1===============
2Pathname lookup
3===============
4
5This write-up is based on three articles published at lwn.net:
6
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
10
11Written by Neil Brown with help from Al Viro and Jon Corbet.
12It has subsequently been updated to reflect changes in the kernel
13including:
14
15- per-directory parallel name lookup.
16- ``openat2()`` resolution restriction flags.
17
18Introduction to pathname lookup
19===============================
20
21The most obvious aspect of pathname lookup, which very little
22exploration is needed to discover, is that it is complex.  There are
23many rules, special cases, and implementation alternatives that all
24combine to confuse the unwary reader.  Computer science has long been
25acquainted with such complexity and has tools to help manage it.  One
26tool that we will make extensive use of is "divide and conquer".  For
27the early parts of the analysis we will divide off symlinks - leaving
28them until the final part.  Well before we get to symlinks we have
29another major division based on the VFS's approach to locking which
30will allow us to review "REF-walk" and "RCU-walk" separately.  But we
31are getting ahead of ourselves.  There are some important low level
32distinctions we need to clarify first.
33
34There are two sorts of ...
35--------------------------
36
37.. _openat: http://man7.org/linux/man-pages/man2/openat.2.html
38
39Pathnames (sometimes "file names"), used to identify objects in the
40filesystem, will be familiar to most readers.  They contain two sorts
41of elements: "slashes" that are sequences of one or more "``/``"
42characters, and "components" that are sequences of one or more
43non-"``/``" characters.  These form two kinds of paths.  Those that
44start with slashes are "absolute" and start from the filesystem root.
45The others are "relative" and start from the current directory, or
46from some other location specified by a file descriptor given to
47"``*at()``" system calls such as `openat() <openat_>`_.
48
49.. _execveat: http://man7.org/linux/man-pages/man2/execveat.2.html
50
51It is tempting to describe the second kind as starting with a
52component, but that isn't always accurate: a pathname can lack both
53slashes and components, it can be empty, in other words.  This is
54generally forbidden in POSIX, but some of those "``*at()``" system calls
55in Linux permit it when the ``AT_EMPTY_PATH`` flag is given.  For
56example, if you have an open file descriptor on an executable file you
57can execute it by calling `execveat() <execveat_>`_ passing
58the file descriptor, an empty path, and the ``AT_EMPTY_PATH`` flag.
59
60These paths can be divided into two sections: the final component and
61everything else.  The "everything else" is the easy bit.  In all cases
62it must identify a directory that already exists, otherwise an error
63such as ``ENOENT`` or ``ENOTDIR`` will be reported.
64
65The final component is not so simple.  Not only do different system
66calls interpret it quite differently (e.g. some create it, some do
67not), but it might not even exist: neither the empty pathname nor the
68pathname that is just slashes have a final component.  If it does
69exist, it could be "``.``" or "``..``" which are handled quite differently
70from other components.
71
72.. _POSIX: https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_12
73
74If a pathname ends with a slash, such as "``/tmp/foo/``" it might be
75tempting to consider that to have an empty final component.  In many
76ways that would lead to correct results, but not always.  In
77particular, ``mkdir()`` and ``rmdir()`` each create or remove a directory named
78by the final component, and they are required to work with pathnames
79ending in "``/``".  According to POSIX_:
80
81  A pathname that contains at least one non-<slash> character and
82  that ends with one or more trailing <slash> characters shall not
83  be resolved successfully unless the last pathname component before
84  the trailing <slash> characters names an existing directory or a
85  directory entry that is to be created for a directory immediately
86  after the pathname is resolved.
87
88The Linux pathname walking code (mostly in ``fs/namei.c``) deals with
89all of these issues: breaking the path into components, handling the
90"everything else" quite separately from the final component, and
91checking that the trailing slash is not used where it isn't
92permitted.  It also addresses the important issue of concurrent
93access.
94
95While one process is looking up a pathname, another might be making
96changes that affect that lookup.  One fairly extreme case is that if
97"a/b" were renamed to "a/c/b" while another process were looking up
98"a/b/..", that process might successfully resolve on "a/c".
99Most races are much more subtle, and a big part of the task of
100pathname lookup is to prevent them from having damaging effects.  Many
101of the possible races are seen most clearly in the context of the
102"dcache" and an understanding of that is central to understanding
103pathname lookup.
104
105More than just a cache
106----------------------
107
108The "dcache" caches information about names in each filesystem to
109make them quickly available for lookup.  Each entry (known as a
110"dentry") contains three significant fields: a component name, a
111pointer to a parent dentry, and a pointer to the "inode" which
112contains further information about the object in that parent with
113the given name.  The inode pointer can be ``NULL`` indicating that the
114name doesn't exist in the parent.  While there can be linkage in the
115dentry of a directory to the dentries of the children, that linkage is
116not used for pathname lookup, and so will not be considered here.
117
118The dcache has a number of uses apart from accelerating lookup.  One
119that will be particularly relevant is that it is closely integrated
120with the mount table that records which filesystem is mounted where.
121What the mount table actually stores is which dentry is mounted on top
122of which other dentry.
123
124When considering the dcache, we have another of our "two types"
125distinctions: there are two types of filesystems.
126
127Some filesystems ensure that the information in the dcache is always
128completely accurate (though not necessarily complete).  This can allow
129the VFS to determine if a particular file does or doesn't exist
130without checking with the filesystem, and means that the VFS can
131protect the filesystem against certain races and other problems.
132These are typically "local" filesystems such as ext3, XFS, and Btrfs.
133
134Other filesystems don't provide that guarantee because they cannot.
135These are typically filesystems that are shared across a network,
136whether remote filesystems like NFS and 9P, or cluster filesystems
137like ocfs2 or cephfs.  These filesystems allow the VFS to revalidate
138cached information, and must provide their own protection against
139awkward races.  The VFS can detect these filesystems by the
140``DCACHE_OP_REVALIDATE`` flag being set in the dentry.
141
142REF-walk: simple concurrency management with refcounts and spinlocks
143--------------------------------------------------------------------
144
145With all of those divisions carefully classified, we can now start
146looking at the actual process of walking along a path.  In particular
147we will start with the handling of the "everything else" part of a
148pathname, and focus on the "REF-walk" approach to concurrency
149management.  This code is found in the ``link_path_walk()`` function, if
150you ignore all the places that only run when "``LOOKUP_RCU``"
151(indicating the use of RCU-walk) is set.
152
153.. _Meet the Lockers: https://lwn.net/Articles/453685/
154
155REF-walk is fairly heavy-handed with locks and reference counts.  Not
156as heavy-handed as in the old "big kernel lock" days, but certainly not
157afraid of taking a lock when one is needed.  It uses a variety of
158different concurrency controls.  A background understanding of the
159various primitives is assumed, or can be gleaned from elsewhere such
160as in `Meet the Lockers`_.
161
162The locking mechanisms used by REF-walk include:
163
164dentry->d_lockref
165~~~~~~~~~~~~~~~~~
166
167This uses the lockref primitive to provide both a spinlock and a
168reference count.  The special-sauce of this primitive is that the
169conceptual sequence "lock; inc_ref; unlock;" can often be performed
170with a single atomic memory operation.
171
172Holding a reference on a dentry ensures that the dentry won't suddenly
173be freed and used for something else, so the values in various fields
174will behave as expected.  It also protects the ``->d_inode`` reference
175to the inode to some extent.
176
177The association between a dentry and its inode is fairly permanent.
178For example, when a file is renamed, the dentry and inode move
179together to the new location.  When a file is created the dentry will
180initially be negative (i.e. ``d_inode`` is ``NULL``), and will be assigned
181to the new inode as part of the act of creation.
182
183When a file is deleted, this can be reflected in the cache either by
184setting ``d_inode`` to ``NULL``, or by removing it from the hash table
185(described shortly) used to look up the name in the parent directory.
186If the dentry is still in use the second option is used as it is
187perfectly legal to keep using an open file after it has been deleted
188and having the dentry around helps.  If the dentry is not otherwise in
189use (i.e. if the refcount in ``d_lockref`` is one), only then will
190``d_inode`` be set to ``NULL``.  Doing it this way is more efficient for a
191very common case.
192
193So as long as a counted reference is held to a dentry, a non-``NULL`` ``->d_inode``
194value will never be changed.
195
196dentry->d_lock
197~~~~~~~~~~~~~~
198
199``d_lock`` is a synonym for the spinlock that is part of ``d_lockref`` above.
200For our purposes, holding this lock protects against the dentry being
201renamed or unlinked.  In particular, its parent (``d_parent``), and its
202name (``d_name``) cannot be changed, and it cannot be removed from the
203dentry hash table.
204
205When looking for a name in a directory, REF-walk takes ``d_lock`` on
206each candidate dentry that it finds in the hash table and then checks
207that the parent and name are correct.  So it doesn't lock the parent
208while searching in the cache; it only locks children.
209
210When looking for the parent for a given name (to handle "``..``"),
211REF-walk can take ``d_lock`` to get a stable reference to ``d_parent``,
212but it first tries a more lightweight approach.  As seen in
213``dget_parent()``, if a reference can be claimed on the parent, and if
214subsequently ``d_parent`` can be seen to have not changed, then there is
215no need to actually take the lock on the child.
216
217rename_lock
218~~~~~~~~~~~
219
220Looking up a given name in a given directory involves computing a hash
221from the two values (the name and the dentry of the directory),
222accessing that slot in a hash table, and searching the linked list
223that is found there.
224
225When a dentry is renamed, the name and the parent dentry can both
226change so the hash will almost certainly change too.  This would move the
227dentry to a different chain in the hash table.  If a filename search
228happened to be looking at a dentry that was moved in this way,
229it might end up continuing the search down the wrong chain,
230and so miss out on part of the correct chain.
231
232The name-lookup process (``d_lookup()``) does *not* try to prevent this
233from happening, but only to detect when it happens.
234``rename_lock`` is a seqlock that is updated whenever any dentry is
235renamed.  If ``d_lookup`` finds that a rename happened while it
236unsuccessfully scanned a chain in the hash table, it simply tries
237again.
238
239``rename_lock`` is also used to detect and defend against potential attacks
240against ``LOOKUP_BENEATH`` and ``LOOKUP_IN_ROOT`` when resolving ".." (where
241the parent directory is moved outside the root, bypassing the ``path_equal()``
242check). If ``rename_lock`` is updated during the lookup and the path encounters
243a "..", a potential attack occurred and ``handle_dots()`` will bail out with
244``-EAGAIN``.
245
246inode->i_rwsem
247~~~~~~~~~~~~~~
248
249``i_rwsem`` is a read/write semaphore that serializes all changes to a particular
250directory.  This ensures that, for example, an ``unlink()`` and a ``rename()``
251cannot both happen at the same time.  It also keeps the directory
252stable while the filesystem is asked to look up a name that is not
253currently in the dcache or, optionally, when the list of entries in a
254directory is being retrieved with ``readdir()``.
255
256This has a complementary role to that of ``d_lock``: ``i_rwsem`` on a
257directory protects all of the names in that directory, while ``d_lock``
258on a name protects just one name in a directory.  Most changes to the
259dcache hold ``i_rwsem`` on the relevant directory inode and briefly take
260``d_lock`` on one or more the dentries while the change happens.  One
261exception is when idle dentries are removed from the dcache due to
262memory pressure.  This uses ``d_lock``, but ``i_rwsem`` plays no role.
263
264The semaphore affects pathname lookup in two distinct ways.  Firstly it
265prevents changes during lookup of a name in a directory.  ``walk_component()`` uses
266``lookup_fast()`` first which, in turn, checks to see if the name is in the cache,
267using only ``d_lock`` locking.  If the name isn't found, then ``walk_component()``
268falls back to ``lookup_slow()`` which takes a shared lock on ``i_rwsem``, checks again that
269the name isn't in the cache, and then calls in to the filesystem to get a
270definitive answer.  A new dentry will be added to the cache regardless of
271the result.
272
273Secondly, when pathname lookup reaches the final component, it will
274sometimes need to take an exclusive lock on ``i_rwsem`` before performing the last lookup so
275that the required exclusion can be achieved.  How path lookup chooses
276to take, or not take, ``i_rwsem`` is one of the
277issues addressed in a subsequent section.
278
279If two threads attempt to look up the same name at the same time - a
280name that is not yet in the dcache - the shared lock on ``i_rwsem`` will
281not prevent them both adding new dentries with the same name.  As this
282would result in confusion an extra level of interlocking is used,
283based around a secondary hash table (``in_lookup_hashtable``) and a
284per-dentry flag bit (``DCACHE_PAR_LOOKUP``).
285
286To add a new dentry to the cache while only holding a shared lock on
287``i_rwsem``, a thread must call ``d_alloc_parallel()``.  This allocates a
288dentry, stores the required name and parent in it, checks if there
289is already a matching dentry in the primary or secondary hash
290tables, and if not, stores the newly allocated dentry in the secondary
291hash table, with ``DCACHE_PAR_LOOKUP`` set.
292
293If a matching dentry was found in the primary hash table then that is
294returned and the caller can know that it lost a race with some other
295thread adding the entry.  If no matching dentry is found in either
296cache, the newly allocated dentry is returned and the caller can
297detect this from the presence of ``DCACHE_PAR_LOOKUP``.  In this case it
298knows that it has won any race and now is responsible for asking the
299filesystem to perform the lookup and find the matching inode.  When
300the lookup is complete, it must call ``d_lookup_done()`` which clears
301the flag and does some other house keeping, including removing the
302dentry from the secondary hash table - it will normally have been
303added to the primary hash table already.  Note that a ``struct
304waitqueue_head`` is passed to ``d_alloc_parallel()``, and
305``d_lookup_done()`` must be called while this ``waitqueue_head`` is still
306in scope.
307
308If a matching dentry is found in the secondary hash table,
309``d_alloc_parallel()`` has a little more work to do. It first waits for
310``DCACHE_PAR_LOOKUP`` to be cleared, using a wait_queue that was passed
311to the instance of ``d_alloc_parallel()`` that won the race and that
312will be woken by the call to ``d_lookup_done()``.  It then checks to see
313if the dentry has now been added to the primary hash table.  If it
314has, the dentry is returned and the caller just sees that it lost any
315race.  If it hasn't been added to the primary hash table, the most
316likely explanation is that some other dentry was added instead using
317``d_splice_alias()``.  In any case, ``d_alloc_parallel()`` repeats all the
318look ups from the start and will normally return something from the
319primary hash table.
320
321mnt->mnt_count
322~~~~~~~~~~~~~~
323
324``mnt_count`` is a per-CPU reference counter on "``mount``" structures.
325Per-CPU here means that incrementing the count is cheap as it only
326uses CPU-local memory, but checking if the count is zero is expensive as
327it needs to check with every CPU.  Taking a ``mnt_count`` reference
328prevents the mount structure from disappearing as the result of regular
329unmount operations, but does not prevent a "lazy" unmount.  So holding
330``mnt_count`` doesn't ensure that the mount remains in the namespace and,
331in particular, doesn't stabilize the link to the mounted-on dentry.  It
332does, however, ensure that the ``mount`` data structure remains coherent,
333and it provides a reference to the root dentry of the mounted
334filesystem.  So a reference through ``->mnt_count`` provides a stable
335reference to the mounted dentry, but not the mounted-on dentry.
336
337mount_lock
338~~~~~~~~~~
339
340``mount_lock`` is a global seqlock, a bit like ``rename_lock``.  It can be used to
341check if any change has been made to any mount points.
342
343While walking down the tree (away from the root) this lock is used when
344crossing a mount point to check that the crossing was safe.  That is,
345the value in the seqlock is read, then the code finds the mount that
346is mounted on the current directory, if there is one, and increments
347the ``mnt_count``.  Finally the value in ``mount_lock`` is checked against
348the old value.  If there is no change, then the crossing was safe.  If there
349was a change, the ``mnt_count`` is decremented and the whole process is
350retried.
351
352When walking up the tree (towards the root) by following a ".." link,
353a little more care is needed.  In this case the seqlock (which
354contains both a counter and a spinlock) is fully locked to prevent
355any changes to any mount points while stepping up.  This locking is
356needed to stabilize the link to the mounted-on dentry, which the
357refcount on the mount itself doesn't ensure.
358
359``mount_lock`` is also used to detect and defend against potential attacks
360against ``LOOKUP_BENEATH`` and ``LOOKUP_IN_ROOT`` when resolving ".." (where
361the parent directory is moved outside the root, bypassing the ``path_equal()``
362check). If ``mount_lock`` is updated during the lookup and the path encounters
363a "..", a potential attack occurred and ``handle_dots()`` will bail out with
364``-EAGAIN``.
365
366RCU
367~~~
368
369Finally the global (but extremely lightweight) RCU read lock is held
370from time to time to ensure certain data structures don't get freed
371unexpectedly.
372
373In particular it is held while scanning chains in the dcache hash
374table, and the mount point hash table.
375
376Bringing it together with ``struct nameidata``
377----------------------------------------------
378
379.. _First edition Unix: https://minnie.tuhs.org/cgi-bin/utree.pl?file=V1/u2.s
380
381Throughout the process of walking a path, the current status is stored
382in a ``struct nameidata``, "namei" being the traditional name - dating
383all the way back to `First Edition Unix`_ - of the function that
384converts a "name" to an "inode".  ``struct nameidata`` contains (among
385other fields):
386
387``struct path path``
388~~~~~~~~~~~~~~~~~~~~
389
390A ``path`` contains a ``struct vfsmount`` (which is
391embedded in a ``struct mount``) and a ``struct dentry``.  Together these
392record the current status of the walk.  They start out referring to the
393starting point (the current working directory, the root directory, or some other
394directory identified by a file descriptor), and are updated on each
395step.  A reference through ``d_lockref`` and ``mnt_count`` is always
396held.
397
398``struct qstr last``
399~~~~~~~~~~~~~~~~~~~~
400
401This is a string together with a length (i.e. *not* ``nul`` terminated)
402that is the "next" component in the pathname.
403
404``int last_type``
405~~~~~~~~~~~~~~~~~
406
407This is one of ``LAST_NORM``, ``LAST_ROOT``, ``LAST_DOT`` or ``LAST_DOTDOT``.
408The ``last`` field is only valid if the type is ``LAST_NORM``.
409
410``struct path root``
411~~~~~~~~~~~~~~~~~~~~
412
413This is used to hold a reference to the effective root of the
414filesystem.  Often that reference won't be needed, so this field is
415only assigned the first time it is used, or when a non-standard root
416is requested.  Keeping a reference in the ``nameidata`` ensures that
417only one root is in effect for the entire path walk, even if it races
418with a ``chroot()`` system call.
419
420It should be noted that in the case of ``LOOKUP_IN_ROOT`` or
421``LOOKUP_BENEATH``, the effective root becomes the directory file descriptor
422passed to ``openat2()`` (which exposes these ``LOOKUP_`` flags).
423
424The root is needed when either of two conditions holds: (1) either the
425pathname or a symbolic link starts with a "'/'", or (2) a "``..``"
426component is being handled, since "``..``" from the root must always stay
427at the root.  The value used is usually the current root directory of
428the calling process.  An alternate root can be provided as when
429``sysctl()`` calls ``file_open_root()``, and when NFSv4 or Btrfs call
430``mount_subtree()``.  In each case a pathname is being looked up in a very
431specific part of the filesystem, and the lookup must not be allowed to
432escape that subtree.  It works a bit like a local ``chroot()``.
433
434Ignoring the handling of symbolic links, we can now describe the
435"``link_path_walk()``" function, which handles the lookup of everything
436except the final component as:
437
438   Given a path (``name``) and a nameidata structure (``nd``), check that the
439   current directory has execute permission and then advance ``name``
440   over one component while updating ``last_type`` and ``last``.  If that
441   was the final component, then return, otherwise call
442   ``walk_component()`` and repeat from the top.
443
444``walk_component()`` is even easier.  If the component is ``LAST_DOTS``,
445it calls ``handle_dots()`` which does the necessary locking as already
446described.  If it finds a ``LAST_NORM`` component it first calls
447"``lookup_fast()``" which only looks in the dcache, but will ask the
448filesystem to revalidate the result if it is that sort of filesystem.
449If that doesn't get a good result, it calls "``lookup_slow()``" which
450takes ``i_rwsem``, rechecks the cache, and then asks the filesystem
451to find a definitive answer.  Each of these will call
452``follow_managed()`` (as described below) to handle any mount points.
453
454In the absence of symbolic links, ``walk_component()`` creates a new
455``struct path`` containing a counted reference to the new dentry and a
456reference to the new ``vfsmount`` which is only counted if it is
457different from the previous ``vfsmount``.  It then calls
458``path_to_nameidata()`` to install the new ``struct path`` in the
459``struct nameidata`` and drop the unneeded references.
460
461This "hand-over-hand" sequencing of getting a reference to the new
462dentry before dropping the reference to the previous dentry may
463seem obvious, but is worth pointing out so that we will recognize its
464analogue in the "RCU-walk" version.
465
466Handling the final component
467----------------------------
468
469``link_path_walk()`` only walks as far as setting ``nd->last`` and
470``nd->last_type`` to refer to the final component of the path.  It does
471not call ``walk_component()`` that last time.  Handling that final
472component remains for the caller to sort out. Those callers are
473``path_lookupat()``, ``path_parentat()``, ``path_mountpoint()`` and
474``path_openat()`` each of which handles the differing requirements of
475different system calls.
476
477``path_parentat()`` is clearly the simplest - it just wraps a little bit
478of housekeeping around ``link_path_walk()`` and returns the parent
479directory and final component to the caller.  The caller will be either
480aiming to create a name (via ``filename_create()``) or remove or rename
481a name (in which case ``user_path_parent()`` is used).  They will use
482``i_rwsem`` to exclude other changes while they validate and then
483perform their operation.
484
485``path_lookupat()`` is nearly as simple - it is used when an existing
486object is wanted such as by ``stat()`` or ``chmod()``.  It essentially just
487calls ``walk_component()`` on the final component through a call to
488``lookup_last()``.  ``path_lookupat()`` returns just the final dentry.
489
490``path_mountpoint()`` handles the special case of unmounting which must
491not try to revalidate the mounted filesystem.  It effectively
492contains, through a call to ``mountpoint_last()``, an alternate
493implementation of ``lookup_slow()`` which skips that step.  This is
494important when unmounting a filesystem that is inaccessible, such as
495one provided by a dead NFS server.
496
497Finally ``path_openat()`` is used for the ``open()`` system call; it
498contains, in support functions starting with "``do_last()``", all the
499complexity needed to handle the different subtleties of O_CREAT (with
500or without O_EXCL), final "``/``" characters, and trailing symbolic
501links.  We will revisit this in the final part of this series, which
502focuses on those symbolic links.  "``do_last()``" will sometimes, but
503not always, take ``i_rwsem``, depending on what it finds.
504
505Each of these, or the functions which call them, need to be alert to
506the possibility that the final component is not ``LAST_NORM``.  If the
507goal of the lookup is to create something, then any value for
508``last_type`` other than ``LAST_NORM`` will result in an error.  For
509example if ``path_parentat()`` reports ``LAST_DOTDOT``, then the caller
510won't try to create that name.  They also check for trailing slashes
511by testing ``last.name[last.len]``.  If there is any character beyond
512the final component, it must be a trailing slash.
513
514Revalidation and automounts
515---------------------------
516
517Apart from symbolic links, there are only two parts of the "REF-walk"
518process not yet covered.  One is the handling of stale cache entries
519and the other is automounts.
520
521On filesystems that require it, the lookup routines will call the
522``->d_revalidate()`` dentry method to ensure that the cached information
523is current.  This will often confirm validity or update a few details
524from a server.  In some cases it may find that there has been change
525further up the path and that something that was thought to be valid
526previously isn't really.  When this happens the lookup of the whole
527path is aborted and retried with the "``LOOKUP_REVAL``" flag set.  This
528forces revalidation to be more thorough.  We will see more details of
529this retry process in the next article.
530
531Automount points are locations in the filesystem where an attempt to
532lookup a name can trigger changes to how that lookup should be
533handled, in particular by mounting a filesystem there.  These are
534covered in greater detail in autofs.txt in the Linux documentation
535tree, but a few notes specifically related to path lookup are in order
536here.
537
538The Linux VFS has a concept of "managed" dentries which is reflected
539in function names such as "``follow_managed()``".  There are three
540potentially interesting things about these dentries corresponding
541to three different flags that might be set in ``dentry->d_flags``:
542
543``DCACHE_MANAGE_TRANSIT``
544~~~~~~~~~~~~~~~~~~~~~~~~~
545
546If this flag has been set, then the filesystem has requested that the
547``d_manage()`` dentry operation be called before handling any possible
548mount point.  This can perform two particular services:
549
550It can block to avoid races.  If an automount point is being
551unmounted, the ``d_manage()`` function will usually wait for that
552process to complete before letting the new lookup proceed and possibly
553trigger a new automount.
554
555It can selectively allow only some processes to transit through a
556mount point.  When a server process is managing automounts, it may
557need to access a directory without triggering normal automount
558processing.  That server process can identify itself to the ``autofs``
559filesystem, which will then give it a special pass through
560``d_manage()`` by returning ``-EISDIR``.
561
562``DCACHE_MOUNTED``
563~~~~~~~~~~~~~~~~~~
564
565This flag is set on every dentry that is mounted on.  As Linux
566supports multiple filesystem namespaces, it is possible that the
567dentry may not be mounted on in *this* namespace, just in some
568other.  So this flag is seen as a hint, not a promise.
569
570If this flag is set, and ``d_manage()`` didn't return ``-EISDIR``,
571``lookup_mnt()`` is called to examine the mount hash table (honoring the
572``mount_lock`` described earlier) and possibly return a new ``vfsmount``
573and a new ``dentry`` (both with counted references).
574
575``DCACHE_NEED_AUTOMOUNT``
576~~~~~~~~~~~~~~~~~~~~~~~~~
577
578If ``d_manage()`` allowed us to get this far, and ``lookup_mnt()`` didn't
579find a mount point, then this flag causes the ``d_automount()`` dentry
580operation to be called.
581
582The ``d_automount()`` operation can be arbitrarily complex and may
583communicate with server processes etc. but it should ultimately either
584report that there was an error, that there was nothing to mount, or
585should provide an updated ``struct path`` with new ``dentry`` and ``vfsmount``.
586
587In the latter case, ``finish_automount()`` will be called to safely
588install the new mount point into the mount table.
589
590There is no new locking of import here and it is important that no
591locks (only counted references) are held over this processing due to
592the very real possibility of extended delays.
593This will become more important next time when we examine RCU-walk
594which is particularly sensitive to delays.
595
596RCU-walk - faster pathname lookup in Linux
597==========================================
598
599RCU-walk is another algorithm for performing pathname lookup in Linux.
600It is in many ways similar to REF-walk and the two share quite a bit
601of code.  The significant difference in RCU-walk is how it allows for
602the possibility of concurrent access.
603
604We noted that REF-walk is complex because there are numerous details
605and special cases.  RCU-walk reduces this complexity by simply
606refusing to handle a number of cases -- it instead falls back to
607REF-walk.  The difficulty with RCU-walk comes from a different
608direction: unfamiliarity.  The locking rules when depending on RCU are
609quite different from traditional locking, so we will spend a little extra
610time when we come to those.
611
612Clear demarcation of roles
613--------------------------
614
615The easiest way to manage concurrency is to forcibly stop any other
616thread from changing the data structures that a given thread is
617looking at.  In cases where no other thread would even think of
618changing the data and lots of different threads want to read at the
619same time, this can be very costly.  Even when using locks that permit
620multiple concurrent readers, the simple act of updating the count of
621the number of current readers can impose an unwanted cost.  So the
622goal when reading a shared data structure that no other process is
623changing is to avoid writing anything to memory at all.  Take no
624locks, increment no counts, leave no footprints.
625
626The REF-walk mechanism already described certainly doesn't follow this
627principle, but then it is really designed to work when there may well
628be other threads modifying the data.  RCU-walk, in contrast, is
629designed for the common situation where there are lots of frequent
630readers and only occasional writers.  This may not be common in all
631parts of the filesystem tree, but in many parts it will be.  For the
632other parts it is important that RCU-walk can quickly fall back to
633using REF-walk.
634
635Pathname lookup always starts in RCU-walk mode but only remains there
636as long as what it is looking for is in the cache and is stable.  It
637dances lightly down the cached filesystem image, leaving no footprints
638and carefully watching where it is, to be sure it doesn't trip.  If it
639notices that something has changed or is changing, or if something
640isn't in the cache, then it tries to stop gracefully and switch to
641REF-walk.
642
643This stopping requires getting a counted reference on the current
644``vfsmount`` and ``dentry``, and ensuring that these are still valid -
645that a path walk with REF-walk would have found the same entries.
646This is an invariant that RCU-walk must guarantee.  It can only make
647decisions, such as selecting the next step, that are decisions which
648REF-walk could also have made if it were walking down the tree at the
649same time.  If the graceful stop succeeds, the rest of the path is
650processed with the reliable, if slightly sluggish, REF-walk.  If
651RCU-walk finds it cannot stop gracefully, it simply gives up and
652restarts from the top with REF-walk.
653
654This pattern of "try RCU-walk, if that fails try REF-walk" can be
655clearly seen in functions like ``filename_lookup()``,
656``filename_parentat()``, ``filename_mountpoint()``,
657``do_filp_open()``, and ``do_file_open_root()``.  These five
658correspond roughly to the four ``path_*()`` functions we met earlier,
659each of which calls ``link_path_walk()``.  The ``path_*()`` functions are
660called using different mode flags until a mode is found which works.
661They are first called with ``LOOKUP_RCU`` set to request "RCU-walk".  If
662that fails with the error ``ECHILD`` they are called again with no
663special flag to request "REF-walk".  If either of those report the
664error ``ESTALE`` a final attempt is made with ``LOOKUP_REVAL`` set (and no
665``LOOKUP_RCU``) to ensure that entries found in the cache are forcibly
666revalidated - normally entries are only revalidated if the filesystem
667determines that they are too old to trust.
668
669The ``LOOKUP_RCU`` attempt may drop that flag internally and switch to
670REF-walk, but will never then try to switch back to RCU-walk.  Places
671that trip up RCU-walk are much more likely to be near the leaves and
672so it is very unlikely that there will be much, if any, benefit from
673switching back.
674
675RCU and seqlocks: fast and light
676--------------------------------
677
678RCU is, unsurprisingly, critical to RCU-walk mode.  The
679``rcu_read_lock()`` is held for the entire time that RCU-walk is walking
680down a path.  The particular guarantee it provides is that the key
681data structures - dentries, inodes, super_blocks, and mounts - will
682not be freed while the lock is held.  They might be unlinked or
683invalidated in one way or another, but the memory will not be
684repurposed so values in various fields will still be meaningful.  This
685is the only guarantee that RCU provides; everything else is done using
686seqlocks.
687
688As we saw above, REF-walk holds a counted reference to the current
689dentry and the current vfsmount, and does not release those references
690before taking references to the "next" dentry or vfsmount.  It also
691sometimes takes the ``d_lock`` spinlock.  These references and locks are
692taken to prevent certain changes from happening.  RCU-walk must not
693take those references or locks and so cannot prevent such changes.
694Instead, it checks to see if a change has been made, and aborts or
695retries if it has.
696
697To preserve the invariant mentioned above (that RCU-walk may only make
698decisions that REF-walk could have made), it must make the checks at
699or near the same places that REF-walk holds the references.  So, when
700REF-walk increments a reference count or takes a spinlock, RCU-walk
701samples the status of a seqlock using ``read_seqcount_begin()`` or a
702similar function.  When REF-walk decrements the count or drops the
703lock, RCU-walk checks if the sampled status is still valid using
704``read_seqcount_retry()`` or similar.
705
706However, there is a little bit more to seqlocks than that.  If
707RCU-walk accesses two different fields in a seqlock-protected
708structure, or accesses the same field twice, there is no a priori
709guarantee of any consistency between those accesses.  When consistency
710is needed - which it usually is - RCU-walk must take a copy and then
711use ``read_seqcount_retry()`` to validate that copy.
712
713``read_seqcount_retry()`` not only checks the sequence number, but also
714imposes a memory barrier so that no memory-read instruction from
715*before* the call can be delayed until *after* the call, either by the
716CPU or by the compiler.  A simple example of this can be seen in
717``slow_dentry_cmp()`` which, for filesystems which do not use simple
718byte-wise name equality, calls into the filesystem to compare a name
719against a dentry.  The length and name pointer are copied into local
720variables, then ``read_seqcount_retry()`` is called to confirm the two
721are consistent, and only then is ``->d_compare()`` called.  When
722standard filename comparison is used, ``dentry_cmp()`` is called
723instead.  Notably it does *not* use ``read_seqcount_retry()``, but
724instead has a large comment explaining why the consistency guarantee
725isn't necessary.  A subsequent ``read_seqcount_retry()`` will be
726sufficient to catch any problem that could occur at this point.
727
728With that little refresher on seqlocks out of the way we can look at
729the bigger picture of how RCU-walk uses seqlocks.
730
731``mount_lock`` and ``nd->m_seq``
732~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
733
734We already met the ``mount_lock`` seqlock when REF-walk used it to
735ensure that crossing a mount point is performed safely.  RCU-walk uses
736it for that too, but for quite a bit more.
737
738Instead of taking a counted reference to each ``vfsmount`` as it
739descends the tree, RCU-walk samples the state of ``mount_lock`` at the
740start of the walk and stores this initial sequence number in the
741``struct nameidata`` in the ``m_seq`` field.  This one lock and one
742sequence number are used to validate all accesses to all ``vfsmounts``,
743and all mount point crossings.  As changes to the mount table are
744relatively rare, it is reasonable to fall back on REF-walk any time
745that any "mount" or "unmount" happens.
746
747``m_seq`` is checked (using ``read_seqretry()``) at the end of an RCU-walk
748sequence, whether switching to REF-walk for the rest of the path or
749when the end of the path is reached.  It is also checked when stepping
750down over a mount point (in ``__follow_mount_rcu()``) or up (in
751``follow_dotdot_rcu()``).  If it is ever found to have changed, the
752whole RCU-walk sequence is aborted and the path is processed again by
753REF-walk.
754
755If RCU-walk finds that ``mount_lock`` hasn't changed then it can be sure
756that, had REF-walk taken counted references on each vfsmount, the
757results would have been the same.  This ensures the invariant holds,
758at least for vfsmount structures.
759
760``dentry->d_seq`` and ``nd->seq``
761~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
762
763In place of taking a count or lock on ``d_reflock``, RCU-walk samples
764the per-dentry ``d_seq`` seqlock, and stores the sequence number in the
765``seq`` field of the nameidata structure, so ``nd->seq`` should always be
766the current sequence number of ``nd->dentry``.  This number needs to be
767revalidated after copying, and before using, the name, parent, or
768inode of the dentry.
769
770The handling of the name we have already looked at, and the parent is
771only accessed in ``follow_dotdot_rcu()`` which fairly trivially follows
772the required pattern, though it does so for three different cases.
773
774When not at a mount point, ``d_parent`` is followed and its ``d_seq`` is
775collected.  When we are at a mount point, we instead follow the
776``mnt->mnt_mountpoint`` link to get a new dentry and collect its
777``d_seq``.  Then, after finally finding a ``d_parent`` to follow, we must
778check if we have landed on a mount point and, if so, must find that
779mount point and follow the ``mnt->mnt_root`` link.  This would imply a
780somewhat unusual, but certainly possible, circumstance where the
781starting point of the path lookup was in part of the filesystem that
782was mounted on, and so not visible from the root.
783
784The inode pointer, stored in ``->d_inode``, is a little more
785interesting.  The inode will always need to be accessed at least
786twice, once to determine if it is NULL and once to verify access
787permissions.  Symlink handling requires a validated inode pointer too.
788Rather than revalidating on each access, a copy is made on the first
789access and it is stored in the ``inode`` field of ``nameidata`` from where
790it can be safely accessed without further validation.
791
792``lookup_fast()`` is the only lookup routine that is used in RCU-mode,
793``lookup_slow()`` being too slow and requiring locks.  It is in
794``lookup_fast()`` that we find the important "hand over hand" tracking
795of the current dentry.
796
797The current ``dentry`` and current ``seq`` number are passed to
798``__d_lookup_rcu()`` which, on success, returns a new ``dentry`` and a
799new ``seq`` number.  ``lookup_fast()`` then copies the inode pointer and
800revalidates the new ``seq`` number.  It then validates the old ``dentry``
801with the old ``seq`` number one last time and only then continues.  This
802process of getting the ``seq`` number of the new dentry and then
803checking the ``seq`` number of the old exactly mirrors the process of
804getting a counted reference to the new dentry before dropping that for
805the old dentry which we saw in REF-walk.
806
807No ``inode->i_rwsem`` or even ``rename_lock``
808~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
809
810A semaphore is a fairly heavyweight lock that can only be taken when it is
811permissible to sleep.  As ``rcu_read_lock()`` forbids sleeping,
812``inode->i_rwsem`` plays no role in RCU-walk.  If some other thread does
813take ``i_rwsem`` and modifies the directory in a way that RCU-walk needs
814to notice, the result will be either that RCU-walk fails to find the
815dentry that it is looking for, or it will find a dentry which
816``read_seqretry()`` won't validate.  In either case it will drop down to
817REF-walk mode which can take whatever locks are needed.
818
819Though ``rename_lock`` could be used by RCU-walk as it doesn't require
820any sleeping, RCU-walk doesn't bother.  REF-walk uses ``rename_lock`` to
821protect against the possibility of hash chains in the dcache changing
822while they are being searched.  This can result in failing to find
823something that actually is there.  When RCU-walk fails to find
824something in the dentry cache, whether it is really there or not, it
825already drops down to REF-walk and tries again with appropriate
826locking.  This neatly handles all cases, so adding extra checks on
827rename_lock would bring no significant value.
828
829``unlazy walk()`` and ``complete_walk()``
830-----------------------------------------
831
832That "dropping down to REF-walk" typically involves a call to
833``unlazy_walk()``, so named because "RCU-walk" is also sometimes
834referred to as "lazy walk".  ``unlazy_walk()`` is called when
835following the path down to the current vfsmount/dentry pair seems to
836have proceeded successfully, but the next step is problematic.  This
837can happen if the next name cannot be found in the dcache, if
838permission checking or name revalidation couldn't be achieved while
839the ``rcu_read_lock()`` is held (which forbids sleeping), if an
840automount point is found, or in a couple of cases involving symlinks.
841It is also called from ``complete_walk()`` when the lookup has reached
842the final component, or the very end of the path, depending on which
843particular flavor of lookup is used.
844
845Other reasons for dropping out of RCU-walk that do not trigger a call
846to ``unlazy_walk()`` are when some inconsistency is found that cannot be
847handled immediately, such as ``mount_lock`` or one of the ``d_seq``
848seqlocks reporting a change.  In these cases the relevant function
849will return ``-ECHILD`` which will percolate up until it triggers a new
850attempt from the top using REF-walk.
851
852For those cases where ``unlazy_walk()`` is an option, it essentially
853takes a reference on each of the pointers that it holds (vfsmount,
854dentry, and possibly some symbolic links) and then verifies that the
855relevant seqlocks have not been changed.  If there have been changes,
856it, too, aborts with ``-ECHILD``, otherwise the transition to REF-walk
857has been a success and the lookup process continues.
858
859Taking a reference on those pointers is not quite as simple as just
860incrementing a counter.  That works to take a second reference if you
861already have one (often indirectly through another object), but it
862isn't sufficient if you don't actually have a counted reference at
863all.  For ``dentry->d_lockref``, it is safe to increment the reference
864counter to get a reference unless it has been explicitly marked as
865"dead" which involves setting the counter to ``-128``.
866``lockref_get_not_dead()`` achieves this.
867
868For ``mnt->mnt_count`` it is safe to take a reference as long as
869``mount_lock`` is then used to validate the reference.  If that
870validation fails, it may *not* be safe to just drop that reference in
871the standard way of calling ``mnt_put()`` - an unmount may have
872progressed too far.  So the code in ``legitimize_mnt()``, when it
873finds that the reference it got might not be safe, checks the
874``MNT_SYNC_UMOUNT`` flag to determine if a simple ``mnt_put()`` is
875correct, or if it should just decrement the count and pretend none of
876this ever happened.
877
878Taking care in filesystems
879--------------------------
880
881RCU-walk depends almost entirely on cached information and often will
882not call into the filesystem at all.  However there are two places,
883besides the already-mentioned component-name comparison, where the
884file system might be included in RCU-walk, and it must know to be
885careful.
886
887If the filesystem has non-standard permission-checking requirements -
888such as a networked filesystem which may need to check with the server
889- the ``i_op->permission`` interface might be called during RCU-walk.
890In this case an extra "``MAY_NOT_BLOCK``" flag is passed so that it
891knows not to sleep, but to return ``-ECHILD`` if it cannot complete
892promptly.  ``i_op->permission`` is given the inode pointer, not the
893dentry, so it doesn't need to worry about further consistency checks.
894However if it accesses any other filesystem data structures, it must
895ensure they are safe to be accessed with only the ``rcu_read_lock()``
896held.  This typically means they must be freed using ``kfree_rcu()`` or
897similar.
898
899.. _READ_ONCE: https://lwn.net/Articles/624126/
900
901If the filesystem may need to revalidate dcache entries, then
902``d_op->d_revalidate`` may be called in RCU-walk too.  This interface
903*is* passed the dentry but does not have access to the ``inode`` or the
904``seq`` number from the ``nameidata``, so it needs to be extra careful
905when accessing fields in the dentry.  This "extra care" typically
906involves using  `READ_ONCE() <READ_ONCE_>`_ to access fields, and verifying the
907result is not NULL before using it.  This pattern can be seen in
908``nfs_lookup_revalidate()``.
909
910A pair of patterns
911------------------
912
913In various places in the details of REF-walk and RCU-walk, and also in
914the big picture, there are a couple of related patterns that are worth
915being aware of.
916
917The first is "try quickly and check, if that fails try slowly".  We
918can see that in the high-level approach of first trying RCU-walk and
919then trying REF-walk, and in places where ``unlazy_walk()`` is used to
920switch to REF-walk for the rest of the path.  We also saw it earlier
921in ``dget_parent()`` when following a "``..``" link.  It tries a quick way
922to get a reference, then falls back to taking locks if needed.
923
924The second pattern is "try quickly and check, if that fails try
925again - repeatedly".  This is seen with the use of ``rename_lock`` and
926``mount_lock`` in REF-walk.  RCU-walk doesn't make use of this pattern -
927if anything goes wrong it is much safer to just abort and try a more
928sedate approach.
929
930The emphasis here is "try quickly and check".  It should probably be
931"try quickly *and carefully*, then check".  The fact that checking is
932needed is a reminder that the system is dynamic and only a limited
933number of things are safe at all.  The most likely cause of errors in
934this whole process is assuming something is safe when in reality it
935isn't.  Careful consideration of what exactly guarantees the safety of
936each access is sometimes necessary.
937
938A walk among the symlinks
939=========================
940
941There are several basic issues that we will examine to understand the
942handling of symbolic links:  the symlink stack, together with cache
943lifetimes, will help us understand the overall recursive handling of
944symlinks and lead to the special care needed for the final component.
945Then a consideration of access-time updates and summary of the various
946flags controlling lookup will finish the story.
947
948The symlink stack
949-----------------
950
951There are only two sorts of filesystem objects that can usefully
952appear in a path prior to the final component: directories and symlinks.
953Handling directories is quite straightforward: the new directory
954simply becomes the starting point at which to interpret the next
955component on the path.  Handling symbolic links requires a bit more
956work.
957
958Conceptually, symbolic links could be handled by editing the path.  If
959a component name refers to a symbolic link, then that component is
960replaced by the body of the link and, if that body starts with a '/',
961then all preceding parts of the path are discarded.  This is what the
962"``readlink -f``" command does, though it also edits out "``.``" and
963"``..``" components.
964
965Directly editing the path string is not really necessary when looking
966up a path, and discarding early components is pointless as they aren't
967looked at anyway.  Keeping track of all remaining components is
968important, but they can of course be kept separately; there is no need
969to concatenate them.  As one symlink may easily refer to another,
970which in turn can refer to a third, we may need to keep the remaining
971components of several paths, each to be processed when the preceding
972ones are completed.  These path remnants are kept on a stack of
973limited size.
974
975There are two reasons for placing limits on how many symlinks can
976occur in a single path lookup.  The most obvious is to avoid loops.
977If a symlink referred to itself either directly or through
978intermediaries, then following the symlink can never complete
979successfully - the error ``ELOOP`` must be returned.  Loops can be
980detected without imposing limits, but limits are the simplest solution
981and, given the second reason for restriction, quite sufficient.
982
983.. _outlined recently: http://thread.gmane.org/gmane.linux.kernel/1934390/focus=1934550
984
985The second reason was `outlined recently`_ by Linus:
986
987   Because it's a latency and DoS issue too. We need to react well to
988   true loops, but also to "very deep" non-loops. It's not about memory
989   use, it's about users triggering unreasonable CPU resources.
990
991Linux imposes a limit on the length of any pathname: ``PATH_MAX``, which
992is 4096.  There are a number of reasons for this limit; not letting the
993kernel spend too much time on just one path is one of them.  With
994symbolic links you can effectively generate much longer paths so some
995sort of limit is needed for the same reason.  Linux imposes a limit of
996at most 40 symlinks in any one path lookup.  It previously imposed a
997further limit of eight on the maximum depth of recursion, but that was
998raised to 40 when a separate stack was implemented, so there is now
999just the one limit.
1000
1001The ``nameidata`` structure that we met in an earlier article contains a
1002small stack that can be used to store the remaining part of up to two
1003symlinks.  In many cases this will be sufficient.  If it isn't, a
1004separate stack is allocated with room for 40 symlinks.  Pathname
1005lookup will never exceed that stack as, once the 40th symlink is
1006detected, an error is returned.
1007
1008It might seem that the name remnants are all that needs to be stored on
1009this stack, but we need a bit more.  To see that, we need to move on to
1010cache lifetimes.
1011
1012Storage and lifetime of cached symlinks
1013---------------------------------------
1014
1015Like other filesystem resources, such as inodes and directory
1016entries, symlinks are cached by Linux to avoid repeated costly access
1017to external storage.  It is particularly important for RCU-walk to be
1018able to find and temporarily hold onto these cached entries, so that
1019it doesn't need to drop down into REF-walk.
1020
1021.. _object-oriented design pattern: https://lwn.net/Articles/446317/
1022
1023While each filesystem is free to make its own choice, symlinks are
1024typically stored in one of two places.  Short symlinks are often
1025stored directly in the inode.  When a filesystem allocates a ``struct
1026inode`` it typically allocates extra space to store private data (a
1027common `object-oriented design pattern`_ in the kernel).  This will
1028sometimes include space for a symlink.  The other common location is
1029in the page cache, which normally stores the content of files.  The
1030pathname in a symlink can be seen as the content of that symlink and
1031can easily be stored in the page cache just like file content.
1032
1033When neither of these is suitable, the next most likely scenario is
1034that the filesystem will allocate some temporary memory and copy or
1035construct the symlink content into that memory whenever it is needed.
1036
1037When the symlink is stored in the inode, it has the same lifetime as
1038the inode which, itself, is protected by RCU or by a counted reference
1039on the dentry.  This means that the mechanisms that pathname lookup
1040uses to access the dcache and icache (inode cache) safely are quite
1041sufficient for accessing some cached symlinks safely.  In these cases,
1042the ``i_link`` pointer in the inode is set to point to wherever the
1043symlink is stored and it can be accessed directly whenever needed.
1044
1045When the symlink is stored in the page cache or elsewhere, the
1046situation is not so straightforward.  A reference on a dentry or even
1047on an inode does not imply any reference on cached pages of that
1048inode, and even an ``rcu_read_lock()`` is not sufficient to ensure that
1049a page will not disappear.  So for these symlinks the pathname lookup
1050code needs to ask the filesystem to provide a stable reference and,
1051significantly, needs to release that reference when it is finished
1052with it.
1053
1054Taking a reference to a cache page is often possible even in RCU-walk
1055mode.  It does require making changes to memory, which is best avoided,
1056but that isn't necessarily a big cost and it is better than dropping
1057out of RCU-walk mode completely.  Even filesystems that allocate
1058space to copy the symlink into can use ``GFP_ATOMIC`` to often successfully
1059allocate memory without the need to drop out of RCU-walk.  If a
1060filesystem cannot successfully get a reference in RCU-walk mode, it
1061must return ``-ECHILD`` and ``unlazy_walk()`` will be called to return to
1062REF-walk mode in which the filesystem is allowed to sleep.
1063
1064The place for all this to happen is the ``i_op->follow_link()`` inode
1065method.  In the present mainline code this is never actually called in
1066RCU-walk mode as the rewrite is not quite complete.  It is likely that
1067in a future release this method will be passed an ``inode`` pointer when
1068called in RCU-walk mode so it both (1) knows to be careful, and (2) has the
1069validated pointer.  Much like the ``i_op->permission()`` method we
1070looked at previously, ``->follow_link()`` would need to be careful that
1071all the data structures it references are safe to be accessed while
1072holding no counted reference, only the RCU lock.  Though getting a
1073reference with ``->follow_link()`` is not yet done in RCU-walk mode, the
1074code is ready to release the reference when that does happen.
1075
1076This need to drop the reference to a symlink adds significant
1077complexity.  It requires a reference to the inode so that the
1078``i_op->put_link()`` inode operation can be called.  In REF-walk, that
1079reference is kept implicitly through a reference to the dentry, so
1080keeping the ``struct path`` of the symlink is easiest.  For RCU-walk,
1081the pointer to the inode is kept separately.  To allow switching from
1082RCU-walk back to REF-walk in the middle of processing nested symlinks
1083we also need the seq number for the dentry so we can confirm that
1084switching back was safe.
1085
1086Finally, when providing a reference to a symlink, the filesystem also
1087provides an opaque "cookie" that must be passed to ``->put_link()`` so that it
1088knows what to free.  This might be the allocated memory area, or a
1089pointer to the ``struct page`` in the page cache, or something else
1090completely.  Only the filesystem knows what it is.
1091
1092In order for the reference to each symlink to be dropped when the walk completes,
1093whether in RCU-walk or REF-walk, the symlink stack needs to contain,
1094along with the path remnants:
1095
1096- the ``struct path`` to provide a reference to the inode in REF-walk
1097- the ``struct inode *`` to provide a reference to the inode in RCU-walk
1098- the ``seq`` to allow the path to be safely switched from RCU-walk to REF-walk
1099- the ``cookie`` that tells ``->put_path()`` what to put.
1100
1101This means that each entry in the symlink stack needs to hold five
1102pointers and an integer instead of just one pointer (the path
1103remnant).  On a 64-bit system, this is about 40 bytes per entry;
1104with 40 entries it adds up to 1600 bytes total, which is less than
1105half a page.  So it might seem like a lot, but is by no means
1106excessive.
1107
1108Note that, in a given stack frame, the path remnant (``name``) is not
1109part of the symlink that the other fields refer to.  It is the remnant
1110to be followed once that symlink has been fully parsed.
1111
1112Following the symlink
1113---------------------
1114
1115The main loop in ``link_path_walk()`` iterates seamlessly over all
1116components in the path and all of the non-final symlinks.  As symlinks
1117are processed, the ``name`` pointer is adjusted to point to a new
1118symlink, or is restored from the stack, so that much of the loop
1119doesn't need to notice.  Getting this ``name`` variable on and off the
1120stack is very straightforward; pushing and popping the references is
1121a little more complex.
1122
1123When a symlink is found, ``walk_component()`` returns the value ``1``
1124(``0`` is returned for any other sort of success, and a negative number
1125is, as usual, an error indicator).  This causes ``get_link()`` to be
1126called; it then gets the link from the filesystem.  Providing that
1127operation is successful, the old path ``name`` is placed on the stack,
1128and the new value is used as the ``name`` for a while.  When the end of
1129the path is found (i.e. ``*name`` is ``'\0'``) the old ``name`` is restored
1130off the stack and path walking continues.
1131
1132Pushing and popping the reference pointers (inode, cookie, etc.) is more
1133complex in part because of the desire to handle tail recursion.  When
1134the last component of a symlink itself points to a symlink, we
1135want to pop the symlink-just-completed off the stack before pushing
1136the symlink-just-found to avoid leaving empty path remnants that would
1137just get in the way.
1138
1139It is most convenient to push the new symlink references onto the
1140stack in ``walk_component()`` immediately when the symlink is found;
1141``walk_component()`` is also the last piece of code that needs to look at the
1142old symlink as it walks that last component.  So it is quite
1143convenient for ``walk_component()`` to release the old symlink and pop
1144the references just before pushing the reference information for the
1145new symlink.  It is guided in this by two flags; ``WALK_GET``, which
1146gives it permission to follow a symlink if it finds one, and
1147``WALK_PUT``, which tells it to release the current symlink after it has been
1148followed.  ``WALK_PUT`` is tested first, leading to a call to
1149``put_link()``.  ``WALK_GET`` is tested subsequently (by
1150``should_follow_link()``) leading to a call to ``pick_link()`` which sets
1151up the stack frame.
1152
1153Symlinks with no final component
1154~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1155
1156A pair of special-case symlinks deserve a little further explanation.
1157Both result in a new ``struct path`` (with mount and dentry) being set
1158up in the ``nameidata``, and result in ``get_link()`` returning ``NULL``.
1159
1160The more obvious case is a symlink to "``/``".  All symlinks starting
1161with "``/``" are detected in ``get_link()`` which resets the ``nameidata``
1162to point to the effective filesystem root.  If the symlink only
1163contains "``/``" then there is nothing more to do, no components at all,
1164so ``NULL`` is returned to indicate that the symlink can be released and
1165the stack frame discarded.
1166
1167The other case involves things in ``/proc`` that look like symlinks but
1168aren't really (and are therefore commonly referred to as "magic-links")::
1169
1170     $ ls -l /proc/self/fd/1
1171     lrwx------ 1 neilb neilb 64 Jun 13 10:19 /proc/self/fd/1 -> /dev/pts/4
1172
1173Every open file descriptor in any process is represented in ``/proc`` by
1174something that looks like a symlink.  It is really a reference to the
1175target file, not just the name of it.  When you ``readlink`` these
1176objects you get a name that might refer to the same file - unless it
1177has been unlinked or mounted over.  When ``walk_component()`` follows
1178one of these, the ``->follow_link()`` method in "procfs" doesn't return
1179a string name, but instead calls ``nd_jump_link()`` which updates the
1180``nameidata`` in place to point to that target.  ``->follow_link()`` then
1181returns ``NULL``.  Again there is no final component and ``get_link()``
1182reports this by leaving the ``last_type`` field of ``nameidata`` as
1183``LAST_BIND``.
1184
1185Following the symlink in the final component
1186--------------------------------------------
1187
1188All this leads to ``link_path_walk()`` walking down every component, and
1189following all symbolic links it finds, until it reaches the final
1190component.  This is just returned in the ``last`` field of ``nameidata``.
1191For some callers, this is all they need; they want to create that
1192``last`` name if it doesn't exist or give an error if it does.  Other
1193callers will want to follow a symlink if one is found, and possibly
1194apply special handling to the last component of that symlink, rather
1195than just the last component of the original file name.  These callers
1196potentially need to call ``link_path_walk()`` again and again on
1197successive symlinks until one is found that doesn't point to another
1198symlink.
1199
1200This case is handled by the relevant caller of ``link_path_walk()``, such as
1201``path_lookupat()`` using a loop that calls ``link_path_walk()``, and then
1202handles the final component.  If the final component is a symlink
1203that needs to be followed, then ``trailing_symlink()`` is called to set
1204things up properly and the loop repeats, calling ``link_path_walk()``
1205again.  This could loop as many as 40 times if the last component of
1206each symlink is another symlink.
1207
1208The various functions that examine the final component and possibly
1209report that it is a symlink are ``lookup_last()``, ``mountpoint_last()``
1210and ``do_last()``, each of which use the same convention as
1211``walk_component()`` of returning ``1`` if a symlink was found that needs
1212to be followed.
1213
1214Of these, ``do_last()`` is the most interesting as it is used for
1215opening a file.  Part of ``do_last()`` runs with ``i_rwsem`` held and this
1216part is in a separate function: ``lookup_open()``.
1217
1218Explaining ``do_last()`` completely is beyond the scope of this article,
1219but a few highlights should help those interested in exploring the
1220code.
1221
12221. Rather than just finding the target file, ``do_last()`` needs to open
1223   it.  If the file was found in the dcache, then ``vfs_open()`` is used for
1224   this.  If not, then ``lookup_open()`` will either call ``atomic_open()`` (if
1225   the filesystem provides it) to combine the final lookup with the open, or
1226   will perform the separate ``lookup_real()`` and ``vfs_create()`` steps
1227   directly.  In the later case the actual "open" of this newly found or
1228   created file will be performed by ``vfs_open()``, just as if the name
1229   were found in the dcache.
1230
12312. ``vfs_open()`` can fail with ``-EOPENSTALE`` if the cached information
1232   wasn't quite current enough.  Rather than restarting the lookup from
1233   the top with ``LOOKUP_REVAL`` set, ``lookup_open()`` is called instead,
1234   giving the filesystem a chance to resolve small inconsistencies.
1235   If that doesn't work, only then is the lookup restarted from the top.
1236
12373. An open with O_CREAT **does** follow a symlink in the final component,
1238   unlike other creation system calls (like ``mkdir``).  So the sequence::
1239
1240          ln -s bar /tmp/foo
1241          echo hello > /tmp/foo
1242
1243   will create a file called ``/tmp/bar``.  This is not permitted if
1244   ``O_EXCL`` is set but otherwise is handled for an O_CREAT open much
1245   like for a non-creating open: ``should_follow_link()`` returns ``1``, and
1246   so does ``do_last()`` so that ``trailing_symlink()`` gets called and the
1247   open process continues on the symlink that was found.
1248
1249Updating the access time
1250------------------------
1251
1252We previously said of RCU-walk that it would "take no locks, increment
1253no counts, leave no footprints."  We have since seen that some
1254"footprints" can be needed when handling symlinks as a counted
1255reference (or even a memory allocation) may be needed.  But these
1256footprints are best kept to a minimum.
1257
1258One other place where walking down a symlink can involve leaving
1259footprints in a way that doesn't affect directories is in updating access times.
1260In Unix (and Linux) every filesystem object has a "last accessed
1261time", or "``atime``".  Passing through a directory to access a file
1262within is not considered to be an access for the purposes of
1263``atime``; only listing the contents of a directory can update its ``atime``.
1264Symlinks are different it seems.  Both reading a symlink (with ``readlink()``)
1265and looking up a symlink on the way to some other destination can
1266update the atime on that symlink.
1267
1268.. _clearest statement: https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_08
1269
1270It is not clear why this is the case; POSIX has little to say on the
1271subject.  The `clearest statement`_ is that, if a particular implementation
1272updates a timestamp in a place not specified by POSIX, this must be
1273documented "except that any changes caused by pathname resolution need
1274not be documented".  This seems to imply that POSIX doesn't really
1275care about access-time updates during pathname lookup.
1276
1277.. _Linux 1.3.87: https://git.kernel.org/cgit/linux/kernel/git/history/history.git/diff/fs/ext2/symlink.c?id=f806c6db77b8eaa6e00dcfb6b567706feae8dbb8
1278
1279An examination of history shows that prior to `Linux 1.3.87`_, the ext2
1280filesystem, at least, didn't update atime when following a link.
1281Unfortunately we have no record of why that behavior was changed.
1282
1283In any case, access time must now be updated and that operation can be
1284quite complex.  Trying to stay in RCU-walk while doing it is best
1285avoided.  Fortunately it is often permitted to skip the ``atime``
1286update.  Because ``atime`` updates cause performance problems in various
1287areas, Linux supports the ``relatime`` mount option, which generally
1288limits the updates of ``atime`` to once per day on files that aren't
1289being changed (and symlinks never change once created).  Even without
1290``relatime``, many filesystems record ``atime`` with a one-second
1291granularity, so only one update per second is required.
1292
1293It is easy to test if an ``atime`` update is needed while in RCU-walk
1294mode and, if it isn't, the update can be skipped and RCU-walk mode
1295continues.  Only when an ``atime`` update is actually required does the
1296path walk drop down to REF-walk.  All of this is handled in the
1297``get_link()`` function.
1298
1299A few flags
1300-----------
1301
1302A suitable way to wrap up this tour of pathname walking is to list
1303the various flags that can be stored in the ``nameidata`` to guide the
1304lookup process.  Many of these are only meaningful on the final
1305component, others reflect the current state of the pathname lookup, and some
1306apply restrictions to all path components encountered in the path lookup.
1307
1308And then there is ``LOOKUP_EMPTY``, which doesn't fit conceptually with
1309the others.  If this is not set, an empty pathname causes an error
1310very early on.  If it is set, empty pathnames are not considered to be
1311an error.
1312
1313Global state flags
1314~~~~~~~~~~~~~~~~~~
1315
1316We have already met two global state flags: ``LOOKUP_RCU`` and
1317``LOOKUP_REVAL``.  These select between one of three overall approaches
1318to lookup: RCU-walk, REF-walk, and REF-walk with forced revalidation.
1319
1320``LOOKUP_PARENT`` indicates that the final component hasn't been reached
1321yet.  This is primarily used to tell the audit subsystem the full
1322context of a particular access being audited.
1323
1324``LOOKUP_ROOT`` indicates that the ``root`` field in the ``nameidata`` was
1325provided by the caller, so it shouldn't be released when it is no
1326longer needed.
1327
1328``LOOKUP_JUMPED`` means that the current dentry was chosen not because
1329it had the right name but for some other reason.  This happens when
1330following "``..``", following a symlink to ``/``, crossing a mount point
1331or accessing a "``/proc/$PID/fd/$FD``" symlink (also known as a "magic
1332link"). In this case the filesystem has not been asked to revalidate the
1333name (with ``d_revalidate()``).  In such cases the inode may still need
1334to be revalidated, so ``d_op->d_weak_revalidate()`` is called if
1335``LOOKUP_JUMPED`` is set when the look completes - which may be at the
1336final component or, when creating, unlinking, or renaming, at the penultimate component.
1337
1338Resolution-restriction flags
1339~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1340
1341In order to allow userspace to protect itself against certain race conditions
1342and attack scenarios involving changing path components, a series of flags are
1343available which apply restrictions to all path components encountered during
1344path lookup. These flags are exposed through ``openat2()``'s ``resolve`` field.
1345
1346``LOOKUP_NO_SYMLINKS`` blocks all symlink traversals (including magic-links).
1347This is distinctly different from ``LOOKUP_FOLLOW``, because the latter only
1348relates to restricting the following of trailing symlinks.
1349
1350``LOOKUP_NO_MAGICLINKS`` blocks all magic-link traversals. Filesystems must
1351ensure that they return errors from ``nd_jump_link()``, because that is how
1352``LOOKUP_NO_MAGICLINKS`` and other magic-link restrictions are implemented.
1353
1354``LOOKUP_NO_XDEV`` blocks all ``vfsmount`` traversals (this includes both
1355bind-mounts and ordinary mounts). Note that the ``vfsmount`` which contains the
1356lookup is determined by the first mountpoint the path lookup reaches --
1357absolute paths start with the ``vfsmount`` of ``/``, and relative paths start
1358with the ``dfd``'s ``vfsmount``. Magic-links are only permitted if the
1359``vfsmount`` of the path is unchanged.
1360
1361``LOOKUP_BENEATH`` blocks any path components which resolve outside the
1362starting point of the resolution. This is done by blocking ``nd_jump_root()``
1363as well as blocking ".." if it would jump outside the starting point.
1364``rename_lock`` and ``mount_lock`` are used to detect attacks against the
1365resolution of "..". Magic-links are also blocked.
1366
1367``LOOKUP_IN_ROOT`` resolves all path components as though the starting point
1368were the filesystem root. ``nd_jump_root()`` brings the resolution back to
1369the starting point, and ".." at the starting point will act as a no-op. As with
1370``LOOKUP_BENEATH``, ``rename_lock`` and ``mount_lock`` are used to detect
1371attacks against ".." resolution. Magic-links are also blocked.
1372
1373Final-component flags
1374~~~~~~~~~~~~~~~~~~~~~
1375
1376Some of these flags are only set when the final component is being
1377considered.  Others are only checked for when considering that final
1378component.
1379
1380``LOOKUP_AUTOMOUNT`` ensures that, if the final component is an automount
1381point, then the mount is triggered.  Some operations would trigger it
1382anyway, but operations like ``stat()`` deliberately don't.  ``statfs()``
1383needs to trigger the mount but otherwise behaves a lot like ``stat()``, so
1384it sets ``LOOKUP_AUTOMOUNT``, as does "``quotactl()``" and the handling of
1385"``mount --bind``".
1386
1387``LOOKUP_FOLLOW`` has a similar function to ``LOOKUP_AUTOMOUNT`` but for
1388symlinks.  Some system calls set or clear it implicitly, while
1389others have API flags such as ``AT_SYMLINK_FOLLOW`` and
1390``UMOUNT_NOFOLLOW`` to control it.  Its effect is similar to
1391``WALK_GET`` that we already met, but it is used in a different way.
1392
1393``LOOKUP_DIRECTORY`` insists that the final component is a directory.
1394Various callers set this and it is also set when the final component
1395is found to be followed by a slash.
1396
1397Finally ``LOOKUP_OPEN``, ``LOOKUP_CREATE``, ``LOOKUP_EXCL``, and
1398``LOOKUP_RENAME_TARGET`` are not used directly by the VFS but are made
1399available to the filesystem and particularly the ``->d_revalidate()``
1400method.  A filesystem can choose not to bother revalidating too hard
1401if it knows that it will be asked to open or create the file soon.
1402These flags were previously useful for ``->lookup()`` too but with the
1403introduction of ``->atomic_open()`` they are less relevant there.
1404
1405End of the road
1406---------------
1407
1408Despite its complexity, all this pathname lookup code appears to be
1409in good shape - various parts are certainly easier to understand now
1410than even a couple of releases ago.  But that doesn't mean it is
1411"finished".   As already mentioned, RCU-walk currently only follows
1412symlinks that are stored in the inode so, while it handles many ext4
1413symlinks, it doesn't help with NFS, XFS, or Btrfs.  That support
1414is not likely to be long delayed.
1415