Re: Drivers: scsi: FLUSH timeout

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Fri, 2013-10-04 at 15:02 +0000, KY Srinivasan wrote:
> 
> > -----Original Message-----
> > From: Eric Seppanen [mailto:eric@purestorage.com]
> > Sent: Thursday, October 03, 2013 1:49 PM
> > To: Nicholas A. Bellinger
> > Cc: KY Srinivasan; linux-kernel@vger.kernel.org; devel@linuxdriverproject.org;
> > linux-scsi@vger.kernel.org
> > Subject: Re: Drivers: scsi: FLUSH timeout
> > 
> > On Thu, Oct 3, 2013 at 5:09 AM, Nicholas A. Bellinger
> > <nab@linux-iscsi.org> wrote:
> > >
> > > On Wed, 2013-10-02 at 18:29 +0000, KY Srinivasan wrote:
> > > > Ideally, I want this to be adjustable like the way we can change the I/O
> > timeout.
> > > > Since that has been attempted earlier and rejected (not clear what the
> > reasons were),
> > > > I was suggesting that we pick a larger number. James, let me know how I
> > should proceed here.
> > > >
> > >
> > > I think the objection was to making a module parameter for doing this
> > > globally for all struct scsi_disk, and not the idea of making it
> > > adjustable on an individual basis per-say..
> > >
> > > What about adding a /sys/class/scsi_disk/$HCTL/flush_timeout..?
> > 
> > Do I/O timeouts and flush timeouts need to be independently adjusted?
> > If you're having trouble with slow operations, it seems likely to be
> > across the board.
> > 
> > Flush timeout could be defined as 2x the read/write timeout.  Any
> > other command-specific timeouts could be scaled the same way.
> 
> I like this idea and would result in minimal changes. James, if it ok with you,
> I could send you the patch.

Depends: I still prefer the per-target override, but if the proposal is
to take the existing variable timeout for the queue and have 2x that for
the flush, so you plan to increase the per-device timeout with hyper-v
to 90s via sysfs, then I'm OK with it.

James