SCSI device order changed

SCSI device order changed

[Yuri van Oers]

Hi,

I've come across an issue which might be unintentional and in need of
fixing, in 2.6.17:

My machine has two SCSI adapters, one of which always has some disks
connected to it. One of those holds the root partition. The other card
occasionally gets some devices connected (burners and such), when I have
need for such a device.

Hard disks are sda thru sdd, the devices start at sde. That is, up until I
booted 2.6.17 _with_ the external devices attached. The kernel swapped
the order of the cards: devices start at sda, and the disks come after,
which means it can't find / on sda.

I found a related post here: http://lkml.org/lkml/2005/12/3/192
which suggests this situation arose around 2.6.15. It also says ordering
can't be guaranteed. If that's the final verdict, I'll simply swap the
cards on the PCI bus and be done with it.

So, is this a bug/problem?

Please CC me, I'm not on the list.

Thanks,
Yuri van Oers

Re: SCSI device order changed

[Jan Engelhardt]

>
>Hard disks are sda thru sdd, the devices start at sde. That is, up until I
>booted 2.6.17 _with_ the external devices attached. The kernel swapped
>the order of the cards: devices start at sda, and the disks come after,
>which means it can't find / on sda.
>
>I found a related post here: http://lkml.org/lkml/2005/12/3/192
>which suggests this situation arose around 2.6.15. It also says ordering
>can't be guaranteed. If that's the final verdict, I'll simply swap the
>cards on the PCI bus and be done with it.
>
And then, hope it does not change again?

It usually depends on link time order (when compiled in) or module load 
order (with initrd), both of which, esp. the latter, can be fine-tuned to 
the user's needs.

Also note CONFIG_BLK_DEV_OFFBOARD.



Jan Engelhardt
-- 

Re: SCSI device order changed

[Yuri van Oers]

On Sun, 16 Jul 2006, Jan Engelhardt wrote:

> >Hard disks are sda thru sdd, the devices start at sde. That is, up until I
> >booted 2.6.17 _with_ the external devices attached. The kernel swapped
> >the order of the cards: devices start at sda, and the disks come after,
> >which means it can't find / on sda.
>
> It usually depends on link time order (when compiled in) or module load
> order (with initrd), both of which, esp. the latter, can be fine-tuned to
> the user's needs.
>
> Also note CONFIG_BLK_DEV_OFFBOARD.

Ok, but I think none of the above applies in this situation, because the
cards use the same driver (AIC7xxx).


Regards,
Yuri van Oers
(please CC me, I'm not on the list)

Re: SCSI device order changed

[Jan Engelhardt]

>
>Ok, but I think none of the above applies in this situation, because the
>cards use the same driver (AIC7xxx).
>
In which case the ... usual(!) ... behavior is "the lowest PCI 
slot/address" gets the first c0 (as in c0t0d0s0) and therefore the first 
disk spots sda, sdb, etc.
But anything may happen :>


Jan Engelhardt
-- 

Re: SCSI device order changed

[Yuri van Oers]

On Mon, 17 Jul 2006, Jan Engelhardt wrote:

> >Ok, but I think none of the above applies in this situation, because the
> >cards use the same driver (AIC7xxx).
>
> In which case the ... usual(!) ... behavior is "the lowest PCI
> slot/address" gets the first c0 (as in c0t0d0s0) and therefore the first
> disk spots sda, sdb, etc.
> But anything may happen :>

Indeed. I tracked the change to 2.6.13-rc4, which includes a patch to
always probe in bus order.

So I'll turn my machine off now and swap the cards.

Regards,
Yuri

RE: SCSI device order changed

[Gary Funck]

Yuri van Oers wrote:
> I found a related post here: http://lkml.org/lkml/2005/12/3/192
> which suggests this situation arose around 2.6.15. It also says ordering
> can't be guaranteed. If that's the final verdict, I'll simply swap the
> cards on the PCI bus and be done with it.
> 

We saw this change in behavior also.  Labeling the root volume, and
mounting via label may help to keep things more portable, however,
I recall that labels and the device manager managed devices interacted
in some way that made mounting / via a label more difficult, however.