[Unionfs] Re: rw nfs branch (Re: unionfs root on nfs)

[Martin Kreiner]

Hans-Peter Jansen schrieb:
> Hi Junjiro-San, hi *,
> 
> Am Freitag, 4. November 2005 16:47 schrieb hooanon05 at yahoo.co.jp:
> 
>>Here is my patch for nfs and procfs, it is against
>>unionfs-20051025-2113. I applied it after my rename2.patch + lookup
>>patch, both patches are posted to this ml on 27 October.
>>nfs patch calls some kernel internal functions, so you need to link
>>unionfs statically instead of dynamic module.
>>If you meet a problem again, try '#define Debug' in proc.c (new
>>file). It will print which process accessed to proc fs.
>>Although I tried procfs patch will be available in non-nfs
>>environment, I didnt test it much.
> 
> 
> I'm desperately seeking a (short time) fix for the /proc problem to get 
> my SuSE 9.3 diskless NFS setup going until the unionfs crew come along 
> with an "official" fix. How did you solved this in your environment? 
> Still based on this patch? The "statically linking" order isn't very 
> appealing, since I try very hard to keep the distribution kernel 
> untouched, if possible somehow.
> 
> I already evaluated these possibilies:
>  - server based union:
>    no problem with /proc, but server crashes forces all clients to
>    reboot, which isn't an option in my setup.. 
>  - client based union:
>    we all know about the consequences of missing mmap support
>    looked into patching startproc, but this is getting somewhat hairy, 
>    and only fixes one problem source
>    certain tricks with mount --move and pivot_root
> 
if startproc does not check inode consistency (like debians start-stop-daemon),
and pivot_root is not an option (because you are using initramfs whith
run-init) you can trick startproc (and other programms relying on "proper"
/proc/<PID>/exe links) by moving your nfsroot right after init starts (e.g. in
/etc/rcS.d/S00-unionfs):

mount --move "$NFSROOT" /

this works for me (with some additions).

martin

> None of these approaches fulfils my needs, but startproc patching still 
> looks like the most attractive ATM, which I will tackle tomorrow.
> 
> Regards,
> Pete