Re: [PATCH 1/2]block cfq: make queue preempt work for queues from different workload

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

On Mon, 2011-01-17 at 22:26 +0800, Vivek Goyal wrote:
> On Tue, Jan 11, 2011 at 04:51:56PM +0800, Shaohua Li wrote:
> > I got this:
> >              fio-874   [007]  2157.724514:   8,32   m   N cfq874 preempt
> >              fio-874   [007]  2157.724519:   8,32   m   N cfq830 slice expired t=1
> >              fio-874   [007]  2157.724520:   8,32   m   N cfq830 sl_used=1 disp=0 charge=1 iops=0 sect=0
> >              fio-874   [007]  2157.724521:   8,32   m   N cfq830 set_active wl_prio:0 wl_type:0
> >              fio-874   [007]  2157.724522:   8,32   m   N cfq830 Not idling. st->count:1
> > cfq830 is an async queue, and preempted by a sync queue cfq874. But since we
> > have cfqg->saved_workload_slice mechanism, the preempt is a nop.
> > Looks currently our preempt is totally broken if the two queues are not from
> > the same workload type.
> > Below patch fixes it. This will might make async queue starvation, but it's
> > what our old code does before cgroup is added.
> 
> I am not sure how good a idea this is. So effectively now if there is a
> preemtion across workload, the workload being preempted can lose its
> share completely and be starved. So it is not just about async WRITES.
> sync-noidle can starve sync-idle workload too as meta data request will
> preempt any regular sync-idle queue and then sync-idle queue workload 
> can be starved. So this is not exactly going back to old CFQ behavior
> but something more than that too.
> 
> On one hand you are going to great lengths to fix the issues where if a
> cfqq is preempted, it does not lose its share (patch 2 in the series) and
> on the other hand you don't mind all the queues in a workload losing their
> share due to preemption.
> 
> I think atleast we should limit this to async workload. Or solve the
> problem by giving even smaller slice length to async workload.
I haven't too much idea about this, but disabling preempt might harm
performance even for sync too, for example, the cfq_rq_close() case.
can we do something like this:
sync preempts async: expire the whole workload slice
sync queue A preempt sync queue B: the workload of queue A will
immediately expire after queue A is expired.

Thanks,
Shaohua