[Linux-cachefs] [PATCH 00/14] Permit filesystem local caching and NFS superblock sharing [try #10]

[David Howells]

These patches make it possible to share NFS superblocks between related
mounts, where "related" means on the same server and FSID. Inodes and dentries
will be shared where the NFS filehandles are the same (for example if two NFS3
files come from the same export but from different mounts, such as is not
uncommon with autofs on /home).

These patches also add local caching for network filesystems such as NFS and
AFS.

The first six patches (NFS superblock sharing) can be applied without the
remaining patches (filesystem local caching).

Changes [try #9] that have been made:

 (*) [PATCH] NFS: Permit filesystem to perform statfs with a known root dentry

     [*] Inclusions of linux/mount.h have been added where necessary to make
       	 allyesconfig build successfully.

 (*) [PATCH] NFS: Share NFS superblocks per-protocol per-server per-FSID

     [*] The exports from fs/namespace.c and fs/namei.c are no longer required.

 (*) [PATCH] FS-Cache: Release page->private in failed readahead

     [*] The try_to_release_page() is called instead of calling the
     	 releasepage() op directly.

     [*] The page is locked before try_to_release_page() is called.

     [*] The call to try_to_release_page() and page_cache_release() have been
     	 abstracted out into a helper function as this bit of code occurs
     	 twice..

Changes [try #10] that have been made:

 (*) [PATCH] NFS: Permit filesystem to perform statfs with a known root dentry

     [*] Pass a dentry rather than a vfsmount to the statfs() op as the key by
     	 which to determine the filesystem.

 (*) [PATCH] NFS: Share NFS superblocks per-protocol per-server per-FSID

     [*] nfs4_pathname_string() needed an extra const.

 (*) [PATCH] FS-Cache: Release page->private in failed readahead

     [*] The comment header on the helper function is much expanded.  This
     	 states why there's a need to call the releasepage() op in the event of
     	 an error.

     [*] BUG() if the page is already locked when we try and lock it.

     [*] Don't set the page mapping pointer until we've locked the page.

     [*] The page is unlocked after try_to_release_page() is called.

 (*) The release-page patch now comes before the fscache-afs patch as well as
     the fscache-nfs patch.

---

In response to those who've asked, there are at least three reasons for
implementing superblock sharing:

 (1) As I understand what I've been told, NFSv4 requires a 1:1 mapping between
     server files and client files.  I suspect this has to do with the
     management of leases.

 (2) We can reduce the resource consumption on NFSv2 and NFSv3 clients as well
     as on NFSv4 clients by sharing superblocks that cover overlapping segments
     of the file space.

     Consider a machine that's used by a lot of people at the same time, each
     of whom has an automounted NFS homedir off of the same server - and in
     fact off of the same disk on the that server.  Currently, with Linus's
     tree, each one will get a separate superblock to represent them; with
     Trond's tree, each one will still get a separate superblock unless they
     share the same root filehandle; and with my patches, they'll get the same
     superblock.

     If two homedirs have a hard link between them (unlikely, I know, but by no
     means impossible, and probably more likely with, say, data such as NFS
     mounted git repositories), then you have the possibility of aliasing.
     This means that you can have two or more inodes in core that refer to the
     same server object, and each of these inodes can have pages that refer to
     the same remote pages on the server - aliasing again.  You _have_ to have
     two inodes because they're covered by separate superblocks.

     Aliasing is bad, generally, because you end up using more storage than
     you need to (pagecache and inode cache in this case), and you have the
     problem of keeping them in sync.  It's also twice as hard to keep two
     inodes up to date when they change on the server as to keep one up to
     date.

     If you can use the same superblock where possible, then you can cut out
     aliasing on that client since you can share dentries that have the same
     file handle (hard links or subtrees).

     Part of the problem with NFSv2 and NFSv3 is that you invoke mountd to get
     the filehandle to a subtree, but you may not be able to work out how two
     different subtrees relate.  The getsb patch permits the superblock to
     have more than one root, which allows us to defer this problem until we
     see the root of one subtree cropping up in another subtree - at which
     point we can splice the former into the latter.

 (3) In my local file caching patches (FS-Cache), I have two reasons for
     wanting this:

     (a) Unique keys.  I need a unique key to find an object in the cache.  If
     	 we can get inode aliases, then I end up with several inodes referring
     	 to the same cache object.  This also means that I have to use a fair
     	 bit of extra memory to keep track of the multiple cookie mappings in
     	 FS-Cache, and have to compare keys a lot to find duplicate mappings.

	 If I can assume that the _netfs_ will manage the 1:1 mapping, I can
	 use a lot less memory and save some processing capacity also.

	 I don't want to invent random keys to differentiate aliased
	 superblocks or inodes as that destroys the persistence capabilities
	 of the cache across power failures and reboots.

     (b) Callbacks.  I want a callback that the netfs passes to FS-Cache to
     	 permit the cache to update the metadata in the cache from netfs
     	 metadata at convenient times.  However, if there's more than one
     	 inode alias in core, which one should the cache use?

AFS doesn't have anything like these problems because mounts are always made
from the root of a volume, and AFS was designed with local caching in mind.

The getsb and statfs patches are a consequence of NFS being permitted to mount
arbitrary subtrees from the server.

David