5 This write-up is based on three articles published at lwn.net:
7 - <https://lwn.net/Articles/649115/> Pathname lookup in Linux
8 - <https://lwn.net/Articles/649729/> RCU-walk: faster pathname lookup in Linux
9 - <https://lwn.net/Articles/650786/> A walk among the symlinks
11 Written by Neil Brown with help from Al Viro and Jon Corbet.
12 It has subsequently been updated to reflect changes in the kernel
15 - per-directory parallel name lookup.
17 Introduction to pathname lookup
18 ===============================
20 The most obvious aspect of pathname lookup, which very little
21 exploration is needed to discover, is that it is complex. There are
22 many rules, special cases, and implementation alternatives that all
23 combine to confuse the unwary reader. Computer science has long been
24 acquainted with such complexity and has tools to help manage it. One
25 tool that we will make extensive use of is "divide and conquer". For
26 the early parts of the analysis we will divide off symlinks - leaving
27 them until the final part. Well before we get to symlinks we have
28 another major division based on the VFS's approach to locking which
29 will allow us to review "REF-walk" and "RCU-walk" separately. But we
30 are getting ahead of ourselves. There are some important low level
31 distinctions we need to clarify first.
33 There are two sorts of ...
34 --------------------------
36 .. _openat: http://man7.org/linux/man-pages/man2/openat.2.html
38 Pathnames (sometimes "file names"), used to identify objects in the
39 filesystem, will be familiar to most readers. They contain two sorts
40 of elements: "slashes" that are sequences of one or more "``/``"
41 characters, and "components" that are sequences of one or more
42 non-"``/``" characters. These form two kinds of paths. Those that
43 start with slashes are "absolute" and start from the filesystem root.
44 The others are "relative" and start from the current directory, or
45 from some other location specified by a file descriptor given to a
46 "``XXXat``" system call such as `openat() <openat_>`_.
48 .. _execveat: http://man7.org/linux/man-pages/man2/execveat.2.html
50 It is tempting to describe the second kind as starting with a
51 component, but that isn't always accurate: a pathname can lack both
52 slashes and components, it can be empty, in other words. This is
53 generally forbidden in POSIX, but some of those "xxx``at``" system calls
54 in Linux permit it when the ``AT_EMPTY_PATH`` flag is given. For
55 example, if you have an open file descriptor on an executable file you
56 can execute it by calling `execveat() <execveat_>`_ passing
57 the file descriptor, an empty path, and the ``AT_EMPTY_PATH`` flag.
59 These paths can be divided into two sections: the final component and
60 everything else. The "everything else" is the easy bit. In all cases
61 it must identify a directory that already exists, otherwise an error
62 such as ``ENOENT`` or ``ENOTDIR`` will be reported.
64 The final component is not so simple. Not only do different system
65 calls interpret it quite differently (e.g. some create it, some do
66 not), but it might not even exist: neither the empty pathname nor the
67 pathname that is just slashes have a final component. If it does
68 exist, it could be "``.``" or "``..``" which are handled quite differently
69 from other components.
71 .. _POSIX: http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_12
73 If a pathname ends with a slash, such as "``/tmp/foo/``" it might be
74 tempting to consider that to have an empty final component. In many
75 ways that would lead to correct results, but not always. In
76 particular, ``mkdir()`` and ``rmdir()`` each create or remove a directory named
77 by the final component, and they are required to work with pathnames
78 ending in "``/``". According to POSIX_
80 A pathname that contains at least one non- <slash> character and
81 that ends with one or more trailing <slash> characters shall not
82 be resolved successfully unless the last pathname component before
83 the trailing <slash> characters names an existing directory or a
84 directory entry that is to be created for a directory immediately
85 after the pathname is resolved.
87 The Linux pathname walking code (mostly in ``fs/namei.c``) deals with
88 all of these issues: breaking the path into components, handling the
89 "everything else" quite separately from the final component, and
90 checking that the trailing slash is not used where it isn't
91 permitted. It also addresses the important issue of concurrent
94 While one process is looking up a pathname, another might be making
95 changes that affect that lookup. One fairly extreme case is that if
96 "a/b" were renamed to "a/c/b" while another process were looking up
97 "a/b/..", that process might successfully resolve on "a/c".
98 Most races are much more subtle, and a big part of the task of
99 pathname lookup is to prevent them from having damaging effects. Many
100 of the possible races are seen most clearly in the context of the
101 "dcache" and an understanding of that is central to understanding
104 More than just a cache
105 ----------------------
107 The "dcache" caches information about names in each filesystem to
108 make them quickly available for lookup. Each entry (known as a
109 "dentry") contains three significant fields: a component name, a
110 pointer to a parent dentry, and a pointer to the "inode" which
111 contains further information about the object in that parent with
112 the given name. The inode pointer can be ``NULL`` indicating that the
113 name doesn't exist in the parent. While there can be linkage in the
114 dentry of a directory to the dentries of the children, that linkage is
115 not used for pathname lookup, and so will not be considered here.
117 The dcache has a number of uses apart from accelerating lookup. One
118 that will be particularly relevant is that it is closely integrated
119 with the mount table that records which filesystem is mounted where.
120 What the mount table actually stores is which dentry is mounted on top
121 of which other dentry.
123 When considering the dcache, we have another of our "two types"
124 distinctions: there are two types of filesystems.
126 Some filesystems ensure that the information in the dcache is always
127 completely accurate (though not necessarily complete). This can allow
128 the VFS to determine if a particular file does or doesn't exist
129 without checking with the filesystem, and means that the VFS can
130 protect the filesystem against certain races and other problems.
131 These are typically "local" filesystems such as ext3, XFS, and Btrfs.
133 Other filesystems don't provide that guarantee because they cannot.
134 These are typically filesystems that are shared across a network,
135 whether remote filesystems like NFS and 9P, or cluster filesystems
136 like ocfs2 or cephfs. These filesystems allow the VFS to revalidate
137 cached information, and must provide their own protection against
138 awkward races. The VFS can detect these filesystems by the
139 ``DCACHE_OP_REVALIDATE`` flag being set in the dentry.
141 REF-walk: simple concurrency management with refcounts and spinlocks
142 --------------------------------------------------------------------
144 With all of those divisions carefully classified, we can now start
145 looking at the actual process of walking along a path. In particular
146 we will start with the handling of the "everything else" part of a
147 pathname, and focus on the "REF-walk" approach to concurrency
148 management. This code is found in the ``link_path_walk()`` function, if
149 you ignore all the places that only run when "``LOOKUP_RCU``"
150 (indicating the use of RCU-walk) is set.
152 .. _Meet the Lockers: https://lwn.net/Articles/453685/
154 REF-walk is fairly heavy-handed with locks and reference counts. Not
155 as heavy-handed as in the old "big kernel lock" days, but certainly not
156 afraid of taking a lock when one is needed. It uses a variety of
157 different concurrency controls. A background understanding of the
158 various primitives is assumed, or can be gleaned from elsewhere such
159 as in `Meet the Lockers`_.
161 The locking mechanisms used by REF-walk include:
166 This uses the lockref primitive to provide both a spinlock and a
167 reference count. The special-sauce of this primitive is that the
168 conceptual sequence "lock; inc_ref; unlock;" can often be performed
169 with a single atomic memory operation.
171 Holding a reference on a dentry ensures that the dentry won't suddenly
172 be freed and used for something else, so the values in various fields
173 will behave as expected. It also protects the ``->d_inode`` reference
174 to the inode to some extent.
176 The association between a dentry and its inode is fairly permanent.
177 For example, when a file is renamed, the dentry and inode move
178 together to the new location. When a file is created the dentry will
179 initially be negative (i.e. ``d_inode`` is ``NULL``), and will be assigned
180 to the new inode as part of the act of creation.
182 When a file is deleted, this can be reflected in the cache either by
183 setting ``d_inode`` to ``NULL``, or by removing it from the hash table
184 (described shortly) used to look up the name in the parent directory.
185 If the dentry is still in use the second option is used as it is
186 perfectly legal to keep using an open file after it has been deleted
187 and having the dentry around helps. If the dentry is not otherwise in
188 use (i.e. if the refcount in ``d_lockref`` is one), only then will
189 ``d_inode`` be set to ``NULL``. Doing it this way is more efficient for a
192 So as long as a counted reference is held to a dentry, a non-``NULL`` ``->d_inode``
193 value will never be changed.
198 ``d_lock`` is a synonym for the spinlock that is part of ``d_lockref`` above.
199 For our purposes, holding this lock protects against the dentry being
200 renamed or unlinked. In particular, its parent (``d_parent``), and its
201 name (``d_name``) cannot be changed, and it cannot be removed from the
204 When looking for a name in a directory, REF-walk takes ``d_lock`` on
205 each candidate dentry that it finds in the hash table and then checks
206 that the parent and name are correct. So it doesn't lock the parent
207 while searching in the cache; it only locks children.
209 When looking for the parent for a given name (to handle "``..``"),
210 REF-walk can take ``d_lock`` to get a stable reference to ``d_parent``,
211 but it first tries a more lightweight approach. As seen in
212 ``dget_parent()``, if a reference can be claimed on the parent, and if
213 subsequently ``d_parent`` can be seen to have not changed, then there is
214 no need to actually take the lock on the child.
219 Looking up a given name in a given directory involves computing a hash
220 from the two values (the name and the dentry of the directory),
221 accessing that slot in a hash table, and searching the linked list
224 When a dentry is renamed, the name and the parent dentry can both
225 change so the hash will almost certainly change too. This would move the
226 dentry to a different chain in the hash table. If a filename search
227 happened to be looking at a dentry that was moved in this way,
228 it might end up continuing the search down the wrong chain,
229 and so miss out on part of the correct chain.
231 The name-lookup process (``d_lookup()``) does _not_ try to prevent this
232 from happening, but only to detect when it happens.
233 ``rename_lock`` is a seqlock that is updated whenever any dentry is
234 renamed. If ``d_lookup`` finds that a rename happened while it
235 unsuccessfully scanned a chain in the hash table, it simply tries
241 ``i_rwsem`` is a read/write semaphore that serializes all changes to a particular
242 directory. This ensures that, for example, an ``unlink()`` and a ``rename()``
243 cannot both happen at the same time. It also keeps the directory
244 stable while the filesystem is asked to look up a name that is not
245 currently in the dcache or, optionally, when the list of entries in a
246 directory is being retrieved with ``readdir()``.
248 This has a complementary role to that of ``d_lock``: ``i_rwsem`` on a
249 directory protects all of the names in that directory, while ``d_lock``
250 on a name protects just one name in a directory. Most changes to the
251 dcache hold ``i_rwsem`` on the relevant directory inode and briefly take
252 ``d_lock`` on one or more the dentries while the change happens. One
253 exception is when idle dentries are removed from the dcache due to
254 memory pressure. This uses ``d_lock``, but ``i_rwsem`` plays no role.
256 The semaphore affects pathname lookup in two distinct ways. Firstly it
257 prevents changes during lookup of a name in a directory. ``walk_component()`` uses
258 ``lookup_fast()`` first which, in turn, checks to see if the name is in the cache,
259 using only ``d_lock`` locking. If the name isn't found, then ``walk_component()``
260 falls back to ``lookup_slow()`` which takes a shared lock on ``i_rwsem``, checks again that
261 the name isn't in the cache, and then calls in to the filesystem to get a
262 definitive answer. A new dentry will be added to the cache regardless of
265 Secondly, when pathname lookup reaches the final component, it will
266 sometimes need to take an exclusive lock on ``i_rwsem`` before performing the last lookup so
267 that the required exclusion can be achieved. How path lookup chooses
268 to take, or not take, ``i_rwsem`` is one of the
269 issues addressed in a subsequent section.
271 If two threads attempt to look up the same name at the same time - a
272 name that is not yet in the dcache - the shared lock on ``i_rwsem`` will
273 not prevent them both adding new dentries with the same name. As this
274 would result in confusion an extra level of interlocking is used,
275 based around a secondary hash table (``in_lookup_hashtable``) and a
276 per-dentry flag bit (``DCACHE_PAR_LOOKUP``).
278 To add a new dentry to the cache while only holding a shared lock on
279 ``i_rwsem``, a thread must call ``d_alloc_parallel()``. This allocates a
280 dentry, stores the required name and parent in it, checks if there
281 is already a matching dentry in the primary or secondary hash
282 tables, and if not, stores the newly allocated dentry in the secondary
283 hash table, with ``DCACHE_PAR_LOOKUP`` set.
285 If a matching dentry was found in the primary hash table then that is
286 returned and the caller can know that it lost a race with some other
287 thread adding the entry. If no matching dentry is found in either
288 cache, the newly allocated dentry is returned and the caller can
289 detect this from the presence of ``DCACHE_PAR_LOOKUP``. In this case it
290 knows that it has won any race and now is responsible for asking the
291 filesystem to perform the lookup and find the matching inode. When
292 the lookup is complete, it must call ``d_lookup_done()`` which clears
293 the flag and does some other house keeping, including removing the
294 dentry from the secondary hash table - it will normally have been
295 added to the primary hash table already. Note that a ``struct
296 waitqueue_head`` is passed to ``d_alloc_parallel()``, and
297 ``d_lookup_done()`` must be called while this ``waitqueue_head`` is still
300 If a matching dentry is found in the secondary hash table,
301 ``d_alloc_parallel()`` has a little more work to do. It first waits for
302 ``DCACHE_PAR_LOOKUP`` to be cleared, using a wait_queue that was passed
303 to the instance of ``d_alloc_parallel()`` that won the race and that
304 will be woken by the call to ``d_lookup_done()``. It then checks to see
305 if the dentry has now been added to the primary hash table. If it
306 has, the dentry is returned and the caller just sees that it lost any
307 race. If it hasn't been added to the primary hash table, the most
308 likely explanation is that some other dentry was added instead using
309 ``d_splice_alias()``. In any case, ``d_alloc_parallel()`` repeats all the
310 look ups from the start and will normally return something from the
316 ``mnt_count`` is a per-CPU reference counter on "``mount``" structures.
317 Per-CPU here means that incrementing the count is cheap as it only
318 uses CPU-local memory, but checking if the count is zero is expensive as
319 it needs to check with every CPU. Taking a ``mnt_count`` reference
320 prevents the mount structure from disappearing as the result of regular
321 unmount operations, but does not prevent a "lazy" unmount. So holding
322 ``mnt_count`` doesn't ensure that the mount remains in the namespace and,
323 in particular, doesn't stabilize the link to the mounted-on dentry. It
324 does, however, ensure that the ``mount`` data structure remains coherent,
325 and it provides a reference to the root dentry of the mounted
326 filesystem. So a reference through ``->mnt_count`` provides a stable
327 reference to the mounted dentry, but not the mounted-on dentry.
332 ``mount_lock`` is a global seqlock, a bit like ``rename_lock``. It can be used to
333 check if any change has been made to any mount points.
335 While walking down the tree (away from the root) this lock is used when
336 crossing a mount point to check that the crossing was safe. That is,
337 the value in the seqlock is read, then the code finds the mount that
338 is mounted on the current directory, if there is one, and increments
339 the ``mnt_count``. Finally the value in ``mount_lock`` is checked against
340 the old value. If there is no change, then the crossing was safe. If there
341 was a change, the ``mnt_count`` is decremented and the whole process is
344 When walking up the tree (towards the root) by following a ".." link,
345 a little more care is needed. In this case the seqlock (which
346 contains both a counter and a spinlock) is fully locked to prevent
347 any changes to any mount points while stepping up. This locking is
348 needed to stabilize the link to the mounted-on dentry, which the
349 refcount on the mount itself doesn't ensure.
354 Finally the global (but extremely lightweight) RCU read lock is held
355 from time to time to ensure certain data structures don't get freed
358 In particular it is held while scanning chains in the dcache hash
359 table, and the mount point hash table.
361 Bringing it together with ``struct nameidata``
362 ----------------------------------------------
364 .. _First edition Unix: http://minnie.tuhs.org/cgi-bin/utree.pl?file=V1/u2.s
366 Throughout the process of walking a path, the current status is stored
367 in a ``struct nameidata``, "namei" being the traditional name - dating
368 all the way back to `First Edition Unix`_ - of the function that
369 converts a "name" to an "inode". ``struct nameidata`` contains (among
375 A ``path`` contains a ``struct vfsmount`` (which is
376 embedded in a ``struct mount``) and a ``struct dentry``. Together these
377 record the current status of the walk. They start out referring to the
378 starting point (the current working directory, the root directory, or some other
379 directory identified by a file descriptor), and are updated on each
380 step. A reference through ``d_lockref`` and ``mnt_count`` is always
386 This is a string together with a length (i.e. _not_ ``nul`` terminated)
387 that is the "next" component in the pathname.
392 This is one of ``LAST_NORM``, ``LAST_ROOT``, ``LAST_DOT``, ``LAST_DOTDOT``, or
393 ``LAST_BIND``. The ``last`` field is only valid if the type is
394 ``LAST_NORM``. ``LAST_BIND`` is used when following a symlink and no
395 components of the symlink have been processed yet. Others should be
396 fairly self-explanatory.
401 This is used to hold a reference to the effective root of the
402 filesystem. Often that reference won't be needed, so this field is
403 only assigned the first time it is used, or when a non-standard root
404 is requested. Keeping a reference in the ``nameidata`` ensures that
405 only one root is in effect for the entire path walk, even if it races
406 with a ``chroot()`` system call.
408 The root is needed when either of two conditions holds: (1) either the
409 pathname or a symbolic link starts with a "'/'", or (2) a "``..``"
410 component is being handled, since "``..``" from the root must always stay
411 at the root. The value used is usually the current root directory of
412 the calling process. An alternate root can be provided as when
413 ``sysctl()`` calls ``file_open_root()``, and when NFSv4 or Btrfs call
414 ``mount_subtree()``. In each case a pathname is being looked up in a very
415 specific part of the filesystem, and the lookup must not be allowed to
416 escape that subtree. It works a bit like a local ``chroot()``.
418 Ignoring the handling of symbolic links, we can now describe the
419 "``link_path_walk()``" function, which handles the lookup of everything
420 except the final component as:
422 Given a path (``name``) and a nameidata structure (``nd``), check that the
423 current directory has execute permission and then advance ``name``
424 over one component while updating ``last_type`` and ``last``. If that
425 was the final component, then return, otherwise call
426 ``walk_component()`` and repeat from the top.
428 ``walk_component()`` is even easier. If the component is ``LAST_DOTS``,
429 it calls ``handle_dots()`` which does the necessary locking as already
430 described. If it finds a ``LAST_NORM`` component it first calls
431 "``lookup_fast()``" which only looks in the dcache, but will ask the
432 filesystem to revalidate the result if it is that sort of filesystem.
433 If that doesn't get a good result, it calls "``lookup_slow()``" which
434 takes ``i_rwsem``, rechecks the cache, and then asks the filesystem
435 to find a definitive answer. Each of these will call
436 ``follow_managed()`` (as described below) to handle any mount points.
438 In the absence of symbolic links, ``walk_component()`` creates a new
439 ``struct path`` containing a counted reference to the new dentry and a
440 reference to the new ``vfsmount`` which is only counted if it is
441 different from the previous ``vfsmount``. It then calls
442 ``path_to_nameidata()`` to install the new ``struct path`` in the
443 ``struct nameidata`` and drop the unneeded references.
445 This "hand-over-hand" sequencing of getting a reference to the new
446 dentry before dropping the reference to the previous dentry may
447 seem obvious, but is worth pointing out so that we will recognize its
448 analogue in the "RCU-walk" version.
450 Handling the final component
451 ----------------------------
453 ``link_path_walk()`` only walks as far as setting ``nd->last`` and
454 ``nd->last_type`` to refer to the final component of the path. It does
455 not call ``walk_component()`` that last time. Handling that final
456 component remains for the caller to sort out. Those callers are
457 ``path_lookupat()``, ``path_parentat()``, ``path_mountpoint()`` and
458 ``path_openat()`` each of which handles the differing requirements of
459 different system calls.
461 ``path_parentat()`` is clearly the simplest - it just wraps a little bit
462 of housekeeping around ``link_path_walk()`` and returns the parent
463 directory and final component to the caller. The caller will be either
464 aiming to create a name (via ``filename_create()``) or remove or rename
465 a name (in which case ``user_path_parent()`` is used). They will use
466 ``i_rwsem`` to exclude other changes while they validate and then
467 perform their operation.
469 ``path_lookupat()`` is nearly as simple - it is used when an existing
470 object is wanted such as by ``stat()`` or ``chmod()``. It essentially just
471 calls ``walk_component()`` on the final component through a call to
472 ``lookup_last()``. ``path_lookupat()`` returns just the final dentry.
474 ``path_mountpoint()`` handles the special case of unmounting which must
475 not try to revalidate the mounted filesystem. It effectively
476 contains, through a call to ``mountpoint_last()``, an alternate
477 implementation of ``lookup_slow()`` which skips that step. This is
478 important when unmounting a filesystem that is inaccessible, such as
479 one provided by a dead NFS server.
481 Finally ``path_openat()`` is used for the ``open()`` system call; it
482 contains, in support functions starting with "``do_last()``", all the
483 complexity needed to handle the different subtleties of O_CREAT (with
484 or without O_EXCL), final "``/``" characters, and trailing symbolic
485 links. We will revisit this in the final part of this series, which
486 focuses on those symbolic links. "``do_last()``" will sometimes, but
487 not always, take ``i_rwsem``, depending on what it finds.
489 Each of these, or the functions which call them, need to be alert to
490 the possibility that the final component is not ``LAST_NORM``. If the
491 goal of the lookup is to create something, then any value for
492 ``last_type`` other than ``LAST_NORM`` will result in an error. For
493 example if ``path_parentat()`` reports ``LAST_DOTDOT``, then the caller
494 won't try to create that name. They also check for trailing slashes
495 by testing ``last.name[last.len]``. If there is any character beyond
496 the final component, it must be a trailing slash.
498 Revalidation and automounts
499 ---------------------------
501 Apart from symbolic links, there are only two parts of the "REF-walk"
502 process not yet covered. One is the handling of stale cache entries
503 and the other is automounts.
505 On filesystems that require it, the lookup routines will call the
506 ``->d_revalidate()`` dentry method to ensure that the cached information
507 is current. This will often confirm validity or update a few details
508 from a server. In some cases it may find that there has been change
509 further up the path and that something that was thought to be valid
510 previously isn't really. When this happens the lookup of the whole
511 path is aborted and retried with the "``LOOKUP_REVAL``" flag set. This
512 forces revalidation to be more thorough. We will see more details of
513 this retry process in the next article.
515 Automount points are locations in the filesystem where an attempt to
516 lookup a name can trigger changes to how that lookup should be
517 handled, in particular by mounting a filesystem there. These are
518 covered in greater detail in autofs.txt in the Linux documentation
519 tree, but a few notes specifically related to path lookup are in order
522 The Linux VFS has a concept of "managed" dentries which is reflected
523 in function names such as "``follow_managed()``". There are three
524 potentially interesting things about these dentries corresponding
525 to three different flags that might be set in ``dentry->d_flags``:
527 ``DCACHE_MANAGE_TRANSIT``
528 ~~~~~~~~~~~~~~~~~~~~~~~~~
530 If this flag has been set, then the filesystem has requested that the
531 ``d_manage()`` dentry operation be called before handling any possible
532 mount point. This can perform two particular services:
534 It can block to avoid races. If an automount point is being
535 unmounted, the ``d_manage()`` function will usually wait for that
536 process to complete before letting the new lookup proceed and possibly
537 trigger a new automount.
539 It can selectively allow only some processes to transit through a
540 mount point. When a server process is managing automounts, it may
541 need to access a directory without triggering normal automount
542 processing. That server process can identify itself to the ``autofs``
543 filesystem, which will then give it a special pass through
544 ``d_manage()`` by returning ``-EISDIR``.
549 This flag is set on every dentry that is mounted on. As Linux
550 supports multiple filesystem namespaces, it is possible that the
551 dentry may not be mounted on in *this* namespace, just in some
552 other. So this flag is seen as a hint, not a promise.
554 If this flag is set, and ``d_manage()`` didn't return ``-EISDIR``,
555 ``lookup_mnt()`` is called to examine the mount hash table (honoring the
556 ``mount_lock`` described earlier) and possibly return a new ``vfsmount``
557 and a new ``dentry`` (both with counted references).
559 ``DCACHE_NEED_AUTOMOUNT``
560 ~~~~~~~~~~~~~~~~~~~~~~~~~
562 If ``d_manage()`` allowed us to get this far, and ``lookup_mnt()`` didn't
563 find a mount point, then this flag causes the ``d_automount()`` dentry
564 operation to be called.
566 The ``d_automount()`` operation can be arbitrarily complex and may
567 communicate with server processes etc. but it should ultimately either
568 report that there was an error, that there was nothing to mount, or
569 should provide an updated ``struct path`` with new ``dentry`` and ``vfsmount``.
571 In the latter case, ``finish_automount()`` will be called to safely
572 install the new mount point into the mount table.
574 There is no new locking of import here and it is important that no
575 locks (only counted references) are held over this processing due to
576 the very real possibility of extended delays.
577 This will become more important next time when we examine RCU-walk
578 which is particularly sensitive to delays.
580 RCU-walk - faster pathname lookup in Linux
581 ==========================================
583 RCU-walk is another algorithm for performing pathname lookup in Linux.
584 It is in many ways similar to REF-walk and the two share quite a bit
585 of code. The significant difference in RCU-walk is how it allows for
586 the possibility of concurrent access.
588 We noted that REF-walk is complex because there are numerous details
589 and special cases. RCU-walk reduces this complexity by simply
590 refusing to handle a number of cases -- it instead falls back to
591 REF-walk. The difficulty with RCU-walk comes from a different
592 direction: unfamiliarity. The locking rules when depending on RCU are
593 quite different from traditional locking, so we will spend a little extra
594 time when we come to those.
596 Clear demarcation of roles
597 --------------------------
599 The easiest way to manage concurrency is to forcibly stop any other
600 thread from changing the data structures that a given thread is
601 looking at. In cases where no other thread would even think of
602 changing the data and lots of different threads want to read at the
603 same time, this can be very costly. Even when using locks that permit
604 multiple concurrent readers, the simple act of updating the count of
605 the number of current readers can impose an unwanted cost. So the
606 goal when reading a shared data structure that no other process is
607 changing is to avoid writing anything to memory at all. Take no
608 locks, increment no counts, leave no footprints.
610 The REF-walk mechanism already described certainly doesn't follow this
611 principle, but then it is really designed to work when there may well
612 be other threads modifying the data. RCU-walk, in contrast, is
613 designed for the common situation where there are lots of frequent
614 readers and only occasional writers. This may not be common in all
615 parts of the filesystem tree, but in many parts it will be. For the
616 other parts it is important that RCU-walk can quickly fall back to
619 Pathname lookup always starts in RCU-walk mode but only remains there
620 as long as what it is looking for is in the cache and is stable. It
621 dances lightly down the cached filesystem image, leaving no footprints
622 and carefully watching where it is, to be sure it doesn't trip. If it
623 notices that something has changed or is changing, or if something
624 isn't in the cache, then it tries to stop gracefully and switch to
627 This stopping requires getting a counted reference on the current
628 ``vfsmount`` and ``dentry``, and ensuring that these are still valid -
629 that a path walk with REF-walk would have found the same entries.
630 This is an invariant that RCU-walk must guarantee. It can only make
631 decisions, such as selecting the next step, that are decisions which
632 REF-walk could also have made if it were walking down the tree at the
633 same time. If the graceful stop succeeds, the rest of the path is
634 processed with the reliable, if slightly sluggish, REF-walk. If
635 RCU-walk finds it cannot stop gracefully, it simply gives up and
636 restarts from the top with REF-walk.
638 This pattern of "try RCU-walk, if that fails try REF-walk" can be
639 clearly seen in functions like ``filename_lookup()``,
640 ``filename_parentat()``, ``filename_mountpoint()``,
641 ``do_filp_open()``, and ``do_file_open_root()``. These five
642 correspond roughly to the four ``path_``* functions we met earlier,
643 each of which calls ``link_path_walk()``. The ``path_*`` functions are
644 called using different mode flags until a mode is found which works.
645 They are first called with ``LOOKUP_RCU`` set to request "RCU-walk". If
646 that fails with the error ``ECHILD`` they are called again with no
647 special flag to request "REF-walk". If either of those report the
648 error ``ESTALE`` a final attempt is made with ``LOOKUP_REVAL`` set (and no
649 ``LOOKUP_RCU``) to ensure that entries found in the cache are forcibly
650 revalidated - normally entries are only revalidated if the filesystem
651 determines that they are too old to trust.
653 The ``LOOKUP_RCU`` attempt may drop that flag internally and switch to
654 REF-walk, but will never then try to switch back to RCU-walk. Places
655 that trip up RCU-walk are much more likely to be near the leaves and
656 so it is very unlikely that there will be much, if any, benefit from
659 RCU and seqlocks: fast and light
660 --------------------------------
662 RCU is, unsurprisingly, critical to RCU-walk mode. The
663 ``rcu_read_lock()`` is held for the entire time that RCU-walk is walking
664 down a path. The particular guarantee it provides is that the key
665 data structures - dentries, inodes, super_blocks, and mounts - will
666 not be freed while the lock is held. They might be unlinked or
667 invalidated in one way or another, but the memory will not be
668 repurposed so values in various fields will still be meaningful. This
669 is the only guarantee that RCU provides; everything else is done using
672 As we saw above, REF-walk holds a counted reference to the current
673 dentry and the current vfsmount, and does not release those references
674 before taking references to the "next" dentry or vfsmount. It also
675 sometimes takes the ``d_lock`` spinlock. These references and locks are
676 taken to prevent certain changes from happening. RCU-walk must not
677 take those references or locks and so cannot prevent such changes.
678 Instead, it checks to see if a change has been made, and aborts or
681 To preserve the invariant mentioned above (that RCU-walk may only make
682 decisions that REF-walk could have made), it must make the checks at
683 or near the same places that REF-walk holds the references. So, when
684 REF-walk increments a reference count or takes a spinlock, RCU-walk
685 samples the status of a seqlock using ``read_seqcount_begin()`` or a
686 similar function. When REF-walk decrements the count or drops the
687 lock, RCU-walk checks if the sampled status is still valid using
688 ``read_seqcount_retry()`` or similar.
690 However, there is a little bit more to seqlocks than that. If
691 RCU-walk accesses two different fields in a seqlock-protected
692 structure, or accesses the same field twice, there is no a priori
693 guarantee of any consistency between those accesses. When consistency
694 is needed - which it usually is - RCU-walk must take a copy and then
695 use ``read_seqcount_retry()`` to validate that copy.
697 ``read_seqcount_retry()`` not only checks the sequence number, but also
698 imposes a memory barrier so that no memory-read instruction from
699 *before* the call can be delayed until *after* the call, either by the
700 CPU or by the compiler. A simple example of this can be seen in
701 ``slow_dentry_cmp()`` which, for filesystems which do not use simple
702 byte-wise name equality, calls into the filesystem to compare a name
703 against a dentry. The length and name pointer are copied into local
704 variables, then ``read_seqcount_retry()`` is called to confirm the two
705 are consistent, and only then is ``->d_compare()`` called. When
706 standard filename comparison is used, ``dentry_cmp()`` is called
707 instead. Notably it does _not_ use ``read_seqcount_retry()``, but
708 instead has a large comment explaining why the consistency guarantee
709 isn't necessary. A subsequent ``read_seqcount_retry()`` will be
710 sufficient to catch any problem that could occur at this point.
712 With that little refresher on seqlocks out of the way we can look at
713 the bigger picture of how RCU-walk uses seqlocks.
715 ``mount_lock`` and ``nd->m_seq``
716 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
718 We already met the ``mount_lock`` seqlock when REF-walk used it to
719 ensure that crossing a mount point is performed safely. RCU-walk uses
720 it for that too, but for quite a bit more.
722 Instead of taking a counted reference to each ``vfsmount`` as it
723 descends the tree, RCU-walk samples the state of ``mount_lock`` at the
724 start of the walk and stores this initial sequence number in the
725 ``struct nameidata`` in the ``m_seq`` field. This one lock and one
726 sequence number are used to validate all accesses to all ``vfsmounts``,
727 and all mount point crossings. As changes to the mount table are
728 relatively rare, it is reasonable to fall back on REF-walk any time
729 that any "mount" or "unmount" happens.
731 ``m_seq`` is checked (using ``read_seqretry()``) at the end of an RCU-walk
732 sequence, whether switching to REF-walk for the rest of the path or
733 when the end of the path is reached. It is also checked when stepping
734 down over a mount point (in ``__follow_mount_rcu()``) or up (in
735 ``follow_dotdot_rcu()``). If it is ever found to have changed, the
736 whole RCU-walk sequence is aborted and the path is processed again by
739 If RCU-walk finds that ``mount_lock`` hasn't changed then it can be sure
740 that, had REF-walk taken counted references on each vfsmount, the
741 results would have been the same. This ensures the invariant holds,
742 at least for vfsmount structures.
744 ``dentry->d_seq`` and ``nd->seq``
745 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
747 In place of taking a count or lock on ``d_reflock``, RCU-walk samples
748 the per-dentry ``d_seq`` seqlock, and stores the sequence number in the
749 ``seq`` field of the nameidata structure, so ``nd->seq`` should always be
750 the current sequence number of ``nd->dentry``. This number needs to be
751 revalidated after copying, and before using, the name, parent, or
754 The handling of the name we have already looked at, and the parent is
755 only accessed in ``follow_dotdot_rcu()`` which fairly trivially follows
756 the required pattern, though it does so for three different cases.
758 When not at a mount point, ``d_parent`` is followed and its ``d_seq`` is
759 collected. When we are at a mount point, we instead follow the
760 ``mnt->mnt_mountpoint`` link to get a new dentry and collect its
761 ``d_seq``. Then, after finally finding a ``d_parent`` to follow, we must
762 check if we have landed on a mount point and, if so, must find that
763 mount point and follow the ``mnt->mnt_root`` link. This would imply a
764 somewhat unusual, but certainly possible, circumstance where the
765 starting point of the path lookup was in part of the filesystem that
766 was mounted on, and so not visible from the root.
768 The inode pointer, stored in ``->d_inode``, is a little more
769 interesting. The inode will always need to be accessed at least
770 twice, once to determine if it is NULL and once to verify access
771 permissions. Symlink handling requires a validated inode pointer too.
772 Rather than revalidating on each access, a copy is made on the first
773 access and it is stored in the ``inode`` field of ``nameidata`` from where
774 it can be safely accessed without further validation.
776 ``lookup_fast()`` is the only lookup routine that is used in RCU-mode,
777 ``lookup_slow()`` being too slow and requiring locks. It is in
778 ``lookup_fast()`` that we find the important "hand over hand" tracking
779 of the current dentry.
781 The current ``dentry`` and current ``seq`` number are passed to
782 ``__d_lookup_rcu()`` which, on success, returns a new ``dentry`` and a
783 new ``seq`` number. ``lookup_fast()`` then copies the inode pointer and
784 revalidates the new ``seq`` number. It then validates the old ``dentry``
785 with the old ``seq`` number one last time and only then continues. This
786 process of getting the ``seq`` number of the new dentry and then
787 checking the ``seq`` number of the old exactly mirrors the process of
788 getting a counted reference to the new dentry before dropping that for
789 the old dentry which we saw in REF-walk.
791 No ``inode->i_rwsem`` or even ``rename_lock``
792 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
794 A semaphore is a fairly heavyweight lock that can only be taken when it is
795 permissible to sleep. As ``rcu_read_lock()`` forbids sleeping,
796 ``inode->i_rwsem`` plays no role in RCU-walk. If some other thread does
797 take ``i_rwsem`` and modifies the directory in a way that RCU-walk needs
798 to notice, the result will be either that RCU-walk fails to find the
799 dentry that it is looking for, or it will find a dentry which
800 ``read_seqretry()`` won't validate. In either case it will drop down to
801 REF-walk mode which can take whatever locks are needed.
803 Though ``rename_lock`` could be used by RCU-walk as it doesn't require
804 any sleeping, RCU-walk doesn't bother. REF-walk uses ``rename_lock`` to
805 protect against the possibility of hash chains in the dcache changing
806 while they are being searched. This can result in failing to find
807 something that actually is there. When RCU-walk fails to find
808 something in the dentry cache, whether it is really there or not, it
809 already drops down to REF-walk and tries again with appropriate
810 locking. This neatly handles all cases, so adding extra checks on
811 rename_lock would bring no significant value.
813 ``unlazy walk()`` and ``complete_walk()``
814 -----------------------------------------
816 That "dropping down to REF-walk" typically involves a call to
817 ``unlazy_walk()``, so named because "RCU-walk" is also sometimes
818 referred to as "lazy walk". ``unlazy_walk()`` is called when
819 following the path down to the current vfsmount/dentry pair seems to
820 have proceeded successfully, but the next step is problematic. This
821 can happen if the next name cannot be found in the dcache, if
822 permission checking or name revalidation couldn't be achieved while
823 the ``rcu_read_lock()`` is held (which forbids sleeping), if an
824 automount point is found, or in a couple of cases involving symlinks.
825 It is also called from ``complete_walk()`` when the lookup has reached
826 the final component, or the very end of the path, depending on which
827 particular flavor of lookup is used.
829 Other reasons for dropping out of RCU-walk that do not trigger a call
830 to ``unlazy_walk()`` are when some inconsistency is found that cannot be
831 handled immediately, such as ``mount_lock`` or one of the ``d_seq``
832 seqlocks reporting a change. In these cases the relevant function
833 will return ``-ECHILD`` which will percolate up until it triggers a new
834 attempt from the top using REF-walk.
836 For those cases where ``unlazy_walk()`` is an option, it essentially
837 takes a reference on each of the pointers that it holds (vfsmount,
838 dentry, and possibly some symbolic links) and then verifies that the
839 relevant seqlocks have not been changed. If there have been changes,
840 it, too, aborts with ``-ECHILD``, otherwise the transition to REF-walk
841 has been a success and the lookup process continues.
843 Taking a reference on those pointers is not quite as simple as just
844 incrementing a counter. That works to take a second reference if you
845 already have one (often indirectly through another object), but it
846 isn't sufficient if you don't actually have a counted reference at
847 all. For ``dentry->d_lockref``, it is safe to increment the reference
848 counter to get a reference unless it has been explicitly marked as
849 "dead" which involves setting the counter to ``-128``.
850 ``lockref_get_not_dead()`` achieves this.
852 For ``mnt->mnt_count`` it is safe to take a reference as long as
853 ``mount_lock`` is then used to validate the reference. If that
854 validation fails, it may *not* be safe to just drop that reference in
855 the standard way of calling ``mnt_put()`` - an unmount may have
856 progressed too far. So the code in ``legitimize_mnt()``, when it
857 finds that the reference it got might not be safe, checks the
858 ``MNT_SYNC_UMOUNT`` flag to determine if a simple ``mnt_put()`` is
859 correct, or if it should just decrement the count and pretend none of
862 Taking care in filesystems
863 --------------------------
865 RCU-walk depends almost entirely on cached information and often will
866 not call into the filesystem at all. However there are two places,
867 besides the already-mentioned component-name comparison, where the
868 file system might be included in RCU-walk, and it must know to be
871 If the filesystem has non-standard permission-checking requirements -
872 such as a networked filesystem which may need to check with the server
873 - the ``i_op->permission`` interface might be called during RCU-walk.
874 In this case an extra "``MAY_NOT_BLOCK``" flag is passed so that it
875 knows not to sleep, but to return ``-ECHILD`` if it cannot complete
876 promptly. ``i_op->permission`` is given the inode pointer, not the
877 dentry, so it doesn't need to worry about further consistency checks.
878 However if it accesses any other filesystem data structures, it must
879 ensure they are safe to be accessed with only the ``rcu_read_lock()``
880 held. This typically means they must be freed using ``kfree_rcu()`` or
883 .. _READ_ONCE: https://lwn.net/Articles/624126/
885 If the filesystem may need to revalidate dcache entries, then
886 ``d_op->d_revalidate`` may be called in RCU-walk too. This interface
887 *is* passed the dentry but does not have access to the ``inode`` or the
888 ``seq`` number from the ``nameidata``, so it needs to be extra careful
889 when accessing fields in the dentry. This "extra care" typically
890 involves using `READ_ONCE() <READ_ONCE_>`_ to access fields, and verifying the
891 result is not NULL before using it. This pattern can be seen in
892 ``nfs_lookup_revalidate()``.
897 In various places in the details of REF-walk and RCU-walk, and also in
898 the big picture, there are a couple of related patterns that are worth
901 The first is "try quickly and check, if that fails try slowly". We
902 can see that in the high-level approach of first trying RCU-walk and
903 then trying REF-walk, and in places where ``unlazy_walk()`` is used to
904 switch to REF-walk for the rest of the path. We also saw it earlier
905 in ``dget_parent()`` when following a "``..``" link. It tries a quick way
906 to get a reference, then falls back to taking locks if needed.
908 The second pattern is "try quickly and check, if that fails try
909 again - repeatedly". This is seen with the use of ``rename_lock`` and
910 ``mount_lock`` in REF-walk. RCU-walk doesn't make use of this pattern -
911 if anything goes wrong it is much safer to just abort and try a more
914 The emphasis here is "try quickly and check". It should probably be
915 "try quickly _and carefully,_ then check". The fact that checking is
916 needed is a reminder that the system is dynamic and only a limited
917 number of things are safe at all. The most likely cause of errors in
918 this whole process is assuming something is safe when in reality it
919 isn't. Careful consideration of what exactly guarantees the safety of
920 each access is sometimes necessary.
922 A walk among the symlinks
923 =========================
925 There are several basic issues that we will examine to understand the
926 handling of symbolic links: the symlink stack, together with cache
927 lifetimes, will help us understand the overall recursive handling of
928 symlinks and lead to the special care needed for the final component.
929 Then a consideration of access-time updates and summary of the various
930 flags controlling lookup will finish the story.
935 There are only two sorts of filesystem objects that can usefully
936 appear in a path prior to the final component: directories and symlinks.
937 Handling directories is quite straightforward: the new directory
938 simply becomes the starting point at which to interpret the next
939 component on the path. Handling symbolic links requires a bit more
942 Conceptually, symbolic links could be handled by editing the path. If
943 a component name refers to a symbolic link, then that component is
944 replaced by the body of the link and, if that body starts with a '/',
945 then all preceding parts of the path are discarded. This is what the
946 "``readlink -f``" command does, though it also edits out "``.``" and
949 Directly editing the path string is not really necessary when looking
950 up a path, and discarding early components is pointless as they aren't
951 looked at anyway. Keeping track of all remaining components is
952 important, but they can of course be kept separately; there is no need
953 to concatenate them. As one symlink may easily refer to another,
954 which in turn can refer to a third, we may need to keep the remaining
955 components of several paths, each to be processed when the preceding
956 ones are completed. These path remnants are kept on a stack of
959 There are two reasons for placing limits on how many symlinks can
960 occur in a single path lookup. The most obvious is to avoid loops.
961 If a symlink referred to itself either directly or through
962 intermediaries, then following the symlink can never complete
963 successfully - the error ``ELOOP`` must be returned. Loops can be
964 detected without imposing limits, but limits are the simplest solution
965 and, given the second reason for restriction, quite sufficient.
967 .. _outlined recently: http://thread.gmane.org/gmane.linux.kernel/1934390/focus=1934550
969 The second reason was `outlined recently`_ by Linus:
971 Because it's a latency and DoS issue too. We need to react well to
972 true loops, but also to "very deep" non-loops. It's not about memory
973 use, it's about users triggering unreasonable CPU resources.
975 Linux imposes a limit on the length of any pathname: ``PATH_MAX``, which
976 is 4096. There are a number of reasons for this limit; not letting the
977 kernel spend too much time on just one path is one of them. With
978 symbolic links you can effectively generate much longer paths so some
979 sort of limit is needed for the same reason. Linux imposes a limit of
980 at most 40 symlinks in any one path lookup. It previously imposed a
981 further limit of eight on the maximum depth of recursion, but that was
982 raised to 40 when a separate stack was implemented, so there is now
985 The ``nameidata`` structure that we met in an earlier article contains a
986 small stack that can be used to store the remaining part of up to two
987 symlinks. In many cases this will be sufficient. If it isn't, a
988 separate stack is allocated with room for 40 symlinks. Pathname
989 lookup will never exceed that stack as, once the 40th symlink is
990 detected, an error is returned.
992 It might seem that the name remnants are all that needs to be stored on
993 this stack, but we need a bit more. To see that, we need to move on to
996 Storage and lifetime of cached symlinks
997 ---------------------------------------
999 Like other filesystem resources, such as inodes and directory
1000 entries, symlinks are cached by Linux to avoid repeated costly access
1001 to external storage. It is particularly important for RCU-walk to be
1002 able to find and temporarily hold onto these cached entries, so that
1003 it doesn't need to drop down into REF-walk.
1005 .. _object-oriented design pattern: https://lwn.net/Articles/446317/
1007 While each filesystem is free to make its own choice, symlinks are
1008 typically stored in one of two places. Short symlinks are often
1009 stored directly in the inode. When a filesystem allocates a ``struct
1010 inode`` it typically allocates extra space to store private data (a
1011 common `object-oriented design pattern`_ in the kernel). This will
1012 sometimes include space for a symlink. The other common location is
1013 in the page cache, which normally stores the content of files. The
1014 pathname in a symlink can be seen as the content of that symlink and
1015 can easily be stored in the page cache just like file content.
1017 When neither of these is suitable, the next most likely scenario is
1018 that the filesystem will allocate some temporary memory and copy or
1019 construct the symlink content into that memory whenever it is needed.
1021 When the symlink is stored in the inode, it has the same lifetime as
1022 the inode which, itself, is protected by RCU or by a counted reference
1023 on the dentry. This means that the mechanisms that pathname lookup
1024 uses to access the dcache and icache (inode cache) safely are quite
1025 sufficient for accessing some cached symlinks safely. In these cases,
1026 the ``i_link`` pointer in the inode is set to point to wherever the
1027 symlink is stored and it can be accessed directly whenever needed.
1029 When the symlink is stored in the page cache or elsewhere, the
1030 situation is not so straightforward. A reference on a dentry or even
1031 on an inode does not imply any reference on cached pages of that
1032 inode, and even an ``rcu_read_lock()`` is not sufficient to ensure that
1033 a page will not disappear. So for these symlinks the pathname lookup
1034 code needs to ask the filesystem to provide a stable reference and,
1035 significantly, needs to release that reference when it is finished
1038 Taking a reference to a cache page is often possible even in RCU-walk
1039 mode. It does require making changes to memory, which is best avoided,
1040 but that isn't necessarily a big cost and it is better than dropping
1041 out of RCU-walk mode completely. Even filesystems that allocate
1042 space to copy the symlink into can use ``GFP_ATOMIC`` to often successfully
1043 allocate memory without the need to drop out of RCU-walk. If a
1044 filesystem cannot successfully get a reference in RCU-walk mode, it
1045 must return ``-ECHILD`` and ``unlazy_walk()`` will be called to return to
1046 REF-walk mode in which the filesystem is allowed to sleep.
1048 The place for all this to happen is the ``i_op->follow_link()`` inode
1049 method. In the present mainline code this is never actually called in
1050 RCU-walk mode as the rewrite is not quite complete. It is likely that
1051 in a future release this method will be passed an ``inode`` pointer when
1052 called in RCU-walk mode so it both (1) knows to be careful, and (2) has the
1053 validated pointer. Much like the ``i_op->permission()`` method we
1054 looked at previously, ``->follow_link()`` would need to be careful that
1055 all the data structures it references are safe to be accessed while
1056 holding no counted reference, only the RCU lock. Though getting a
1057 reference with ``->follow_link()`` is not yet done in RCU-walk mode, the
1058 code is ready to release the reference when that does happen.
1060 This need to drop the reference to a symlink adds significant
1061 complexity. It requires a reference to the inode so that the
1062 ``i_op->put_link()`` inode operation can be called. In REF-walk, that
1063 reference is kept implicitly through a reference to the dentry, so
1064 keeping the ``struct path`` of the symlink is easiest. For RCU-walk,
1065 the pointer to the inode is kept separately. To allow switching from
1066 RCU-walk back to REF-walk in the middle of processing nested symlinks
1067 we also need the seq number for the dentry so we can confirm that
1068 switching back was safe.
1070 Finally, when providing a reference to a symlink, the filesystem also
1071 provides an opaque "cookie" that must be passed to ``->put_link()`` so that it
1072 knows what to free. This might be the allocated memory area, or a
1073 pointer to the ``struct page`` in the page cache, or something else
1074 completely. Only the filesystem knows what it is.
1076 In order for the reference to each symlink to be dropped when the walk completes,
1077 whether in RCU-walk or REF-walk, the symlink stack needs to contain,
1078 along with the path remnants:
1080 - the ``struct path`` to provide a reference to the inode in REF-walk
1081 - the ``struct inode *`` to provide a reference to the inode in RCU-walk
1082 - the ``seq`` to allow the path to be safely switched from RCU-walk to REF-walk
1083 - the ``cookie`` that tells ``->put_path()`` what to put.
1085 This means that each entry in the symlink stack needs to hold five
1086 pointers and an integer instead of just one pointer (the path
1087 remnant). On a 64-bit system, this is about 40 bytes per entry;
1088 with 40 entries it adds up to 1600 bytes total, which is less than
1089 half a page. So it might seem like a lot, but is by no means
1092 Note that, in a given stack frame, the path remnant (``name``) is not
1093 part of the symlink that the other fields refer to. It is the remnant
1094 to be followed once that symlink has been fully parsed.
1096 Following the symlink
1097 ---------------------
1099 The main loop in ``link_path_walk()`` iterates seamlessly over all
1100 components in the path and all of the non-final symlinks. As symlinks
1101 are processed, the ``name`` pointer is adjusted to point to a new
1102 symlink, or is restored from the stack, so that much of the loop
1103 doesn't need to notice. Getting this ``name`` variable on and off the
1104 stack is very straightforward; pushing and popping the references is
1105 a little more complex.
1107 When a symlink is found, ``walk_component()`` returns the value ``1``
1108 (``0`` is returned for any other sort of success, and a negative number
1109 is, as usual, an error indicator). This causes ``get_link()`` to be
1110 called; it then gets the link from the filesystem. Providing that
1111 operation is successful, the old path ``name`` is placed on the stack,
1112 and the new value is used as the ``name`` for a while. When the end of
1113 the path is found (i.e. ``*name`` is ``'\0'``) the old ``name`` is restored
1114 off the stack and path walking continues.
1116 Pushing and popping the reference pointers (inode, cookie, etc.) is more
1117 complex in part because of the desire to handle tail recursion. When
1118 the last component of a symlink itself points to a symlink, we
1119 want to pop the symlink-just-completed off the stack before pushing
1120 the symlink-just-found to avoid leaving empty path remnants that would
1121 just get in the way.
1123 It is most convenient to push the new symlink references onto the
1124 stack in ``walk_component()`` immediately when the symlink is found;
1125 ``walk_component()`` is also the last piece of code that needs to look at the
1126 old symlink as it walks that last component. So it is quite
1127 convenient for ``walk_component()`` to release the old symlink and pop
1128 the references just before pushing the reference information for the
1129 new symlink. It is guided in this by two flags; ``WALK_GET``, which
1130 gives it permission to follow a symlink if it finds one, and
1131 ``WALK_PUT``, which tells it to release the current symlink after it has been
1132 followed. ``WALK_PUT`` is tested first, leading to a call to
1133 ``put_link()``. ``WALK_GET`` is tested subsequently (by
1134 ``should_follow_link()``) leading to a call to ``pick_link()`` which sets
1137 Symlinks with no final component
1138 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1140 A pair of special-case symlinks deserve a little further explanation.
1141 Both result in a new ``struct path`` (with mount and dentry) being set
1142 up in the ``nameidata``, and result in ``get_link()`` returning ``NULL``.
1144 The more obvious case is a symlink to "``/``". All symlinks starting
1145 with "``/``" are detected in ``get_link()`` which resets the ``nameidata``
1146 to point to the effective filesystem root. If the symlink only
1147 contains "``/``" then there is nothing more to do, no components at all,
1148 so ``NULL`` is returned to indicate that the symlink can be released and
1149 the stack frame discarded.
1151 The other case involves things in ``/proc`` that look like symlinks but
1154 $ ls -l /proc/self/fd/1
1155 lrwx------ 1 neilb neilb 64 Jun 13 10:19 /proc/self/fd/1 -> /dev/pts/4
1157 Every open file descriptor in any process is represented in ``/proc`` by
1158 something that looks like a symlink. It is really a reference to the
1159 target file, not just the name of it. When you ``readlink`` these
1160 objects you get a name that might refer to the same file - unless it
1161 has been unlinked or mounted over. When ``walk_component()`` follows
1162 one of these, the ``->follow_link()`` method in "procfs" doesn't return
1163 a string name, but instead calls ``nd_jump_link()`` which updates the
1164 ``nameidata`` in place to point to that target. ``->follow_link()`` then
1165 returns ``NULL``. Again there is no final component and ``get_link()``
1166 reports this by leaving the ``last_type`` field of ``nameidata`` as
1169 Following the symlink in the final component
1170 --------------------------------------------
1172 All this leads to ``link_path_walk()`` walking down every component, and
1173 following all symbolic links it finds, until it reaches the final
1174 component. This is just returned in the ``last`` field of ``nameidata``.
1175 For some callers, this is all they need; they want to create that
1176 ``last`` name if it doesn't exist or give an error if it does. Other
1177 callers will want to follow a symlink if one is found, and possibly
1178 apply special handling to the last component of that symlink, rather
1179 than just the last component of the original file name. These callers
1180 potentially need to call ``link_path_walk()`` again and again on
1181 successive symlinks until one is found that doesn't point to another
1184 This case is handled by the relevant caller of ``link_path_walk()``, such as
1185 ``path_lookupat()`` using a loop that calls ``link_path_walk()``, and then
1186 handles the final component. If the final component is a symlink
1187 that needs to be followed, then ``trailing_symlink()`` is called to set
1188 things up properly and the loop repeats, calling ``link_path_walk()``
1189 again. This could loop as many as 40 times if the last component of
1190 each symlink is another symlink.
1192 The various functions that examine the final component and possibly
1193 report that it is a symlink are ``lookup_last()``, ``mountpoint_last()``
1194 and ``do_last()``, each of which use the same convention as
1195 ``walk_component()`` of returning ``1`` if a symlink was found that needs
1198 Of these, ``do_last()`` is the most interesting as it is used for
1199 opening a file. Part of ``do_last()`` runs with ``i_rwsem`` held and this
1200 part is in a separate function: ``lookup_open()``.
1202 Explaining ``do_last()`` completely is beyond the scope of this article,
1203 but a few highlights should help those interested in exploring the
1206 1. Rather than just finding the target file, ``do_last()`` needs to open
1207 it. If the file was found in the dcache, then ``vfs_open()`` is used for
1208 this. If not, then ``lookup_open()`` will either call ``atomic_open()`` (if
1209 the filesystem provides it) to combine the final lookup with the open, or
1210 will perform the separate ``lookup_real()`` and ``vfs_create()`` steps
1211 directly. In the later case the actual "open" of this newly found or
1212 created file will be performed by ``vfs_open()``, just as if the name
1213 were found in the dcache.
1215 2. ``vfs_open()`` can fail with ``-EOPENSTALE`` if the cached information
1216 wasn't quite current enough. Rather than restarting the lookup from
1217 the top with ``LOOKUP_REVAL`` set, ``lookup_open()`` is called instead,
1218 giving the filesystem a chance to resolve small inconsistencies.
1219 If that doesn't work, only then is the lookup restarted from the top.
1221 3. An open with O_CREAT **does** follow a symlink in the final component,
1222 unlike other creation system calls (like ``mkdir``). So the sequence::
1225 echo hello > /tmp/foo
1227 will create a file called ``/tmp/bar``. This is not permitted if
1228 ``O_EXCL`` is set but otherwise is handled for an O_CREAT open much
1229 like for a non-creating open: ``should_follow_link()`` returns ``1``, and
1230 so does ``do_last()`` so that ``trailing_symlink()`` gets called and the
1231 open process continues on the symlink that was found.
1233 Updating the access time
1234 ------------------------
1236 We previously said of RCU-walk that it would "take no locks, increment
1237 no counts, leave no footprints." We have since seen that some
1238 "footprints" can be needed when handling symlinks as a counted
1239 reference (or even a memory allocation) may be needed. But these
1240 footprints are best kept to a minimum.
1242 One other place where walking down a symlink can involve leaving
1243 footprints in a way that doesn't affect directories is in updating access times.
1244 In Unix (and Linux) every filesystem object has a "last accessed
1245 time", or "``atime``". Passing through a directory to access a file
1246 within is not considered to be an access for the purposes of
1247 ``atime``; only listing the contents of a directory can update its ``atime``.
1248 Symlinks are different it seems. Both reading a symlink (with ``readlink()``)
1249 and looking up a symlink on the way to some other destination can
1250 update the atime on that symlink.
1252 .. _clearest statement: http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_08
1254 It is not clear why this is the case; POSIX has little to say on the
1255 subject. The `clearest statement`_ is that, if a particular implementation
1256 updates a timestamp in a place not specified by POSIX, this must be
1257 documented "except that any changes caused by pathname resolution need
1258 not be documented". This seems to imply that POSIX doesn't really
1259 care about access-time updates during pathname lookup.
1261 .. _Linux 1.3.87: https://git.kernel.org/cgit/linux/kernel/git/history/history.git/diff/fs/ext2/symlink.c?id=f806c6db77b8eaa6e00dcfb6b567706feae8dbb8
1263 An examination of history shows that prior to `Linux 1.3.87`_, the ext2
1264 filesystem, at least, didn't update atime when following a link.
1265 Unfortunately we have no record of why that behavior was changed.
1267 In any case, access time must now be updated and that operation can be
1268 quite complex. Trying to stay in RCU-walk while doing it is best
1269 avoided. Fortunately it is often permitted to skip the ``atime``
1270 update. Because ``atime`` updates cause performance problems in various
1271 areas, Linux supports the ``relatime`` mount option, which generally
1272 limits the updates of ``atime`` to once per day on files that aren't
1273 being changed (and symlinks never change once created). Even without
1274 ``relatime``, many filesystems record ``atime`` with a one-second
1275 granularity, so only one update per second is required.
1277 It is easy to test if an ``atime`` update is needed while in RCU-walk
1278 mode and, if it isn't, the update can be skipped and RCU-walk mode
1279 continues. Only when an ``atime`` update is actually required does the
1280 path walk drop down to REF-walk. All of this is handled in the
1281 ``get_link()`` function.
1286 A suitable way to wrap up this tour of pathname walking is to list
1287 the various flags that can be stored in the ``nameidata`` to guide the
1288 lookup process. Many of these are only meaningful on the final
1289 component, others reflect the current state of the pathname lookup.
1290 And then there is ``LOOKUP_EMPTY``, which doesn't fit conceptually with
1291 the others. If this is not set, an empty pathname causes an error
1292 very early on. If it is set, empty pathnames are not considered to be
1298 We have already met two global state flags: ``LOOKUP_RCU`` and
1299 ``LOOKUP_REVAL``. These select between one of three overall approaches
1300 to lookup: RCU-walk, REF-walk, and REF-walk with forced revalidation.
1302 ``LOOKUP_PARENT`` indicates that the final component hasn't been reached
1303 yet. This is primarily used to tell the audit subsystem the full
1304 context of a particular access being audited.
1306 ``LOOKUP_ROOT`` indicates that the ``root`` field in the ``nameidata`` was
1307 provided by the caller, so it shouldn't be released when it is no
1310 ``LOOKUP_JUMPED`` means that the current dentry was chosen not because
1311 it had the right name but for some other reason. This happens when
1312 following "``..``", following a symlink to ``/``, crossing a mount point
1313 or accessing a "``/proc/$PID/fd/$FD``" symlink. In this case the
1314 filesystem has not been asked to revalidate the name (with
1315 ``d_revalidate()``). In such cases the inode may still need to be
1316 revalidated, so ``d_op->d_weak_revalidate()`` is called if
1317 ``LOOKUP_JUMPED`` is set when the look completes - which may be at the
1318 final component or, when creating, unlinking, or renaming, at the penultimate component.
1320 Final-component flags
1321 ~~~~~~~~~~~~~~~~~~~~~
1323 Some of these flags are only set when the final component is being
1324 considered. Others are only checked for when considering that final
1327 ``LOOKUP_AUTOMOUNT`` ensures that, if the final component is an automount
1328 point, then the mount is triggered. Some operations would trigger it
1329 anyway, but operations like ``stat()`` deliberately don't. ``statfs()``
1330 needs to trigger the mount but otherwise behaves a lot like ``stat()``, so
1331 it sets ``LOOKUP_AUTOMOUNT``, as does "``quotactl()``" and the handling of
1334 ``LOOKUP_FOLLOW`` has a similar function to ``LOOKUP_AUTOMOUNT`` but for
1335 symlinks. Some system calls set or clear it implicitly, while
1336 others have API flags such as ``AT_SYMLINK_FOLLOW`` and
1337 ``UMOUNT_NOFOLLOW`` to control it. Its effect is similar to
1338 ``WALK_GET`` that we already met, but it is used in a different way.
1340 ``LOOKUP_DIRECTORY`` insists that the final component is a directory.
1341 Various callers set this and it is also set when the final component
1342 is found to be followed by a slash.
1344 Finally ``LOOKUP_OPEN``, ``LOOKUP_CREATE``, ``LOOKUP_EXCL``, and
1345 ``LOOKUP_RENAME_TARGET`` are not used directly by the VFS but are made
1346 available to the filesystem and particularly the ``->d_revalidate()``
1347 method. A filesystem can choose not to bother revalidating too hard
1348 if it knows that it will be asked to open or create the file soon.
1349 These flags were previously useful for ``->lookup()`` too but with the
1350 introduction of ``->atomic_open()`` they are less relevant there.
1355 Despite its complexity, all this pathname lookup code appears to be
1356 in good shape - various parts are certainly easier to understand now
1357 than even a couple of releases ago. But that doesn't mean it is
1358 "finished". As already mentioned, RCU-walk currently only follows
1359 symlinks that are stored in the inode so, while it handles many ext4
1360 symlinks, it doesn't help with NFS, XFS, or Btrfs. That support
1361 is not likely to be long delayed.