will be fixed in the near future but
> >>> this doesn't seem likely:
> >>>
> >>> http://www.spinics.net/lists/linux-nfs/msg20462.html
> >> The code to use preadv/pwritev has been in qemu for over 2 years,
> >> and it took people to notice the NFS slowdown until now, so don't
> >> expect it to be fixed three days layer.
> > True. That brings up a different question - whether we are doing
> > enough testing on mainline QEMU :(
> 
> The issue here is NFS, not QEMU.  

Sure. But we should have caught the regression on NFS when
preadv/pwritev change went into QEMU or before going in -- Isn't it ?
Since it was 2 years ago (like hch indicated) - we could have
fixed NFS long ago :) 

> Moreover, the real problem is that 
> we're using O_DIRECT.  O_DIRECT seems to result in nothing but problems 
> and it never seems to be tested well on any file system.

O_DIRECT was added for a specific use-case in mind - and its supposed
to handle only that case well (pre-allocated files with database usage -
where db has their own caching layer). That case is well tested 
by various DB vendors on most (important) local filesystems. 

You know very well why we are on O_DIRECT path :)

> 
> I think the fundamental problem we keep running into really boils down 
> to O_DIRECT being a second class interface within Linux.

Its by design. Its a special case for specific use.

Thanks,
Badari