[dm-devel] dm howto?

[christophe varoqui]

Le dim 13/06/2004 à 05:15, gfbarros at ualr.edu a écrit :
> is there documentation somewhere on making dm work with multipath in rhel3? im just wondering if most of this is already in the kernel provided by redhat before i start tearing things apart and rebuilding :)
> 
I guess RedHat packages LVM1
and in no way it packages dm-multipath.
They plan to do so for RHEL4.
They are still lagging to integrate this in Fedora Core, so I'm not that
confident. Alastair should know better.

Right now, it's securepath for RHEL or nothing. DM will void your
support agreement.

> ive played with using mdadm a bit but it doesnt seems to deal with noticing dead paths and re-adding them too well...
> 
> Ive also noticed that disks presented by one of my controllers (hsv100) are not ready:
> 
> SCSI device sdb: 2097152 512-byte hdwr sectors (1074 MB)
> sdb: asking for cache data failed
> sdb: assuming drive cache: write through
>  sdb:<6>Device sdb not ready.
> end_request: I/O error, dev sdb, sector 0
> Buffer I/O error on device sdb, logical block 0
> Device sdb not ready.
> end_request: I/O error, dev sdb, sector 0
> Buffer I/O error on device sdb, logical block 0
>  unable to read partition table
> 
Yes, these are the ghosts paths :
each LUN is assigned to a single controler, so valid paths are all
through this LUN-master-controler. The other controler present ghost
paths, so that activating them doesn't need action on the host's side
(scsi rescan not needed).

> but i only see this in my Debian box, the RHEL3 one only actually sees one path per controller pair... which im guessing is the one to the active controller. Im guessing that bad things would happen if that controller died, as the kernel would need me to reload the qla2300 module before seeing the new presentation... im assuming report_lun fixes this, how do i set that in the stock kernel? or am i sol?
> 
Yes, Debian and standard kernel have to right behaviour.
Not sure what cause to different one in the RedHat kernel ... may be the
Qlogic failover implementation.

> would this have something to do with the HSV's not being really active-active? how is this dealt with? (is this the group_by_tur feature in multipath-tools?)
> 
Some call this controler family "assymetric". IBM have the same case
with FastT ctlrs. Yes, group_by_tur was created for assymetric ctlrs.
The multipath policy works best with symmetric ones.

> if this is not the right list for these questions, please steer me in the right direction :) 
> 
This is the one.

regards,
cvaroqui