[Linux-cachefs] Re: [PATCH][RFC] NFS: Adding hooks to enable the use of FS-Cache

[Trond Myklebust]

On Fri, 2006-04-21 at 11:14 -0400, Steve Dickson wrote:
> This patch (based on Trond's nfs-2.6 git tree) adds the hooks need to
> allow NFS to use the generic filesystem caching facility FS-Cache.
> Caching is done on a per mount bases. To activate caching, the 'fsc'
> flag is need as a mount option.
> Ex:
>      mount -o fsc server:/export /mnt
> 
> (binary and source rpms as well as the actual patch are available at
>   http://people.redhat.com/steved/cachefs/util-linux/)

Hmm... Are you sure that is based on the latest trees? AFAICS it
shouldn't even apply to stock 2.6.17-rc2 since it doesn't take into
account the fact that we already use PagePrivate()+nfs_invalidatepage in
order to fix truncate races.
Could you look into that please?

Cheers,
  Trond

> NFS uses the cache on a readonly bases; Meaning only reads are caches
> not writes. But this *does not* mean filesystems have to be mounted
> read only. Filesystems are mounted read/write but only data from
> reads will be cached. When a file is opened with either RW or WR
> bits, the caching will be turned off (see nfs_open for the details).
> The reason for read only is became quite apparent that keeping the cache 
> coherent while jugging both reads and writes is pretty hard. So it was
> decided to start with a read only implementation and then move on to a
> read/write implementation in the future...
> 
> There are still a few rough spots in this patch. One being the
> sysctl variables that show what, if any, caching is being done need
> to be rolled into NFS I/O metrics code that recently went in.
> 
> The second one being, what David termed as, NFS aliasing. The way
> NFS manages its super blocks would cause double caching since the
> cookies returned by the cache would not be unique. We got around
> this problem by making each nfs mount have unique 'fsctag' which
> would be used to generate the caching cookie. The default fsctag
> value is the mount point but it also can be set on the command line
> (which is undocumented at this point since its not clear we really
> want to do this).
> 
> Unfortunately having the mount command passed the kernel a string turned
> out to a bit difficult since NFS used a binary interface (something
> that will hopefully change this summer). So David came up with the idea
> (not too surprising) of using the keyrings to passed down the string
> which did work very well... But as I said, we have every intending of
> change the binary interface into an ascii one and when that happens,
> this interface will change.
> 
> Finally there still needs to be some work on mmapped pages. Its not
> clear NFS is used the cache as well as it could when pages are mmapped.
> I'm still investigating...
> 
> Now I'm sure Trond will find several more rough spots... ;-) but
> these are all of the ones I know about....
> 
> steved.
> 
>