[dm-devel] Multipathing hints request

[christophe varoqui]

On jeu, 2005-07-21 at 16:31 +0200, Philipp Niemann wrote:
> Hello list, nice to meet you.
> 
> As I am new to the list I am not quite shure if the topic is correct,
> but...
> 
> Around here I have two EMC Clariion 4500 Storage Arrays, two Brocade Silkworms
> each building a fabric, and two QLogic 2340 HBA in my PC. For now I am
> using debian stable (sarge) with a vanilla kernel-2.6.12.3 compiled with
> all I can get from the md/devmapper part of config.
> 
> I want the OS to use the two HBAs for path failover. The Clariion is an
> activce/passive device.
> 
> I use the multipath-tools written by Christophe Varoqui and others,
> version 0.4.2.4-2 as packaged with debian. There is a more up to date
> one on the project homepage but I haved bothered yet.
> 
> Is this the right list to ask for help/suggestions? What details would
> you like to have then? If not the right list, where to post?
> 
> More details:
> 
> HBA1 is connected to Clariion1_SPA and Clariion2_SPA via fabric1
> HBA2 is connected to Clariion1_SPB and Clariion2_SPB via fabric2
> 
> I have the devices in /proc/partitions (total of 5 LUNs configured, each
> with 2 path makes 10 devices in /proc/partitions)
> 
> I have multipath configured like this:
> 
> # multipath -lv2 # should be group-by-node-name policy
> 3600601602001f1a665653dbe010465c5
> [size=22 GB][features="0"][hwhandler="0"]
> \_ round-robin 0 [enabled][first]
>   \_ 6:0:0:0 sda  8:0     [ready ][active]
>   \_ 7:0:1:0 sdh  8:112   [faulty][active]
> 
> 360060160200161a037f90d75dfca84f7
> [size=20 GB][features="0"][hwhandler="0"]
> \_ round-robin 0 [active][first]
>   \_ 6:0:1:1 sde  8:64    [faulty][active]
>   \_ 7:0:0:1 sdg  8:96    [ready ][active]
> 
> 3600601602001f1a600a524d18275c0b2
> [size=20 GB][features="0"][hwhandler="0"]
> \_ round-robin 0 [enabled][first]
>   \_ 6:0:0:1 sdb  8:16    [ready ][active]
>   \_ 7:0:1:1 sdi  8:128   [faulty][active]
> 
> 3600601602001f1a6e0c6fdbae081bd89
> [size=20 GB][features="0"][hwhandler="0"]
> \_ round-robin 0 [enabled][first]
>   \_ 6:0:0:2 sdc  8:32    [faulty][active]
>   \_ 7:0:1:2 sdj  8:144   [ready ][active]
> 
> 360060160200161a06215f38ffbb2d6e9
> [size=20 GB][features="0"][hwhandler="0"]
> \_ round-robin 0 [enabled][first]
>   \_ 6:0:1:0 sdd  8:48    [faulty][active]
>   \_ 7:0:0:0 sdf  8:80    [ready ][active]
> 
> 
> I think this doesn't look too bad. Ready/Faulty match the default owner for the
> LUN on the Clariion. I am able to create a filesystem say on
> 360060160200161a037f90d75dfca84f7, which might appear as dm-4 in /proc/partitions
> 
It does *not* look good. You ought to have 2 paths group per multipath.
One for active paths, the other for inactive paths. As you'll see this
is why the failover does not work for you.

>     BTW: Anyone to know what "hwhandler" tells me? Can I change it? Same for
>         "features".

The "hardware handler" is an optional additional kernel module used to
trigger a specific operation when a new Path Group get activated. This
is how the host will ask your Clariion controler to activate the
inactive paths, for example.

So yes, you need hwhandler="1 emc". "0" meaning no hwhandler at all.

>     BTW2: I'd prefer to use "emc" instead of "round-robin". How?
>         I concluded something like that might be possible as dm-round-robin
>             and dm-emc exist
> 
bzz. round-robin is the io speading policy, whereas "emc" is a
hwhandler. Different beast.

> I am able to mount that filesystem and to create files on it. I prefer doing something
> like dd if=/dev/zero of=/mnt/test bs=1024 count=$((1024*1024*19)). That way I have
> plenty of time to do some failover tests.
> 
> My problem (finally): If I pull the cable establishing the path to sdg, I get
> almost immediate hard IO errors, corrupting the filesystem. The devmapper
> won't switch to the second path, though it recognizes the failing
> disk sdg and disables the path 8:96. How do I do that properly?
> 
You should now know what to do :/

Regards,
-- 
christophe varoqui <christophe.varoqui at free.fr>