1
2Introduction to pathname lookup
3===============================
4
5The most obvious aspect of pathname lookup, which very little
6exploration is needed to discover, is that it is complex.  There are
7many rules, special cases, and implementation alternatives that all
8combine to confuse the unwary reader.  Computer science has long been
9acquainted with such complexity and has tools to help manage it.  One
10tool that we will make extensive use of is "divide and conquer".  For
11the early parts of the analysis we will divide off symlinks - leaving
12them until the final part.  Well before we get to symlinks we have
13another major division based on the VFS's approach to locking which
14will allow us to review "REF-walk" and "RCU-walk" separately.  But we
15are getting ahead of ourselves.  There are some important low level
16distinctions we need to clarify first.
17
18There are two sorts of ...
19--------------------------
20
21.. _openat: http://man7.org/linux/man-pages/man2/openat.2.html
22
23Pathnames (sometimes "file names"), used to identify objects in the
24filesystem, will be familiar to most readers.  They contain two sorts
25of elements: "slashes" that are sequences of one or more "``/``"
26characters, and "components" that are sequences of one or more
27non-"``/``" characters.  These form two kinds of paths.  Those that
28start with slashes are "absolute" and start from the filesystem root.
29The others are "relative" and start from the current directory, or
30from some other location specified by a file descriptor given to a
31"``XXXat``" system call such as `openat() <openat_>`_.
32
33.. _execveat: http://man7.org/linux/man-pages/man2/execveat.2.html
34
35It is tempting to describe the second kind as starting with a
36component, but that isn't always accurate: a pathname can lack both
37slashes and components, it can be empty, in other words.  This is
38generally forbidden in POSIX, but some of those "xxx``at``" system calls
39in Linux permit it when the ``AT_EMPTY_PATH`` flag is given.  For
40example, if you have an open file descriptor on an executable file you
41can execute it by calling `execveat() <execveat_>`_ passing
42the file descriptor, an empty path, and the ``AT_EMPTY_PATH`` flag.
43
44These paths can be divided into two sections: the final component and
45everything else.  The "everything else" is the easy bit.  In all cases
46it must identify a directory that already exists, otherwise an error
47such as ``ENOENT`` or ``ENOTDIR`` will be reported.
48
49The final component is not so simple.  Not only do different system
50calls interpret it quite differently (e.g. some create it, some do
51not), but it might not even exist: neither the empty pathname nor the
52pathname that is just slashes have a final component.  If it does
53exist, it could be "``.``" or "``..``" which are handled quite differently
54from other components.
55
56.. _POSIX: http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_12
57
58If a pathname ends with a slash, such as "``/tmp/foo/``" it might be
59tempting to consider that to have an empty final component.  In many
60ways that would lead to correct results, but not always.  In
61particular, ``mkdir()`` and ``rmdir()`` each create or remove a directory named
62by the final component, and they are required to work with pathnames
63ending in "``/``".  According to POSIX_
64
65  A pathname that contains at least one non- &lt;slash> character and
66  that ends with one or more trailing &lt;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.
71
72The Linux pathname walking code (mostly in ``fs/namei.c``) deals with
73all of these issues: breaking the path into components, handling the
74"everything else" quite separately from the final component, and
75checking that the trailing slash is not used where it isn't
76permitted.  It also addresses the important issue of concurrent
77access.
78
79While one process is looking up a pathname, another might be making
80changes 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".
83Most races are much more subtle, and a big part of the task of
84pathname lookup is to prevent them from having damaging effects.  Many
85of the possible races are seen most clearly in the context of the
86"dcache" and an understanding of that is central to understanding
87pathname lookup.
88
89More than just a cache
90----------------------
91
92The "dcache" caches information about names in each filesystem to
93make them quickly available for lookup.  Each entry (known as a
94"dentry") contains three significant fields: a component name, a
95pointer to a parent dentry, and a pointer to the "inode" which
96contains further information about the object in that parent with
97the given name.  The inode pointer can be ``NULL`` indicating that the
98name doesn't exist in the parent.  While there can be linkage in the
99dentry of a directory to the dentries of the children, that linkage is
100not used for pathname lookup, and so will not be considered here.
101
102The dcache has a number of uses apart from accelerating lookup.  One
103that will be particularly relevant is that it is closely integrated
104with the mount table that records which filesystem is mounted where.
105What the mount table actually stores is which dentry is mounted on top
106of which other dentry.
107
108When considering the dcache, we have another of our "two types"
109distinctions: there are two types of filesystems.
110
111Some filesystems ensure that the information in the dcache is always
112completely accurate (though not necessarily complete).  This can allow
113the VFS to determine if a particular file does or doesn't exist
114without checking with the filesystem, and means that the VFS can
115protect the filesystem against certain races and other problems.
116These are typically "local" filesystems such as ext3, XFS, and Btrfs.
117
118Other filesystems don't provide that guarantee because they cannot.
119These are typically filesystems that are shared across a network,
120whether remote filesystems like NFS and 9P, or cluster filesystems
121like ocfs2 or cephfs.  These filesystems allow the VFS to revalidate
122cached information, and must provide their own protection against
123awkward races.  The VFS can detect these filesystems by the
124``DCACHE_OP_REVALIDATE`` flag being set in the dentry.
125
126REF-walk: simple concurrency management with refcounts and spinlocks
127--------------------------------------------------------------------
128
129With all of those divisions carefully classified, we can now start
130looking at the actual process of walking along a path.  In particular
131we will start with the handling of the "everything else" part of a
132pathname, and focus on the "REF-walk" approach to concurrency
133management.  This code is found in the ``link_path_walk()`` function, if
134you ignore all the places that only run when "``LOOKUP_RCU``"
135(indicating the use of RCU-walk) is set.
136
137.. _Meet the Lockers: https://lwn.net/Articles/453685/
138
139REF-walk is fairly heavy-handed with locks and reference counts.  Not
140as heavy-handed as in the old "big kernel lock" days, but certainly not
141afraid of taking a lock when one is needed.  It uses a variety of
142different concurrency controls.  A background understanding of the
143various primitives is assumed, or can be gleaned from elsewhere such
144as in `Meet the Lockers`_.
145
146The locking mechanisms used by REF-walk include:
147
148dentry->d_lockref
149~~~~~~~~~~~~~~~~~
150
151This uses the lockref primitive to provide both a spinlock and a
152reference count.  The special-sauce of this primitive is that the
153conceptual sequence "lock; inc_ref; unlock;" can often be performed
154with a single atomic memory operation.
155
156Holding a reference on a dentry ensures that the dentry won't suddenly
157be freed and used for something else, so the values in various fields
158will behave as expected.  It also protects the ``->d_inode`` reference
159to the inode to some extent.
160
161The association between a dentry and its inode is fairly permanent.
162For example, when a file is renamed, the dentry and inode move
163together to the new location.  When a file is created the dentry will
164initially be negative (i.e. ``d_inode`` is ``NULL``), and will be assigned
165to the new inode as part of the act of creation.
166
167When a file is deleted, this can be reflected in the cache either by
168setting ``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.
170If the dentry is still in use the second option is used as it is
171perfectly legal to keep using an open file after it has been deleted
172and having the dentry around helps.  If the dentry is not otherwise in
173use (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
175very common case.
176
177So as long as a counted reference is held to a dentry, a non-``NULL`` ``->d_inode``
178value will never be changed.
179
180dentry->d_lock
181~~~~~~~~~~~~~~
182
183``d_lock`` is a synonym for the spinlock that is part of ``d_lockref`` above.
184For our purposes, holding this lock protects against the dentry being
185renamed or unlinked.  In particular, its parent (``d_parent``), and its
186name (``d_name``) cannot be changed, and it cannot be removed from the
187dentry hash table.
188
189When looking for a name in a directory, REF-walk takes ``d_lock`` on
190each candidate dentry that it finds in the hash table and then checks
191that the parent and name are correct.  So it doesn't lock the parent
192while searching in the cache; it only locks children.
193
194When looking for the parent for a given name (to handle "``..``"),
195REF-walk can take ``d_lock`` to get a stable reference to ``d_parent``,
196but it first tries a more lightweight approach.  As seen in
197``dget_parent()``, if a reference can be claimed on the parent, and if
198subsequently ``d_parent`` can be seen to have not changed, then there is
199no need to actually take the lock on the child.
200
201rename_lock
202~~~~~~~~~~~
203
204Looking up a given name in a given directory involves computing a hash
205from the two values (the name and the dentry of the directory),
206accessing that slot in a hash table, and searching the linked list
207that is found there.
208
209When a dentry is renamed, the name and the parent dentry can both
210change so the hash will almost certainly change too.  This would move the
211dentry to a different chain in the hash table.  If a filename search
212happened to be looking at a dentry that was moved in this way,
213it might end up continuing the search down the wrong chain,
214and so miss out on part of the correct chain.
215
216The name-lookup process (``d_lookup()``) does _not_ try to prevent this
217from happening, but only to detect when it happens.
218``rename_lock`` is a seqlock that is updated whenever any dentry is
219renamed.  If ``d_lookup`` finds that a rename happened while it
220unsuccessfully scanned a chain in the hash table, it simply tries
221again.
222
223inode->i_rwsem
224~~~~~~~~~~~~~~
225
226``i_rwsem`` is a read/write semaphore that serializes all changes to a particular
227directory.  This ensures that, for example, an ``unlink()`` and a ``rename()``
228cannot both happen at the same time.  It also keeps the directory
229stable while the filesystem is asked to look up a name that is not
230currently in the dcache or, optionally, when the list of entries in a
231directory is being retrieved with ``readdir()``.
232
233This has a complementary role to that of ``d_lock``: ``i_rwsem`` on a
234directory protects all of the names in that directory, while ``d_lock``
235on a name protects just one name in a directory.  Most changes to the
236dcache 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
238exception is when idle dentries are removed from the dcache due to
239memory pressure.  This uses ``d_lock``, but ``i_rwsem`` plays no role.
240
241The semaphore affects pathname lookup in two distinct ways.  Firstly it
242prevents 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,
244using only ``d_lock`` locking.  If the name isn't found, then ``walk_component()``
245falls back to ``lookup_slow()`` which takes a shared lock on ``i_rwsem``, checks again that
246the name isn't in the cache, and then calls in to the filesystem to get a
247definitive answer.  A new dentry will be added to the cache regardless of
248the result.
249
250Secondly, when pathname lookup reaches the final component, it will
251sometimes need to take an exclusive lock on ``i_rwsem`` before performing the last lookup so
252that the required exclusion can be achieved.  How path lookup chooses
253to take, or not take, ``i_rwsem`` is one of the
254issues addressed in a subsequent section.
255
256If two threads attempt to look up the same name at the same time - a
257name that is not yet in the dcache - the shared lock on ``i_rwsem`` will
258not prevent them both adding new dentries with the same name.  As this
259would result in confusion an extra level of interlocking is used,
260based around a secondary hash table (``in_lookup_hashtable``) and a
261per-dentry flag bit (``DCACHE_PAR_LOOKUP``).
262
263To 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
265dentry, stores the required name and parent in it, checks if there
266is already a matching dentry in the primary or secondary hash
267tables, and if not, stores the newly allocated dentry in the secondary
268hash table, with ``DCACHE_PAR_LOOKUP`` set.
269
270If a matching dentry was found in the primary hash table then that is
271returned and the caller can know that it lost a race with some other
272thread adding the entry.  If no matching dentry is found in either
273cache, the newly allocated dentry is returned and the caller can
274detect this from the presence of ``DCACHE_PAR_LOOKUP``.  In this case it
275knows that it has won any race and now is responsible for asking the
276filesystem to perform the lookup and find the matching inode.  When
277the lookup is complete, it must call ``d_lookup_done()`` which clears
278the flag and does some other house keeping, including removing the
279dentry from the secondary hash table - it will normally have been
280added to the primary hash table already.  Note that a ``struct
281waitqueue_head`` is passed to ``d_alloc_parallel()``, and
282``d_lookup_done()`` must be called while this ``waitqueue_head`` is still
283in scope.
284
285If 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
288to the instance of ``d_alloc_parallel()`` that won the race and that
289will be woken by the call to ``d_lookup_done()``.  It then checks to see
290if the dentry has now been added to the primary hash table.  If it
291has, the dentry is returned and the caller just sees that it lost any
292race.  If it hasn't been added to the primary hash table, the most
293likely explanation is that some other dentry was added instead using
294``d_splice_alias()``.  In any case, ``d_alloc_parallel()`` repeats all the
295look ups from the start and will normally return something from the
296primary hash table.
297
298mnt->mnt_count
299~~~~~~~~~~~~~~
300
301``mnt_count`` is a per-CPU reference counter on "``mount``" structures.
302Per-CPU here means that incrementing the count is cheap as it only
303uses CPU-local memory, but checking if the count is zero is expensive as
304it needs to check with every CPU.  Taking a ``mnt_count`` reference
305prevents the mount structure from disappearing as the result of regular
306unmount operations, but does not prevent a "lazy" unmount.  So holding
307``mnt_count`` doesn't ensure that the mount remains in the namespace and,
308in particular, doesn't stabilize the link to the mounted-on dentry.  It
309does, however, ensure that the ``mount`` data structure remains coherent,
310and it provides a reference to the root dentry of the mounted
311filesystem.  So a reference through ``->mnt_count`` provides a stable
312reference to the mounted dentry, but not the mounted-on dentry.
313
314mount_lock
315~~~~~~~~~~
316
317``mount_lock`` is a global seqlock, a bit like ``rename_lock``.  It can be used to
318check if any change has been made to any mount points.
319
320While walking down the tree (away from the root) this lock is used when
321crossing a mount point to check that the crossing was safe.  That is,
322the value in the seqlock is read, then the code finds the mount that
323is mounted on the current directory, if there is one, and increments
324the ``mnt_count``.  Finally the value in ``mount_lock`` is checked against
325the old value.  If there is no change, then the crossing was safe.  If there
326was a change, the ``mnt_count`` is decremented and the whole process is
327retried.
328
329When walking up the tree (towards the root) by following a ".." link,
330a little more care is needed.  In this case the seqlock (which
331contains both a counter and a spinlock) is fully locked to prevent
332any changes to any mount points while stepping up.  This locking is
333needed to stabilize the link to the mounted-on dentry, which the
334refcount on the mount itself doesn't ensure.
335
336RCU
337~~~
338
339Finally the global (but extremely lightweight) RCU read lock is held
340from time to time to ensure certain data structures don't get freed
341unexpectedly.
342
343In particular it is held while scanning chains in the dcache hash
344table, and the mount point hash table.
345
346Bringing it together with ``struct nameidata``
347--------------------------------------------
348
349.. _First edition Unix: http://minnie.tuhs.org/cgi-bin/utree.pl?file=V1/u2.s
350
351Throughout the process of walking a path, the current status is stored
352in a ``struct nameidata``, "namei" being the traditional name - dating
353all the way back to `First Edition Unix`_ - of the function that
354converts a "name" to an "inode".  ``struct nameidata`` contains (among
355other fields):
356
357``struct path path``
358~~~~~~~~~~~~~~~~~~
359
360A ``path`` contains a ``struct vfsmount`` (which is
361embedded in a ``struct mount``) and a ``struct dentry``.  Together these
362record the current status of the walk.  They start out referring to the
363starting point (the current working directory, the root directory, or some other
364directory identified by a file descriptor), and are updated on each
365step.  A reference through ``d_lockref`` and ``mnt_count`` is always
366held.
367
368``struct qstr last``
369~~~~~~~~~~~~~~~~~~
370
371This is a string together with a length (i.e. _not_ ``nul`` terminated)
372that is the "next" component in the pathname.
373
374``int last_type``
375~~~~~~~~~~~~~~~
376
377This 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
380components of the symlink have been processed yet.  Others should be
381fairly self-explanatory.
382
383``struct path root``
384~~~~~~~~~~~~~~~~~~
385
386This is used to hold a reference to the effective root of the
387filesystem.  Often that reference won't be needed, so this field is
388only assigned the first time it is used, or when a non-standard root
389is requested.  Keeping a reference in the ``nameidata`` ensures that
390only one root is in effect for the entire path walk, even if it races
391with a ``chroot()`` system call.
392
393The root is needed when either of two conditions holds: (1) either the
394pathname or a symbolic link starts with a "'/'", or (2) a "``..``"
395component is being handled, since "``..``" from the root must always stay
396at the root.  The value used is usually the current root directory of
397the 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
400specific part of the filesystem, and the lookup must not be allowed to
401escape that subtree.  It works a bit like a local ``chroot()``.
402
403Ignoring the handling of symbolic links, we can now describe the
404"``link_path_walk()``" function, which handles the lookup of everything
405except the final component as:
406
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.
412
413``walk_component()`` is even easier.  If the component is ``LAST_DOTS``,
414it calls ``handle_dots()`` which does the necessary locking as already
415described.  If it finds a ``LAST_NORM`` component it first calls
416"``lookup_fast()``" which only looks in the dcache, but will ask the
417filesystem to revalidate the result if it is that sort of filesystem.
418If that doesn't get a good result, it calls "``lookup_slow()``" which
419takes ``i_rwsem``, rechecks the cache, and then asks the filesystem
420to find a definitive answer.  Each of these will call
421``follow_managed()`` (as described below) to handle any mount points.
422
423In the absence of symbolic links, ``walk_component()`` creates a new
424``struct path`` containing a counted reference to the new dentry and a
425reference to the new ``vfsmount`` which is only counted if it is
426different 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.
429
430This "hand-over-hand" sequencing of getting a reference to the new
431dentry before dropping the reference to the previous dentry may
432seem obvious, but is worth pointing out so that we will recognize its
433analogue in the "RCU-walk" version.
434
435Handling the final component
436----------------------------
437
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
440not call ``walk_component()`` that last time.  Handling that final
441component 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
444different system calls.
445
446``path_parentat()`` is clearly the simplest - it just wraps a little bit
447of housekeeping around ``link_path_walk()`` and returns the parent
448directory and final component to the caller.  The caller will be either
449aiming to create a name (via ``filename_create()``) or remove or rename
450a name (in which case ``user_path_parent()`` is used).  They will use
451``i_rwsem`` to exclude other changes while they validate and then
452perform their operation.
453
454``path_lookupat()`` is nearly as simple - it is used when an existing
455object is wanted such as by ``stat()`` or ``chmod()``.  It essentially just
456calls ``walk_component()`` on the final component through a call to
457``lookup_last()``.  ``path_lookupat()`` returns just the final dentry.
458
459``path_mountpoint()`` handles the special case of unmounting which must
460not try to revalidate the mounted filesystem.  It effectively
461contains, through a call to ``mountpoint_last()``, an alternate
462implementation of ``lookup_slow()`` which skips that step.  This is
463important when unmounting a filesystem that is inaccessible, such as
464one provided by a dead NFS server.
465
466Finally ``path_openat()`` is used for the ``open()`` system call; it
467contains, in support functions starting with "``do_last()``", all the
468complexity needed to handle the different subtleties of O_CREAT (with
469or without O_EXCL), final "``/``" characters, and trailing symbolic
470links.  We will revisit this in the final part of this series, which
471focuses on those symbolic links.  "``do_last()``" will sometimes, but
472not always, take ``i_rwsem``, depending on what it finds.
473
474Each of these, or the functions which call them, need to be alert to
475the possibility that the final component is not ``LAST_NORM``.  If the
476goal of the lookup is to create something, then any value for
477``last_type`` other than ``LAST_NORM`` will result in an error.  For
478example if ``path_parentat()`` reports ``LAST_DOTDOT``, then the caller
479won't try to create that name.  They also check for trailing slashes
480by testing ``last.name[last.len]``.  If there is any character beyond
481the final component, it must be a trailing slash.
482
483Revalidation and automounts
484---------------------------
485
486Apart from symbolic links, there are only two parts of the "REF-walk"
487process not yet covered.  One is the handling of stale cache entries
488and the other is automounts.
489
490On filesystems that require it, the lookup routines will call the
491``->d_revalidate()`` dentry method to ensure that the cached information
492is current.  This will often confirm validity or update a few details
493from a server.  In some cases it may find that there has been change
494further up the path and that something that was thought to be valid
495previously isn't really.  When this happens the lookup of the whole
496path is aborted and retried with the "``LOOKUP_REVAL``" flag set.  This
497forces revalidation to be more thorough.  We will see more details of
498this retry process in the next article.
499
500Automount points are locations in the filesystem where an attempt to
501lookup a name can trigger changes to how that lookup should be
502handled, in particular by mounting a filesystem there.  These are
503covered in greater detail in autofs.txt in the Linux documentation
504tree, but a few notes specifically related to path lookup are in order
505here.
506
507The Linux VFS has a concept of "managed" dentries which is reflected
508in function names such as "``follow_managed()``".  There are three
509potentially interesting things about these dentries corresponding
510to three different flags that might be set in ``dentry->d_flags``:
511
512``DCACHE_MANAGE_TRANSIT``
513~~~~~~~~~~~~~~~~~~~~~~~
514
515If this flag has been set, then the filesystem has requested that the
516``d_manage()`` dentry operation be called before handling any possible
517mount point.  This can perform two particular services:
518
519It can block to avoid races.  If an automount point is being
520unmounted, the ``d_manage()`` function will usually wait for that
521process to complete before letting the new lookup proceed and possibly
522trigger a new automount.
523
524It can selectively allow only some processes to transit through a
525mount point.  When a server process is managing automounts, it may
526need to access a directory without triggering normal automount
527processing.  That server process can identify itself to the ``autofs``
528filesystem, which will then give it a special pass through
529``d_manage()`` by returning ``-EISDIR``.
530
531``DCACHE_MOUNTED``
532~~~~~~~~~~~~~~~~
533
534This flag is set on every dentry that is mounted on.  As Linux
535supports multiple filesystem namespaces, it is possible that the
536dentry may not be mounted on in *this* namespace, just in some
537other.  So this flag is seen as a hint, not a promise.
538
539If 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``
542and a new ``dentry`` (both with counted references).
543
544``DCACHE_NEED_AUTOMOUNT``
545~~~~~~~~~~~~~~~~~~~~~~~
546
547If ``d_manage()`` allowed us to get this far, and ``lookup_mnt()`` didn't
548find a mount point, then this flag causes the ``d_automount()`` dentry
549operation to be called.
550
551The ``d_automount()`` operation can be arbitrarily complex and may
552communicate with server processes etc. but it should ultimately either
553report that there was an error, that there was nothing to mount, or
554should provide an updated ``struct path`` with new ``dentry`` and ``vfsmount``.
555
556In the latter case, ``finish_automount()`` will be called to safely
557install the new mount point into the mount table.
558
559There is no new locking of import here and it is important that no
560locks (only counted references) are held over this processing due to
561the very real possibility of extended delays.
562This will become more important next time when we examine RCU-walk
563which is particularly sensitive to delays.
564
565RCU-walk - faster pathname lookup in Linux
566==========================================
567
568RCU-walk is another algorithm for performing pathname lookup in Linux.
569It is in many ways similar to REF-walk and the two share quite a bit
570of code.  The significant difference in RCU-walk is how it allows for
571the possibility of concurrent access.
572
573We noted that REF-walk is complex because there are numerous details
574and special cases.  RCU-walk reduces this complexity by simply
575refusing to handle a number of cases -- it instead falls back to
576REF-walk.  The difficulty with RCU-walk comes from a different
577direction: unfamiliarity.  The locking rules when depending on RCU are
578quite different from traditional locking, so we will spend a little extra
579time when we come to those.
580
581Clear demarcation of roles
582--------------------------
583
584The easiest way to manage concurrency is to forcibly stop any other
585thread from changing the data structures that a given thread is
586looking at.  In cases where no other thread would even think of
587changing the data and lots of different threads want to read at the
588same time, this can be very costly.  Even when using locks that permit
589multiple concurrent readers, the simple act of updating the count of
590the number of current readers can impose an unwanted cost.  So the
591goal when reading a shared data structure that no other process is
592changing is to avoid writing anything to memory at all.  Take no
593locks, increment no counts, leave no footprints.
594
595The REF-walk mechanism already described certainly doesn't follow this
596principle, but then it is really designed to work when there may well
597be other threads modifying the data.  RCU-walk, in contrast, is
598designed for the common situation where there are lots of frequent
599readers and only occasional writers.  This may not be common in all
600parts of the filesystem tree, but in many parts it will be.  For the
601other parts it is important that RCU-walk can quickly fall back to
602using REF-walk.
603
604Pathname lookup always starts in RCU-walk mode but only remains there
605as long as what it is looking for is in the cache and is stable.  It
606dances lightly down the cached filesystem image, leaving no footprints
607and carefully watching where it is, to be sure it doesn't trip.  If it
608notices that something has changed or is changing, or if something
609isn't in the cache, then it tries to stop gracefully and switch to
610REF-walk.
611
612This stopping requires getting a counted reference on the current
613``vfsmount`` and ``dentry``, and ensuring that these are still valid -
614that a path walk with REF-walk would have found the same entries.
615This is an invariant that RCU-walk must guarantee.  It can only make
616decisions, such as selecting the next step, that are decisions which
617REF-walk could also have made if it were walking down the tree at the
618same time.  If the graceful stop succeeds, the rest of the path is
619processed with the reliable, if slightly sluggish, REF-walk.  If
620RCU-walk finds it cannot stop gracefully, it simply gives up and
621restarts from the top with REF-walk.
622
623This pattern of "try RCU-walk, if that fails try REF-walk" can be
624clearly seen in functions like ``filename_lookup()``,
625``filename_parentat()``, ``filename_mountpoint()``,
626``do_filp_open()``, and ``do_file_open_root()``.  These five
627correspond roughly to the four ``path_``* functions we met earlier,
628each of which calls ``link_path_walk()``.  The ``path_*`` functions are
629called using different mode flags until a mode is found which works.
630They are first called with ``LOOKUP_RCU`` set to request "RCU-walk".  If
631that fails with the error ``ECHILD`` they are called again with no
632special flag to request "REF-walk".  If either of those report the
633error ``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
635revalidated - normally entries are only revalidated if the filesystem
636determines that they are too old to trust.
637
638The ``LOOKUP_RCU`` attempt may drop that flag internally and switch to
639REF-walk, but will never then try to switch back to RCU-walk.  Places
640that trip up RCU-walk are much more likely to be near the leaves and
641so it is very unlikely that there will be much, if any, benefit from
642switching back.
643
644RCU and seqlocks: fast and light
645--------------------------------
646
647RCU is, unsurprisingly, critical to RCU-walk mode.  The
648``rcu_read_lock()`` is held for the entire time that RCU-walk is walking
649down a path.  The particular guarantee it provides is that the key
650data structures - dentries, inodes, super_blocks, and mounts - will
651not be freed while the lock is held.  They might be unlinked or
652invalidated in one way or another, but the memory will not be
653repurposed so values in various fields will still be meaningful.  This
654is the only guarantee that RCU provides; everything else is done using
655seqlocks.
656
657As we saw above, REF-walk holds a counted reference to the current
658dentry and the current vfsmount, and does not release those references
659before taking references to the "next" dentry or vfsmount.  It also
660sometimes takes the ``d_lock`` spinlock.  These references and locks are
661taken to prevent certain changes from happening.  RCU-walk must not
662take those references or locks and so cannot prevent such changes.
663Instead, it checks to see if a change has been made, and aborts or
664retries if it has.
665
666To preserve the invariant mentioned above (that RCU-walk may only make
667decisions that REF-walk could have made), it must make the checks at
668or near the same places that REF-walk holds the references.  So, when
669REF-walk increments a reference count or takes a spinlock, RCU-walk
670samples the status of a seqlock using ``read_seqcount_begin()`` or a
671similar function.  When REF-walk decrements the count or drops the
672lock, RCU-walk checks if the sampled status is still valid using
673``read_seqcount_retry()`` or similar.
674
675However, there is a little bit more to seqlocks than that.  If
676RCU-walk accesses two different fields in a seqlock-protected
677structure, or accesses the same field twice, there is no a priori
678guarantee of any consistency between those accesses.  When consistency
679is needed - which it usually is - RCU-walk must take a copy and then
680use ``read_seqcount_retry()`` to validate that copy.
681
682``read_seqcount_retry()`` not only checks the sequence number, but also
683imposes a memory barrier so that no memory-read instruction from
684*before* the call can be delayed until *after* the call, either by the
685CPU 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
687byte-wise name equality, calls into the filesystem to compare a name
688against a dentry.  The length and name pointer are copied into local
689variables, then ``read_seqcount_retry()`` is called to confirm the two
690are consistent, and only then is ``->d_compare()`` called.  When
691standard filename comparison is used, ``dentry_cmp()`` is called
692instead.  Notably it does _not_ use ``read_seqcount_retry()``, but
693instead has a large comment explaining why the consistency guarantee
694isn't necessary.  A subsequent ``read_seqcount_retry()`` will be
695sufficient to catch any problem that could occur at this point.
696
697With that little refresher on seqlocks out of the way we can look at
698the bigger picture of how RCU-walk uses seqlocks.
699
700``mount_lock`` and ``nd->m_seq``
701~~~~~~~~~~~~~~~~~~~~~~~~~~~~
702
703We already met the ``mount_lock`` seqlock when REF-walk used it to
704ensure that crossing a mount point is performed safely.  RCU-walk uses
705it for that too, but for quite a bit more.
706
707Instead of taking a counted reference to each ``vfsmount`` as it
708descends the tree, RCU-walk samples the state of ``mount_lock`` at the
709start of the walk and stores this initial sequence number in the
710``struct nameidata`` in the ``m_seq`` field.  This one lock and one
711sequence number are used to validate all accesses to all ``vfsmounts``,
712and all mount point crossings.  As changes to the mount table are
713relatively rare, it is reasonable to fall back on REF-walk any time
714that any "mount" or "unmount" happens.
715
716``m_seq`` is checked (using ``read_seqretry()``) at the end of an RCU-walk
717sequence, whether switching to REF-walk for the rest of the path or
718when the end of the path is reached.  It is also checked when stepping
719down over a mount point (in ``__follow_mount_rcu()``) or up (in
720``follow_dotdot_rcu()``).  If it is ever found to have changed, the
721whole RCU-walk sequence is aborted and the path is processed again by
722REF-walk.
723
724If RCU-walk finds that ``mount_lock`` hasn't changed then it can be sure
725that, had REF-walk taken counted references on each vfsmount, the
726results would have been the same.  This ensures the invariant holds,
727at least for vfsmount structures.
728
729``dentry->d_seq`` and ``nd->seq``
730~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
731
732In place of taking a count or lock on ``d_reflock``, RCU-walk samples
733the 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
735the current sequence number of ``nd->dentry``.  This number needs to be
736revalidated after copying, and before using, the name, parent, or
737inode of the dentry.
738
739The handling of the name we have already looked at, and the parent is
740only accessed in ``follow_dotdot_rcu()`` which fairly trivially follows
741the required pattern, though it does so for three different cases.
742
743When not at a mount point, ``d_parent`` is followed and its ``d_seq`` is
744collected.  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
747check if we have landed on a mount point and, if so, must find that
748mount point and follow the ``mnt->mnt_root`` link.  This would imply a
749somewhat unusual, but certainly possible, circumstance where the
750starting point of the path lookup was in part of the filesystem that
751was mounted on, and so not visible from the root.
752
753The inode pointer, stored in ``->d_inode``, is a little more
754interesting.  The inode will always need to be accessed at least
755twice, once to determine if it is NULL and once to verify access
756permissions.  Symlink handling requires a validated inode pointer too.
757Rather than revalidating on each access, a copy is made on the first
758access and it is stored in the ``inode`` field of ``nameidata`` from where
759it can be safely accessed without further validation.
760
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
764of the current dentry.
765
766The current ``dentry`` and current ``seq`` number are passed to
767``__d_lookup_rcu()`` which, on success, returns a new ``dentry`` and a
768new ``seq`` number.  ``lookup_fast()`` then copies the inode pointer and
769revalidates the new ``seq`` number.  It then validates the old ``dentry``
770with the old ``seq`` number one last time and only then continues.  This
771process of getting the ``seq`` number of the new dentry and then
772checking the ``seq`` number of the old exactly mirrors the process of
773getting a counted reference to the new dentry before dropping that for
774the old dentry which we saw in REF-walk.
775
776No ``inode->i_rwsem`` or even ``rename_lock``
777~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
778
779A semaphore is a fairly heavyweight lock that can only be taken when it is
780permissible to sleep.  As ``rcu_read_lock()`` forbids sleeping,
781``inode->i_rwsem`` plays no role in RCU-walk.  If some other thread does
782take ``i_rwsem`` and modifies the directory in a way that RCU-walk needs
783to notice, the result will be either that RCU-walk fails to find the
784dentry 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
786REF-walk mode which can take whatever locks are needed.
787
788Though ``rename_lock`` could be used by RCU-walk as it doesn't require
789any sleeping, RCU-walk doesn't bother.  REF-walk uses ``rename_lock`` to
790protect against the possibility of hash chains in the dcache changing
791while they are being searched.  This can result in failing to find
792something that actually is there.  When RCU-walk fails to find
793something in the dentry cache, whether it is really there or not, it
794already drops down to REF-walk and tries again with appropriate
795locking.  This neatly handles all cases, so adding extra checks on
796rename_lock would bring no significant value.
797
798``unlazy walk()`` and ``complete_walk()``
799-------------------------------------
800
801That "dropping down to REF-walk" typically involves a call to
802``unlazy_walk()``, so named because "RCU-walk" is also sometimes
803referred to as "lazy walk".  ``unlazy_walk()`` is called when
804following the path down to the current vfsmount/dentry pair seems to
805have proceeded successfully, but the next step is problematic.  This
806can happen if the next name cannot be found in the dcache, if
807permission checking or name revalidation couldn't be achieved while
808the ``rcu_read_lock()`` is held (which forbids sleeping), if an
809automount point is found, or in a couple of cases involving symlinks.
810It is also called from ``complete_walk()`` when the lookup has reached
811the final component, or the very end of the path, depending on which
812particular flavor of lookup is used.
813
814Other reasons for dropping out of RCU-walk that do not trigger a call
815to ``unlazy_walk()`` are when some inconsistency is found that cannot be
816handled immediately, such as ``mount_lock`` or one of the ``d_seq``
817seqlocks reporting a change.  In these cases the relevant function
818will return ``-ECHILD`` which will percolate up until it triggers a new
819attempt from the top using REF-walk.
820
821For those cases where ``unlazy_walk()`` is an option, it essentially
822takes a reference on each of the pointers that it holds (vfsmount,
823dentry, and possibly some symbolic links) and then verifies that the
824relevant seqlocks have not been changed.  If there have been changes,
825it, too, aborts with ``-ECHILD``, otherwise the transition to REF-walk
826has been a success and the lookup process continues.
827
828Taking a reference on those pointers is not quite as simple as just
829incrementing a counter.  That works to take a second reference if you
830already have one (often indirectly through another object), but it
831isn't sufficient if you don't actually have a counted reference at
832all.  For ``dentry->d_lockref``, it is safe to increment the reference
833counter 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.
836
837For ``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
839validation fails, it may *not* be safe to just drop that reference in
840the standard way of calling ``mnt_put()`` - an unmount may have
841progressed too far.  So the code in ``legitimize_mnt()``, when it
842finds that the reference it got might not be safe, checks the
843``MNT_SYNC_UMOUNT`` flag to determine if a simple ``mnt_put()`` is
844correct, or if it should just decrement the count and pretend none of
845this ever happened.
846
847Taking care in filesystems
848--------------------------
849
850RCU-walk depends almost entirely on cached information and often will
851not call into the filesystem at all.  However there are two places,
852besides the already-mentioned component-name comparison, where the
853file system might be included in RCU-walk, and it must know to be
854careful.
855
856If the filesystem has non-standard permission-checking requirements -
857such as a networked filesystem which may need to check with the server
858- the ``i_op->permission`` interface might be called during RCU-walk.
859In this case an extra "``MAY_NOT_BLOCK``" flag is passed so that it
860knows not to sleep, but to return ``-ECHILD`` if it cannot complete
861promptly.  ``i_op->permission`` is given the inode pointer, not the
862dentry, so it doesn't need to worry about further consistency checks.
863However if it accesses any other filesystem data structures, it must
864ensure they are safe to be accessed with only the ``rcu_read_lock()``
865held.  This typically means they must be freed using ``kfree_rcu()`` or
866similar.
867
868.. _READ_ONCE: https://lwn.net/Articles/624126/
869
870If 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
874when accessing fields in the dentry.  This "extra care" typically
875involves using  `READ_ONCE() <READ_ONCE_>`_ to access fields, and verifying the
876result is not NULL before using it.  This pattern can be seen in
877``nfs_lookup_revalidate()``.
878
879A pair of patterns
880------------------
881
882In various places in the details of REF-walk and RCU-walk, and also in
883the big picture, there are a couple of related patterns that are worth
884being aware of.
885
886The first is "try quickly and check, if that fails try slowly".  We
887can see that in the high-level approach of first trying RCU-walk and
888then trying REF-walk, and in places where ``unlazy_walk()`` is used to
889switch to REF-walk for the rest of the path.  We also saw it earlier
890in ``dget_parent()`` when following a "``..``" link.  It tries a quick way
891to get a reference, then falls back to taking locks if needed.
892
893The second pattern is "try quickly and check, if that fails try
894again - 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 -
896if anything goes wrong it is much safer to just abort and try a more
897sedate approach.
898
899The emphasis here is "try quickly and check".  It should probably be
900"try quickly _and carefully,_ then check".  The fact that checking is
901needed is a reminder that the system is dynamic and only a limited
902number of things are safe at all.  The most likely cause of errors in
903this whole process is assuming something is safe when in reality it
904isn't.  Careful consideration of what exactly guarantees the safety of
905each access is sometimes necessary.
906
907A walk among the symlinks
908=========================
909
910There are several basic issues that we will examine to understand the
911handling of symbolic links:  the symlink stack, together with cache
912lifetimes, will help us understand the overall recursive handling of
913symlinks and lead to the special care needed for the final component.
914Then a consideration of access-time updates and summary of the various
915flags controlling lookup will finish the story.
916
917The symlink stack
918-----------------
919
920There are only two sorts of filesystem objects that can usefully
921appear in a path prior to the final component: directories and symlinks.
922Handling directories is quite straightforward: the new directory
923simply becomes the starting point at which to interpret the next
924component on the path.  Handling symbolic links requires a bit more
925work.
926
927Conceptually, symbolic links could be handled by editing the path.  If
928a component name refers to a symbolic link, then that component is
929replaced by the body of the link and, if that body starts with a '/',
930then all preceding parts of the path are discarded.  This is what the
931"``readlink -f``" command does, though it also edits out "``.``" and
932"``..``" components.
933
934Directly editing the path string is not really necessary when looking
935up a path, and discarding early components is pointless as they aren't
936looked at anyway.  Keeping track of all remaining components is
937important, but they can of course be kept separately; there is no need
938to concatenate them.  As one symlink may easily refer to another,
939which in turn can refer to a third, we may need to keep the remaining
940components of several paths, each to be processed when the preceding
941ones are completed.  These path remnants are kept on a stack of
942limited size.
943
944There are two reasons for placing limits on how many symlinks can
945occur in a single path lookup.  The most obvious is to avoid loops.
946If a symlink referred to itself either directly or through
947intermediaries, then following the symlink can never complete
948successfully - the error ``ELOOP`` must be returned.  Loops can be
949detected without imposing limits, but limits are the simplest solution
950and, given the second reason for restriction, quite sufficient.
951
952.. _outlined recently: http://thread.gmane.org/gmane.linux.kernel/1934390/focus=1934550
953
954The second reason was `outlined recently`_ by Linus:
955
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.
959
960Linux imposes a limit on the length of any pathname: ``PATH_MAX``, which
961is 4096.  There are a number of reasons for this limit; not letting the
962kernel spend too much time on just one path is one of them.  With
963symbolic links you can effectively generate much longer paths so some
964sort of limit is needed for the same reason.  Linux imposes a limit of
965at most 40 symlinks in any one path lookup.  It previously imposed a
966further limit of eight on the maximum depth of recursion, but that was
967raised to 40 when a separate stack was implemented, so there is now
968just the one limit.
969
970The ``nameidata`` structure that we met in an earlier article contains a
971small stack that can be used to store the remaining part of up to two
972symlinks.  In many cases this will be sufficient.  If it isn't, a
973separate stack is allocated with room for 40 symlinks.  Pathname
974lookup will never exceed that stack as, once the 40th symlink is
975detected, an error is returned.
976
977It might seem that the name remnants are all that needs to be stored on
978this stack, but we need a bit more.  To see that, we need to move on to
979cache lifetimes.
980
981Storage and lifetime of cached symlinks
982---------------------------------------
983
984Like other filesystem resources, such as inodes and directory
985entries, symlinks are cached by Linux to avoid repeated costly access
986to external storage.  It is particularly important for RCU-walk to be
987able to find and temporarily hold onto these cached entries, so that
988it doesn't need to drop down into REF-walk.
989
990.. _object-oriented design pattern: https://lwn.net/Articles/446317/
991
992While each filesystem is free to make its own choice, symlinks are
993typically stored in one of two places.  Short symlinks are often
994stored directly in the inode.  When a filesystem allocates a ``struct
995inode`` it typically allocates extra space to store private data (a
996common `object-oriented design pattern`_ in the kernel).  This will
997sometimes include space for a symlink.  The other common location is
998in the page cache, which normally stores the content of files.  The
999pathname in a symlink can be seen as the content of that symlink and
1000can easily be stored in the page cache just like file content.
1001
1002When neither of these is suitable, the next most likely scenario is
1003that the filesystem will allocate some temporary memory and copy or
1004construct the symlink content into that memory whenever it is needed.
1005
1006When the symlink is stored in the inode, it has the same lifetime as
1007the inode which, itself, is protected by RCU or by a counted reference
1008on the dentry.  This means that the mechanisms that pathname lookup
1009uses to access the dcache and icache (inode cache) safely are quite
1010sufficient for accessing some cached symlinks safely.  In these cases,
1011the ``i_link`` pointer in the inode is set to point to wherever the
1012symlink is stored and it can be accessed directly whenever needed.
1013
1014When the symlink is stored in the page cache or elsewhere, the
1015situation is not so straightforward.  A reference on a dentry or even
1016on an inode does not imply any reference on cached pages of that
1017inode, and even an ``rcu_read_lock()`` is not sufficient to ensure that
1018a page will not disappear.  So for these symlinks the pathname lookup
1019code needs to ask the filesystem to provide a stable reference and,
1020significantly, needs to release that reference when it is finished
1021with it.
1022
1023Taking a reference to a cache page is often possible even in RCU-walk
1024mode.  It does require making changes to memory, which is best avoided,
1025but that isn't necessarily a big cost and it is better than dropping
1026out of RCU-walk mode completely.  Even filesystems that allocate
1027space to copy the symlink into can use ``GFP_ATOMIC`` to often successfully
1028allocate memory without the need to drop out of RCU-walk.  If a
1029filesystem cannot successfully get a reference in RCU-walk mode, it
1030must return ``-ECHILD`` and ``unlazy_walk()`` will be called to return to
1031REF-walk mode in which the filesystem is allowed to sleep.
1032
1033The place for all this to happen is the ``i_op->follow_link()`` inode
1034method.  In the present mainline code this is never actually called in
1035RCU-walk mode as the rewrite is not quite complete.  It is likely that
1036in a future release this method will be passed an ``inode`` pointer when
1037called in RCU-walk mode so it both (1) knows to be careful, and (2) has the
1038validated pointer.  Much like the ``i_op->permission()`` method we
1039looked at previously, ``->follow_link()`` would need to be careful that
1040all the data structures it references are safe to be accessed while
1041holding no counted reference, only the RCU lock.  Though getting a
1042reference with ``->follow_link()`` is not yet done in RCU-walk mode, the
1043code is ready to release the reference when that does happen.
1044
1045This need to drop the reference to a symlink adds significant
1046complexity.  It requires a reference to the inode so that the
1047``i_op->put_link()`` inode operation can be called.  In REF-walk, that
1048reference is kept implicitly through a reference to the dentry, so
1049keeping the ``struct path`` of the symlink is easiest.  For RCU-walk,
1050the pointer to the inode is kept separately.  To allow switching from
1051RCU-walk back to REF-walk in the middle of processing nested symlinks
1052we also need the seq number for the dentry so we can confirm that
1053switching back was safe.
1054
1055Finally, when providing a reference to a symlink, the filesystem also
1056provides an opaque "cookie" that must be passed to ``->put_link()`` so that it
1057knows what to free.  This might be the allocated memory area, or a
1058pointer to the ``struct page`` in the page cache, or something else
1059completely.  Only the filesystem knows what it is.
1060
1061In order for the reference to each symlink to be dropped when the walk completes,
1062whether in RCU-walk or REF-walk, the symlink stack needs to contain,
1063along with the path remnants:
1064
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.
1069
1070This means that each entry in the symlink stack needs to hold five
1071pointers and an integer instead of just one pointer (the path
1072remnant).  On a 64-bit system, this is about 40 bytes per entry;
1073with 40 entries it adds up to 1600 bytes total, which is less than
1074half a page.  So it might seem like a lot, but is by no means
1075excessive.
1076
1077Note that, in a given stack frame, the path remnant (``name``) is not
1078part of the symlink that the other fields refer to.  It is the remnant
1079to be followed once that symlink has been fully parsed.
1080
1081Following the symlink
1082---------------------
1083
1084The main loop in ``link_path_walk()`` iterates seamlessly over all
1085components in the path and all of the non-final symlinks.  As symlinks
1086are processed, the ``name`` pointer is adjusted to point to a new
1087symlink, or is restored from the stack, so that much of the loop
1088doesn't need to notice.  Getting this ``name`` variable on and off the
1089stack is very straightforward; pushing and popping the references is
1090a little more complex.
1091
1092When a symlink is found, ``walk_component()`` returns the value ``1``
1093(``0`` is returned for any other sort of success, and a negative number
1094is, as usual, an error indicator).  This causes ``get_link()`` to be
1095called; it then gets the link from the filesystem.  Providing that
1096operation is successful, the old path ``name`` is placed on the stack,
1097and the new value is used as the ``name`` for a while.  When the end of
1098the path is found (i.e. ``*name`` is ``'\0'``) the old ``name`` is restored
1099off the stack and path walking continues.
1100
1101Pushing and popping the reference pointers (inode, cookie, etc.) is more
1102complex in part because of the desire to handle tail recursion.  When
1103the last component of a symlink itself points to a symlink, we
1104want to pop the symlink-just-completed off the stack before pushing
1105the symlink-just-found to avoid leaving empty path remnants that would
1106just get in the way.
1107
1108It is most convenient to push the new symlink references onto the
1109stack 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
1111old symlink as it walks that last component.  So it is quite
1112convenient for ``walk_component()`` to release the old symlink and pop
1113the references just before pushing the reference information for the
1114new symlink.  It is guided in this by two flags; ``WALK_GET``, which
1115gives 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
1117followed.  ``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
1120up the stack frame.
1121
1122Symlinks with no final component
1123~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1124
1125A pair of special-case symlinks deserve a little further explanation.
1126Both result in a new ``struct path`` (with mount and dentry) being set
1127up in the ``nameidata``, and result in ``get_link()`` returning ``NULL``.
1128
1129The more obvious case is a symlink to "``/``".  All symlinks starting
1130with "``/``" are detected in ``get_link()`` which resets the ``nameidata``
1131to point to the effective filesystem root.  If the symlink only
1132contains "``/``" then there is nothing more to do, no components at all,
1133so ``NULL`` is returned to indicate that the symlink can be released and
1134the stack frame discarded.
1135
1136The other case involves things in ``/proc`` that look like symlinks but
1137aren't really::
1138
1139     $ ls -l /proc/self/fd/1
1140     lrwx------ 1 neilb neilb 64 Jun 13 10:19 /proc/self/fd/1 -> /dev/pts/4
1141
1142Every open file descriptor in any process is represented in ``/proc`` by
1143something that looks like a symlink.  It is really a reference to the
1144target file, not just the name of it.  When you ``readlink`` these
1145objects you get a name that might refer to the same file - unless it
1146has been unlinked or mounted over.  When ``walk_component()`` follows
1147one of these, the ``->follow_link()`` method in "procfs" doesn't return
1148a string name, but instead calls ``nd_jump_link()`` which updates the
1149``nameidata`` in place to point to that target.  ``->follow_link()`` then
1150returns ``NULL``.  Again there is no final component and ``get_link()``
1151reports this by leaving the ``last_type`` field of ``nameidata`` as
1152``LAST_BIND``.
1153
1154Following the symlink in the final component
1155--------------------------------------------
1156
1157All this leads to ``link_path_walk()`` walking down every component, and
1158following all symbolic links it finds, until it reaches the final
1159component.  This is just returned in the ``last`` field of ``nameidata``.
1160For 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
1162callers will want to follow a symlink if one is found, and possibly
1163apply special handling to the last component of that symlink, rather
1164than just the last component of the original file name.  These callers
1165potentially need to call ``link_path_walk()`` again and again on
1166successive symlinks until one is found that doesn't point to another
1167symlink.
1168
1169This 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
1171handles the final component.  If the final component is a symlink
1172that needs to be followed, then ``trailing_symlink()`` is called to set
1173things up properly and the loop repeats, calling ``link_path_walk()``
1174again.  This could loop as many as 40 times if the last component of
1175each symlink is another symlink.
1176
1177The various functions that examine the final component and possibly
1178report that it is a symlink are ``lookup_last()``, ``mountpoint_last()``
1179and ``do_last()``, each of which use the same convention as
1180``walk_component()`` of returning ``1`` if a symlink was found that needs
1181to be followed.
1182
1183Of these, ``do_last()`` is the most interesting as it is used for
1184opening a file.  Part of ``do_last()`` runs with ``i_rwsem`` held and this
1185part is in a separate function: ``lookup_open()``.
1186
1187Explaining ``do_last()`` completely is beyond the scope of this article,
1188but a few highlights should help those interested in exploring the
1189code.
1190
11911. 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.
1199
12002. ``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.
1205
12063. An open with O_CREAT **does** follow a symlink in the final component,
1207   unlike other creation system calls (like ``mkdir``).  So the sequence::
1208
1209          ln -s bar /tmp/foo
1210          echo hello > /tmp/foo
1211
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.
1217
1218Updating the access time
1219------------------------
1220
1221We previously said of RCU-walk that it would "take no locks, increment
1222no counts, leave no footprints."  We have since seen that some
1223"footprints" can be needed when handling symlinks as a counted
1224reference (or even a memory allocation) may be needed.  But these
1225footprints are best kept to a minimum.
1226
1227One other place where walking down a symlink can involve leaving
1228footprints in a way that doesn't affect directories is in updating access times.
1229In Unix (and Linux) every filesystem object has a "last accessed
1230time", or "``atime``".  Passing through a directory to access a file
1231within is not considered to be an access for the purposes of
1232``atime``; only listing the contents of a directory can update its ``atime``.
1233Symlinks are different it seems.  Both reading a symlink (with ``readlink()``)
1234and looking up a symlink on the way to some other destination can
1235update the atime on that symlink.
1236
1237.. _clearest statement: http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_08
1238
1239It is not clear why this is the case; POSIX has little to say on the
1240subject.  The `clearest statement`_ is that, if a particular implementation
1241updates a timestamp in a place not specified by POSIX, this must be
1242documented "except that any changes caused by pathname resolution need
1243not be documented".  This seems to imply that POSIX doesn't really
1244care about access-time updates during pathname lookup.
1245
1246.. _Linux 1.3.87: https://git.kernel.org/cgit/linux/kernel/git/history/history.git/diff/fs/ext2/symlink.c?id=f806c6db77b8eaa6e00dcfb6b567706feae8dbb8
1247
1248An examination of history shows that prior to `Linux 1.3.87`_, the ext2
1249filesystem, at least, didn't update atime when following a link.
1250Unfortunately we have no record of why that behavior was changed.
1251
1252In any case, access time must now be updated and that operation can be
1253quite complex.  Trying to stay in RCU-walk while doing it is best
1254avoided.  Fortunately it is often permitted to skip the ``atime``
1255update.  Because ``atime`` updates cause performance problems in various
1256areas, Linux supports the ``relatime`` mount option, which generally
1257limits the updates of ``atime`` to once per day on files that aren't
1258being changed (and symlinks never change once created).  Even without
1259``relatime``, many filesystems record ``atime`` with a one-second
1260granularity, so only one update per second is required.
1261
1262It is easy to test if an ``atime`` update is needed while in RCU-walk
1263mode and, if it isn't, the update can be skipped and RCU-walk mode
1264continues.  Only when an ``atime`` update is actually required does the
1265path walk drop down to REF-walk.  All of this is handled in the
1266``get_link()`` function.
1267
1268A few flags
1269-----------
1270
1271A suitable way to wrap up this tour of pathname walking is to list
1272the various flags that can be stored in the ``nameidata`` to guide the
1273lookup process.  Many of these are only meaningful on the final
1274component, others reflect the current state of the pathname lookup.
1275And then there is ``LOOKUP_EMPTY``, which doesn't fit conceptually with
1276the others.  If this is not set, an empty pathname causes an error
1277very early on.  If it is set, empty pathnames are not considered to be
1278an error.
1279
1280Global state flags
1281~~~~~~~~~~~~~~~~~~
1282
1283We have already met two global state flags: ``LOOKUP_RCU`` and
1284``LOOKUP_REVAL``.  These select between one of three overall approaches
1285to lookup: RCU-walk, REF-walk, and REF-walk with forced revalidation.
1286
1287``LOOKUP_PARENT`` indicates that the final component hasn't been reached
1288yet.  This is primarily used to tell the audit subsystem the full
1289context of a particular access being audited.
1290
1291``LOOKUP_ROOT`` indicates that the ``root`` field in the ``nameidata`` was
1292provided by the caller, so it shouldn't be released when it is no
1293longer needed.
1294
1295``LOOKUP_JUMPED`` means that the current dentry was chosen not because
1296it had the right name but for some other reason.  This happens when
1297following "``..``", following a symlink to ``/``, crossing a mount point
1298or accessing a "``/proc/$PID/fd/$FD``" symlink.  In this case the
1299filesystem has not been asked to revalidate the name (with
1300``d_revalidate()``).  In such cases the inode may still need to be
1301revalidated, so ``d_op->d_weak_revalidate()`` is called if
1302``LOOKUP_JUMPED`` is set when the look completes - which may be at the
1303final component or, when creating, unlinking, or renaming, at the penultimate component.
1304
1305Final-component flags
1306~~~~~~~~~~~~~~~~~~~~~
1307
1308Some of these flags are only set when the final component is being
1309considered.  Others are only checked for when considering that final
1310component.
1311
1312``LOOKUP_AUTOMOUNT`` ensures that, if the final component is an automount
1313point, then the mount is triggered.  Some operations would trigger it
1314anyway, but operations like ``stat()`` deliberately don't.  ``statfs()``
1315needs to trigger the mount but otherwise behaves a lot like ``stat()``, so
1316it sets ``LOOKUP_AUTOMOUNT``, as does "``quotactl()``" and the handling of
1317"``mount --bind``".
1318
1319``LOOKUP_FOLLOW`` has a similar function to ``LOOKUP_AUTOMOUNT`` but for
1320symlinks.  Some system calls set or clear it implicitly, while
1321others 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.
1324
1325``LOOKUP_DIRECTORY`` insists that the final component is a directory.
1326Various callers set this and it is also set when the final component
1327is found to be followed by a slash.
1328
1329Finally ``LOOKUP_OPEN``, ``LOOKUP_CREATE``, ``LOOKUP_EXCL``, and
1330``LOOKUP_RENAME_TARGET`` are not used directly by the VFS but are made
1331available to the filesystem and particularly the ``->d_revalidate()``
1332method.  A filesystem can choose not to bother revalidating too hard
1333if it knows that it will be asked to open or create the file soon.
1334These flags were previously useful for ``->lookup()`` too but with the
1335introduction of ``->atomic_open()`` they are less relevant there.
1336
1337End of the road
1338---------------
1339
1340Despite its complexity, all this pathname lookup code appears to be
1341in good shape - various parts are certainly easier to understand now
1342than even a couple of releases ago.  But that doesn't mean it is
1343"finished".   As already mentioned, RCU-walk currently only follows
1344symlinks that are stored in the inode so, while it handles many ext4
1345symlinks, it doesn't help with NFS, XFS, or Btrfs.  That support
1346is not likely to be long delayed.
1347