[dm-devel] [PATCH 0/2] dm: Add no_abort_q feature flag to dm-mpath

[Moger, Babu]

> -----Original Message-----
> From: Mike Anderson [mailto:andmike at linux.vnet.ibm.com]
> Sent: Wednesday, May 05, 2010 8:39 PM
> To: Moger, Babu
> Cc: dm-devel at redhat.com; linux-scsi at vger.kernel.org
> Subject: Re: [PATCH 0/2] dm: Add no_abort_q feature flag to dm-mpath
> 
> Moger, Babu <Babu.Moger at lsi.com> wrote:
> > > -----Original Message-----
> > > From: linux-scsi-owner at vger.kernel.org [mailto:linux-scsi-
> > > owner at vger.kernel.org] On Behalf Of Mike Anderson
> > > Sent: Monday, May 03, 2010 11:01 PM
> > > To: dm-devel at redhat.com
> > > Cc: linux-scsi at vger.kernel.org
> > > Subject: [PATCH 0/2] dm: Add no_abort_q feature flag to dm-mpath
> > >
> > > This patch series contains two patches.
> > >
> > > 1.) The first patch adds a feature flags bit field to the multipath
> > > structure for selected features.
> > >
> > > 2.) The second patch allows a user to set a feature flag for a dm-
> mpath
> > > device of "no_abort_q" to indicate that deactivate_path should not
> call
> > > blk_abort_queue. I tried to select a short feature name to feature
> > > output
> > > listed with multipath -l would not be too long "features='2
> > > queue_if_no_path
> > > no_abort_q'
> >
> > Mike,
> >  In what special circumstances you recommend to use this feature? It
> would be great if you could add that explanation in your patch
> descriptions.
> 
> Yes I will update the descriptions.
> 
> To answer your question in general it would be circumstance where the
> block_abort_queue during failover is leading to longer recovery instead
> of
> shorter. This could be because of aborting / transport / device
> interactions (the work by others in the iSCSI and FC transports have
> made
> things better here).
> 
> While there are case where abort makes a big difference (being able to
> failover in less than max_retries * IO timeout value). The latency
> numbers
> for simple RDAC initiator failover show only small improvements on my
> configurations (others may have better data). A assumption would be
> that
> there could be combinations of transports and devices out there where
> this
> might not give the fastest failover so the user may want control.
> 
> On a historically note the call to block_abort_queue came in somewhat
> sideways by me linked to the timeout movement from SCSI to blk so it
> could
> have integrated better. I should have had a way to disable / enable it
> then and I am trying to provide that now so that the user has some
> control.

Thanks for the details.