Loading...
1
2 Overview of the Linux Virtual File System
3
4 Original author: Richard Gooch <rgooch@atnf.csiro.au>
5
6 Last updated on June 24, 2007.
7
8 Copyright (C) 1999 Richard Gooch
9 Copyright (C) 2005 Pekka Enberg
10
11 This file is released under the GPLv2.
12
13
14Introduction
15============
16
17The Virtual File System (also known as the Virtual Filesystem Switch)
18is the software layer in the kernel that provides the filesystem
19interface to userspace programs. It also provides an abstraction
20within the kernel which allows different filesystem implementations to
21coexist.
22
23VFS system calls open(2), stat(2), read(2), write(2), chmod(2) and so
24on are called from a process context. Filesystem locking is described
25in the document Documentation/filesystems/Locking.
26
27
28Directory Entry Cache (dcache)
29------------------------------
30
31The VFS implements the open(2), stat(2), chmod(2), and similar system
32calls. The pathname argument that is passed to them is used by the VFS
33to search through the directory entry cache (also known as the dentry
34cache or dcache). This provides a very fast look-up mechanism to
35translate a pathname (filename) into a specific dentry. Dentries live
36in RAM and are never saved to disc: they exist only for performance.
37
38The dentry cache is meant to be a view into your entire filespace. As
39most computers cannot fit all dentries in the RAM at the same time,
40some bits of the cache are missing. In order to resolve your pathname
41into a dentry, the VFS may have to resort to creating dentries along
42the way, and then loading the inode. This is done by looking up the
43inode.
44
45
46The Inode Object
47----------------
48
49An individual dentry usually has a pointer to an inode. Inodes are
50filesystem objects such as regular files, directories, FIFOs and other
51beasts. They live either on the disc (for block device filesystems)
52or in the memory (for pseudo filesystems). Inodes that live on the
53disc are copied into the memory when required and changes to the inode
54are written back to disc. A single inode can be pointed to by multiple
55dentries (hard links, for example, do this).
56
57To look up an inode requires that the VFS calls the lookup() method of
58the parent directory inode. This method is installed by the specific
59filesystem implementation that the inode lives in. Once the VFS has
60the required dentry (and hence the inode), we can do all those boring
61things like open(2) the file, or stat(2) it to peek at the inode
62data. The stat(2) operation is fairly simple: once the VFS has the
63dentry, it peeks at the inode data and passes some of it back to
64userspace.
65
66
67The File Object
68---------------
69
70Opening a file requires another operation: allocation of a file
71structure (this is the kernel-side implementation of file
72descriptors). The freshly allocated file structure is initialized with
73a pointer to the dentry and a set of file operation member functions.
74These are taken from the inode data. The open() file method is then
75called so the specific filesystem implementation can do its work. You
76can see that this is another switch performed by the VFS. The file
77structure is placed into the file descriptor table for the process.
78
79Reading, writing and closing files (and other assorted VFS operations)
80is done by using the userspace file descriptor to grab the appropriate
81file structure, and then calling the required file structure method to
82do whatever is required. For as long as the file is open, it keeps the
83dentry in use, which in turn means that the VFS inode is still in use.
84
85
86Registering and Mounting a Filesystem
87=====================================
88
89To register and unregister a filesystem, use the following API
90functions:
91
92 #include <linux/fs.h>
93
94 extern int register_filesystem(struct file_system_type *);
95 extern int unregister_filesystem(struct file_system_type *);
96
97The passed struct file_system_type describes your filesystem. When a
98request is made to mount a filesystem onto a directory in your namespace,
99the VFS will call the appropriate mount() method for the specific
100filesystem. New vfsmount referring to the tree returned by ->mount()
101will be attached to the mountpoint, so that when pathname resolution
102reaches the mountpoint it will jump into the root of that vfsmount.
103
104You can see all filesystems that are registered to the kernel in the
105file /proc/filesystems.
106
107
108struct file_system_type
109-----------------------
110
111This describes the filesystem. As of kernel 2.6.39, the following
112members are defined:
113
114struct file_system_type {
115 const char *name;
116 int fs_flags;
117 struct dentry *(*mount) (struct file_system_type *, int,
118 const char *, void *);
119 void (*kill_sb) (struct super_block *);
120 struct module *owner;
121 struct file_system_type * next;
122 struct list_head fs_supers;
123 struct lock_class_key s_lock_key;
124 struct lock_class_key s_umount_key;
125};
126
127 name: the name of the filesystem type, such as "ext2", "iso9660",
128 "msdos" and so on
129
130 fs_flags: various flags (i.e. FS_REQUIRES_DEV, FS_NO_DCACHE, etc.)
131
132 mount: the method to call when a new instance of this
133 filesystem should be mounted
134
135 kill_sb: the method to call when an instance of this filesystem
136 should be shut down
137
138 owner: for internal VFS use: you should initialize this to THIS_MODULE in
139 most cases.
140
141 next: for internal VFS use: you should initialize this to NULL
142
143 s_lock_key, s_umount_key: lockdep-specific
144
145The mount() method has the following arguments:
146
147 struct file_system_type *fs_type: describes the filesystem, partly initialized
148 by the specific filesystem code
149
150 int flags: mount flags
151
152 const char *dev_name: the device name we are mounting.
153
154 void *data: arbitrary mount options, usually comes as an ASCII
155 string (see "Mount Options" section)
156
157The mount() method must return the root dentry of the tree requested by
158caller. An active reference to its superblock must be grabbed and the
159superblock must be locked. On failure it should return ERR_PTR(error).
160
161The arguments match those of mount(2) and their interpretation
162depends on filesystem type. E.g. for block filesystems, dev_name is
163interpreted as block device name, that device is opened and if it
164contains a suitable filesystem image the method creates and initializes
165struct super_block accordingly, returning its root dentry to caller.
166
167->mount() may choose to return a subtree of existing filesystem - it
168doesn't have to create a new one. The main result from the caller's
169point of view is a reference to dentry at the root of (sub)tree to
170be attached; creation of new superblock is a common side effect.
171
172The most interesting member of the superblock structure that the
173mount() method fills in is the "s_op" field. This is a pointer to
174a "struct super_operations" which describes the next level of the
175filesystem implementation.
176
177Usually, a filesystem uses one of the generic mount() implementations
178and provides a fill_super() callback instead. The generic variants are:
179
180 mount_bdev: mount a filesystem residing on a block device
181
182 mount_nodev: mount a filesystem that is not backed by a device
183
184 mount_single: mount a filesystem which shares the instance between
185 all mounts
186
187A fill_super() callback implementation has the following arguments:
188
189 struct super_block *sb: the superblock structure. The callback
190 must initialize this properly.
191
192 void *data: arbitrary mount options, usually comes as an ASCII
193 string (see "Mount Options" section)
194
195 int silent: whether or not to be silent on error
196
197
198The Superblock Object
199=====================
200
201A superblock object represents a mounted filesystem.
202
203
204struct super_operations
205-----------------------
206
207This describes how the VFS can manipulate the superblock of your
208filesystem. As of kernel 2.6.22, the following members are defined:
209
210struct super_operations {
211 struct inode *(*alloc_inode)(struct super_block *sb);
212 void (*destroy_inode)(struct inode *);
213
214 void (*dirty_inode) (struct inode *, int flags);
215 int (*write_inode) (struct inode *, int);
216 void (*drop_inode) (struct inode *);
217 void (*delete_inode) (struct inode *);
218 void (*put_super) (struct super_block *);
219 int (*sync_fs)(struct super_block *sb, int wait);
220 int (*freeze_fs) (struct super_block *);
221 int (*unfreeze_fs) (struct super_block *);
222 int (*statfs) (struct dentry *, struct kstatfs *);
223 int (*remount_fs) (struct super_block *, int *, char *);
224 void (*clear_inode) (struct inode *);
225 void (*umount_begin) (struct super_block *);
226
227 int (*show_options)(struct seq_file *, struct dentry *);
228
229 ssize_t (*quota_read)(struct super_block *, int, char *, size_t, loff_t);
230 ssize_t (*quota_write)(struct super_block *, int, const char *, size_t, loff_t);
231 int (*nr_cached_objects)(struct super_block *);
232 void (*free_cached_objects)(struct super_block *, int);
233};
234
235All methods are called without any locks being held, unless otherwise
236noted. This means that most methods can block safely. All methods are
237only called from a process context (i.e. not from an interrupt handler
238or bottom half).
239
240 alloc_inode: this method is called by alloc_inode() to allocate memory
241 for struct inode and initialize it. If this function is not
242 defined, a simple 'struct inode' is allocated. Normally
243 alloc_inode will be used to allocate a larger structure which
244 contains a 'struct inode' embedded within it.
245
246 destroy_inode: this method is called by destroy_inode() to release
247 resources allocated for struct inode. It is only required if
248 ->alloc_inode was defined and simply undoes anything done by
249 ->alloc_inode.
250
251 dirty_inode: this method is called by the VFS to mark an inode dirty.
252
253 write_inode: this method is called when the VFS needs to write an
254 inode to disc. The second parameter indicates whether the write
255 should be synchronous or not, not all filesystems check this flag.
256
257 drop_inode: called when the last access to the inode is dropped,
258 with the inode->i_lock spinlock held.
259
260 This method should be either NULL (normal UNIX filesystem
261 semantics) or "generic_delete_inode" (for filesystems that do not
262 want to cache inodes - causing "delete_inode" to always be
263 called regardless of the value of i_nlink)
264
265 The "generic_delete_inode()" behavior is equivalent to the
266 old practice of using "force_delete" in the put_inode() case,
267 but does not have the races that the "force_delete()" approach
268 had.
269
270 delete_inode: called when the VFS wants to delete an inode
271
272 put_super: called when the VFS wishes to free the superblock
273 (i.e. unmount). This is called with the superblock lock held
274
275 sync_fs: called when VFS is writing out all dirty data associated with
276 a superblock. The second parameter indicates whether the method
277 should wait until the write out has been completed. Optional.
278
279 freeze_fs: called when VFS is locking a filesystem and
280 forcing it into a consistent state. This method is currently
281 used by the Logical Volume Manager (LVM).
282
283 unfreeze_fs: called when VFS is unlocking a filesystem and making it writable
284 again.
285
286 statfs: called when the VFS needs to get filesystem statistics.
287
288 remount_fs: called when the filesystem is remounted. This is called
289 with the kernel lock held
290
291 clear_inode: called then the VFS clears the inode. Optional
292
293 umount_begin: called when the VFS is unmounting a filesystem.
294
295 show_options: called by the VFS to show mount options for
296 /proc/<pid>/mounts. (see "Mount Options" section)
297
298 quota_read: called by the VFS to read from filesystem quota file.
299
300 quota_write: called by the VFS to write to filesystem quota file.
301
302 nr_cached_objects: called by the sb cache shrinking function for the
303 filesystem to return the number of freeable cached objects it contains.
304 Optional.
305
306 free_cache_objects: called by the sb cache shrinking function for the
307 filesystem to scan the number of objects indicated to try to free them.
308 Optional, but any filesystem implementing this method needs to also
309 implement ->nr_cached_objects for it to be called correctly.
310
311 We can't do anything with any errors that the filesystem might
312 encountered, hence the void return type. This will never be called if
313 the VM is trying to reclaim under GFP_NOFS conditions, hence this
314 method does not need to handle that situation itself.
315
316 Implementations must include conditional reschedule calls inside any
317 scanning loop that is done. This allows the VFS to determine
318 appropriate scan batch sizes without having to worry about whether
319 implementations will cause holdoff problems due to large scan batch
320 sizes.
321
322Whoever sets up the inode is responsible for filling in the "i_op" field. This
323is a pointer to a "struct inode_operations" which describes the methods that
324can be performed on individual inodes.
325
326struct xattr_handlers
327---------------------
328
329On filesystems that support extended attributes (xattrs), the s_xattr
330superblock field points to a NULL-terminated array of xattr handlers. Extended
331attributes are name:value pairs.
332
333 name: Indicates that the handler matches attributes with the specified name
334 (such as "system.posix_acl_access"); the prefix field must be NULL.
335
336 prefix: Indicates that the handler matches all attributes with the specified
337 name prefix (such as "user."); the name field must be NULL.
338
339 list: Determine if attributes matching this xattr handler should be listed
340 for a particular dentry. Used by some listxattr implementations like
341 generic_listxattr.
342
343 get: Called by the VFS to get the value of a particular extended attribute.
344 This method is called by the getxattr(2) system call.
345
346 set: Called by the VFS to set the value of a particular extended attribute.
347 When the new value is NULL, called to remove a particular extended
348 attribute. This method is called by the the setxattr(2) and
349 removexattr(2) system calls.
350
351When none of the xattr handlers of a filesystem match the specified attribute
352name or when a filesystem doesn't support extended attributes, the various
353*xattr(2) system calls return -EOPNOTSUPP.
354
355
356The Inode Object
357================
358
359An inode object represents an object within the filesystem.
360
361
362struct inode_operations
363-----------------------
364
365This describes how the VFS can manipulate an inode in your
366filesystem. As of kernel 2.6.22, the following members are defined:
367
368struct inode_operations {
369 int (*create) (struct inode *,struct dentry *, umode_t, bool);
370 struct dentry * (*lookup) (struct inode *,struct dentry *, unsigned int);
371 int (*link) (struct dentry *,struct inode *,struct dentry *);
372 int (*unlink) (struct inode *,struct dentry *);
373 int (*symlink) (struct inode *,struct dentry *,const char *);
374 int (*mkdir) (struct inode *,struct dentry *,umode_t);
375 int (*rmdir) (struct inode *,struct dentry *);
376 int (*mknod) (struct inode *,struct dentry *,umode_t,dev_t);
377 int (*rename) (struct inode *, struct dentry *,
378 struct inode *, struct dentry *, unsigned int);
379 int (*readlink) (struct dentry *, char __user *,int);
380 const char *(*get_link) (struct dentry *, struct inode *,
381 struct delayed_call *);
382 int (*permission) (struct inode *, int);
383 int (*get_acl)(struct inode *, int);
384 int (*setattr) (struct dentry *, struct iattr *);
385 int (*getattr) (const struct path *, struct kstat *, u32, unsigned int);
386 ssize_t (*listxattr) (struct dentry *, char *, size_t);
387 void (*update_time)(struct inode *, struct timespec *, int);
388 int (*atomic_open)(struct inode *, struct dentry *, struct file *,
389 unsigned open_flag, umode_t create_mode, int *opened);
390 int (*tmpfile) (struct inode *, struct dentry *, umode_t);
391};
392
393Again, all methods are called without any locks being held, unless
394otherwise noted.
395
396 create: called by the open(2) and creat(2) system calls. Only
397 required if you want to support regular files. The dentry you
398 get should not have an inode (i.e. it should be a negative
399 dentry). Here you will probably call d_instantiate() with the
400 dentry and the newly created inode
401
402 lookup: called when the VFS needs to look up an inode in a parent
403 directory. The name to look for is found in the dentry. This
404 method must call d_add() to insert the found inode into the
405 dentry. The "i_count" field in the inode structure should be
406 incremented. If the named inode does not exist a NULL inode
407 should be inserted into the dentry (this is called a negative
408 dentry). Returning an error code from this routine must only
409 be done on a real error, otherwise creating inodes with system
410 calls like create(2), mknod(2), mkdir(2) and so on will fail.
411 If you wish to overload the dentry methods then you should
412 initialise the "d_dop" field in the dentry; this is a pointer
413 to a struct "dentry_operations".
414 This method is called with the directory inode semaphore held
415
416 link: called by the link(2) system call. Only required if you want
417 to support hard links. You will probably need to call
418 d_instantiate() just as you would in the create() method
419
420 unlink: called by the unlink(2) system call. Only required if you
421 want to support deleting inodes
422
423 symlink: called by the symlink(2) system call. Only required if you
424 want to support symlinks. You will probably need to call
425 d_instantiate() just as you would in the create() method
426
427 mkdir: called by the mkdir(2) system call. Only required if you want
428 to support creating subdirectories. You will probably need to
429 call d_instantiate() just as you would in the create() method
430
431 rmdir: called by the rmdir(2) system call. Only required if you want
432 to support deleting subdirectories
433
434 mknod: called by the mknod(2) system call to create a device (char,
435 block) inode or a named pipe (FIFO) or socket. Only required
436 if you want to support creating these types of inodes. You
437 will probably need to call d_instantiate() just as you would
438 in the create() method
439
440 rename: called by the rename(2) system call to rename the object to
441 have the parent and name given by the second inode and dentry.
442
443 The filesystem must return -EINVAL for any unsupported or
444 unknown flags. Currently the following flags are implemented:
445 (1) RENAME_NOREPLACE: this flag indicates that if the target
446 of the rename exists the rename should fail with -EEXIST
447 instead of replacing the target. The VFS already checks for
448 existence, so for local filesystems the RENAME_NOREPLACE
449 implementation is equivalent to plain rename.
450 (2) RENAME_EXCHANGE: exchange source and target. Both must
451 exist; this is checked by the VFS. Unlike plain rename,
452 source and target may be of different type.
453
454 get_link: called by the VFS to follow a symbolic link to the
455 inode it points to. Only required if you want to support
456 symbolic links. This method returns the symlink body
457 to traverse (and possibly resets the current position with
458 nd_jump_link()). If the body won't go away until the inode
459 is gone, nothing else is needed; if it needs to be otherwise
460 pinned, arrange for its release by having get_link(..., ..., done)
461 do set_delayed_call(done, destructor, argument).
462 In that case destructor(argument) will be called once VFS is
463 done with the body you've returned.
464 May be called in RCU mode; that is indicated by NULL dentry
465 argument. If request can't be handled without leaving RCU mode,
466 have it return ERR_PTR(-ECHILD).
467
468 readlink: this is now just an override for use by readlink(2) for the
469 cases when ->get_link uses nd_jump_link() or object is not in
470 fact a symlink. Normally filesystems should only implement
471 ->get_link for symlinks and readlink(2) will automatically use
472 that.
473
474 permission: called by the VFS to check for access rights on a POSIX-like
475 filesystem.
476
477 May be called in rcu-walk mode (mask & MAY_NOT_BLOCK). If in rcu-walk
478 mode, the filesystem must check the permission without blocking or
479 storing to the inode.
480
481 If a situation is encountered that rcu-walk cannot handle, return
482 -ECHILD and it will be called again in ref-walk mode.
483
484 setattr: called by the VFS to set attributes for a file. This method
485 is called by chmod(2) and related system calls.
486
487 getattr: called by the VFS to get attributes of a file. This method
488 is called by stat(2) and related system calls.
489
490 listxattr: called by the VFS to list all extended attributes for a
491 given file. This method is called by the listxattr(2) system call.
492
493 update_time: called by the VFS to update a specific time or the i_version of
494 an inode. If this is not defined the VFS will update the inode itself
495 and call mark_inode_dirty_sync.
496
497 atomic_open: called on the last component of an open. Using this optional
498 method the filesystem can look up, possibly create and open the file in
499 one atomic operation. If it cannot perform this (e.g. the file type
500 turned out to be wrong) it may signal this by returning 1 instead of
501 usual 0 or -ve . This method is only called if the last component is
502 negative or needs lookup. Cached positive dentries are still handled by
503 f_op->open(). If the file was created, the FILE_CREATED flag should be
504 set in "opened". In case of O_EXCL the method must only succeed if the
505 file didn't exist and hence FILE_CREATED shall always be set on success.
506
507 tmpfile: called in the end of O_TMPFILE open(). Optional, equivalent to
508 atomically creating, opening and unlinking a file in given directory.
509
510The Address Space Object
511========================
512
513The address space object is used to group and manage pages in the page
514cache. It can be used to keep track of the pages in a file (or
515anything else) and also track the mapping of sections of the file into
516process address spaces.
517
518There are a number of distinct yet related services that an
519address-space can provide. These include communicating memory
520pressure, page lookup by address, and keeping track of pages tagged as
521Dirty or Writeback.
522
523The first can be used independently to the others. The VM can try to
524either write dirty pages in order to clean them, or release clean
525pages in order to reuse them. To do this it can call the ->writepage
526method on dirty pages, and ->releasepage on clean pages with
527PagePrivate set. Clean pages without PagePrivate and with no external
528references will be released without notice being given to the
529address_space.
530
531To achieve this functionality, pages need to be placed on an LRU with
532lru_cache_add and mark_page_active needs to be called whenever the
533page is used.
534
535Pages are normally kept in a radix tree index by ->index. This tree
536maintains information about the PG_Dirty and PG_Writeback status of
537each page, so that pages with either of these flags can be found
538quickly.
539
540The Dirty tag is primarily used by mpage_writepages - the default
541->writepages method. It uses the tag to find dirty pages to call
542->writepage on. If mpage_writepages is not used (i.e. the address
543provides its own ->writepages) , the PAGECACHE_TAG_DIRTY tag is
544almost unused. write_inode_now and sync_inode do use it (through
545__sync_single_inode) to check if ->writepages has been successful in
546writing out the whole address_space.
547
548The Writeback tag is used by filemap*wait* and sync_page* functions,
549via filemap_fdatawait_range, to wait for all writeback to complete.
550
551An address_space handler may attach extra information to a page,
552typically using the 'private' field in the 'struct page'. If such
553information is attached, the PG_Private flag should be set. This will
554cause various VM routines to make extra calls into the address_space
555handler to deal with that data.
556
557An address space acts as an intermediate between storage and
558application. Data is read into the address space a whole page at a
559time, and provided to the application either by copying of the page,
560or by memory-mapping the page.
561Data is written into the address space by the application, and then
562written-back to storage typically in whole pages, however the
563address_space has finer control of write sizes.
564
565The read process essentially only requires 'readpage'. The write
566process is more complicated and uses write_begin/write_end or
567set_page_dirty to write data into the address_space, and writepage
568and writepages to writeback data to storage.
569
570Adding and removing pages to/from an address_space is protected by the
571inode's i_mutex.
572
573When data is written to a page, the PG_Dirty flag should be set. It
574typically remains set until writepage asks for it to be written. This
575should clear PG_Dirty and set PG_Writeback. It can be actually
576written at any point after PG_Dirty is clear. Once it is known to be
577safe, PG_Writeback is cleared.
578
579Writeback makes use of a writeback_control structure to direct the
580operations. This gives the the writepage and writepages operations some
581information about the nature of and reason for the writeback request,
582and the constraints under which it is being done. It is also used to
583return information back to the caller about the result of a writepage or
584writepages request.
585
586Handling errors during writeback
587--------------------------------
588Most applications that do buffered I/O will periodically call a file
589synchronization call (fsync, fdatasync, msync or sync_file_range) to
590ensure that data written has made it to the backing store. When there
591is an error during writeback, they expect that error to be reported when
592a file sync request is made. After an error has been reported on one
593request, subsequent requests on the same file descriptor should return
5940, unless further writeback errors have occurred since the previous file
595syncronization.
596
597Ideally, the kernel would report errors only on file descriptions on
598which writes were done that subsequently failed to be written back. The
599generic pagecache infrastructure does not track the file descriptions
600that have dirtied each individual page however, so determining which
601file descriptors should get back an error is not possible.
602
603Instead, the generic writeback error tracking infrastructure in the
604kernel settles for reporting errors to fsync on all file descriptions
605that were open at the time that the error occurred. In a situation with
606multiple writers, all of them will get back an error on a subsequent fsync,
607even if all of the writes done through that particular file descriptor
608succeeded (or even if there were no writes on that file descriptor at all).
609
610Filesystems that wish to use this infrastructure should call
611mapping_set_error to record the error in the address_space when it
612occurs. Then, after writing back data from the pagecache in their
613file->fsync operation, they should call file_check_and_advance_wb_err to
614ensure that the struct file's error cursor has advanced to the correct
615point in the stream of errors emitted by the backing device(s).
616
617struct address_space_operations
618-------------------------------
619
620This describes how the VFS can manipulate mapping of a file to page cache in
621your filesystem. The following members are defined:
622
623struct address_space_operations {
624 int (*writepage)(struct page *page, struct writeback_control *wbc);
625 int (*readpage)(struct file *, struct page *);
626 int (*writepages)(struct address_space *, struct writeback_control *);
627 int (*set_page_dirty)(struct page *page);
628 int (*readpages)(struct file *filp, struct address_space *mapping,
629 struct list_head *pages, unsigned nr_pages);
630 int (*write_begin)(struct file *, struct address_space *mapping,
631 loff_t pos, unsigned len, unsigned flags,
632 struct page **pagep, void **fsdata);
633 int (*write_end)(struct file *, struct address_space *mapping,
634 loff_t pos, unsigned len, unsigned copied,
635 struct page *page, void *fsdata);
636 sector_t (*bmap)(struct address_space *, sector_t);
637 void (*invalidatepage) (struct page *, unsigned int, unsigned int);
638 int (*releasepage) (struct page *, int);
639 void (*freepage)(struct page *);
640 ssize_t (*direct_IO)(struct kiocb *, struct iov_iter *iter);
641 /* isolate a page for migration */
642 bool (*isolate_page) (struct page *, isolate_mode_t);
643 /* migrate the contents of a page to the specified target */
644 int (*migratepage) (struct page *, struct page *);
645 /* put migration-failed page back to right list */
646 void (*putback_page) (struct page *);
647 int (*launder_page) (struct page *);
648
649 int (*is_partially_uptodate) (struct page *, unsigned long,
650 unsigned long);
651 void (*is_dirty_writeback) (struct page *, bool *, bool *);
652 int (*error_remove_page) (struct mapping *mapping, struct page *page);
653 int (*swap_activate)(struct file *);
654 int (*swap_deactivate)(struct file *);
655};
656
657 writepage: called by the VM to write a dirty page to backing store.
658 This may happen for data integrity reasons (i.e. 'sync'), or
659 to free up memory (flush). The difference can be seen in
660 wbc->sync_mode.
661 The PG_Dirty flag has been cleared and PageLocked is true.
662 writepage should start writeout, should set PG_Writeback,
663 and should make sure the page is unlocked, either synchronously
664 or asynchronously when the write operation completes.
665
666 If wbc->sync_mode is WB_SYNC_NONE, ->writepage doesn't have to
667 try too hard if there are problems, and may choose to write out
668 other pages from the mapping if that is easier (e.g. due to
669 internal dependencies). If it chooses not to start writeout, it
670 should return AOP_WRITEPAGE_ACTIVATE so that the VM will not keep
671 calling ->writepage on that page.
672
673 See the file "Locking" for more details.
674
675 readpage: called by the VM to read a page from backing store.
676 The page will be Locked when readpage is called, and should be
677 unlocked and marked uptodate once the read completes.
678 If ->readpage discovers that it needs to unlock the page for
679 some reason, it can do so, and then return AOP_TRUNCATED_PAGE.
680 In this case, the page will be relocated, relocked and if
681 that all succeeds, ->readpage will be called again.
682
683 writepages: called by the VM to write out pages associated with the
684 address_space object. If wbc->sync_mode is WBC_SYNC_ALL, then
685 the writeback_control will specify a range of pages that must be
686 written out. If it is WBC_SYNC_NONE, then a nr_to_write is given
687 and that many pages should be written if possible.
688 If no ->writepages is given, then mpage_writepages is used
689 instead. This will choose pages from the address space that are
690 tagged as DIRTY and will pass them to ->writepage.
691
692 set_page_dirty: called by the VM to set a page dirty.
693 This is particularly needed if an address space attaches
694 private data to a page, and that data needs to be updated when
695 a page is dirtied. This is called, for example, when a memory
696 mapped page gets modified.
697 If defined, it should set the PageDirty flag, and the
698 PAGECACHE_TAG_DIRTY tag in the radix tree.
699
700 readpages: called by the VM to read pages associated with the address_space
701 object. This is essentially just a vector version of
702 readpage. Instead of just one page, several pages are
703 requested.
704 readpages is only used for read-ahead, so read errors are
705 ignored. If anything goes wrong, feel free to give up.
706
707 write_begin:
708 Called by the generic buffered write code to ask the filesystem to
709 prepare to write len bytes at the given offset in the file. The
710 address_space should check that the write will be able to complete,
711 by allocating space if necessary and doing any other internal
712 housekeeping. If the write will update parts of any basic-blocks on
713 storage, then those blocks should be pre-read (if they haven't been
714 read already) so that the updated blocks can be written out properly.
715
716 The filesystem must return the locked pagecache page for the specified
717 offset, in *pagep, for the caller to write into.
718
719 It must be able to cope with short writes (where the length passed to
720 write_begin is greater than the number of bytes copied into the page).
721
722 flags is a field for AOP_FLAG_xxx flags, described in
723 include/linux/fs.h.
724
725 A void * may be returned in fsdata, which then gets passed into
726 write_end.
727
728 Returns 0 on success; < 0 on failure (which is the error code), in
729 which case write_end is not called.
730
731 write_end: After a successful write_begin, and data copy, write_end must
732 be called. len is the original len passed to write_begin, and copied
733 is the amount that was able to be copied.
734
735 The filesystem must take care of unlocking the page and releasing it
736 refcount, and updating i_size.
737
738 Returns < 0 on failure, otherwise the number of bytes (<= 'copied')
739 that were able to be copied into pagecache.
740
741 bmap: called by the VFS to map a logical block offset within object to
742 physical block number. This method is used by the FIBMAP
743 ioctl and for working with swap-files. To be able to swap to
744 a file, the file must have a stable mapping to a block
745 device. The swap system does not go through the filesystem
746 but instead uses bmap to find out where the blocks in the file
747 are and uses those addresses directly.
748
749 invalidatepage: If a page has PagePrivate set, then invalidatepage
750 will be called when part or all of the page is to be removed
751 from the address space. This generally corresponds to either a
752 truncation, punch hole or a complete invalidation of the address
753 space (in the latter case 'offset' will always be 0 and 'length'
754 will be PAGE_SIZE). Any private data associated with the page
755 should be updated to reflect this truncation. If offset is 0 and
756 length is PAGE_SIZE, then the private data should be released,
757 because the page must be able to be completely discarded. This may
758 be done by calling the ->releasepage function, but in this case the
759 release MUST succeed.
760
761 releasepage: releasepage is called on PagePrivate pages to indicate
762 that the page should be freed if possible. ->releasepage
763 should remove any private data from the page and clear the
764 PagePrivate flag. If releasepage() fails for some reason, it must
765 indicate failure with a 0 return value.
766 releasepage() is used in two distinct though related cases. The
767 first is when the VM finds a clean page with no active users and
768 wants to make it a free page. If ->releasepage succeeds, the
769 page will be removed from the address_space and become free.
770
771 The second case is when a request has been made to invalidate
772 some or all pages in an address_space. This can happen
773 through the fadvise(POSIX_FADV_DONTNEED) system call or by the
774 filesystem explicitly requesting it as nfs and 9fs do (when
775 they believe the cache may be out of date with storage) by
776 calling invalidate_inode_pages2().
777 If the filesystem makes such a call, and needs to be certain
778 that all pages are invalidated, then its releasepage will
779 need to ensure this. Possibly it can clear the PageUptodate
780 bit if it cannot free private data yet.
781
782 freepage: freepage is called once the page is no longer visible in
783 the page cache in order to allow the cleanup of any private
784 data. Since it may be called by the memory reclaimer, it
785 should not assume that the original address_space mapping still
786 exists, and it should not block.
787
788 direct_IO: called by the generic read/write routines to perform
789 direct_IO - that is IO requests which bypass the page cache
790 and transfer data directly between the storage and the
791 application's address space.
792
793 isolate_page: Called by the VM when isolating a movable non-lru page.
794 If page is successfully isolated, VM marks the page as PG_isolated
795 via __SetPageIsolated.
796
797 migrate_page: This is used to compact the physical memory usage.
798 If the VM wants to relocate a page (maybe off a memory card
799 that is signalling imminent failure) it will pass a new page
800 and an old page to this function. migrate_page should
801 transfer any private data across and update any references
802 that it has to the page.
803
804 putback_page: Called by the VM when isolated page's migration fails.
805
806 launder_page: Called before freeing a page - it writes back the dirty page. To
807 prevent redirtying the page, it is kept locked during the whole
808 operation.
809
810 is_partially_uptodate: Called by the VM when reading a file through the
811 pagecache when the underlying blocksize != pagesize. If the required
812 block is up to date then the read can complete without needing the IO
813 to bring the whole page up to date.
814
815 is_dirty_writeback: Called by the VM when attempting to reclaim a page.
816 The VM uses dirty and writeback information to determine if it needs
817 to stall to allow flushers a chance to complete some IO. Ordinarily
818 it can use PageDirty and PageWriteback but some filesystems have
819 more complex state (unstable pages in NFS prevent reclaim) or
820 do not set those flags due to locking problems. This callback
821 allows a filesystem to indicate to the VM if a page should be
822 treated as dirty or writeback for the purposes of stalling.
823
824 error_remove_page: normally set to generic_error_remove_page if truncation
825 is ok for this address space. Used for memory failure handling.
826 Setting this implies you deal with pages going away under you,
827 unless you have them locked or reference counts increased.
828
829 swap_activate: Called when swapon is used on a file to allocate
830 space if necessary and pin the block lookup information in
831 memory. A return value of zero indicates success,
832 in which case this file can be used to back swapspace.
833
834 swap_deactivate: Called during swapoff on files where swap_activate
835 was successful.
836
837
838The File Object
839===============
840
841A file object represents a file opened by a process. This is also known
842as an "open file description" in POSIX parlance.
843
844
845struct file_operations
846----------------------
847
848This describes how the VFS can manipulate an open file. As of kernel
8494.1, the following members are defined:
850
851struct file_operations {
852 struct module *owner;
853 loff_t (*llseek) (struct file *, loff_t, int);
854 ssize_t (*read) (struct file *, char __user *, size_t, loff_t *);
855 ssize_t (*write) (struct file *, const char __user *, size_t, loff_t *);
856 ssize_t (*read_iter) (struct kiocb *, struct iov_iter *);
857 ssize_t (*write_iter) (struct kiocb *, struct iov_iter *);
858 int (*iterate) (struct file *, struct dir_context *);
859 unsigned int (*poll) (struct file *, struct poll_table_struct *);
860 long (*unlocked_ioctl) (struct file *, unsigned int, unsigned long);
861 long (*compat_ioctl) (struct file *, unsigned int, unsigned long);
862 int (*mmap) (struct file *, struct vm_area_struct *);
863 int (*mremap)(struct file *, struct vm_area_struct *);
864 int (*open) (struct inode *, struct file *);
865 int (*flush) (struct file *, fl_owner_t id);
866 int (*release) (struct inode *, struct file *);
867 int (*fsync) (struct file *, loff_t, loff_t, int datasync);
868 int (*fasync) (int, struct file *, int);
869 int (*lock) (struct file *, int, struct file_lock *);
870 ssize_t (*sendpage) (struct file *, struct page *, int, size_t, loff_t *, int);
871 unsigned long (*get_unmapped_area)(struct file *, unsigned long, unsigned long, unsigned long, unsigned long);
872 int (*check_flags)(int);
873 int (*flock) (struct file *, int, struct file_lock *);
874 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *, loff_t *, size_t, unsigned int);
875 ssize_t (*splice_read)(struct file *, loff_t *, struct pipe_inode_info *, size_t, unsigned int);
876 int (*setlease)(struct file *, long, struct file_lock **, void **);
877 long (*fallocate)(struct file *file, int mode, loff_t offset,
878 loff_t len);
879 void (*show_fdinfo)(struct seq_file *m, struct file *f);
880#ifndef CONFIG_MMU
881 unsigned (*mmap_capabilities)(struct file *);
882#endif
883};
884
885Again, all methods are called without any locks being held, unless
886otherwise noted.
887
888 llseek: called when the VFS needs to move the file position index
889
890 read: called by read(2) and related system calls
891
892 read_iter: possibly asynchronous read with iov_iter as destination
893
894 write: called by write(2) and related system calls
895
896 write_iter: possibly asynchronous write with iov_iter as source
897
898 iterate: called when the VFS needs to read the directory contents
899
900 poll: called by the VFS when a process wants to check if there is
901 activity on this file and (optionally) go to sleep until there
902 is activity. Called by the select(2) and poll(2) system calls
903
904 unlocked_ioctl: called by the ioctl(2) system call.
905
906 compat_ioctl: called by the ioctl(2) system call when 32 bit system calls
907 are used on 64 bit kernels.
908
909 mmap: called by the mmap(2) system call
910
911 open: called by the VFS when an inode should be opened. When the VFS
912 opens a file, it creates a new "struct file". It then calls the
913 open method for the newly allocated file structure. You might
914 think that the open method really belongs in
915 "struct inode_operations", and you may be right. I think it's
916 done the way it is because it makes filesystems simpler to
917 implement. The open() method is a good place to initialize the
918 "private_data" member in the file structure if you want to point
919 to a device structure
920
921 flush: called by the close(2) system call to flush a file
922
923 release: called when the last reference to an open file is closed
924
925 fsync: called by the fsync(2) system call. Also see the section above
926 entitled "Handling errors during writeback".
927
928 fasync: called by the fcntl(2) system call when asynchronous
929 (non-blocking) mode is enabled for a file
930
931 lock: called by the fcntl(2) system call for F_GETLK, F_SETLK, and F_SETLKW
932 commands
933
934 get_unmapped_area: called by the mmap(2) system call
935
936 check_flags: called by the fcntl(2) system call for F_SETFL command
937
938 flock: called by the flock(2) system call
939
940 splice_write: called by the VFS to splice data from a pipe to a file. This
941 method is used by the splice(2) system call
942
943 splice_read: called by the VFS to splice data from file to a pipe. This
944 method is used by the splice(2) system call
945
946 setlease: called by the VFS to set or release a file lock lease. setlease
947 implementations should call generic_setlease to record or remove
948 the lease in the inode after setting it.
949
950 fallocate: called by the VFS to preallocate blocks or punch a hole.
951
952Note that the file operations are implemented by the specific
953filesystem in which the inode resides. When opening a device node
954(character or block special) most filesystems will call special
955support routines in the VFS which will locate the required device
956driver information. These support routines replace the filesystem file
957operations with those for the device driver, and then proceed to call
958the new open() method for the file. This is how opening a device file
959in the filesystem eventually ends up calling the device driver open()
960method.
961
962
963Directory Entry Cache (dcache)
964==============================
965
966
967struct dentry_operations
968------------------------
969
970This describes how a filesystem can overload the standard dentry
971operations. Dentries and the dcache are the domain of the VFS and the
972individual filesystem implementations. Device drivers have no business
973here. These methods may be set to NULL, as they are either optional or
974the VFS uses a default. As of kernel 2.6.22, the following members are
975defined:
976
977struct dentry_operations {
978 int (*d_revalidate)(struct dentry *, unsigned int);
979 int (*d_weak_revalidate)(struct dentry *, unsigned int);
980 int (*d_hash)(const struct dentry *, struct qstr *);
981 int (*d_compare)(const struct dentry *,
982 unsigned int, const char *, const struct qstr *);
983 int (*d_delete)(const struct dentry *);
984 int (*d_init)(struct dentry *);
985 void (*d_release)(struct dentry *);
986 void (*d_iput)(struct dentry *, struct inode *);
987 char *(*d_dname)(struct dentry *, char *, int);
988 struct vfsmount *(*d_automount)(struct path *);
989 int (*d_manage)(const struct path *, bool);
990 struct dentry *(*d_real)(struct dentry *, const struct inode *,
991 unsigned int, unsigned int);
992};
993
994 d_revalidate: called when the VFS needs to revalidate a dentry. This
995 is called whenever a name look-up finds a dentry in the
996 dcache. Most local filesystems leave this as NULL, because all their
997 dentries in the dcache are valid. Network filesystems are different
998 since things can change on the server without the client necessarily
999 being aware of it.
1000
1001 This function should return a positive value if the dentry is still
1002 valid, and zero or a negative error code if it isn't.
1003
1004 d_revalidate may be called in rcu-walk mode (flags & LOOKUP_RCU).
1005 If in rcu-walk mode, the filesystem must revalidate the dentry without
1006 blocking or storing to the dentry, d_parent and d_inode should not be
1007 used without care (because they can change and, in d_inode case, even
1008 become NULL under us).
1009
1010 If a situation is encountered that rcu-walk cannot handle, return
1011 -ECHILD and it will be called again in ref-walk mode.
1012
1013 d_weak_revalidate: called when the VFS needs to revalidate a "jumped" dentry.
1014 This is called when a path-walk ends at dentry that was not acquired by
1015 doing a lookup in the parent directory. This includes "/", "." and "..",
1016 as well as procfs-style symlinks and mountpoint traversal.
1017
1018 In this case, we are less concerned with whether the dentry is still
1019 fully correct, but rather that the inode is still valid. As with
1020 d_revalidate, most local filesystems will set this to NULL since their
1021 dcache entries are always valid.
1022
1023 This function has the same return code semantics as d_revalidate.
1024
1025 d_weak_revalidate is only called after leaving rcu-walk mode.
1026
1027 d_hash: called when the VFS adds a dentry to the hash table. The first
1028 dentry passed to d_hash is the parent directory that the name is
1029 to be hashed into.
1030
1031 Same locking and synchronisation rules as d_compare regarding
1032 what is safe to dereference etc.
1033
1034 d_compare: called to compare a dentry name with a given name. The first
1035 dentry is the parent of the dentry to be compared, the second is
1036 the child dentry. len and name string are properties of the dentry
1037 to be compared. qstr is the name to compare it with.
1038
1039 Must be constant and idempotent, and should not take locks if
1040 possible, and should not or store into the dentry.
1041 Should not dereference pointers outside the dentry without
1042 lots of care (eg. d_parent, d_inode, d_name should not be used).
1043
1044 However, our vfsmount is pinned, and RCU held, so the dentries and
1045 inodes won't disappear, neither will our sb or filesystem module.
1046 ->d_sb may be used.
1047
1048 It is a tricky calling convention because it needs to be called under
1049 "rcu-walk", ie. without any locks or references on things.
1050
1051 d_delete: called when the last reference to a dentry is dropped and the
1052 dcache is deciding whether or not to cache it. Return 1 to delete
1053 immediately, or 0 to cache the dentry. Default is NULL which means to
1054 always cache a reachable dentry. d_delete must be constant and
1055 idempotent.
1056
1057 d_init: called when a dentry is allocated
1058
1059 d_release: called when a dentry is really deallocated
1060
1061 d_iput: called when a dentry loses its inode (just prior to its
1062 being deallocated). The default when this is NULL is that the
1063 VFS calls iput(). If you define this method, you must call
1064 iput() yourself
1065
1066 d_dname: called when the pathname of a dentry should be generated.
1067 Useful for some pseudo filesystems (sockfs, pipefs, ...) to delay
1068 pathname generation. (Instead of doing it when dentry is created,
1069 it's done only when the path is needed.). Real filesystems probably
1070 dont want to use it, because their dentries are present in global
1071 dcache hash, so their hash should be an invariant. As no lock is
1072 held, d_dname() should not try to modify the dentry itself, unless
1073 appropriate SMP safety is used. CAUTION : d_path() logic is quite
1074 tricky. The correct way to return for example "Hello" is to put it
1075 at the end of the buffer, and returns a pointer to the first char.
1076 dynamic_dname() helper function is provided to take care of this.
1077
1078 Example :
1079
1080 static char *pipefs_dname(struct dentry *dent, char *buffer, int buflen)
1081 {
1082 return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
1083 dentry->d_inode->i_ino);
1084 }
1085
1086 d_automount: called when an automount dentry is to be traversed (optional).
1087 This should create a new VFS mount record and return the record to the
1088 caller. The caller is supplied with a path parameter giving the
1089 automount directory to describe the automount target and the parent
1090 VFS mount record to provide inheritable mount parameters. NULL should
1091 be returned if someone else managed to make the automount first. If
1092 the vfsmount creation failed, then an error code should be returned.
1093 If -EISDIR is returned, then the directory will be treated as an
1094 ordinary directory and returned to pathwalk to continue walking.
1095
1096 If a vfsmount is returned, the caller will attempt to mount it on the
1097 mountpoint and will remove the vfsmount from its expiration list in
1098 the case of failure. The vfsmount should be returned with 2 refs on
1099 it to prevent automatic expiration - the caller will clean up the
1100 additional ref.
1101
1102 This function is only used if DCACHE_NEED_AUTOMOUNT is set on the
1103 dentry. This is set by __d_instantiate() if S_AUTOMOUNT is set on the
1104 inode being added.
1105
1106 d_manage: called to allow the filesystem to manage the transition from a
1107 dentry (optional). This allows autofs, for example, to hold up clients
1108 waiting to explore behind a 'mountpoint' whilst letting the daemon go
1109 past and construct the subtree there. 0 should be returned to let the
1110 calling process continue. -EISDIR can be returned to tell pathwalk to
1111 use this directory as an ordinary directory and to ignore anything
1112 mounted on it and not to check the automount flag. Any other error
1113 code will abort pathwalk completely.
1114
1115 If the 'rcu_walk' parameter is true, then the caller is doing a
1116 pathwalk in RCU-walk mode. Sleeping is not permitted in this mode,
1117 and the caller can be asked to leave it and call again by returning
1118 -ECHILD. -EISDIR may also be returned to tell pathwalk to
1119 ignore d_automount or any mounts.
1120
1121 This function is only used if DCACHE_MANAGE_TRANSIT is set on the
1122 dentry being transited from.
1123
1124 d_real: overlay/union type filesystems implement this method to return one of
1125 the underlying dentries hidden by the overlay. It is used in three
1126 different modes:
1127
1128 Called from open it may need to copy-up the file depending on the
1129 supplied open flags. This mode is selected with a non-zero flags
1130 argument. In this mode the d_real method can return an error.
1131
1132 Called from file_dentry() it returns the real dentry matching the inode
1133 argument. The real dentry may be from a lower layer already copied up,
1134 but still referenced from the file. This mode is selected with a
1135 non-NULL inode argument. This will always succeed.
1136
1137 With NULL inode and zero flags the topmost real underlying dentry is
1138 returned. This will always succeed.
1139
1140 This method is never called with both non-NULL inode and non-zero flags.
1141
1142Each dentry has a pointer to its parent dentry, as well as a hash list
1143of child dentries. Child dentries are basically like files in a
1144directory.
1145
1146
1147Directory Entry Cache API
1148--------------------------
1149
1150There are a number of functions defined which permit a filesystem to
1151manipulate dentries:
1152
1153 dget: open a new handle for an existing dentry (this just increments
1154 the usage count)
1155
1156 dput: close a handle for a dentry (decrements the usage count). If
1157 the usage count drops to 0, and the dentry is still in its
1158 parent's hash, the "d_delete" method is called to check whether
1159 it should be cached. If it should not be cached, or if the dentry
1160 is not hashed, it is deleted. Otherwise cached dentries are put
1161 into an LRU list to be reclaimed on memory shortage.
1162
1163 d_drop: this unhashes a dentry from its parents hash list. A
1164 subsequent call to dput() will deallocate the dentry if its
1165 usage count drops to 0
1166
1167 d_delete: delete a dentry. If there are no other open references to
1168 the dentry then the dentry is turned into a negative dentry
1169 (the d_iput() method is called). If there are other
1170 references, then d_drop() is called instead
1171
1172 d_add: add a dentry to its parents hash list and then calls
1173 d_instantiate()
1174
1175 d_instantiate: add a dentry to the alias hash list for the inode and
1176 updates the "d_inode" member. The "i_count" member in the
1177 inode structure should be set/incremented. If the inode
1178 pointer is NULL, the dentry is called a "negative
1179 dentry". This function is commonly called when an inode is
1180 created for an existing negative dentry
1181
1182 d_lookup: look up a dentry given its parent and path name component
1183 It looks up the child of that given name from the dcache
1184 hash table. If it is found, the reference count is incremented
1185 and the dentry is returned. The caller must use dput()
1186 to free the dentry when it finishes using it.
1187
1188Mount Options
1189=============
1190
1191Parsing options
1192---------------
1193
1194On mount and remount the filesystem is passed a string containing a
1195comma separated list of mount options. The options can have either of
1196these forms:
1197
1198 option
1199 option=value
1200
1201The <linux/parser.h> header defines an API that helps parse these
1202options. There are plenty of examples on how to use it in existing
1203filesystems.
1204
1205Showing options
1206---------------
1207
1208If a filesystem accepts mount options, it must define show_options()
1209to show all the currently active options. The rules are:
1210
1211 - options MUST be shown which are not default or their values differ
1212 from the default
1213
1214 - options MAY be shown which are enabled by default or have their
1215 default value
1216
1217Options used only internally between a mount helper and the kernel
1218(such as file descriptors), or which only have an effect during the
1219mounting (such as ones controlling the creation of a journal) are exempt
1220from the above rules.
1221
1222The underlying reason for the above rules is to make sure, that a
1223mount can be accurately replicated (e.g. umounting and mounting again)
1224based on the information found in /proc/mounts.
1225
1226Resources
1227=========
1228
1229(Note some of these resources are not up-to-date with the latest kernel
1230 version.)
1231
1232Creating Linux virtual filesystems. 2002
1233 <http://lwn.net/Articles/13325/>
1234
1235The Linux Virtual File-system Layer by Neil Brown. 1999
1236 <http://www.cse.unsw.edu.au/~neilb/oss/linux-commentary/vfs.html>
1237
1238A tour of the Linux VFS by Michael K. Johnson. 1996
1239 <http://www.tldp.org/LDP/khg/HyperNews/get/fs/vfstour.html>
1240
1241A small trail through the Linux kernel by Andries Brouwer. 2001
1242 <http://www.win.tue.nl/~aeb/linux/vfs/trail.html>