In message <[EMAIL PROTECTED]>,
"=?ISO-8859-1?Q?P=E1sztor_L=E9n=E1rd_Zolt=E1n?=" writes:
> Hi!
>
> The newest release with 2.6.22.1 kernel works stably for me, thanks for the
> fixes!
> The cache coherency feature also working fine with the specified mount
> option which increments the generation number of the superblock (incgen).
>
> I have multiple servers with low nfs traffic.
> I'd love to see if the unionfs do not cache any objects in it's branches, so
> every file access means an object-reget.
> Is it possible to implement as a mount option?
> Or is there any solution to get always the latest filesystem state without
> user interaction?
>
> For example if I do a remount with incgen option in every 10 seconds it may
> solve the problem but I think it's an ugly hack.
Lenard, what version of Unionfs have you been using? I assume by "newest"
you mean something in which we added the mtime-based cache-coherency
support. Are you? All of our recent releases include that automatic
cache-coherency support.
With the mtime-based cache-coherency support, you do NOT need to use the
incgen explicitly any longer. That was needed in the past in order to force
Unionfs to invalidate all lower objects and re-get them; it had to be done
manually by users after they've modified lower objects. But now, Unionfs
automatically figures out when a lower object had changed, and it will
re-get the lower objects as needed.
Using the 'incgen' remount option should be a last-resort; it will cause
unionfs to purge all lower objects, and then they have to be re-gotten
incrementally; this could affect your performance. And if we add an option
to always re-get lower objects, performance would suffer greatly.
So I'm curious, if you're using the latest unionfs2 code, why you still
find it necessary to use the incgen option?
One possibility is that you mount unionfs on top of nfs mounts, and the nfs
mounts don't reflect server-side changes fast enough for you? This would be
an NFS issue, since it caches attributes from the server onto the client,
and doesn't purge them frequently enough. In that case, I'd recommend that
you mount all of your NFS mounts with "-o noac" (No Attribute Cache), or
play with the four options ac{reg,dir}{min,max}.
In other words, if the lower file system doesn't see new objects' data and
attributes (esp. mtime/ctime), then Unionfs, sitting above, cannot see them
either -- thus you have to first make the lower file systems see the updated
file attributes.
Cheers,
Erez.
_______________________________________________
unionfs mailing list: http://unionfs.filesystems.org/
unionfs@mail.fsl.cs.sunysb.edu
http://www.fsl.cs.sunysb.edu/mailman/listinfo/unionfs
