Re: let md auto-detect 128+ raid members, fix potential race condition

From: Neil Brown
Date: Mon Jul 31 2006 - 21:17:31 EST

On Monday July 31, aoliva@xxxxxxxxxx wrote:
> On Jul 30, 2006, Neil Brown <neilb@xxxxxxx> wrote:
> > 1/
> > It just isn't "right". We don't mount filesystems from partitions
> > just because they have type 'Linux'. We don't enable swap on
> > partitions just because they have type 'Linux swap'. So why do we
> > assemble md/raid from partitions that have type 'Linux raid
> > autodetect'?
> Similar reason to why vgscan finds and attempts to use any partitions
> that have the appropriate type/signature (difference being that raid
> auto-detect looks at the actual partition type, whereas vgscan looks
> at the actual data, just like mdadm, IIRC): when you have to bootstrap
> from an initrd, you don't want to be forced to have the correct data
> in the initrd image, since then any reconfiguration requires the info
> to be introduced in the initrd image before the machine goes down.
> Sometimes, especially in case of disk failures, you just can't do
> that.

The initrd need to 'know' how to find the root filesystem, whether by
devnum or uuid or whatever.
In exactly the same way it needs to know how to find the components
for the root md array - uuid is the best. There is no need to
reconfigure this in the case of a disk failure.

Current mdadm will assemble arrays for you given only a hostname. You
still need to get the hostname into the initrd, but that is no
different from a root device number.

> > 2/
> > It can cause problems when moving devices.
> It can, indeed, and it has caused such problems to me before, but
> they're the exception, not the rule, and one should optimize for the
> rule, not the exception.

We aren't talking about optimisation. We are talking about whether it
actually works or not. A system that stops booting just because you
plugged a couple of extra drives in is a badly configured system.

> > 3/
> > The information redundancy can cause a problem when it gets out of
> > sync. i.e. you add a partition to a raid array without setting
> > the partition type to 'fd'. This works, but on the next reboot
> > the partition doesn't get added back into the array and you have
> > to manually add it yourself.
> > This too is not purely theory - it has been reported slightly more
> > often than '2'.
> This has happened to me as well, and I remember it was extremely
> confusing when it first happened :-) But that's an argument to change
> the behavior so as to look for the superblock instead of trusting the
> partition type, not an argument to remove the auto-detection
> feature.

As has been said, I don't want to remove auto-detection. I want to do
it right, and do it from userspace. It is in-kernel autodetection
that I have no interest in improving.

> And then, the reliance on partition type has been useful at times as
> well, when I explicitly did *not* want a certain raid device or raid
> member to be brought up on boot.

Well, at boot it should only bring up the raid array containing the
root filesystem. Everything else is best done by /etc/init.d
scripts. And you can stop those from running by booting with -s (or
whatever it is to get single-user).

> > So my preferred solution to the problem is to tell people not to use
> > autodetect. Quite possibly this should be documented in the code, and
> > maybe even have a KERN_INFO message if more than 64 devices are
> > autodetected.
> I wouldn't have a problem with that, since then distros would probably
> switch to a more recommended mechanism that works just as well, i.e.,
> ideally without requiring initrd-regeneration after reconfigurations
> such as adding one more raid device to the logical volume group
> containing the root filesystem.
> > If we were to 'fix' this problem, I think the cleanest approach (which
> > I haven't actually coded, so it might not work...) would be to define
> > a new flag to go in hd_struct->policy to say if the partition type
> > suggested auto-detect, and get partitions/check.c to set this.
> > Then have md iterate all partitions looking for this flag.
> AFAICT we'd still need a list or an array, since we add stuff back to
> the list in various situations.

No. We just need a list of partitions of the appropriate type.
Taking items of the list and putting them back on later is simply a
non-essential optimisation.

> > So: Do you *really* need to *fix* this, or can you just use 'mdadm'
> > to assemble you arrays instead?
> I'm not sure. I'd expect not to need it, but the limited feature
> currently in place, that initrd uses to bring up the raid1 devices
> containing the physical volumes that form the volume group where the
> logical volume with my root filesystem is also brings up various raid6
> physical volumes that form an unrelated volume group, and it does so
> in such a way that the last of them, containing the 128th fd-type
> partition in the box, ends up being left out, so the raid device it's
> a member of is brought up either degraded or missing the spare member,
> none of which are good.
> I don't know that I can easily get initrd to replace nash's
> raidautorun for mdadm unless mdadm has a mode to bring up any arrays
> it can find, as opposed to bringing up a specific array out of a given
> list of members or scanning for members. Either way, this won't fix
> the problem 2) that you mentioned, but requiring initrd-regeneration
> after extending the volume group containing the root device is another
> problem that the current modes of operation of mdadm AFAIK won't
> contemplate, so switching to it will trade one problem for another,
> and the latter is IMHO more common than the former.

Get mdadm 2.5.2 (or 2.5.3 if I get that out soon enough) and try

mdadm --assemble --scan --homehost='<system>' --auto-update-homehost \
--auto=yes --run

in your initrd, having set the hostname correctly first. It might do
exactly what you want.

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at