1:orphan:
2
3Making Filesystems Exportable
4=============================
5
6Overview
7--------
8
9All filesystem operations require a dentry (or two) as a starting
10point.  Local applications have a reference-counted hold on suitable
11dentries via open file descriptors or cwd/root.  However remote
12applications that access a filesystem via a remote filesystem protocol
13such as NFS may not be able to hold such a reference, and so need a
14different way to refer to a particular dentry.  As the alternative
15form of reference needs to be stable across renames, truncates, and
16server-reboot (among other things, though these tend to be the most
17problematic), there is no simple answer like 'filename'.
18
19The mechanism discussed here allows each filesystem implementation to
20specify how to generate an opaque (outside of the filesystem) byte
21string for any dentry, and how to find an appropriate dentry for any
22given opaque byte string.
23This byte string will be called a "filehandle fragment" as it
24corresponds to part of an NFS filehandle.
25
26A filesystem which supports the mapping between filehandle fragments
27and dentries will be termed "exportable".
28
29
30
31Dcache Issues
32-------------
33
34The dcache normally contains a proper prefix of any given filesystem
35tree.  This means that if any filesystem object is in the dcache, then
36all of the ancestors of that filesystem object are also in the dcache.
37As normal access is by filename this prefix is created naturally and
38maintained easily (by each object maintaining a reference count on
39its parent).
40
41However when objects are included into the dcache by interpreting a
42filehandle fragment, there is no automatic creation of a path prefix
43for the object.  This leads to two related but distinct features of
44the dcache that are not needed for normal filesystem access.
45
461. The dcache must sometimes contain objects that are not part of the
47   proper prefix. i.e that are not connected to the root.
482. The dcache must be prepared for a newly found (via ->lookup) directory
49   to already have a (non-connected) dentry, and must be able to move
50   that dentry into place (based on the parent and name in the
51   ->lookup).   This is particularly needed for directories as
52   it is a dcache invariant that directories only have one dentry.
53
54To implement these features, the dcache has:
55
56a. A dentry flag DCACHE_DISCONNECTED which is set on
57   any dentry that might not be part of the proper prefix.
58   This is set when anonymous dentries are created, and cleared when a
59   dentry is noticed to be a child of a dentry which is in the proper
60   prefix.  If the refcount on a dentry with this flag set
61   becomes zero, the dentry is immediately discarded, rather than being
62   kept in the dcache.  If a dentry that is not already in the dcache
63   is repeatedly accessed by filehandle (as NFSD might do), an new dentry
64   will be a allocated for each access, and discarded at the end of
65   the access.
66
67   Note that such a dentry can acquire children, name, ancestors, etc.
68   without losing DCACHE_DISCONNECTED - that flag is only cleared when
69   subtree is successfully reconnected to root.  Until then dentries
70   in such subtree are retained only as long as there are references;
71   refcount reaching zero means immediate eviction, same as for unhashed
72   dentries.  That guarantees that we won't need to hunt them down upon
73   umount.
74
75b. A primitive for creation of secondary roots - d_obtain_root(inode).
76   Those do _not_ bear DCACHE_DISCONNECTED.  They are placed on the
77   per-superblock list (->s_roots), so they can be located at umount
78   time for eviction purposes.
79
80c. Helper routines to allocate anonymous dentries, and to help attach
81   loose directory dentries at lookup time. They are:
82
83    d_obtain_alias(inode) will return a dentry for the given inode.
84      If the inode already has a dentry, one of those is returned.
85
86      If it doesn't, a new anonymous (IS_ROOT and
87      DCACHE_DISCONNECTED) dentry is allocated and attached.
88
89      In the case of a directory, care is taken that only one dentry
90      can ever be attached.
91
92    d_splice_alias(inode, dentry) will introduce a new dentry into the tree;
93      either the passed-in dentry or a preexisting alias for the given inode
94      (such as an anonymous one created by d_obtain_alias), if appropriate.
95      It returns NULL when the passed-in dentry is used, following the calling
96      convention of ->lookup.
97
98Filesystem Issues
99-----------------
100
101For a filesystem to be exportable it must:
102
103   1. provide the filehandle fragment routines described below.
104   2. make sure that d_splice_alias is used rather than d_add
105      when ->lookup finds an inode for a given parent and name.
106
107      If inode is NULL, d_splice_alias(inode, dentry) is equivalent to::
108
109		d_add(dentry, inode), NULL
110
111      Similarly, d_splice_alias(ERR_PTR(err), dentry) = ERR_PTR(err)
112
113      Typically the ->lookup routine will simply end with a::
114
115		return d_splice_alias(inode, dentry);
116	}
117
118
119
120A file system implementation declares that instances of the filesystem
121are exportable by setting the s_export_op field in the struct
122super_block.  This field must point to a "struct export_operations"
123struct which has the following members:
124
125  encode_fh (optional)
126    Takes a dentry and creates a filehandle fragment which may later be used
127    to find or create a dentry for the same object.  The default
128    implementation creates a filehandle fragment that encodes a 32bit inode
129    and generation number for the inode encoded, and if necessary the
130    same information for the parent.
131
132  fh_to_dentry (mandatory)
133    Given a filehandle fragment, this should find the implied object and
134    create a dentry for it (possibly with d_obtain_alias).
135
136  fh_to_parent (optional but strongly recommended)
137    Given a filehandle fragment, this should find the parent of the
138    implied object and create a dentry for it (possibly with
139    d_obtain_alias).  May fail if the filehandle fragment is too small.
140
141  get_parent (optional but strongly recommended)
142    When given a dentry for a directory, this should return  a dentry for
143    the parent.  Quite possibly the parent dentry will have been allocated
144    by d_alloc_anon.  The default get_parent function just returns an error
145    so any filehandle lookup that requires finding a parent will fail.
146    ->lookup("..") is *not* used as a default as it can leave ".." entries
147    in the dcache which are too messy to work with.
148
149  get_name (optional)
150    When given a parent dentry and a child dentry, this should find a name
151    in the directory identified by the parent dentry, which leads to the
152    object identified by the child dentry.  If no get_name function is
153    supplied, a default implementation is provided which uses vfs_readdir
154    to find potential names, and matches inode numbers to find the correct
155    match.
156
157  flags
158    Some filesystems may need to be handled differently than others. The
159    export_operations struct also includes a flags field that allows the
160    filesystem to communicate such information to nfsd. See the Export
161    Operations Flags section below for more explanation.
162
163A filehandle fragment consists of an array of 1 or more 4byte words,
164together with a one byte "type".
165The decode_fh routine should not depend on the stated size that is
166passed to it.  This size may be larger than the original filehandle
167generated by encode_fh, in which case it will have been padded with
168nuls.  Rather, the encode_fh routine should choose a "type" which
169indicates the decode_fh how much of the filehandle is valid, and how
170it should be interpreted.
171
172Export Operations Flags
173-----------------------
174In addition to the operation vector pointers, struct export_operations also
175contains a "flags" field that allows the filesystem to communicate to nfsd
176that it may want to do things differently when dealing with it. The
177following flags are defined:
178
179  EXPORT_OP_NOWCC - disable NFSv3 WCC attributes on this filesystem
180    RFC 1813 recommends that servers always send weak cache consistency
181    (WCC) data to the client after each operation. The server should
182    atomically collect attributes about the inode, do an operation on it,
183    and then collect the attributes afterward. This allows the client to
184    skip issuing GETATTRs in some situations but means that the server
185    is calling vfs_getattr for almost all RPCs. On some filesystems
186    (particularly those that are clustered or networked) this is expensive
187    and atomicity is difficult to guarantee. This flag indicates to nfsd
188    that it should skip providing WCC attributes to the client in NFSv3
189    replies when doing operations on this filesystem. Consider enabling
190    this on filesystems that have an expensive ->getattr inode operation,
191    or when atomicity between pre and post operation attribute collection
192    is impossible to guarantee.
193
194  EXPORT_OP_NOSUBTREECHK - disallow subtree checking on this fs
195    Many NFS operations deal with filehandles, which the server must then
196    vet to ensure that they live inside of an exported tree. When the
197    export consists of an entire filesystem, this is trivial. nfsd can just
198    ensure that the filehandle live on the filesystem. When only part of a
199    filesystem is exported however, then nfsd must walk the ancestors of the
200    inode to ensure that it's within an exported subtree. This is an
201    expensive operation and not all filesystems can support it properly.
202    This flag exempts the filesystem from subtree checking and causes
203    exportfs to get back an error if it tries to enable subtree checking
204    on it.
205
206  EXPORT_OP_CLOSE_BEFORE_UNLINK - always close cached files before unlinking
207    On some exportable filesystems (such as NFS) unlinking a file that
208    is still open can cause a fair bit of extra work. For instance,
209    the NFS client will do a "sillyrename" to ensure that the file
210    sticks around while it's still open. When reexporting, that open
211    file is held by nfsd so we usually end up doing a sillyrename, and
212    then immediately deleting the sillyrenamed file just afterward when
213    the link count actually goes to zero. Sometimes this delete can race
214    with other operations (for instance an rmdir of the parent directory).
215    This flag causes nfsd to close any open files for this inode _before_
216    calling into the vfs to do an unlink or a rename that would replace
217    an existing file.
218
219  EXPORT_OP_REMOTE_FS - Backing storage for this filesystem is remote
220    PF_LOCAL_THROTTLE exists for loopback NFSD, where a thread needs to
221    write to one bdi (the final bdi) in order to free up writes queued
222    to another bdi (the client bdi). Such threads get a private balance
223    of dirty pages so that dirty pages for the client bdi do not imact
224    the daemon writing to the final bdi. For filesystems whose durable
225    storage is not local (such as exported NFS filesystems), this
226    constraint has negative consequences. EXPORT_OP_REMOTE_FS enables
227    an export to disable writeback throttling.
228
229  EXPORT_OP_NOATOMIC_ATTR - Filesystem does not update attributes atomically
230    EXPORT_OP_NOATOMIC_ATTR indicates that the exported filesystem
231    cannot provide the semantics required by the "atomic" boolean in
232    NFSv4's change_info4. This boolean indicates to a client whether the
233    returned before and after change attributes were obtained atomically
234    with the respect to the requested metadata operation (UNLINK,
235    OPEN/CREATE, MKDIR, etc).
236
237  EXPORT_OP_FLUSH_ON_CLOSE - Filesystem flushes file data on close(2)
238    On most filesystems, inodes can remain under writeback after the
239    file is closed. NFSD relies on client activity or local flusher
240    threads to handle writeback. Certain filesystems, such as NFS, flush
241    all of an inode's dirty data on last close. Exports that behave this
242    way should set EXPORT_OP_FLUSH_ON_CLOSE so that NFSD knows to skip
243    waiting for writeback when closing such files.
244