Loading...
Note: File does not exist in v6.13.7.
1<head>
2<style> p { max-width:50em} ol, ul {max-width: 40em}</style>
3</head>
4
5Pathname lookup in Linux.
6=========================
7
8This write-up is based on three articles published at lwn.net:
9
10- <https://lwn.net/Articles/649115/> Pathname lookup in Linux
11- <https://lwn.net/Articles/649729/> RCU-walk: faster pathname lookup in Linux
12- <https://lwn.net/Articles/650786/> A walk among the symlinks
13
14Written by Neil Brown with help from Al Viro and Jon Corbet.
15
16Introduction
17------------
18
19The most obvious aspect of pathname lookup, which very little
20exploration is needed to discover, is that it is complex. There are
21many rules, special cases, and implementation alternatives that all
22combine to confuse the unwary reader. Computer science has long been
23acquainted with such complexity and has tools to help manage it. One
24tool that we will make extensive use of is "divide and conquer". For
25the early parts of the analysis we will divide off symlinks - leaving
26them until the final part. Well before we get to symlinks we have
27another major division based on the VFS's approach to locking which
28will allow us to review "REF-walk" and "RCU-walk" separately. But we
29are getting ahead of ourselves. There are some important low level
30distinctions we need to clarify first.
31
32There are two sorts of ...
33--------------------------
34
35[`openat()`]: http://man7.org/linux/man-pages/man2/openat.2.html
36
37Pathnames (sometimes "file names"), used to identify objects in the
38filesystem, will be familiar to most readers. They contain two sorts
39of elements: "slashes" that are sequences of one or more "`/`"
40characters, and "components" that are sequences of one or more
41non-"`/`" characters. These form two kinds of paths. Those that
42start with slashes are "absolute" and start from the filesystem root.
43The others are "relative" and start from the current directory, or
44from some other location specified by a file descriptor given to a
45"xxx`at`" system call such as "[`openat()`]".
46
47[`execveat()`]: http://man7.org/linux/man-pages/man2/execveat.2.html
48
49It is tempting to describe the second kind as starting with a
50component, but that isn't always accurate: a pathname can lack both
51slashes and components, it can be empty, in other words. This is
52generally forbidden in POSIX, but some of those "xxx`at`" system calls
53in Linux permit it when the `AT_EMPTY_PATH` flag is given. For
54example, if you have an open file descriptor on an executable file you
55can execute it by calling [`execveat()`] passing the file descriptor,
56an empty path, and the `AT_EMPTY_PATH` flag.
57
58These paths can be divided into two sections: the final component and
59everything else. The "everything else" is the easy bit. In all cases
60it must identify a directory that already exists, otherwise an error
61such as `ENOENT` or `ENOTDIR` will be reported.
62
63The final component is not so simple. Not only do different system
64calls interpret it quite differently (e.g. some create it, some do
65not), but it might not even exist: neither the empty pathname nor the
66pathname that is just slashes have a final component. If it does
67exist, it could be "`.`" or "`..`" which are handled quite differently
68from other components.
69
70[POSIX]: http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_12
71
72If a pathname ends with a slash, such as "`/tmp/foo/`" it might be
73tempting to consider that to have an empty final component. In many
74ways that would lead to correct results, but not always. In
75particular, `mkdir()` and `rmdir()` each create or remove a directory named
76by the final component, and they are required to work with pathnames
77ending in "`/`". According to [POSIX]
78
79> A pathname that contains at least one non- <slash> character and
80> that ends with one or more trailing <slash> characters shall not
81> be resolved successfully unless the last pathname component before
82> the trailing <slash> characters names an existing directory or a
83> directory entry that is to be created for a directory immediately
84> after the pathname is resolved.
85
86The Linux pathname walking code (mostly in `fs/namei.c`) deals with
87all of these issues: breaking the path into components, handling the
88"everything else" quite separately from the final component, and
89checking that the trailing slash is not used where it isn't
90permitted. It also addresses the important issue of concurrent
91access.
92
93While one process is looking up a pathname, another might be making
94changes that affect that lookup. One fairly extreme case is that if
95"a/b" were renamed to "a/c/b" while another process were looking up
96"a/b/..", that process might successfully resolve on "a/c".
97Most races are much more subtle, and a big part of the task of
98pathname lookup is to prevent them from having damaging effects. Many
99of the possible races are seen most clearly in the context of the
100"dcache" and an understanding of that is central to understanding
101pathname lookup.
102
103More than just a cache.
104-----------------------
105
106The "dcache" caches information about names in each filesystem to
107make them quickly available for lookup. Each entry (known as a
108"dentry") contains three significant fields: a component name, a
109pointer to a parent dentry, and a pointer to the "inode" which
110contains further information about the object in that parent with
111the given name. The inode pointer can be `NULL` indicating that the
112name doesn't exist in the parent. While there can be linkage in the
113dentry of a directory to the dentries of the children, that linkage is
114not used for pathname lookup, and so will not be considered here.
115
116The dcache has a number of uses apart from accelerating lookup. One
117that will be particularly relevant is that it is closely integrated
118with the mount table that records which filesystem is mounted where.
119What the mount table actually stores is which dentry is mounted on top
120of which other dentry.
121
122When considering the dcache, we have another of our "two types"
123distinctions: there are two types of filesystems.
124
125Some filesystems ensure that the information in the dcache is always
126completely accurate (though not necessarily complete). This can allow
127the VFS to determine if a particular file does or doesn't exist
128without checking with the filesystem, and means that the VFS can
129protect the filesystem against certain races and other problems.
130These are typically "local" filesystems such as ext3, XFS, and Btrfs.
131
132Other filesystems don't provide that guarantee because they cannot.
133These are typically filesystems that are shared across a network,
134whether remote filesystems like NFS and 9P, or cluster filesystems
135like ocfs2 or cephfs. These filesystems allow the VFS to revalidate
136cached information, and must provide their own protection against
137awkward races. The VFS can detect these filesystems by the
138`DCACHE_OP_REVALIDATE` flag being set in the dentry.
139
140REF-walk: simple concurrency management with refcounts and spinlocks
141--------------------------------------------------------------------
142
143With all of those divisions carefully classified, we can now start
144looking at the actual process of walking along a path. In particular
145we will start with the handling of the "everything else" part of a
146pathname, and focus on the "REF-walk" approach to concurrency
147management. This code is found in the `link_path_walk()` function, if
148you ignore all the places that only run when "`LOOKUP_RCU`"
149(indicating the use of RCU-walk) is set.
150
151[Meet the Lockers]: https://lwn.net/Articles/453685/
152
153REF-walk is fairly heavy-handed with locks and reference counts. Not
154as heavy-handed as in the old "big kernel lock" days, but certainly not
155afraid of taking a lock when one is needed. It uses a variety of
156different concurrency controls. A background understanding of the
157various primitives is assumed, or can be gleaned from elsewhere such
158as in [Meet the Lockers].
159
160The locking mechanisms used by REF-walk include:
161
162### dentry->d_lockref ###
163
164This uses the lockref primitive to provide both a spinlock and a
165reference count. The special-sauce of this primitive is that the
166conceptual sequence "lock; inc_ref; unlock;" can often be performed
167with a single atomic memory operation.
168
169Holding a reference on a dentry ensures that the dentry won't suddenly
170be freed and used for something else, so the values in various fields
171will behave as expected. It also protects the `->d_inode` reference
172to the inode to some extent.
173
174The association between a dentry and its inode is fairly permanent.
175For example, when a file is renamed, the dentry and inode move
176together to the new location. When a file is created the dentry will
177initially be negative (i.e. `d_inode` is `NULL`), and will be assigned
178to the new inode as part of the act of creation.
179
180When a file is deleted, this can be reflected in the cache either by
181setting `d_inode` to `NULL`, or by removing it from the hash table
182(described shortly) used to look up the name in the parent directory.
183If the dentry is still in use the second option is used as it is
184perfectly legal to keep using an open file after it has been deleted
185and having the dentry around helps. If the dentry is not otherwise in
186use (i.e. if the refcount in `d_lockref` is one), only then will
187`d_inode` be set to `NULL`. Doing it this way is more efficient for a
188very common case.
189
190So as long as a counted reference is held to a dentry, a non-`NULL` `->d_inode`
191value will never be changed.
192
193### dentry->d_lock ###
194
195`d_lock` is a synonym for the spinlock that is part of `d_lockref` above.
196For our purposes, holding this lock protects against the dentry being
197renamed or unlinked. In particular, its parent (`d_parent`), and its
198name (`d_name`) cannot be changed, and it cannot be removed from the
199dentry hash table.
200
201When looking for a name in a directory, REF-walk takes `d_lock` on
202each candidate dentry that it finds in the hash table and then checks
203that the parent and name are correct. So it doesn't lock the parent
204while searching in the cache; it only locks children.
205
206When looking for the parent for a given name (to handle "`..`"),
207REF-walk can take `d_lock` to get a stable reference to `d_parent`,
208but it first tries a more lightweight approach. As seen in
209`dget_parent()`, if a reference can be claimed on the parent, and if
210subsequently `d_parent` can be seen to have not changed, then there is
211no need to actually take the lock on the child.
212
213### rename_lock ###
214
215Looking up a given name in a given directory involves computing a hash
216from the two values (the name and the dentry of the directory),
217accessing that slot in a hash table, and searching the linked list
218that is found there.
219
220When a dentry is renamed, the name and the parent dentry can both
221change so the hash will almost certainly change too. This would move the
222dentry to a different chain in the hash table. If a filename search
223happened to be looking at a dentry that was moved in this way,
224it might end up continuing the search down the wrong chain,
225and so miss out on part of the correct chain.
226
227The name-lookup process (`d_lookup()`) does _not_ try to prevent this
228from happening, but only to detect when it happens.
229`rename_lock` is a seqlock that is updated whenever any dentry is
230renamed. If `d_lookup` finds that a rename happened while it
231unsuccessfully scanned a chain in the hash table, it simply tries
232again.
233
234### inode->i_mutex ###
235
236`i_mutex` is a mutex that serializes all changes to a particular
237directory. This ensures that, for example, an `unlink()` and a `rename()`
238cannot both happen at the same time. It also keeps the directory
239stable while the filesystem is asked to look up a name that is not
240currently in the dcache.
241
242This has a complementary role to that of `d_lock`: `i_mutex` on a
243directory protects all of the names in that directory, while `d_lock`
244on a name protects just one name in a directory. Most changes to the
245dcache hold `i_mutex` on the relevant directory inode and briefly take
246`d_lock` on one or more the dentries while the change happens. One
247exception is when idle dentries are removed from the dcache due to
248memory pressure. This uses `d_lock`, but `i_mutex` plays no role.
249
250The mutex affects pathname lookup in two distinct ways. Firstly it
251serializes lookup of a name in a directory. `walk_component()` uses
252`lookup_fast()` first which, in turn, checks to see if the name is in the cache,
253using only `d_lock` locking. If the name isn't found, then `walk_component()`
254falls back to `lookup_slow()` which takes `i_mutex`, checks again that
255the name isn't in the cache, and then calls in to the filesystem to get a
256definitive answer. A new dentry will be added to the cache regardless of
257the result.
258
259Secondly, when pathname lookup reaches the final component, it will
260sometimes need to take `i_mutex` before performing the last lookup so
261that the required exclusion can be achieved. How path lookup chooses
262to take, or not take, `i_mutex` is one of the
263issues addressed in a subsequent section.
264
265### mnt->mnt_count ###
266
267`mnt_count` is a per-CPU reference counter on "`mount`" structures.
268Per-CPU here means that incrementing the count is cheap as it only
269uses CPU-local memory, but checking if the count is zero is expensive as
270it needs to check with every CPU. Taking a `mnt_count` reference
271prevents the mount structure from disappearing as the result of regular
272unmount operations, but does not prevent a "lazy" unmount. So holding
273`mnt_count` doesn't ensure that the mount remains in the namespace and,
274in particular, doesn't stabilize the link to the mounted-on dentry. It
275does, however, ensure that the `mount` data structure remains coherent,
276and it provides a reference to the root dentry of the mounted
277filesystem. So a reference through `->mnt_count` provides a stable
278reference to the mounted dentry, but not the mounted-on dentry.
279
280### mount_lock ###
281
282`mount_lock` is a global seqlock, a bit like `rename_lock`. It can be used to
283check if any change has been made to any mount points.
284
285While walking down the tree (away from the root) this lock is used when
286crossing a mount point to check that the crossing was safe. That is,
287the value in the seqlock is read, then the code finds the mount that
288is mounted on the current directory, if there is one, and increments
289the `mnt_count`. Finally the value in `mount_lock` is checked against
290the old value. If there is no change, then the crossing was safe. If there
291was a change, the `mnt_count` is decremented and the whole process is
292retried.
293
294When walking up the tree (towards the root) by following a ".." link,
295a little more care is needed. In this case the seqlock (which
296contains both a counter and a spinlock) is fully locked to prevent
297any changes to any mount points while stepping up. This locking is
298needed to stabilize the link to the mounted-on dentry, which the
299refcount on the mount itself doesn't ensure.
300
301### RCU ###
302
303Finally the global (but extremely lightweight) RCU read lock is held
304from time to time to ensure certain data structures don't get freed
305unexpectedly.
306
307In particular it is held while scanning chains in the dcache hash
308table, and the mount point hash table.
309
310Bringing it together with `struct nameidata`
311--------------------------------------------
312
313[First edition Unix]: http://minnie.tuhs.org/cgi-bin/utree.pl?file=V1/u2.s
314
315Throughout the process of walking a path, the current status is stored
316in a `struct nameidata`, "namei" being the traditional name - dating
317all the way back to [First Edition Unix] - of the function that
318converts a "name" to an "inode". `struct nameidata` contains (among
319other fields):
320
321### `struct path path` ###
322
323A `path` contains a `struct vfsmount` (which is
324embedded in a `struct mount`) and a `struct dentry`. Together these
325record the current status of the walk. They start out referring to the
326starting point (the current working directory, the root directory, or some other
327directory identified by a file descriptor), and are updated on each
328step. A reference through `d_lockref` and `mnt_count` is always
329held.
330
331### `struct qstr last` ###
332
333This is a string together with a length (i.e. _not_ `nul` terminated)
334that is the "next" component in the pathname.
335
336### `int last_type` ###
337
338This is one of `LAST_NORM`, `LAST_ROOT`, `LAST_DOT`, `LAST_DOTDOT`, or
339`LAST_BIND`. The `last` field is only valid if the type is
340`LAST_NORM`. `LAST_BIND` is used when following a symlink and no
341components of the symlink have been processed yet. Others should be
342fairly self-explanatory.
343
344### `struct path root` ###
345
346This is used to hold a reference to the effective root of the
347filesystem. Often that reference won't be needed, so this field is
348only assigned the first time it is used, or when a non-standard root
349is requested. Keeping a reference in the `nameidata` ensures that
350only one root is in effect for the entire path walk, even if it races
351with a `chroot()` system call.
352
353The root is needed when either of two conditions holds: (1) either the
354pathname or a symbolic link starts with a "'/'", or (2) a "`..`"
355component is being handled, since "`..`" from the root must always stay
356at the root. The value used is usually the current root directory of
357the calling process. An alternate root can be provided as when
358`sysctl()` calls `file_open_root()`, and when NFSv4 or Btrfs call
359`mount_subtree()`. In each case a pathname is being looked up in a very
360specific part of the filesystem, and the lookup must not be allowed to
361escape that subtree. It works a bit like a local `chroot()`.
362
363Ignoring the handling of symbolic links, we can now describe the
364"`link_path_walk()`" function, which handles the lookup of everything
365except the final component as:
366
367> Given a path (`name`) and a nameidata structure (`nd`), check that the
368> current directory has execute permission and then advance `name`
369> over one component while updating `last_type` and `last`. If that
370> was the final component, then return, otherwise call
371> `walk_component()` and repeat from the top.
372
373`walk_component()` is even easier. If the component is `LAST_DOTS`,
374it calls `handle_dots()` which does the necessary locking as already
375described. If it finds a `LAST_NORM` component it first calls
376"`lookup_fast()`" which only looks in the dcache, but will ask the
377filesystem to revalidate the result if it is that sort of filesystem.
378If that doesn't get a good result, it calls "`lookup_slow()`" which
379takes the `i_mutex`, rechecks the cache, and then asks the filesystem
380to find a definitive answer. Each of these will call
381`follow_managed()` (as described below) to handle any mount points.
382
383In the absence of symbolic links, `walk_component()` creates a new
384`struct path` containing a counted reference to the new dentry and a
385reference to the new `vfsmount` which is only counted if it is
386different from the previous `vfsmount`. It then calls
387`path_to_nameidata()` to install the new `struct path` in the
388`struct nameidata` and drop the unneeded references.
389
390This "hand-over-hand" sequencing of getting a reference to the new
391dentry before dropping the reference to the previous dentry may
392seem obvious, but is worth pointing out so that we will recognize its
393analogue in the "RCU-walk" version.
394
395Handling the final component.
396-----------------------------
397
398`link_path_walk()` only walks as far as setting `nd->last` and
399`nd->last_type` to refer to the final component of the path. It does
400not call `walk_component()` that last time. Handling that final
401component remains for the caller to sort out. Those callers are
402`path_lookupat()`, `path_parentat()`, `path_mountpoint()` and
403`path_openat()` each of which handles the differing requirements of
404different system calls.
405
406`path_parentat()` is clearly the simplest - it just wraps a little bit
407of housekeeping around `link_path_walk()` and returns the parent
408directory and final component to the caller. The caller will be either
409aiming to create a name (via `filename_create()`) or remove or rename
410a name (in which case `user_path_parent()` is used). They will use
411`i_mutex` to exclude other changes while they validate and then
412perform their operation.
413
414`path_lookupat()` is nearly as simple - it is used when an existing
415object is wanted such as by `stat()` or `chmod()`. It essentially just
416calls `walk_component()` on the final component through a call to
417`lookup_last()`. `path_lookupat()` returns just the final dentry.
418
419`path_mountpoint()` handles the special case of unmounting which must
420not try to revalidate the mounted filesystem. It effectively
421contains, through a call to `mountpoint_last()`, an alternate
422implementation of `lookup_slow()` which skips that step. This is
423important when unmounting a filesystem that is inaccessible, such as
424one provided by a dead NFS server.
425
426Finally `path_openat()` is used for the `open()` system call; it
427contains, in support functions starting with "`do_last()`", all the
428complexity needed to handle the different subtleties of O_CREAT (with
429or without O_EXCL), final "`/`" characters, and trailing symbolic
430links. We will revisit this in the final part of this series, which
431focuses on those symbolic links. "`do_last()`" will sometimes, but
432not always, take `i_mutex`, depending on what it finds.
433
434Each of these, or the functions which call them, need to be alert to
435the possibility that the final component is not `LAST_NORM`. If the
436goal of the lookup is to create something, then any value for
437`last_type` other than `LAST_NORM` will result in an error. For
438example if `path_parentat()` reports `LAST_DOTDOT`, then the caller
439won't try to create that name. They also check for trailing slashes
440by testing `last.name[last.len]`. If there is any character beyond
441the final component, it must be a trailing slash.
442
443Revalidation and automounts
444---------------------------
445
446Apart from symbolic links, there are only two parts of the "REF-walk"
447process not yet covered. One is the handling of stale cache entries
448and the other is automounts.
449
450On filesystems that require it, the lookup routines will call the
451`->d_revalidate()` dentry method to ensure that the cached information
452is current. This will often confirm validity or update a few details
453from a server. In some cases it may find that there has been change
454further up the path and that something that was thought to be valid
455previously isn't really. When this happens the lookup of the whole
456path is aborted and retried with the "`LOOKUP_REVAL`" flag set. This
457forces revalidation to be more thorough. We will see more details of
458this retry process in the next article.
459
460Automount points are locations in the filesystem where an attempt to
461lookup a name can trigger changes to how that lookup should be
462handled, in particular by mounting a filesystem there. These are
463covered in greater detail in autofs4.txt in the Linux documentation
464tree, but a few notes specifically related to path lookup are in order
465here.
466
467The Linux VFS has a concept of "managed" dentries which is reflected
468in function names such as "`follow_managed()`". There are three
469potentially interesting things about these dentries corresponding
470to three different flags that might be set in `dentry->d_flags`:
471
472### `DCACHE_MANAGE_TRANSIT` ###
473
474If this flag has been set, then the filesystem has requested that the
475`d_manage()` dentry operation be called before handling any possible
476mount point. This can perform two particular services:
477
478It can block to avoid races. If an automount point is being
479unmounted, the `d_manage()` function will usually wait for that
480process to complete before letting the new lookup proceed and possibly
481trigger a new automount.
482
483It can selectively allow only some processes to transit through a
484mount point. When a server process is managing automounts, it may
485need to access a directory without triggering normal automount
486processing. That server process can identify itself to the `autofs`
487filesystem, which will then give it a special pass through
488`d_manage()` by returning `-EISDIR`.
489
490### `DCACHE_MOUNTED` ###
491
492This flag is set on every dentry that is mounted on. As Linux
493supports multiple filesystem namespaces, it is possible that the
494dentry may not be mounted on in *this* namespace, just in some
495other. So this flag is seen as a hint, not a promise.
496
497If this flag is set, and `d_manage()` didn't return `-EISDIR`,
498`lookup_mnt()` is called to examine the mount hash table (honoring the
499`mount_lock` described earlier) and possibly return a new `vfsmount`
500and a new `dentry` (both with counted references).
501
502### `DCACHE_NEED_AUTOMOUNT` ###
503
504If `d_manage()` allowed us to get this far, and `lookup_mnt()` didn't
505find a mount point, then this flag causes the `d_automount()` dentry
506operation to be called.
507
508The `d_automount()` operation can be arbitrarily complex and may
509communicate with server processes etc. but it should ultimately either
510report that there was an error, that there was nothing to mount, or
511should provide an updated `struct path` with new `dentry` and `vfsmount`.
512
513In the latter case, `finish_automount()` will be called to safely
514install the new mount point into the mount table.
515
516There is no new locking of import here and it is important that no
517locks (only counted references) are held over this processing due to
518the very real possibility of extended delays.
519This will become more important next time when we examine RCU-walk
520which is particularly sensitive to delays.
521
522RCU-walk - faster pathname lookup in Linux
523==========================================
524
525RCU-walk is another algorithm for performing pathname lookup in Linux.
526It is in many ways similar to REF-walk and the two share quite a bit
527of code. The significant difference in RCU-walk is how it allows for
528the possibility of concurrent access.
529
530We noted that REF-walk is complex because there are numerous details
531and special cases. RCU-walk reduces this complexity by simply
532refusing to handle a number of cases -- it instead falls back to
533REF-walk. The difficulty with RCU-walk comes from a different
534direction: unfamiliarity. The locking rules when depending on RCU are
535quite different from traditional locking, so we will spend a little extra
536time when we come to those.
537
538Clear demarcation of roles
539--------------------------
540
541The easiest way to manage concurrency is to forcibly stop any other
542thread from changing the data structures that a given thread is
543looking at. In cases where no other thread would even think of
544changing the data and lots of different threads want to read at the
545same time, this can be very costly. Even when using locks that permit
546multiple concurrent readers, the simple act of updating the count of
547the number of current readers can impose an unwanted cost. So the
548goal when reading a shared data structure that no other process is
549changing is to avoid writing anything to memory at all. Take no
550locks, increment no counts, leave no footprints.
551
552The REF-walk mechanism already described certainly doesn't follow this
553principle, but then it is really designed to work when there may well
554be other threads modifying the data. RCU-walk, in contrast, is
555designed for the common situation where there are lots of frequent
556readers and only occasional writers. This may not be common in all
557parts of the filesystem tree, but in many parts it will be. For the
558other parts it is important that RCU-walk can quickly fall back to
559using REF-walk.
560
561Pathname lookup always starts in RCU-walk mode but only remains there
562as long as what it is looking for is in the cache and is stable. It
563dances lightly down the cached filesystem image, leaving no footprints
564and carefully watching where it is, to be sure it doesn't trip. If it
565notices that something has changed or is changing, or if something
566isn't in the cache, then it tries to stop gracefully and switch to
567REF-walk.
568
569This stopping requires getting a counted reference on the current
570`vfsmount` and `dentry`, and ensuring that these are still valid -
571that a path walk with REF-walk would have found the same entries.
572This is an invariant that RCU-walk must guarantee. It can only make
573decisions, such as selecting the next step, that are decisions which
574REF-walk could also have made if it were walking down the tree at the
575same time. If the graceful stop succeeds, the rest of the path is
576processed with the reliable, if slightly sluggish, REF-walk. If
577RCU-walk finds it cannot stop gracefully, it simply gives up and
578restarts from the top with REF-walk.
579
580This pattern of "try RCU-walk, if that fails try REF-walk" can be
581clearly seen in functions like `filename_lookup()`,
582`filename_parentat()`, `filename_mountpoint()`,
583`do_filp_open()`, and `do_file_open_root()`. These five
584correspond roughly to the four `path_`* functions we met earlier,
585each of which calls `link_path_walk()`. The `path_*` functions are
586called using different mode flags until a mode is found which works.
587They are first called with `LOOKUP_RCU` set to request "RCU-walk". If
588that fails with the error `ECHILD` they are called again with no
589special flag to request "REF-walk". If either of those report the
590error `ESTALE` a final attempt is made with `LOOKUP_REVAL` set (and no
591`LOOKUP_RCU`) to ensure that entries found in the cache are forcibly
592revalidated - normally entries are only revalidated if the filesystem
593determines that they are too old to trust.
594
595The `LOOKUP_RCU` attempt may drop that flag internally and switch to
596REF-walk, but will never then try to switch back to RCU-walk. Places
597that trip up RCU-walk are much more likely to be near the leaves and
598so it is very unlikely that there will be much, if any, benefit from
599switching back.
600
601RCU and seqlocks: fast and light
602--------------------------------
603
604RCU is, unsurprisingly, critical to RCU-walk mode. The
605`rcu_read_lock()` is held for the entire time that RCU-walk is walking
606down a path. The particular guarantee it provides is that the key
607data structures - dentries, inodes, super_blocks, and mounts - will
608not be freed while the lock is held. They might be unlinked or
609invalidated in one way or another, but the memory will not be
610repurposed so values in various fields will still be meaningful. This
611is the only guarantee that RCU provides; everything else is done using
612seqlocks.
613
614As we saw above, REF-walk holds a counted reference to the current
615dentry and the current vfsmount, and does not release those references
616before taking references to the "next" dentry or vfsmount. It also
617sometimes takes the `d_lock` spinlock. These references and locks are
618taken to prevent certain changes from happening. RCU-walk must not
619take those references or locks and so cannot prevent such changes.
620Instead, it checks to see if a change has been made, and aborts or
621retries if it has.
622
623To preserve the invariant mentioned above (that RCU-walk may only make
624decisions that REF-walk could have made), it must make the checks at
625or near the same places that REF-walk holds the references. So, when
626REF-walk increments a reference count or takes a spinlock, RCU-walk
627samples the status of a seqlock using `read_seqcount_begin()` or a
628similar function. When REF-walk decrements the count or drops the
629lock, RCU-walk checks if the sampled status is still valid using
630`read_seqcount_retry()` or similar.
631
632However, there is a little bit more to seqlocks than that. If
633RCU-walk accesses two different fields in a seqlock-protected
634structure, or accesses the same field twice, there is no a priori
635guarantee of any consistency between those accesses. When consistency
636is needed - which it usually is - RCU-walk must take a copy and then
637use `read_seqcount_retry()` to validate that copy.
638
639`read_seqcount_retry()` not only checks the sequence number, but also
640imposes a memory barrier so that no memory-read instruction from
641*before* the call can be delayed until *after* the call, either by the
642CPU or by the compiler. A simple example of this can be seen in
643`slow_dentry_cmp()` which, for filesystems which do not use simple
644byte-wise name equality, calls into the filesystem to compare a name
645against a dentry. The length and name pointer are copied into local
646variables, then `read_seqcount_retry()` is called to confirm the two
647are consistent, and only then is `->d_compare()` called. When
648standard filename comparison is used, `dentry_cmp()` is called
649instead. Notably it does _not_ use `read_seqcount_retry()`, but
650instead has a large comment explaining why the consistency guarantee
651isn't necessary. A subsequent `read_seqcount_retry()` will be
652sufficient to catch any problem that could occur at this point.
653
654With that little refresher on seqlocks out of the way we can look at
655the bigger picture of how RCU-walk uses seqlocks.
656
657### `mount_lock` and `nd->m_seq` ###
658
659We already met the `mount_lock` seqlock when REF-walk used it to
660ensure that crossing a mount point is performed safely. RCU-walk uses
661it for that too, but for quite a bit more.
662
663Instead of taking a counted reference to each `vfsmount` as it
664descends the tree, RCU-walk samples the state of `mount_lock` at the
665start of the walk and stores this initial sequence number in the
666`struct nameidata` in the `m_seq` field. This one lock and one
667sequence number are used to validate all accesses to all `vfsmounts`,
668and all mount point crossings. As changes to the mount table are
669relatively rare, it is reasonable to fall back on REF-walk any time
670that any "mount" or "unmount" happens.
671
672`m_seq` is checked (using `read_seqretry()`) at the end of an RCU-walk
673sequence, whether switching to REF-walk for the rest of the path or
674when the end of the path is reached. It is also checked when stepping
675down over a mount point (in `__follow_mount_rcu()`) or up (in
676`follow_dotdot_rcu()`). If it is ever found to have changed, the
677whole RCU-walk sequence is aborted and the path is processed again by
678REF-walk.
679
680If RCU-walk finds that `mount_lock` hasn't changed then it can be sure
681that, had REF-walk taken counted references on each vfsmount, the
682results would have been the same. This ensures the invariant holds,
683at least for vfsmount structures.
684
685### `dentry->d_seq` and `nd->seq`. ###
686
687In place of taking a count or lock on `d_reflock`, RCU-walk samples
688the per-dentry `d_seq` seqlock, and stores the sequence number in the
689`seq` field of the nameidata structure, so `nd->seq` should always be
690the current sequence number of `nd->dentry`. This number needs to be
691revalidated after copying, and before using, the name, parent, or
692inode of the dentry.
693
694The handling of the name we have already looked at, and the parent is
695only accessed in `follow_dotdot_rcu()` which fairly trivially follows
696the required pattern, though it does so for three different cases.
697
698When not at a mount point, `d_parent` is followed and its `d_seq` is
699collected. When we are at a mount point, we instead follow the
700`mnt->mnt_mountpoint` link to get a new dentry and collect its
701`d_seq`. Then, after finally finding a `d_parent` to follow, we must
702check if we have landed on a mount point and, if so, must find that
703mount point and follow the `mnt->mnt_root` link. This would imply a
704somewhat unusual, but certainly possible, circumstance where the
705starting point of the path lookup was in part of the filesystem that
706was mounted on, and so not visible from the root.
707
708The inode pointer, stored in `->d_inode`, is a little more
709interesting. The inode will always need to be accessed at least
710twice, once to determine if it is NULL and once to verify access
711permissions. Symlink handling requires a validated inode pointer too.
712Rather than revalidating on each access, a copy is made on the first
713access and it is stored in the `inode` field of `nameidata` from where
714it can be safely accessed without further validation.
715
716`lookup_fast()` is the only lookup routine that is used in RCU-mode,
717`lookup_slow()` being too slow and requiring locks. It is in
718`lookup_fast()` that we find the important "hand over hand" tracking
719of the current dentry.
720
721The current `dentry` and current `seq` number are passed to
722`__d_lookup_rcu()` which, on success, returns a new `dentry` and a
723new `seq` number. `lookup_fast()` then copies the inode pointer and
724revalidates the new `seq` number. It then validates the old `dentry`
725with the old `seq` number one last time and only then continues. This
726process of getting the `seq` number of the new dentry and then
727checking the `seq` number of the old exactly mirrors the process of
728getting a counted reference to the new dentry before dropping that for
729the old dentry which we saw in REF-walk.
730
731### No `inode->i_mutex` or even `rename_lock` ###
732
733A mutex is a fairly heavyweight lock that can only be taken when it is
734permissible to sleep. As `rcu_read_lock()` forbids sleeping,
735`inode->i_mutex` plays no role in RCU-walk. If some other thread does
736take `i_mutex` and modifies the directory in a way that RCU-walk needs
737to notice, the result will be either that RCU-walk fails to find the
738dentry that it is looking for, or it will find a dentry which
739`read_seqretry()` won't validate. In either case it will drop down to
740REF-walk mode which can take whatever locks are needed.
741
742Though `rename_lock` could be used by RCU-walk as it doesn't require
743any sleeping, RCU-walk doesn't bother. REF-walk uses `rename_lock` to
744protect against the possibility of hash chains in the dcache changing
745while they are being searched. This can result in failing to find
746something that actually is there. When RCU-walk fails to find
747something in the dentry cache, whether it is really there or not, it
748already drops down to REF-walk and tries again with appropriate
749locking. This neatly handles all cases, so adding extra checks on
750rename_lock would bring no significant value.
751
752`unlazy walk()` and `complete_walk()`
753-------------------------------------
754
755That "dropping down to REF-walk" typically involves a call to
756`unlazy_walk()`, so named because "RCU-walk" is also sometimes
757referred to as "lazy walk". `unlazy_walk()` is called when
758following the path down to the current vfsmount/dentry pair seems to
759have proceeded successfully, but the next step is problematic. This
760can happen if the next name cannot be found in the dcache, if
761permission checking or name revalidation couldn't be achieved while
762the `rcu_read_lock()` is held (which forbids sleeping), if an
763automount point is found, or in a couple of cases involving symlinks.
764It is also called from `complete_walk()` when the lookup has reached
765the final component, or the very end of the path, depending on which
766particular flavor of lookup is used.
767
768Other reasons for dropping out of RCU-walk that do not trigger a call
769to `unlazy_walk()` are when some inconsistency is found that cannot be
770handled immediately, such as `mount_lock` or one of the `d_seq`
771seqlocks reporting a change. In these cases the relevant function
772will return `-ECHILD` which will percolate up until it triggers a new
773attempt from the top using REF-walk.
774
775For those cases where `unlazy_walk()` is an option, it essentially
776takes a reference on each of the pointers that it holds (vfsmount,
777dentry, and possibly some symbolic links) and then verifies that the
778relevant seqlocks have not been changed. If there have been changes,
779it, too, aborts with `-ECHILD`, otherwise the transition to REF-walk
780has been a success and the lookup process continues.
781
782Taking a reference on those pointers is not quite as simple as just
783incrementing a counter. That works to take a second reference if you
784already have one (often indirectly through another object), but it
785isn't sufficient if you don't actually have a counted reference at
786all. For `dentry->d_lockref`, it is safe to increment the reference
787counter to get a reference unless it has been explicitly marked as
788"dead" which involves setting the counter to `-128`.
789`lockref_get_not_dead()` achieves this.
790
791For `mnt->mnt_count` it is safe to take a reference as long as
792`mount_lock` is then used to validate the reference. If that
793validation fails, it may *not* be safe to just drop that reference in
794the standard way of calling `mnt_put()` - an unmount may have
795progressed too far. So the code in `legitimize_mnt()`, when it
796finds that the reference it got might not be safe, checks the
797`MNT_SYNC_UMOUNT` flag to determine if a simple `mnt_put()` is
798correct, or if it should just decrement the count and pretend none of
799this ever happened.
800
801Taking care in filesystems
802---------------------------
803
804RCU-walk depends almost entirely on cached information and often will
805not call into the filesystem at all. However there are two places,
806besides the already-mentioned component-name comparison, where the
807file system might be included in RCU-walk, and it must know to be
808careful.
809
810If the filesystem has non-standard permission-checking requirements -
811such as a networked filesystem which may need to check with the server
812- the `i_op->permission` interface might be called during RCU-walk.
813In this case an extra "`MAY_NOT_BLOCK`" flag is passed so that it
814knows not to sleep, but to return `-ECHILD` if it cannot complete
815promptly. `i_op->permission` is given the inode pointer, not the
816dentry, so it doesn't need to worry about further consistency checks.
817However if it accesses any other filesystem data structures, it must
818ensure they are safe to be accessed with only the `rcu_read_lock()`
819held. This typically means they must be freed using `kfree_rcu()` or
820similar.
821
822[`READ_ONCE()`]: https://lwn.net/Articles/624126/
823
824If the filesystem may need to revalidate dcache entries, then
825`d_op->d_revalidate` may be called in RCU-walk too. This interface
826*is* passed the dentry but does not have access to the `inode` or the
827`seq` number from the `nameidata`, so it needs to be extra careful
828when accessing fields in the dentry. This "extra care" typically
829involves using `ACCESS_ONCE()` or the newer [`READ_ONCE()`] to access
830fields, and verifying the result is not NULL before using it. This
831pattern can be see in `nfs_lookup_revalidate()`.
832
833A pair of patterns
834------------------
835
836In various places in the details of REF-walk and RCU-walk, and also in
837the big picture, there are a couple of related patterns that are worth
838being aware of.
839
840The first is "try quickly and check, if that fails try slowly". We
841can see that in the high-level approach of first trying RCU-walk and
842then trying REF-walk, and in places where `unlazy_walk()` is used to
843switch to REF-walk for the rest of the path. We also saw it earlier
844in `dget_parent()` when following a "`..`" link. It tries a quick way
845to get a reference, then falls back to taking locks if needed.
846
847The second pattern is "try quickly and check, if that fails try
848again - repeatedly". This is seen with the use of `rename_lock` and
849`mount_lock` in REF-walk. RCU-walk doesn't make use of this pattern -
850if anything goes wrong it is much safer to just abort and try a more
851sedate approach.
852
853The emphasis here is "try quickly and check". It should probably be
854"try quickly _and carefully,_ then check". The fact that checking is
855needed is a reminder that the system is dynamic and only a limited
856number of things are safe at all. The most likely cause of errors in
857this whole process is assuming something is safe when in reality it
858isn't. Careful consideration of what exactly guarantees the safety of
859each access is sometimes necessary.
860
861A walk among the symlinks
862=========================
863
864There are several basic issues that we will examine to understand the
865handling of symbolic links: the symlink stack, together with cache
866lifetimes, will help us understand the overall recursive handling of
867symlinks and lead to the special care needed for the final component.
868Then a consideration of access-time updates and summary of the various
869flags controlling lookup will finish the story.
870
871The symlink stack
872-----------------
873
874There are only two sorts of filesystem objects that can usefully
875appear in a path prior to the final component: directories and symlinks.
876Handling directories is quite straightforward: the new directory
877simply becomes the starting point at which to interpret the next
878component on the path. Handling symbolic links requires a bit more
879work.
880
881Conceptually, symbolic links could be handled by editing the path. If
882a component name refers to a symbolic link, then that component is
883replaced by the body of the link and, if that body starts with a '/',
884then all preceding parts of the path are discarded. This is what the
885"`readlink -f`" command does, though it also edits out "`.`" and
886"`..`" components.
887
888Directly editing the path string is not really necessary when looking
889up a path, and discarding early components is pointless as they aren't
890looked at anyway. Keeping track of all remaining components is
891important, but they can of course be kept separately; there is no need
892to concatenate them. As one symlink may easily refer to another,
893which in turn can refer to a third, we may need to keep the remaining
894components of several paths, each to be processed when the preceding
895ones are completed. These path remnants are kept on a stack of
896limited size.
897
898There are two reasons for placing limits on how many symlinks can
899occur in a single path lookup. The most obvious is to avoid loops.
900If a symlink referred to itself either directly or through
901intermediaries, then following the symlink can never complete
902successfully - the error `ELOOP` must be returned. Loops can be
903detected without imposing limits, but limits are the simplest solution
904and, given the second reason for restriction, quite sufficient.
905
906[outlined recently]: http://thread.gmane.org/gmane.linux.kernel/1934390/focus=1934550
907
908The second reason was [outlined recently] by Linus:
909
910> Because it's a latency and DoS issue too. We need to react well to
911> true loops, but also to "very deep" non-loops. It's not about memory
912> use, it's about users triggering unreasonable CPU resources.
913
914Linux imposes a limit on the length of any pathname: `PATH_MAX`, which
915is 4096. There are a number of reasons for this limit; not letting the
916kernel spend too much time on just one path is one of them. With
917symbolic links you can effectively generate much longer paths so some
918sort of limit is needed for the same reason. Linux imposes a limit of
919at most 40 symlinks in any one path lookup. It previously imposed a
920further limit of eight on the maximum depth of recursion, but that was
921raised to 40 when a separate stack was implemented, so there is now
922just the one limit.
923
924The `nameidata` structure that we met in an earlier article contains a
925small stack that can be used to store the remaining part of up to two
926symlinks. In many cases this will be sufficient. If it isn't, a
927separate stack is allocated with room for 40 symlinks. Pathname
928lookup will never exceed that stack as, once the 40th symlink is
929detected, an error is returned.
930
931It might seem that the name remnants are all that needs to be stored on
932this stack, but we need a bit more. To see that, we need to move on to
933cache lifetimes.
934
935Storage and lifetime of cached symlinks
936---------------------------------------
937
938Like other filesystem resources, such as inodes and directory
939entries, symlinks are cached by Linux to avoid repeated costly access
940to external storage. It is particularly important for RCU-walk to be
941able to find and temporarily hold onto these cached entries, so that
942it doesn't need to drop down into REF-walk.
943
944[object-oriented design pattern]: https://lwn.net/Articles/446317/
945
946While each filesystem is free to make its own choice, symlinks are
947typically stored in one of two places. Short symlinks are often
948stored directly in the inode. When a filesystem allocates a `struct
949inode` it typically allocates extra space to store private data (a
950common [object-oriented design pattern] in the kernel). This will
951sometimes include space for a symlink. The other common location is
952in the page cache, which normally stores the content of files. The
953pathname in a symlink can be seen as the content of that symlink and
954can easily be stored in the page cache just like file content.
955
956When neither of these is suitable, the next most likely scenario is
957that the filesystem will allocate some temporary memory and copy or
958construct the symlink content into that memory whenever it is needed.
959
960When the symlink is stored in the inode, it has the same lifetime as
961the inode which, itself, is protected by RCU or by a counted reference
962on the dentry. This means that the mechanisms that pathname lookup
963uses to access the dcache and icache (inode cache) safely are quite
964sufficient for accessing some cached symlinks safely. In these cases,
965the `i_link` pointer in the inode is set to point to wherever the
966symlink is stored and it can be accessed directly whenever needed.
967
968When the symlink is stored in the page cache or elsewhere, the
969situation is not so straightforward. A reference on a dentry or even
970on an inode does not imply any reference on cached pages of that
971inode, and even an `rcu_read_lock()` is not sufficient to ensure that
972a page will not disappear. So for these symlinks the pathname lookup
973code needs to ask the filesystem to provide a stable reference and,
974significantly, needs to release that reference when it is finished
975with it.
976
977Taking a reference to a cache page is often possible even in RCU-walk
978mode. It does require making changes to memory, which is best avoided,
979but that isn't necessarily a big cost and it is better than dropping
980out of RCU-walk mode completely. Even filesystems that allocate
981space to copy the symlink into can use `GFP_ATOMIC` to often successfully
982allocate memory without the need to drop out of RCU-walk. If a
983filesystem cannot successfully get a reference in RCU-walk mode, it
984must return `-ECHILD` and `unlazy_walk()` will be called to return to
985REF-walk mode in which the filesystem is allowed to sleep.
986
987The place for all this to happen is the `i_op->follow_link()` inode
988method. In the present mainline code this is never actually called in
989RCU-walk mode as the rewrite is not quite complete. It is likely that
990in a future release this method will be passed an `inode` pointer when
991called in RCU-walk mode so it both (1) knows to be careful, and (2) has the
992validated pointer. Much like the `i_op->permission()` method we
993looked at previously, `->follow_link()` would need to be careful that
994all the data structures it references are safe to be accessed while
995holding no counted reference, only the RCU lock. Though getting a
996reference with `->follow_link()` is not yet done in RCU-walk mode, the
997code is ready to release the reference when that does happen.
998
999This need to drop the reference to a symlink adds significant
1000complexity. It requires a reference to the inode so that the
1001`i_op->put_link()` inode operation can be called. In REF-walk, that
1002reference is kept implicitly through a reference to the dentry, so
1003keeping the `struct path` of the symlink is easiest. For RCU-walk,
1004the pointer to the inode is kept separately. To allow switching from
1005RCU-walk back to REF-walk in the middle of processing nested symlinks
1006we also need the seq number for the dentry so we can confirm that
1007switching back was safe.
1008
1009Finally, when providing a reference to a symlink, the filesystem also
1010provides an opaque "cookie" that must be passed to `->put_link()` so that it
1011knows what to free. This might be the allocated memory area, or a
1012pointer to the `struct page` in the page cache, or something else
1013completely. Only the filesystem knows what it is.
1014
1015In order for the reference to each symlink to be dropped when the walk completes,
1016whether in RCU-walk or REF-walk, the symlink stack needs to contain,
1017along with the path remnants:
1018
1019- the `struct path` to provide a reference to the inode in REF-walk
1020- the `struct inode *` to provide a reference to the inode in RCU-walk
1021- the `seq` to allow the path to be safely switched from RCU-walk to REF-walk
1022- the `cookie` that tells `->put_path()` what to put.
1023
1024This means that each entry in the symlink stack needs to hold five
1025pointers and an integer instead of just one pointer (the path
1026remnant). On a 64-bit system, this is about 40 bytes per entry;
1027with 40 entries it adds up to 1600 bytes total, which is less than
1028half a page. So it might seem like a lot, but is by no means
1029excessive.
1030
1031Note that, in a given stack frame, the path remnant (`name`) is not
1032part of the symlink that the other fields refer to. It is the remnant
1033to be followed once that symlink has been fully parsed.
1034
1035Following the symlink
1036---------------------
1037
1038The main loop in `link_path_walk()` iterates seamlessly over all
1039components in the path and all of the non-final symlinks. As symlinks
1040are processed, the `name` pointer is adjusted to point to a new
1041symlink, or is restored from the stack, so that much of the loop
1042doesn't need to notice. Getting this `name` variable on and off the
1043stack is very straightforward; pushing and popping the references is
1044a little more complex.
1045
1046When a symlink is found, `walk_component()` returns the value `1`
1047(`0` is returned for any other sort of success, and a negative number
1048is, as usual, an error indicator). This causes `get_link()` to be
1049called; it then gets the link from the filesystem. Providing that
1050operation is successful, the old path `name` is placed on the stack,
1051and the new value is used as the `name` for a while. When the end of
1052the path is found (i.e. `*name` is `'\0'`) the old `name` is restored
1053off the stack and path walking continues.
1054
1055Pushing and popping the reference pointers (inode, cookie, etc.) is more
1056complex in part because of the desire to handle tail recursion. When
1057the last component of a symlink itself points to a symlink, we
1058want to pop the symlink-just-completed off the stack before pushing
1059the symlink-just-found to avoid leaving empty path remnants that would
1060just get in the way.
1061
1062It is most convenient to push the new symlink references onto the
1063stack in `walk_component()` immediately when the symlink is found;
1064`walk_component()` is also the last piece of code that needs to look at the
1065old symlink as it walks that last component. So it is quite
1066convenient for `walk_component()` to release the old symlink and pop
1067the references just before pushing the reference information for the
1068new symlink. It is guided in this by two flags; `WALK_GET`, which
1069gives it permission to follow a symlink if it finds one, and
1070`WALK_PUT`, which tells it to release the current symlink after it has been
1071followed. `WALK_PUT` is tested first, leading to a call to
1072`put_link()`. `WALK_GET` is tested subsequently (by
1073`should_follow_link()`) leading to a call to `pick_link()` which sets
1074up the stack frame.
1075
1076### Symlinks with no final component ###
1077
1078A pair of special-case symlinks deserve a little further explanation.
1079Both result in a new `struct path` (with mount and dentry) being set
1080up in the `nameidata`, and result in `get_link()` returning `NULL`.
1081
1082The more obvious case is a symlink to "`/`". All symlinks starting
1083with "`/`" are detected in `get_link()` which resets the `nameidata`
1084to point to the effective filesystem root. If the symlink only
1085contains "`/`" then there is nothing more to do, no components at all,
1086so `NULL` is returned to indicate that the symlink can be released and
1087the stack frame discarded.
1088
1089The other case involves things in `/proc` that look like symlinks but
1090aren't really.
1091
1092> $ ls -l /proc/self/fd/1
1093> lrwx------ 1 neilb neilb 64 Jun 13 10:19 /proc/self/fd/1 -> /dev/pts/4
1094
1095Every open file descriptor in any process is represented in `/proc` by
1096something that looks like a symlink. It is really a reference to the
1097target file, not just the name of it. When you `readlink` these
1098objects you get a name that might refer to the same file - unless it
1099has been unlinked or mounted over. When `walk_component()` follows
1100one of these, the `->follow_link()` method in "procfs" doesn't return
1101a string name, but instead calls `nd_jump_link()` which updates the
1102`nameidata` in place to point to that target. `->follow_link()` then
1103returns `NULL`. Again there is no final component and `get_link()`
1104reports this by leaving the `last_type` field of `nameidata` as
1105`LAST_BIND`.
1106
1107Following the symlink in the final component
1108--------------------------------------------
1109
1110All this leads to `link_path_walk()` walking down every component, and
1111following all symbolic links it finds, until it reaches the final
1112component. This is just returned in the `last` field of `nameidata`.
1113For some callers, this is all they need; they want to create that
1114`last` name if it doesn't exist or give an error if it does. Other
1115callers will want to follow a symlink if one is found, and possibly
1116apply special handling to the last component of that symlink, rather
1117than just the last component of the original file name. These callers
1118potentially need to call `link_path_walk()` again and again on
1119successive symlinks until one is found that doesn't point to another
1120symlink.
1121
1122This case is handled by the relevant caller of `link_path_walk()`, such as
1123`path_lookupat()` using a loop that calls `link_path_walk()`, and then
1124handles the final component. If the final component is a symlink
1125that needs to be followed, then `trailing_symlink()` is called to set
1126things up properly and the loop repeats, calling `link_path_walk()`
1127again. This could loop as many as 40 times if the last component of
1128each symlink is another symlink.
1129
1130The various functions that examine the final component and possibly
1131report that it is a symlink are `lookup_last()`, `mountpoint_last()`
1132and `do_last()`, each of which use the same convention as
1133`walk_component()` of returning `1` if a symlink was found that needs
1134to be followed.
1135
1136Of these, `do_last()` is the most interesting as it is used for
1137opening a file. Part of `do_last()` runs with `i_mutex` held and this
1138part is in a separate function: `lookup_open()`.
1139
1140Explaining `do_last()` completely is beyond the scope of this article,
1141but a few highlights should help those interested in exploring the
1142code.
1143
11441. Rather than just finding the target file, `do_last()` needs to open
1145 it. If the file was found in the dcache, then `vfs_open()` is used for
1146 this. If not, then `lookup_open()` will either call `atomic_open()` (if
1147 the filesystem provides it) to combine the final lookup with the open, or
1148 will perform the separate `lookup_real()` and `vfs_create()` steps
1149 directly. In the later case the actual "open" of this newly found or
1150 created file will be performed by `vfs_open()`, just as if the name
1151 were found in the dcache.
1152
11532. `vfs_open()` can fail with `-EOPENSTALE` if the cached information
1154 wasn't quite current enough. Rather than restarting the lookup from
1155 the top with `LOOKUP_REVAL` set, `lookup_open()` is called instead,
1156 giving the filesystem a chance to resolve small inconsistencies.
1157 If that doesn't work, only then is the lookup restarted from the top.
1158
11593. An open with O_CREAT **does** follow a symlink in the final component,
1160 unlike other creation system calls (like `mkdir`). So the sequence:
1161
1162 > ln -s bar /tmp/foo
1163 > echo hello > /tmp/foo
1164
1165 will create a file called `/tmp/bar`. This is not permitted if
1166 `O_EXCL` is set but otherwise is handled for an O_CREAT open much
1167 like for a non-creating open: `should_follow_link()` returns `1`, and
1168 so does `do_last()` so that `trailing_symlink()` gets called and the
1169 open process continues on the symlink that was found.
1170
1171Updating the access time
1172------------------------
1173
1174We previously said of RCU-walk that it would "take no locks, increment
1175no counts, leave no footprints." We have since seen that some
1176"footprints" can be needed when handling symlinks as a counted
1177reference (or even a memory allocation) may be needed. But these
1178footprints are best kept to a minimum.
1179
1180One other place where walking down a symlink can involve leaving
1181footprints in a way that doesn't affect directories is in updating access times.
1182In Unix (and Linux) every filesystem object has a "last accessed
1183time", or "`atime`". Passing through a directory to access a file
1184within is not considered to be an access for the purposes of
1185`atime`; only listing the contents of a directory can update its `atime`.
1186Symlinks are different it seems. Both reading a symlink (with `readlink()`)
1187and looking up a symlink on the way to some other destination can
1188update the atime on that symlink.
1189
1190[clearest statement]: http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_08
1191
1192It is not clear why this is the case; POSIX has little to say on the
1193subject. The [clearest statement] is that, if a particular implementation
1194updates a timestamp in a place not specified by POSIX, this must be
1195documented "except that any changes caused by pathname resolution need
1196not be documented". This seems to imply that POSIX doesn't really
1197care about access-time updates during pathname lookup.
1198
1199[Linux 1.3.87]: https://git.kernel.org/cgit/linux/kernel/git/history/history.git/diff/fs/ext2/symlink.c?id=f806c6db77b8eaa6e00dcfb6b567706feae8dbb8
1200
1201An examination of history shows that prior to [Linux 1.3.87], the ext2
1202filesystem, at least, didn't update atime when following a link.
1203Unfortunately we have no record of why that behavior was changed.
1204
1205In any case, access time must now be updated and that operation can be
1206quite complex. Trying to stay in RCU-walk while doing it is best
1207avoided. Fortunately it is often permitted to skip the `atime`
1208update. Because `atime` updates cause performance problems in various
1209areas, Linux supports the `relatime` mount option, which generally
1210limits the updates of `atime` to once per day on files that aren't
1211being changed (and symlinks never change once created). Even without
1212`relatime`, many filesystems record `atime` with a one-second
1213granularity, so only one update per second is required.
1214
1215It is easy to test if an `atime` update is needed while in RCU-walk
1216mode and, if it isn't, the update can be skipped and RCU-walk mode
1217continues. Only when an `atime` update is actually required does the
1218path walk drop down to REF-walk. All of this is handled in the
1219`get_link()` function.
1220
1221A few flags
1222-----------
1223
1224A suitable way to wrap up this tour of pathname walking is to list
1225the various flags that can be stored in the `nameidata` to guide the
1226lookup process. Many of these are only meaningful on the final
1227component, others reflect the current state of the pathname lookup.
1228And then there is `LOOKUP_EMPTY`, which doesn't fit conceptually with
1229the others. If this is not set, an empty pathname causes an error
1230very early on. If it is set, empty pathnames are not considered to be
1231an error.
1232
1233### Global state flags ###
1234
1235We have already met two global state flags: `LOOKUP_RCU` and
1236`LOOKUP_REVAL`. These select between one of three overall approaches
1237to lookup: RCU-walk, REF-walk, and REF-walk with forced revalidation.
1238
1239`LOOKUP_PARENT` indicates that the final component hasn't been reached
1240yet. This is primarily used to tell the audit subsystem the full
1241context of a particular access being audited.
1242
1243`LOOKUP_ROOT` indicates that the `root` field in the `nameidata` was
1244provided by the caller, so it shouldn't be released when it is no
1245longer needed.
1246
1247`LOOKUP_JUMPED` means that the current dentry was chosen not because
1248it had the right name but for some other reason. This happens when
1249following "`..`", following a symlink to `/`, crossing a mount point
1250or accessing a "`/proc/$PID/fd/$FD`" symlink. In this case the
1251filesystem has not been asked to revalidate the name (with
1252`d_revalidate()`). In such cases the inode may still need to be
1253revalidated, so `d_op->d_weak_revalidate()` is called if
1254`LOOKUP_JUMPED` is set when the look completes - which may be at the
1255final component or, when creating, unlinking, or renaming, at the penultimate component.
1256
1257### Final-component flags ###
1258
1259Some of these flags are only set when the final component is being
1260considered. Others are only checked for when considering that final
1261component.
1262
1263`LOOKUP_AUTOMOUNT` ensures that, if the final component is an automount
1264point, then the mount is triggered. Some operations would trigger it
1265anyway, but operations like `stat()` deliberately don't. `statfs()`
1266needs to trigger the mount but otherwise behaves a lot like `stat()`, so
1267it sets `LOOKUP_AUTOMOUNT`, as does "`quotactl()`" and the handling of
1268"`mount --bind`".
1269
1270`LOOKUP_FOLLOW` has a similar function to `LOOKUP_AUTOMOUNT` but for
1271symlinks. Some system calls set or clear it implicitly, while
1272others have API flags such as `AT_SYMLINK_FOLLOW` and
1273`UMOUNT_NOFOLLOW` to control it. Its effect is similar to
1274`WALK_GET` that we already met, but it is used in a different way.
1275
1276`LOOKUP_DIRECTORY` insists that the final component is a directory.
1277Various callers set this and it is also set when the final component
1278is found to be followed by a slash.
1279
1280Finally `LOOKUP_OPEN`, `LOOKUP_CREATE`, `LOOKUP_EXCL`, and
1281`LOOKUP_RENAME_TARGET` are not used directly by the VFS but are made
1282available to the filesystem and particularly the `->d_revalidate()`
1283method. A filesystem can choose not to bother revalidating too hard
1284if it knows that it will be asked to open or create the file soon.
1285These flags were previously useful for `->lookup()` too but with the
1286introduction of `->atomic_open()` they are less relevant there.
1287
1288End of the road
1289---------------
1290
1291Despite its complexity, all this pathname lookup code appears to be
1292in good shape - various parts are certainly easier to understand now
1293than even a couple of releases ago. But that doesn't mean it is
1294"finished". As already mentioned, RCU-walk currently only follows
1295symlinks that are stored in the inode so, while it handles many ext4
1296symlinks, it doesn't help with NFS, XFS, or Btrfs. That support
1297is not likely to be long delayed.