Re: [PATCH 0/3] scsi_dh: Make scsi device handler modules automatically inserted

[Chandra Seetharaman <sekharan@us.ibm.com>]

On Tue, 2009-07-07 at 18:14 +0000, James Bottomley wrote:
> On Tue, 2009-07-07 at 13:51 -0400, Peter Jones wrote:
> > On 07/07/2009 01:12 PM, James Bottomley wrote:
> > > On Fri, 2009-06-26 at 09:56 -0400, Peter Jones wrote:
> > >> James, I think Chandra and I have responded to most if not all of your points, 
> > >> and would appreciate your thoughts on what we've said.
> > > 
> > > Well, you didn't respond to the important one:
> > > 
> > > You're seeking to change the binding of these helpers from manual to
> > > automatic.  This would mean that the modules are loaded in the single
> > > controller case, for which they may not be wanted and also when some
> > > multi path tool other than dm-mp is managing the device, in which case
> > > they may actively interfere with operations.  Your basic contention is
> > > that you "don't see any concern here".
> > 
> > I think Chandra addressed this in his reply to your previous email: 
> 
> I don't think so:
> 
> > [From him]
> > > This is by design (of SCSI DH). We do want the device to be attached to
> > > its handler _irrespective_ of whether multipath comes along or not.
> > > 
> > > BTW, there is _no_ infrastructure in multipath for handlers. They were
> > > removed from multipath when scsi dh came along. So, no worries about
> > > proprietary multipath handlers. Also, multipath _can_ attach a device to
> > > a different (SCSI) device handler if it finds that the one that is
> > > already attached is not the right one.
> 
> I want to know what happens when some multipath system other than dm-mp
> tries to use a system with a device handler attached ... I don't see how
> that addresses the issue at all. The device handlers, when they're
> attached can alter sense and return code processing .. this has the
> potential to interfere with how the other multipath code is expecting
> things to happen

When I responded to your earlier email I thought you meant the
interaction between dm-mp and the hardware handler ( and not other
out-of-tree multipath solutions). My answer was in that context alone.

Yes, your conclusion is correct. There are few functionalities in the
hardware handler interface that might affect the above layers (block and
above). One is the prep_fn() function, second is the check_sense()
function, and third is the activate() function.

Since scsi_dh is designed for dm-mp :). 

I do not know of all the out-of-tree multipath solutions and how they
behave and at what layer they interact. In effect, I haven't tested
hardware handler with other multipath solutions.

> 
> If we have to get really concrete: what happens with something like
> powerpath and scsi_dh_emc attached?
> 

They _might_ be affected as mentioned above.

> > [From you again]
> > > When I ask what testing you've done for either of these, the studied
> > > silence eloquently illustrates "none". A policy change like this
> > > can't be made without being incredibly sure we're not going to screw
> > > up existing installations.
> > 
> > I'll let Chandra address this, as it is my understanding that he has
> > hardware and has tested this code with it.
> > 
> > > The second point I made speaks to the technical ugliness of this: what
> > > you're basically doing is open coding multiple binding for a device
> > > handler specific case.  If you can persuade me the policy above is
> > > correct, then technically all this should be done correctly via multiple
> > > binding in the generic device core ... before this interface nastiness
> > > you're constructing propagates outwards and becomes part of the user
> > > ABI.
> > 
> > I'm willing to re-implement the functionality with a different mechanism
> > if it has your blessing, if you can be specific about what it is you think
> > would be better.  Obviously I'll hold off on that until we've come to some
> > agreement about the other aspects.
> 
> Well, the two things aren't so dependent: this dh_state attribute that
> hannes just NAK'd is precisely a binding attribute for your hand rolled

Just to clarify:

dh_state attribute is not added new, it was originally added by Hannes. 

I just fixed a problem to make it work as it is intended, and sent an
explanation (about my fix) for Hannes to reconsider is NAK.

> multiple driver attachment, so actually getting generic device multiple
> binding sorted out would help out regardless of what the final
> attachment policy decision is.
> 
> James
> 
>