• D
    NFS: Share NFS superblocks per-protocol per-server per-FSID · 54ceac45
    David Howells 提交于
    The attached patch makes NFS share superblocks between mounts from the same
    server and FSID over the same protocol.
    
    It does this by creating each superblock with a false root and returning the
    real root dentry in the vfsmount presented by get_sb(). The root dentry set
    starts off as an anonymous dentry if we don't already have the dentry for its
    inode, otherwise it simply returns the dentry we already have.
    
    We may thus end up with several trees of dentries in the superblock, and if at
    some later point one of anonymous tree roots is discovered by normal filesystem
    activity to be located in another tree within the superblock, the anonymous
    root is named and materialises attached to the second tree at the appropriate
    point.
    
    Why do it this way? Why not pass an extra argument to the mount() syscall to
    indicate the subpath and then pathwalk from the server root to the desired
    directory? You can't guarantee this will work for two reasons:
    
     (1) The root and intervening nodes may not be accessible to the client.
    
         With NFS2 and NFS3, for instance, mountd is called on the server to get
         the filehandle for the tip of a path. mountd won't give us handles for
         anything we don't have permission to access, and so we can't set up NFS
         inodes for such nodes, and so can't easily set up dentries (we'd have to
         have ghost inodes or something).
    
         With this patch we don't actually create dentries until we get handles
         from the server that we can use to set up their inodes, and we don't
         actually bind them into the tree until we know for sure where they go.
    
     (2) Inaccessible symbolic links.
    
         If we're asked to mount two exports from the server, eg:
    
    	mount warthog:/warthog/aaa/xxx /mmm
    	mount warthog:/warthog/bbb/yyy /nnn
    
         We may not be able to access anything nearer the root than xxx and yyy,
         but we may find out later that /mmm/www/yyy, say, is actually the same
         directory as the one mounted on /nnn. What we might then find out, for
         example, is that /warthog/bbb was actually a symbolic link to
         /warthog/aaa/xxx/www, but we can't actually determine that by talking to
         the server until /warthog is made available by NFS.
    
         This would lead to having constructed an errneous dentry tree which we
         can't easily fix. We can end up with a dentry marked as a directory when
         it should actually be a symlink, or we could end up with an apparently
         hardlinked directory.
    
         With this patch we need not make assumptions about the type of a dentry
         for which we can't retrieve information, nor need we assume we know its
         place in the grand scheme of things until we actually see that place.
    
    This patch reduces the possibility of aliasing in the inode and page caches for
    inodes that may be accessed by more than one NFS export. It also reduces the
    number of superblocks required for NFS where there are many NFS exports being
    used from a server (home directory server + autofs for example).
    
    This in turn makes it simpler to do local caching of network filesystems, as it
    can then be guaranteed that there won't be links from multiple inodes in
    separate superblocks to the same cache file.
    
    Obviously, cache aliasing between different levels of NFS protocol could still
    be a problem, but at least that gives us another key to use when indexing the
    cache.
    
    This patch makes the following changes:
    
     (1) The server record construction/destruction has been abstracted out into
         its own set of functions to make things easier to get right.  These have
         been moved into fs/nfs/client.c.
    
         All the code in fs/nfs/client.c has to do with the management of
         connections to servers, and doesn't touch superblocks in any way; the
         remaining code in fs/nfs/super.c has to do with VFS superblock management.
    
     (2) The sequence of events undertaken by NFS mount is now reordered:
    
         (a) A volume representation (struct nfs_server) is allocated.
    
         (b) A server representation (struct nfs_client) is acquired.  This may be
         	 allocated or shared, and is keyed on server address, port and NFS
         	 version.
    
         (c) If allocated, the client representation is initialised.  The state
         	 member variable of nfs_client is used to prevent a race during
         	 initialisation from two mounts.
    
         (d) For NFS4 a simple pathwalk is performed, walking from FH to FH to find
         	 the root filehandle for the mount (fs/nfs/getroot.c).  For NFS2/3 we
         	 are given the root FH in advance.
    
         (e) The volume FSID is probed for on the root FH.
    
         (f) The volume representation is initialised from the FSINFO record
         	 retrieved on the root FH.
    
         (g) sget() is called to acquire a superblock.  This may be allocated or
         	 shared, keyed on client pointer and FSID.
    
         (h) If allocated, the superblock is initialised.
    
         (i) If the superblock is shared, then the new nfs_server record is
         	 discarded.
    
         (j) The root dentry for this mount is looked up from the root FH.
    
         (k) The root dentry for this mount is assigned to the vfsmount.
    
     (3) nfs_readdir_lookup() creates dentries for each of the entries readdir()
         returns; this function now attaches disconnected trees from alternate
         roots that happen to be discovered attached to a directory being read (in
         the same way nfs_lookup() is made to do for lookup ops).
    
         The new d_materialise_unique() function is now used to do this, thus
         permitting the whole thing to be done under one set of locks, and thus
         avoiding any race between mount and lookup operations on the same
         directory.
    
     (4) The client management code uses a new debug facility: NFSDBG_CLIENT which
         is set by echoing 1024 to /proc/net/sunrpc/nfs_debug.
    
     (5) Clone mounts are now called xdev mounts.
    
     (6) Use the dentry passed to the statfs() op as the handle for retrieving fs
         statistics rather than the root dentry of the superblock (which is now a
         dummy).
    Signed-Off-By: NDavid Howells <dhowells@redhat.com>
    Signed-off-by: NTrond Myklebust <Trond.Myklebust@netapp.com>
    54ceac45
Makefile 641 字节