2 Introduction to pathname lookup
3 ===============================
5 The most obvious aspect of pathname lookup, which very little
6 exploration is needed to discover, is that it is complex. There are
7 many rules, special cases, and implementation alternatives that all
8 combine to confuse the unwary reader. Computer science has long been
9 acquainted with such complexity and has tools to help manage it. One
10 tool that we will make extensive use of is "divide and conquer". For
11 the early parts of the analysis we will divide off symlinks - leaving
12 them until the final part. Well before we get to symlinks we have
13 another major division based on the VFS's approach to locking which
14 will allow us to review "REF-walk" and "RCU-walk" separately. But we
15 are getting ahead of ourselves. There are some important low level
16 distinctions we need to clarify first.
18 There are two sorts of ...
19 --------------------------
21 .. _openat: http://man7.org/linux/man-pages/man2/openat.2.html
23 Pathnames (sometimes "file names"), used to identify objects in the
24 filesystem, will be familiar to most readers. They contain two sorts
25 of elements: "slashes" that are sequences of one or more "``/``"
26 characters, and "components" that are sequences of one or more
27 non-"``/``" characters. These form two kinds of paths. Those that
28 start with slashes are "absolute" and start from the filesystem root.
29 The others are "relative" and start from the current directory, or
30 from some other location specified by a file descriptor given to a
31 "``XXXat``" system call such as `openat() <openat_>`_.
33 .. _execveat: http://man7.org/linux/man-pages/man2/execveat.2.html
35 It is tempting to describe the second kind as starting with a
36 component, but that isn't always accurate: a pathname can lack both
37 slashes and components, it can be empty, in other words. This is
38 generally forbidden in POSIX, but some of those "xxx``at``" system calls
39 in Linux permit it when the ``AT_EMPTY_PATH`` flag is given. For
40 example, if you have an open file descriptor on an executable file you
41 can execute it by calling `execveat() <execveat_>`_ passing
42 the file descriptor, an empty path, and the ``AT_EMPTY_PATH`` flag.
44 These paths can be divided into two sections: the final component and
45 everything else. The "everything else" is the easy bit. In all cases
46 it must identify a directory that already exists, otherwise an error
47 such as ``ENOENT`` or ``ENOTDIR`` will be reported.
49 The final component is not so simple. Not only do different system
50 calls interpret it quite differently (e.g. some create it, some do
51 not), but it might not even exist: neither the empty pathname nor the
52 pathname that is just slashes have a final component. If it does
53 exist, it could be "``.``" or "``..``" which are handled quite differently
54 from other components.
56 .. _POSIX: http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_12
58 If a pathname ends with a slash, such as "``/tmp/foo/``" it might be
59 tempting to consider that to have an empty final component. In many
60 ways that would lead to correct results, but not always. In
61 particular, ``mkdir()`` and ``rmdir()`` each create or remove a directory named
62 by the final component, and they are required to work with pathnames
63 ending in "``/``". According to POSIX_
65 A pathname that contains at least one non- <slash> character and
66 that ends with one or more trailing <slash> characters shall not
67 be resolved successfully unless the last pathname component before
68 the trailing <slash> characters names an existing directory or a
69 directory entry that is to be created for a directory immediately
70 after the pathname is resolved.
72 The Linux pathname walking code (mostly in ``fs/namei.c``) deals with
73 all of these issues: breaking the path into components, handling the
74 "everything else" quite separately from the final component, and
75 checking that the trailing slash is not used where it isn't
76 permitted. It also addresses the important issue of concurrent
79 While one process is looking up a pathname, another might be making
80 changes that affect that lookup. One fairly extreme case is that if
81 "a/b" were renamed to "a/c/b" while another process were looking up
82 "a/b/..", that process might successfully resolve on "a/c".
83 Most races are much more subtle, and a big part of the task of
84 pathname lookup is to prevent them from having damaging effects. Many
85 of the possible races are seen most clearly in the context of the
86 "dcache" and an understanding of that is central to understanding
89 More than just a cache
90 ----------------------
92 The "dcache" caches information about names in each filesystem to
93 make them quickly available for lookup. Each entry (known as a
94 "dentry") contains three significant fields: a component name, a
95 pointer to a parent dentry, and a pointer to the "inode" which
96 contains further information about the object in that parent with
97 the given name. The inode pointer can be ``NULL`` indicating that the
98 name doesn't exist in the parent. While there can be linkage in the
99 dentry of a directory to the dentries of the children, that linkage is
100 not used for pathname lookup, and so will not be considered here.
102 The dcache has a number of uses apart from accelerating lookup. One
103 that will be particularly relevant is that it is closely integrated
104 with the mount table that records which filesystem is mounted where.
105 What the mount table actually stores is which dentry is mounted on top
106 of which other dentry.
108 When considering the dcache, we have another of our "two types"
109 distinctions: there are two types of filesystems.
111 Some filesystems ensure that the information in the dcache is always
112 completely accurate (though not necessarily complete). This can allow
113 the VFS to determine if a particular file does or doesn't exist
114 without checking with the filesystem, and means that the VFS can
115 protect the filesystem against certain races and other problems.
116 These are typically "local" filesystems such as ext3, XFS, and Btrfs.
118 Other filesystems don't provide that guarantee because they cannot.
119 These are typically filesystems that are shared across a network,
120 whether remote filesystems like NFS and 9P, or cluster filesystems
121 like ocfs2 or cephfs. These filesystems allow the VFS to revalidate
122 cached information, and must provide their own protection against
123 awkward races. The VFS can detect these filesystems by the
124 ``DCACHE_OP_REVALIDATE`` flag being set in the dentry.
126 REF-walk: simple concurrency management with refcounts and spinlocks
127 --------------------------------------------------------------------
129 With all of those divisions carefully classified, we can now start
130 looking at the actual process of walking along a path. In particular
131 we will start with the handling of the "everything else" part of a
132 pathname, and focus on the "REF-walk" approach to concurrency
133 management. This code is found in the ``link_path_walk()`` function, if
134 you ignore all the places that only run when "``LOOKUP_RCU``"
135 (indicating the use of RCU-walk) is set.
137 .. _Meet the Lockers: https://lwn.net/Articles/453685/
139 REF-walk is fairly heavy-handed with locks and reference counts. Not
140 as heavy-handed as in the old "big kernel lock" days, but certainly not
141 afraid of taking a lock when one is needed. It uses a variety of
142 different concurrency controls. A background understanding of the
143 various primitives is assumed, or can be gleaned from elsewhere such
144 as in `Meet the Lockers`_.
146 The locking mechanisms used by REF-walk include:
151 This uses the lockref primitive to provide both a spinlock and a
152 reference count. The special-sauce of this primitive is that the
153 conceptual sequence "lock; inc_ref; unlock;" can often be performed
154 with a single atomic memory operation.
156 Holding a reference on a dentry ensures that the dentry won't suddenly
157 be freed and used for something else, so the values in various fields
158 will behave as expected. It also protects the ``->d_inode`` reference
159 to the inode to some extent.
161 The association between a dentry and its inode is fairly permanent.
162 For example, when a file is renamed, the dentry and inode move
163 together to the new location. When a file is created the dentry will
164 initially be negative (i.e. ``d_inode`` is ``NULL``), and will be assigned
165 to the new inode as part of the act of creation.
167 When a file is deleted, this can be reflected in the cache either by
168 setting ``d_inode`` to ``NULL``, or by removing it from the hash table
169 (described shortly) used to look up the name in the parent directory.
170 If the dentry is still in use the second option is used as it is
171 perfectly legal to keep using an open file after it has been deleted
172 and having the dentry around helps. If the dentry is not otherwise in
173 use (i.e. if the refcount in ``d_lockref`` is one), only then will
174 ``d_inode`` be set to ``NULL``. Doing it this way is more efficient for a
177 So as long as a counted reference is held to a dentry, a non-``NULL`` ``->d_inode``
178 value will never be changed.
183 ``d_lock`` is a synonym for the spinlock that is part of ``d_lockref`` above.
184 For our purposes, holding this lock protects against the dentry being
185 renamed or unlinked. In particular, its parent (``d_parent``), and its
186 name (``d_name``) cannot be changed, and it cannot be removed from the
189 When looking for a name in a directory, REF-walk takes ``d_lock`` on
190 each candidate dentry that it finds in the hash table and then checks
191 that the parent and name are correct. So it doesn't lock the parent
192 while searching in the cache; it only locks children.
194 When looking for the parent for a given name (to handle "``..``"),
195 REF-walk can take ``d_lock`` to get a stable reference to ``d_parent``,
196 but it first tries a more lightweight approach. As seen in
197 ``dget_parent()``, if a reference can be claimed on the parent, and if
198 subsequently ``d_parent`` can be seen to have not changed, then there is
199 no need to actually take the lock on the child.
204 Looking up a given name in a given directory involves computing a hash
205 from the two values (the name and the dentry of the directory),
206 accessing that slot in a hash table, and searching the linked list
209 When a dentry is renamed, the name and the parent dentry can both
210 change so the hash will almost certainly change too. This would move the
211 dentry to a different chain in the hash table. If a filename search
212 happened to be looking at a dentry that was moved in this way,
213 it might end up continuing the search down the wrong chain,
214 and so miss out on part of the correct chain.
216 The name-lookup process (``d_lookup()``) does _not_ try to prevent this
217 from happening, but only to detect when it happens.
218 ``rename_lock`` is a seqlock that is updated whenever any dentry is
219 renamed. If ``d_lookup`` finds that a rename happened while it
220 unsuccessfully scanned a chain in the hash table, it simply tries
226 ``i_rwsem`` is a read/write semaphore that serializes all changes to a particular
227 directory. This ensures that, for example, an ``unlink()`` and a ``rename()``
228 cannot both happen at the same time. It also keeps the directory
229 stable while the filesystem is asked to look up a name that is not
230 currently in the dcache or, optionally, when the list of entries in a
231 directory is being retrieved with ``readdir()``.
233 This has a complementary role to that of ``d_lock``: ``i_rwsem`` on a
234 directory protects all of the names in that directory, while ``d_lock``
235 on a name protects just one name in a directory. Most changes to the
236 dcache hold ``i_rwsem`` on the relevant directory inode and briefly take
237 ``d_lock`` on one or more the dentries while the change happens. One
238 exception is when idle dentries are removed from the dcache due to
239 memory pressure. This uses ``d_lock``, but ``i_rwsem`` plays no role.
241 The semaphore affects pathname lookup in two distinct ways. Firstly it
242 prevents changes during lookup of a name in a directory. ``walk_component()`` uses
243 ``lookup_fast()`` first which, in turn, checks to see if the name is in the cache,
244 using only ``d_lock`` locking. If the name isn't found, then ``walk_component()``
245 falls back to ``lookup_slow()`` which takes a shared lock on ``i_rwsem``, checks again that
246 the name isn't in the cache, and then calls in to the filesystem to get a
247 definitive answer. A new dentry will be added to the cache regardless of
250 Secondly, when pathname lookup reaches the final component, it will
251 sometimes need to take an exclusive lock on ``i_rwsem`` before performing the last lookup so
252 that the required exclusion can be achieved. How path lookup chooses
253 to take, or not take, ``i_rwsem`` is one of the
254 issues addressed in a subsequent section.
256 If two threads attempt to look up the same name at the same time - a
257 name that is not yet in the dcache - the shared lock on ``i_rwsem`` will
258 not prevent them both adding new dentries with the same name. As this
259 would result in confusion an extra level of interlocking is used,
260 based around a secondary hash table (``in_lookup_hashtable``) and a
261 per-dentry flag bit (``DCACHE_PAR_LOOKUP``).
263 To add a new dentry to the cache while only holding a shared lock on
264 ``i_rwsem``, a thread must call ``d_alloc_parallel()``. This allocates a
265 dentry, stores the required name and parent in it, checks if there
266 is already a matching dentry in the primary or secondary hash
267 tables, and if not, stores the newly allocated dentry in the secondary
268 hash table, with ``DCACHE_PAR_LOOKUP`` set.
270 If a matching dentry was found in the primary hash table then that is
271 returned and the caller can know that it lost a race with some other
272 thread adding the entry. If no matching dentry is found in either
273 cache, the newly allocated dentry is returned and the caller can
274 detect this from the presence of ``DCACHE_PAR_LOOKUP``. In this case it
275 knows that it has won any race and now is responsible for asking the
276 filesystem to perform the lookup and find the matching inode. When
277 the lookup is complete, it must call ``d_lookup_done()`` which clears
278 the flag and does some other house keeping, including removing the
279 dentry from the secondary hash table - it will normally have been
280 added to the primary hash table already. Note that a ``struct
281 waitqueue_head`` is passed to ``d_alloc_parallel()``, and
282 ``d_lookup_done()`` must be called while this ``waitqueue_head`` is still
285 If a matching dentry is found in the secondary hash table,
286 ``d_alloc_parallel()`` has a little more work to do. It first waits for
287 ``DCACHE_PAR_LOOKUP`` to be cleared, using a wait_queue that was passed
288 to the instance of ``d_alloc_parallel()`` that won the race and that
289 will be woken by the call to ``d_lookup_done()``. It then checks to see
290 if the dentry has now been added to the primary hash table. If it
291 has, the dentry is returned and the caller just sees that it lost any
292 race. If it hasn't been added to the primary hash table, the most
293 likely explanation is that some other dentry was added instead using
294 ``d_splice_alias()``. In any case, ``d_alloc_parallel()`` repeats all the
295 look ups from the start and will normally return something from the
301 ``mnt_count`` is a per-CPU reference counter on "``mount``" structures.
302 Per-CPU here means that incrementing the count is cheap as it only
303 uses CPU-local memory, but checking if the count is zero is expensive as
304 it needs to check with every CPU. Taking a ``mnt_count`` reference
305 prevents the mount structure from disappearing as the result of regular
306 unmount operations, but does not prevent a "lazy" unmount. So holding
307 ``mnt_count`` doesn't ensure that the mount remains in the namespace and,
308 in particular, doesn't stabilize the link to the mounted-on dentry. It
309 does, however, ensure that the ``mount`` data structure remains coherent,
310 and it provides a reference to the root dentry of the mounted
311 filesystem. So a reference through ``->mnt_count`` provides a stable
312 reference to the mounted dentry, but not the mounted-on dentry.
317 ``mount_lock`` is a global seqlock, a bit like ``rename_lock``. It can be used to
318 check if any change has been made to any mount points.
320 While walking down the tree (away from the root) this lock is used when
321 crossing a mount point to check that the crossing was safe. That is,
322 the value in the seqlock is read, then the code finds the mount that
323 is mounted on the current directory, if there is one, and increments
324 the ``mnt_count``. Finally the value in ``mount_lock`` is checked against
325 the old value. If there is no change, then the crossing was safe. If there
326 was a change, the ``mnt_count`` is decremented and the whole process is
329 When walking up the tree (towards the root) by following a ".." link,
330 a little more care is needed. In this case the seqlock (which
331 contains both a counter and a spinlock) is fully locked to prevent
332 any changes to any mount points while stepping up. This locking is
333 needed to stabilize the link to the mounted-on dentry, which the
334 refcount on the mount itself doesn't ensure.
339 Finally the global (but extremely lightweight) RCU read lock is held
340 from time to time to ensure certain data structures don't get freed
343 In particular it is held while scanning chains in the dcache hash
344 table, and the mount point hash table.
346 Bringing it together with ``struct nameidata``
347 --------------------------------------------
349 .. _First edition Unix: http://minnie.tuhs.org/cgi-bin/utree.pl?file=V1/u2.s
351 Throughout the process of walking a path, the current status is stored
352 in a ``struct nameidata``, "namei" being the traditional name - dating
353 all the way back to `First Edition Unix`_ - of the function that
354 converts a "name" to an "inode". ``struct nameidata`` contains (among
360 A ``path`` contains a ``struct vfsmount`` (which is
361 embedded in a ``struct mount``) and a ``struct dentry``. Together these
362 record the current status of the walk. They start out referring to the
363 starting point (the current working directory, the root directory, or some other
364 directory identified by a file descriptor), and are updated on each
365 step. A reference through ``d_lockref`` and ``mnt_count`` is always
371 This is a string together with a length (i.e. _not_ ``nul`` terminated)
372 that is the "next" component in the pathname.
377 This is one of ``LAST_NORM``, ``LAST_ROOT``, ``LAST_DOT``, ``LAST_DOTDOT``, or
378 ``LAST_BIND``. The ``last`` field is only valid if the type is
379 ``LAST_NORM``. ``LAST_BIND`` is used when following a symlink and no
380 components of the symlink have been processed yet. Others should be
381 fairly self-explanatory.
386 This is used to hold a reference to the effective root of the
387 filesystem. Often that reference won't be needed, so this field is
388 only assigned the first time it is used, or when a non-standard root
389 is requested. Keeping a reference in the ``nameidata`` ensures that
390 only one root is in effect for the entire path walk, even if it races
391 with a ``chroot()`` system call.
393 The root is needed when either of two conditions holds: (1) either the
394 pathname or a symbolic link starts with a "'/'", or (2) a "``..``"
395 component is being handled, since "``..``" from the root must always stay
396 at the root. The value used is usually the current root directory of
397 the calling process. An alternate root can be provided as when
398 ``sysctl()`` calls ``file_open_root()``, and when NFSv4 or Btrfs call
399 ``mount_subtree()``. In each case a pathname is being looked up in a very
400 specific part of the filesystem, and the lookup must not be allowed to
401 escape that subtree. It works a bit like a local ``chroot()``.
403 Ignoring the handling of symbolic links, we can now describe the
404 "``link_path_walk()``" function, which handles the lookup of everything
405 except the final component as:
407 Given a path (``name``) and a nameidata structure (``nd``), check that the
408 current directory has execute permission and then advance ``name``
409 over one component while updating ``last_type`` and ``last``. If that
410 was the final component, then return, otherwise call
411 ``walk_component()`` and repeat from the top.
413 ``walk_component()`` is even easier. If the component is ``LAST_DOTS``,
414 it calls ``handle_dots()`` which does the necessary locking as already
415 described. If it finds a ``LAST_NORM`` component it first calls
416 "``lookup_fast()``" which only looks in the dcache, but will ask the
417 filesystem to revalidate the result if it is that sort of filesystem.
418 If that doesn't get a good result, it calls "``lookup_slow()``" which
419 takes ``i_rwsem``, rechecks the cache, and then asks the filesystem
420 to find a definitive answer. Each of these will call
421 ``follow_managed()`` (as described below) to handle any mount points.
423 In the absence of symbolic links, ``walk_component()`` creates a new
424 ``struct path`` containing a counted reference to the new dentry and a
425 reference to the new ``vfsmount`` which is only counted if it is
426 different from the previous ``vfsmount``. It then calls
427 ``path_to_nameidata()`` to install the new ``struct path`` in the
428 ``struct nameidata`` and drop the unneeded references.
430 This "hand-over-hand" sequencing of getting a reference to the new
431 dentry before dropping the reference to the previous dentry may
432 seem obvious, but is worth pointing out so that we will recognize its
433 analogue in the "RCU-walk" version.
435 Handling the final component
436 ----------------------------
438 ``link_path_walk()`` only walks as far as setting ``nd->last`` and
439 ``nd->last_type`` to refer to the final component of the path. It does
440 not call ``walk_component()`` that last time. Handling that final
441 component remains for the caller to sort out. Those callers are
442 ``path_lookupat()``, ``path_parentat()``, ``path_mountpoint()`` and
443 ``path_openat()`` each of which handles the differing requirements of
444 different system calls.
446 ``path_parentat()`` is clearly the simplest - it just wraps a little bit
447 of housekeeping around ``link_path_walk()`` and returns the parent
448 directory and final component to the caller. The caller will be either
449 aiming to create a name (via ``filename_create()``) or remove or rename
450 a name (in which case ``user_path_parent()`` is used). They will use
451 ``i_rwsem`` to exclude other changes while they validate and then
452 perform their operation.
454 ``path_lookupat()`` is nearly as simple - it is used when an existing
455 object is wanted such as by ``stat()`` or ``chmod()``. It essentially just
456 calls ``walk_component()`` on the final component through a call to
457 ``lookup_last()``. ``path_lookupat()`` returns just the final dentry.
459 ``path_mountpoint()`` handles the special case of unmounting which must
460 not try to revalidate the mounted filesystem. It effectively
461 contains, through a call to ``mountpoint_last()``, an alternate
462 implementation of ``lookup_slow()`` which skips that step. This is
463 important when unmounting a filesystem that is inaccessible, such as
464 one provided by a dead NFS server.
466 Finally ``path_openat()`` is used for the ``open()`` system call; it
467 contains, in support functions starting with "``do_last()``", all the
468 complexity needed to handle the different subtleties of O_CREAT (with
469 or without O_EXCL), final "``/``" characters, and trailing symbolic
470 links. We will revisit this in the final part of this series, which
471 focuses on those symbolic links. "``do_last()``" will sometimes, but
472 not always, take ``i_rwsem``, depending on what it finds.
474 Each of these, or the functions which call them, need to be alert to
475 the possibility that the final component is not ``LAST_NORM``. If the
476 goal of the lookup is to create something, then any value for
477 ``last_type`` other than ``LAST_NORM`` will result in an error. For
478 example if ``path_parentat()`` reports ``LAST_DOTDOT``, then the caller
479 won't try to create that name. They also check for trailing slashes
480 by testing ``last.name[last.len]``. If there is any character beyond
481 the final component, it must be a trailing slash.
483 Revalidation and automounts
484 ---------------------------
486 Apart from symbolic links, there are only two parts of the "REF-walk"
487 process not yet covered. One is the handling of stale cache entries
488 and the other is automounts.
490 On filesystems that require it, the lookup routines will call the
491 ``->d_revalidate()`` dentry method to ensure that the cached information
492 is current. This will often confirm validity or update a few details
493 from a server. In some cases it may find that there has been change
494 further up the path and that something that was thought to be valid
495 previously isn't really. When this happens the lookup of the whole
496 path is aborted and retried with the "``LOOKUP_REVAL``" flag set. This
497 forces revalidation to be more thorough. We will see more details of
498 this retry process in the next article.
500 Automount points are locations in the filesystem where an attempt to
501 lookup a name can trigger changes to how that lookup should be
502 handled, in particular by mounting a filesystem there. These are
503 covered in greater detail in autofs.txt in the Linux documentation
504 tree, but a few notes specifically related to path lookup are in order
507 The Linux VFS has a concept of "managed" dentries which is reflected
508 in function names such as "``follow_managed()``". There are three
509 potentially interesting things about these dentries corresponding
510 to three different flags that might be set in ``dentry->d_flags``:
512 ``DCACHE_MANAGE_TRANSIT``
513 ~~~~~~~~~~~~~~~~~~~~~~~
515 If this flag has been set, then the filesystem has requested that the
516 ``d_manage()`` dentry operation be called before handling any possible
517 mount point. This can perform two particular services:
519 It can block to avoid races. If an automount point is being
520 unmounted, the ``d_manage()`` function will usually wait for that
521 process to complete before letting the new lookup proceed and possibly
522 trigger a new automount.
524 It can selectively allow only some processes to transit through a
525 mount point. When a server process is managing automounts, it may
526 need to access a directory without triggering normal automount
527 processing. That server process can identify itself to the ``autofs``
528 filesystem, which will then give it a special pass through
529 ``d_manage()`` by returning ``-EISDIR``.
534 This flag is set on every dentry that is mounted on. As Linux
535 supports multiple filesystem namespaces, it is possible that the
536 dentry may not be mounted on in *this* namespace, just in some
537 other. So this flag is seen as a hint, not a promise.
539 If this flag is set, and ``d_manage()`` didn't return ``-EISDIR``,
540 ``lookup_mnt()`` is called to examine the mount hash table (honoring the
541 ``mount_lock`` described earlier) and possibly return a new ``vfsmount``
542 and a new ``dentry`` (both with counted references).
544 ``DCACHE_NEED_AUTOMOUNT``
545 ~~~~~~~~~~~~~~~~~~~~~~~
547 If ``d_manage()`` allowed us to get this far, and ``lookup_mnt()`` didn't
548 find a mount point, then this flag causes the ``d_automount()`` dentry
549 operation to be called.
551 The ``d_automount()`` operation can be arbitrarily complex and may
552 communicate with server processes etc. but it should ultimately either
553 report that there was an error, that there was nothing to mount, or
554 should provide an updated ``struct path`` with new ``dentry`` and ``vfsmount``.
556 In the latter case, ``finish_automount()`` will be called to safely
557 install the new mount point into the mount table.
559 There is no new locking of import here and it is important that no
560 locks (only counted references) are held over this processing due to
561 the very real possibility of extended delays.
562 This will become more important next time when we examine RCU-walk
563 which is particularly sensitive to delays.
565 RCU-walk - faster pathname lookup in Linux
566 ==========================================
568 RCU-walk is another algorithm for performing pathname lookup in Linux.
569 It is in many ways similar to REF-walk and the two share quite a bit
570 of code. The significant difference in RCU-walk is how it allows for
571 the possibility of concurrent access.
573 We noted that REF-walk is complex because there are numerous details
574 and special cases. RCU-walk reduces this complexity by simply
575 refusing to handle a number of cases -- it instead falls back to
576 REF-walk. The difficulty with RCU-walk comes from a different
577 direction: unfamiliarity. The locking rules when depending on RCU are
578 quite different from traditional locking, so we will spend a little extra
579 time when we come to those.
581 Clear demarcation of roles
582 --------------------------
584 The easiest way to manage concurrency is to forcibly stop any other
585 thread from changing the data structures that a given thread is
586 looking at. In cases where no other thread would even think of
587 changing the data and lots of different threads want to read at the
588 same time, this can be very costly. Even when using locks that permit
589 multiple concurrent readers, the simple act of updating the count of
590 the number of current readers can impose an unwanted cost. So the
591 goal when reading a shared data structure that no other process is
592 changing is to avoid writing anything to memory at all. Take no
593 locks, increment no counts, leave no footprints.
595 The REF-walk mechanism already described certainly doesn't follow this
596 principle, but then it is really designed to work when there may well
597 be other threads modifying the data. RCU-walk, in contrast, is
598 designed for the common situation where there are lots of frequent
599 readers and only occasional writers. This may not be common in all
600 parts of the filesystem tree, but in many parts it will be. For the
601 other parts it is important that RCU-walk can quickly fall back to
604 Pathname lookup always starts in RCU-walk mode but only remains there
605 as long as what it is looking for is in the cache and is stable. It
606 dances lightly down the cached filesystem image, leaving no footprints
607 and carefully watching where it is, to be sure it doesn't trip. If it
608 notices that something has changed or is changing, or if something
609 isn't in the cache, then it tries to stop gracefully and switch to
612 This stopping requires getting a counted reference on the current
613 ``vfsmount`` and ``dentry``, and ensuring that these are still valid -
614 that a path walk with REF-walk would have found the same entries.
615 This is an invariant that RCU-walk must guarantee. It can only make
616 decisions, such as selecting the next step, that are decisions which
617 REF-walk could also have made if it were walking down the tree at the
618 same time. If the graceful stop succeeds, the rest of the path is
619 processed with the reliable, if slightly sluggish, REF-walk. If
620 RCU-walk finds it cannot stop gracefully, it simply gives up and
621 restarts from the top with REF-walk.
623 This pattern of "try RCU-walk, if that fails try REF-walk" can be
624 clearly seen in functions like ``filename_lookup()``,
625 ``filename_parentat()``, ``filename_mountpoint()``,
626 ``do_filp_open()``, and ``do_file_open_root()``. These five
627 correspond roughly to the four ``path_``* functions we met earlier,
628 each of which calls ``link_path_walk()``. The ``path_*`` functions are
629 called using different mode flags until a mode is found which works.
630 They are first called with ``LOOKUP_RCU`` set to request "RCU-walk". If
631 that fails with the error ``ECHILD`` they are called again with no
632 special flag to request "REF-walk". If either of those report the
633 error ``ESTALE`` a final attempt is made with ``LOOKUP_REVAL`` set (and no
634 ``LOOKUP_RCU``) to ensure that entries found in the cache are forcibly
635 revalidated - normally entries are only revalidated if the filesystem
636 determines that they are too old to trust.
638 The ``LOOKUP_RCU`` attempt may drop that flag internally and switch to
639 REF-walk, but will never then try to switch back to RCU-walk. Places
640 that trip up RCU-walk are much more likely to be near the leaves and
641 so it is very unlikely that there will be much, if any, benefit from
644 RCU and seqlocks: fast and light
645 --------------------------------
647 RCU is, unsurprisingly, critical to RCU-walk mode. The
648 ``rcu_read_lock()`` is held for the entire time that RCU-walk is walking
649 down a path. The particular guarantee it provides is that the key
650 data structures - dentries, inodes, super_blocks, and mounts - will
651 not be freed while the lock is held. They might be unlinked or
652 invalidated in one way or another, but the memory will not be
653 repurposed so values in various fields will still be meaningful. This
654 is the only guarantee that RCU provides; everything else is done using
657 As we saw above, REF-walk holds a counted reference to the current
658 dentry and the current vfsmount, and does not release those references
659 before taking references to the "next" dentry or vfsmount. It also
660 sometimes takes the ``d_lock`` spinlock. These references and locks are
661 taken to prevent certain changes from happening. RCU-walk must not
662 take those references or locks and so cannot prevent such changes.
663 Instead, it checks to see if a change has been made, and aborts or
666 To preserve the invariant mentioned above (that RCU-walk may only make
667 decisions that REF-walk could have made), it must make the checks at
668 or near the same places that REF-walk holds the references. So, when
669 REF-walk increments a reference count or takes a spinlock, RCU-walk
670 samples the status of a seqlock using ``read_seqcount_begin()`` or a
671 similar function. When REF-walk decrements the count or drops the
672 lock, RCU-walk checks if the sampled status is still valid using
673 ``read_seqcount_retry()`` or similar.
675 However, there is a little bit more to seqlocks than that. If
676 RCU-walk accesses two different fields in a seqlock-protected
677 structure, or accesses the same field twice, there is no a priori
678 guarantee of any consistency between those accesses. When consistency
679 is needed - which it usually is - RCU-walk must take a copy and then
680 use ``read_seqcount_retry()`` to validate that copy.
682 ``read_seqcount_retry()`` not only checks the sequence number, but also
683 imposes a memory barrier so that no memory-read instruction from
684 *before* the call can be delayed until *after* the call, either by the
685 CPU or by the compiler. A simple example of this can be seen in
686 ``slow_dentry_cmp()`` which, for filesystems which do not use simple
687 byte-wise name equality, calls into the filesystem to compare a name
688 against a dentry. The length and name pointer are copied into local
689 variables, then ``read_seqcount_retry()`` is called to confirm the two
690 are consistent, and only then is ``->d_compare()`` called. When
691 standard filename comparison is used, ``dentry_cmp()`` is called
692 instead. Notably it does _not_ use ``read_seqcount_retry()``, but
693 instead has a large comment explaining why the consistency guarantee
694 isn't necessary. A subsequent ``read_seqcount_retry()`` will be
695 sufficient to catch any problem that could occur at this point.
697 With that little refresher on seqlocks out of the way we can look at
698 the bigger picture of how RCU-walk uses seqlocks.
700 ``mount_lock`` and ``nd->m_seq``
701 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
703 We already met the ``mount_lock`` seqlock when REF-walk used it to
704 ensure that crossing a mount point is performed safely. RCU-walk uses
705 it for that too, but for quite a bit more.
707 Instead of taking a counted reference to each ``vfsmount`` as it
708 descends the tree, RCU-walk samples the state of ``mount_lock`` at the
709 start of the walk and stores this initial sequence number in the
710 ``struct nameidata`` in the ``m_seq`` field. This one lock and one
711 sequence number are used to validate all accesses to all ``vfsmounts``,
712 and all mount point crossings. As changes to the mount table are
713 relatively rare, it is reasonable to fall back on REF-walk any time
714 that any "mount" or "unmount" happens.
716 ``m_seq`` is checked (using ``read_seqretry()``) at the end of an RCU-walk
717 sequence, whether switching to REF-walk for the rest of the path or
718 when the end of the path is reached. It is also checked when stepping
719 down over a mount point (in ``__follow_mount_rcu()``) or up (in
720 ``follow_dotdot_rcu()``). If it is ever found to have changed, the
721 whole RCU-walk sequence is aborted and the path is processed again by
724 If RCU-walk finds that ``mount_lock`` hasn't changed then it can be sure
725 that, had REF-walk taken counted references on each vfsmount, the
726 results would have been the same. This ensures the invariant holds,
727 at least for vfsmount structures.
729 ``dentry->d_seq`` and ``nd->seq``
730 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
732 In place of taking a count or lock on ``d_reflock``, RCU-walk samples
733 the per-dentry ``d_seq`` seqlock, and stores the sequence number in the
734 ``seq`` field of the nameidata structure, so ``nd->seq`` should always be
735 the current sequence number of ``nd->dentry``. This number needs to be
736 revalidated after copying, and before using, the name, parent, or
739 The handling of the name we have already looked at, and the parent is
740 only accessed in ``follow_dotdot_rcu()`` which fairly trivially follows
741 the required pattern, though it does so for three different cases.
743 When not at a mount point, ``d_parent`` is followed and its ``d_seq`` is
744 collected. When we are at a mount point, we instead follow the
745 ``mnt->mnt_mountpoint`` link to get a new dentry and collect its
746 ``d_seq``. Then, after finally finding a ``d_parent`` to follow, we must
747 check if we have landed on a mount point and, if so, must find that
748 mount point and follow the ``mnt->mnt_root`` link. This would imply a
749 somewhat unusual, but certainly possible, circumstance where the
750 starting point of the path lookup was in part of the filesystem that
751 was mounted on, and so not visible from the root.
753 The inode pointer, stored in ``->d_inode``, is a little more
754 interesting. The inode will always need to be accessed at least
755 twice, once to determine if it is NULL and once to verify access
756 permissions. Symlink handling requires a validated inode pointer too.
757 Rather than revalidating on each access, a copy is made on the first
758 access and it is stored in the ``inode`` field of ``nameidata`` from where
759 it can be safely accessed without further validation.
761 ``lookup_fast()`` is the only lookup routine that is used in RCU-mode,
762 ``lookup_slow()`` being too slow and requiring locks. It is in
763 ``lookup_fast()`` that we find the important "hand over hand" tracking
764 of the current dentry.
766 The current ``dentry`` and current ``seq`` number are passed to
767 ``__d_lookup_rcu()`` which, on success, returns a new ``dentry`` and a
768 new ``seq`` number. ``lookup_fast()`` then copies the inode pointer and
769 revalidates the new ``seq`` number. It then validates the old ``dentry``
770 with the old ``seq`` number one last time and only then continues. This
771 process of getting the ``seq`` number of the new dentry and then
772 checking the ``seq`` number of the old exactly mirrors the process of
773 getting a counted reference to the new dentry before dropping that for
774 the old dentry which we saw in REF-walk.
776 No ``inode->i_rwsem`` or even ``rename_lock``
777 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
779 A semaphore is a fairly heavyweight lock that can only be taken when it is
780 permissible to sleep. As ``rcu_read_lock()`` forbids sleeping,
781 ``inode->i_rwsem`` plays no role in RCU-walk. If some other thread does
782 take ``i_rwsem`` and modifies the directory in a way that RCU-walk needs
783 to notice, the result will be either that RCU-walk fails to find the
784 dentry that it is looking for, or it will find a dentry which
785 ``read_seqretry()`` won't validate. In either case it will drop down to
786 REF-walk mode which can take whatever locks are needed.
788 Though ``rename_lock`` could be used by RCU-walk as it doesn't require
789 any sleeping, RCU-walk doesn't bother. REF-walk uses ``rename_lock`` to
790 protect against the possibility of hash chains in the dcache changing
791 while they are being searched. This can result in failing to find
792 something that actually is there. When RCU-walk fails to find
793 something in the dentry cache, whether it is really there or not, it
794 already drops down to REF-walk and tries again with appropriate
795 locking. This neatly handles all cases, so adding extra checks on
796 rename_lock would bring no significant value.
798 ``unlazy walk()`` and ``complete_walk()``
799 -------------------------------------
801 That "dropping down to REF-walk" typically involves a call to
802 ``unlazy_walk()``, so named because "RCU-walk" is also sometimes
803 referred to as "lazy walk". ``unlazy_walk()`` is called when
804 following the path down to the current vfsmount/dentry pair seems to
805 have proceeded successfully, but the next step is problematic. This
806 can happen if the next name cannot be found in the dcache, if
807 permission checking or name revalidation couldn't be achieved while
808 the ``rcu_read_lock()`` is held (which forbids sleeping), if an
809 automount point is found, or in a couple of cases involving symlinks.
810 It is also called from ``complete_walk()`` when the lookup has reached
811 the final component, or the very end of the path, depending on which
812 particular flavor of lookup is used.
814 Other reasons for dropping out of RCU-walk that do not trigger a call
815 to ``unlazy_walk()`` are when some inconsistency is found that cannot be
816 handled immediately, such as ``mount_lock`` or one of the ``d_seq``
817 seqlocks reporting a change. In these cases the relevant function
818 will return ``-ECHILD`` which will percolate up until it triggers a new
819 attempt from the top using REF-walk.
821 For those cases where ``unlazy_walk()`` is an option, it essentially
822 takes a reference on each of the pointers that it holds (vfsmount,
823 dentry, and possibly some symbolic links) and then verifies that the
824 relevant seqlocks have not been changed. If there have been changes,
825 it, too, aborts with ``-ECHILD``, otherwise the transition to REF-walk
826 has been a success and the lookup process continues.
828 Taking a reference on those pointers is not quite as simple as just
829 incrementing a counter. That works to take a second reference if you
830 already have one (often indirectly through another object), but it
831 isn't sufficient if you don't actually have a counted reference at
832 all. For ``dentry->d_lockref``, it is safe to increment the reference
833 counter to get a reference unless it has been explicitly marked as
834 "dead" which involves setting the counter to ``-128``.
835 ``lockref_get_not_dead()`` achieves this.
837 For ``mnt->mnt_count`` it is safe to take a reference as long as
838 ``mount_lock`` is then used to validate the reference. If that
839 validation fails, it may *not* be safe to just drop that reference in
840 the standard way of calling ``mnt_put()`` - an unmount may have
841 progressed too far. So the code in ``legitimize_mnt()``, when it
842 finds that the reference it got might not be safe, checks the
843 ``MNT_SYNC_UMOUNT`` flag to determine if a simple ``mnt_put()`` is
844 correct, or if it should just decrement the count and pretend none of
847 Taking care in filesystems
848 --------------------------
850 RCU-walk depends almost entirely on cached information and often will
851 not call into the filesystem at all. However there are two places,
852 besides the already-mentioned component-name comparison, where the
853 file system might be included in RCU-walk, and it must know to be
856 If the filesystem has non-standard permission-checking requirements -
857 such as a networked filesystem which may need to check with the server
858 - the ``i_op->permission`` interface might be called during RCU-walk.
859 In this case an extra "``MAY_NOT_BLOCK``" flag is passed so that it
860 knows not to sleep, but to return ``-ECHILD`` if it cannot complete
861 promptly. ``i_op->permission`` is given the inode pointer, not the
862 dentry, so it doesn't need to worry about further consistency checks.
863 However if it accesses any other filesystem data structures, it must
864 ensure they are safe to be accessed with only the ``rcu_read_lock()``
865 held. This typically means they must be freed using ``kfree_rcu()`` or
868 .. _READ_ONCE: https://lwn.net/Articles/624126/
870 If the filesystem may need to revalidate dcache entries, then
871 ``d_op->d_revalidate`` may be called in RCU-walk too. This interface
872 *is* passed the dentry but does not have access to the ``inode`` or the
873 ``seq`` number from the ``nameidata``, so it needs to be extra careful
874 when accessing fields in the dentry. This "extra care" typically
875 involves using `READ_ONCE() <READ_ONCE_>`_ to access fields, and verifying the
876 result is not NULL before using it. This pattern can be seen in
877 ``nfs_lookup_revalidate()``.
882 In various places in the details of REF-walk and RCU-walk, and also in
883 the big picture, there are a couple of related patterns that are worth
886 The first is "try quickly and check, if that fails try slowly". We
887 can see that in the high-level approach of first trying RCU-walk and
888 then trying REF-walk, and in places where ``unlazy_walk()`` is used to
889 switch to REF-walk for the rest of the path. We also saw it earlier
890 in ``dget_parent()`` when following a "``..``" link. It tries a quick way
891 to get a reference, then falls back to taking locks if needed.
893 The second pattern is "try quickly and check, if that fails try
894 again - repeatedly". This is seen with the use of ``rename_lock`` and
895 ``mount_lock`` in REF-walk. RCU-walk doesn't make use of this pattern -
896 if anything goes wrong it is much safer to just abort and try a more
899 The emphasis here is "try quickly and check". It should probably be
900 "try quickly _and carefully,_ then check". The fact that checking is
901 needed is a reminder that the system is dynamic and only a limited
902 number of things are safe at all. The most likely cause of errors in
903 this whole process is assuming something is safe when in reality it
904 isn't. Careful consideration of what exactly guarantees the safety of
905 each access is sometimes necessary.
907 A walk among the symlinks
908 =========================
910 There are several basic issues that we will examine to understand the
911 handling of symbolic links: the symlink stack, together with cache
912 lifetimes, will help us understand the overall recursive handling of
913 symlinks and lead to the special care needed for the final component.
914 Then a consideration of access-time updates and summary of the various
915 flags controlling lookup will finish the story.
920 There are only two sorts of filesystem objects that can usefully
921 appear in a path prior to the final component: directories and symlinks.
922 Handling directories is quite straightforward: the new directory
923 simply becomes the starting point at which to interpret the next
924 component on the path. Handling symbolic links requires a bit more
927 Conceptually, symbolic links could be handled by editing the path. If
928 a component name refers to a symbolic link, then that component is
929 replaced by the body of the link and, if that body starts with a '/',
930 then all preceding parts of the path are discarded. This is what the
931 "``readlink -f``" command does, though it also edits out "``.``" and
934 Directly editing the path string is not really necessary when looking
935 up a path, and discarding early components is pointless as they aren't
936 looked at anyway. Keeping track of all remaining components is
937 important, but they can of course be kept separately; there is no need
938 to concatenate them. As one symlink may easily refer to another,
939 which in turn can refer to a third, we may need to keep the remaining
940 components of several paths, each to be processed when the preceding
941 ones are completed. These path remnants are kept on a stack of
944 There are two reasons for placing limits on how many symlinks can
945 occur in a single path lookup. The most obvious is to avoid loops.
946 If a symlink referred to itself either directly or through
947 intermediaries, then following the symlink can never complete
948 successfully - the error ``ELOOP`` must be returned. Loops can be
949 detected without imposing limits, but limits are the simplest solution
950 and, given the second reason for restriction, quite sufficient.
952 .. _outlined recently: http://thread.gmane.org/gmane.linux.kernel/1934390/focus=1934550
954 The second reason was `outlined recently`_ by Linus:
956 Because it's a latency and DoS issue too. We need to react well to
957 true loops, but also to "very deep" non-loops. It's not about memory
958 use, it's about users triggering unreasonable CPU resources.
960 Linux imposes a limit on the length of any pathname: ``PATH_MAX``, which
961 is 4096. There are a number of reasons for this limit; not letting the
962 kernel spend too much time on just one path is one of them. With
963 symbolic links you can effectively generate much longer paths so some
964 sort of limit is needed for the same reason. Linux imposes a limit of
965 at most 40 symlinks in any one path lookup. It previously imposed a
966 further limit of eight on the maximum depth of recursion, but that was
967 raised to 40 when a separate stack was implemented, so there is now
970 The ``nameidata`` structure that we met in an earlier article contains a
971 small stack that can be used to store the remaining part of up to two
972 symlinks. In many cases this will be sufficient. If it isn't, a
973 separate stack is allocated with room for 40 symlinks. Pathname
974 lookup will never exceed that stack as, once the 40th symlink is
975 detected, an error is returned.
977 It might seem that the name remnants are all that needs to be stored on
978 this stack, but we need a bit more. To see that, we need to move on to
981 Storage and lifetime of cached symlinks
982 ---------------------------------------
984 Like other filesystem resources, such as inodes and directory
985 entries, symlinks are cached by Linux to avoid repeated costly access
986 to external storage. It is particularly important for RCU-walk to be
987 able to find and temporarily hold onto these cached entries, so that
988 it doesn't need to drop down into REF-walk.
990 .. _object-oriented design pattern: https://lwn.net/Articles/446317/
992 While each filesystem is free to make its own choice, symlinks are
993 typically stored in one of two places. Short symlinks are often
994 stored directly in the inode. When a filesystem allocates a ``struct
995 inode`` it typically allocates extra space to store private data (a
996 common `object-oriented design pattern`_ in the kernel). This will
997 sometimes include space for a symlink. The other common location is
998 in the page cache, which normally stores the content of files. The
999 pathname in a symlink can be seen as the content of that symlink and
1000 can easily be stored in the page cache just like file content.
1002 When neither of these is suitable, the next most likely scenario is
1003 that the filesystem will allocate some temporary memory and copy or
1004 construct the symlink content into that memory whenever it is needed.
1006 When the symlink is stored in the inode, it has the same lifetime as
1007 the inode which, itself, is protected by RCU or by a counted reference
1008 on the dentry. This means that the mechanisms that pathname lookup
1009 uses to access the dcache and icache (inode cache) safely are quite
1010 sufficient for accessing some cached symlinks safely. In these cases,
1011 the ``i_link`` pointer in the inode is set to point to wherever the
1012 symlink is stored and it can be accessed directly whenever needed.
1014 When the symlink is stored in the page cache or elsewhere, the
1015 situation is not so straightforward. A reference on a dentry or even
1016 on an inode does not imply any reference on cached pages of that
1017 inode, and even an ``rcu_read_lock()`` is not sufficient to ensure that
1018 a page will not disappear. So for these symlinks the pathname lookup
1019 code needs to ask the filesystem to provide a stable reference and,
1020 significantly, needs to release that reference when it is finished
1023 Taking a reference to a cache page is often possible even in RCU-walk
1024 mode. It does require making changes to memory, which is best avoided,
1025 but that isn't necessarily a big cost and it is better than dropping
1026 out of RCU-walk mode completely. Even filesystems that allocate
1027 space to copy the symlink into can use ``GFP_ATOMIC`` to often successfully
1028 allocate memory without the need to drop out of RCU-walk. If a
1029 filesystem cannot successfully get a reference in RCU-walk mode, it
1030 must return ``-ECHILD`` and ``unlazy_walk()`` will be called to return to
1031 REF-walk mode in which the filesystem is allowed to sleep.
1033 The place for all this to happen is the ``i_op->follow_link()`` inode
1034 method. In the present mainline code this is never actually called in
1035 RCU-walk mode as the rewrite is not quite complete. It is likely that
1036 in a future release this method will be passed an ``inode`` pointer when
1037 called in RCU-walk mode so it both (1) knows to be careful, and (2) has the
1038 validated pointer. Much like the ``i_op->permission()`` method we
1039 looked at previously, ``->follow_link()`` would need to be careful that
1040 all the data structures it references are safe to be accessed while
1041 holding no counted reference, only the RCU lock. Though getting a
1042 reference with ``->follow_link()`` is not yet done in RCU-walk mode, the
1043 code is ready to release the reference when that does happen.
1045 This need to drop the reference to a symlink adds significant
1046 complexity. It requires a reference to the inode so that the
1047 ``i_op->put_link()`` inode operation can be called. In REF-walk, that
1048 reference is kept implicitly through a reference to the dentry, so
1049 keeping the ``struct path`` of the symlink is easiest. For RCU-walk,
1050 the pointer to the inode is kept separately. To allow switching from
1051 RCU-walk back to REF-walk in the middle of processing nested symlinks
1052 we also need the seq number for the dentry so we can confirm that
1053 switching back was safe.
1055 Finally, when providing a reference to a symlink, the filesystem also
1056 provides an opaque "cookie" that must be passed to ``->put_link()`` so that it
1057 knows what to free. This might be the allocated memory area, or a
1058 pointer to the ``struct page`` in the page cache, or something else
1059 completely. Only the filesystem knows what it is.
1061 In order for the reference to each symlink to be dropped when the walk completes,
1062 whether in RCU-walk or REF-walk, the symlink stack needs to contain,
1063 along with the path remnants:
1065 - the ``struct path`` to provide a reference to the inode in REF-walk
1066 - the ``struct inode *`` to provide a reference to the inode in RCU-walk
1067 - the ``seq`` to allow the path to be safely switched from RCU-walk to REF-walk
1068 - the ``cookie`` that tells ``->put_path()`` what to put.
1070 This means that each entry in the symlink stack needs to hold five
1071 pointers and an integer instead of just one pointer (the path
1072 remnant). On a 64-bit system, this is about 40 bytes per entry;
1073 with 40 entries it adds up to 1600 bytes total, which is less than
1074 half a page. So it might seem like a lot, but is by no means
1077 Note that, in a given stack frame, the path remnant (``name``) is not
1078 part of the symlink that the other fields refer to. It is the remnant
1079 to be followed once that symlink has been fully parsed.
1081 Following the symlink
1082 ---------------------
1084 The main loop in ``link_path_walk()`` iterates seamlessly over all
1085 components in the path and all of the non-final symlinks. As symlinks
1086 are processed, the ``name`` pointer is adjusted to point to a new
1087 symlink, or is restored from the stack, so that much of the loop
1088 doesn't need to notice. Getting this ``name`` variable on and off the
1089 stack is very straightforward; pushing and popping the references is
1090 a little more complex.
1092 When a symlink is found, ``walk_component()`` returns the value ``1``
1093 (``0`` is returned for any other sort of success, and a negative number
1094 is, as usual, an error indicator). This causes ``get_link()`` to be
1095 called; it then gets the link from the filesystem. Providing that
1096 operation is successful, the old path ``name`` is placed on the stack,
1097 and the new value is used as the ``name`` for a while. When the end of
1098 the path is found (i.e. ``*name`` is ``'\0'``) the old ``name`` is restored
1099 off the stack and path walking continues.
1101 Pushing and popping the reference pointers (inode, cookie, etc.) is more
1102 complex in part because of the desire to handle tail recursion. When
1103 the last component of a symlink itself points to a symlink, we
1104 want to pop the symlink-just-completed off the stack before pushing
1105 the symlink-just-found to avoid leaving empty path remnants that would
1106 just get in the way.
1108 It is most convenient to push the new symlink references onto the
1109 stack in ``walk_component()`` immediately when the symlink is found;
1110 ``walk_component()`` is also the last piece of code that needs to look at the
1111 old symlink as it walks that last component. So it is quite
1112 convenient for ``walk_component()`` to release the old symlink and pop
1113 the references just before pushing the reference information for the
1114 new symlink. It is guided in this by two flags; ``WALK_GET``, which
1115 gives it permission to follow a symlink if it finds one, and
1116 ``WALK_PUT``, which tells it to release the current symlink after it has been
1117 followed. ``WALK_PUT`` is tested first, leading to a call to
1118 ``put_link()``. ``WALK_GET`` is tested subsequently (by
1119 ``should_follow_link()``) leading to a call to ``pick_link()`` which sets
1122 Symlinks with no final component
1123 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1125 A pair of special-case symlinks deserve a little further explanation.
1126 Both result in a new ``struct path`` (with mount and dentry) being set
1127 up in the ``nameidata``, and result in ``get_link()`` returning ``NULL``.
1129 The more obvious case is a symlink to "``/``". All symlinks starting
1130 with "``/``" are detected in ``get_link()`` which resets the ``nameidata``
1131 to point to the effective filesystem root. If the symlink only
1132 contains "``/``" then there is nothing more to do, no components at all,
1133 so ``NULL`` is returned to indicate that the symlink can be released and
1134 the stack frame discarded.
1136 The other case involves things in ``/proc`` that look like symlinks but
1139 $ ls -l /proc/self/fd/1
1140 lrwx------ 1 neilb neilb 64 Jun 13 10:19 /proc/self/fd/1 -> /dev/pts/4
1142 Every open file descriptor in any process is represented in ``/proc`` by
1143 something that looks like a symlink. It is really a reference to the
1144 target file, not just the name of it. When you ``readlink`` these
1145 objects you get a name that might refer to the same file - unless it
1146 has been unlinked or mounted over. When ``walk_component()`` follows
1147 one of these, the ``->follow_link()`` method in "procfs" doesn't return
1148 a string name, but instead calls ``nd_jump_link()`` which updates the
1149 ``nameidata`` in place to point to that target. ``->follow_link()`` then
1150 returns ``NULL``. Again there is no final component and ``get_link()``
1151 reports this by leaving the ``last_type`` field of ``nameidata`` as
1154 Following the symlink in the final component
1155 --------------------------------------------
1157 All this leads to ``link_path_walk()`` walking down every component, and
1158 following all symbolic links it finds, until it reaches the final
1159 component. This is just returned in the ``last`` field of ``nameidata``.
1160 For some callers, this is all they need; they want to create that
1161 ``last`` name if it doesn't exist or give an error if it does. Other
1162 callers will want to follow a symlink if one is found, and possibly
1163 apply special handling to the last component of that symlink, rather
1164 than just the last component of the original file name. These callers
1165 potentially need to call ``link_path_walk()`` again and again on
1166 successive symlinks until one is found that doesn't point to another
1169 This case is handled by the relevant caller of ``link_path_walk()``, such as
1170 ``path_lookupat()`` using a loop that calls ``link_path_walk()``, and then
1171 handles the final component. If the final component is a symlink
1172 that needs to be followed, then ``trailing_symlink()`` is called to set
1173 things up properly and the loop repeats, calling ``link_path_walk()``
1174 again. This could loop as many as 40 times if the last component of
1175 each symlink is another symlink.
1177 The various functions that examine the final component and possibly
1178 report that it is a symlink are ``lookup_last()``, ``mountpoint_last()``
1179 and ``do_last()``, each of which use the same convention as
1180 ``walk_component()`` of returning ``1`` if a symlink was found that needs
1183 Of these, ``do_last()`` is the most interesting as it is used for
1184 opening a file. Part of ``do_last()`` runs with ``i_rwsem`` held and this
1185 part is in a separate function: ``lookup_open()``.
1187 Explaining ``do_last()`` completely is beyond the scope of this article,
1188 but a few highlights should help those interested in exploring the
1191 1. Rather than just finding the target file, ``do_last()`` needs to open
1192 it. If the file was found in the dcache, then ``vfs_open()`` is used for
1193 this. If not, then ``lookup_open()`` will either call ``atomic_open()`` (if
1194 the filesystem provides it) to combine the final lookup with the open, or
1195 will perform the separate ``lookup_real()`` and ``vfs_create()`` steps
1196 directly. In the later case the actual "open" of this newly found or
1197 created file will be performed by ``vfs_open()``, just as if the name
1198 were found in the dcache.
1200 2. ``vfs_open()`` can fail with ``-EOPENSTALE`` if the cached information
1201 wasn't quite current enough. Rather than restarting the lookup from
1202 the top with ``LOOKUP_REVAL`` set, ``lookup_open()`` is called instead,
1203 giving the filesystem a chance to resolve small inconsistencies.
1204 If that doesn't work, only then is the lookup restarted from the top.
1206 3. An open with O_CREAT **does** follow a symlink in the final component,
1207 unlike other creation system calls (like ``mkdir``). So the sequence::
1210 echo hello > /tmp/foo
1212 will create a file called ``/tmp/bar``. This is not permitted if
1213 ``O_EXCL`` is set but otherwise is handled for an O_CREAT open much
1214 like for a non-creating open: ``should_follow_link()`` returns ``1``, and
1215 so does ``do_last()`` so that ``trailing_symlink()`` gets called and the
1216 open process continues on the symlink that was found.
1218 Updating the access time
1219 ------------------------
1221 We previously said of RCU-walk that it would "take no locks, increment
1222 no counts, leave no footprints." We have since seen that some
1223 "footprints" can be needed when handling symlinks as a counted
1224 reference (or even a memory allocation) may be needed. But these
1225 footprints are best kept to a minimum.
1227 One other place where walking down a symlink can involve leaving
1228 footprints in a way that doesn't affect directories is in updating access times.
1229 In Unix (and Linux) every filesystem object has a "last accessed
1230 time", or "``atime``". Passing through a directory to access a file
1231 within is not considered to be an access for the purposes of
1232 ``atime``; only listing the contents of a directory can update its ``atime``.
1233 Symlinks are different it seems. Both reading a symlink (with ``readlink()``)
1234 and looking up a symlink on the way to some other destination can
1235 update the atime on that symlink.
1237 .. _clearest statement: http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_08
1239 It is not clear why this is the case; POSIX has little to say on the
1240 subject. The `clearest statement`_ is that, if a particular implementation
1241 updates a timestamp in a place not specified by POSIX, this must be
1242 documented "except that any changes caused by pathname resolution need
1243 not be documented". This seems to imply that POSIX doesn't really
1244 care about access-time updates during pathname lookup.
1246 .. _Linux 1.3.87: https://git.kernel.org/cgit/linux/kernel/git/history/history.git/diff/fs/ext2/symlink.c?id=f806c6db77b8eaa6e00dcfb6b567706feae8dbb8
1248 An examination of history shows that prior to `Linux 1.3.87`_, the ext2
1249 filesystem, at least, didn't update atime when following a link.
1250 Unfortunately we have no record of why that behavior was changed.
1252 In any case, access time must now be updated and that operation can be
1253 quite complex. Trying to stay in RCU-walk while doing it is best
1254 avoided. Fortunately it is often permitted to skip the ``atime``
1255 update. Because ``atime`` updates cause performance problems in various
1256 areas, Linux supports the ``relatime`` mount option, which generally
1257 limits the updates of ``atime`` to once per day on files that aren't
1258 being changed (and symlinks never change once created). Even without
1259 ``relatime``, many filesystems record ``atime`` with a one-second
1260 granularity, so only one update per second is required.
1262 It is easy to test if an ``atime`` update is needed while in RCU-walk
1263 mode and, if it isn't, the update can be skipped and RCU-walk mode
1264 continues. Only when an ``atime`` update is actually required does the
1265 path walk drop down to REF-walk. All of this is handled in the
1266 ``get_link()`` function.
1271 A suitable way to wrap up this tour of pathname walking is to list
1272 the various flags that can be stored in the ``nameidata`` to guide the
1273 lookup process. Many of these are only meaningful on the final
1274 component, others reflect the current state of the pathname lookup.
1275 And then there is ``LOOKUP_EMPTY``, which doesn't fit conceptually with
1276 the others. If this is not set, an empty pathname causes an error
1277 very early on. If it is set, empty pathnames are not considered to be
1283 We have already met two global state flags: ``LOOKUP_RCU`` and
1284 ``LOOKUP_REVAL``. These select between one of three overall approaches
1285 to lookup: RCU-walk, REF-walk, and REF-walk with forced revalidation.
1287 ``LOOKUP_PARENT`` indicates that the final component hasn't been reached
1288 yet. This is primarily used to tell the audit subsystem the full
1289 context of a particular access being audited.
1291 ``LOOKUP_ROOT`` indicates that the ``root`` field in the ``nameidata`` was
1292 provided by the caller, so it shouldn't be released when it is no
1295 ``LOOKUP_JUMPED`` means that the current dentry was chosen not because
1296 it had the right name but for some other reason. This happens when
1297 following "``..``", following a symlink to ``/``, crossing a mount point
1298 or accessing a "``/proc/$PID/fd/$FD``" symlink. In this case the
1299 filesystem has not been asked to revalidate the name (with
1300 ``d_revalidate()``). In such cases the inode may still need to be
1301 revalidated, so ``d_op->d_weak_revalidate()`` is called if
1302 ``LOOKUP_JUMPED`` is set when the look completes - which may be at the
1303 final component or, when creating, unlinking, or renaming, at the penultimate component.
1305 Final-component flags
1306 ~~~~~~~~~~~~~~~~~~~~~
1308 Some of these flags are only set when the final component is being
1309 considered. Others are only checked for when considering that final
1312 ``LOOKUP_AUTOMOUNT`` ensures that, if the final component is an automount
1313 point, then the mount is triggered. Some operations would trigger it
1314 anyway, but operations like ``stat()`` deliberately don't. ``statfs()``
1315 needs to trigger the mount but otherwise behaves a lot like ``stat()``, so
1316 it sets ``LOOKUP_AUTOMOUNT``, as does "``quotactl()``" and the handling of
1319 ``LOOKUP_FOLLOW`` has a similar function to ``LOOKUP_AUTOMOUNT`` but for
1320 symlinks. Some system calls set or clear it implicitly, while
1321 others have API flags such as ``AT_SYMLINK_FOLLOW`` and
1322 ``UMOUNT_NOFOLLOW`` to control it. Its effect is similar to
1323 ``WALK_GET`` that we already met, but it is used in a different way.
1325 ``LOOKUP_DIRECTORY`` insists that the final component is a directory.
1326 Various callers set this and it is also set when the final component
1327 is found to be followed by a slash.
1329 Finally ``LOOKUP_OPEN``, ``LOOKUP_CREATE``, ``LOOKUP_EXCL``, and
1330 ``LOOKUP_RENAME_TARGET`` are not used directly by the VFS but are made
1331 available to the filesystem and particularly the ``->d_revalidate()``
1332 method. A filesystem can choose not to bother revalidating too hard
1333 if it knows that it will be asked to open or create the file soon.
1334 These flags were previously useful for ``->lookup()`` too but with the
1335 introduction of ``->atomic_open()`` they are less relevant there.
1340 Despite its complexity, all this pathname lookup code appears to be
1341 in good shape - various parts are certainly easier to understand now
1342 than even a couple of releases ago. But that doesn't mean it is
1343 "finished". As already mentioned, RCU-walk currently only follows
1344 symlinks that are stored in the inode so, while it handles many ext4
1345 symlinks, it doesn't help with NFS, XFS, or Btrfs. That support
1346 is not likely to be long delayed.
git.samba.org / sfrench / cifs-2.6.git / blob