KNFSD - The Kernel NFS Daemon for Linux

Intro

"G. Allen Morris III" <gam3@ixlabs.com> has a web page which claims to be the Kernal NFS Daemon Maintainer's Page, but it is a bit out of date now.

H.J. Lu of VA-Research did a lot of work getting knfsd in 2.2 to pass Connect-a-thon tests and did some SPEC-SFS benchmarks. He also maintains the user-land support programs. The can be found in in the linuxnfs project on SourceForge: http://www.linuxnfs.sourceforge.org

SGI have an "NFSv3" project at http://oss.sgi.com/projects/nfs3/.

Patches

My current project for 2.4.0-test is providing an nfsd_operations interface between knfsd and filesystems. Currently, ext2fs is the only filesystem in the main tree that can be effectively exported, however Chris Mason has patches to make reiserfs work.

I have one outstanding patches at the moment:

A

Add nfsd_operations support, including some support added to fs/dcache.c. With this, ext2fs is definately the only filesystem type that can be exported, but others can provide support through a well defined interface

Documentation

ToDo List

1. try to mount an mounted-on but not exported directory dies.(DONE)
2. UDP check sums aren't checked (DONE)
3. work around bug in Solaris7 wherebytop bit of mtime/atime gets set by excl create, but annoyes client. (DONE)
4. fix stable/non-stable write code in vfs.c (DONE)
5. make reply cache check port+proto aswell as ip+xid (DONE E)
6. really sync inodes now that we can (DONE)
7. make sure lockd is running each time we start nfsd- it might have been shot - already there.
8. check nfsd and lock are registered each time we start nfsd. Well, if you kill them off and then restart, it will be fine
9. If a filesystem is mounted in two places, make it possible to export the two differently
10. Linux currently uses 1-second resolution mtime. NFS can benefit from higher resolution to do cache consistancy checking. This could be provided without underlying filesystems being changed by:
    • Adding microsecond (or nanosecond?) fields the the inode structure
    • Having VFS set these appropriately
    • When an inode is loaded - preset the nanosecond field to the maximum possible value. (10^9-1) If the filesystem supports better resolution it can set it, otherwise th default value should do the closest possible to the "right thing"
11. In Linux 2.3, NFSv3 cannot return pre-op attributes for writes as it doesn't lock the inode (the lock happens lower down). This should be fixed, but it is not clear how.
12. The "nohide" export option is not fully supported by exportfs/mountd. When an export entry is given to the kernel (e.g. in response to a mount request), all export entries for child filesystems which are marked "nohide" should also be given to the kernel.
13. NFSD makes more demands on filesystems than VFS does. In some cases, NFSD could use certain services that a filesystem could supply, but doesn't because they aren't defined in the various _operations structures. There should be an nfsd_operations structure that filesystems could provide that would define operations like "convert filehandle to dentry" and "convert dentry to filehandle" and others.
14. NFSEXP_ASYNC should be OFF by default. This is controlled by the user-land utilities. This needs to be agreed to by users before the change is made.
15. NFSv4 support <grin>
16. Restructure export table to:
    1. Allow non-independant subtrees of a filesystem (e.g. / and /tmp) to be exported with different options (assuming subtreecheck).
    2. Be friendly to authentication styles other than AUTH_UNIX.
    3. work well with an upcall mechanism to tell mountd if an request arrives for an unknown client or export point. This would require negative entries, and timing out of unused enties.
17. Provide an upcall mechanism to mountd in case of a request or filehandle that cannot be authenticated/authorised by info in the kernel. It would have to allow for:
    1. Unknown client
    2. Unknown authentication scheme
    3. uninitialised authentication context.
    4. unknown filesystem
    5. unknown export options for client/filesystem

File System Interface

1. The filesystem must be a FS_REQUIRES_DEV filesystem - partly because a stable device number is needed for the handle, partly to exclude PROCFS and NFS from being exported.
2. The filesystem must provide a read_inode method which will take an inode number and provide the inode.
3. If read_inode is presented with an inode number which refers to an inactive (never used or deleted) inode, it should return a bad_inode created with make_bad_inode. Correspondingly, if delete_inode is presented with an is_bad_inode inode, it should quietly ignore it.
4. The filesystem should set and maintain the i_generation field in the inode structure. Without this nfsd might provide access to the wrong file when it should return ESTALE.
5. The filesystem must support lookup of ".." to find the parent of a directory. VFS never actually uses this, but most filesystems seem to provide it anyway.