]> git.karo-electronics.de Git - karo-tx-linux.git/blob - fs/namei.c
Merge tag 'cifs-fixes-for-4.13-rc7-and-stable' of git://git.samba.org/sfrench/cifs-2.6
[karo-tx-linux.git] / fs / namei.c
1 /*
2  *  linux/fs/namei.c
3  *
4  *  Copyright (C) 1991, 1992  Linus Torvalds
5  */
6
7 /*
8  * Some corrections by tytso.
9  */
10
11 /* [Feb 1997 T. Schoebel-Theuer] Complete rewrite of the pathname
12  * lookup logic.
13  */
14 /* [Feb-Apr 2000, AV] Rewrite to the new namespace architecture.
15  */
16
17 #include <linux/init.h>
18 #include <linux/export.h>
19 #include <linux/kernel.h>
20 #include <linux/slab.h>
21 #include <linux/fs.h>
22 #include <linux/namei.h>
23 #include <linux/pagemap.h>
24 #include <linux/fsnotify.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/ima.h>
28 #include <linux/syscalls.h>
29 #include <linux/mount.h>
30 #include <linux/audit.h>
31 #include <linux/capability.h>
32 #include <linux/file.h>
33 #include <linux/fcntl.h>
34 #include <linux/device_cgroup.h>
35 #include <linux/fs_struct.h>
36 #include <linux/posix_acl.h>
37 #include <linux/hash.h>
38 #include <linux/bitops.h>
39 #include <linux/init_task.h>
40 #include <linux/uaccess.h>
41
42 #include "internal.h"
43 #include "mount.h"
44
45 /* [Feb-1997 T. Schoebel-Theuer]
46  * Fundamental changes in the pathname lookup mechanisms (namei)
47  * were necessary because of omirr.  The reason is that omirr needs
48  * to know the _real_ pathname, not the user-supplied one, in case
49  * of symlinks (and also when transname replacements occur).
50  *
51  * The new code replaces the old recursive symlink resolution with
52  * an iterative one (in case of non-nested symlink chains).  It does
53  * this with calls to <fs>_follow_link().
54  * As a side effect, dir_namei(), _namei() and follow_link() are now 
55  * replaced with a single function lookup_dentry() that can handle all 
56  * the special cases of the former code.
57  *
58  * With the new dcache, the pathname is stored at each inode, at least as
59  * long as the refcount of the inode is positive.  As a side effect, the
60  * size of the dcache depends on the inode cache and thus is dynamic.
61  *
62  * [29-Apr-1998 C. Scott Ananian] Updated above description of symlink
63  * resolution to correspond with current state of the code.
64  *
65  * Note that the symlink resolution is not *completely* iterative.
66  * There is still a significant amount of tail- and mid- recursion in
67  * the algorithm.  Also, note that <fs>_readlink() is not used in
68  * lookup_dentry(): lookup_dentry() on the result of <fs>_readlink()
69  * may return different results than <fs>_follow_link().  Many virtual
70  * filesystems (including /proc) exhibit this behavior.
71  */
72
73 /* [24-Feb-97 T. Schoebel-Theuer] Side effects caused by new implementation:
74  * New symlink semantics: when open() is called with flags O_CREAT | O_EXCL
75  * and the name already exists in form of a symlink, try to create the new
76  * name indicated by the symlink. The old code always complained that the
77  * name already exists, due to not following the symlink even if its target
78  * is nonexistent.  The new semantics affects also mknod() and link() when
79  * the name is a symlink pointing to a non-existent name.
80  *
81  * I don't know which semantics is the right one, since I have no access
82  * to standards. But I found by trial that HP-UX 9.0 has the full "new"
83  * semantics implemented, while SunOS 4.1.1 and Solaris (SunOS 5.4) have the
84  * "old" one. Personally, I think the new semantics is much more logical.
85  * Note that "ln old new" where "new" is a symlink pointing to a non-existing
86  * file does succeed in both HP-UX and SunOs, but not in Solaris
87  * and in the old Linux semantics.
88  */
89
90 /* [16-Dec-97 Kevin Buhr] For security reasons, we change some symlink
91  * semantics.  See the comments in "open_namei" and "do_link" below.
92  *
93  * [10-Sep-98 Alan Modra] Another symlink change.
94  */
95
96 /* [Feb-Apr 2000 AV] Complete rewrite. Rules for symlinks:
97  *      inside the path - always follow.
98  *      in the last component in creation/removal/renaming - never follow.
99  *      if LOOKUP_FOLLOW passed - follow.
100  *      if the pathname has trailing slashes - follow.
101  *      otherwise - don't follow.
102  * (applied in that order).
103  *
104  * [Jun 2000 AV] Inconsistent behaviour of open() in case if flags==O_CREAT
105  * restored for 2.4. This is the last surviving part of old 4.2BSD bug.
106  * During the 2.4 we need to fix the userland stuff depending on it -
107  * hopefully we will be able to get rid of that wart in 2.5. So far only
108  * XEmacs seems to be relying on it...
109  */
110 /*
111  * [Sep 2001 AV] Single-semaphore locking scheme (kudos to David Holland)
112  * implemented.  Let's see if raised priority of ->s_vfs_rename_mutex gives
113  * any extra contention...
114  */
115
116 /* In order to reduce some races, while at the same time doing additional
117  * checking and hopefully speeding things up, we copy filenames to the
118  * kernel data space before using them..
119  *
120  * POSIX.1 2.4: an empty pathname is invalid (ENOENT).
121  * PATH_MAX includes the nul terminator --RR.
122  */
123
124 #define EMBEDDED_NAME_MAX       (PATH_MAX - offsetof(struct filename, iname))
125
126 struct filename *
127 getname_flags(const char __user *filename, int flags, int *empty)
128 {
129         struct filename *result;
130         char *kname;
131         int len;
132
133         result = audit_reusename(filename);
134         if (result)
135                 return result;
136
137         result = __getname();
138         if (unlikely(!result))
139                 return ERR_PTR(-ENOMEM);
140
141         /*
142          * First, try to embed the struct filename inside the names_cache
143          * allocation
144          */
145         kname = (char *)result->iname;
146         result->name = kname;
147
148         len = strncpy_from_user(kname, filename, EMBEDDED_NAME_MAX);
149         if (unlikely(len < 0)) {
150                 __putname(result);
151                 return ERR_PTR(len);
152         }
153
154         /*
155          * Uh-oh. We have a name that's approaching PATH_MAX. Allocate a
156          * separate struct filename so we can dedicate the entire
157          * names_cache allocation for the pathname, and re-do the copy from
158          * userland.
159          */
160         if (unlikely(len == EMBEDDED_NAME_MAX)) {
161                 const size_t size = offsetof(struct filename, iname[1]);
162                 kname = (char *)result;
163
164                 /*
165                  * size is chosen that way we to guarantee that
166                  * result->iname[0] is within the same object and that
167                  * kname can't be equal to result->iname, no matter what.
168                  */
169                 result = kzalloc(size, GFP_KERNEL);
170                 if (unlikely(!result)) {
171                         __putname(kname);
172                         return ERR_PTR(-ENOMEM);
173                 }
174                 result->name = kname;
175                 len = strncpy_from_user(kname, filename, PATH_MAX);
176                 if (unlikely(len < 0)) {
177                         __putname(kname);
178                         kfree(result);
179                         return ERR_PTR(len);
180                 }
181                 if (unlikely(len == PATH_MAX)) {
182                         __putname(kname);
183                         kfree(result);
184                         return ERR_PTR(-ENAMETOOLONG);
185                 }
186         }
187
188         result->refcnt = 1;
189         /* The empty path is special. */
190         if (unlikely(!len)) {
191                 if (empty)
192                         *empty = 1;
193                 if (!(flags & LOOKUP_EMPTY)) {
194                         putname(result);
195                         return ERR_PTR(-ENOENT);
196                 }
197         }
198
199         result->uptr = filename;
200         result->aname = NULL;
201         audit_getname(result);
202         return result;
203 }
204
205 struct filename *
206 getname(const char __user * filename)
207 {
208         return getname_flags(filename, 0, NULL);
209 }
210
211 struct filename *
212 getname_kernel(const char * filename)
213 {
214         struct filename *result;
215         int len = strlen(filename) + 1;
216
217         result = __getname();
218         if (unlikely(!result))
219                 return ERR_PTR(-ENOMEM);
220
221         if (len <= EMBEDDED_NAME_MAX) {
222                 result->name = (char *)result->iname;
223         } else if (len <= PATH_MAX) {
224                 struct filename *tmp;
225
226                 tmp = kmalloc(sizeof(*tmp), GFP_KERNEL);
227                 if (unlikely(!tmp)) {
228                         __putname(result);
229                         return ERR_PTR(-ENOMEM);
230                 }
231                 tmp->name = (char *)result;
232                 result = tmp;
233         } else {
234                 __putname(result);
235                 return ERR_PTR(-ENAMETOOLONG);
236         }
237         memcpy((char *)result->name, filename, len);
238         result->uptr = NULL;
239         result->aname = NULL;
240         result->refcnt = 1;
241         audit_getname(result);
242
243         return result;
244 }
245
246 void putname(struct filename *name)
247 {
248         BUG_ON(name->refcnt <= 0);
249
250         if (--name->refcnt > 0)
251                 return;
252
253         if (name->name != name->iname) {
254                 __putname(name->name);
255                 kfree(name);
256         } else
257                 __putname(name);
258 }
259
260 static int check_acl(struct inode *inode, int mask)
261 {
262 #ifdef CONFIG_FS_POSIX_ACL
263         struct posix_acl *acl;
264
265         if (mask & MAY_NOT_BLOCK) {
266                 acl = get_cached_acl_rcu(inode, ACL_TYPE_ACCESS);
267                 if (!acl)
268                         return -EAGAIN;
269                 /* no ->get_acl() calls in RCU mode... */
270                 if (is_uncached_acl(acl))
271                         return -ECHILD;
272                 return posix_acl_permission(inode, acl, mask & ~MAY_NOT_BLOCK);
273         }
274
275         acl = get_acl(inode, ACL_TYPE_ACCESS);
276         if (IS_ERR(acl))
277                 return PTR_ERR(acl);
278         if (acl) {
279                 int error = posix_acl_permission(inode, acl, mask);
280                 posix_acl_release(acl);
281                 return error;
282         }
283 #endif
284
285         return -EAGAIN;
286 }
287
288 /*
289  * This does the basic permission checking
290  */
291 static int acl_permission_check(struct inode *inode, int mask)
292 {
293         unsigned int mode = inode->i_mode;
294
295         if (likely(uid_eq(current_fsuid(), inode->i_uid)))
296                 mode >>= 6;
297         else {
298                 if (IS_POSIXACL(inode) && (mode & S_IRWXG)) {
299                         int error = check_acl(inode, mask);
300                         if (error != -EAGAIN)
301                                 return error;
302                 }
303
304                 if (in_group_p(inode->i_gid))
305                         mode >>= 3;
306         }
307
308         /*
309          * If the DACs are ok we don't need any capability check.
310          */
311         if ((mask & ~mode & (MAY_READ | MAY_WRITE | MAY_EXEC)) == 0)
312                 return 0;
313         return -EACCES;
314 }
315
316 /**
317  * generic_permission -  check for access rights on a Posix-like filesystem
318  * @inode:      inode to check access rights for
319  * @mask:       right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC, ...)
320  *
321  * Used to check for read/write/execute permissions on a file.
322  * We use "fsuid" for this, letting us set arbitrary permissions
323  * for filesystem access without changing the "normal" uids which
324  * are used for other things.
325  *
326  * generic_permission is rcu-walk aware. It returns -ECHILD in case an rcu-walk
327  * request cannot be satisfied (eg. requires blocking or too much complexity).
328  * It would then be called again in ref-walk mode.
329  */
330 int generic_permission(struct inode *inode, int mask)
331 {
332         int ret;
333
334         /*
335          * Do the basic permission checks.
336          */
337         ret = acl_permission_check(inode, mask);
338         if (ret != -EACCES)
339                 return ret;
340
341         if (S_ISDIR(inode->i_mode)) {
342                 /* DACs are overridable for directories */
343                 if (!(mask & MAY_WRITE))
344                         if (capable_wrt_inode_uidgid(inode,
345                                                      CAP_DAC_READ_SEARCH))
346                                 return 0;
347                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
348                         return 0;
349                 return -EACCES;
350         }
351
352         /*
353          * Searching includes executable on directories, else just read.
354          */
355         mask &= MAY_READ | MAY_WRITE | MAY_EXEC;
356         if (mask == MAY_READ)
357                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_READ_SEARCH))
358                         return 0;
359         /*
360          * Read/write DACs are always overridable.
361          * Executable DACs are overridable when there is
362          * at least one exec bit set.
363          */
364         if (!(mask & MAY_EXEC) || (inode->i_mode & S_IXUGO))
365                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
366                         return 0;
367
368         return -EACCES;
369 }
370 EXPORT_SYMBOL(generic_permission);
371
372 /*
373  * We _really_ want to just do "generic_permission()" without
374  * even looking at the inode->i_op values. So we keep a cache
375  * flag in inode->i_opflags, that says "this has not special
376  * permission function, use the fast case".
377  */
378 static inline int do_inode_permission(struct inode *inode, int mask)
379 {
380         if (unlikely(!(inode->i_opflags & IOP_FASTPERM))) {
381                 if (likely(inode->i_op->permission))
382                         return inode->i_op->permission(inode, mask);
383
384                 /* This gets set once for the inode lifetime */
385                 spin_lock(&inode->i_lock);
386                 inode->i_opflags |= IOP_FASTPERM;
387                 spin_unlock(&inode->i_lock);
388         }
389         return generic_permission(inode, mask);
390 }
391
392 /**
393  * __inode_permission - Check for access rights to a given inode
394  * @inode: Inode to check permission on
395  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
396  *
397  * Check for read/write/execute permissions on an inode.
398  *
399  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
400  *
401  * This does not check for a read-only file system.  You probably want
402  * inode_permission().
403  */
404 int __inode_permission(struct inode *inode, int mask)
405 {
406         int retval;
407
408         if (unlikely(mask & MAY_WRITE)) {
409                 /*
410                  * Nobody gets write access to an immutable file.
411                  */
412                 if (IS_IMMUTABLE(inode))
413                         return -EPERM;
414
415                 /*
416                  * Updating mtime will likely cause i_uid and i_gid to be
417                  * written back improperly if their true value is unknown
418                  * to the vfs.
419                  */
420                 if (HAS_UNMAPPED_ID(inode))
421                         return -EACCES;
422         }
423
424         retval = do_inode_permission(inode, mask);
425         if (retval)
426                 return retval;
427
428         retval = devcgroup_inode_permission(inode, mask);
429         if (retval)
430                 return retval;
431
432         return security_inode_permission(inode, mask);
433 }
434 EXPORT_SYMBOL(__inode_permission);
435
436 /**
437  * sb_permission - Check superblock-level permissions
438  * @sb: Superblock of inode to check permission on
439  * @inode: Inode to check permission on
440  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
441  *
442  * Separate out file-system wide checks from inode-specific permission checks.
443  */
444 static int sb_permission(struct super_block *sb, struct inode *inode, int mask)
445 {
446         if (unlikely(mask & MAY_WRITE)) {
447                 umode_t mode = inode->i_mode;
448
449                 /* Nobody gets write access to a read-only fs. */
450                 if ((sb->s_flags & MS_RDONLY) &&
451                     (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
452                         return -EROFS;
453         }
454         return 0;
455 }
456
457 /**
458  * inode_permission - Check for access rights to a given inode
459  * @inode: Inode to check permission on
460  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
461  *
462  * Check for read/write/execute permissions on an inode.  We use fs[ug]id for
463  * this, letting us set arbitrary permissions for filesystem access without
464  * changing the "normal" UIDs which are used for other things.
465  *
466  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
467  */
468 int inode_permission(struct inode *inode, int mask)
469 {
470         int retval;
471
472         retval = sb_permission(inode->i_sb, inode, mask);
473         if (retval)
474                 return retval;
475         return __inode_permission(inode, mask);
476 }
477 EXPORT_SYMBOL(inode_permission);
478
479 /**
480  * path_get - get a reference to a path
481  * @path: path to get the reference to
482  *
483  * Given a path increment the reference count to the dentry and the vfsmount.
484  */
485 void path_get(const struct path *path)
486 {
487         mntget(path->mnt);
488         dget(path->dentry);
489 }
490 EXPORT_SYMBOL(path_get);
491
492 /**
493  * path_put - put a reference to a path
494  * @path: path to put the reference to
495  *
496  * Given a path decrement the reference count to the dentry and the vfsmount.
497  */
498 void path_put(const struct path *path)
499 {
500         dput(path->dentry);
501         mntput(path->mnt);
502 }
503 EXPORT_SYMBOL(path_put);
504
505 #define EMBEDDED_LEVELS 2
506 struct nameidata {
507         struct path     path;
508         struct qstr     last;
509         struct path     root;
510         struct inode    *inode; /* path.dentry.d_inode */
511         unsigned int    flags;
512         unsigned        seq, m_seq;
513         int             last_type;
514         unsigned        depth;
515         int             total_link_count;
516         struct saved {
517                 struct path link;
518                 struct delayed_call done;
519                 const char *name;
520                 unsigned seq;
521         } *stack, internal[EMBEDDED_LEVELS];
522         struct filename *name;
523         struct nameidata *saved;
524         struct inode    *link_inode;
525         unsigned        root_seq;
526         int             dfd;
527 } __randomize_layout;
528
529 static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
530 {
531         struct nameidata *old = current->nameidata;
532         p->stack = p->internal;
533         p->dfd = dfd;
534         p->name = name;
535         p->total_link_count = old ? old->total_link_count : 0;
536         p->saved = old;
537         current->nameidata = p;
538 }
539
540 static void restore_nameidata(void)
541 {
542         struct nameidata *now = current->nameidata, *old = now->saved;
543
544         current->nameidata = old;
545         if (old)
546                 old->total_link_count = now->total_link_count;
547         if (now->stack != now->internal)
548                 kfree(now->stack);
549 }
550
551 static int __nd_alloc_stack(struct nameidata *nd)
552 {
553         struct saved *p;
554
555         if (nd->flags & LOOKUP_RCU) {
556                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
557                                   GFP_ATOMIC);
558                 if (unlikely(!p))
559                         return -ECHILD;
560         } else {
561                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
562                                   GFP_KERNEL);
563                 if (unlikely(!p))
564                         return -ENOMEM;
565         }
566         memcpy(p, nd->internal, sizeof(nd->internal));
567         nd->stack = p;
568         return 0;
569 }
570
571 /**
572  * path_connected - Verify that a path->dentry is below path->mnt.mnt_root
573  * @path: nameidate to verify
574  *
575  * Rename can sometimes move a file or directory outside of a bind
576  * mount, path_connected allows those cases to be detected.
577  */
578 static bool path_connected(const struct path *path)
579 {
580         struct vfsmount *mnt = path->mnt;
581
582         /* Only bind mounts can have disconnected paths */
583         if (mnt->mnt_root == mnt->mnt_sb->s_root)
584                 return true;
585
586         return is_subdir(path->dentry, mnt->mnt_root);
587 }
588
589 static inline int nd_alloc_stack(struct nameidata *nd)
590 {
591         if (likely(nd->depth != EMBEDDED_LEVELS))
592                 return 0;
593         if (likely(nd->stack != nd->internal))
594                 return 0;
595         return __nd_alloc_stack(nd);
596 }
597
598 static void drop_links(struct nameidata *nd)
599 {
600         int i = nd->depth;
601         while (i--) {
602                 struct saved *last = nd->stack + i;
603                 do_delayed_call(&last->done);
604                 clear_delayed_call(&last->done);
605         }
606 }
607
608 static void terminate_walk(struct nameidata *nd)
609 {
610         drop_links(nd);
611         if (!(nd->flags & LOOKUP_RCU)) {
612                 int i;
613                 path_put(&nd->path);
614                 for (i = 0; i < nd->depth; i++)
615                         path_put(&nd->stack[i].link);
616                 if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
617                         path_put(&nd->root);
618                         nd->root.mnt = NULL;
619                 }
620         } else {
621                 nd->flags &= ~LOOKUP_RCU;
622                 if (!(nd->flags & LOOKUP_ROOT))
623                         nd->root.mnt = NULL;
624                 rcu_read_unlock();
625         }
626         nd->depth = 0;
627 }
628
629 /* path_put is needed afterwards regardless of success or failure */
630 static bool legitimize_path(struct nameidata *nd,
631                             struct path *path, unsigned seq)
632 {
633         int res = __legitimize_mnt(path->mnt, nd->m_seq);
634         if (unlikely(res)) {
635                 if (res > 0)
636                         path->mnt = NULL;
637                 path->dentry = NULL;
638                 return false;
639         }
640         if (unlikely(!lockref_get_not_dead(&path->dentry->d_lockref))) {
641                 path->dentry = NULL;
642                 return false;
643         }
644         return !read_seqcount_retry(&path->dentry->d_seq, seq);
645 }
646
647 static bool legitimize_links(struct nameidata *nd)
648 {
649         int i;
650         for (i = 0; i < nd->depth; i++) {
651                 struct saved *last = nd->stack + i;
652                 if (unlikely(!legitimize_path(nd, &last->link, last->seq))) {
653                         drop_links(nd);
654                         nd->depth = i + 1;
655                         return false;
656                 }
657         }
658         return true;
659 }
660
661 /*
662  * Path walking has 2 modes, rcu-walk and ref-walk (see
663  * Documentation/filesystems/path-lookup.txt).  In situations when we can't
664  * continue in RCU mode, we attempt to drop out of rcu-walk mode and grab
665  * normal reference counts on dentries and vfsmounts to transition to ref-walk
666  * mode.  Refcounts are grabbed at the last known good point before rcu-walk
667  * got stuck, so ref-walk may continue from there. If this is not successful
668  * (eg. a seqcount has changed), then failure is returned and it's up to caller
669  * to restart the path walk from the beginning in ref-walk mode.
670  */
671
672 /**
673  * unlazy_walk - try to switch to ref-walk mode.
674  * @nd: nameidata pathwalk data
675  * Returns: 0 on success, -ECHILD on failure
676  *
677  * unlazy_walk attempts to legitimize the current nd->path and nd->root
678  * for ref-walk mode.
679  * Must be called from rcu-walk context.
680  * Nothing should touch nameidata between unlazy_walk() failure and
681  * terminate_walk().
682  */
683 static int unlazy_walk(struct nameidata *nd)
684 {
685         struct dentry *parent = nd->path.dentry;
686
687         BUG_ON(!(nd->flags & LOOKUP_RCU));
688
689         nd->flags &= ~LOOKUP_RCU;
690         if (unlikely(!legitimize_links(nd)))
691                 goto out2;
692         if (unlikely(!legitimize_path(nd, &nd->path, nd->seq)))
693                 goto out1;
694         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
695                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq)))
696                         goto out;
697         }
698         rcu_read_unlock();
699         BUG_ON(nd->inode != parent->d_inode);
700         return 0;
701
702 out2:
703         nd->path.mnt = NULL;
704         nd->path.dentry = NULL;
705 out1:
706         if (!(nd->flags & LOOKUP_ROOT))
707                 nd->root.mnt = NULL;
708 out:
709         rcu_read_unlock();
710         return -ECHILD;
711 }
712
713 /**
714  * unlazy_child - try to switch to ref-walk mode.
715  * @nd: nameidata pathwalk data
716  * @dentry: child of nd->path.dentry
717  * @seq: seq number to check dentry against
718  * Returns: 0 on success, -ECHILD on failure
719  *
720  * unlazy_child attempts to legitimize the current nd->path, nd->root and dentry
721  * for ref-walk mode.  @dentry must be a path found by a do_lookup call on
722  * @nd.  Must be called from rcu-walk context.
723  * Nothing should touch nameidata between unlazy_child() failure and
724  * terminate_walk().
725  */
726 static int unlazy_child(struct nameidata *nd, struct dentry *dentry, unsigned seq)
727 {
728         BUG_ON(!(nd->flags & LOOKUP_RCU));
729
730         nd->flags &= ~LOOKUP_RCU;
731         if (unlikely(!legitimize_links(nd)))
732                 goto out2;
733         if (unlikely(!legitimize_mnt(nd->path.mnt, nd->m_seq)))
734                 goto out2;
735         if (unlikely(!lockref_get_not_dead(&nd->path.dentry->d_lockref)))
736                 goto out1;
737
738         /*
739          * We need to move both the parent and the dentry from the RCU domain
740          * to be properly refcounted. And the sequence number in the dentry
741          * validates *both* dentry counters, since we checked the sequence
742          * number of the parent after we got the child sequence number. So we
743          * know the parent must still be valid if the child sequence number is
744          */
745         if (unlikely(!lockref_get_not_dead(&dentry->d_lockref)))
746                 goto out;
747         if (unlikely(read_seqcount_retry(&dentry->d_seq, seq))) {
748                 rcu_read_unlock();
749                 dput(dentry);
750                 goto drop_root_mnt;
751         }
752         /*
753          * Sequence counts matched. Now make sure that the root is
754          * still valid and get it if required.
755          */
756         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
757                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq))) {
758                         rcu_read_unlock();
759                         dput(dentry);
760                         return -ECHILD;
761                 }
762         }
763
764         rcu_read_unlock();
765         return 0;
766
767 out2:
768         nd->path.mnt = NULL;
769 out1:
770         nd->path.dentry = NULL;
771 out:
772         rcu_read_unlock();
773 drop_root_mnt:
774         if (!(nd->flags & LOOKUP_ROOT))
775                 nd->root.mnt = NULL;
776         return -ECHILD;
777 }
778
779 static inline int d_revalidate(struct dentry *dentry, unsigned int flags)
780 {
781         if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
782                 return dentry->d_op->d_revalidate(dentry, flags);
783         else
784                 return 1;
785 }
786
787 /**
788  * complete_walk - successful completion of path walk
789  * @nd:  pointer nameidata
790  *
791  * If we had been in RCU mode, drop out of it and legitimize nd->path.
792  * Revalidate the final result, unless we'd already done that during
793  * the path walk or the filesystem doesn't ask for it.  Return 0 on
794  * success, -error on failure.  In case of failure caller does not
795  * need to drop nd->path.
796  */
797 static int complete_walk(struct nameidata *nd)
798 {
799         struct dentry *dentry = nd->path.dentry;
800         int status;
801
802         if (nd->flags & LOOKUP_RCU) {
803                 if (!(nd->flags & LOOKUP_ROOT))
804                         nd->root.mnt = NULL;
805                 if (unlikely(unlazy_walk(nd)))
806                         return -ECHILD;
807         }
808
809         if (likely(!(nd->flags & LOOKUP_JUMPED)))
810                 return 0;
811
812         if (likely(!(dentry->d_flags & DCACHE_OP_WEAK_REVALIDATE)))
813                 return 0;
814
815         status = dentry->d_op->d_weak_revalidate(dentry, nd->flags);
816         if (status > 0)
817                 return 0;
818
819         if (!status)
820                 status = -ESTALE;
821
822         return status;
823 }
824
825 static void set_root(struct nameidata *nd)
826 {
827         struct fs_struct *fs = current->fs;
828
829         if (nd->flags & LOOKUP_RCU) {
830                 unsigned seq;
831
832                 do {
833                         seq = read_seqcount_begin(&fs->seq);
834                         nd->root = fs->root;
835                         nd->root_seq = __read_seqcount_begin(&nd->root.dentry->d_seq);
836                 } while (read_seqcount_retry(&fs->seq, seq));
837         } else {
838                 get_fs_root(fs, &nd->root);
839         }
840 }
841
842 static void path_put_conditional(struct path *path, struct nameidata *nd)
843 {
844         dput(path->dentry);
845         if (path->mnt != nd->path.mnt)
846                 mntput(path->mnt);
847 }
848
849 static inline void path_to_nameidata(const struct path *path,
850                                         struct nameidata *nd)
851 {
852         if (!(nd->flags & LOOKUP_RCU)) {
853                 dput(nd->path.dentry);
854                 if (nd->path.mnt != path->mnt)
855                         mntput(nd->path.mnt);
856         }
857         nd->path.mnt = path->mnt;
858         nd->path.dentry = path->dentry;
859 }
860
861 static int nd_jump_root(struct nameidata *nd)
862 {
863         if (nd->flags & LOOKUP_RCU) {
864                 struct dentry *d;
865                 nd->path = nd->root;
866                 d = nd->path.dentry;
867                 nd->inode = d->d_inode;
868                 nd->seq = nd->root_seq;
869                 if (unlikely(read_seqcount_retry(&d->d_seq, nd->seq)))
870                         return -ECHILD;
871         } else {
872                 path_put(&nd->path);
873                 nd->path = nd->root;
874                 path_get(&nd->path);
875                 nd->inode = nd->path.dentry->d_inode;
876         }
877         nd->flags |= LOOKUP_JUMPED;
878         return 0;
879 }
880
881 /*
882  * Helper to directly jump to a known parsed path from ->get_link,
883  * caller must have taken a reference to path beforehand.
884  */
885 void nd_jump_link(struct path *path)
886 {
887         struct nameidata *nd = current->nameidata;
888         path_put(&nd->path);
889
890         nd->path = *path;
891         nd->inode = nd->path.dentry->d_inode;
892         nd->flags |= LOOKUP_JUMPED;
893 }
894
895 static inline void put_link(struct nameidata *nd)
896 {
897         struct saved *last = nd->stack + --nd->depth;
898         do_delayed_call(&last->done);
899         if (!(nd->flags & LOOKUP_RCU))
900                 path_put(&last->link);
901 }
902
903 int sysctl_protected_symlinks __read_mostly = 0;
904 int sysctl_protected_hardlinks __read_mostly = 0;
905
906 /**
907  * may_follow_link - Check symlink following for unsafe situations
908  * @nd: nameidata pathwalk data
909  *
910  * In the case of the sysctl_protected_symlinks sysctl being enabled,
911  * CAP_DAC_OVERRIDE needs to be specifically ignored if the symlink is
912  * in a sticky world-writable directory. This is to protect privileged
913  * processes from failing races against path names that may change out
914  * from under them by way of other users creating malicious symlinks.
915  * It will permit symlinks to be followed only when outside a sticky
916  * world-writable directory, or when the uid of the symlink and follower
917  * match, or when the directory owner matches the symlink's owner.
918  *
919  * Returns 0 if following the symlink is allowed, -ve on error.
920  */
921 static inline int may_follow_link(struct nameidata *nd)
922 {
923         const struct inode *inode;
924         const struct inode *parent;
925         kuid_t puid;
926
927         if (!sysctl_protected_symlinks)
928                 return 0;
929
930         /* Allowed if owner and follower match. */
931         inode = nd->link_inode;
932         if (uid_eq(current_cred()->fsuid, inode->i_uid))
933                 return 0;
934
935         /* Allowed if parent directory not sticky and world-writable. */
936         parent = nd->inode;
937         if ((parent->i_mode & (S_ISVTX|S_IWOTH)) != (S_ISVTX|S_IWOTH))
938                 return 0;
939
940         /* Allowed if parent directory and link owner match. */
941         puid = parent->i_uid;
942         if (uid_valid(puid) && uid_eq(puid, inode->i_uid))
943                 return 0;
944
945         if (nd->flags & LOOKUP_RCU)
946                 return -ECHILD;
947
948         audit_log_link_denied("follow_link", &nd->stack[0].link);
949         return -EACCES;
950 }
951
952 /**
953  * safe_hardlink_source - Check for safe hardlink conditions
954  * @inode: the source inode to hardlink from
955  *
956  * Return false if at least one of the following conditions:
957  *    - inode is not a regular file
958  *    - inode is setuid
959  *    - inode is setgid and group-exec
960  *    - access failure for read and write
961  *
962  * Otherwise returns true.
963  */
964 static bool safe_hardlink_source(struct inode *inode)
965 {
966         umode_t mode = inode->i_mode;
967
968         /* Special files should not get pinned to the filesystem. */
969         if (!S_ISREG(mode))
970                 return false;
971
972         /* Setuid files should not get pinned to the filesystem. */
973         if (mode & S_ISUID)
974                 return false;
975
976         /* Executable setgid files should not get pinned to the filesystem. */
977         if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP))
978                 return false;
979
980         /* Hardlinking to unreadable or unwritable sources is dangerous. */
981         if (inode_permission(inode, MAY_READ | MAY_WRITE))
982                 return false;
983
984         return true;
985 }
986
987 /**
988  * may_linkat - Check permissions for creating a hardlink
989  * @link: the source to hardlink from
990  *
991  * Block hardlink when all of:
992  *  - sysctl_protected_hardlinks enabled
993  *  - fsuid does not match inode
994  *  - hardlink source is unsafe (see safe_hardlink_source() above)
995  *  - not CAP_FOWNER in a namespace with the inode owner uid mapped
996  *
997  * Returns 0 if successful, -ve on error.
998  */
999 static int may_linkat(struct path *link)
1000 {
1001         struct inode *inode;
1002
1003         if (!sysctl_protected_hardlinks)
1004                 return 0;
1005
1006         inode = link->dentry->d_inode;
1007
1008         /* Source inode owner (or CAP_FOWNER) can hardlink all they like,
1009          * otherwise, it must be a safe source.
1010          */
1011         if (safe_hardlink_source(inode) || inode_owner_or_capable(inode))
1012                 return 0;
1013
1014         audit_log_link_denied("linkat", link);
1015         return -EPERM;
1016 }
1017
1018 static __always_inline
1019 const char *get_link(struct nameidata *nd)
1020 {
1021         struct saved *last = nd->stack + nd->depth - 1;
1022         struct dentry *dentry = last->link.dentry;
1023         struct inode *inode = nd->link_inode;
1024         int error;
1025         const char *res;
1026
1027         if (!(nd->flags & LOOKUP_RCU)) {
1028                 touch_atime(&last->link);
1029                 cond_resched();
1030         } else if (atime_needs_update_rcu(&last->link, inode)) {
1031                 if (unlikely(unlazy_walk(nd)))
1032                         return ERR_PTR(-ECHILD);
1033                 touch_atime(&last->link);
1034         }
1035
1036         error = security_inode_follow_link(dentry, inode,
1037                                            nd->flags & LOOKUP_RCU);
1038         if (unlikely(error))
1039                 return ERR_PTR(error);
1040
1041         nd->last_type = LAST_BIND;
1042         res = inode->i_link;
1043         if (!res) {
1044                 const char * (*get)(struct dentry *, struct inode *,
1045                                 struct delayed_call *);
1046                 get = inode->i_op->get_link;
1047                 if (nd->flags & LOOKUP_RCU) {
1048                         res = get(NULL, inode, &last->done);
1049                         if (res == ERR_PTR(-ECHILD)) {
1050                                 if (unlikely(unlazy_walk(nd)))
1051                                         return ERR_PTR(-ECHILD);
1052                                 res = get(dentry, inode, &last->done);
1053                         }
1054                 } else {
1055                         res = get(dentry, inode, &last->done);
1056                 }
1057                 if (IS_ERR_OR_NULL(res))
1058                         return res;
1059         }
1060         if (*res == '/') {
1061                 if (!nd->root.mnt)
1062                         set_root(nd);
1063                 if (unlikely(nd_jump_root(nd)))
1064                         return ERR_PTR(-ECHILD);
1065                 while (unlikely(*++res == '/'))
1066                         ;
1067         }
1068         if (!*res)
1069                 res = NULL;
1070         return res;
1071 }
1072
1073 /*
1074  * follow_up - Find the mountpoint of path's vfsmount
1075  *
1076  * Given a path, find the mountpoint of its source file system.
1077  * Replace @path with the path of the mountpoint in the parent mount.
1078  * Up is towards /.
1079  *
1080  * Return 1 if we went up a level and 0 if we were already at the
1081  * root.
1082  */
1083 int follow_up(struct path *path)
1084 {
1085         struct mount *mnt = real_mount(path->mnt);
1086         struct mount *parent;
1087         struct dentry *mountpoint;
1088
1089         read_seqlock_excl(&mount_lock);
1090         parent = mnt->mnt_parent;
1091         if (parent == mnt) {
1092                 read_sequnlock_excl(&mount_lock);
1093                 return 0;
1094         }
1095         mntget(&parent->mnt);
1096         mountpoint = dget(mnt->mnt_mountpoint);
1097         read_sequnlock_excl(&mount_lock);
1098         dput(path->dentry);
1099         path->dentry = mountpoint;
1100         mntput(path->mnt);
1101         path->mnt = &parent->mnt;
1102         return 1;
1103 }
1104 EXPORT_SYMBOL(follow_up);
1105
1106 /*
1107  * Perform an automount
1108  * - return -EISDIR to tell follow_managed() to stop and return the path we
1109  *   were called with.
1110  */
1111 static int follow_automount(struct path *path, struct nameidata *nd,
1112                             bool *need_mntput)
1113 {
1114         struct vfsmount *mnt;
1115         int err;
1116
1117         if (!path->dentry->d_op || !path->dentry->d_op->d_automount)
1118                 return -EREMOTE;
1119
1120         /* We don't want to mount if someone's just doing a stat -
1121          * unless they're stat'ing a directory and appended a '/' to
1122          * the name.
1123          *
1124          * We do, however, want to mount if someone wants to open or
1125          * create a file of any type under the mountpoint, wants to
1126          * traverse through the mountpoint or wants to open the
1127          * mounted directory.  Also, autofs may mark negative dentries
1128          * as being automount points.  These will need the attentions
1129          * of the daemon to instantiate them before they can be used.
1130          */
1131         if (!(nd->flags & (LOOKUP_PARENT | LOOKUP_DIRECTORY |
1132                            LOOKUP_OPEN | LOOKUP_CREATE | LOOKUP_AUTOMOUNT)) &&
1133             path->dentry->d_inode)
1134                 return -EISDIR;
1135
1136         if (path->dentry->d_sb->s_user_ns != &init_user_ns)
1137                 return -EACCES;
1138
1139         nd->total_link_count++;
1140         if (nd->total_link_count >= 40)
1141                 return -ELOOP;
1142
1143         mnt = path->dentry->d_op->d_automount(path);
1144         if (IS_ERR(mnt)) {
1145                 /*
1146                  * The filesystem is allowed to return -EISDIR here to indicate
1147                  * it doesn't want to automount.  For instance, autofs would do
1148                  * this so that its userspace daemon can mount on this dentry.
1149                  *
1150                  * However, we can only permit this if it's a terminal point in
1151                  * the path being looked up; if it wasn't then the remainder of
1152                  * the path is inaccessible and we should say so.
1153                  */
1154                 if (PTR_ERR(mnt) == -EISDIR && (nd->flags & LOOKUP_PARENT))
1155                         return -EREMOTE;
1156                 return PTR_ERR(mnt);
1157         }
1158
1159         if (!mnt) /* mount collision */
1160                 return 0;
1161
1162         if (!*need_mntput) {
1163                 /* lock_mount() may release path->mnt on error */
1164                 mntget(path->mnt);
1165                 *need_mntput = true;
1166         }
1167         err = finish_automount(mnt, path);
1168
1169         switch (err) {
1170         case -EBUSY:
1171                 /* Someone else made a mount here whilst we were busy */
1172                 return 0;
1173         case 0:
1174                 path_put(path);
1175                 path->mnt = mnt;
1176                 path->dentry = dget(mnt->mnt_root);
1177                 return 0;
1178         default:
1179                 return err;
1180         }
1181
1182 }
1183
1184 /*
1185  * Handle a dentry that is managed in some way.
1186  * - Flagged for transit management (autofs)
1187  * - Flagged as mountpoint
1188  * - Flagged as automount point
1189  *
1190  * This may only be called in refwalk mode.
1191  *
1192  * Serialization is taken care of in namespace.c
1193  */
1194 static int follow_managed(struct path *path, struct nameidata *nd)
1195 {
1196         struct vfsmount *mnt = path->mnt; /* held by caller, must be left alone */
1197         unsigned managed;
1198         bool need_mntput = false;
1199         int ret = 0;
1200
1201         /* Given that we're not holding a lock here, we retain the value in a
1202          * local variable for each dentry as we look at it so that we don't see
1203          * the components of that value change under us */
1204         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1205                managed &= DCACHE_MANAGED_DENTRY,
1206                unlikely(managed != 0)) {
1207                 /* Allow the filesystem to manage the transit without i_mutex
1208                  * being held. */
1209                 if (managed & DCACHE_MANAGE_TRANSIT) {
1210                         BUG_ON(!path->dentry->d_op);
1211                         BUG_ON(!path->dentry->d_op->d_manage);
1212                         ret = path->dentry->d_op->d_manage(path, false);
1213                         if (ret < 0)
1214                                 break;
1215                 }
1216
1217                 /* Transit to a mounted filesystem. */
1218                 if (managed & DCACHE_MOUNTED) {
1219                         struct vfsmount *mounted = lookup_mnt(path);
1220                         if (mounted) {
1221                                 dput(path->dentry);
1222                                 if (need_mntput)
1223                                         mntput(path->mnt);
1224                                 path->mnt = mounted;
1225                                 path->dentry = dget(mounted->mnt_root);
1226                                 need_mntput = true;
1227                                 continue;
1228                         }
1229
1230                         /* Something is mounted on this dentry in another
1231                          * namespace and/or whatever was mounted there in this
1232                          * namespace got unmounted before lookup_mnt() could
1233                          * get it */
1234                 }
1235
1236                 /* Handle an automount point */
1237                 if (managed & DCACHE_NEED_AUTOMOUNT) {
1238                         ret = follow_automount(path, nd, &need_mntput);
1239                         if (ret < 0)
1240                                 break;
1241                         continue;
1242                 }
1243
1244                 /* We didn't change the current path point */
1245                 break;
1246         }
1247
1248         if (need_mntput && path->mnt == mnt)
1249                 mntput(path->mnt);
1250         if (ret == -EISDIR || !ret)
1251                 ret = 1;
1252         if (need_mntput)
1253                 nd->flags |= LOOKUP_JUMPED;
1254         if (unlikely(ret < 0))
1255                 path_put_conditional(path, nd);
1256         return ret;
1257 }
1258
1259 int follow_down_one(struct path *path)
1260 {
1261         struct vfsmount *mounted;
1262
1263         mounted = lookup_mnt(path);
1264         if (mounted) {
1265                 dput(path->dentry);
1266                 mntput(path->mnt);
1267                 path->mnt = mounted;
1268                 path->dentry = dget(mounted->mnt_root);
1269                 return 1;
1270         }
1271         return 0;
1272 }
1273 EXPORT_SYMBOL(follow_down_one);
1274
1275 static inline int managed_dentry_rcu(const struct path *path)
1276 {
1277         return (path->dentry->d_flags & DCACHE_MANAGE_TRANSIT) ?
1278                 path->dentry->d_op->d_manage(path, true) : 0;
1279 }
1280
1281 /*
1282  * Try to skip to top of mountpoint pile in rcuwalk mode.  Fail if
1283  * we meet a managed dentry that would need blocking.
1284  */
1285 static bool __follow_mount_rcu(struct nameidata *nd, struct path *path,
1286                                struct inode **inode, unsigned *seqp)
1287 {
1288         for (;;) {
1289                 struct mount *mounted;
1290                 /*
1291                  * Don't forget we might have a non-mountpoint managed dentry
1292                  * that wants to block transit.
1293                  */
1294                 switch (managed_dentry_rcu(path)) {
1295                 case -ECHILD:
1296                 default:
1297                         return false;
1298                 case -EISDIR:
1299                         return true;
1300                 case 0:
1301                         break;
1302                 }
1303
1304                 if (!d_mountpoint(path->dentry))
1305                         return !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1306
1307                 mounted = __lookup_mnt(path->mnt, path->dentry);
1308                 if (!mounted)
1309                         break;
1310                 path->mnt = &mounted->mnt;
1311                 path->dentry = mounted->mnt.mnt_root;
1312                 nd->flags |= LOOKUP_JUMPED;
1313                 *seqp = read_seqcount_begin(&path->dentry->d_seq);
1314                 /*
1315                  * Update the inode too. We don't need to re-check the
1316                  * dentry sequence number here after this d_inode read,
1317                  * because a mount-point is always pinned.
1318                  */
1319                 *inode = path->dentry->d_inode;
1320         }
1321         return !read_seqretry(&mount_lock, nd->m_seq) &&
1322                 !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1323 }
1324
1325 static int follow_dotdot_rcu(struct nameidata *nd)
1326 {
1327         struct inode *inode = nd->inode;
1328
1329         while (1) {
1330                 if (path_equal(&nd->path, &nd->root))
1331                         break;
1332                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1333                         struct dentry *old = nd->path.dentry;
1334                         struct dentry *parent = old->d_parent;
1335                         unsigned seq;
1336
1337                         inode = parent->d_inode;
1338                         seq = read_seqcount_begin(&parent->d_seq);
1339                         if (unlikely(read_seqcount_retry(&old->d_seq, nd->seq)))
1340                                 return -ECHILD;
1341                         nd->path.dentry = parent;
1342                         nd->seq = seq;
1343                         if (unlikely(!path_connected(&nd->path)))
1344                                 return -ENOENT;
1345                         break;
1346                 } else {
1347                         struct mount *mnt = real_mount(nd->path.mnt);
1348                         struct mount *mparent = mnt->mnt_parent;
1349                         struct dentry *mountpoint = mnt->mnt_mountpoint;
1350                         struct inode *inode2 = mountpoint->d_inode;
1351                         unsigned seq = read_seqcount_begin(&mountpoint->d_seq);
1352                         if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1353                                 return -ECHILD;
1354                         if (&mparent->mnt == nd->path.mnt)
1355                                 break;
1356                         /* we know that mountpoint was pinned */
1357                         nd->path.dentry = mountpoint;
1358                         nd->path.mnt = &mparent->mnt;
1359                         inode = inode2;
1360                         nd->seq = seq;
1361                 }
1362         }
1363         while (unlikely(d_mountpoint(nd->path.dentry))) {
1364                 struct mount *mounted;
1365                 mounted = __lookup_mnt(nd->path.mnt, nd->path.dentry);
1366                 if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1367                         return -ECHILD;
1368                 if (!mounted)
1369                         break;
1370                 nd->path.mnt = &mounted->mnt;
1371                 nd->path.dentry = mounted->mnt.mnt_root;
1372                 inode = nd->path.dentry->d_inode;
1373                 nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
1374         }
1375         nd->inode = inode;
1376         return 0;
1377 }
1378
1379 /*
1380  * Follow down to the covering mount currently visible to userspace.  At each
1381  * point, the filesystem owning that dentry may be queried as to whether the
1382  * caller is permitted to proceed or not.
1383  */
1384 int follow_down(struct path *path)
1385 {
1386         unsigned managed;
1387         int ret;
1388
1389         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1390                unlikely(managed & DCACHE_MANAGED_DENTRY)) {
1391                 /* Allow the filesystem to manage the transit without i_mutex
1392                  * being held.
1393                  *
1394                  * We indicate to the filesystem if someone is trying to mount
1395                  * something here.  This gives autofs the chance to deny anyone
1396                  * other than its daemon the right to mount on its
1397                  * superstructure.
1398                  *
1399                  * The filesystem may sleep at this point.
1400                  */
1401                 if (managed & DCACHE_MANAGE_TRANSIT) {
1402                         BUG_ON(!path->dentry->d_op);
1403                         BUG_ON(!path->dentry->d_op->d_manage);
1404                         ret = path->dentry->d_op->d_manage(path, false);
1405                         if (ret < 0)
1406                                 return ret == -EISDIR ? 0 : ret;
1407                 }
1408
1409                 /* Transit to a mounted filesystem. */
1410                 if (managed & DCACHE_MOUNTED) {
1411                         struct vfsmount *mounted = lookup_mnt(path);
1412                         if (!mounted)
1413                                 break;
1414                         dput(path->dentry);
1415                         mntput(path->mnt);
1416                         path->mnt = mounted;
1417                         path->dentry = dget(mounted->mnt_root);
1418                         continue;
1419                 }
1420
1421                 /* Don't handle automount points here */
1422                 break;
1423         }
1424         return 0;
1425 }
1426 EXPORT_SYMBOL(follow_down);
1427
1428 /*
1429  * Skip to top of mountpoint pile in refwalk mode for follow_dotdot()
1430  */
1431 static void follow_mount(struct path *path)
1432 {
1433         while (d_mountpoint(path->dentry)) {
1434                 struct vfsmount *mounted = lookup_mnt(path);
1435                 if (!mounted)
1436                         break;
1437                 dput(path->dentry);
1438                 mntput(path->mnt);
1439                 path->mnt = mounted;
1440                 path->dentry = dget(mounted->mnt_root);
1441         }
1442 }
1443
1444 static int path_parent_directory(struct path *path)
1445 {
1446         struct dentry *old = path->dentry;
1447         /* rare case of legitimate dget_parent()... */
1448         path->dentry = dget_parent(path->dentry);
1449         dput(old);
1450         if (unlikely(!path_connected(path)))
1451                 return -ENOENT;
1452         return 0;
1453 }
1454
1455 static int follow_dotdot(struct nameidata *nd)
1456 {
1457         while(1) {
1458                 if (nd->path.dentry == nd->root.dentry &&
1459                     nd->path.mnt == nd->root.mnt) {
1460                         break;
1461                 }
1462                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1463                         int ret = path_parent_directory(&nd->path);
1464                         if (ret)
1465                                 return ret;
1466                         break;
1467                 }
1468                 if (!follow_up(&nd->path))
1469                         break;
1470         }
1471         follow_mount(&nd->path);
1472         nd->inode = nd->path.dentry->d_inode;
1473         return 0;
1474 }
1475
1476 /*
1477  * This looks up the name in dcache and possibly revalidates the found dentry.
1478  * NULL is returned if the dentry does not exist in the cache.
1479  */
1480 static struct dentry *lookup_dcache(const struct qstr *name,
1481                                     struct dentry *dir,
1482                                     unsigned int flags)
1483 {
1484         struct dentry *dentry = d_lookup(dir, name);
1485         if (dentry) {
1486                 int error = d_revalidate(dentry, flags);
1487                 if (unlikely(error <= 0)) {
1488                         if (!error)
1489                                 d_invalidate(dentry);
1490                         dput(dentry);
1491                         return ERR_PTR(error);
1492                 }
1493         }
1494         return dentry;
1495 }
1496
1497 /*
1498  * Call i_op->lookup on the dentry.  The dentry must be negative and
1499  * unhashed.
1500  *
1501  * dir->d_inode->i_mutex must be held
1502  */
1503 static struct dentry *lookup_real(struct inode *dir, struct dentry *dentry,
1504                                   unsigned int flags)
1505 {
1506         struct dentry *old;
1507
1508         /* Don't create child dentry for a dead directory. */
1509         if (unlikely(IS_DEADDIR(dir))) {
1510                 dput(dentry);
1511                 return ERR_PTR(-ENOENT);
1512         }
1513
1514         old = dir->i_op->lookup(dir, dentry, flags);
1515         if (unlikely(old)) {
1516                 dput(dentry);
1517                 dentry = old;
1518         }
1519         return dentry;
1520 }
1521
1522 static struct dentry *__lookup_hash(const struct qstr *name,
1523                 struct dentry *base, unsigned int flags)
1524 {
1525         struct dentry *dentry = lookup_dcache(name, base, flags);
1526
1527         if (dentry)
1528                 return dentry;
1529
1530         dentry = d_alloc(base, name);
1531         if (unlikely(!dentry))
1532                 return ERR_PTR(-ENOMEM);
1533
1534         return lookup_real(base->d_inode, dentry, flags);
1535 }
1536
1537 static int lookup_fast(struct nameidata *nd,
1538                        struct path *path, struct inode **inode,
1539                        unsigned *seqp)
1540 {
1541         struct vfsmount *mnt = nd->path.mnt;
1542         struct dentry *dentry, *parent = nd->path.dentry;
1543         int status = 1;
1544         int err;
1545
1546         /*
1547          * Rename seqlock is not required here because in the off chance
1548          * of a false negative due to a concurrent rename, the caller is
1549          * going to fall back to non-racy lookup.
1550          */
1551         if (nd->flags & LOOKUP_RCU) {
1552                 unsigned seq;
1553                 bool negative;
1554                 dentry = __d_lookup_rcu(parent, &nd->last, &seq);
1555                 if (unlikely(!dentry)) {
1556                         if (unlazy_walk(nd))
1557                                 return -ECHILD;
1558                         return 0;
1559                 }
1560
1561                 /*
1562                  * This sequence count validates that the inode matches
1563                  * the dentry name information from lookup.
1564                  */
1565                 *inode = d_backing_inode(dentry);
1566                 negative = d_is_negative(dentry);
1567                 if (unlikely(read_seqcount_retry(&dentry->d_seq, seq)))
1568                         return -ECHILD;
1569
1570                 /*
1571                  * This sequence count validates that the parent had no
1572                  * changes while we did the lookup of the dentry above.
1573                  *
1574                  * The memory barrier in read_seqcount_begin of child is
1575                  *  enough, we can use __read_seqcount_retry here.
1576                  */
1577                 if (unlikely(__read_seqcount_retry(&parent->d_seq, nd->seq)))
1578                         return -ECHILD;
1579
1580                 *seqp = seq;
1581                 status = d_revalidate(dentry, nd->flags);
1582                 if (likely(status > 0)) {
1583                         /*
1584                          * Note: do negative dentry check after revalidation in
1585                          * case that drops it.
1586                          */
1587                         if (unlikely(negative))
1588                                 return -ENOENT;
1589                         path->mnt = mnt;
1590                         path->dentry = dentry;
1591                         if (likely(__follow_mount_rcu(nd, path, inode, seqp)))
1592                                 return 1;
1593                 }
1594                 if (unlazy_child(nd, dentry, seq))
1595                         return -ECHILD;
1596                 if (unlikely(status == -ECHILD))
1597                         /* we'd been told to redo it in non-rcu mode */
1598                         status = d_revalidate(dentry, nd->flags);
1599         } else {
1600                 dentry = __d_lookup(parent, &nd->last);
1601                 if (unlikely(!dentry))
1602                         return 0;
1603                 status = d_revalidate(dentry, nd->flags);
1604         }
1605         if (unlikely(status <= 0)) {
1606                 if (!status)
1607                         d_invalidate(dentry);
1608                 dput(dentry);
1609                 return status;
1610         }
1611         if (unlikely(d_is_negative(dentry))) {
1612                 dput(dentry);
1613                 return -ENOENT;
1614         }
1615
1616         path->mnt = mnt;
1617         path->dentry = dentry;
1618         err = follow_managed(path, nd);
1619         if (likely(err > 0))
1620                 *inode = d_backing_inode(path->dentry);
1621         return err;
1622 }
1623
1624 /* Fast lookup failed, do it the slow way */
1625 static struct dentry *lookup_slow(const struct qstr *name,
1626                                   struct dentry *dir,
1627                                   unsigned int flags)
1628 {
1629         struct dentry *dentry = ERR_PTR(-ENOENT), *old;
1630         struct inode *inode = dir->d_inode;
1631         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
1632
1633         inode_lock_shared(inode);
1634         /* Don't go there if it's already dead */
1635         if (unlikely(IS_DEADDIR(inode)))
1636                 goto out;
1637 again:
1638         dentry = d_alloc_parallel(dir, name, &wq);
1639         if (IS_ERR(dentry))
1640                 goto out;
1641         if (unlikely(!d_in_lookup(dentry))) {
1642                 if (!(flags & LOOKUP_NO_REVAL)) {
1643                         int error = d_revalidate(dentry, flags);
1644                         if (unlikely(error <= 0)) {
1645                                 if (!error) {
1646                                         d_invalidate(dentry);
1647                                         dput(dentry);
1648                                         goto again;
1649                                 }
1650                                 dput(dentry);
1651                                 dentry = ERR_PTR(error);
1652                         }
1653                 }
1654         } else {
1655                 old = inode->i_op->lookup(inode, dentry, flags);
1656                 d_lookup_done(dentry);
1657                 if (unlikely(old)) {
1658                         dput(dentry);
1659                         dentry = old;
1660                 }
1661         }
1662 out:
1663         inode_unlock_shared(inode);
1664         return dentry;
1665 }
1666
1667 static inline int may_lookup(struct nameidata *nd)
1668 {
1669         if (nd->flags & LOOKUP_RCU) {
1670                 int err = inode_permission(nd->inode, MAY_EXEC|MAY_NOT_BLOCK);
1671                 if (err != -ECHILD)
1672                         return err;
1673                 if (unlazy_walk(nd))
1674                         return -ECHILD;
1675         }
1676         return inode_permission(nd->inode, MAY_EXEC);
1677 }
1678
1679 static inline int handle_dots(struct nameidata *nd, int type)
1680 {
1681         if (type == LAST_DOTDOT) {
1682                 if (!nd->root.mnt)
1683                         set_root(nd);
1684                 if (nd->flags & LOOKUP_RCU) {
1685                         return follow_dotdot_rcu(nd);
1686                 } else
1687                         return follow_dotdot(nd);
1688         }
1689         return 0;
1690 }
1691
1692 static int pick_link(struct nameidata *nd, struct path *link,
1693                      struct inode *inode, unsigned seq)
1694 {
1695         int error;
1696         struct saved *last;
1697         if (unlikely(nd->total_link_count++ >= MAXSYMLINKS)) {
1698                 path_to_nameidata(link, nd);
1699                 return -ELOOP;
1700         }
1701         if (!(nd->flags & LOOKUP_RCU)) {
1702                 if (link->mnt == nd->path.mnt)
1703                         mntget(link->mnt);
1704         }
1705         error = nd_alloc_stack(nd);
1706         if (unlikely(error)) {
1707                 if (error == -ECHILD) {
1708                         if (unlikely(!legitimize_path(nd, link, seq))) {
1709                                 drop_links(nd);
1710                                 nd->depth = 0;
1711                                 nd->flags &= ~LOOKUP_RCU;
1712                                 nd->path.mnt = NULL;
1713                                 nd->path.dentry = NULL;
1714                                 if (!(nd->flags & LOOKUP_ROOT))
1715                                         nd->root.mnt = NULL;
1716                                 rcu_read_unlock();
1717                         } else if (likely(unlazy_walk(nd)) == 0)
1718                                 error = nd_alloc_stack(nd);
1719                 }
1720                 if (error) {
1721                         path_put(link);
1722                         return error;
1723                 }
1724         }
1725
1726         last = nd->stack + nd->depth++;
1727         last->link = *link;
1728         clear_delayed_call(&last->done);
1729         nd->link_inode = inode;
1730         last->seq = seq;
1731         return 1;
1732 }
1733
1734 enum {WALK_FOLLOW = 1, WALK_MORE = 2};
1735
1736 /*
1737  * Do we need to follow links? We _really_ want to be able
1738  * to do this check without having to look at inode->i_op,
1739  * so we keep a cache of "no, this doesn't need follow_link"
1740  * for the common case.
1741  */
1742 static inline int step_into(struct nameidata *nd, struct path *path,
1743                             int flags, struct inode *inode, unsigned seq)
1744 {
1745         if (!(flags & WALK_MORE) && nd->depth)
1746                 put_link(nd);
1747         if (likely(!d_is_symlink(path->dentry)) ||
1748            !(flags & WALK_FOLLOW || nd->flags & LOOKUP_FOLLOW)) {
1749                 /* not a symlink or should not follow */
1750                 path_to_nameidata(path, nd);
1751                 nd->inode = inode;
1752                 nd->seq = seq;
1753                 return 0;
1754         }
1755         /* make sure that d_is_symlink above matches inode */
1756         if (nd->flags & LOOKUP_RCU) {
1757                 if (read_seqcount_retry(&path->dentry->d_seq, seq))
1758                         return -ECHILD;
1759         }
1760         return pick_link(nd, path, inode, seq);
1761 }
1762
1763 static int walk_component(struct nameidata *nd, int flags)
1764 {
1765         struct path path;
1766         struct inode *inode;
1767         unsigned seq;
1768         int err;
1769         /*
1770          * "." and ".." are special - ".." especially so because it has
1771          * to be able to know about the current root directory and
1772          * parent relationships.
1773          */
1774         if (unlikely(nd->last_type != LAST_NORM)) {
1775                 err = handle_dots(nd, nd->last_type);
1776                 if (!(flags & WALK_MORE) && nd->depth)
1777                         put_link(nd);
1778                 return err;
1779         }
1780         err = lookup_fast(nd, &path, &inode, &seq);
1781         if (unlikely(err <= 0)) {
1782                 if (err < 0)
1783                         return err;
1784                 path.dentry = lookup_slow(&nd->last, nd->path.dentry,
1785                                           nd->flags);
1786                 if (IS_ERR(path.dentry))
1787                         return PTR_ERR(path.dentry);
1788
1789                 path.mnt = nd->path.mnt;
1790                 err = follow_managed(&path, nd);
1791                 if (unlikely(err < 0))
1792                         return err;
1793
1794                 if (unlikely(d_is_negative(path.dentry))) {
1795                         path_to_nameidata(&path, nd);
1796                         return -ENOENT;
1797                 }
1798
1799                 seq = 0;        /* we are already out of RCU mode */
1800                 inode = d_backing_inode(path.dentry);
1801         }
1802
1803         return step_into(nd, &path, flags, inode, seq);
1804 }
1805
1806 /*
1807  * We can do the critical dentry name comparison and hashing
1808  * operations one word at a time, but we are limited to:
1809  *
1810  * - Architectures with fast unaligned word accesses. We could
1811  *   do a "get_unaligned()" if this helps and is sufficiently
1812  *   fast.
1813  *
1814  * - non-CONFIG_DEBUG_PAGEALLOC configurations (so that we
1815  *   do not trap on the (extremely unlikely) case of a page
1816  *   crossing operation.
1817  *
1818  * - Furthermore, we need an efficient 64-bit compile for the
1819  *   64-bit case in order to generate the "number of bytes in
1820  *   the final mask". Again, that could be replaced with a
1821  *   efficient population count instruction or similar.
1822  */
1823 #ifdef CONFIG_DCACHE_WORD_ACCESS
1824
1825 #include <asm/word-at-a-time.h>
1826
1827 #ifdef HASH_MIX
1828
1829 /* Architecture provides HASH_MIX and fold_hash() in <asm/hash.h> */
1830
1831 #elif defined(CONFIG_64BIT)
1832 /*
1833  * Register pressure in the mixing function is an issue, particularly
1834  * on 32-bit x86, but almost any function requires one state value and
1835  * one temporary.  Instead, use a function designed for two state values
1836  * and no temporaries.
1837  *
1838  * This function cannot create a collision in only two iterations, so
1839  * we have two iterations to achieve avalanche.  In those two iterations,
1840  * we have six layers of mixing, which is enough to spread one bit's
1841  * influence out to 2^6 = 64 state bits.
1842  *
1843  * Rotate constants are scored by considering either 64 one-bit input
1844  * deltas or 64*63/2 = 2016 two-bit input deltas, and finding the
1845  * probability of that delta causing a change to each of the 128 output
1846  * bits, using a sample of random initial states.
1847  *
1848  * The Shannon entropy of the computed probabilities is then summed
1849  * to produce a score.  Ideally, any input change has a 50% chance of
1850  * toggling any given output bit.
1851  *
1852  * Mixing scores (in bits) for (12,45):
1853  * Input delta: 1-bit      2-bit
1854  * 1 round:     713.3    42542.6
1855  * 2 rounds:   2753.7   140389.8
1856  * 3 rounds:   5954.1   233458.2
1857  * 4 rounds:   7862.6   256672.2
1858  * Perfect:    8192     258048
1859  *            (64*128) (64*63/2 * 128)
1860  */
1861 #define HASH_MIX(x, y, a)       \
1862         (       x ^= (a),       \
1863         y ^= x, x = rol64(x,12),\
1864         x += y, y = rol64(y,45),\
1865         y *= 9                  )
1866
1867 /*
1868  * Fold two longs into one 32-bit hash value.  This must be fast, but
1869  * latency isn't quite as critical, as there is a fair bit of additional
1870  * work done before the hash value is used.
1871  */
1872 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1873 {
1874         y ^= x * GOLDEN_RATIO_64;
1875         y *= GOLDEN_RATIO_64;
1876         return y >> 32;
1877 }
1878
1879 #else   /* 32-bit case */
1880
1881 /*
1882  * Mixing scores (in bits) for (7,20):
1883  * Input delta: 1-bit      2-bit
1884  * 1 round:     330.3     9201.6
1885  * 2 rounds:   1246.4    25475.4
1886  * 3 rounds:   1907.1    31295.1
1887  * 4 rounds:   2042.3    31718.6
1888  * Perfect:    2048      31744
1889  *            (32*64)   (32*31/2 * 64)
1890  */
1891 #define HASH_MIX(x, y, a)       \
1892         (       x ^= (a),       \
1893         y ^= x, x = rol32(x, 7),\
1894         x += y, y = rol32(y,20),\
1895         y *= 9                  )
1896
1897 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1898 {
1899         /* Use arch-optimized multiply if one exists */
1900         return __hash_32(y ^ __hash_32(x));
1901 }
1902
1903 #endif
1904
1905 /*
1906  * Return the hash of a string of known length.  This is carfully
1907  * designed to match hash_name(), which is the more critical function.
1908  * In particular, we must end by hashing a final word containing 0..7
1909  * payload bytes, to match the way that hash_name() iterates until it
1910  * finds the delimiter after the name.
1911  */
1912 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1913 {
1914         unsigned long a, x = 0, y = (unsigned long)salt;
1915
1916         for (;;) {
1917                 if (!len)
1918                         goto done;
1919                 a = load_unaligned_zeropad(name);
1920                 if (len < sizeof(unsigned long))
1921                         break;
1922                 HASH_MIX(x, y, a);
1923                 name += sizeof(unsigned long);
1924                 len -= sizeof(unsigned long);
1925         }
1926         x ^= a & bytemask_from_count(len);
1927 done:
1928         return fold_hash(x, y);
1929 }
1930 EXPORT_SYMBOL(full_name_hash);
1931
1932 /* Return the "hash_len" (hash and length) of a null-terminated string */
1933 u64 hashlen_string(const void *salt, const char *name)
1934 {
1935         unsigned long a = 0, x = 0, y = (unsigned long)salt;
1936         unsigned long adata, mask, len;
1937         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1938
1939         len = 0;
1940         goto inside;
1941
1942         do {
1943                 HASH_MIX(x, y, a);
1944                 len += sizeof(unsigned long);
1945 inside:
1946                 a = load_unaligned_zeropad(name+len);
1947         } while (!has_zero(a, &adata, &constants));
1948
1949         adata = prep_zero_mask(a, adata, &constants);
1950         mask = create_zero_mask(adata);
1951         x ^= a & zero_bytemask(mask);
1952
1953         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1954 }
1955 EXPORT_SYMBOL(hashlen_string);
1956
1957 /*
1958  * Calculate the length and hash of the path component, and
1959  * return the "hash_len" as the result.
1960  */
1961 static inline u64 hash_name(const void *salt, const char *name)
1962 {
1963         unsigned long a = 0, b, x = 0, y = (unsigned long)salt;
1964         unsigned long adata, bdata, mask, len;
1965         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1966
1967         len = 0;
1968         goto inside;
1969
1970         do {
1971                 HASH_MIX(x, y, a);
1972                 len += sizeof(unsigned long);
1973 inside:
1974                 a = load_unaligned_zeropad(name+len);
1975                 b = a ^ REPEAT_BYTE('/');
1976         } while (!(has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)));
1977
1978         adata = prep_zero_mask(a, adata, &constants);
1979         bdata = prep_zero_mask(b, bdata, &constants);
1980         mask = create_zero_mask(adata | bdata);
1981         x ^= a & zero_bytemask(mask);
1982
1983         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1984 }
1985
1986 #else   /* !CONFIG_DCACHE_WORD_ACCESS: Slow, byte-at-a-time version */
1987
1988 /* Return the hash of a string of known length */
1989 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1990 {
1991         unsigned long hash = init_name_hash(salt);
1992         while (len--)
1993                 hash = partial_name_hash((unsigned char)*name++, hash);
1994         return end_name_hash(hash);
1995 }
1996 EXPORT_SYMBOL(full_name_hash);
1997
1998 /* Return the "hash_len" (hash and length) of a null-terminated string */
1999 u64 hashlen_string(const void *salt, const char *name)
2000 {
2001         unsigned long hash = init_name_hash(salt);
2002         unsigned long len = 0, c;
2003
2004         c = (unsigned char)*name;
2005         while (c) {
2006                 len++;
2007                 hash = partial_name_hash(c, hash);
2008                 c = (unsigned char)name[len];
2009         }
2010         return hashlen_create(end_name_hash(hash), len);
2011 }
2012 EXPORT_SYMBOL(hashlen_string);
2013
2014 /*
2015  * We know there's a real path component here of at least
2016  * one character.
2017  */
2018 static inline u64 hash_name(const void *salt, const char *name)
2019 {
2020         unsigned long hash = init_name_hash(salt);
2021         unsigned long len = 0, c;
2022
2023         c = (unsigned char)*name;
2024         do {
2025                 len++;
2026                 hash = partial_name_hash(c, hash);
2027                 c = (unsigned char)name[len];
2028         } while (c && c != '/');
2029         return hashlen_create(end_name_hash(hash), len);
2030 }
2031
2032 #endif
2033
2034 /*
2035  * Name resolution.
2036  * This is the basic name resolution function, turning a pathname into
2037  * the final dentry. We expect 'base' to be positive and a directory.
2038  *
2039  * Returns 0 and nd will have valid dentry and mnt on success.
2040  * Returns error and drops reference to input namei data on failure.
2041  */
2042 static int link_path_walk(const char *name, struct nameidata *nd)
2043 {
2044         int err;
2045
2046         while (*name=='/')
2047                 name++;
2048         if (!*name)
2049                 return 0;
2050
2051         /* At this point we know we have a real path component. */
2052         for(;;) {
2053                 u64 hash_len;
2054                 int type;
2055
2056                 err = may_lookup(nd);
2057                 if (err)
2058                         return err;
2059
2060                 hash_len = hash_name(nd->path.dentry, name);
2061
2062                 type = LAST_NORM;
2063                 if (name[0] == '.') switch (hashlen_len(hash_len)) {
2064                         case 2:
2065                                 if (name[1] == '.') {
2066                                         type = LAST_DOTDOT;
2067                                         nd->flags |= LOOKUP_JUMPED;
2068                                 }
2069                                 break;
2070                         case 1:
2071                                 type = LAST_DOT;
2072                 }
2073                 if (likely(type == LAST_NORM)) {
2074                         struct dentry *parent = nd->path.dentry;
2075                         nd->flags &= ~LOOKUP_JUMPED;
2076                         if (unlikely(parent->d_flags & DCACHE_OP_HASH)) {
2077                                 struct qstr this = { { .hash_len = hash_len }, .name = name };
2078                                 err = parent->d_op->d_hash(parent, &this);
2079                                 if (err < 0)
2080                                         return err;
2081                                 hash_len = this.hash_len;
2082                                 name = this.name;
2083                         }
2084                 }
2085
2086                 nd->last.hash_len = hash_len;
2087                 nd->last.name = name;
2088                 nd->last_type = type;
2089
2090                 name += hashlen_len(hash_len);
2091                 if (!*name)
2092                         goto OK;
2093                 /*
2094                  * If it wasn't NUL, we know it was '/'. Skip that
2095                  * slash, and continue until no more slashes.
2096                  */
2097                 do {
2098                         name++;
2099                 } while (unlikely(*name == '/'));
2100                 if (unlikely(!*name)) {
2101 OK:
2102                         /* pathname body, done */
2103                         if (!nd->depth)
2104                                 return 0;
2105                         name = nd->stack[nd->depth - 1].name;
2106                         /* trailing symlink, done */
2107                         if (!name)
2108                                 return 0;
2109                         /* last component of nested symlink */
2110                         err = walk_component(nd, WALK_FOLLOW);
2111                 } else {
2112                         /* not the last component */
2113                         err = walk_component(nd, WALK_FOLLOW | WALK_MORE);
2114                 }
2115                 if (err < 0)
2116                         return err;
2117
2118                 if (err) {
2119                         const char *s = get_link(nd);
2120
2121                         if (IS_ERR(s))
2122                                 return PTR_ERR(s);
2123                         err = 0;
2124                         if (unlikely(!s)) {
2125                                 /* jumped */
2126                                 put_link(nd);
2127                         } else {
2128                                 nd->stack[nd->depth - 1].name = name;
2129                                 name = s;
2130                                 continue;
2131                         }
2132                 }
2133                 if (unlikely(!d_can_lookup(nd->path.dentry))) {
2134                         if (nd->flags & LOOKUP_RCU) {
2135                                 if (unlazy_walk(nd))
2136                                         return -ECHILD;
2137                         }
2138                         return -ENOTDIR;
2139                 }
2140         }
2141 }
2142
2143 static const char *path_init(struct nameidata *nd, unsigned flags)
2144 {
2145         const char *s = nd->name->name;
2146
2147         if (!*s)
2148                 flags &= ~LOOKUP_RCU;
2149
2150         nd->last_type = LAST_ROOT; /* if there are only slashes... */
2151         nd->flags = flags | LOOKUP_JUMPED | LOOKUP_PARENT;
2152         nd->depth = 0;
2153         if (flags & LOOKUP_ROOT) {
2154                 struct dentry *root = nd->root.dentry;
2155                 struct inode *inode = root->d_inode;
2156                 if (*s && unlikely(!d_can_lookup(root)))
2157                         return ERR_PTR(-ENOTDIR);
2158                 nd->path = nd->root;
2159                 nd->inode = inode;
2160                 if (flags & LOOKUP_RCU) {
2161                         rcu_read_lock();
2162                         nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2163                         nd->root_seq = nd->seq;
2164                         nd->m_seq = read_seqbegin(&mount_lock);
2165                 } else {
2166                         path_get(&nd->path);
2167                 }
2168                 return s;
2169         }
2170
2171         nd->root.mnt = NULL;
2172         nd->path.mnt = NULL;
2173         nd->path.dentry = NULL;
2174
2175         nd->m_seq = read_seqbegin(&mount_lock);
2176         if (*s == '/') {
2177                 if (flags & LOOKUP_RCU)
2178                         rcu_read_lock();
2179                 set_root(nd);
2180                 if (likely(!nd_jump_root(nd)))
2181                         return s;
2182                 nd->root.mnt = NULL;
2183                 rcu_read_unlock();
2184                 return ERR_PTR(-ECHILD);
2185         } else if (nd->dfd == AT_FDCWD) {
2186                 if (flags & LOOKUP_RCU) {
2187                         struct fs_struct *fs = current->fs;
2188                         unsigned seq;
2189
2190                         rcu_read_lock();
2191
2192                         do {
2193                                 seq = read_seqcount_begin(&fs->seq);
2194                                 nd->path = fs->pwd;
2195                                 nd->inode = nd->path.dentry->d_inode;
2196                                 nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2197                         } while (read_seqcount_retry(&fs->seq, seq));
2198                 } else {
2199                         get_fs_pwd(current->fs, &nd->path);
2200                         nd->inode = nd->path.dentry->d_inode;
2201                 }
2202                 return s;
2203         } else {
2204                 /* Caller must check execute permissions on the starting path component */
2205                 struct fd f = fdget_raw(nd->dfd);
2206                 struct dentry *dentry;
2207
2208                 if (!f.file)
2209                         return ERR_PTR(-EBADF);
2210
2211                 dentry = f.file->f_path.dentry;
2212
2213                 if (*s) {
2214                         if (!d_can_lookup(dentry)) {
2215                                 fdput(f);
2216                                 return ERR_PTR(-ENOTDIR);
2217                         }
2218                 }
2219
2220                 nd->path = f.file->f_path;
2221                 if (flags & LOOKUP_RCU) {
2222                         rcu_read_lock();
2223                         nd->inode = nd->path.dentry->d_inode;
2224                         nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
2225                 } else {
2226                         path_get(&nd->path);
2227                         nd->inode = nd->path.dentry->d_inode;
2228                 }
2229                 fdput(f);
2230                 return s;
2231         }
2232 }
2233
2234 static const char *trailing_symlink(struct nameidata *nd)
2235 {
2236         const char *s;
2237         int error = may_follow_link(nd);
2238         if (unlikely(error))
2239                 return ERR_PTR(error);
2240         nd->flags |= LOOKUP_PARENT;
2241         nd->stack[0].name = NULL;
2242         s = get_link(nd);
2243         return s ? s : "";
2244 }
2245
2246 static inline int lookup_last(struct nameidata *nd)
2247 {
2248         if (nd->last_type == LAST_NORM && nd->last.name[nd->last.len])
2249                 nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
2250
2251         nd->flags &= ~LOOKUP_PARENT;
2252         return walk_component(nd, 0);
2253 }
2254
2255 static int handle_lookup_down(struct nameidata *nd)
2256 {
2257         struct path path = nd->path;
2258         struct inode *inode = nd->inode;
2259         unsigned seq = nd->seq;
2260         int err;
2261
2262         if (nd->flags & LOOKUP_RCU) {
2263                 /*
2264                  * don't bother with unlazy_walk on failure - we are
2265                  * at the very beginning of walk, so we lose nothing
2266                  * if we simply redo everything in non-RCU mode
2267                  */
2268                 if (unlikely(!__follow_mount_rcu(nd, &path, &inode, &seq)))
2269                         return -ECHILD;
2270         } else {
2271                 dget(path.dentry);
2272                 err = follow_managed(&path, nd);
2273                 if (unlikely(err < 0))
2274                         return err;
2275                 inode = d_backing_inode(path.dentry);
2276                 seq = 0;
2277         }
2278         path_to_nameidata(&path, nd);
2279         nd->inode = inode;
2280         nd->seq = seq;
2281         return 0;
2282 }
2283
2284 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2285 static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
2286 {
2287         const char *s = path_init(nd, flags);
2288         int err;
2289
2290         if (IS_ERR(s))
2291                 return PTR_ERR(s);
2292
2293         if (unlikely(flags & LOOKUP_DOWN)) {
2294                 err = handle_lookup_down(nd);
2295                 if (unlikely(err < 0)) {
2296                         terminate_walk(nd);
2297                         return err;
2298                 }
2299         }
2300
2301         while (!(err = link_path_walk(s, nd))
2302                 && ((err = lookup_last(nd)) > 0)) {
2303                 s = trailing_symlink(nd);
2304                 if (IS_ERR(s)) {
2305                         err = PTR_ERR(s);
2306                         break;
2307                 }
2308         }
2309         if (!err)
2310                 err = complete_walk(nd);
2311
2312         if (!err && nd->flags & LOOKUP_DIRECTORY)
2313                 if (!d_can_lookup(nd->path.dentry))
2314                         err = -ENOTDIR;
2315         if (!err) {
2316                 *path = nd->path;
2317                 nd->path.mnt = NULL;
2318                 nd->path.dentry = NULL;
2319         }
2320         terminate_walk(nd);
2321         return err;
2322 }
2323
2324 static int filename_lookup(int dfd, struct filename *name, unsigned flags,
2325                            struct path *path, struct path *root)
2326 {
2327         int retval;
2328         struct nameidata nd;
2329         if (IS_ERR(name))
2330                 return PTR_ERR(name);
2331         if (unlikely(root)) {
2332                 nd.root = *root;
2333                 flags |= LOOKUP_ROOT;
2334         }
2335         set_nameidata(&nd, dfd, name);
2336         retval = path_lookupat(&nd, flags | LOOKUP_RCU, path);
2337         if (unlikely(retval == -ECHILD))
2338                 retval = path_lookupat(&nd, flags, path);
2339         if (unlikely(retval == -ESTALE))
2340                 retval = path_lookupat(&nd, flags | LOOKUP_REVAL, path);
2341
2342         if (likely(!retval))
2343                 audit_inode(name, path->dentry, flags & LOOKUP_PARENT);
2344         restore_nameidata();
2345         putname(name);
2346         return retval;
2347 }
2348
2349 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2350 static int path_parentat(struct nameidata *nd, unsigned flags,
2351                                 struct path *parent)
2352 {
2353         const char *s = path_init(nd, flags);
2354         int err;
2355         if (IS_ERR(s))
2356                 return PTR_ERR(s);
2357         err = link_path_walk(s, nd);
2358         if (!err)
2359                 err = complete_walk(nd);
2360         if (!err) {
2361                 *parent = nd->path;
2362                 nd->path.mnt = NULL;
2363                 nd->path.dentry = NULL;
2364         }
2365         terminate_walk(nd);
2366         return err;
2367 }
2368
2369 static struct filename *filename_parentat(int dfd, struct filename *name,
2370                                 unsigned int flags, struct path *parent,
2371                                 struct qstr *last, int *type)
2372 {
2373         int retval;
2374         struct nameidata nd;
2375
2376         if (IS_ERR(name))
2377                 return name;
2378         set_nameidata(&nd, dfd, name);
2379         retval = path_parentat(&nd, flags | LOOKUP_RCU, parent);
2380         if (unlikely(retval == -ECHILD))
2381                 retval = path_parentat(&nd, flags, parent);
2382         if (unlikely(retval == -ESTALE))
2383                 retval = path_parentat(&nd, flags | LOOKUP_REVAL, parent);
2384         if (likely(!retval)) {
2385                 *last = nd.last;
2386                 *type = nd.last_type;
2387                 audit_inode(name, parent->dentry, LOOKUP_PARENT);
2388         } else {
2389                 putname(name);
2390                 name = ERR_PTR(retval);
2391         }
2392         restore_nameidata();
2393         return name;
2394 }
2395
2396 /* does lookup, returns the object with parent locked */
2397 struct dentry *kern_path_locked(const char *name, struct path *path)
2398 {
2399         struct filename *filename;
2400         struct dentry *d;
2401         struct qstr last;
2402         int type;
2403
2404         filename = filename_parentat(AT_FDCWD, getname_kernel(name), 0, path,
2405                                     &last, &type);
2406         if (IS_ERR(filename))
2407                 return ERR_CAST(filename);
2408         if (unlikely(type != LAST_NORM)) {
2409                 path_put(path);
2410                 putname(filename);
2411                 return ERR_PTR(-EINVAL);
2412         }
2413         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
2414         d = __lookup_hash(&last, path->dentry, 0);
2415         if (IS_ERR(d)) {
2416                 inode_unlock(path->dentry->d_inode);
2417                 path_put(path);
2418         }
2419         putname(filename);
2420         return d;
2421 }
2422
2423 int kern_path(const char *name, unsigned int flags, struct path *path)
2424 {
2425         return filename_lookup(AT_FDCWD, getname_kernel(name),
2426                                flags, path, NULL);
2427 }
2428 EXPORT_SYMBOL(kern_path);
2429
2430 /**
2431  * vfs_path_lookup - lookup a file path relative to a dentry-vfsmount pair
2432  * @dentry:  pointer to dentry of the base directory
2433  * @mnt: pointer to vfs mount of the base directory
2434  * @name: pointer to file name
2435  * @flags: lookup flags
2436  * @path: pointer to struct path to fill
2437  */
2438 int vfs_path_lookup(struct dentry *dentry, struct vfsmount *mnt,
2439                     const char *name, unsigned int flags,
2440                     struct path *path)
2441 {
2442         struct path root = {.mnt = mnt, .dentry = dentry};
2443         /* the first argument of filename_lookup() is ignored with root */
2444         return filename_lookup(AT_FDCWD, getname_kernel(name),
2445                                flags , path, &root);
2446 }
2447 EXPORT_SYMBOL(vfs_path_lookup);
2448
2449 /**
2450  * lookup_one_len - filesystem helper to lookup single pathname component
2451  * @name:       pathname component to lookup
2452  * @base:       base directory to lookup from
2453  * @len:        maximum length @len should be interpreted to
2454  *
2455  * Note that this routine is purely a helper for filesystem usage and should
2456  * not be called by generic code.
2457  *
2458  * The caller must hold base->i_mutex.
2459  */
2460 struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
2461 {
2462         struct qstr this;
2463         unsigned int c;
2464         int err;
2465
2466         WARN_ON_ONCE(!inode_is_locked(base->d_inode));
2467
2468         this.name = name;
2469         this.len = len;
2470         this.hash = full_name_hash(base, name, len);
2471         if (!len)
2472                 return ERR_PTR(-EACCES);
2473
2474         if (unlikely(name[0] == '.')) {
2475                 if (len < 2 || (len == 2 && name[1] == '.'))
2476                         return ERR_PTR(-EACCES);
2477         }
2478
2479         while (len--) {
2480                 c = *(const unsigned char *)name++;
2481                 if (c == '/' || c == '\0')
2482                         return ERR_PTR(-EACCES);
2483         }
2484         /*
2485          * See if the low-level filesystem might want
2486          * to use its own hash..
2487          */
2488         if (base->d_flags & DCACHE_OP_HASH) {
2489                 int err = base->d_op->d_hash(base, &this);
2490                 if (err < 0)
2491                         return ERR_PTR(err);
2492         }
2493
2494         err = inode_permission(base->d_inode, MAY_EXEC);
2495         if (err)
2496                 return ERR_PTR(err);
2497
2498         return __lookup_hash(&this, base, 0);
2499 }
2500 EXPORT_SYMBOL(lookup_one_len);
2501
2502 /**
2503  * lookup_one_len_unlocked - filesystem helper to lookup single pathname component
2504  * @name:       pathname component to lookup
2505  * @base:       base directory to lookup from
2506  * @len:        maximum length @len should be interpreted to
2507  *
2508  * Note that this routine is purely a helper for filesystem usage and should
2509  * not be called by generic code.
2510  *
2511  * Unlike lookup_one_len, it should be called without the parent
2512  * i_mutex held, and will take the i_mutex itself if necessary.
2513  */
2514 struct dentry *lookup_one_len_unlocked(const char *name,
2515                                        struct dentry *base, int len)
2516 {
2517         struct qstr this;
2518         unsigned int c;
2519         int err;
2520         struct dentry *ret;
2521
2522         this.name = name;
2523         this.len = len;
2524         this.hash = full_name_hash(base, name, len);
2525         if (!len)
2526                 return ERR_PTR(-EACCES);
2527
2528         if (unlikely(name[0] == '.')) {
2529                 if (len < 2 || (len == 2 && name[1] == '.'))
2530                         return ERR_PTR(-EACCES);
2531         }
2532
2533         while (len--) {
2534                 c = *(const unsigned char *)name++;
2535                 if (c == '/' || c == '\0')
2536                         return ERR_PTR(-EACCES);
2537         }
2538         /*
2539          * See if the low-level filesystem might want
2540          * to use its own hash..
2541          */
2542         if (base->d_flags & DCACHE_OP_HASH) {
2543                 int err = base->d_op->d_hash(base, &this);
2544                 if (err < 0)
2545                         return ERR_PTR(err);
2546         }
2547
2548         err = inode_permission(base->d_inode, MAY_EXEC);
2549         if (err)
2550                 return ERR_PTR(err);
2551
2552         ret = lookup_dcache(&this, base, 0);
2553         if (!ret)
2554                 ret = lookup_slow(&this, base, 0);
2555         return ret;
2556 }
2557 EXPORT_SYMBOL(lookup_one_len_unlocked);
2558
2559 #ifdef CONFIG_UNIX98_PTYS
2560 int path_pts(struct path *path)
2561 {
2562         /* Find something mounted on "pts" in the same directory as
2563          * the input path.
2564          */
2565         struct dentry *child, *parent;
2566         struct qstr this;
2567         int ret;
2568
2569         ret = path_parent_directory(path);
2570         if (ret)
2571                 return ret;
2572
2573         parent = path->dentry;
2574         this.name = "pts";
2575         this.len = 3;
2576         child = d_hash_and_lookup(parent, &this);
2577         if (!child)
2578                 return -ENOENT;
2579
2580         path->dentry = child;
2581         dput(parent);
2582         follow_mount(path);
2583         return 0;
2584 }
2585 #endif
2586
2587 int user_path_at_empty(int dfd, const char __user *name, unsigned flags,
2588                  struct path *path, int *empty)
2589 {
2590         return filename_lookup(dfd, getname_flags(name, flags, empty),
2591                                flags, path, NULL);
2592 }
2593 EXPORT_SYMBOL(user_path_at_empty);
2594
2595 /**
2596  * mountpoint_last - look up last component for umount
2597  * @nd:   pathwalk nameidata - currently pointing at parent directory of "last"
2598  *
2599  * This is a special lookup_last function just for umount. In this case, we
2600  * need to resolve the path without doing any revalidation.
2601  *
2602  * The nameidata should be the result of doing a LOOKUP_PARENT pathwalk. Since
2603  * mountpoints are always pinned in the dcache, their ancestors are too. Thus,
2604  * in almost all cases, this lookup will be served out of the dcache. The only
2605  * cases where it won't are if nd->last refers to a symlink or the path is
2606  * bogus and it doesn't exist.
2607  *
2608  * Returns:
2609  * -error: if there was an error during lookup. This includes -ENOENT if the
2610  *         lookup found a negative dentry.
2611  *
2612  * 0:      if we successfully resolved nd->last and found it to not to be a
2613  *         symlink that needs to be followed.
2614  *
2615  * 1:      if we successfully resolved nd->last and found it to be a symlink
2616  *         that needs to be followed.
2617  */
2618 static int
2619 mountpoint_last(struct nameidata *nd)
2620 {
2621         int error = 0;
2622         struct dentry *dir = nd->path.dentry;
2623         struct path path;
2624
2625         /* If we're in rcuwalk, drop out of it to handle last component */
2626         if (nd->flags & LOOKUP_RCU) {
2627                 if (unlazy_walk(nd))
2628                         return -ECHILD;
2629         }
2630
2631         nd->flags &= ~LOOKUP_PARENT;
2632
2633         if (unlikely(nd->last_type != LAST_NORM)) {
2634                 error = handle_dots(nd, nd->last_type);
2635                 if (error)
2636                         return error;
2637                 path.dentry = dget(nd->path.dentry);
2638         } else {
2639                 path.dentry = d_lookup(dir, &nd->last);
2640                 if (!path.dentry) {
2641                         /*
2642                          * No cached dentry. Mounted dentries are pinned in the
2643                          * cache, so that means that this dentry is probably
2644                          * a symlink or the path doesn't actually point
2645                          * to a mounted dentry.
2646                          */
2647                         path.dentry = lookup_slow(&nd->last, dir,
2648                                              nd->flags | LOOKUP_NO_REVAL);
2649                         if (IS_ERR(path.dentry))
2650                                 return PTR_ERR(path.dentry);
2651                 }
2652         }
2653         if (d_is_negative(path.dentry)) {
2654                 dput(path.dentry);
2655                 return -ENOENT;
2656         }
2657         path.mnt = nd->path.mnt;
2658         return step_into(nd, &path, 0, d_backing_inode(path.dentry), 0);
2659 }
2660
2661 /**
2662  * path_mountpoint - look up a path to be umounted
2663  * @nd:         lookup context
2664  * @flags:      lookup flags
2665  * @path:       pointer to container for result
2666  *
2667  * Look up the given name, but don't attempt to revalidate the last component.
2668  * Returns 0 and "path" will be valid on success; Returns error otherwise.
2669  */
2670 static int
2671 path_mountpoint(struct nameidata *nd, unsigned flags, struct path *path)
2672 {
2673         const char *s = path_init(nd, flags);
2674         int err;
2675         if (IS_ERR(s))
2676                 return PTR_ERR(s);
2677         while (!(err = link_path_walk(s, nd)) &&
2678                 (err = mountpoint_last(nd)) > 0) {
2679                 s = trailing_symlink(nd);
2680                 if (IS_ERR(s)) {
2681                         err = PTR_ERR(s);
2682                         break;
2683                 }
2684         }
2685         if (!err) {
2686                 *path = nd->path;
2687                 nd->path.mnt = NULL;
2688                 nd->path.dentry = NULL;
2689                 follow_mount(path);
2690         }
2691         terminate_walk(nd);
2692         return err;
2693 }
2694
2695 static int
2696 filename_mountpoint(int dfd, struct filename *name, struct path *path,
2697                         unsigned int flags)
2698 {
2699         struct nameidata nd;
2700         int error;
2701         if (IS_ERR(name))
2702                 return PTR_ERR(name);
2703         set_nameidata(&nd, dfd, name);
2704         error = path_mountpoint(&nd, flags | LOOKUP_RCU, path);
2705         if (unlikely(error == -ECHILD))
2706                 error = path_mountpoint(&nd, flags, path);
2707         if (unlikely(error == -ESTALE))
2708                 error = path_mountpoint(&nd, flags | LOOKUP_REVAL, path);
2709         if (likely(!error))
2710                 audit_inode(name, path->dentry, 0);
2711         restore_nameidata();
2712         putname(name);
2713         return error;
2714 }
2715
2716 /**
2717  * user_path_mountpoint_at - lookup a path from userland in order to umount it
2718  * @dfd:        directory file descriptor
2719  * @name:       pathname from userland
2720  * @flags:      lookup flags
2721  * @path:       pointer to container to hold result
2722  *
2723  * A umount is a special case for path walking. We're not actually interested
2724  * in the inode in this situation, and ESTALE errors can be a problem. We
2725  * simply want track down the dentry and vfsmount attached at the mountpoint
2726  * and avoid revalidating the last component.
2727  *
2728  * Returns 0 and populates "path" on success.
2729  */
2730 int
2731 user_path_mountpoint_at(int dfd, const char __user *name, unsigned int flags,
2732                         struct path *path)
2733 {
2734         return filename_mountpoint(dfd, getname(name), path, flags);
2735 }
2736
2737 int
2738 kern_path_mountpoint(int dfd, const char *name, struct path *path,
2739                         unsigned int flags)
2740 {
2741         return filename_mountpoint(dfd, getname_kernel(name), path, flags);
2742 }
2743 EXPORT_SYMBOL(kern_path_mountpoint);
2744
2745 int __check_sticky(struct inode *dir, struct inode *inode)
2746 {
2747         kuid_t fsuid = current_fsuid();
2748
2749         if (uid_eq(inode->i_uid, fsuid))
2750                 return 0;
2751         if (uid_eq(dir->i_uid, fsuid))
2752                 return 0;
2753         return !capable_wrt_inode_uidgid(inode, CAP_FOWNER);
2754 }
2755 EXPORT_SYMBOL(__check_sticky);
2756
2757 /*
2758  *      Check whether we can remove a link victim from directory dir, check
2759  *  whether the type of victim is right.
2760  *  1. We can't do it if dir is read-only (done in permission())
2761  *  2. We should have write and exec permissions on dir
2762  *  3. We can't remove anything from append-only dir
2763  *  4. We can't do anything with immutable dir (done in permission())
2764  *  5. If the sticky bit on dir is set we should either
2765  *      a. be owner of dir, or
2766  *      b. be owner of victim, or
2767  *      c. have CAP_FOWNER capability
2768  *  6. If the victim is append-only or immutable we can't do antyhing with
2769  *     links pointing to it.
2770  *  7. If the victim has an unknown uid or gid we can't change the inode.
2771  *  8. If we were asked to remove a directory and victim isn't one - ENOTDIR.
2772  *  9. If we were asked to remove a non-directory and victim isn't one - EISDIR.
2773  * 10. We can't remove a root or mountpoint.
2774  * 11. We don't allow removal of NFS sillyrenamed files; it's handled by
2775  *     nfs_async_unlink().
2776  */
2777 static int may_delete(struct inode *dir, struct dentry *victim, bool isdir)
2778 {
2779         struct inode *inode = d_backing_inode(victim);
2780         int error;
2781
2782         if (d_is_negative(victim))
2783                 return -ENOENT;
2784         BUG_ON(!inode);
2785
2786         BUG_ON(victim->d_parent->d_inode != dir);
2787         audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);
2788
2789         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
2790         if (error)
2791                 return error;
2792         if (IS_APPEND(dir))
2793                 return -EPERM;
2794
2795         if (check_sticky(dir, inode) || IS_APPEND(inode) ||
2796             IS_IMMUTABLE(inode) || IS_SWAPFILE(inode) || HAS_UNMAPPED_ID(inode))
2797                 return -EPERM;
2798         if (isdir) {
2799                 if (!d_is_dir(victim))
2800                         return -ENOTDIR;
2801                 if (IS_ROOT(victim))
2802                         return -EBUSY;
2803         } else if (d_is_dir(victim))
2804                 return -EISDIR;
2805         if (IS_DEADDIR(dir))
2806                 return -ENOENT;
2807         if (victim->d_flags & DCACHE_NFSFS_RENAMED)
2808                 return -EBUSY;
2809         return 0;
2810 }
2811
2812 /*      Check whether we can create an object with dentry child in directory
2813  *  dir.
2814  *  1. We can't do it if child already exists (open has special treatment for
2815  *     this case, but since we are inlined it's OK)
2816  *  2. We can't do it if dir is read-only (done in permission())
2817  *  3. We can't do it if the fs can't represent the fsuid or fsgid.
2818  *  4. We should have write and exec permissions on dir
2819  *  5. We can't do it if dir is immutable (done in permission())
2820  */
2821 static inline int may_create(struct inode *dir, struct dentry *child)
2822 {
2823         struct user_namespace *s_user_ns;
2824         audit_inode_child(dir, child, AUDIT_TYPE_CHILD_CREATE);
2825         if (child->d_inode)
2826                 return -EEXIST;
2827         if (IS_DEADDIR(dir))
2828                 return -ENOENT;
2829         s_user_ns = dir->i_sb->s_user_ns;
2830         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
2831             !kgid_has_mapping(s_user_ns, current_fsgid()))
2832                 return -EOVERFLOW;
2833         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
2834 }
2835
2836 /*
2837  * p1 and p2 should be directories on the same fs.
2838  */
2839 struct dentry *lock_rename(struct dentry *p1, struct dentry *p2)
2840 {
2841         struct dentry *p;
2842
2843         if (p1 == p2) {
2844                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2845                 return NULL;
2846         }
2847
2848         mutex_lock(&p1->d_sb->s_vfs_rename_mutex);
2849
2850         p = d_ancestor(p2, p1);
2851         if (p) {
2852                 inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
2853                 inode_lock_nested(p1->d_inode, I_MUTEX_CHILD);
2854                 return p;
2855         }
2856
2857         p = d_ancestor(p1, p2);
2858         if (p) {
2859                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2860                 inode_lock_nested(p2->d_inode, I_MUTEX_CHILD);
2861                 return p;
2862         }
2863
2864         inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2865         inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
2866         return NULL;
2867 }
2868 EXPORT_SYMBOL(lock_rename);
2869
2870 void unlock_rename(struct dentry *p1, struct dentry *p2)
2871 {
2872         inode_unlock(p1->d_inode);
2873         if (p1 != p2) {
2874                 inode_unlock(p2->d_inode);
2875                 mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
2876         }
2877 }
2878 EXPORT_SYMBOL(unlock_rename);
2879
2880 int vfs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2881                 bool want_excl)
2882 {
2883         int error = may_create(dir, dentry);
2884         if (error)
2885                 return error;
2886
2887         if (!dir->i_op->create)
2888                 return -EACCES; /* shouldn't it be ENOSYS? */
2889         mode &= S_IALLUGO;
2890         mode |= S_IFREG;
2891         error = security_inode_create(dir, dentry, mode);
2892         if (error)
2893                 return error;
2894         error = dir->i_op->create(dir, dentry, mode, want_excl);
2895         if (!error)
2896                 fsnotify_create(dir, dentry);
2897         return error;
2898 }
2899 EXPORT_SYMBOL(vfs_create);
2900
2901 bool may_open_dev(const struct path *path)
2902 {
2903         return !(path->mnt->mnt_flags & MNT_NODEV) &&
2904                 !(path->mnt->mnt_sb->s_iflags & SB_I_NODEV);
2905 }
2906
2907 static int may_open(const struct path *path, int acc_mode, int flag)
2908 {
2909         struct dentry *dentry = path->dentry;
2910         struct inode *inode = dentry->d_inode;
2911         int error;
2912
2913         if (!inode)
2914                 return -ENOENT;
2915
2916         switch (inode->i_mode & S_IFMT) {
2917         case S_IFLNK:
2918                 return -ELOOP;
2919         case S_IFDIR:
2920                 if (acc_mode & MAY_WRITE)
2921                         return -EISDIR;
2922                 break;
2923         case S_IFBLK:
2924         case S_IFCHR:
2925                 if (!may_open_dev(path))
2926                         return -EACCES;
2927                 /*FALLTHRU*/
2928         case S_IFIFO:
2929         case S_IFSOCK:
2930                 flag &= ~O_TRUNC;
2931                 break;
2932         }
2933
2934         error = inode_permission(inode, MAY_OPEN | acc_mode);
2935         if (error)
2936                 return error;
2937
2938         /*
2939          * An append-only file must be opened in append mode for writing.
2940          */
2941         if (IS_APPEND(inode)) {
2942                 if  ((flag & O_ACCMODE) != O_RDONLY && !(flag & O_APPEND))
2943                         return -EPERM;
2944                 if (flag & O_TRUNC)
2945                         return -EPERM;
2946         }
2947
2948         /* O_NOATIME can only be set by the owner or superuser */
2949         if (flag & O_NOATIME && !inode_owner_or_capable(inode))
2950                 return -EPERM;
2951
2952         return 0;
2953 }
2954
2955 static int handle_truncate(struct file *filp)
2956 {
2957         const struct path *path = &filp->f_path;
2958         struct inode *inode = path->dentry->d_inode;
2959         int error = get_write_access(inode);
2960         if (error)
2961                 return error;
2962         /*
2963          * Refuse to truncate files with mandatory locks held on them.
2964          */
2965         error = locks_verify_locked(filp);
2966         if (!error)
2967                 error = security_path_truncate(path);
2968         if (!error) {
2969                 error = do_truncate(path->dentry, 0,
2970                                     ATTR_MTIME|ATTR_CTIME|ATTR_OPEN,
2971                                     filp);
2972         }
2973         put_write_access(inode);
2974         return error;
2975 }
2976
2977 static inline int open_to_namei_flags(int flag)
2978 {
2979         if ((flag & O_ACCMODE) == 3)
2980                 flag--;
2981         return flag;
2982 }
2983
2984 static int may_o_create(const struct path *dir, struct dentry *dentry, umode_t mode)
2985 {
2986         struct user_namespace *s_user_ns;
2987         int error = security_path_mknod(dir, dentry, mode, 0);
2988         if (error)
2989                 return error;
2990
2991         s_user_ns = dir->dentry->d_sb->s_user_ns;
2992         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
2993             !kgid_has_mapping(s_user_ns, current_fsgid()))
2994                 return -EOVERFLOW;
2995
2996         error = inode_permission(dir->dentry->d_inode, MAY_WRITE | MAY_EXEC);
2997         if (error)
2998                 return error;
2999
3000         return security_inode_create(dir->dentry->d_inode, dentry, mode);
3001 }
3002
3003 /*
3004  * Attempt to atomically look up, create and open a file from a negative
3005  * dentry.
3006  *
3007  * Returns 0 if successful.  The file will have been created and attached to
3008  * @file by the filesystem calling finish_open().
3009  *
3010  * Returns 1 if the file was looked up only or didn't need creating.  The
3011  * caller will need to perform the open themselves.  @path will have been
3012  * updated to point to the new dentry.  This may be negative.
3013  *
3014  * Returns an error code otherwise.
3015  */
3016 static int atomic_open(struct nameidata *nd, struct dentry *dentry,
3017                         struct path *path, struct file *file,
3018                         const struct open_flags *op,
3019                         int open_flag, umode_t mode,
3020                         int *opened)
3021 {
3022         struct dentry *const DENTRY_NOT_SET = (void *) -1UL;
3023         struct inode *dir =  nd->path.dentry->d_inode;
3024         int error;
3025
3026         if (!(~open_flag & (O_EXCL | O_CREAT))) /* both O_EXCL and O_CREAT */
3027                 open_flag &= ~O_TRUNC;
3028
3029         if (nd->flags & LOOKUP_DIRECTORY)
3030                 open_flag |= O_DIRECTORY;
3031
3032         file->f_path.dentry = DENTRY_NOT_SET;
3033         file->f_path.mnt = nd->path.mnt;
3034         error = dir->i_op->atomic_open(dir, dentry, file,
3035                                        open_to_namei_flags(open_flag),
3036                                        mode, opened);
3037         d_lookup_done(dentry);
3038         if (!error) {
3039                 /*
3040                  * We didn't have the inode before the open, so check open
3041                  * permission here.
3042                  */
3043                 int acc_mode = op->acc_mode;
3044                 if (*opened & FILE_CREATED) {
3045                         WARN_ON(!(open_flag & O_CREAT));
3046                         fsnotify_create(dir, dentry);
3047                         acc_mode = 0;
3048                 }
3049                 error = may_open(&file->f_path, acc_mode, open_flag);
3050                 if (WARN_ON(error > 0))
3051                         error = -EINVAL;
3052         } else if (error > 0) {
3053                 if (WARN_ON(file->f_path.dentry == DENTRY_NOT_SET)) {
3054                         error = -EIO;
3055                 } else {
3056                         if (file->f_path.dentry) {
3057                                 dput(dentry);
3058                                 dentry = file->f_path.dentry;
3059                         }
3060                         if (*opened & FILE_CREATED)
3061                                 fsnotify_create(dir, dentry);
3062                         if (unlikely(d_is_negative(dentry))) {
3063                                 error = -ENOENT;
3064                         } else {
3065                                 path->dentry = dentry;
3066                                 path->mnt = nd->path.mnt;
3067                                 return 1;
3068                         }
3069                 }
3070         }
3071         dput(dentry);
3072         return error;
3073 }
3074
3075 /*
3076  * Look up and maybe create and open the last component.
3077  *
3078  * Must be called with i_mutex held on parent.
3079  *
3080  * Returns 0 if the file was successfully atomically created (if necessary) and
3081  * opened.  In this case the file will be returned attached to @file.
3082  *
3083  * Returns 1 if the file was not completely opened at this time, though lookups
3084  * and creations will have been performed and the dentry returned in @path will
3085  * be positive upon return if O_CREAT was specified.  If O_CREAT wasn't
3086  * specified then a negative dentry may be returned.
3087  *
3088  * An error code is returned otherwise.
3089  *
3090  * FILE_CREATE will be set in @*opened if the dentry was created and will be
3091  * cleared otherwise prior to returning.
3092  */
3093 static int lookup_open(struct nameidata *nd, struct path *path,
3094                         struct file *file,
3095                         const struct open_flags *op,
3096                         bool got_write, int *opened)
3097 {
3098         struct dentry *dir = nd->path.dentry;
3099         struct inode *dir_inode = dir->d_inode;
3100         int open_flag = op->open_flag;
3101         struct dentry *dentry;
3102         int error, create_error = 0;
3103         umode_t mode = op->mode;
3104         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
3105
3106         if (unlikely(IS_DEADDIR(dir_inode)))
3107                 return -ENOENT;
3108
3109         *opened &= ~FILE_CREATED;
3110         dentry = d_lookup(dir, &nd->last);
3111         for (;;) {
3112                 if (!dentry) {
3113                         dentry = d_alloc_parallel(dir, &nd->last, &wq);
3114                         if (IS_ERR(dentry))
3115                                 return PTR_ERR(dentry);
3116                 }
3117                 if (d_in_lookup(dentry))
3118                         break;
3119
3120                 error = d_revalidate(dentry, nd->flags);
3121                 if (likely(error > 0))
3122                         break;
3123                 if (error)
3124                         goto out_dput;
3125                 d_invalidate(dentry);
3126                 dput(dentry);
3127                 dentry = NULL;
3128         }
3129         if (dentry->d_inode) {
3130                 /* Cached positive dentry: will open in f_op->open */
3131                 goto out_no_open;
3132         }
3133
3134         /*
3135          * Checking write permission is tricky, bacuse we don't know if we are
3136          * going to actually need it: O_CREAT opens should work as long as the
3137          * file exists.  But checking existence breaks atomicity.  The trick is
3138          * to check access and if not granted clear O_CREAT from the flags.
3139          *
3140          * Another problem is returing the "right" error value (e.g. for an
3141          * O_EXCL open we want to return EEXIST not EROFS).
3142          */
3143         if (open_flag & O_CREAT) {
3144                 if (!IS_POSIXACL(dir->d_inode))
3145                         mode &= ~current_umask();
3146                 if (unlikely(!got_write)) {
3147                         create_error = -EROFS;
3148                         open_flag &= ~O_CREAT;
3149                         if (open_flag & (O_EXCL | O_TRUNC))
3150                                 goto no_open;
3151                         /* No side effects, safe to clear O_CREAT */
3152                 } else {
3153                         create_error = may_o_create(&nd->path, dentry, mode);
3154                         if (create_error) {
3155                                 open_flag &= ~O_CREAT;
3156                                 if (open_flag & O_EXCL)
3157                                         goto no_open;
3158                         }
3159                 }
3160         } else if ((open_flag & (O_TRUNC|O_WRONLY|O_RDWR)) &&
3161                    unlikely(!got_write)) {
3162                 /*
3163                  * No O_CREATE -> atomicity not a requirement -> fall
3164                  * back to lookup + open
3165                  */
3166                 goto no_open;
3167         }
3168
3169         if (dir_inode->i_op->atomic_open) {
3170                 error = atomic_open(nd, dentry, path, file, op, open_flag,
3171                                     mode, opened);
3172                 if (unlikely(error == -ENOENT) && create_error)
3173                         error = create_error;
3174                 return error;
3175         }
3176
3177 no_open:
3178         if (d_in_lookup(dentry)) {
3179                 struct dentry *res = dir_inode->i_op->lookup(dir_inode, dentry,
3180                                                              nd->flags);
3181                 d_lookup_done(dentry);
3182                 if (unlikely(res)) {
3183                         if (IS_ERR(res)) {
3184                                 error = PTR_ERR(res);
3185                                 goto out_dput;
3186                         }
3187                         dput(dentry);
3188                         dentry = res;
3189                 }
3190         }
3191
3192         /* Negative dentry, just create the file */
3193         if (!dentry->d_inode && (open_flag & O_CREAT)) {
3194                 *opened |= FILE_CREATED;
3195                 audit_inode_child(dir_inode, dentry, AUDIT_TYPE_CHILD_CREATE);
3196                 if (!dir_inode->i_op->create) {
3197                         error = -EACCES;
3198                         goto out_dput;
3199                 }
3200                 error = dir_inode->i_op->create(dir_inode, dentry, mode,
3201                                                 open_flag & O_EXCL);
3202                 if (error)
3203                         goto out_dput;
3204                 fsnotify_create(dir_inode, dentry);
3205         }
3206         if (unlikely(create_error) && !dentry->d_inode) {
3207                 error = create_error;
3208                 goto out_dput;
3209         }
3210 out_no_open:
3211         path->dentry = dentry;
3212         path->mnt = nd->path.mnt;
3213         return 1;
3214
3215 out_dput:
3216         dput(dentry);
3217         return error;
3218 }
3219
3220 /*
3221  * Handle the last step of open()
3222  */
3223 static int do_last(struct nameidata *nd,
3224                    struct file *file, const struct open_flags *op,
3225                    int *opened)
3226 {
3227         struct dentry *dir = nd->path.dentry;
3228         int open_flag = op->open_flag;
3229         bool will_truncate = (open_flag & O_TRUNC) != 0;
3230         bool got_write = false;
3231         int acc_mode = op->acc_mode;
3232         unsigned seq;
3233         struct inode *inode;
3234         struct path path;
3235         int error;
3236
3237         nd->flags &= ~LOOKUP_PARENT;
3238         nd->flags |= op->intent;
3239
3240         if (nd->last_type != LAST_NORM) {
3241                 error = handle_dots(nd, nd->last_type);
3242                 if (unlikely(error))
3243                         return error;
3244                 goto finish_open;
3245         }
3246
3247         if (!(open_flag & O_CREAT)) {
3248                 if (nd->last.name[nd->last.len])
3249                         nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
3250                 /* we _can_ be in RCU mode here */
3251                 error = lookup_fast(nd, &path, &inode, &seq);
3252                 if (likely(error > 0))
3253                         goto finish_lookup;
3254
3255                 if (error < 0)
3256                         return error;
3257
3258                 BUG_ON(nd->inode != dir->d_inode);
3259                 BUG_ON(nd->flags & LOOKUP_RCU);
3260         } else {
3261                 /* create side of things */
3262                 /*
3263                  * This will *only* deal with leaving RCU mode - LOOKUP_JUMPED
3264                  * has been cleared when we got to the last component we are
3265                  * about to look up
3266                  */
3267                 error = complete_walk(nd);
3268                 if (error)
3269                         return error;
3270
3271                 audit_inode(nd->name, dir, LOOKUP_PARENT);
3272                 /* trailing slashes? */
3273                 if (unlikely(nd->last.name[nd->last.len]))
3274                         return -EISDIR;
3275         }
3276
3277         if (open_flag & (O_CREAT | O_TRUNC | O_WRONLY | O_RDWR)) {
3278                 error = mnt_want_write(nd->path.mnt);
3279                 if (!error)
3280                         got_write = true;
3281                 /*
3282                  * do _not_ fail yet - we might not need that or fail with
3283                  * a different error; let lookup_open() decide; we'll be
3284                  * dropping this one anyway.
3285                  */
3286         }
3287         if (open_flag & O_CREAT)
3288                 inode_lock(dir->d_inode);
3289         else
3290                 inode_lock_shared(dir->d_inode);
3291         error = lookup_open(nd, &path, file, op, got_write, opened);
3292         if (open_flag & O_CREAT)
3293                 inode_unlock(dir->d_inode);
3294         else
3295                 inode_unlock_shared(dir->d_inode);
3296
3297         if (error <= 0) {
3298                 if (error)
3299                         goto out;
3300
3301                 if ((*opened & FILE_CREATED) ||
3302                     !S_ISREG(file_inode(file)->i_mode))
3303                         will_truncate = false;
3304
3305                 audit_inode(nd->name, file->f_path.dentry, 0);
3306                 goto opened;
3307         }
3308
3309         if (*opened & FILE_CREATED) {
3310                 /* Don't check for write permission, don't truncate */
3311                 open_flag &= ~O_TRUNC;
3312                 will_truncate = false;
3313                 acc_mode = 0;
3314                 path_to_nameidata(&path, nd);
3315                 goto finish_open_created;
3316         }
3317
3318         /*
3319          * If atomic_open() acquired write access it is dropped now due to
3320          * possible mount and symlink following (this might be optimized away if
3321          * necessary...)
3322          */
3323         if (got_write) {
3324                 mnt_drop_write(nd->path.mnt);
3325                 got_write = false;
3326         }
3327
3328         error = follow_managed(&path, nd);
3329         if (unlikely(error < 0))
3330                 return error;
3331
3332         if (unlikely(d_is_negative(path.dentry))) {
3333                 path_to_nameidata(&path, nd);
3334                 return -ENOENT;
3335         }
3336
3337         /*
3338          * create/update audit record if it already exists.
3339          */
3340         audit_inode(nd->name, path.dentry, 0);
3341
3342         if (unlikely((open_flag & (O_EXCL | O_CREAT)) == (O_EXCL | O_CREAT))) {
3343                 path_to_nameidata(&path, nd);
3344                 return -EEXIST;
3345         }
3346
3347         seq = 0;        /* out of RCU mode, so the value doesn't matter */
3348         inode = d_backing_inode(path.dentry);
3349 finish_lookup:
3350         error = step_into(nd, &path, 0, inode, seq);
3351         if (unlikely(error))
3352                 return error;
3353 finish_open:
3354         /* Why this, you ask?  _Now_ we might have grown LOOKUP_JUMPED... */
3355         error = complete_walk(nd);
3356         if (error)
3357                 return error;
3358         audit_inode(nd->name, nd->path.dentry, 0);
3359         error = -EISDIR;
3360         if ((open_flag & O_CREAT) && d_is_dir(nd->path.dentry))
3361                 goto out;
3362         error = -ENOTDIR;
3363         if ((nd->flags & LOOKUP_DIRECTORY) && !d_can_lookup(nd->path.dentry))
3364                 goto out;
3365         if (!d_is_reg(nd->path.dentry))
3366                 will_truncate = false;
3367
3368         if (will_truncate) {
3369                 error = mnt_want_write(nd->path.mnt);
3370                 if (error)
3371                         goto out;
3372                 got_write = true;
3373         }
3374 finish_open_created:
3375         error = may_open(&nd->path, acc_mode, open_flag);
3376         if (error)
3377                 goto out;
3378         BUG_ON(*opened & FILE_OPENED); /* once it's opened, it's opened */
3379         error = vfs_open(&nd->path, file, current_cred());
3380         if (error)
3381                 goto out;
3382         *opened |= FILE_OPENED;
3383 opened:
3384         error = open_check_o_direct(file);
3385         if (!error)
3386                 error = ima_file_check(file, op->acc_mode, *opened);
3387         if (!error && will_truncate)
3388                 error = handle_truncate(file);
3389 out:
3390         if (unlikely(error) && (*opened & FILE_OPENED))
3391                 fput(file);
3392         if (unlikely(error > 0)) {
3393                 WARN_ON(1);
3394                 error = -EINVAL;
3395         }
3396         if (got_write)
3397                 mnt_drop_write(nd->path.mnt);
3398         return error;
3399 }
3400
3401 struct dentry *vfs_tmpfile(struct dentry *dentry, umode_t mode, int open_flag)
3402 {
3403         struct dentry *child = NULL;
3404         struct inode *dir = dentry->d_inode;
3405         struct inode *inode;
3406         int error;
3407
3408         /* we want directory to be writable */
3409         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
3410         if (error)
3411                 goto out_err;
3412         error = -EOPNOTSUPP;
3413         if (!dir->i_op->tmpfile)
3414                 goto out_err;
3415         error = -ENOMEM;
3416         child = d_alloc(dentry, &slash_name);
3417         if (unlikely(!child))
3418                 goto out_err;
3419         error = dir->i_op->tmpfile(dir, child, mode);
3420         if (error)
3421                 goto out_err;
3422         error = -ENOENT;
3423         inode = child->d_inode;
3424         if (unlikely(!inode))
3425                 goto out_err;
3426         if (!(open_flag & O_EXCL)) {
3427                 spin_lock(&inode->i_lock);
3428                 inode->i_state |= I_LINKABLE;
3429                 spin_unlock(&inode->i_lock);
3430         }
3431         return child;
3432
3433 out_err:
3434         dput(child);
3435         return ERR_PTR(error);
3436 }
3437 EXPORT_SYMBOL(vfs_tmpfile);
3438
3439 static int do_tmpfile(struct nameidata *nd, unsigned flags,
3440                 const struct open_flags *op,
3441                 struct file *file, int *opened)
3442 {
3443         struct dentry *child;
3444         struct path path;
3445         int error = path_lookupat(nd, flags | LOOKUP_DIRECTORY, &path);
3446         if (unlikely(error))
3447                 return error;
3448         error = mnt_want_write(path.mnt);
3449         if (unlikely(error))
3450                 goto out;
3451         child = vfs_tmpfile(path.dentry, op->mode, op->open_flag);
3452         error = PTR_ERR(child);
3453         if (unlikely(IS_ERR(child)))
3454                 goto out2;
3455         dput(path.dentry);
3456         path.dentry = child;
3457         audit_inode(nd->name, child, 0);
3458         /* Don't check for other permissions, the inode was just created */
3459         error = may_open(&path, 0, op->open_flag);
3460         if (error)
3461                 goto out2;
3462         file->f_path.mnt = path.mnt;
3463         error = finish_open(file, child, NULL, opened);
3464         if (error)
3465                 goto out2;
3466         error = open_check_o_direct(file);
3467         if (error)
3468                 fput(file);
3469 out2:
3470         mnt_drop_write(path.mnt);
3471 out:
3472         path_put(&path);
3473         return error;
3474 }
3475
3476 static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
3477 {
3478         struct path path;
3479         int error = path_lookupat(nd, flags, &path);
3480         if (!error) {
3481                 audit_inode(nd->name, path.dentry, 0);
3482                 error = vfs_open(&path, file, current_cred());
3483                 path_put(&path);
3484         }
3485         return error;
3486 }
3487
3488 static struct file *path_openat(struct nameidata *nd,
3489                         const struct open_flags *op, unsigned flags)
3490 {
3491         const char *s;
3492         struct file *file;
3493         int opened = 0;
3494         int error;
3495
3496         file = get_empty_filp();
3497         if (IS_ERR(file))
3498                 return file;
3499
3500         file->f_flags = op->open_flag;
3501
3502         if (unlikely(file->f_flags & __O_TMPFILE)) {
3503                 error = do_tmpfile(nd, flags, op, file, &opened);
3504                 goto out2;
3505         }
3506
3507         if (unlikely(file->f_flags & O_PATH)) {
3508                 error = do_o_path(nd, flags, file);
3509                 if (!error)
3510                         opened |= FILE_OPENED;
3511                 goto out2;
3512         }
3513
3514         s = path_init(nd, flags);
3515         if (IS_ERR(s)) {
3516                 put_filp(file);
3517                 return ERR_CAST(s);
3518         }
3519         while (!(error = link_path_walk(s, nd)) &&
3520                 (error = do_last(nd, file, op, &opened)) > 0) {
3521                 nd->flags &= ~(LOOKUP_OPEN|LOOKUP_CREATE|LOOKUP_EXCL);
3522                 s = trailing_symlink(nd);
3523                 if (IS_ERR(s)) {
3524                         error = PTR_ERR(s);
3525                         break;
3526                 }
3527         }
3528         terminate_walk(nd);
3529 out2:
3530         if (!(opened & FILE_OPENED)) {
3531                 BUG_ON(!error);
3532                 put_filp(file);
3533         }
3534         if (unlikely(error)) {
3535                 if (error == -EOPENSTALE) {
3536                         if (flags & LOOKUP_RCU)
3537                                 error = -ECHILD;
3538                         else
3539                                 error = -ESTALE;
3540                 }
3541                 file = ERR_PTR(error);
3542         }
3543         return file;
3544 }
3545
3546 struct file *do_filp_open(int dfd, struct filename *pathname,
3547                 const struct open_flags *op)
3548 {
3549         struct nameidata nd;
3550         int flags = op->lookup_flags;
3551         struct file *filp;
3552
3553         set_nameidata(&nd, dfd, pathname);
3554         filp = path_openat(&nd, op, flags | LOOKUP_RCU);
3555         if (unlikely(filp == ERR_PTR(-ECHILD)))
3556                 filp = path_openat(&nd, op, flags);
3557         if (unlikely(filp == ERR_PTR(-ESTALE)))
3558                 filp = path_openat(&nd, op, flags | LOOKUP_REVAL);
3559         restore_nameidata();
3560         return filp;
3561 }
3562
3563 struct file *do_file_open_root(struct dentry *dentry, struct vfsmount *mnt,
3564                 const char *name, const struct open_flags *op)
3565 {
3566         struct nameidata nd;
3567         struct file *file;
3568         struct filename *filename;
3569         int flags = op->lookup_flags | LOOKUP_ROOT;
3570
3571         nd.root.mnt = mnt;
3572         nd.root.dentry = dentry;
3573
3574         if (d_is_symlink(dentry) && op->intent & LOOKUP_OPEN)
3575                 return ERR_PTR(-ELOOP);
3576
3577         filename = getname_kernel(name);
3578         if (IS_ERR(filename))
3579                 return ERR_CAST(filename);
3580
3581         set_nameidata(&nd, -1, filename);
3582         file = path_openat(&nd, op, flags | LOOKUP_RCU);
3583         if (unlikely(file == ERR_PTR(-ECHILD)))
3584                 file = path_openat(&nd, op, flags);
3585         if (unlikely(file == ERR_PTR(-ESTALE)))
3586                 file = path_openat(&nd, op, flags | LOOKUP_REVAL);
3587         restore_nameidata();
3588         putname(filename);
3589         return file;
3590 }
3591
3592 static struct dentry *filename_create(int dfd, struct filename *name,
3593                                 struct path *path, unsigned int lookup_flags)
3594 {
3595         struct dentry *dentry = ERR_PTR(-EEXIST);
3596         struct qstr last;
3597         int type;
3598         int err2;
3599         int error;
3600         bool is_dir = (lookup_flags & LOOKUP_DIRECTORY);
3601
3602         /*
3603          * Note that only LOOKUP_REVAL and LOOKUP_DIRECTORY matter here. Any
3604          * other flags passed in are ignored!
3605          */
3606         lookup_flags &= LOOKUP_REVAL;
3607
3608         name = filename_parentat(dfd, name, lookup_flags, path, &last, &type);
3609         if (IS_ERR(name))
3610                 return ERR_CAST(name);
3611
3612         /*
3613          * Yucky last component or no last component at all?
3614          * (foo/., foo/.., /////)
3615          */
3616         if (unlikely(type != LAST_NORM))
3617                 goto out;
3618
3619         /* don't fail immediately if it's r/o, at least try to report other errors */
3620         err2 = mnt_want_write(path->mnt);
3621         /*
3622          * Do the final lookup.
3623          */
3624         lookup_flags |= LOOKUP_CREATE | LOOKUP_EXCL;
3625         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
3626         dentry = __lookup_hash(&last, path->dentry, lookup_flags);
3627         if (IS_ERR(dentry))
3628                 goto unlock;
3629
3630         error = -EEXIST;
3631         if (d_is_positive(dentry))
3632                 goto fail;
3633
3634         /*
3635          * Special case - lookup gave negative, but... we had foo/bar/
3636          * From the vfs_mknod() POV we just have a negative dentry -
3637          * all is fine. Let's be bastards - you had / on the end, you've
3638          * been asking for (non-existent) directory. -ENOENT for you.
3639          */
3640         if (unlikely(!is_dir && last.name[last.len])) {
3641                 error = -ENOENT;
3642                 goto fail;
3643         }
3644         if (unlikely(err2)) {
3645                 error = err2;
3646                 goto fail;
3647         }
3648         putname(name);
3649         return dentry;
3650 fail:
3651         dput(dentry);
3652         dentry = ERR_PTR(error);
3653 unlock:
3654         inode_unlock(path->dentry->d_inode);
3655         if (!err2)
3656                 mnt_drop_write(path->mnt);
3657 out:
3658         path_put(path);
3659         putname(name);
3660         return dentry;
3661 }
3662
3663 struct dentry *kern_path_create(int dfd, const char *pathname,
3664                                 struct path *path, unsigned int lookup_flags)
3665 {
3666         return filename_create(dfd, getname_kernel(pathname),
3667                                 path, lookup_flags);
3668 }
3669 EXPORT_SYMBOL(kern_path_create);
3670
3671 void done_path_create(struct path *path, struct dentry *dentry)
3672 {
3673         dput(dentry);
3674         inode_unlock(path->dentry->d_inode);
3675         mnt_drop_write(path->mnt);
3676         path_put(path);
3677 }
3678 EXPORT_SYMBOL(done_path_create);
3679
3680 inline struct dentry *user_path_create(int dfd, const char __user *pathname,
3681                                 struct path *path, unsigned int lookup_flags)
3682 {
3683         return filename_create(dfd, getname(pathname), path, lookup_flags);
3684 }
3685 EXPORT_SYMBOL(user_path_create);
3686
3687 int vfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3688 {
3689         int error = may_create(dir, dentry);
3690
3691         if (error)
3692                 return error;
3693
3694         if ((S_ISCHR(mode) || S_ISBLK(mode)) && !capable(CAP_MKNOD))
3695                 return -EPERM;
3696
3697         if (!dir->i_op->mknod)
3698                 return -EPERM;
3699
3700         error = devcgroup_inode_mknod(mode, dev);
3701         if (error)
3702                 return error;
3703
3704         error = security_inode_mknod(dir, dentry, mode, dev);
3705         if (error)
3706                 return error;
3707
3708         error = dir->i_op->mknod(dir, dentry, mode, dev);
3709         if (!error)
3710                 fsnotify_create(dir, dentry);
3711         return error;
3712 }
3713 EXPORT_SYMBOL(vfs_mknod);
3714
3715 static int may_mknod(umode_t mode)
3716 {
3717         switch (mode & S_IFMT) {
3718         case S_IFREG:
3719         case S_IFCHR:
3720         case S_IFBLK:
3721         case S_IFIFO:
3722         case S_IFSOCK:
3723         case 0: /* zero mode translates to S_IFREG */
3724                 return 0;
3725         case S_IFDIR:
3726                 return -EPERM;
3727         default:
3728                 return -EINVAL;
3729         }
3730 }
3731
3732 SYSCALL_DEFINE4(mknodat, int, dfd, const char __user *, filename, umode_t, mode,
3733                 unsigned, dev)
3734 {
3735         struct dentry *dentry;
3736         struct path path;
3737         int error;
3738         unsigned int lookup_flags = 0;
3739
3740         error = may_mknod(mode);
3741         if (error)
3742                 return error;
3743 retry:
3744         dentry = user_path_create(dfd, filename, &path, lookup_flags);
3745         if (IS_ERR(dentry))
3746                 return PTR_ERR(dentry);
3747
3748         if (!IS_POSIXACL(path.dentry->d_inode))
3749                 mode &= ~current_umask();
3750         error = security_path_mknod(&path, dentry, mode, dev);
3751         if (error)
3752                 goto out;
3753         switch (mode & S_IFMT) {
3754                 case 0: case S_IFREG:
3755                         error = vfs_create(path.dentry->d_inode,dentry,mode,true);
3756                         if (!error)
3757                                 ima_post_path_mknod(dentry);
3758                         break;
3759                 case S_IFCHR: case S_IFBLK:
3760                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,
3761                                         new_decode_dev(dev));
3762                         break;
3763                 case S_IFIFO: case S_IFSOCK:
3764                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,0);
3765                         break;
3766         }
3767 out:
3768         done_path_create(&path, dentry);
3769         if (retry_estale(error, lookup_flags)) {
3770                 lookup_flags |= LOOKUP_REVAL;
3771                 goto retry;
3772         }
3773         return error;
3774 }
3775
3776 SYSCALL_DEFINE3(mknod, const char __user *, filename, umode_t, mode, unsigned, dev)
3777 {
3778         return sys_mknodat(AT_FDCWD, filename, mode, dev);
3779 }
3780
3781 int vfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3782 {
3783         int error = may_create(dir, dentry);
3784         unsigned max_links = dir->i_sb->s_max_links;
3785
3786         if (error)
3787                 return error;
3788
3789         if (!dir->i_op->mkdir)
3790                 return -EPERM;
3791
3792         mode &= (S_IRWXUGO|S_ISVTX);
3793         error = security_inode_mkdir(dir, dentry, mode);
3794         if (error)
3795                 return error;
3796
3797         if (max_links && dir->i_nlink >= max_links)
3798                 return -EMLINK;
3799
3800         error = dir->i_op->mkdir(dir, dentry, mode);
3801         if (!error)
3802                 fsnotify_mkdir(dir, dentry);
3803         return error;
3804 }
3805 EXPORT_SYMBOL(vfs_mkdir);
3806
3807 SYSCALL_DEFINE3(mkdirat, int, dfd, const char __user *, pathname, umode_t, mode)
3808 {
3809         struct dentry *dentry;
3810         struct path path;
3811         int error;
3812         unsigned int lookup_flags = LOOKUP_DIRECTORY;
3813
3814 retry:
3815         dentry = user_path_create(dfd, pathname, &path, lookup_flags);
3816         if (IS_ERR(dentry))
3817                 return PTR_ERR(dentry);
3818
3819         if (!IS_POSIXACL(path.dentry->d_inode))
3820                 mode &= ~current_umask();
3821         error = security_path_mkdir(&path, dentry, mode);
3822         if (!error)
3823                 error = vfs_mkdir(path.dentry->d_inode, dentry, mode);
3824         done_path_create(&path, dentry);
3825         if (retry_estale(error, lookup_flags)) {
3826                 lookup_flags |= LOOKUP_REVAL;
3827                 goto retry;
3828         }
3829         return error;
3830 }
3831
3832 SYSCALL_DEFINE2(mkdir, const char __user *, pathname, umode_t, mode)
3833 {
3834         return sys_mkdirat(AT_FDCWD, pathname, mode);
3835 }
3836
3837 int vfs_rmdir(struct inode *dir, struct dentry *dentry)
3838 {
3839         int error = may_delete(dir, dentry, 1);
3840
3841         if (error)
3842                 return error;
3843
3844         if (!dir->i_op->rmdir)
3845                 return -EPERM;
3846
3847         dget(dentry);
3848         inode_lock(dentry->d_inode);
3849
3850         error = -EBUSY;
3851         if (is_local_mountpoint(dentry))
3852                 goto out;
3853
3854         error = security_inode_rmdir(dir, dentry);
3855         if (error)
3856                 goto out;
3857
3858         shrink_dcache_parent(dentry);
3859         error = dir->i_op->rmdir(dir, dentry);
3860         if (error)
3861                 goto out;
3862
3863         dentry->d_inode->i_flags |= S_DEAD;
3864         dont_mount(dentry);
3865         detach_mounts(dentry);
3866
3867 out:
3868         inode_unlock(dentry->d_inode);
3869         dput(dentry);
3870         if (!error)
3871                 d_delete(dentry);
3872         return error;
3873 }
3874 EXPORT_SYMBOL(vfs_rmdir);
3875
3876 static long do_rmdir(int dfd, const char __user *pathname)
3877 {
3878         int error = 0;
3879         struct filename *name;
3880         struct dentry *dentry;
3881         struct path path;
3882         struct qstr last;
3883         int type;
3884         unsigned int lookup_flags = 0;
3885 retry:
3886         name = filename_parentat(dfd, getname(pathname), lookup_flags,
3887                                 &path, &last, &type);
3888         if (IS_ERR(name))
3889                 return PTR_ERR(name);
3890
3891         switch (type) {
3892         case LAST_DOTDOT:
3893                 error = -ENOTEMPTY;
3894                 goto exit1;
3895         case LAST_DOT:
3896                 error = -EINVAL;
3897                 goto exit1;
3898         case LAST_ROOT:
3899                 error = -EBUSY;
3900                 goto exit1;
3901         }
3902
3903         error = mnt_want_write(path.mnt);
3904         if (error)
3905                 goto exit1;
3906
3907         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3908         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3909         error = PTR_ERR(dentry);
3910         if (IS_ERR(dentry))
3911                 goto exit2;
3912         if (!dentry->d_inode) {
3913                 error = -ENOENT;
3914                 goto exit3;
3915         }
3916         error = security_path_rmdir(&path, dentry);
3917         if (error)
3918                 goto exit3;
3919         error = vfs_rmdir(path.dentry->d_inode, dentry);
3920 exit3:
3921         dput(dentry);
3922 exit2:
3923         inode_unlock(path.dentry->d_inode);
3924         mnt_drop_write(path.mnt);
3925 exit1:
3926         path_put(&path);
3927         putname(name);
3928         if (retry_estale(error, lookup_flags)) {
3929                 lookup_flags |= LOOKUP_REVAL;
3930                 goto retry;
3931         }
3932         return error;
3933 }
3934
3935 SYSCALL_DEFINE1(rmdir, const char __user *, pathname)
3936 {
3937         return do_rmdir(AT_FDCWD, pathname);
3938 }
3939
3940 /**
3941  * vfs_unlink - unlink a filesystem object
3942  * @dir:        parent directory
3943  * @dentry:     victim
3944  * @delegated_inode: returns victim inode, if the inode is delegated.
3945  *
3946  * The caller must hold dir->i_mutex.
3947  *
3948  * If vfs_unlink discovers a delegation, it will return -EWOULDBLOCK and
3949  * return a reference to the inode in delegated_inode.  The caller
3950  * should then break the delegation on that inode and retry.  Because
3951  * breaking a delegation may take a long time, the caller should drop
3952  * dir->i_mutex before doing so.
3953  *
3954  * Alternatively, a caller may pass NULL for delegated_inode.  This may
3955  * be appropriate for callers that expect the underlying filesystem not
3956  * to be NFS exported.
3957  */
3958 int vfs_unlink(struct inode *dir, struct dentry *dentry, struct inode **delegated_inode)
3959 {
3960         struct inode *target = dentry->d_inode;
3961         int error = may_delete(dir, dentry, 0);
3962
3963         if (error)
3964                 return error;
3965
3966         if (!dir->i_op->unlink)
3967                 return -EPERM;
3968
3969         inode_lock(target);
3970         if (is_local_mountpoint(dentry))
3971                 error = -EBUSY;
3972         else {
3973                 error = security_inode_unlink(dir, dentry);
3974                 if (!error) {
3975                         error = try_break_deleg(target, delegated_inode);
3976                         if (error)
3977                                 goto out;
3978                         error = dir->i_op->unlink(dir, dentry);
3979                         if (!error) {
3980                                 dont_mount(dentry);
3981                                 detach_mounts(dentry);
3982                         }
3983                 }
3984         }
3985 out:
3986         inode_unlock(target);
3987
3988         /* We don't d_delete() NFS sillyrenamed files--they still exist. */
3989         if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
3990                 fsnotify_link_count(target);
3991                 d_delete(dentry);
3992         }
3993
3994         return error;
3995 }
3996 EXPORT_SYMBOL(vfs_unlink);
3997
3998 /*
3999  * Make sure that the actual truncation of the file will occur outside its
4000  * directory's i_mutex.  Truncate can take a long time if there is a lot of
4001  * writeout happening, and we don't want to prevent access to the directory
4002  * while waiting on the I/O.
4003  */
4004 static long do_unlinkat(int dfd, const char __user *pathname)
4005 {
4006         int error;
4007         struct filename *name;
4008         struct dentry *dentry;
4009         struct path path;
4010         struct qstr last;
4011         int type;
4012         struct inode *inode = NULL;
4013         struct inode *delegated_inode = NULL;
4014         unsigned int lookup_flags = 0;
4015 retry:
4016         name = filename_parentat(dfd, getname(pathname), lookup_flags,
4017                                 &path, &last, &type);
4018         if (IS_ERR(name))
4019                 return PTR_ERR(name);
4020
4021         error = -EISDIR;
4022         if (type != LAST_NORM)
4023                 goto exit1;
4024
4025         error = mnt_want_write(path.mnt);
4026         if (error)
4027                 goto exit1;
4028 retry_deleg:
4029         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
4030         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
4031         error = PTR_ERR(dentry);
4032         if (!IS_ERR(dentry)) {
4033                 /* Why not before? Because we want correct error value */
4034                 if (last.name[last.len])
4035                         goto slashes;
4036                 inode = dentry->d_inode;
4037                 if (d_is_negative(dentry))
4038                         goto slashes;
4039                 ihold(inode);
4040                 error = security_path_unlink(&path, dentry);
4041                 if (error)
4042                         goto exit2;
4043                 error = vfs_unlink(path.dentry->d_inode, dentry, &delegated_inode);
4044 exit2:
4045                 dput(dentry);
4046         }
4047         inode_unlock(path.dentry->d_inode);
4048         if (inode)
4049                 iput(inode);    /* truncate the inode here */
4050         inode = NULL;
4051         if (delegated_inode) {
4052                 error = break_deleg_wait(&delegated_inode);
4053                 if (!error)
4054                         goto retry_deleg;
4055         }
4056         mnt_drop_write(path.mnt);
4057 exit1:
4058         path_put(&path);
4059         putname(name);
4060         if (retry_estale(error, lookup_flags)) {
4061                 lookup_flags |= LOOKUP_REVAL;
4062                 inode = NULL;
4063                 goto retry;
4064         }
4065         return error;
4066
4067 slashes:
4068         if (d_is_negative(dentry))
4069                 error = -ENOENT;
4070         else if (d_is_dir(dentry))
4071                 error = -EISDIR;
4072         else
4073                 error = -ENOTDIR;
4074         goto exit2;
4075 }
4076
4077 SYSCALL_DEFINE3(unlinkat, int, dfd, const char __user *, pathname, int, flag)
4078 {
4079         if ((flag & ~AT_REMOVEDIR) != 0)
4080                 return -EINVAL;
4081
4082         if (flag & AT_REMOVEDIR)
4083                 return do_rmdir(dfd, pathname);
4084
4085         return do_unlinkat(dfd, pathname);
4086 }
4087
4088 SYSCALL_DEFINE1(unlink, const char __user *, pathname)
4089 {
4090         return do_unlinkat(AT_FDCWD, pathname);
4091 }
4092
4093 int vfs_symlink(struct inode *dir, struct dentry *dentry, const char *oldname)
4094 {
4095         int error = may_create(dir, dentry);
4096
4097         if (error)
4098                 return error;
4099
4100         if (!dir->i_op->symlink)
4101                 return -EPERM;
4102
4103         error = security_inode_symlink(dir, dentry, oldname);
4104         if (error)
4105                 return error;
4106
4107         error = dir->i_op->symlink(dir, dentry, oldname);
4108         if (!error)
4109                 fsnotify_create(dir, dentry);
4110         return error;
4111 }
4112 EXPORT_SYMBOL(vfs_symlink);
4113
4114 SYSCALL_DEFINE3(symlinkat, const char __user *, oldname,
4115                 int, newdfd, const char __user *, newname)
4116 {
4117         int error;
4118         struct filename *from;
4119         struct dentry *dentry;
4120         struct path path;
4121         unsigned int lookup_flags = 0;
4122
4123         from = getname(oldname);
4124         if (IS_ERR(from))
4125                 return PTR_ERR(from);
4126 retry:
4127         dentry = user_path_create(newdfd, newname, &path, lookup_flags);
4128         error = PTR_ERR(dentry);
4129         if (IS_ERR(dentry))
4130                 goto out_putname;
4131
4132         error = security_path_symlink(&path, dentry, from->name);
4133         if (!error)
4134                 error = vfs_symlink(path.dentry->d_inode, dentry, from->name);
4135         done_path_create(&path, dentry);
4136         if (retry_estale(error, lookup_flags)) {
4137                 lookup_flags |= LOOKUP_REVAL;
4138                 goto retry;
4139         }
4140 out_putname:
4141         putname(from);
4142         return error;
4143 }
4144
4145 SYSCALL_DEFINE2(symlink, const char __user *, oldname, const char __user *, newname)
4146 {
4147         return sys_symlinkat(oldname, AT_FDCWD, newname);
4148 }
4149
4150 /**
4151  * vfs_link - create a new link
4152  * @old_dentry: object to be linked
4153  * @dir:        new parent
4154  * @new_dentry: where to create the new link
4155  * @delegated_inode: returns inode needing a delegation break
4156  *
4157  * The caller must hold dir->i_mutex
4158  *
4159  * If vfs_link discovers a delegation on the to-be-linked file in need
4160  * of breaking, it will return -EWOULDBLOCK and return a reference to the
4161  * inode in delegated_inode.  The caller should then break the delegation
4162  * and retry.  Because breaking a delegation may take a long time, the
4163  * caller should drop the i_mutex before doing so.
4164  *
4165  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4166  * be appropriate for callers that expect the underlying filesystem not
4167  * to be NFS exported.
4168  */
4169 int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_dentry, struct inode **delegated_inode)
4170 {
4171         struct inode *inode = old_dentry->d_inode;
4172         unsigned max_links = dir->i_sb->s_max_links;
4173         int error;
4174
4175         if (!inode)
4176                 return -ENOENT;
4177
4178         error = may_create(dir, new_dentry);
4179         if (error)
4180                 return error;
4181
4182         if (dir->i_sb != inode->i_sb)
4183                 return -EXDEV;
4184
4185         /*
4186          * A link to an append-only or immutable file cannot be created.
4187          */
4188         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4189                 return -EPERM;
4190         /*
4191          * Updating the link count will likely cause i_uid and i_gid to
4192          * be writen back improperly if their true value is unknown to
4193          * the vfs.
4194          */
4195         if (HAS_UNMAPPED_ID(inode))
4196                 return -EPERM;
4197         if (!dir->i_op->link)
4198                 return -EPERM;
4199         if (S_ISDIR(inode->i_mode))
4200                 return -EPERM;
4201
4202         error = security_inode_link(old_dentry, dir, new_dentry);
4203         if (error)
4204                 return error;
4205
4206         inode_lock(inode);
4207         /* Make sure we don't allow creating hardlink to an unlinked file */
4208         if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))
4209                 error =  -ENOENT;
4210         else if (max_links && inode->i_nlink >= max_links)
4211                 error = -EMLINK;
4212         else {
4213                 error = try_break_deleg(inode, delegated_inode);
4214                 if (!error)
4215                         error = dir->i_op->link(old_dentry, dir, new_dentry);
4216         }
4217
4218         if (!error && (inode->i_state & I_LINKABLE)) {
4219                 spin_lock(&inode->i_lock);
4220                 inode->i_state &= ~I_LINKABLE;
4221                 spin_unlock(&inode->i_lock);
4222         }
4223         inode_unlock(inode);
4224         if (!error)
4225                 fsnotify_link(dir, inode, new_dentry);
4226         return error;
4227 }
4228 EXPORT_SYMBOL(vfs_link);
4229
4230 /*
4231  * Hardlinks are often used in delicate situations.  We avoid
4232  * security-related surprises by not following symlinks on the
4233  * newname.  --KAB
4234  *
4235  * We don't follow them on the oldname either to be compatible
4236  * with linux 2.0, and to avoid hard-linking to directories
4237  * and other special files.  --ADM
4238  */
4239 SYSCALL_DEFINE5(linkat, int, olddfd, const char __user *, oldname,
4240                 int, newdfd, const char __user *, newname, int, flags)
4241 {
4242         struct dentry *new_dentry;
4243         struct path old_path, new_path;
4244         struct inode *delegated_inode = NULL;
4245         int how = 0;
4246         int error;
4247
4248         if ((flags & ~(AT_SYMLINK_FOLLOW | AT_EMPTY_PATH)) != 0)
4249                 return -EINVAL;
4250         /*
4251          * To use null names we require CAP_DAC_READ_SEARCH
4252          * This ensures that not everyone will be able to create
4253          * handlink using the passed filedescriptor.
4254          */
4255         if (flags & AT_EMPTY_PATH) {
4256                 if (!capable(CAP_DAC_READ_SEARCH))
4257                         return -ENOENT;
4258                 how = LOOKUP_EMPTY;
4259         }
4260
4261         if (flags & AT_SYMLINK_FOLLOW)
4262                 how |= LOOKUP_FOLLOW;
4263 retry:
4264         error = user_path_at(olddfd, oldname, how, &old_path);
4265         if (error)
4266                 return error;
4267
4268         new_dentry = user_path_create(newdfd, newname, &new_path,
4269                                         (how & LOOKUP_REVAL));
4270         error = PTR_ERR(new_dentry);
4271         if (IS_ERR(new_dentry))
4272                 goto out;
4273
4274         error = -EXDEV;
4275         if (old_path.mnt != new_path.mnt)
4276                 goto out_dput;
4277         error = may_linkat(&old_path);
4278         if (unlikely(error))
4279                 goto out_dput;
4280         error = security_path_link(old_path.dentry, &new_path, new_dentry);
4281         if (error)
4282                 goto out_dput;
4283         error = vfs_link(old_path.dentry, new_path.dentry->d_inode, new_dentry, &delegated_inode);
4284 out_dput:
4285         done_path_create(&new_path, new_dentry);
4286         if (delegated_inode) {
4287                 error = break_deleg_wait(&delegated_inode);
4288                 if (!error) {
4289                         path_put(&old_path);
4290                         goto retry;
4291                 }
4292         }
4293         if (retry_estale(error, how)) {
4294                 path_put(&old_path);
4295                 how |= LOOKUP_REVAL;
4296                 goto retry;
4297         }
4298 out:
4299         path_put(&old_path);
4300
4301         return error;
4302 }
4303
4304 SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname)
4305 {
4306         return sys_linkat(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4307 }
4308
4309 /**
4310  * vfs_rename - rename a filesystem object
4311  * @old_dir:    parent of source
4312  * @old_dentry: source
4313  * @new_dir:    parent of destination
4314  * @new_dentry: destination
4315  * @delegated_inode: returns an inode needing a delegation break
4316  * @flags:      rename flags
4317  *
4318  * The caller must hold multiple mutexes--see lock_rename()).
4319  *
4320  * If vfs_rename discovers a delegation in need of breaking at either
4321  * the source or destination, it will return -EWOULDBLOCK and return a
4322  * reference to the inode in delegated_inode.  The caller should then
4323  * break the delegation and retry.  Because breaking a delegation may
4324  * take a long time, the caller should drop all locks before doing
4325  * so.
4326  *
4327  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4328  * be appropriate for callers that expect the underlying filesystem not
4329  * to be NFS exported.
4330  *
4331  * The worst of all namespace operations - renaming directory. "Perverted"
4332  * doesn't even start to describe it. Somebody in UCB had a heck of a trip...
4333  * Problems:
4334  *
4335  *      a) we can get into loop creation.
4336  *      b) race potential - two innocent renames can create a loop together.
4337  *         That's where 4.4 screws up. Current fix: serialization on
4338  *         sb->s_vfs_rename_mutex. We might be more accurate, but that's another
4339  *         story.
4340  *      c) we have to lock _four_ objects - parents and victim (if it exists),
4341  *         and source (if it is not a directory).
4342  *         And that - after we got ->i_mutex on parents (until then we don't know
4343  *         whether the target exists).  Solution: try to be smart with locking
4344  *         order for inodes.  We rely on the fact that tree topology may change
4345  *         only under ->s_vfs_rename_mutex _and_ that parent of the object we
4346  *         move will be locked.  Thus we can rank directories by the tree
4347  *         (ancestors first) and rank all non-directories after them.
4348  *         That works since everybody except rename does "lock parent, lookup,
4349  *         lock child" and rename is under ->s_vfs_rename_mutex.
4350  *         HOWEVER, it relies on the assumption that any object with ->lookup()
4351  *         has no more than 1 dentry.  If "hybrid" objects will ever appear,
4352  *         we'd better make sure that there's no link(2) for them.
4353  *      d) conversion from fhandle to dentry may come in the wrong moment - when
4354  *         we are removing the target. Solution: we will have to grab ->i_mutex
4355  *         in the fhandle_to_dentry code. [FIXME - current nfsfh.c relies on
4356  *         ->i_mutex on parents, which works but leads to some truly excessive
4357  *         locking].
4358  */
4359 int vfs_rename(struct inode *old_dir, struct dentry *old_dentry,
4360                struct inode *new_dir, struct dentry *new_dentry,
4361                struct inode **delegated_inode, unsigned int flags)
4362 {
4363         int error;
4364         bool is_dir = d_is_dir(old_dentry);
4365         struct inode *source = old_dentry->d_inode;
4366         struct inode *target = new_dentry->d_inode;
4367         bool new_is_dir = false;
4368         unsigned max_links = new_dir->i_sb->s_max_links;
4369         struct name_snapshot old_name;
4370
4371         if (source == target)
4372                 return 0;
4373
4374         error = may_delete(old_dir, old_dentry, is_dir);
4375         if (error)
4376                 return error;
4377
4378         if (!target) {
4379                 error = may_create(new_dir, new_dentry);
4380         } else {
4381                 new_is_dir = d_is_dir(new_dentry);
4382
4383                 if (!(flags & RENAME_EXCHANGE))
4384                         error = may_delete(new_dir, new_dentry, is_dir);
4385                 else
4386                         error = may_delete(new_dir, new_dentry, new_is_dir);
4387         }
4388         if (error)
4389                 return error;
4390
4391         if (!old_dir->i_op->rename)
4392                 return -EPERM;
4393
4394         /*
4395          * If we are going to change the parent - check write permissions,
4396          * we'll need to flip '..'.
4397          */
4398         if (new_dir != old_dir) {
4399                 if (is_dir) {
4400                         error = inode_permission(source, MAY_WRITE);
4401                         if (error)
4402                                 return error;
4403                 }
4404                 if ((flags & RENAME_EXCHANGE) && new_is_dir) {
4405                         error = inode_permission(target, MAY_WRITE);
4406                         if (error)
4407                                 return error;
4408                 }
4409         }
4410
4411         error = security_inode_rename(old_dir, old_dentry, new_dir, new_dentry,
4412                                       flags);
4413         if (error)
4414                 return error;
4415
4416         take_dentry_name_snapshot(&old_name, old_dentry);
4417         dget(new_dentry);
4418         if (!is_dir || (flags & RENAME_EXCHANGE))
4419                 lock_two_nondirectories(source, target);
4420         else if (target)
4421                 inode_lock(target);
4422
4423         error = -EBUSY;
4424         if (is_local_mountpoint(old_dentry) || is_local_mountpoint(new_dentry))
4425                 goto out;
4426
4427         if (max_links && new_dir != old_dir) {
4428                 error = -EMLINK;
4429                 if (is_dir && !new_is_dir && new_dir->i_nlink >= max_links)
4430                         goto out;
4431                 if ((flags & RENAME_EXCHANGE) && !is_dir && new_is_dir &&
4432                     old_dir->i_nlink >= max_links)
4433                         goto out;
4434         }
4435         if (is_dir && !(flags & RENAME_EXCHANGE) && target)
4436                 shrink_dcache_parent(new_dentry);
4437         if (!is_dir) {
4438                 error = try_break_deleg(source, delegated_inode);
4439                 if (error)
4440                         goto out;
4441         }
4442         if (target && !new_is_dir) {
4443                 error = try_break_deleg(target, delegated_inode);
4444                 if (error)
4445                         goto out;
4446         }
4447         error = old_dir->i_op->rename(old_dir, old_dentry,
4448                                        new_dir, new_dentry, flags);
4449         if (error)
4450                 goto out;
4451
4452         if (!(flags & RENAME_EXCHANGE) && target) {
4453                 if (is_dir)
4454                         target->i_flags |= S_DEAD;
4455                 dont_mount(new_dentry);
4456                 detach_mounts(new_dentry);
4457         }
4458         if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE)) {
4459                 if (!(flags & RENAME_EXCHANGE))
4460                         d_move(old_dentry, new_dentry);
4461                 else
4462                         d_exchange(old_dentry, new_dentry);
4463         }
4464 out:
4465         if (!is_dir || (flags & RENAME_EXCHANGE))
4466                 unlock_two_nondirectories(source, target);
4467         else if (target)
4468                 inode_unlock(target);
4469         dput(new_dentry);
4470         if (!error) {
4471                 fsnotify_move(old_dir, new_dir, old_name.name, is_dir,
4472                               !(flags & RENAME_EXCHANGE) ? target : NULL, old_dentry);
4473                 if (flags & RENAME_EXCHANGE) {
4474                         fsnotify_move(new_dir, old_dir, old_dentry->d_name.name,
4475                                       new_is_dir, NULL, new_dentry);
4476                 }
4477         }
4478         release_dentry_name_snapshot(&old_name);
4479
4480         return error;
4481 }
4482 EXPORT_SYMBOL(vfs_rename);
4483
4484 SYSCALL_DEFINE5(renameat2, int, olddfd, const char __user *, oldname,
4485                 int, newdfd, const char __user *, newname, unsigned int, flags)
4486 {
4487         struct dentry *old_dentry, *new_dentry;
4488         struct dentry *trap;
4489         struct path old_path, new_path;
4490         struct qstr old_last, new_last;
4491         int old_type, new_type;
4492         struct inode *delegated_inode = NULL;
4493         struct filename *from;
4494         struct filename *to;
4495         unsigned int lookup_flags = 0, target_flags = LOOKUP_RENAME_TARGET;
4496         bool should_retry = false;
4497         int error;
4498
4499         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
4500                 return -EINVAL;
4501
4502         if ((flags & (RENAME_NOREPLACE | RENAME_WHITEOUT)) &&
4503             (flags & RENAME_EXCHANGE))
4504                 return -EINVAL;
4505
4506         if ((flags & RENAME_WHITEOUT) && !capable(CAP_MKNOD))
4507                 return -EPERM;
4508
4509         if (flags & RENAME_EXCHANGE)
4510                 target_flags = 0;
4511
4512 retry:
4513         from = filename_parentat(olddfd, getname(oldname), lookup_flags,
4514                                 &old_path, &old_last, &old_type);
4515         if (IS_ERR(from)) {
4516                 error = PTR_ERR(from);
4517                 goto exit;
4518         }
4519
4520         to = filename_parentat(newdfd, getname(newname), lookup_flags,
4521                                 &new_path, &new_last, &new_type);
4522         if (IS_ERR(to)) {
4523                 error = PTR_ERR(to);
4524                 goto exit1;
4525         }
4526
4527         error = -EXDEV;
4528         if (old_path.mnt != new_path.mnt)
4529                 goto exit2;
4530
4531         error = -EBUSY;
4532         if (old_type != LAST_NORM)
4533                 goto exit2;
4534
4535         if (flags & RENAME_NOREPLACE)
4536                 error = -EEXIST;
4537         if (new_type != LAST_NORM)
4538                 goto exit2;
4539
4540         error = mnt_want_write(old_path.mnt);
4541         if (error)
4542                 goto exit2;
4543
4544 retry_deleg:
4545         trap = lock_rename(new_path.dentry, old_path.dentry);
4546
4547         old_dentry = __lookup_hash(&old_last, old_path.dentry, lookup_flags);
4548         error = PTR_ERR(old_dentry);
4549         if (IS_ERR(old_dentry))
4550                 goto exit3;
4551         /* source must exist */
4552         error = -ENOENT;
4553         if (d_is_negative(old_dentry))
4554                 goto exit4;
4555         new_dentry = __lookup_hash(&new_last, new_path.dentry, lookup_flags | target_flags);
4556         error = PTR_ERR(new_dentry);
4557         if (IS_ERR(new_dentry))
4558                 goto exit4;
4559         error = -EEXIST;
4560         if ((flags & RENAME_NOREPLACE) && d_is_positive(new_dentry))
4561                 goto exit5;
4562         if (flags & RENAME_EXCHANGE) {
4563                 error = -ENOENT;
4564                 if (d_is_negative(new_dentry))
4565                         goto exit5;
4566
4567                 if (!d_is_dir(new_dentry)) {
4568                         error = -ENOTDIR;
4569                         if (new_last.name[new_last.len])
4570                                 goto exit5;
4571                 }
4572         }
4573         /* unless the source is a directory trailing slashes give -ENOTDIR */
4574         if (!d_is_dir(old_dentry)) {
4575                 error = -ENOTDIR;
4576                 if (old_last.name[old_last.len])
4577                         goto exit5;
4578                 if (!(flags & RENAME_EXCHANGE) && new_last.name[new_last.len])
4579                         goto exit5;
4580         }
4581         /* source should not be ancestor of target */
4582         error = -EINVAL;
4583         if (old_dentry == trap)
4584                 goto exit5;
4585         /* target should not be an ancestor of source */
4586         if (!(flags & RENAME_EXCHANGE))
4587                 error = -ENOTEMPTY;
4588         if (new_dentry == trap)
4589                 goto exit5;
4590
4591         error = security_path_rename(&old_path, old_dentry,
4592                                      &new_path, new_dentry, flags);
4593         if (error)
4594                 goto exit5;
4595         error = vfs_rename(old_path.dentry->d_inode, old_dentry,
4596                            new_path.dentry->d_inode, new_dentry,
4597                            &delegated_inode, flags);
4598 exit5:
4599         dput(new_dentry);
4600 exit4:
4601         dput(old_dentry);
4602 exit3:
4603         unlock_rename(new_path.dentry, old_path.dentry);
4604         if (delegated_inode) {
4605                 error = break_deleg_wait(&delegated_inode);
4606                 if (!error)
4607                         goto retry_deleg;
4608         }
4609         mnt_drop_write(old_path.mnt);
4610 exit2:
4611         if (retry_estale(error, lookup_flags))
4612                 should_retry = true;
4613         path_put(&new_path);
4614         putname(to);
4615 exit1:
4616         path_put(&old_path);
4617         putname(from);
4618         if (should_retry) {
4619                 should_retry = false;
4620                 lookup_flags |= LOOKUP_REVAL;
4621                 goto retry;
4622         }
4623 exit:
4624         return error;
4625 }
4626
4627 SYSCALL_DEFINE4(renameat, int, olddfd, const char __user *, oldname,
4628                 int, newdfd, const char __user *, newname)
4629 {
4630         return sys_renameat2(olddfd, oldname, newdfd, newname, 0);
4631 }
4632
4633 SYSCALL_DEFINE2(rename, const char __user *, oldname, const char __user *, newname)
4634 {
4635         return sys_renameat2(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4636 }
4637
4638 int vfs_whiteout(struct inode *dir, struct dentry *dentry)
4639 {
4640         int error = may_create(dir, dentry);
4641         if (error)
4642                 return error;
4643
4644         if (!dir->i_op->mknod)
4645                 return -EPERM;
4646
4647         return dir->i_op->mknod(dir, dentry,
4648                                 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
4649 }
4650 EXPORT_SYMBOL(vfs_whiteout);
4651
4652 int readlink_copy(char __user *buffer, int buflen, const char *link)
4653 {
4654         int len = PTR_ERR(link);
4655         if (IS_ERR(link))
4656                 goto out;
4657
4658         len = strlen(link);
4659         if (len > (unsigned) buflen)
4660                 len = buflen;
4661         if (copy_to_user(buffer, link, len))
4662                 len = -EFAULT;
4663 out:
4664         return len;
4665 }
4666
4667 /*
4668  * A helper for ->readlink().  This should be used *ONLY* for symlinks that
4669  * have ->get_link() not calling nd_jump_link().  Using (or not using) it
4670  * for any given inode is up to filesystem.
4671  */
4672 static int generic_readlink(struct dentry *dentry, char __user *buffer,
4673                             int buflen)
4674 {
4675         DEFINE_DELAYED_CALL(done);
4676         struct inode *inode = d_inode(dentry);
4677         const char *link = inode->i_link;
4678         int res;
4679
4680         if (!link) {
4681                 link = inode->i_op->get_link(dentry, inode, &done);
4682                 if (IS_ERR(link))
4683                         return PTR_ERR(link);
4684         }
4685         res = readlink_copy(buffer, buflen, link);
4686         do_delayed_call(&done);
4687         return res;
4688 }
4689
4690 /**
4691  * vfs_readlink - copy symlink body into userspace buffer
4692  * @dentry: dentry on which to get symbolic link
4693  * @buffer: user memory pointer
4694  * @buflen: size of buffer
4695  *
4696  * Does not touch atime.  That's up to the caller if necessary
4697  *
4698  * Does not call security hook.
4699  */
4700 int vfs_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4701 {
4702         struct inode *inode = d_inode(dentry);
4703
4704         if (unlikely(!(inode->i_opflags & IOP_DEFAULT_READLINK))) {
4705                 if (unlikely(inode->i_op->readlink))
4706                         return inode->i_op->readlink(dentry, buffer, buflen);
4707
4708                 if (!d_is_symlink(dentry))
4709                         return -EINVAL;
4710
4711                 spin_lock(&inode->i_lock);
4712                 inode->i_opflags |= IOP_DEFAULT_READLINK;
4713                 spin_unlock(&inode->i_lock);
4714         }
4715
4716         return generic_readlink(dentry, buffer, buflen);
4717 }
4718 EXPORT_SYMBOL(vfs_readlink);
4719
4720 /**
4721  * vfs_get_link - get symlink body
4722  * @dentry: dentry on which to get symbolic link
4723  * @done: caller needs to free returned data with this
4724  *
4725  * Calls security hook and i_op->get_link() on the supplied inode.
4726  *
4727  * It does not touch atime.  That's up to the caller if necessary.
4728  *
4729  * Does not work on "special" symlinks like /proc/$$/fd/N
4730  */
4731 const char *vfs_get_link(struct dentry *dentry, struct delayed_call *done)
4732 {
4733         const char *res = ERR_PTR(-EINVAL);
4734         struct inode *inode = d_inode(dentry);
4735
4736         if (d_is_symlink(dentry)) {
4737                 res = ERR_PTR(security_inode_readlink(dentry));
4738                 if (!res)
4739                         res = inode->i_op->get_link(dentry, inode, done);
4740         }
4741         return res;
4742 }
4743 EXPORT_SYMBOL(vfs_get_link);
4744
4745 /* get the link contents into pagecache */
4746 const char *page_get_link(struct dentry *dentry, struct inode *inode,
4747                           struct delayed_call *callback)
4748 {
4749         char *kaddr;
4750         struct page *page;
4751         struct address_space *mapping = inode->i_mapping;
4752
4753         if (!dentry) {
4754                 page = find_get_page(mapping, 0);
4755                 if (!page)
4756                         return ERR_PTR(-ECHILD);
4757                 if (!PageUptodate(page)) {
4758                         put_page(page);
4759                         return ERR_PTR(-ECHILD);
4760                 }
4761         } else {
4762                 page = read_mapping_page(mapping, 0, NULL);
4763                 if (IS_ERR(page))
4764                         return (char*)page;
4765         }
4766         set_delayed_call(callback, page_put_link, page);
4767         BUG_ON(mapping_gfp_mask(mapping) & __GFP_HIGHMEM);
4768         kaddr = page_address(page);
4769         nd_terminate_link(kaddr, inode->i_size, PAGE_SIZE - 1);
4770         return kaddr;
4771 }
4772
4773 EXPORT_SYMBOL(page_get_link);
4774
4775 void page_put_link(void *arg)
4776 {
4777         put_page(arg);
4778 }
4779 EXPORT_SYMBOL(page_put_link);
4780
4781 int page_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4782 {
4783         DEFINE_DELAYED_CALL(done);
4784         int res = readlink_copy(buffer, buflen,
4785                                 page_get_link(dentry, d_inode(dentry),
4786                                               &done));
4787         do_delayed_call(&done);
4788         return res;
4789 }
4790 EXPORT_SYMBOL(page_readlink);
4791
4792 /*
4793  * The nofs argument instructs pagecache_write_begin to pass AOP_FLAG_NOFS
4794  */
4795 int __page_symlink(struct inode *inode, const char *symname, int len, int nofs)
4796 {
4797         struct address_space *mapping = inode->i_mapping;
4798         struct page *page;
4799         void *fsdata;
4800         int err;
4801         unsigned int flags = 0;
4802         if (nofs)
4803                 flags |= AOP_FLAG_NOFS;
4804
4805 retry:
4806         err = pagecache_write_begin(NULL, mapping, 0, len-1,
4807                                 flags, &page, &fsdata);
4808         if (err)
4809                 goto fail;
4810
4811         memcpy(page_address(page), symname, len-1);
4812
4813         err = pagecache_write_end(NULL, mapping, 0, len-1, len-1,
4814                                                         page, fsdata);
4815         if (err < 0)
4816                 goto fail;
4817         if (err < len-1)
4818                 goto retry;
4819
4820         mark_inode_dirty(inode);
4821         return 0;
4822 fail:
4823         return err;
4824 }
4825 EXPORT_SYMBOL(__page_symlink);
4826
4827 int page_symlink(struct inode *inode, const char *symname, int len)
4828 {
4829         return __page_symlink(inode, symname, len,
4830                         !mapping_gfp_constraint(inode->i_mapping, __GFP_FS));
4831 }
4832 EXPORT_SYMBOL(page_symlink);
4833
4834 const struct inode_operations page_symlink_inode_operations = {
4835         .get_link       = page_get_link,
4836 };
4837 EXPORT_SYMBOL(page_symlink_inode_operations);