Re: [PATCH 0/2] cfq-iosched: fixing RQ_NOIDLE handling.

[Vivek Goyal <vgoyal@redhat.com>]

On Tue, Jul 20, 2010 at 09:10:56PM +0200, Corrado Zoccolo wrote:
> On Tue, Jul 20, 2010 at 4:26 PM, Vivek Goyal <vgoyal@redhat.com> wrote:
> > On Tue, Jul 20, 2010 at 10:11:03AM -0400, Christoph Hellwig wrote:
> >> Didn't you guys have a previous iteration of the fixes that gets
> >> rid of REQ_NODILE by improving the heuristics inside cfq?  That
> >> would be much, much preffered from the filesystem point of view.
> I think the previous iteration required more complex heuristics, while
> this one uses existing ones to handle one more class of problems.
> I understand that you still see the complexity from the fs side, but
> Vivek's proposal may help also there. It only needs to be tested thoroughly.
> 
> >
> > Actually in this patch, I was thinking we can probably get rid of
> > RQ_NOIDLE flag and just check for WRITE_SYNC. Any WRITE_SYNC queue
> > gets served on sync-noidle tree. I am wondering will we not face jbd
> > thread issues with direct writes also? If yes, then not special casing
> > direct IO writes and treat them same as O_SYNC writes will make sense.
> 
> Probably it is better to submit this first, since it is already
> tested, and then have a different patch that can finish the work
> This will help when bisecting for possible regressions, since I'm not
> sure why the other writes are not already marked with RQ_NOIDLE (maybe
> it was introduced for some good reason to distinguish the two sets,
> and we won't know unless we find the workload where it helped).
> I'll resend the current patch with Jeff's reviewed and tested tags.
> 

I am fine with pushing this patch as it is first and then once we have
an answer to question whether direct IO path and O_SYNC/fsync path need
same treatment or different treatment in IO scheduler, we can fix RQ_NOIDLE
flag issue also.

Vivek