[Devel] Re: seems to be a flaw in cfq

[Jens Axboe]

On Wed, Dec 20 2006, Jens Axboe wrote:
> On Wed, Dec 20 2006, Vasily Tarasov wrote:
> > Hello, Jens.
> > 
> > Jens Axboe wrote:
> > >
> > > Back after dinner, the fresh energy served it's purpose - I think I know
> > > what the issue is. We allow merging across process queues, which will
> > > effectively serialize some io if they are sync (like this case). I'll
> > > hack up a fix for current git and give it a test spin, to verify that
> > > this is the problem here.
> > >   
> > As far as I understand merging is an often event in this situation only
> > because of sequential
> > reading, Below is a job-file with rw=randread and direct=false. However
> > we got the same bizarre results.
> 
> Can you repeat with the below patch applied? It's against 2.6.20-rc1
> (ish), so you'll need to update to that kernel first. Best would be to
> upgrade to current git for testing purposes. Then verify that you are
> still seeing the unfair results, apply the patch, then try and
> reproduce.
> 
> The patch fixed the scenario for me. I'll retry with the random buffered
> read load you describe here.

I'm still seeing odd results with buffered io, while the direct case is
definitely fixed. So we are either looking at two bugs, or one bug that
the merge fixup hides.

-- 
Jens Axboe