1 =============== 2 Pathname lookup 3 =============== 4 5 This write-up is based on three articles published at lwn.net: 6 7 - <https://lwn.net/Articles/649115/> Pathname lookup in Linux 8 - <https://lwn.net/Articles/649729/> RCU-walk: faster pathname lookup in Linux 9 - <https://lwn.net/Articles/650786/> A walk among the symlinks 10 11 Written by Neil Brown with help from Al Viro and Jon Corbet. 12 It has subsequently been updated to reflect changes in the kernel 13 including: 14 15 - per-directory parallel name lookup. 16 - ``openat2()`` resolution restriction flags. 17 18 Introduction to pathname lookup 19 =============================== 20 21 The most obvious aspect of pathname lookup, which very little 22 exploration is needed to discover, is that it is complex. There are 23 many rules, special cases, and implementation alternatives that all 24 combine to confuse the unwary reader. Computer science has long been 25 acquainted with such complexity and has tools to help manage it. One 26 tool that we will make extensive use of is "divide and conquer". For 27 the early parts of the analysis we will divide off symlinks - leaving 28 them until the final part. Well before we get to symlinks we have 29 another major division based on the VFS's approach to locking which 30 will allow us to review "REF-walk" and "RCU-walk" separately. But we 31 are getting ahead of ourselves. There are some important low level 32 distinctions we need to clarify first. 33 34 There are two sorts of ... 35 -------------------------- 36 37 .. _openat: http://man7.org/linux/man-pages/man2/openat.2.html 38 39 Pathnames (sometimes "file names"), used to identify objects in the 40 filesystem, will be familiar to most readers. They contain two sorts 41 of elements: "slashes" that are sequences of one or more "``/``" 42 characters, and "components" that are sequences of one or more 43 non-"``/``" characters. These form two kinds of paths. Those that 44 start with slashes are "absolute" and start from the filesystem root. 45 The others are "relative" and start from the current directory, or 46 from some other location specified by a file descriptor given to 47 "``*at()``" system calls such as `openat() <openat_>`_. 48 49 .. _execveat: http://man7.org/linux/man-pages/man2/execveat.2.html 50 51 It is tempting to describe the second kind as starting with a 52 component, but that isn't always accurate: a pathname can lack both 53 slashes and components, it can be empty, in other words. This is 54 generally forbidden in POSIX, but some of those "``*at()``" system calls 55 in Linux permit it when the ``AT_EMPTY_PATH`` flag is given. For 56 example, if you have an open file descriptor on an executable file you 57 can execute it by calling `execveat() <execveat_>`_ passing 58 the file descriptor, an empty path, and the ``AT_EMPTY_PATH`` flag. 59 60 These paths can be divided into two sections: the final component and 61 everything else. The "everything else" is the easy bit. In all cases 62 it must identify a directory that already exists, otherwise an error 63 such as ``ENOENT`` or ``ENOTDIR`` will be reported. 64 65 The final component is not so simple. Not only do different system 66 calls interpret it quite differently (e.g. some create it, some do 67 not), but it might not even exist: neither the empty pathname nor the 68 pathname that is just slashes have a final component. If it does 69 exist, it could be "``.``" or "``..``" which are handled quite differently 70 from other components. 71 72 .. _POSIX: https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_12 73 74 If a pathname ends with a slash, such as "``/tmp/foo/``" it might be 75 tempting to consider that to have an empty final component. In many 76 ways that would lead to correct results, but not always. In 77 particular, ``mkdir()`` and ``rmdir()`` each create or remove a directory named 78 by the final component, and they are required to work with pathnames 79 ending in "``/``". According to POSIX_: 80 81 A pathname that contains at least one non-<slash> character and 82 that ends with one or more trailing <slash> characters shall not 83 be resolved successfully unless the last pathname component before 84 the trailing <slash> characters names an existing directory or a 85 directory entry that is to be created for a directory immediately 86 after the pathname is resolved. 87 88 The Linux pathname walking code (mostly in ``fs/namei.c``) deals with 89 all of these issues: breaking the path into components, handling the 90 "everything else" quite separately from the final component, and 91 checking that the trailing slash is not used where it isn't 92 permitted. It also addresses the important issue of concurrent 93 access. 94 95 While one process is looking up a pathname, another might be making 96 changes that affect that lookup. One fairly extreme case is that if 97 "a/b" were renamed to "a/c/b" while another process were looking up 98 "a/b/..", that process might successfully resolve on "a/c". 99 Most races are much more subtle, and a big part of the task of 100 pathname lookup is to prevent them from having damaging effects. Many 101 of the possible races are seen most clearly in the context of the 102 "dcache" and an understanding of that is central to understanding 103 pathname lookup. 104 105 More than just a cache 106 ---------------------- 107 108 The "dcache" caches information about names in each filesystem to 109 make them quickly available for lookup. Each entry (known as a 110 "dentry") contains three significant fields: a component name, a 111 pointer to a parent dentry, and a pointer to the "inode" which 112 contains further information about the object in that parent with 113 the given name. The inode pointer can be ``NULL`` indicating that the 114 name doesn't exist in the parent. While there can be linkage in the 115 dentry of a directory to the dentries of the children, that linkage is 116 not used for pathname lookup, and so will not be considered here. 117 118 The dcache has a number of uses apart from accelerating lookup. One 119 that will be particularly relevant is that it is closely integrated 120 with the mount table that records which filesystem is mounted where. 121 What the mount table actually stores is which dentry is mounted on top 122 of which other dentry. 123 124 When considering the dcache, we have another of our "two types" 125 distinctions: there are two types of filesystems. 126 127 Some filesystems ensure that the information in the dcache is always 128 completely accurate (though not necessarily complete). This can allow 129 the VFS to determine if a particular file does or doesn't exist 130 without checking with the filesystem, and means that the VFS can 131 protect the filesystem against certain races and other problems. 132 These are typically "local" filesystems such as ext3, XFS, and Btrfs. 133 134 Other filesystems don't provide that guarantee because they cannot. 135 These are typically filesystems that are shared across a network, 136 whether remote filesystems like NFS and 9P, or cluster filesystems 137 like ocfs2 or cephfs. These filesystems allow the VFS to revalidate 138 cached information, and must provide their own protection against 139 awkward races. The VFS can detect these filesystems by the 140 ``DCACHE_OP_REVALIDATE`` flag being set in the dentry. 141 142 REF-walk: simple concurrency management with refcounts and spinlocks 143 -------------------------------------------------------------------- 144 145 With all of those divisions carefully classified, we can now start 146 looking at the actual process of walking along a path. In particular 147 we will start with the handling of the "everything else" part of a 148 pathname, and focus on the "REF-walk" approach to concurrency 149 management. This code is found in the ``link_path_walk()`` function, if 150 you ignore all the places that only run when "``LOOKUP_RCU``" 151 (indicating the use of RCU-walk) is set. 152 153 .. _Meet the Lockers: https://lwn.net/Articles/453685/ 154 155 REF-walk is fairly heavy-handed with locks and reference counts. Not 156 as heavy-handed as in the old "big kernel lock" days, but certainly not 157 afraid of taking a lock when one is needed. It uses a variety of 158 different concurrency controls. A background understanding of the 159 various primitives is assumed, or can be gleaned from elsewhere such 160 as in `Meet the Lockers`_. 161 162 The locking mechanisms used by REF-walk include: 163 164 dentry->d_lockref 165 ~~~~~~~~~~~~~~~~~ 166 167 This uses the lockref primitive to provide both a spinlock and a 168 reference count. The special-sauce of this primitive is that the 169 conceptual sequence "lock; inc_ref; unlock;" can often be performed 170 with a single atomic memory operation. 171 172 Holding a reference on a dentry ensures that the dentry won't suddenly 173 be freed and used for something else, so the values in various fields 174 will behave as expected. It also protects the ``->d_inode`` reference 175 to the inode to some extent. 176 177 The association between a dentry and its inode is fairly permanent. 178 For example, when a file is renamed, the dentry and inode move 179 together to the new location. When a file is created the dentry will 180 initially be negative (i.e. ``d_inode`` is ``NULL``), and will be assigned 181 to the new inode as part of the act of creation. 182 183 When a file is deleted, this can be reflected in the cache either by 184 setting ``d_inode`` to ``NULL``, or by removing it from the hash table 185 (described shortly) used to look up the name in the parent directory. 186 If the dentry is still in use the second option is used as it is 187 perfectly legal to keep using an open file after it has been deleted 188 and having the dentry around helps. If the dentry is not otherwise in 189 use (i.e. if the refcount in ``d_lockref`` is one), only then will 190 ``d_inode`` be set to ``NULL``. Doing it this way is more efficient for a 191 very common case. 192 193 So as long as a counted reference is held to a dentry, a non-``NULL`` ``->d_inode`` 194 value will never be changed. 195 196 dentry->d_lock 197 ~~~~~~~~~~~~~~ 198 199 ``d_lock`` is a synonym for the spinlock that is part of ``d_lockref`` above. 200 For our purposes, holding this lock protects against the dentry being 201 renamed or unlinked. In particular, its parent (``d_parent``), and its 202 name (``d_name``) cannot be changed, and it cannot be removed from the 203 dentry hash table. 204 205 When looking for a name in a directory, REF-walk takes ``d_lock`` on 206 each candidate dentry that it finds in the hash table and then checks 207 that the parent and name are correct. So it doesn't lock the parent 208 while searching in the cache; it only locks children. 209 210 When looking for the parent for a given name (to handle "``..``"), 211 REF-walk can take ``d_lock`` to get a stable reference to ``d_parent``, 212 but it first tries a more lightweight approach. As seen in 213 ``dget_parent()``, if a reference can be claimed on the parent, and if 214 subsequently ``d_parent`` can be seen to have not changed, then there is 215 no need to actually take the lock on the child. 216 217 rename_lock 218 ~~~~~~~~~~~ 219 220 Looking up a given name in a given directory involves computing a hash 221 from the two values (the name and the dentry of the directory), 222 accessing that slot in a hash table, and searching the linked list 223 that is found there. 224 225 When a dentry is renamed, the name and the parent dentry can both 226 change so the hash will almost certainly change too. This would move the 227 dentry to a different chain in the hash table. If a filename search 228 happened to be looking at a dentry that was moved in this way, 229 it might end up continuing the search down the wrong chain, 230 and so miss out on part of the correct chain. 231 232 The name-lookup process (``d_lookup()``) does *not* try to prevent this 233 from happening, but only to detect when it happens. 234 ``rename_lock`` is a seqlock that is updated whenever any dentry is 235 renamed. If ``d_lookup`` finds that a rename happened while it 236 unsuccessfully scanned a chain in the hash table, it simply tries 237 again. 238 239 ``rename_lock`` is also used to detect and defend against potential attacks 240 against ``LOOKUP_BENEATH`` and ``LOOKUP_IN_ROOT`` when resolving ".." (where 241 the parent directory is moved outside the root, bypassing the ``path_equal()`` 242 check). If ``rename_lock`` is updated during the lookup and the path encounters 243 a "..", a potential attack occurred and ``handle_dots()`` will bail out with 244 ``-EAGAIN``. 245 246 inode->i_rwsem 247 ~~~~~~~~~~~~~~ 248 249 ``i_rwsem`` is a read/write semaphore that serializes all changes to a particular 250 directory. This ensures that, for example, an ``unlink()`` and a ``rename()`` 251 cannot both happen at the same time. It also keeps the directory 252 stable while the filesystem is asked to look up a name that is not 253 currently in the dcache or, optionally, when the list of entries in a 254 directory is being retrieved with ``readdir()``. 255 256 This has a complementary role to that of ``d_lock``: ``i_rwsem`` on a 257 directory protects all of the names in that directory, while ``d_lock`` 258 on a name protects just one name in a directory. Most changes to the 259 dcache hold ``i_rwsem`` on the relevant directory inode and briefly take 260 ``d_lock`` on one or more the dentries while the change happens. One 261 exception is when idle dentries are removed from the dcache due to 262 memory pressure. This uses ``d_lock``, but ``i_rwsem`` plays no role. 263 264 The semaphore affects pathname lookup in two distinct ways. Firstly it 265 prevents changes during lookup of a name in a directory. ``walk_component()`` uses 266 ``lookup_fast()`` first which, in turn, checks to see if the name is in the cache, 267 using only ``d_lock`` locking. If the name isn't found, then ``walk_component()`` 268 falls back to ``lookup_slow()`` which takes a shared lock on ``i_rwsem``, checks again that 269 the name isn't in the cache, and then calls in to the filesystem to get a 270 definitive answer. A new dentry will be added to the cache regardless of 271 the result. 272 273 Secondly, when pathname lookup reaches the final component, it will 274 sometimes need to take an exclusive lock on ``i_rwsem`` before performing the last lookup so 275 that the required exclusion can be achieved. How path lookup chooses 276 to take, or not take, ``i_rwsem`` is one of the 277 issues addressed in a subsequent section. 278 279 If two threads attempt to look up the same name at the same time - a 280 name that is not yet in the dcache - the shared lock on ``i_rwsem`` will 281 not prevent them both adding new dentries with the same name. As this 282 would result in confusion an extra level of interlocking is used, 283 based around a secondary hash table (``in_lookup_hashtable``) and a 284 per-dentry flag bit (``DCACHE_PAR_LOOKUP``). 285 286 To add a new dentry to the cache while only holding a shared lock on 287 ``i_rwsem``, a thread must call ``d_alloc_parallel()``. This allocates a 288 dentry, stores the required name and parent in it, checks if there 289 is already a matching dentry in the primary or secondary hash 290 tables, and if not, stores the newly allocated dentry in the secondary 291 hash table, with ``DCACHE_PAR_LOOKUP`` set. 292 293 If a matching dentry was found in the primary hash table then that is 294 returned and the caller can know that it lost a race with some other 295 thread adding the entry. If no matching dentry is found in either 296 cache, the newly allocated dentry is returned and the caller can 297 detect this from the presence of ``DCACHE_PAR_LOOKUP``. In this case it 298 knows that it has won any race and now is responsible for asking the 299 filesystem to perform the lookup and find the matching inode. When 300 the lookup is complete, it must call ``d_lookup_done()`` which clears 301 the flag and does some other house keeping, including removing the 302 dentry from the secondary hash table - it will normally have been 303 added to the primary hash table already. Note that a ``struct 304 waitqueue_head`` is passed to ``d_alloc_parallel()``, and 305 ``d_lookup_done()`` must be called while this ``waitqueue_head`` is still 306 in scope. 307 308 If a matching dentry is found in the secondary hash table, 309 ``d_alloc_parallel()`` has a little more work to do. It first waits for 310 ``DCACHE_PAR_LOOKUP`` to be cleared, using a wait_queue that was passed 311 to the instance of ``d_alloc_parallel()`` that won the race and that 312 will be woken by the call to ``d_lookup_done()``. It then checks to see 313 if the dentry has now been added to the primary hash table. If it 314 has, the dentry is returned and the caller just sees that it lost any 315 race. If it hasn't been added to the primary hash table, the most 316 likely explanation is that some other dentry was added instead using 317 ``d_splice_alias()``. In any case, ``d_alloc_parallel()`` repeats all the 318 look ups from the start and will normally return something from the 319 primary hash table. 320 321 mnt->mnt_count 322 ~~~~~~~~~~~~~~ 323 324 ``mnt_count`` is a per-CPU reference counter on "``mount``" structures. 325 Per-CPU here means that incrementing the count is cheap as it only 326 uses CPU-local memory, but checking if the count is zero is expensive as 327 it needs to check with every CPU. Taking a ``mnt_count`` reference 328 prevents the mount structure from disappearing as the result of regular 329 unmount operations, but does not prevent a "lazy" unmount. So holding 330 ``mnt_count`` doesn't ensure that the mount remains in the namespace and, 331 in particular, doesn't stabilize the link to the mounted-on dentry. It 332 does, however, ensure that the ``mount`` data structure remains coherent, 333 and it provides a reference to the root dentry of the mounted 334 filesystem. So a reference through ``->mnt_count`` provides a stable 335 reference to the mounted dentry, but not the mounted-on dentry. 336 337 mount_lock 338 ~~~~~~~~~~ 339 340 ``mount_lock`` is a global seqlock, a bit like ``rename_lock``. It can be used to 341 check if any change has been made to any mount points. 342 343 While walking down the tree (away from the root) this lock is used when 344 crossing a mount point to check that the crossing was safe. That is, 345 the value in the seqlock is read, then the code finds the mount that 346 is mounted on the current directory, if there is one, and increments 347 the ``mnt_count``. Finally the value in ``mount_lock`` is checked against 348 the old value. If there is no change, then the crossing was safe. If there 349 was a change, the ``mnt_count`` is decremented and the whole process is 350 retried. 351 352 When walking up the tree (towards the root) by following a ".." link, 353 a little more care is needed. In this case the seqlock (which 354 contains both a counter and a spinlock) is fully locked to prevent 355 any changes to any mount points while stepping up. This locking is 356 needed to stabilize the link to the mounted-on dentry, which the 357 refcount on the mount itself doesn't ensure. 358 359 ``mount_lock`` is also used to detect and defend against potential attacks 360 against ``LOOKUP_BENEATH`` and ``LOOKUP_IN_ROOT`` when resolving ".." (where 361 the parent directory is moved outside the root, bypassing the ``path_equal()`` 362 check). If ``mount_lock`` is updated during the lookup and the path encounters 363 a "..", a potential attack occurred and ``handle_dots()`` will bail out with 364 ``-EAGAIN``. 365 366 RCU 367 ~~~ 368 369 Finally the global (but extremely lightweight) RCU read lock is held 370 from time to time to ensure certain data structures don't get freed 371 unexpectedly. 372 373 In particular it is held while scanning chains in the dcache hash 374 table, and the mount point hash table. 375 376 Bringing it together with ``struct nameidata`` 377 ---------------------------------------------- 378 379 .. _First edition Unix: https://minnie.tuhs.org/cgi-bin/utree.pl?file=V1/u2.s 380 381 Throughout the process of walking a path, the current status is stored 382 in a ``struct nameidata``, "namei" being the traditional name - dating 383 all the way back to `First Edition Unix`_ - of the function that 384 converts a "name" to an "inode". ``struct nameidata`` contains (among 385 other fields): 386 387 ``struct path path`` 388 ~~~~~~~~~~~~~~~~~~~~ 389 390 A ``path`` contains a ``struct vfsmount`` (which is 391 embedded in a ``struct mount``) and a ``struct dentry``. Together these 392 record the current status of the walk. They start out referring to the 393 starting point (the current working directory, the root directory, or some other 394 directory identified by a file descriptor), and are updated on each 395 step. A reference through ``d_lockref`` and ``mnt_count`` is always 396 held. 397 398 ``struct qstr last`` 399 ~~~~~~~~~~~~~~~~~~~~ 400 401 This is a string together with a length (i.e. *not* ``nul`` terminated) 402 that is the "next" component in the pathname. 403 404 ``int last_type`` 405 ~~~~~~~~~~~~~~~~~ 406 407 This is one of ``LAST_NORM``, ``LAST_ROOT``, ``LAST_DOT`` or ``LAST_DOTDOT``. 408 The ``last`` field is only valid if the type is ``LAST_NORM``. 409 410 ``struct path root`` 411 ~~~~~~~~~~~~~~~~~~~~ 412 413 This is used to hold a reference to the effective root of the 414 filesystem. Often that reference won't be needed, so this field is 415 only assigned the first time it is used, or when a non-standard root 416 is requested. Keeping a reference in the ``nameidata`` ensures that 417 only one root is in effect for the entire path walk, even if it races 418 with a ``chroot()`` system call. 419 420 It should be noted that in the case of ``LOOKUP_IN_ROOT`` or 421 ``LOOKUP_BENEATH``, the effective root becomes the directory file descriptor 422 passed to ``openat2()`` (which exposes these ``LOOKUP_`` flags). 423 424 The root is needed when either of two conditions holds: (1) either the 425 pathname or a symbolic link starts with a "'/'", or (2) a "``..``" 426 component is being handled, since "``..``" from the root must always stay 427 at the root. The value used is usually the current root directory of 428 the calling process. An alternate root can be provided as when 429 ``sysctl()`` calls ``file_open_root()``, and when NFSv4 or Btrfs call 430 ``mount_subtree()``. In each case a pathname is being looked up in a very 431 specific part of the filesystem, and the lookup must not be allowed to 432 escape that subtree. It works a bit like a local ``chroot()``. 433 434 Ignoring the handling of symbolic links, we can now describe the 435 "``link_path_walk()``" function, which handles the lookup of everything 436 except the final component as: 437 438 Given a path (``name``) and a nameidata structure (``nd``), check that the 439 current directory has execute permission and then advance ``name`` 440 over one component while updating ``last_type`` and ``last``. If that 441 was the final component, then return, otherwise call 442 ``walk_component()`` and repeat from the top. 443 444 ``walk_component()`` is even easier. If the component is ``LAST_DOTS``, 445 it calls ``handle_dots()`` which does the necessary locking as already 446 described. If it finds a ``LAST_NORM`` component it first calls 447 "``lookup_fast()``" which only looks in the dcache, but will ask the 448 filesystem to revalidate the result if it is that sort of filesystem. 449 If that doesn't get a good result, it calls "``lookup_slow()``" which 450 takes ``i_rwsem``, rechecks the cache, and then asks the filesystem 451 to find a definitive answer. 452 453 As the last step of walk_component(), step_into() will be called either 454 directly from walk_component() or from handle_dots(). It calls 455 handle_mounts(), to check and handle mount points, in which a new 456 ``struct path`` is created containing a counted reference to the new dentry and 457 a reference to the new ``vfsmount`` which is only counted if it is 458 different from the previous ``vfsmount``. Then if there is 459 a symbolic link, step_into() calls pick_link() to deal with it, 460 otherwise it installs the new ``struct path`` in the ``struct nameidata``, and 461 drops the unneeded references. 462 463 This "hand-over-hand" sequencing of getting a reference to the new 464 dentry before dropping the reference to the previous dentry may 465 seem obvious, but is worth pointing out so that we will recognize its 466 analogue in the "RCU-walk" version. 467 468 Handling the final component 469 ---------------------------- 470 471 ``link_path_walk()`` only walks as far as setting ``nd->last`` and 472 ``nd->last_type`` to refer to the final component of the path. It does 473 not call ``walk_component()`` that last time. Handling that final 474 component remains for the caller to sort out. Those callers are 475 path_lookupat(), path_parentat() and 476 path_openat() each of which handles the differing requirements of 477 different system calls. 478 479 ``path_parentat()`` is clearly the simplest - it just wraps a little bit 480 of housekeeping around ``link_path_walk()`` and returns the parent 481 directory and final component to the caller. The caller will be either 482 aiming to create a name (via ``filename_create()``) or remove or rename 483 a name (in which case ``user_path_parent()`` is used). They will use 484 ``i_rwsem`` to exclude other changes while they validate and then 485 perform their operation. 486 487 ``path_lookupat()`` is nearly as simple - it is used when an existing 488 object is wanted such as by ``stat()`` or ``chmod()``. It essentially just 489 calls ``walk_component()`` on the final component through a call to 490 ``lookup_last()``. ``path_lookupat()`` returns just the final dentry. 491 It is worth noting that when flag ``LOOKUP_MOUNTPOINT`` is set, 492 path_lookupat() will unset LOOKUP_JUMPED in nameidata so that in the 493 subsequent path traversal d_weak_revalidate() won't be called. 494 This is important when unmounting a filesystem that is inaccessible, such as 495 one provided by a dead NFS server. 496 497 Finally ``path_openat()`` is used for the ``open()`` system call; it 498 contains, in support functions starting with "open_last_lookups()", all the 499 complexity needed to handle the different subtleties of O_CREAT (with 500 or without O_EXCL), final "``/``" characters, and trailing symbolic 501 links. We will revisit this in the final part of this series, which 502 focuses on those symbolic links. "open_last_lookups()" will sometimes, but 503 not always, take ``i_rwsem``, depending on what it finds. 504 505 Each of these, or the functions which call them, need to be alert to 506 the possibility that the final component is not ``LAST_NORM``. If the 507 goal of the lookup is to create something, then any value for 508 ``last_type`` other than ``LAST_NORM`` will result in an error. For 509 example if ``path_parentat()`` reports ``LAST_DOTDOT``, then the caller 510 won't try to create that name. They also check for trailing slashes 511 by testing ``last.name[last.len]``. If there is any character beyond 512 the final component, it must be a trailing slash. 513 514 Revalidation and automounts 515 --------------------------- 516 517 Apart from symbolic links, there are only two parts of the "REF-walk" 518 process not yet covered. One is the handling of stale cache entries 519 and the other is automounts. 520 521 On filesystems that require it, the lookup routines will call the 522 ``->d_revalidate()`` dentry method to ensure that the cached information 523 is current. This will often confirm validity or update a few details 524 from a server. In some cases it may find that there has been change 525 further up the path and that something that was thought to be valid 526 previously isn't really. When this happens the lookup of the whole 527 path is aborted and retried with the "``LOOKUP_REVAL``" flag set. This 528 forces revalidation to be more thorough. We will see more details of 529 this retry process in the next article. 530 531 Automount points are locations in the filesystem where an attempt to 532 lookup a name can trigger changes to how that lookup should be 533 handled, in particular by mounting a filesystem there. These are 534 covered in greater detail in autofs.txt in the Linux documentation 535 tree, but a few notes specifically related to path lookup are in order 536 here. 537 538 The Linux VFS has a concept of "managed" dentries. There are three 539 potentially interesting things about these dentries corresponding 540 to three different flags that might be set in ``dentry->d_flags``: 541 542 ``DCACHE_MANAGE_TRANSIT`` 543 ~~~~~~~~~~~~~~~~~~~~~~~~~ 544 545 If this flag has been set, then the filesystem has requested that the 546 ``d_manage()`` dentry operation be called before handling any possible 547 mount point. This can perform two particular services: 548 549 It can block to avoid races. If an automount point is being 550 unmounted, the ``d_manage()`` function will usually wait for that 551 process to complete before letting the new lookup proceed and possibly 552 trigger a new automount. 553 554 It can selectively allow only some processes to transit through a 555 mount point. When a server process is managing automounts, it may 556 need to access a directory without triggering normal automount 557 processing. That server process can identify itself to the ``autofs`` 558 filesystem, which will then give it a special pass through 559 ``d_manage()`` by returning ``-EISDIR``. 560 561 ``DCACHE_MOUNTED`` 562 ~~~~~~~~~~~~~~~~~~ 563 564 This flag is set on every dentry that is mounted on. As Linux 565 supports multiple filesystem namespaces, it is possible that the 566 dentry may not be mounted on in *this* namespace, just in some 567 other. So this flag is seen as a hint, not a promise. 568 569 If this flag is set, and ``d_manage()`` didn't return ``-EISDIR``, 570 ``lookup_mnt()`` is called to examine the mount hash table (honoring the 571 ``mount_lock`` described earlier) and possibly return a new ``vfsmount`` 572 and a new ``dentry`` (both with counted references). 573 574 ``DCACHE_NEED_AUTOMOUNT`` 575 ~~~~~~~~~~~~~~~~~~~~~~~~~ 576 577 If ``d_manage()`` allowed us to get this far, and ``lookup_mnt()`` didn't 578 find a mount point, then this flag causes the ``d_automount()`` dentry 579 operation to be called. 580 581 The ``d_automount()`` operation can be arbitrarily complex and may 582 communicate with server processes etc. but it should ultimately either 583 report that there was an error, that there was nothing to mount, or 584 should provide an updated ``struct path`` with new ``dentry`` and ``vfsmount``. 585 586 In the latter case, ``finish_automount()`` will be called to safely 587 install the new mount point into the mount table. 588 589 There is no new locking of import here and it is important that no 590 locks (only counted references) are held over this processing due to 591 the very real possibility of extended delays. 592 This will become more important next time when we examine RCU-walk 593 which is particularly sensitive to delays. 594 595 RCU-walk - faster pathname lookup in Linux 596 ========================================== 597 598 RCU-walk is another algorithm for performing pathname lookup in Linux. 599 It is in many ways similar to REF-walk and the two share quite a bit 600 of code. The significant difference in RCU-walk is how it allows for 601 the possibility of concurrent access. 602 603 We noted that REF-walk is complex because there are numerous details 604 and special cases. RCU-walk reduces this complexity by simply 605 refusing to handle a number of cases -- it instead falls back to 606 REF-walk. The difficulty with RCU-walk comes from a different 607 direction: unfamiliarity. The locking rules when depending on RCU are 608 quite different from traditional locking, so we will spend a little extra 609 time when we come to those. 610 611 Clear demarcation of roles 612 -------------------------- 613 614 The easiest way to manage concurrency is to forcibly stop any other 615 thread from changing the data structures that a given thread is 616 looking at. In cases where no other thread would even think of 617 changing the data and lots of different threads want to read at the 618 same time, this can be very costly. Even when using locks that permit 619 multiple concurrent readers, the simple act of updating the count of 620 the number of current readers can impose an unwanted cost. So the 621 goal when reading a shared data structure that no other process is 622 changing is to avoid writing anything to memory at all. Take no 623 locks, increment no counts, leave no footprints. 624 625 The REF-walk mechanism already described certainly doesn't follow this 626 principle, but then it is really designed to work when there may well 627 be other threads modifying the data. RCU-walk, in contrast, is 628 designed for the common situation where there are lots of frequent 629 readers and only occasional writers. This may not be common in all 630 parts of the filesystem tree, but in many parts it will be. For the 631 other parts it is important that RCU-walk can quickly fall back to 632 using REF-walk. 633 634 Pathname lookup always starts in RCU-walk mode but only remains there 635 as long as what it is looking for is in the cache and is stable. It 636 dances lightly down the cached filesystem image, leaving no footprints 637 and carefully watching where it is, to be sure it doesn't trip. If it 638 notices that something has changed or is changing, or if something 639 isn't in the cache, then it tries to stop gracefully and switch to 640 REF-walk. 641 642 This stopping requires getting a counted reference on the current 643 ``vfsmount`` and ``dentry``, and ensuring that these are still valid - 644 that a path walk with REF-walk would have found the same entries. 645 This is an invariant that RCU-walk must guarantee. It can only make 646 decisions, such as selecting the next step, that are decisions which 647 REF-walk could also have made if it were walking down the tree at the 648 same time. If the graceful stop succeeds, the rest of the path is 649 processed with the reliable, if slightly sluggish, REF-walk. If 650 RCU-walk finds it cannot stop gracefully, it simply gives up and 651 restarts from the top with REF-walk. 652 653 This pattern of "try RCU-walk, if that fails try REF-walk" can be 654 clearly seen in functions like filename_lookup(), 655 filename_parentat(), 656 do_filp_open(), and do_file_open_root(). These four 657 correspond roughly to the three ``path_*()`` functions we met earlier, 658 each of which calls ``link_path_walk()``. The ``path_*()`` functions are 659 called using different mode flags until a mode is found which works. 660 They are first called with ``LOOKUP_RCU`` set to request "RCU-walk". If 661 that fails with the error ``ECHILD`` they are called again with no 662 special flag to request "REF-walk". If either of those report the 663 error ``ESTALE`` a final attempt is made with ``LOOKUP_REVAL`` set (and no 664 ``LOOKUP_RCU``) to ensure that entries found in the cache are forcibly 665 revalidated - normally entries are only revalidated if the filesystem 666 determines that they are too old to trust. 667 668 The ``LOOKUP_RCU`` attempt may drop that flag internally and switch to 669 REF-walk, but will never then try to switch back to RCU-walk. Places 670 that trip up RCU-walk are much more likely to be near the leaves and 671 so it is very unlikely that there will be much, if any, benefit from 672 switching back. 673 674 RCU and seqlocks: fast and light 675 -------------------------------- 676 677 RCU is, unsurprisingly, critical to RCU-walk mode. The 678 ``rcu_read_lock()`` is held for the entire time that RCU-walk is walking 679 down a path. The particular guarantee it provides is that the key 680 data structures - dentries, inodes, super_blocks, and mounts - will 681 not be freed while the lock is held. They might be unlinked or 682 invalidated in one way or another, but the memory will not be 683 repurposed so values in various fields will still be meaningful. This 684 is the only guarantee that RCU provides; everything else is done using 685 seqlocks. 686 687 As we saw above, REF-walk holds a counted reference to the current 688 dentry and the current vfsmount, and does not release those references 689 before taking references to the "next" dentry or vfsmount. It also 690 sometimes takes the ``d_lock`` spinlock. These references and locks are 691 taken to prevent certain changes from happening. RCU-walk must not 692 take those references or locks and so cannot prevent such changes. 693 Instead, it checks to see if a change has been made, and aborts or 694 retries if it has. 695 696 To preserve the invariant mentioned above (that RCU-walk may only make 697 decisions that REF-walk could have made), it must make the checks at 698 or near the same places that REF-walk holds the references. So, when 699 REF-walk increments a reference count or takes a spinlock, RCU-walk 700 samples the status of a seqlock using ``read_seqcount_begin()`` or a 701 similar function. When REF-walk decrements the count or drops the 702 lock, RCU-walk checks if the sampled status is still valid using 703 ``read_seqcount_retry()`` or similar. 704 705 However, there is a little bit more to seqlocks than that. If 706 RCU-walk accesses two different fields in a seqlock-protected 707 structure, or accesses the same field twice, there is no a priori 708 guarantee of any consistency between those accesses. When consistency 709 is needed - which it usually is - RCU-walk must take a copy and then 710 use ``read_seqcount_retry()`` to validate that copy. 711 712 ``read_seqcount_retry()`` not only checks the sequence number, but also 713 imposes a memory barrier so that no memory-read instruction from 714 *before* the call can be delayed until *after* the call, either by the 715 CPU or by the compiler. A simple example of this can be seen in 716 ``slow_dentry_cmp()`` which, for filesystems which do not use simple 717 byte-wise name equality, calls into the filesystem to compare a name 718 against a dentry. The length and name pointer are copied into local 719 variables, then ``read_seqcount_retry()`` is called to confirm the two 720 are consistent, and only then is ``->d_compare()`` called. When 721 standard filename comparison is used, ``dentry_cmp()`` is called 722 instead. Notably it does *not* use ``read_seqcount_retry()``, but 723 instead has a large comment explaining why the consistency guarantee 724 isn't necessary. A subsequent ``read_seqcount_retry()`` will be 725 sufficient to catch any problem that could occur at this point. 726 727 With that little refresher on seqlocks out of the way we can look at 728 the bigger picture of how RCU-walk uses seqlocks. 729 730 ``mount_lock`` and ``nd->m_seq`` 731 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 732 733 We already met the ``mount_lock`` seqlock when REF-walk used it to 734 ensure that crossing a mount point is performed safely. RCU-walk uses 735 it for that too, but for quite a bit more. 736 737 Instead of taking a counted reference to each ``vfsmount`` as it 738 descends the tree, RCU-walk samples the state of ``mount_lock`` at the 739 start of the walk and stores this initial sequence number in the 740 ``struct nameidata`` in the ``m_seq`` field. This one lock and one 741 sequence number are used to validate all accesses to all ``vfsmounts``, 742 and all mount point crossings. As changes to the mount table are 743 relatively rare, it is reasonable to fall back on REF-walk any time 744 that any "mount" or "unmount" happens. 745 746 ``m_seq`` is checked (using ``read_seqretry()``) at the end of an RCU-walk 747 sequence, whether switching to REF-walk for the rest of the path or 748 when the end of the path is reached. It is also checked when stepping 749 down over a mount point (in ``__follow_mount_rcu()``) or up (in 750 ``follow_dotdot_rcu()``). If it is ever found to have changed, the 751 whole RCU-walk sequence is aborted and the path is processed again by 752 REF-walk. 753 754 If RCU-walk finds that ``mount_lock`` hasn't changed then it can be sure 755 that, had REF-walk taken counted references on each vfsmount, the 756 results would have been the same. This ensures the invariant holds, 757 at least for vfsmount structures. 758 759 ``dentry->d_seq`` and ``nd->seq`` 760 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 761 762 In place of taking a count or lock on ``d_reflock``, RCU-walk samples 763 the per-dentry ``d_seq`` seqlock, and stores the sequence number in the 764 ``seq`` field of the nameidata structure, so ``nd->seq`` should always be 765 the current sequence number of ``nd->dentry``. This number needs to be 766 revalidated after copying, and before using, the name, parent, or 767 inode of the dentry. 768 769 The handling of the name we have already looked at, and the parent is 770 only accessed in ``follow_dotdot_rcu()`` which fairly trivially follows 771 the required pattern, though it does so for three different cases. 772 773 When not at a mount point, ``d_parent`` is followed and its ``d_seq`` is 774 collected. When we are at a mount point, we instead follow the 775 ``mnt->mnt_mountpoint`` link to get a new dentry and collect its 776 ``d_seq``. Then, after finally finding a ``d_parent`` to follow, we must 777 check if we have landed on a mount point and, if so, must find that 778 mount point and follow the ``mnt->mnt_root`` link. This would imply a 779 somewhat unusual, but certainly possible, circumstance where the 780 starting point of the path lookup was in part of the filesystem that 781 was mounted on, and so not visible from the root. 782 783 The inode pointer, stored in ``->d_inode``, is a little more 784 interesting. The inode will always need to be accessed at least 785 twice, once to determine if it is NULL and once to verify access 786 permissions. Symlink handling requires a validated inode pointer too. 787 Rather than revalidating on each access, a copy is made on the first 788 access and it is stored in the ``inode`` field of ``nameidata`` from where 789 it can be safely accessed without further validation. 790 791 ``lookup_fast()`` is the only lookup routine that is used in RCU-mode, 792 ``lookup_slow()`` being too slow and requiring locks. It is in 793 ``lookup_fast()`` that we find the important "hand over hand" tracking 794 of the current dentry. 795 796 The current ``dentry`` and current ``seq`` number are passed to 797 ``__d_lookup_rcu()`` which, on success, returns a new ``dentry`` and a 798 new ``seq`` number. ``lookup_fast()`` then copies the inode pointer and 799 revalidates the new ``seq`` number. It then validates the old ``dentry`` 800 with the old ``seq`` number one last time and only then continues. This 801 process of getting the ``seq`` number of the new dentry and then 802 checking the ``seq`` number of the old exactly mirrors the process of 803 getting a counted reference to the new dentry before dropping that for 804 the old dentry which we saw in REF-walk. 805 806 No ``inode->i_rwsem`` or even ``rename_lock`` 807 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 808 809 A semaphore is a fairly heavyweight lock that can only be taken when it is 810 permissible to sleep. As ``rcu_read_lock()`` forbids sleeping, 811 ``inode->i_rwsem`` plays no role in RCU-walk. If some other thread does 812 take ``i_rwsem`` and modifies the directory in a way that RCU-walk needs 813 to notice, the result will be either that RCU-walk fails to find the 814 dentry that it is looking for, or it will find a dentry which 815 ``read_seqretry()`` won't validate. In either case it will drop down to 816 REF-walk mode which can take whatever locks are needed. 817 818 Though ``rename_lock`` could be used by RCU-walk as it doesn't require 819 any sleeping, RCU-walk doesn't bother. REF-walk uses ``rename_lock`` to 820 protect against the possibility of hash chains in the dcache changing 821 while they are being searched. This can result in failing to find 822 something that actually is there. When RCU-walk fails to find 823 something in the dentry cache, whether it is really there or not, it 824 already drops down to REF-walk and tries again with appropriate 825 locking. This neatly handles all cases, so adding extra checks on 826 rename_lock would bring no significant value. 827 828 ``unlazy walk()`` and ``complete_walk()`` 829 ----------------------------------------- 830 831 That "dropping down to REF-walk" typically involves a call to 832 ``unlazy_walk()``, so named because "RCU-walk" is also sometimes 833 referred to as "lazy walk". ``unlazy_walk()`` is called when 834 following the path down to the current vfsmount/dentry pair seems to 835 have proceeded successfully, but the next step is problematic. This 836 can happen if the next name cannot be found in the dcache, if 837 permission checking or name revalidation couldn't be achieved while 838 the ``rcu_read_lock()`` is held (which forbids sleeping), if an 839 automount point is found, or in a couple of cases involving symlinks. 840 It is also called from ``complete_walk()`` when the lookup has reached 841 the final component, or the very end of the path, depending on which 842 particular flavor of lookup is used. 843 844 Other reasons for dropping out of RCU-walk that do not trigger a call 845 to ``unlazy_walk()`` are when some inconsistency is found that cannot be 846 handled immediately, such as ``mount_lock`` or one of the ``d_seq`` 847 seqlocks reporting a change. In these cases the relevant function 848 will return ``-ECHILD`` which will percolate up until it triggers a new 849 attempt from the top using REF-walk. 850 851 For those cases where ``unlazy_walk()`` is an option, it essentially 852 takes a reference on each of the pointers that it holds (vfsmount, 853 dentry, and possibly some symbolic links) and then verifies that the 854 relevant seqlocks have not been changed. If there have been changes, 855 it, too, aborts with ``-ECHILD``, otherwise the transition to REF-walk 856 has been a success and the lookup process continues. 857 858 Taking a reference on those pointers is not quite as simple as just 859 incrementing a counter. That works to take a second reference if you 860 already have one (often indirectly through another object), but it 861 isn't sufficient if you don't actually have a counted reference at 862 all. For ``dentry->d_lockref``, it is safe to increment the reference 863 counter to get a reference unless it has been explicitly marked as 864 "dead" which involves setting the counter to ``-128``. 865 ``lockref_get_not_dead()`` achieves this. 866 867 For ``mnt->mnt_count`` it is safe to take a reference as long as 868 ``mount_lock`` is then used to validate the reference. If that 869 validation fails, it may *not* be safe to just drop that reference in 870 the standard way of calling ``mnt_put()`` - an unmount may have 871 progressed too far. So the code in ``legitimize_mnt()``, when it 872 finds that the reference it got might not be safe, checks the 873 ``MNT_SYNC_UMOUNT`` flag to determine if a simple ``mnt_put()`` is 874 correct, or if it should just decrement the count and pretend none of 875 this ever happened. 876 877 Taking care in filesystems 878 -------------------------- 879 880 RCU-walk depends almost entirely on cached information and often will 881 not call into the filesystem at all. However there are two places, 882 besides the already-mentioned component-name comparison, where the 883 file system might be included in RCU-walk, and it must know to be 884 careful. 885 886 If the filesystem has non-standard permission-checking requirements - 887 such as a networked filesystem which may need to check with the server 888 - the ``i_op->permission`` interface might be called during RCU-walk. 889 In this case an extra "``MAY_NOT_BLOCK``" flag is passed so that it 890 knows not to sleep, but to return ``-ECHILD`` if it cannot complete 891 promptly. ``i_op->permission`` is given the inode pointer, not the 892 dentry, so it doesn't need to worry about further consistency checks. 893 However if it accesses any other filesystem data structures, it must 894 ensure they are safe to be accessed with only the ``rcu_read_lock()`` 895 held. This typically means they must be freed using ``kfree_rcu()`` or 896 similar. 897 898 .. _READ_ONCE: https://lwn.net/Articles/624126/ 899 900 If the filesystem may need to revalidate dcache entries, then 901 ``d_op->d_revalidate`` may be called in RCU-walk too. This interface 902 *is* passed the dentry but does not have access to the ``inode`` or the 903 ``seq`` number from the ``nameidata``, so it needs to be extra careful 904 when accessing fields in the dentry. This "extra care" typically 905 involves using `READ_ONCE() <READ_ONCE_>`_ to access fields, and verifying the 906 result is not NULL before using it. This pattern can be seen in 907 ``nfs_lookup_revalidate()``. 908 909 A pair of patterns 910 ------------------ 911 912 In various places in the details of REF-walk and RCU-walk, and also in 913 the big picture, there are a couple of related patterns that are worth 914 being aware of. 915 916 The first is "try quickly and check, if that fails try slowly". We 917 can see that in the high-level approach of first trying RCU-walk and 918 then trying REF-walk, and in places where ``unlazy_walk()`` is used to 919 switch to REF-walk for the rest of the path. We also saw it earlier 920 in ``dget_parent()`` when following a "``..``" link. It tries a quick way 921 to get a reference, then falls back to taking locks if needed. 922 923 The second pattern is "try quickly and check, if that fails try 924 again - repeatedly". This is seen with the use of ``rename_lock`` and 925 ``mount_lock`` in REF-walk. RCU-walk doesn't make use of this pattern - 926 if anything goes wrong it is much safer to just abort and try a more 927 sedate approach. 928 929 The emphasis here is "try quickly and check". It should probably be 930 "try quickly *and carefully*, then check". The fact that checking is 931 needed is a reminder that the system is dynamic and only a limited 932 number of things are safe at all. The most likely cause of errors in 933 this whole process is assuming something is safe when in reality it 934 isn't. Careful consideration of what exactly guarantees the safety of 935 each access is sometimes necessary. 936 937 A walk among the symlinks 938 ========================= 939 940 There are several basic issues that we will examine to understand the 941 handling of symbolic links: the symlink stack, together with cache 942 lifetimes, will help us understand the overall recursive handling of 943 symlinks and lead to the special care needed for the final component. 944 Then a consideration of access-time updates and summary of the various 945 flags controlling lookup will finish the story. 946 947 The symlink stack 948 ----------------- 949 950 There are only two sorts of filesystem objects that can usefully 951 appear in a path prior to the final component: directories and symlinks. 952 Handling directories is quite straightforward: the new directory 953 simply becomes the starting point at which to interpret the next 954 component on the path. Handling symbolic links requires a bit more 955 work. 956 957 Conceptually, symbolic links could be handled by editing the path. If 958 a component name refers to a symbolic link, then that component is 959 replaced by the body of the link and, if that body starts with a '/', 960 then all preceding parts of the path are discarded. This is what the 961 "``readlink -f``" command does, though it also edits out "``.``" and 962 "``..``" components. 963 964 Directly editing the path string is not really necessary when looking 965 up a path, and discarding early components is pointless as they aren't 966 looked at anyway. Keeping track of all remaining components is 967 important, but they can of course be kept separately; there is no need 968 to concatenate them. As one symlink may easily refer to another, 969 which in turn can refer to a third, we may need to keep the remaining 970 components of several paths, each to be processed when the preceding 971 ones are completed. These path remnants are kept on a stack of 972 limited size. 973 974 There are two reasons for placing limits on how many symlinks can 975 occur in a single path lookup. The most obvious is to avoid loops. 976 If a symlink referred to itself either directly or through 977 intermediaries, then following the symlink can never complete 978 successfully - the error ``ELOOP`` must be returned. Loops can be 979 detected without imposing limits, but limits are the simplest solution 980 and, given the second reason for restriction, quite sufficient. 981 982 .. _outlined recently: http://thread.gmane.org/gmane.linux.kernel/1934390/focus=1934550 983 984 The second reason was `outlined recently`_ by Linus: 985 986 Because it's a latency and DoS issue too. We need to react well to 987 true loops, but also to "very deep" non-loops. It's not about memory 988 use, it's about users triggering unreasonable CPU resources. 989 990 Linux imposes a limit on the length of any pathname: ``PATH_MAX``, which 991 is 4096. There are a number of reasons for this limit; not letting the 992 kernel spend too much time on just one path is one of them. With 993 symbolic links you can effectively generate much longer paths so some 994 sort of limit is needed for the same reason. Linux imposes a limit of 995 at most 40 (MAXSYMLINKS) symlinks in any one path lookup. It previously imposed 996 a further limit of eight on the maximum depth of recursion, but that was 997 raised to 40 when a separate stack was implemented, so there is now 998 just the one limit. 999 1000 The ``nameidata`` structure that we met in an earlier article contains a 1001 small stack that can be used to store the remaining part of up to two 1002 symlinks. In many cases this will be sufficient. If it isn't, a 1003 separate stack is allocated with room for 40 symlinks. Pathname 1004 lookup will never exceed that stack as, once the 40th symlink is 1005 detected, an error is returned. 1006 1007 It might seem that the name remnants are all that needs to be stored on 1008 this stack, but we need a bit more. To see that, we need to move on to 1009 cache lifetimes. 1010 1011 Storage and lifetime of cached symlinks 1012 --------------------------------------- 1013 1014 Like other filesystem resources, such as inodes and directory 1015 entries, symlinks are cached by Linux to avoid repeated costly access 1016 to external storage. It is particularly important for RCU-walk to be 1017 able to find and temporarily hold onto these cached entries, so that 1018 it doesn't need to drop down into REF-walk. 1019 1020 .. _object-oriented design pattern: https://lwn.net/Articles/446317/ 1021 1022 While each filesystem is free to make its own choice, symlinks are 1023 typically stored in one of two places. Short symlinks are often 1024 stored directly in the inode. When a filesystem allocates a ``struct 1025 inode`` it typically allocates extra space to store private data (a 1026 common `object-oriented design pattern`_ in the kernel). This will 1027 sometimes include space for a symlink. The other common location is 1028 in the page cache, which normally stores the content of files. The 1029 pathname in a symlink can be seen as the content of that symlink and 1030 can easily be stored in the page cache just like file content. 1031 1032 When neither of these is suitable, the next most likely scenario is 1033 that the filesystem will allocate some temporary memory and copy or 1034 construct the symlink content into that memory whenever it is needed. 1035 1036 When the symlink is stored in the inode, it has the same lifetime as 1037 the inode which, itself, is protected by RCU or by a counted reference 1038 on the dentry. This means that the mechanisms that pathname lookup 1039 uses to access the dcache and icache (inode cache) safely are quite 1040 sufficient for accessing some cached symlinks safely. In these cases, 1041 the ``i_link`` pointer in the inode is set to point to wherever the 1042 symlink is stored and it can be accessed directly whenever needed. 1043 1044 When the symlink is stored in the page cache or elsewhere, the 1045 situation is not so straightforward. A reference on a dentry or even 1046 on an inode does not imply any reference on cached pages of that 1047 inode, and even an ``rcu_read_lock()`` is not sufficient to ensure that 1048 a page will not disappear. So for these symlinks the pathname lookup 1049 code needs to ask the filesystem to provide a stable reference and, 1050 significantly, needs to release that reference when it is finished 1051 with it. 1052 1053 Taking a reference to a cache page is often possible even in RCU-walk 1054 mode. It does require making changes to memory, which is best avoided, 1055 but that isn't necessarily a big cost and it is better than dropping 1056 out of RCU-walk mode completely. Even filesystems that allocate 1057 space to copy the symlink into can use ``GFP_ATOMIC`` to often successfully 1058 allocate memory without the need to drop out of RCU-walk. If a 1059 filesystem cannot successfully get a reference in RCU-walk mode, it 1060 must return ``-ECHILD`` and ``unlazy_walk()`` will be called to return to 1061 REF-walk mode in which the filesystem is allowed to sleep. 1062 1063 The place for all this to happen is the ``i_op->get_link()`` inode 1064 method. This is called both in RCU-walk and REF-walk. In RCU-walk the 1065 ``dentry*`` argument is NULL, ``->get_link()`` can return -ECHILD to drop out of 1066 RCU-walk. Much like the ``i_op->permission()`` method we 1067 looked at previously, ``->get_link()`` would need to be careful that 1068 all the data structures it references are safe to be accessed while 1069 holding no counted reference, only the RCU lock. A callback 1070 ``struct delayed_called`` will be passed to ``->get_link()``: 1071 file systems can set their own put_link function and argument through 1072 set_delayed_call(). Later on, when VFS wants to put link, it will call 1073 do_delayed_call() to invoke that callback function with the argument. 1074 1075 In order for the reference to each symlink to be dropped when the walk completes, 1076 whether in RCU-walk or REF-walk, the symlink stack needs to contain, 1077 along with the path remnants: 1078 1079 - the ``struct path`` to provide a reference to the previous path 1080 - the ``const char *`` to provide a reference to the to previous name 1081 - the ``seq`` to allow the path to be safely switched from RCU-walk to REF-walk 1082 - the ``struct delayed_call`` for later invocation. 1083 1084 This means that each entry in the symlink stack needs to hold five 1085 pointers and an integer instead of just one pointer (the path 1086 remnant). On a 64-bit system, this is about 40 bytes per entry; 1087 with 40 entries it adds up to 1600 bytes total, which is less than 1088 half a page. So it might seem like a lot, but is by no means 1089 excessive. 1090 1091 Note that, in a given stack frame, the path remnant (``name``) is not 1092 part of the symlink that the other fields refer to. It is the remnant 1093 to be followed once that symlink has been fully parsed. 1094 1095 Following the symlink 1096 --------------------- 1097 1098 The main loop in ``link_path_walk()`` iterates seamlessly over all 1099 components in the path and all of the non-final symlinks. As symlinks 1100 are processed, the ``name`` pointer is adjusted to point to a new 1101 symlink, or is restored from the stack, so that much of the loop 1102 doesn't need to notice. Getting this ``name`` variable on and off the 1103 stack is very straightforward; pushing and popping the references is 1104 a little more complex. 1105 1106 When a symlink is found, walk_component() calls pick_link() via step_into() 1107 which returns the link from the filesystem. 1108 Providing that operation is successful, the old path ``name`` is placed on the 1109 stack, and the new value is used as the ``name`` for a while. When the end of 1110 the path is found (i.e. ``*name`` is ``'\0'``) the old ``name`` is restored 1111 off the stack and path walking continues. 1112 1113 Pushing and popping the reference pointers (inode, cookie, etc.) is more 1114 complex in part because of the desire to handle tail recursion. When 1115 the last component of a symlink itself points to a symlink, we 1116 want to pop the symlink-just-completed off the stack before pushing 1117 the symlink-just-found to avoid leaving empty path remnants that would 1118 just get in the way. 1119 1120 It is most convenient to push the new symlink references onto the 1121 stack in ``walk_component()`` immediately when the symlink is found; 1122 ``walk_component()`` is also the last piece of code that needs to look at the 1123 old symlink as it walks that last component. So it is quite 1124 convenient for ``walk_component()`` to release the old symlink and pop 1125 the references just before pushing the reference information for the 1126 new symlink. It is guided in this by three flags: ``WALK_NOFOLLOW`` which 1127 forbids it from following a symlink if it finds one, ``WALK_MORE`` 1128 which indicates that it is yet too early to release the 1129 current symlink, and ``WALK_TRAILING`` which indicates that it is on the final 1130 component of the lookup, so we will check userspace flag ``LOOKUP_FOLLOW`` to 1131 decide whether follow it when it is a symlink and call ``may_follow_link()`` to 1132 check if we have privilege to follow it. 1133 1134 Symlinks with no final component 1135 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 1136 1137 A pair of special-case symlinks deserve a little further explanation. 1138 Both result in a new ``struct path`` (with mount and dentry) being set 1139 up in the ``nameidata``, and result in pick_link() returning ``NULL``. 1140 1141 The more obvious case is a symlink to "``/``". All symlinks starting 1142 with "``/``" are detected in pick_link() which resets the ``nameidata`` 1143 to point to the effective filesystem root. If the symlink only 1144 contains "``/``" then there is nothing more to do, no components at all, 1145 so ``NULL`` is returned to indicate that the symlink can be released and 1146 the stack frame discarded. 1147 1148 The other case involves things in ``/proc`` that look like symlinks but 1149 aren't really (and are therefore commonly referred to as "magic-links"):: 1150 1151 $ ls -l /proc/self/fd/1 1152 lrwx------ 1 neilb neilb 64 Jun 13 10:19 /proc/self/fd/1 -> /dev/pts/4 1153 1154 Every open file descriptor in any process is represented in ``/proc`` by 1155 something that looks like a symlink. It is really a reference to the 1156 target file, not just the name of it. When you ``readlink`` these 1157 objects you get a name that might refer to the same file - unless it 1158 has been unlinked or mounted over. When ``walk_component()`` follows 1159 one of these, the ``->get_link()`` method in "procfs" doesn't return 1160 a string name, but instead calls nd_jump_link() which updates the 1161 ``nameidata`` in place to point to that target. ``->get_link()`` then 1162 returns ``NULL``. Again there is no final component and pick_link() 1163 returns ``NULL``. 1164 1165 Following the symlink in the final component 1166 -------------------------------------------- 1167 1168 All this leads to ``link_path_walk()`` walking down every component, and 1169 following all symbolic links it finds, until it reaches the final 1170 component. This is just returned in the ``last`` field of ``nameidata``. 1171 For some callers, this is all they need; they want to create that 1172 ``last`` name if it doesn't exist or give an error if it does. Other 1173 callers will want to follow a symlink if one is found, and possibly 1174 apply special handling to the last component of that symlink, rather 1175 than just the last component of the original file name. These callers 1176 potentially need to call ``link_path_walk()`` again and again on 1177 successive symlinks until one is found that doesn't point to another 1178 symlink. 1179 1180 This case is handled by relevant callers of link_path_walk(), such as 1181 path_lookupat(), path_openat() using a loop that calls link_path_walk(), 1182 and then handles the final component by calling open_last_lookups() or 1183 lookup_last(). If it is a symlink that needs to be followed, 1184 open_last_lookups() or lookup_last() will set things up properly and 1185 return the path so that the loop repeats, calling 1186 link_path_walk() again. This could loop as many as 40 times if the last 1187 component of each symlink is another symlink. 1188 1189 Of the various functions that examine the final component, 1190 open_last_lookups() is the most interesting as it works in tandem 1191 with do_open() for opening a file. Part of open_last_lookups() runs 1192 with ``i_rwsem`` held and this part is in a separate function: lookup_open(). 1193 1194 Explaining open_last_lookups() and do_open() completely is beyond the scope 1195 of this article, but a few highlights should help those interested in exploring 1196 the code. 1197 1198 1. Rather than just finding the target file, do_open() is used after 1199 open_last_lookup() to open 1200 it. If the file was found in the dcache, then ``vfs_open()`` is used for 1201 this. If not, then ``lookup_open()`` will either call ``atomic_open()`` (if 1202 the filesystem provides it) to combine the final lookup with the open, or 1203 will perform the separate ``i_op->lookup()`` and ``i_op->create()`` steps 1204 directly. In the later case the actual "open" of this newly found or 1205 created file will be performed by vfs_open(), just as if the name 1206 were found in the dcache. 1207 1208 2. vfs_open() can fail with ``-EOPENSTALE`` if the cached information 1209 wasn't quite current enough. If it's in RCU-walk ``-ECHILD`` will be returned 1210 otherwise ``-ESTALE`` is returned. When ``-ESTALE`` is returned, the caller may 1211 retry with ``LOOKUP_REVAL`` flag set. 1212 1213 3. An open with O_CREAT **does** follow a symlink in the final component, 1214 unlike other creation system calls (like ``mkdir``). So the sequence:: 1215 1216 ln -s bar /tmp/foo 1217 echo hello > /tmp/foo 1218 1219 will create a file called ``/tmp/bar``. This is not permitted if 1220 ``O_EXCL`` is set but otherwise is handled for an O_CREAT open much 1221 like for a non-creating open: lookup_last() or open_last_lookup() 1222 returns a non ``NULL`` value, and link_path_walk() gets called and the 1223 open process continues on the symlink that was found. 1224 1225 Updating the access time 1226 ------------------------ 1227 1228 We previously said of RCU-walk that it would "take no locks, increment 1229 no counts, leave no footprints." We have since seen that some 1230 "footprints" can be needed when handling symlinks as a counted 1231 reference (or even a memory allocation) may be needed. But these 1232 footprints are best kept to a minimum. 1233 1234 One other place where walking down a symlink can involve leaving 1235 footprints in a way that doesn't affect directories is in updating access times. 1236 In Unix (and Linux) every filesystem object has a "last accessed 1237 time", or "``atime``". Passing through a directory to access a file 1238 within is not considered to be an access for the purposes of 1239 ``atime``; only listing the contents of a directory can update its ``atime``. 1240 Symlinks are different it seems. Both reading a symlink (with ``readlink()``) 1241 and looking up a symlink on the way to some other destination can 1242 update the atime on that symlink. 1243 1244 .. _clearest statement: https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap04.html#tag_04_08 1245 1246 It is not clear why this is the case; POSIX has little to say on the 1247 subject. The `clearest statement`_ is that, if a particular implementation 1248 updates a timestamp in a place not specified by POSIX, this must be 1249 documented "except that any changes caused by pathname resolution need 1250 not be documented". This seems to imply that POSIX doesn't really 1251 care about access-time updates during pathname lookup. 1252 1253 .. _Linux 1.3.87: https://git.kernel.org/cgit/linux/kernel/git/history/history.git/diff/fs/ext2/symlink.c?id=f806c6db77b8eaa6e00dcfb6b567706feae8dbb8 1254 1255 An examination of history shows that prior to `Linux 1.3.87`_, the ext2 1256 filesystem, at least, didn't update atime when following a link. 1257 Unfortunately we have no record of why that behavior was changed. 1258 1259 In any case, access time must now be updated and that operation can be 1260 quite complex. Trying to stay in RCU-walk while doing it is best 1261 avoided. Fortunately it is often permitted to skip the ``atime`` 1262 update. Because ``atime`` updates cause performance problems in various 1263 areas, Linux supports the ``relatime`` mount option, which generally 1264 limits the updates of ``atime`` to once per day on files that aren't 1265 being changed (and symlinks never change once created). Even without 1266 ``relatime``, many filesystems record ``atime`` with a one-second 1267 granularity, so only one update per second is required. 1268 1269 It is easy to test if an ``atime`` update is needed while in RCU-walk 1270 mode and, if it isn't, the update can be skipped and RCU-walk mode 1271 continues. Only when an ``atime`` update is actually required does the 1272 path walk drop down to REF-walk. All of this is handled in the 1273 ``get_link()`` function. 1274 1275 A few flags 1276 ----------- 1277 1278 A suitable way to wrap up this tour of pathname walking is to list 1279 the various flags that can be stored in the ``nameidata`` to guide the 1280 lookup process. Many of these are only meaningful on the final 1281 component, others reflect the current state of the pathname lookup, and some 1282 apply restrictions to all path components encountered in the path lookup. 1283 1284 And then there is ``LOOKUP_EMPTY``, which doesn't fit conceptually with 1285 the others. If this is not set, an empty pathname causes an error 1286 very early on. If it is set, empty pathnames are not considered to be 1287 an error. 1288 1289 Global state flags 1290 ~~~~~~~~~~~~~~~~~~ 1291 1292 We have already met two global state flags: ``LOOKUP_RCU`` and 1293 ``LOOKUP_REVAL``. These select between one of three overall approaches 1294 to lookup: RCU-walk, REF-walk, and REF-walk with forced revalidation. 1295 1296 ``LOOKUP_PARENT`` indicates that the final component hasn't been reached 1297 yet. This is primarily used to tell the audit subsystem the full 1298 context of a particular access being audited. 1299 1300 ``ND_ROOT_PRESET`` indicates that the ``root`` field in the ``nameidata`` was 1301 provided by the caller, so it shouldn't be released when it is no 1302 longer needed. 1303 1304 ``ND_JUMPED`` means that the current dentry was chosen not because 1305 it had the right name but for some other reason. This happens when 1306 following "``..``", following a symlink to ``/``, crossing a mount point 1307 or accessing a "``/proc/$PID/fd/$FD``" symlink (also known as a "magic 1308 link"). In this case the filesystem has not been asked to revalidate the 1309 name (with ``d_revalidate()``). In such cases the inode may still need 1310 to be revalidated, so ``d_op->d_weak_revalidate()`` is called if 1311 ``ND_JUMPED`` is set when the look completes - which may be at the 1312 final component or, when creating, unlinking, or renaming, at the penultimate component. 1313 1314 Resolution-restriction flags 1315 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 1316 1317 In order to allow userspace to protect itself against certain race conditions 1318 and attack scenarios involving changing path components, a series of flags are 1319 available which apply restrictions to all path components encountered during 1320 path lookup. These flags are exposed through ``openat2()``'s ``resolve`` field. 1321 1322 ``LOOKUP_NO_SYMLINKS`` blocks all symlink traversals (including magic-links). 1323 This is distinctly different from ``LOOKUP_FOLLOW``, because the latter only 1324 relates to restricting the following of trailing symlinks. 1325 1326 ``LOOKUP_NO_MAGICLINKS`` blocks all magic-link traversals. Filesystems must 1327 ensure that they return errors from ``nd_jump_link()``, because that is how 1328 ``LOOKUP_NO_MAGICLINKS`` and other magic-link restrictions are implemented. 1329 1330 ``LOOKUP_NO_XDEV`` blocks all ``vfsmount`` traversals (this includes both 1331 bind-mounts and ordinary mounts). Note that the ``vfsmount`` which contains the 1332 lookup is determined by the first mountpoint the path lookup reaches -- 1333 absolute paths start with the ``vfsmount`` of ``/``, and relative paths start 1334 with the ``dfd``'s ``vfsmount``. Magic-links are only permitted if the 1335 ``vfsmount`` of the path is unchanged. 1336 1337 ``LOOKUP_BENEATH`` blocks any path components which resolve outside the 1338 starting point of the resolution. This is done by blocking ``nd_jump_root()`` 1339 as well as blocking ".." if it would jump outside the starting point. 1340 ``rename_lock`` and ``mount_lock`` are used to detect attacks against the 1341 resolution of "..". Magic-links are also blocked. 1342 1343 ``LOOKUP_IN_ROOT`` resolves all path components as though the starting point 1344 were the filesystem root. ``nd_jump_root()`` brings the resolution back to 1345 the starting point, and ".." at the starting point will act as a no-op. As with 1346 ``LOOKUP_BENEATH``, ``rename_lock`` and ``mount_lock`` are used to detect 1347 attacks against ".." resolution. Magic-links are also blocked. 1348 1349 Final-component flags 1350 ~~~~~~~~~~~~~~~~~~~~~ 1351 1352 Some of these flags are only set when the final component is being 1353 considered. Others are only checked for when considering that final 1354 component. 1355 1356 ``LOOKUP_AUTOMOUNT`` ensures that, if the final component is an automount 1357 point, then the mount is triggered. Some operations would trigger it 1358 anyway, but operations like ``stat()`` deliberately don't. ``statfs()`` 1359 needs to trigger the mount but otherwise behaves a lot like ``stat()``, so 1360 it sets ``LOOKUP_AUTOMOUNT``, as does "``quotactl()``" and the handling of 1361 "``mount --bind``". 1362 1363 ``LOOKUP_FOLLOW`` has a similar function to ``LOOKUP_AUTOMOUNT`` but for 1364 symlinks. Some system calls set or clear it implicitly, while 1365 others have API flags such as ``AT_SYMLINK_FOLLOW`` and 1366 ``UMOUNT_NOFOLLOW`` to control it. Its effect is similar to 1367 ``WALK_GET`` that we already met, but it is used in a different way. 1368 1369 ``LOOKUP_DIRECTORY`` insists that the final component is a directory. 1370 Various callers set this and it is also set when the final component 1371 is found to be followed by a slash. 1372 1373 Finally ``LOOKUP_OPEN``, ``LOOKUP_CREATE``, ``LOOKUP_EXCL``, and 1374 ``LOOKUP_RENAME_TARGET`` are not used directly by the VFS but are made 1375 available to the filesystem and particularly the ``->d_revalidate()`` 1376 method. A filesystem can choose not to bother revalidating too hard 1377 if it knows that it will be asked to open or create the file soon. 1378 These flags were previously useful for ``->lookup()`` too but with the 1379 introduction of ``->atomic_open()`` they are less relevant there. 1380 1381 End of the road 1382 --------------- 1383 1384 Despite its complexity, all this pathname lookup code appears to be 1385 in good shape - various parts are certainly easier to understand now 1386 than even a couple of releases ago. But that doesn't mean it is 1387 "finished". As already mentioned, RCU-walk currently only follows 1388 symlinks that are stored in the inode so, while it handles many ext4 1389 symlinks, it doesn't help with NFS, XFS, or Btrfs. That support 1390 is not likely to be long delayed.
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.