On Wed, Jan 30, 2013 at 03:09:38PM -0500, J. Bruce Fields wrote: > On Wed, Jan 30, 2013 at 02:31:09PM -0500, bfields wrote: > > On Fri, Jan 25, 2013 at 03:23:28PM +0100, Niels de Vos wrote: > > > Hi all, > > > > > > the last few days I have been looking into making the tracking of > > > NFS-clients more persistent. As it is today, the NFS-clients are kept in > > > a list in memory on the NFS-server. When the NFS-server restarts, the > > > list is recreated from scratch and does not contain the NFS-clients that > > > still have the export mounted (Bug 904065). > > > > > > NFSv3 depends on the MOUNT protocol. When an NFS-client mounts an > > > export, the MOUNT protocol is used to get the initial file-handle. With > > > this handle, the NFS-service can be contacted. The actual services > > > providing the MOUNT and NFSv3 protocol can be separate (Linux kernel > > > NFSd) or implemented closely together (Gluster NFS-server). > > > > > > Now, when the Linux kernel NFS-server is used, the NFS-clients are saved > > > my the rpc.mountd process (which handles the MOUNT protocol) in > > > /var/lib/nfs/rwtab. This file is modified on mounting and unmounting. > > > Implementing a persistent cache similar to this is pretty straight > > > forward and is available for testing and review in [1]. > > > > > > There are however some use-cases that may require some different > > > handling. When an NFS-server starts to mount an export, the MOUNT > > > protocol is handled on a specific server. After getting the initial > > > file-handle for the export, any Gluster NFS-server can be used to talk > > > NFSv3 and do I/O. When the NFS-clients are kept only on the NFS-server > > > that handled the initial MOUNT request, and due to fail-over (think CTDB > > > and similar here) an other NFS-server is used, the persistent cache of > > > 'connected' NFS-clients is inaccurate. > > > > > > The easiest way I can think of to remedy this issue, is to place the > > > persistent NFS-client cache on a GlusterFS volume. When CTDB is used, > > > the locking-file and is placed on a shared storage as well, so the same > > This is the statd data? That's the more important thing to get right. Uhm, no. The locking-file I meant is for CTDB itself (I think). From my understanding the statd/NFS-locking is done through the GlusterFS-client (the NFS-server is a client, just like a FUSE-mount). For all I know the statd/NFS-locking is working as it should. > > > volume can be used for the NFS-client cache. Providing an option > > > to set the volume/path of the NFS-client cache would be needed for > > > this. I guess that this could result in a chicken-and-egg > > > situation (NFS-server is started, but no volume mounted yet)? > > I don't think there should be any problem here: the exported filesystems > need to be available before the server starts anyway. (Otherwise the > only response the server could give to operations on filehandles would > be ESTALE.) Well, the NFS-server dynamically gets exports (GlusterFS volumes) added when these are started or newly created. There is no hard requirement that a specific volume is available for the NFS-server to place a shared files with a list of NFS-clients. Probably easily solved by making the path to the file configurable and only accessing it when needed (and not at startup of the NFS-server). > > --b. > > > > > > > Any ideas or recommendations are welcome. The patch in [1] is not final > > > yet and I'd like some feedback before I proceed any further. > > > > My only comment is that this doesn't need to be perfect or even all that > > good. > > > > The list can already get out of sync in other ways: clients can just > > fail to unmount, for example. > > > > I don't think it's used by anything other than showmount. > -- Niels de Vos Sr. Software Maintenance Engineer Support Engineering Group Red Hat Global Support Services
