Multipath I/O with Linux

[Kodjo Agbenu]

Hello,

Has someone tried the multipath md driver with Linux, ie accessing the same scsi or fibre channel disk device via 2 or more adapters ?
Could you please post a sample config ?

Thank you very much for that.

Kodjo

[Dan Wanek · Accepted solution]

Attached is a raidtab for a multipath config on linux. Play with it a bit because it is a much different concept to multipathing than HP-UX if you are used to that.

[U.SivaKumar_2]

Hi,

read this thread

Alan Cox wrote:
> On Tue, 2002-09-10 at 15:06, Cameron, Steve wrote:
> > We can use the md driver for this. However, we cannot boot from
> > such a multipath device. Lilo and grub do not understand md multipath
> > devices, nor do anaconda or other installers. (Enhancing all of those,
> > I'd like to avoid. Cramming multipath i/o into the low level driver
> > would accomplish that, but, too yucky.)
> >
> > If there is work going on to enhance the multipath support in linux
> > it would be nice if you could boot from and install to such devices.
>
> Booting from them is a BIOS matter. If the BIOS can do the block loads
> off the multipath device we can do the rest. The probe stuff
> can be done
> in the boot initrd - as some vendors handle raid for example.


Well, the BIOS can do it if it has one working path, right?
(I think the md information is at the end of the partition,)
Maybe it will have some trouble if the primary path is failed,
but ignoring that for now. In the normal case, the bios shouldn't
even have to know it's a multipath device, right?


lilo and grub as they stand today, and anaconda (et al) as it
stands today, cannot do it. They can do RAID1 md devices only.
lilo for example will complain if you try to run it with
boot=/dev/md0, and /dev/md0 is not a RAID1 device. At least
it did when I tried it. When I looked at the lilo source, it
goes through each disk in the raid device and puts the boot
code on each one, a la RAID1. No provision is made for any other
kind of raid. Raid 5, naturally is much harder, and is unlikely
to have BIOS support, so this is understandable. Multipath seems
much closer to raid1 in terms of what's necessary for booting,
that is, much much easier than raid 5. I tried hacking on lilo a
bit, but so far, I am unsuccessful. I think grub cannot even do
RAID1.


I agree in principle, the initrd scripts can insmod multipath.o
to get things rolling, etc. The trouble comes from lilo, grub
and install time configuration.


-- steve


regards,
U.SivaKumar

[Kodjo Agbenu]

Thank you for this interesting information. However, I'm not yet interested in booting over multipath I/O.

Another point : I guess to be able to bypass lilo's limitations, one could do mirroring (RAID1) with disk1 seen via path1 and disk2 seen via path2.

Regards,

Kodjo

[Serviceguard for Linux]

You should also be aware of a couple of things.

1. Recovery after a repair is not automatic. If you are familiar with md for raid, you know that the status of a failed path will be marked with an "F". This is not cleared when that path is fixed. You need to manually go in and "repair" it with raidhotadd.
2. I believe I've seen cases where the partition information is not correct after a repair. Before the raidhotadd above, I'd recommend doing an sfdisk -R to the drive(s) to make sure the OS has this information correct.
3. Finally, I've heard reports, but not yet verified, that if one path is down when the raid is started, it will not start. I suspect but have not yet tested that you need to swap the "raid-disk" identification.

Hope this helps,
Rick