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

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

On Sat, 2011-04-23 at 06:57 +0800, Tejun Heo wrote:
> Hello,
> 
> On Mon, Apr 18, 2011 at 04:25:57PM +0800, Shaohua Li wrote:
> > then why requeue adds request at head? we could have the similar issue.
> 
> SCSI doesn't seem to do it anymore but it used to cache scmd at
> rq->special over requeues so that it doesn't have to re-initialize
> requests across requeues, which means that unprepped request getting
> ahead of requeued ones may lead to deadlock due to resource
> starvation, so that's why requeue uses front queueing.
> 
> The code changed over time and the above requirement might not be
> necessary at this point.  I don't know.  However, block layer doesn't
> have any method to enforce that requests can't hold any extra resource
> on requeue and having such difficult to trigger deadlock condition
> dormant is scary.
> 
> What kind of benchmarking are we talking about on which kernel?
> blk-flush support has been revamped twice recently.  2.6.38 stripped
> out the block layer barrier thing and then it got re-reimplemented for
> 2.6.39 to support advanced flush merging.  If the regression (for
> which benchmark btw?) was visible on the older reimplementation, I'd
> really like to know how it behaves on 2.6.39-rcX.
> 
> If the problem is localized to 2.6.38, oh well, too bad, but I don't
> think we care too much.  If some distro is basing their kernel on
> 2.6.38 and the flush regression is hurting them, backporting the new
> implementation from 2.6.39 shouldn't be too difficult after all.  The
> reimplementation was almost self-contained.
> 
> If the regression affects 2.6.39 implementation too, eh well, we need
> to think of something, but I'd really like to know what kind of
> workload we're talking about.
> 
> > I'll look at this. Optimizing this one should fix the regression too. On
> > the other hand, adding flush request at head if it just follows a flush
> > still has its advantage, because drive cache is already flushed out.
> 
> New implementation wouldn't issue two flushes back to back like that,
> it doesn't make any sense to begin with.  Again, what have you been
> testing and how?
Hi Tejun,
this is a regression from 2.6.39-rc2 compared to 2.6.39-rc1, so this
isn't related to the flush rewritten. Workload is sysbench fileio,
please see the first mail at the thread for detail.

Thanks,
Shaohua