Re: spurious completions during NCQ on rt kernel

[Tejun Heo <htejun@gmail.com>]

Hello,

Mark Lord wrote:
> Mmmm.. that's interesting.  I suppose a big difference between -rt and
> vanilla
> is in latency between interrupt signal and the interrupt handler
> actually running.
> 
> I think in -rt it is slightly *longer*, since the handler is run by a
> scheduled
> thread instead of directly.
> 
> Anyway, for me, the intriguing point of data here, is that there could
> be some
> tight soft timing issue that sometimes can make NCQ fail.  And our
> response to date
> with this is to blacklist the drives.. which may or may not be correct now.

I think it's interrupt timing problem too.  I just received my order of
problematic drives.  I'll try to reproduce it and find out.  At any
rate, I think we'll probably have to drop spurious NCQ completion check.

Thanks.

-- 
tejun