Re: [RFC][PATCH] Handle multipath paths in a path group properly during pg_init (was "RE: [dm-devel] DM does not activate the paths if there are more than one path in path group during failover")

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

On Fri, 2008-11-14 at 00:12 +0000, Alasdair G Kergon wrote:
> On Thu, Nov 13, 2008 at 03:48:58PM -0800, Chandra Seetharaman wrote:
> > The problem reported by Moger Babu was caused due to the architectural
> > change made when we moved from dm hardware handler to SCSI hardware
> > handler.
> 
> pg_init is single initialisation for the group.
> It only needs to happen on one path.

Yes, that was my initial response too.

dm hardware handlers did not maintain a "state" for each path in which
case activating/initializing the controller from one path was sufficient
for all the paths connected to the controller to work as expected.

SCSI Hardware Handlers maintain a "state" for each path, so that they
can fail an I/O at prep_fn, if the path is not an "active" path, which
is what lead to the bug Babu saw.

> 
> It sounds to me that what you're adding here is a *per-path* initialisation,
> something *separate* from (and additional to) the existing *per-group* pg_init.

No, I "replaced" the "per-group" pg_init with "per-path" init.

> 
> Has this patch been tested on EMC?

No, we have tested it with LSI RDAC. I will send a separate email to EMC
folks (on Cc here) describing the problem and the fix and ask their help
in testing.

> 
> Alasdair