Re: x86, mce, Use user return notifier in mce

[Huang Ying <ying.huang@intel.com>]

On Fri, 2012-01-20 at 21:56 -0800, Tejun Heo wrote:
> Hello,
> 
> On Fri, Jan 20, 2012 at 6:44 PM, Huang Ying <ying.huang@intel.com> wrote:
> > It appears that WQ_HIGHPRI only provides priority between work queue,
> > not between the work queue backing kthread and other tasks.  Is there
> > any mechanism for that?
> 
> No, it doesn't.
> 
> > If my understanding was correct, WQ_MEM_RECLAIM has some side-effect for
> > that.  Because hardware errors occurs seldom, the reserved kthread for
> > WQ_MEME_RECLAIM just sleeps most of the time.  When first hardware error
> > occurs and the work item is queued, the reserved kthread is waked up.
> > Because the reserved kthread sleeps for long time, it is highly possible
> > for it to be scheduled at the next schedule point.
> 
> But rescuer is used only under memory pressure. It doesn't help latency at all.
> 
> > Because hardware error usually has no locality, WQ_UNBOUND can be used
> > for it so that the work item can be put on relative low-load CPU.  From
> > the document, it is said WQ_UNBOUND work items will be executed ASAP
> > too.  Compared with WQ_HIGHPRI, how about the priority of WQ_UNBOUND?
> 
> Maybe, maybe not. I suggest just using WQ_HIGHPRI for now and worrying
> about it later if the scheduling latency actually turns out to matter.

This is a performance issue.  So maybe we need to measure the actual
latency firstly.  The first step can be using WQ_HIGHPRI as you
suggested.

Thanks for your information!

Best Regards,
Huang Ying