Re: [RFC]block: add flush request at head

[Shaohua Li <shaohua.li@intel.com>]

On Mon, 2011-04-18 at 17:26 +0800, Christoph Hellwig wrote:
> On Mon, Apr 18, 2011 at 10:08:52AM +0200, Jens Axboe wrote:
> > Might be worth adding something for this special case, seems like the
> > NCQ restrictions will continue to be around forever (or a long time, at
> > least).
> 
> I heared people are working on adding a queued FLUSH to the standard,
> but it's going to take a long time for it to get into real life systems.
> 
> What would help now is allowing libata to actually use the FUA bit,
> given that every common disk and controller supports it these days.
> 
> Shaohua, does adding a
> 
> 	libata.fua = 1
> 
> to the kernel command line help your benchmark in any way?  It should
> if you flushes are mostly from journal writes, but not from fsync
> that didn't change any metadata.
This is a workload with fsync. I tested libata.fua=1, but nothing changed.

I also hacked the code when we proceed queue running flush list, also
proceed queue pending flush list, which doesn't change correctness for
sata. This improved a little, around 5%, but doesn't recover the whole
regression. so I still need add the flush request at queue head.

Thanks,
Shaohua