Live module for root=nfs using Aufs

Live module for root=nfs using Aufs

[Jon Ander Hernandez]

Hello,

We are currently developing a tool for creating simple beowulf
clusters, and we decided to use Dracut instead of initramfs-tools on
Debian. We found Dracut to be very well designed.

We have developed a very simple module, which allows us to have a
writable overlay over a read-only exported nfs4 using aufs :

http://kestrelhpc.svn.sourceforge.net/viewvc/kestrelhpc/Kestrelhpc/Src/dracut/union_mount_with_aufs.diff

Well, it is so general that also allows to get a writable overlay over
any root mounted by Dracut.

I've seen that you are changing Dracut's parameter's namespace, and I
would like to know if this kind of "live" use case could fit in them
somehow, because it seems that the actual live namespace implies the
classic live cd-rom use case.

We also have several patches for make Dracut 007's network work on
Ubuntu Maverick. Should I open a bug on Debian's bugzilla or should
they be reviewed on this list instead?

Cheers,

Jon Ander

Re: Live module for root=nfs using Aufs

[Seewer Philippe]

On 11/08/2010 01:43 AM, Jon Ander Hernandez wrote:
> Hello,
>
> We are currently developing a tool for creating simple beowulf
> clusters, and we decided to use Dracut instead of initramfs-tools on
> Debian. We found Dracut to be very well designed.
>
> We have developed a very simple module, which allows us to have a
> writable overlay over a read-only exported nfs4 using aufs :
>
> http://kestrelhpc.svn.sourceforge.net/viewvc/kestrelhpc/Kestrelhpc/Src/dracut/union_mount_with_aufs.diff
>
> Well, it is so general that also allows to get a writable overlay over
> any root mounted by Dracut.

I just had a quick look over the patch and in itself it looks ok. 
Although I'm not sure if having the aufs at this stage is necessary. 
What if you just mount it later inside the rootfs? (Background: I'm 
using aufs overlays for our read only nfs-roots, and I just put the 
necessary aufs mounts inside fstab)

> I've seen that you are changing Dracut's parameter's namespace, and I
> would like to know if this kind of "live" use case could fit in them
> somehow, because it seems that the actual live namespace implies the
> classic live cd-rom use case.
>
> We also have several patches for make Dracut 007's network work on
> Ubuntu Maverick. Should I open a bug on Debian's bugzilla or should
> they be reviewed on this list instead?

Please post your patches here. There's more eyes here I think :-)


Regards,
Philippe

Re: Live module for root=nfs using Aufs

[Jon Ander Hernandez]

2010/11/8 Seewer Philippe <philippe.seewer-omB+W0Dpw2o@public.gmane.org>:

> I just had a quick look over the patch and in itself it looks ok. Although
> I'm not sure if having the aufs at this stage is necessary. What if you just
> mount it later inside the rootfs? (Background: I'm using aufs overlays for
> our read only nfs-roots, and I just put the necessary aufs mounts inside
> fstab)

I have to say that I haven't thought about it before.... It is really
a good idea, and is probably better than have a specific module. :-)

But something that worries me, is if the init process will cause
problems, since we will be mounting a new root over the old one once
we have processes which have previously opened files on the old /.

Regards,

Jon Ander.

Re: Live module for root=nfs using Aufs

[Harald Hoyer]

On 11/08/2010 08:03 PM, Jon Ander Hernandez wrote:
> 2010/11/8 Seewer Philippe<philippe.seewer-omB+W0Dpw2o@public.gmane.org>:
>
>> I just had a quick look over the patch and in itself it looks ok. Although
>> I'm not sure if having the aufs at this stage is necessary. What if you just
>> mount it later inside the rootfs? (Background: I'm using aufs overlays for
>> our read only nfs-roots, and I just put the necessary aufs mounts inside
>> fstab)
>
> I have to say that I haven't thought about it before.... It is really
> a good idea, and is probably better than have a specific module. :-)
>
> But something that worries me, is if the init process will cause
> problems, since we will be mounting a new root over the old one once
> we have processes which have previously opened files on the old /.
>
> Regards,
>
> Jon Ander.

So, do you still need this dracut module? Or is it good enough, to do this in 
the real root?

Re: Live module for root=nfs using Aufs

[Jon Ander Hernandez]

2010/11/10 Harald Hoyer <harald-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>:
> On 11/08/2010 08:03 PM, Jon Ander Hernandez wrote:
>>
>> 2010/11/8 Seewer Philippe<philippe.seewer-omB+W0Dpw2o@public.gmane.org>:
>>
>>> I just had a quick look over the patch and in itself it looks ok.
>>> Although
>>> I'm not sure if having the aufs at this stage is necessary. What if you
>>> just
>>> mount it later inside the rootfs? (Background: I'm using aufs overlays
>>> for
>>> our read only nfs-roots, and I just put the necessary aufs mounts inside
>>> fstab)
>>
>> I have to say that I haven't thought about it before.... It is really
>> a good idea, and is probably better than have a specific module. :-)
>>
>> But something that worries me, is if the init process will cause
>> problems, since we will be mounting a new root over the old one once
>> we have processes which have previously opened files on the old /.
>>
>> Regards,
>>
>> Jon Ander.
>
> So, do you still need this dracut module? Or is it good enough, to do this
> in the real root?

The short answer is I still require something which makes the union
mount very early.
The long answer is I don't know if it is feasible to do the union
mount on the fstab or if the initramfs is the only option. The idea
doesn't seem to be so bad, but I find terrible dificul to debug the
result when in that ecuation you have nfs4 and upstart. So I don't
exactly know what is going on before my nodes get freezed...

I'll need to make more tests, and debug a little more.