Merge tag 'dmaengine-4.8-rc1' of git://git.infradead.org/users/vkoul/slave-dma
[cascardo/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 <asm/uaccess.h>
40
41 #include "internal.h"
42 #include "mount.h"
43
44 /* [Feb-1997 T. Schoebel-Theuer]
45  * Fundamental changes in the pathname lookup mechanisms (namei)
46  * were necessary because of omirr.  The reason is that omirr needs
47  * to know the _real_ pathname, not the user-supplied one, in case
48  * of symlinks (and also when transname replacements occur).
49  *
50  * The new code replaces the old recursive symlink resolution with
51  * an iterative one (in case of non-nested symlink chains).  It does
52  * this with calls to <fs>_follow_link().
53  * As a side effect, dir_namei(), _namei() and follow_link() are now 
54  * replaced with a single function lookup_dentry() that can handle all 
55  * the special cases of the former code.
56  *
57  * With the new dcache, the pathname is stored at each inode, at least as
58  * long as the refcount of the inode is positive.  As a side effect, the
59  * size of the dcache depends on the inode cache and thus is dynamic.
60  *
61  * [29-Apr-1998 C. Scott Ananian] Updated above description of symlink
62  * resolution to correspond with current state of the code.
63  *
64  * Note that the symlink resolution is not *completely* iterative.
65  * There is still a significant amount of tail- and mid- recursion in
66  * the algorithm.  Also, note that <fs>_readlink() is not used in
67  * lookup_dentry(): lookup_dentry() on the result of <fs>_readlink()
68  * may return different results than <fs>_follow_link().  Many virtual
69  * filesystems (including /proc) exhibit this behavior.
70  */
71
72 /* [24-Feb-97 T. Schoebel-Theuer] Side effects caused by new implementation:
73  * New symlink semantics: when open() is called with flags O_CREAT | O_EXCL
74  * and the name already exists in form of a symlink, try to create the new
75  * name indicated by the symlink. The old code always complained that the
76  * name already exists, due to not following the symlink even if its target
77  * is nonexistent.  The new semantics affects also mknod() and link() when
78  * the name is a symlink pointing to a non-existent name.
79  *
80  * I don't know which semantics is the right one, since I have no access
81  * to standards. But I found by trial that HP-UX 9.0 has the full "new"
82  * semantics implemented, while SunOS 4.1.1 and Solaris (SunOS 5.4) have the
83  * "old" one. Personally, I think the new semantics is much more logical.
84  * Note that "ln old new" where "new" is a symlink pointing to a non-existing
85  * file does succeed in both HP-UX and SunOs, but not in Solaris
86  * and in the old Linux semantics.
87  */
88
89 /* [16-Dec-97 Kevin Buhr] For security reasons, we change some symlink
90  * semantics.  See the comments in "open_namei" and "do_link" below.
91  *
92  * [10-Sep-98 Alan Modra] Another symlink change.
93  */
94
95 /* [Feb-Apr 2000 AV] Complete rewrite. Rules for symlinks:
96  *      inside the path - always follow.
97  *      in the last component in creation/removal/renaming - never follow.
98  *      if LOOKUP_FOLLOW passed - follow.
99  *      if the pathname has trailing slashes - follow.
100  *      otherwise - don't follow.
101  * (applied in that order).
102  *
103  * [Jun 2000 AV] Inconsistent behaviour of open() in case if flags==O_CREAT
104  * restored for 2.4. This is the last surviving part of old 4.2BSD bug.
105  * During the 2.4 we need to fix the userland stuff depending on it -
106  * hopefully we will be able to get rid of that wart in 2.5. So far only
107  * XEmacs seems to be relying on it...
108  */
109 /*
110  * [Sep 2001 AV] Single-semaphore locking scheme (kudos to David Holland)
111  * implemented.  Let's see if raised priority of ->s_vfs_rename_mutex gives
112  * any extra contention...
113  */
114
115 /* In order to reduce some races, while at the same time doing additional
116  * checking and hopefully speeding things up, we copy filenames to the
117  * kernel data space before using them..
118  *
119  * POSIX.1 2.4: an empty pathname is invalid (ENOENT).
120  * PATH_MAX includes the nul terminator --RR.
121  */
122
123 #define EMBEDDED_NAME_MAX       (PATH_MAX - offsetof(struct filename, iname))
124
125 struct filename *
126 getname_flags(const char __user *filename, int flags, int *empty)
127 {
128         struct filename *result;
129         char *kname;
130         int len;
131
132         result = audit_reusename(filename);
133         if (result)
134                 return result;
135
136         result = __getname();
137         if (unlikely(!result))
138                 return ERR_PTR(-ENOMEM);
139
140         /*
141          * First, try to embed the struct filename inside the names_cache
142          * allocation
143          */
144         kname = (char *)result->iname;
145         result->name = kname;
146
147         len = strncpy_from_user(kname, filename, EMBEDDED_NAME_MAX);
148         if (unlikely(len < 0)) {
149                 __putname(result);
150                 return ERR_PTR(len);
151         }
152
153         /*
154          * Uh-oh. We have a name that's approaching PATH_MAX. Allocate a
155          * separate struct filename so we can dedicate the entire
156          * names_cache allocation for the pathname, and re-do the copy from
157          * userland.
158          */
159         if (unlikely(len == EMBEDDED_NAME_MAX)) {
160                 const size_t size = offsetof(struct filename, iname[1]);
161                 kname = (char *)result;
162
163                 /*
164                  * size is chosen that way we to guarantee that
165                  * result->iname[0] is within the same object and that
166                  * kname can't be equal to result->iname, no matter what.
167                  */
168                 result = kzalloc(size, GFP_KERNEL);
169                 if (unlikely(!result)) {
170                         __putname(kname);
171                         return ERR_PTR(-ENOMEM);
172                 }
173                 result->name = kname;
174                 len = strncpy_from_user(kname, filename, PATH_MAX);
175                 if (unlikely(len < 0)) {
176                         __putname(kname);
177                         kfree(result);
178                         return ERR_PTR(len);
179                 }
180                 if (unlikely(len == PATH_MAX)) {
181                         __putname(kname);
182                         kfree(result);
183                         return ERR_PTR(-ENAMETOOLONG);
184                 }
185         }
186
187         result->refcnt = 1;
188         /* The empty path is special. */
189         if (unlikely(!len)) {
190                 if (empty)
191                         *empty = 1;
192                 if (!(flags & LOOKUP_EMPTY)) {
193                         putname(result);
194                         return ERR_PTR(-ENOENT);
195                 }
196         }
197
198         result->uptr = filename;
199         result->aname = NULL;
200         audit_getname(result);
201         return result;
202 }
203
204 struct filename *
205 getname(const char __user * filename)
206 {
207         return getname_flags(filename, 0, NULL);
208 }
209
210 struct filename *
211 getname_kernel(const char * filename)
212 {
213         struct filename *result;
214         int len = strlen(filename) + 1;
215
216         result = __getname();
217         if (unlikely(!result))
218                 return ERR_PTR(-ENOMEM);
219
220         if (len <= EMBEDDED_NAME_MAX) {
221                 result->name = (char *)result->iname;
222         } else if (len <= PATH_MAX) {
223                 struct filename *tmp;
224
225                 tmp = kmalloc(sizeof(*tmp), GFP_KERNEL);
226                 if (unlikely(!tmp)) {
227                         __putname(result);
228                         return ERR_PTR(-ENOMEM);
229                 }
230                 tmp->name = (char *)result;
231                 result = tmp;
232         } else {
233                 __putname(result);
234                 return ERR_PTR(-ENAMETOOLONG);
235         }
236         memcpy((char *)result->name, filename, len);
237         result->uptr = NULL;
238         result->aname = NULL;
239         result->refcnt = 1;
240         audit_getname(result);
241
242         return result;
243 }
244
245 void putname(struct filename *name)
246 {
247         BUG_ON(name->refcnt <= 0);
248
249         if (--name->refcnt > 0)
250                 return;
251
252         if (name->name != name->iname) {
253                 __putname(name->name);
254                 kfree(name);
255         } else
256                 __putname(name);
257 }
258
259 static int check_acl(struct inode *inode, int mask)
260 {
261 #ifdef CONFIG_FS_POSIX_ACL
262         struct posix_acl *acl;
263
264         if (mask & MAY_NOT_BLOCK) {
265                 acl = get_cached_acl_rcu(inode, ACL_TYPE_ACCESS);
266                 if (!acl)
267                         return -EAGAIN;
268                 /* no ->get_acl() calls in RCU mode... */
269                 if (is_uncached_acl(acl))
270                         return -ECHILD;
271                 return posix_acl_permission(inode, acl, mask & ~MAY_NOT_BLOCK);
272         }
273
274         acl = get_acl(inode, ACL_TYPE_ACCESS);
275         if (IS_ERR(acl))
276                 return PTR_ERR(acl);
277         if (acl) {
278                 int error = posix_acl_permission(inode, acl, mask);
279                 posix_acl_release(acl);
280                 return error;
281         }
282 #endif
283
284         return -EAGAIN;
285 }
286
287 /*
288  * This does the basic permission checking
289  */
290 static int acl_permission_check(struct inode *inode, int mask)
291 {
292         unsigned int mode = inode->i_mode;
293
294         if (likely(uid_eq(current_fsuid(), inode->i_uid)))
295                 mode >>= 6;
296         else {
297                 if (IS_POSIXACL(inode) && (mode & S_IRWXG)) {
298                         int error = check_acl(inode, mask);
299                         if (error != -EAGAIN)
300                                 return error;
301                 }
302
303                 if (in_group_p(inode->i_gid))
304                         mode >>= 3;
305         }
306
307         /*
308          * If the DACs are ok we don't need any capability check.
309          */
310         if ((mask & ~mode & (MAY_READ | MAY_WRITE | MAY_EXEC)) == 0)
311                 return 0;
312         return -EACCES;
313 }
314
315 /**
316  * generic_permission -  check for access rights on a Posix-like filesystem
317  * @inode:      inode to check access rights for
318  * @mask:       right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC, ...)
319  *
320  * Used to check for read/write/execute permissions on a file.
321  * We use "fsuid" for this, letting us set arbitrary permissions
322  * for filesystem access without changing the "normal" uids which
323  * are used for other things.
324  *
325  * generic_permission is rcu-walk aware. It returns -ECHILD in case an rcu-walk
326  * request cannot be satisfied (eg. requires blocking or too much complexity).
327  * It would then be called again in ref-walk mode.
328  */
329 int generic_permission(struct inode *inode, int mask)
330 {
331         int ret;
332
333         /*
334          * Do the basic permission checks.
335          */
336         ret = acl_permission_check(inode, mask);
337         if (ret != -EACCES)
338                 return ret;
339
340         if (S_ISDIR(inode->i_mode)) {
341                 /* DACs are overridable for directories */
342                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
343                         return 0;
344                 if (!(mask & MAY_WRITE))
345                         if (capable_wrt_inode_uidgid(inode,
346                                                      CAP_DAC_READ_SEARCH))
347                                 return 0;
348                 return -EACCES;
349         }
350         /*
351          * Read/write DACs are always overridable.
352          * Executable DACs are overridable when there is
353          * at least one exec bit set.
354          */
355         if (!(mask & MAY_EXEC) || (inode->i_mode & S_IXUGO))
356                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
357                         return 0;
358
359         /*
360          * Searching includes executable on directories, else just read.
361          */
362         mask &= MAY_READ | MAY_WRITE | MAY_EXEC;
363         if (mask == MAY_READ)
364                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_READ_SEARCH))
365                         return 0;
366
367         return -EACCES;
368 }
369 EXPORT_SYMBOL(generic_permission);
370
371 /*
372  * We _really_ want to just do "generic_permission()" without
373  * even looking at the inode->i_op values. So we keep a cache
374  * flag in inode->i_opflags, that says "this has not special
375  * permission function, use the fast case".
376  */
377 static inline int do_inode_permission(struct inode *inode, int mask)
378 {
379         if (unlikely(!(inode->i_opflags & IOP_FASTPERM))) {
380                 if (likely(inode->i_op->permission))
381                         return inode->i_op->permission(inode, mask);
382
383                 /* This gets set once for the inode lifetime */
384                 spin_lock(&inode->i_lock);
385                 inode->i_opflags |= IOP_FASTPERM;
386                 spin_unlock(&inode->i_lock);
387         }
388         return generic_permission(inode, mask);
389 }
390
391 /**
392  * __inode_permission - Check for access rights to a given inode
393  * @inode: Inode to check permission on
394  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
395  *
396  * Check for read/write/execute permissions on an inode.
397  *
398  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
399  *
400  * This does not check for a read-only file system.  You probably want
401  * inode_permission().
402  */
403 int __inode_permission(struct inode *inode, int mask)
404 {
405         int retval;
406
407         if (unlikely(mask & MAY_WRITE)) {
408                 /*
409                  * Nobody gets write access to an immutable file.
410                  */
411                 if (IS_IMMUTABLE(inode))
412                         return -EACCES;
413         }
414
415         retval = do_inode_permission(inode, mask);
416         if (retval)
417                 return retval;
418
419         retval = devcgroup_inode_permission(inode, mask);
420         if (retval)
421                 return retval;
422
423         return security_inode_permission(inode, mask);
424 }
425 EXPORT_SYMBOL(__inode_permission);
426
427 /**
428  * sb_permission - Check superblock-level permissions
429  * @sb: Superblock of inode to check permission on
430  * @inode: Inode to check permission on
431  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
432  *
433  * Separate out file-system wide checks from inode-specific permission checks.
434  */
435 static int sb_permission(struct super_block *sb, struct inode *inode, int mask)
436 {
437         if (unlikely(mask & MAY_WRITE)) {
438                 umode_t mode = inode->i_mode;
439
440                 /* Nobody gets write access to a read-only fs. */
441                 if ((sb->s_flags & MS_RDONLY) &&
442                     (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
443                         return -EROFS;
444         }
445         return 0;
446 }
447
448 /**
449  * inode_permission - Check for access rights to a given inode
450  * @inode: Inode to check permission on
451  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
452  *
453  * Check for read/write/execute permissions on an inode.  We use fs[ug]id for
454  * this, letting us set arbitrary permissions for filesystem access without
455  * changing the "normal" UIDs which are used for other things.
456  *
457  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
458  */
459 int inode_permission(struct inode *inode, int mask)
460 {
461         int retval;
462
463         retval = sb_permission(inode->i_sb, inode, mask);
464         if (retval)
465                 return retval;
466         return __inode_permission(inode, mask);
467 }
468 EXPORT_SYMBOL(inode_permission);
469
470 /**
471  * path_get - get a reference to a path
472  * @path: path to get the reference to
473  *
474  * Given a path increment the reference count to the dentry and the vfsmount.
475  */
476 void path_get(const struct path *path)
477 {
478         mntget(path->mnt);
479         dget(path->dentry);
480 }
481 EXPORT_SYMBOL(path_get);
482
483 /**
484  * path_put - put a reference to a path
485  * @path: path to put the reference to
486  *
487  * Given a path decrement the reference count to the dentry and the vfsmount.
488  */
489 void path_put(const struct path *path)
490 {
491         dput(path->dentry);
492         mntput(path->mnt);
493 }
494 EXPORT_SYMBOL(path_put);
495
496 #define EMBEDDED_LEVELS 2
497 struct nameidata {
498         struct path     path;
499         struct qstr     last;
500         struct path     root;
501         struct inode    *inode; /* path.dentry.d_inode */
502         unsigned int    flags;
503         unsigned        seq, m_seq;
504         int             last_type;
505         unsigned        depth;
506         int             total_link_count;
507         struct saved {
508                 struct path link;
509                 struct delayed_call done;
510                 const char *name;
511                 unsigned seq;
512         } *stack, internal[EMBEDDED_LEVELS];
513         struct filename *name;
514         struct nameidata *saved;
515         struct inode    *link_inode;
516         unsigned        root_seq;
517         int             dfd;
518 };
519
520 static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
521 {
522         struct nameidata *old = current->nameidata;
523         p->stack = p->internal;
524         p->dfd = dfd;
525         p->name = name;
526         p->total_link_count = old ? old->total_link_count : 0;
527         p->saved = old;
528         current->nameidata = p;
529 }
530
531 static void restore_nameidata(void)
532 {
533         struct nameidata *now = current->nameidata, *old = now->saved;
534
535         current->nameidata = old;
536         if (old)
537                 old->total_link_count = now->total_link_count;
538         if (now->stack != now->internal)
539                 kfree(now->stack);
540 }
541
542 static int __nd_alloc_stack(struct nameidata *nd)
543 {
544         struct saved *p;
545
546         if (nd->flags & LOOKUP_RCU) {
547                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
548                                   GFP_ATOMIC);
549                 if (unlikely(!p))
550                         return -ECHILD;
551         } else {
552                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
553                                   GFP_KERNEL);
554                 if (unlikely(!p))
555                         return -ENOMEM;
556         }
557         memcpy(p, nd->internal, sizeof(nd->internal));
558         nd->stack = p;
559         return 0;
560 }
561
562 /**
563  * path_connected - Verify that a path->dentry is below path->mnt.mnt_root
564  * @path: nameidate to verify
565  *
566  * Rename can sometimes move a file or directory outside of a bind
567  * mount, path_connected allows those cases to be detected.
568  */
569 static bool path_connected(const struct path *path)
570 {
571         struct vfsmount *mnt = path->mnt;
572
573         /* Only bind mounts can have disconnected paths */
574         if (mnt->mnt_root == mnt->mnt_sb->s_root)
575                 return true;
576
577         return is_subdir(path->dentry, mnt->mnt_root);
578 }
579
580 static inline int nd_alloc_stack(struct nameidata *nd)
581 {
582         if (likely(nd->depth != EMBEDDED_LEVELS))
583                 return 0;
584         if (likely(nd->stack != nd->internal))
585                 return 0;
586         return __nd_alloc_stack(nd);
587 }
588
589 static void drop_links(struct nameidata *nd)
590 {
591         int i = nd->depth;
592         while (i--) {
593                 struct saved *last = nd->stack + i;
594                 do_delayed_call(&last->done);
595                 clear_delayed_call(&last->done);
596         }
597 }
598
599 static void terminate_walk(struct nameidata *nd)
600 {
601         drop_links(nd);
602         if (!(nd->flags & LOOKUP_RCU)) {
603                 int i;
604                 path_put(&nd->path);
605                 for (i = 0; i < nd->depth; i++)
606                         path_put(&nd->stack[i].link);
607                 if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
608                         path_put(&nd->root);
609                         nd->root.mnt = NULL;
610                 }
611         } else {
612                 nd->flags &= ~LOOKUP_RCU;
613                 if (!(nd->flags & LOOKUP_ROOT))
614                         nd->root.mnt = NULL;
615                 rcu_read_unlock();
616         }
617         nd->depth = 0;
618 }
619
620 /* path_put is needed afterwards regardless of success or failure */
621 static bool legitimize_path(struct nameidata *nd,
622                             struct path *path, unsigned seq)
623 {
624         int res = __legitimize_mnt(path->mnt, nd->m_seq);
625         if (unlikely(res)) {
626                 if (res > 0)
627                         path->mnt = NULL;
628                 path->dentry = NULL;
629                 return false;
630         }
631         if (unlikely(!lockref_get_not_dead(&path->dentry->d_lockref))) {
632                 path->dentry = NULL;
633                 return false;
634         }
635         return !read_seqcount_retry(&path->dentry->d_seq, seq);
636 }
637
638 static bool legitimize_links(struct nameidata *nd)
639 {
640         int i;
641         for (i = 0; i < nd->depth; i++) {
642                 struct saved *last = nd->stack + i;
643                 if (unlikely(!legitimize_path(nd, &last->link, last->seq))) {
644                         drop_links(nd);
645                         nd->depth = i + 1;
646                         return false;
647                 }
648         }
649         return true;
650 }
651
652 /*
653  * Path walking has 2 modes, rcu-walk and ref-walk (see
654  * Documentation/filesystems/path-lookup.txt).  In situations when we can't
655  * continue in RCU mode, we attempt to drop out of rcu-walk mode and grab
656  * normal reference counts on dentries and vfsmounts to transition to ref-walk
657  * mode.  Refcounts are grabbed at the last known good point before rcu-walk
658  * got stuck, so ref-walk may continue from there. If this is not successful
659  * (eg. a seqcount has changed), then failure is returned and it's up to caller
660  * to restart the path walk from the beginning in ref-walk mode.
661  */
662
663 /**
664  * unlazy_walk - try to switch to ref-walk mode.
665  * @nd: nameidata pathwalk data
666  * @dentry: child of nd->path.dentry or NULL
667  * @seq: seq number to check dentry against
668  * Returns: 0 on success, -ECHILD on failure
669  *
670  * unlazy_walk attempts to legitimize the current nd->path, nd->root and dentry
671  * for ref-walk mode.  @dentry must be a path found by a do_lookup call on
672  * @nd or NULL.  Must be called from rcu-walk context.
673  * Nothing should touch nameidata between unlazy_walk() failure and
674  * terminate_walk().
675  */
676 static int unlazy_walk(struct nameidata *nd, struct dentry *dentry, unsigned seq)
677 {
678         struct dentry *parent = nd->path.dentry;
679
680         BUG_ON(!(nd->flags & LOOKUP_RCU));
681
682         nd->flags &= ~LOOKUP_RCU;
683         if (unlikely(!legitimize_links(nd)))
684                 goto out2;
685         if (unlikely(!legitimize_mnt(nd->path.mnt, nd->m_seq)))
686                 goto out2;
687         if (unlikely(!lockref_get_not_dead(&parent->d_lockref)))
688                 goto out1;
689
690         /*
691          * For a negative lookup, the lookup sequence point is the parents
692          * sequence point, and it only needs to revalidate the parent dentry.
693          *
694          * For a positive lookup, we need to move both the parent and the
695          * dentry from the RCU domain to be properly refcounted. And the
696          * sequence number in the dentry validates *both* dentry counters,
697          * since we checked the sequence number of the parent after we got
698          * the child sequence number. So we know the parent must still
699          * be valid if the child sequence number is still valid.
700          */
701         if (!dentry) {
702                 if (read_seqcount_retry(&parent->d_seq, nd->seq))
703                         goto out;
704                 BUG_ON(nd->inode != parent->d_inode);
705         } else {
706                 if (!lockref_get_not_dead(&dentry->d_lockref))
707                         goto out;
708                 if (read_seqcount_retry(&dentry->d_seq, seq))
709                         goto drop_dentry;
710         }
711
712         /*
713          * Sequence counts matched. Now make sure that the root is
714          * still valid and get it if required.
715          */
716         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
717                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq))) {
718                         rcu_read_unlock();
719                         dput(dentry);
720                         return -ECHILD;
721                 }
722         }
723
724         rcu_read_unlock();
725         return 0;
726
727 drop_dentry:
728         rcu_read_unlock();
729         dput(dentry);
730         goto drop_root_mnt;
731 out2:
732         nd->path.mnt = NULL;
733 out1:
734         nd->path.dentry = NULL;
735 out:
736         rcu_read_unlock();
737 drop_root_mnt:
738         if (!(nd->flags & LOOKUP_ROOT))
739                 nd->root.mnt = NULL;
740         return -ECHILD;
741 }
742
743 static int unlazy_link(struct nameidata *nd, struct path *link, unsigned seq)
744 {
745         if (unlikely(!legitimize_path(nd, link, seq))) {
746                 drop_links(nd);
747                 nd->depth = 0;
748                 nd->flags &= ~LOOKUP_RCU;
749                 nd->path.mnt = NULL;
750                 nd->path.dentry = NULL;
751                 if (!(nd->flags & LOOKUP_ROOT))
752                         nd->root.mnt = NULL;
753                 rcu_read_unlock();
754         } else if (likely(unlazy_walk(nd, NULL, 0)) == 0) {
755                 return 0;
756         }
757         path_put(link);
758         return -ECHILD;
759 }
760
761 static inline int d_revalidate(struct dentry *dentry, unsigned int flags)
762 {
763         return dentry->d_op->d_revalidate(dentry, flags);
764 }
765
766 /**
767  * complete_walk - successful completion of path walk
768  * @nd:  pointer nameidata
769  *
770  * If we had been in RCU mode, drop out of it and legitimize nd->path.
771  * Revalidate the final result, unless we'd already done that during
772  * the path walk or the filesystem doesn't ask for it.  Return 0 on
773  * success, -error on failure.  In case of failure caller does not
774  * need to drop nd->path.
775  */
776 static int complete_walk(struct nameidata *nd)
777 {
778         struct dentry *dentry = nd->path.dentry;
779         int status;
780
781         if (nd->flags & LOOKUP_RCU) {
782                 if (!(nd->flags & LOOKUP_ROOT))
783                         nd->root.mnt = NULL;
784                 if (unlikely(unlazy_walk(nd, NULL, 0)))
785                         return -ECHILD;
786         }
787
788         if (likely(!(nd->flags & LOOKUP_JUMPED)))
789                 return 0;
790
791         if (likely(!(dentry->d_flags & DCACHE_OP_WEAK_REVALIDATE)))
792                 return 0;
793
794         status = dentry->d_op->d_weak_revalidate(dentry, nd->flags);
795         if (status > 0)
796                 return 0;
797
798         if (!status)
799                 status = -ESTALE;
800
801         return status;
802 }
803
804 static void set_root(struct nameidata *nd)
805 {
806         struct fs_struct *fs = current->fs;
807
808         if (nd->flags & LOOKUP_RCU) {
809                 unsigned seq;
810
811                 do {
812                         seq = read_seqcount_begin(&fs->seq);
813                         nd->root = fs->root;
814                         nd->root_seq = __read_seqcount_begin(&nd->root.dentry->d_seq);
815                 } while (read_seqcount_retry(&fs->seq, seq));
816         } else {
817                 get_fs_root(fs, &nd->root);
818         }
819 }
820
821 static void path_put_conditional(struct path *path, struct nameidata *nd)
822 {
823         dput(path->dentry);
824         if (path->mnt != nd->path.mnt)
825                 mntput(path->mnt);
826 }
827
828 static inline void path_to_nameidata(const struct path *path,
829                                         struct nameidata *nd)
830 {
831         if (!(nd->flags & LOOKUP_RCU)) {
832                 dput(nd->path.dentry);
833                 if (nd->path.mnt != path->mnt)
834                         mntput(nd->path.mnt);
835         }
836         nd->path.mnt = path->mnt;
837         nd->path.dentry = path->dentry;
838 }
839
840 static int nd_jump_root(struct nameidata *nd)
841 {
842         if (nd->flags & LOOKUP_RCU) {
843                 struct dentry *d;
844                 nd->path = nd->root;
845                 d = nd->path.dentry;
846                 nd->inode = d->d_inode;
847                 nd->seq = nd->root_seq;
848                 if (unlikely(read_seqcount_retry(&d->d_seq, nd->seq)))
849                         return -ECHILD;
850         } else {
851                 path_put(&nd->path);
852                 nd->path = nd->root;
853                 path_get(&nd->path);
854                 nd->inode = nd->path.dentry->d_inode;
855         }
856         nd->flags |= LOOKUP_JUMPED;
857         return 0;
858 }
859
860 /*
861  * Helper to directly jump to a known parsed path from ->get_link,
862  * caller must have taken a reference to path beforehand.
863  */
864 void nd_jump_link(struct path *path)
865 {
866         struct nameidata *nd = current->nameidata;
867         path_put(&nd->path);
868
869         nd->path = *path;
870         nd->inode = nd->path.dentry->d_inode;
871         nd->flags |= LOOKUP_JUMPED;
872 }
873
874 static inline void put_link(struct nameidata *nd)
875 {
876         struct saved *last = nd->stack + --nd->depth;
877         do_delayed_call(&last->done);
878         if (!(nd->flags & LOOKUP_RCU))
879                 path_put(&last->link);
880 }
881
882 int sysctl_protected_symlinks __read_mostly = 0;
883 int sysctl_protected_hardlinks __read_mostly = 0;
884
885 /**
886  * may_follow_link - Check symlink following for unsafe situations
887  * @nd: nameidata pathwalk data
888  *
889  * In the case of the sysctl_protected_symlinks sysctl being enabled,
890  * CAP_DAC_OVERRIDE needs to be specifically ignored if the symlink is
891  * in a sticky world-writable directory. This is to protect privileged
892  * processes from failing races against path names that may change out
893  * from under them by way of other users creating malicious symlinks.
894  * It will permit symlinks to be followed only when outside a sticky
895  * world-writable directory, or when the uid of the symlink and follower
896  * match, or when the directory owner matches the symlink's owner.
897  *
898  * Returns 0 if following the symlink is allowed, -ve on error.
899  */
900 static inline int may_follow_link(struct nameidata *nd)
901 {
902         const struct inode *inode;
903         const struct inode *parent;
904
905         if (!sysctl_protected_symlinks)
906                 return 0;
907
908         /* Allowed if owner and follower match. */
909         inode = nd->link_inode;
910         if (uid_eq(current_cred()->fsuid, inode->i_uid))
911                 return 0;
912
913         /* Allowed if parent directory not sticky and world-writable. */
914         parent = nd->inode;
915         if ((parent->i_mode & (S_ISVTX|S_IWOTH)) != (S_ISVTX|S_IWOTH))
916                 return 0;
917
918         /* Allowed if parent directory and link owner match. */
919         if (uid_eq(parent->i_uid, inode->i_uid))
920                 return 0;
921
922         if (nd->flags & LOOKUP_RCU)
923                 return -ECHILD;
924
925         audit_log_link_denied("follow_link", &nd->stack[0].link);
926         return -EACCES;
927 }
928
929 /**
930  * safe_hardlink_source - Check for safe hardlink conditions
931  * @inode: the source inode to hardlink from
932  *
933  * Return false if at least one of the following conditions:
934  *    - inode is not a regular file
935  *    - inode is setuid
936  *    - inode is setgid and group-exec
937  *    - access failure for read and write
938  *
939  * Otherwise returns true.
940  */
941 static bool safe_hardlink_source(struct inode *inode)
942 {
943         umode_t mode = inode->i_mode;
944
945         /* Special files should not get pinned to the filesystem. */
946         if (!S_ISREG(mode))
947                 return false;
948
949         /* Setuid files should not get pinned to the filesystem. */
950         if (mode & S_ISUID)
951                 return false;
952
953         /* Executable setgid files should not get pinned to the filesystem. */
954         if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP))
955                 return false;
956
957         /* Hardlinking to unreadable or unwritable sources is dangerous. */
958         if (inode_permission(inode, MAY_READ | MAY_WRITE))
959                 return false;
960
961         return true;
962 }
963
964 /**
965  * may_linkat - Check permissions for creating a hardlink
966  * @link: the source to hardlink from
967  *
968  * Block hardlink when all of:
969  *  - sysctl_protected_hardlinks enabled
970  *  - fsuid does not match inode
971  *  - hardlink source is unsafe (see safe_hardlink_source() above)
972  *  - not CAP_FOWNER in a namespace with the inode owner uid mapped
973  *
974  * Returns 0 if successful, -ve on error.
975  */
976 static int may_linkat(struct path *link)
977 {
978         struct inode *inode;
979
980         if (!sysctl_protected_hardlinks)
981                 return 0;
982
983         inode = link->dentry->d_inode;
984
985         /* Source inode owner (or CAP_FOWNER) can hardlink all they like,
986          * otherwise, it must be a safe source.
987          */
988         if (inode_owner_or_capable(inode) || safe_hardlink_source(inode))
989                 return 0;
990
991         audit_log_link_denied("linkat", link);
992         return -EPERM;
993 }
994
995 static __always_inline
996 const char *get_link(struct nameidata *nd)
997 {
998         struct saved *last = nd->stack + nd->depth - 1;
999         struct dentry *dentry = last->link.dentry;
1000         struct inode *inode = nd->link_inode;
1001         int error;
1002         const char *res;
1003
1004         if (!(nd->flags & LOOKUP_RCU)) {
1005                 touch_atime(&last->link);
1006                 cond_resched();
1007         } else if (atime_needs_update(&last->link, inode)) {
1008                 if (unlikely(unlazy_walk(nd, NULL, 0)))
1009                         return ERR_PTR(-ECHILD);
1010                 touch_atime(&last->link);
1011         }
1012
1013         error = security_inode_follow_link(dentry, inode,
1014                                            nd->flags & LOOKUP_RCU);
1015         if (unlikely(error))
1016                 return ERR_PTR(error);
1017
1018         nd->last_type = LAST_BIND;
1019         res = inode->i_link;
1020         if (!res) {
1021                 const char * (*get)(struct dentry *, struct inode *,
1022                                 struct delayed_call *);
1023                 get = inode->i_op->get_link;
1024                 if (nd->flags & LOOKUP_RCU) {
1025                         res = get(NULL, inode, &last->done);
1026                         if (res == ERR_PTR(-ECHILD)) {
1027                                 if (unlikely(unlazy_walk(nd, NULL, 0)))
1028                                         return ERR_PTR(-ECHILD);
1029                                 res = get(dentry, inode, &last->done);
1030                         }
1031                 } else {
1032                         res = get(dentry, inode, &last->done);
1033                 }
1034                 if (IS_ERR_OR_NULL(res))
1035                         return res;
1036         }
1037         if (*res == '/') {
1038                 if (!nd->root.mnt)
1039                         set_root(nd);
1040                 if (unlikely(nd_jump_root(nd)))
1041                         return ERR_PTR(-ECHILD);
1042                 while (unlikely(*++res == '/'))
1043                         ;
1044         }
1045         if (!*res)
1046                 res = NULL;
1047         return res;
1048 }
1049
1050 /*
1051  * follow_up - Find the mountpoint of path's vfsmount
1052  *
1053  * Given a path, find the mountpoint of its source file system.
1054  * Replace @path with the path of the mountpoint in the parent mount.
1055  * Up is towards /.
1056  *
1057  * Return 1 if we went up a level and 0 if we were already at the
1058  * root.
1059  */
1060 int follow_up(struct path *path)
1061 {
1062         struct mount *mnt = real_mount(path->mnt);
1063         struct mount *parent;
1064         struct dentry *mountpoint;
1065
1066         read_seqlock_excl(&mount_lock);
1067         parent = mnt->mnt_parent;
1068         if (parent == mnt) {
1069                 read_sequnlock_excl(&mount_lock);
1070                 return 0;
1071         }
1072         mntget(&parent->mnt);
1073         mountpoint = dget(mnt->mnt_mountpoint);
1074         read_sequnlock_excl(&mount_lock);
1075         dput(path->dentry);
1076         path->dentry = mountpoint;
1077         mntput(path->mnt);
1078         path->mnt = &parent->mnt;
1079         return 1;
1080 }
1081 EXPORT_SYMBOL(follow_up);
1082
1083 /*
1084  * Perform an automount
1085  * - return -EISDIR to tell follow_managed() to stop and return the path we
1086  *   were called with.
1087  */
1088 static int follow_automount(struct path *path, struct nameidata *nd,
1089                             bool *need_mntput)
1090 {
1091         struct vfsmount *mnt;
1092         int err;
1093
1094         if (!path->dentry->d_op || !path->dentry->d_op->d_automount)
1095                 return -EREMOTE;
1096
1097         /* We don't want to mount if someone's just doing a stat -
1098          * unless they're stat'ing a directory and appended a '/' to
1099          * the name.
1100          *
1101          * We do, however, want to mount if someone wants to open or
1102          * create a file of any type under the mountpoint, wants to
1103          * traverse through the mountpoint or wants to open the
1104          * mounted directory.  Also, autofs may mark negative dentries
1105          * as being automount points.  These will need the attentions
1106          * of the daemon to instantiate them before they can be used.
1107          */
1108         if (!(nd->flags & (LOOKUP_PARENT | LOOKUP_DIRECTORY |
1109                            LOOKUP_OPEN | LOOKUP_CREATE | LOOKUP_AUTOMOUNT)) &&
1110             path->dentry->d_inode)
1111                 return -EISDIR;
1112
1113         nd->total_link_count++;
1114         if (nd->total_link_count >= 40)
1115                 return -ELOOP;
1116
1117         mnt = path->dentry->d_op->d_automount(path);
1118         if (IS_ERR(mnt)) {
1119                 /*
1120                  * The filesystem is allowed to return -EISDIR here to indicate
1121                  * it doesn't want to automount.  For instance, autofs would do
1122                  * this so that its userspace daemon can mount on this dentry.
1123                  *
1124                  * However, we can only permit this if it's a terminal point in
1125                  * the path being looked up; if it wasn't then the remainder of
1126                  * the path is inaccessible and we should say so.
1127                  */
1128                 if (PTR_ERR(mnt) == -EISDIR && (nd->flags & LOOKUP_PARENT))
1129                         return -EREMOTE;
1130                 return PTR_ERR(mnt);
1131         }
1132
1133         if (!mnt) /* mount collision */
1134                 return 0;
1135
1136         if (!*need_mntput) {
1137                 /* lock_mount() may release path->mnt on error */
1138                 mntget(path->mnt);
1139                 *need_mntput = true;
1140         }
1141         err = finish_automount(mnt, path);
1142
1143         switch (err) {
1144         case -EBUSY:
1145                 /* Someone else made a mount here whilst we were busy */
1146                 return 0;
1147         case 0:
1148                 path_put(path);
1149                 path->mnt = mnt;
1150                 path->dentry = dget(mnt->mnt_root);
1151                 return 0;
1152         default:
1153                 return err;
1154         }
1155
1156 }
1157
1158 /*
1159  * Handle a dentry that is managed in some way.
1160  * - Flagged for transit management (autofs)
1161  * - Flagged as mountpoint
1162  * - Flagged as automount point
1163  *
1164  * This may only be called in refwalk mode.
1165  *
1166  * Serialization is taken care of in namespace.c
1167  */
1168 static int follow_managed(struct path *path, struct nameidata *nd)
1169 {
1170         struct vfsmount *mnt = path->mnt; /* held by caller, must be left alone */
1171         unsigned managed;
1172         bool need_mntput = false;
1173         int ret = 0;
1174
1175         /* Given that we're not holding a lock here, we retain the value in a
1176          * local variable for each dentry as we look at it so that we don't see
1177          * the components of that value change under us */
1178         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1179                managed &= DCACHE_MANAGED_DENTRY,
1180                unlikely(managed != 0)) {
1181                 /* Allow the filesystem to manage the transit without i_mutex
1182                  * being held. */
1183                 if (managed & DCACHE_MANAGE_TRANSIT) {
1184                         BUG_ON(!path->dentry->d_op);
1185                         BUG_ON(!path->dentry->d_op->d_manage);
1186                         ret = path->dentry->d_op->d_manage(path->dentry, false);
1187                         if (ret < 0)
1188                                 break;
1189                 }
1190
1191                 /* Transit to a mounted filesystem. */
1192                 if (managed & DCACHE_MOUNTED) {
1193                         struct vfsmount *mounted = lookup_mnt(path);
1194                         if (mounted) {
1195                                 dput(path->dentry);
1196                                 if (need_mntput)
1197                                         mntput(path->mnt);
1198                                 path->mnt = mounted;
1199                                 path->dentry = dget(mounted->mnt_root);
1200                                 need_mntput = true;
1201                                 continue;
1202                         }
1203
1204                         /* Something is mounted on this dentry in another
1205                          * namespace and/or whatever was mounted there in this
1206                          * namespace got unmounted before lookup_mnt() could
1207                          * get it */
1208                 }
1209
1210                 /* Handle an automount point */
1211                 if (managed & DCACHE_NEED_AUTOMOUNT) {
1212                         ret = follow_automount(path, nd, &need_mntput);
1213                         if (ret < 0)
1214                                 break;
1215                         continue;
1216                 }
1217
1218                 /* We didn't change the current path point */
1219                 break;
1220         }
1221
1222         if (need_mntput && path->mnt == mnt)
1223                 mntput(path->mnt);
1224         if (ret == -EISDIR || !ret)
1225                 ret = 1;
1226         if (need_mntput)
1227                 nd->flags |= LOOKUP_JUMPED;
1228         if (unlikely(ret < 0))
1229                 path_put_conditional(path, nd);
1230         return ret;
1231 }
1232
1233 int follow_down_one(struct path *path)
1234 {
1235         struct vfsmount *mounted;
1236
1237         mounted = lookup_mnt(path);
1238         if (mounted) {
1239                 dput(path->dentry);
1240                 mntput(path->mnt);
1241                 path->mnt = mounted;
1242                 path->dentry = dget(mounted->mnt_root);
1243                 return 1;
1244         }
1245         return 0;
1246 }
1247 EXPORT_SYMBOL(follow_down_one);
1248
1249 static inline int managed_dentry_rcu(struct dentry *dentry)
1250 {
1251         return (dentry->d_flags & DCACHE_MANAGE_TRANSIT) ?
1252                 dentry->d_op->d_manage(dentry, true) : 0;
1253 }
1254
1255 /*
1256  * Try to skip to top of mountpoint pile in rcuwalk mode.  Fail if
1257  * we meet a managed dentry that would need blocking.
1258  */
1259 static bool __follow_mount_rcu(struct nameidata *nd, struct path *path,
1260                                struct inode **inode, unsigned *seqp)
1261 {
1262         for (;;) {
1263                 struct mount *mounted;
1264                 /*
1265                  * Don't forget we might have a non-mountpoint managed dentry
1266                  * that wants to block transit.
1267                  */
1268                 switch (managed_dentry_rcu(path->dentry)) {
1269                 case -ECHILD:
1270                 default:
1271                         return false;
1272                 case -EISDIR:
1273                         return true;
1274                 case 0:
1275                         break;
1276                 }
1277
1278                 if (!d_mountpoint(path->dentry))
1279                         return !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1280
1281                 mounted = __lookup_mnt(path->mnt, path->dentry);
1282                 if (!mounted)
1283                         break;
1284                 path->mnt = &mounted->mnt;
1285                 path->dentry = mounted->mnt.mnt_root;
1286                 nd->flags |= LOOKUP_JUMPED;
1287                 *seqp = read_seqcount_begin(&path->dentry->d_seq);
1288                 /*
1289                  * Update the inode too. We don't need to re-check the
1290                  * dentry sequence number here after this d_inode read,
1291                  * because a mount-point is always pinned.
1292                  */
1293                 *inode = path->dentry->d_inode;
1294         }
1295         return !read_seqretry(&mount_lock, nd->m_seq) &&
1296                 !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1297 }
1298
1299 static int follow_dotdot_rcu(struct nameidata *nd)
1300 {
1301         struct inode *inode = nd->inode;
1302
1303         while (1) {
1304                 if (path_equal(&nd->path, &nd->root))
1305                         break;
1306                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1307                         struct dentry *old = nd->path.dentry;
1308                         struct dentry *parent = old->d_parent;
1309                         unsigned seq;
1310
1311                         inode = parent->d_inode;
1312                         seq = read_seqcount_begin(&parent->d_seq);
1313                         if (unlikely(read_seqcount_retry(&old->d_seq, nd->seq)))
1314                                 return -ECHILD;
1315                         nd->path.dentry = parent;
1316                         nd->seq = seq;
1317                         if (unlikely(!path_connected(&nd->path)))
1318                                 return -ENOENT;
1319                         break;
1320                 } else {
1321                         struct mount *mnt = real_mount(nd->path.mnt);
1322                         struct mount *mparent = mnt->mnt_parent;
1323                         struct dentry *mountpoint = mnt->mnt_mountpoint;
1324                         struct inode *inode2 = mountpoint->d_inode;
1325                         unsigned seq = read_seqcount_begin(&mountpoint->d_seq);
1326                         if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1327                                 return -ECHILD;
1328                         if (&mparent->mnt == nd->path.mnt)
1329                                 break;
1330                         /* we know that mountpoint was pinned */
1331                         nd->path.dentry = mountpoint;
1332                         nd->path.mnt = &mparent->mnt;
1333                         inode = inode2;
1334                         nd->seq = seq;
1335                 }
1336         }
1337         while (unlikely(d_mountpoint(nd->path.dentry))) {
1338                 struct mount *mounted;
1339                 mounted = __lookup_mnt(nd->path.mnt, nd->path.dentry);
1340                 if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1341                         return -ECHILD;
1342                 if (!mounted)
1343                         break;
1344                 nd->path.mnt = &mounted->mnt;
1345                 nd->path.dentry = mounted->mnt.mnt_root;
1346                 inode = nd->path.dentry->d_inode;
1347                 nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
1348         }
1349         nd->inode = inode;
1350         return 0;
1351 }
1352
1353 /*
1354  * Follow down to the covering mount currently visible to userspace.  At each
1355  * point, the filesystem owning that dentry may be queried as to whether the
1356  * caller is permitted to proceed or not.
1357  */
1358 int follow_down(struct path *path)
1359 {
1360         unsigned managed;
1361         int ret;
1362
1363         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1364                unlikely(managed & DCACHE_MANAGED_DENTRY)) {
1365                 /* Allow the filesystem to manage the transit without i_mutex
1366                  * being held.
1367                  *
1368                  * We indicate to the filesystem if someone is trying to mount
1369                  * something here.  This gives autofs the chance to deny anyone
1370                  * other than its daemon the right to mount on its
1371                  * superstructure.
1372                  *
1373                  * The filesystem may sleep at this point.
1374                  */
1375                 if (managed & DCACHE_MANAGE_TRANSIT) {
1376                         BUG_ON(!path->dentry->d_op);
1377                         BUG_ON(!path->dentry->d_op->d_manage);
1378                         ret = path->dentry->d_op->d_manage(
1379                                 path->dentry, false);
1380                         if (ret < 0)
1381                                 return ret == -EISDIR ? 0 : ret;
1382                 }
1383
1384                 /* Transit to a mounted filesystem. */
1385                 if (managed & DCACHE_MOUNTED) {
1386                         struct vfsmount *mounted = lookup_mnt(path);
1387                         if (!mounted)
1388                                 break;
1389                         dput(path->dentry);
1390                         mntput(path->mnt);
1391                         path->mnt = mounted;
1392                         path->dentry = dget(mounted->mnt_root);
1393                         continue;
1394                 }
1395
1396                 /* Don't handle automount points here */
1397                 break;
1398         }
1399         return 0;
1400 }
1401 EXPORT_SYMBOL(follow_down);
1402
1403 /*
1404  * Skip to top of mountpoint pile in refwalk mode for follow_dotdot()
1405  */
1406 static void follow_mount(struct path *path)
1407 {
1408         while (d_mountpoint(path->dentry)) {
1409                 struct vfsmount *mounted = lookup_mnt(path);
1410                 if (!mounted)
1411                         break;
1412                 dput(path->dentry);
1413                 mntput(path->mnt);
1414                 path->mnt = mounted;
1415                 path->dentry = dget(mounted->mnt_root);
1416         }
1417 }
1418
1419 static int path_parent_directory(struct path *path)
1420 {
1421         struct dentry *old = path->dentry;
1422         /* rare case of legitimate dget_parent()... */
1423         path->dentry = dget_parent(path->dentry);
1424         dput(old);
1425         if (unlikely(!path_connected(path)))
1426                 return -ENOENT;
1427         return 0;
1428 }
1429
1430 static int follow_dotdot(struct nameidata *nd)
1431 {
1432         while(1) {
1433                 if (nd->path.dentry == nd->root.dentry &&
1434                     nd->path.mnt == nd->root.mnt) {
1435                         break;
1436                 }
1437                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1438                         int ret = path_parent_directory(&nd->path);
1439                         if (ret)
1440                                 return ret;
1441                         break;
1442                 }
1443                 if (!follow_up(&nd->path))
1444                         break;
1445         }
1446         follow_mount(&nd->path);
1447         nd->inode = nd->path.dentry->d_inode;
1448         return 0;
1449 }
1450
1451 /*
1452  * This looks up the name in dcache and possibly revalidates the found dentry.
1453  * NULL is returned if the dentry does not exist in the cache.
1454  */
1455 static struct dentry *lookup_dcache(const struct qstr *name,
1456                                     struct dentry *dir,
1457                                     unsigned int flags)
1458 {
1459         struct dentry *dentry;
1460         int error;
1461
1462         dentry = d_lookup(dir, name);
1463         if (dentry) {
1464                 if (dentry->d_flags & DCACHE_OP_REVALIDATE) {
1465                         error = d_revalidate(dentry, flags);
1466                         if (unlikely(error <= 0)) {
1467                                 if (!error)
1468                                         d_invalidate(dentry);
1469                                 dput(dentry);
1470                                 return ERR_PTR(error);
1471                         }
1472                 }
1473         }
1474         return dentry;
1475 }
1476
1477 /*
1478  * Call i_op->lookup on the dentry.  The dentry must be negative and
1479  * unhashed.
1480  *
1481  * dir->d_inode->i_mutex must be held
1482  */
1483 static struct dentry *lookup_real(struct inode *dir, struct dentry *dentry,
1484                                   unsigned int flags)
1485 {
1486         struct dentry *old;
1487
1488         /* Don't create child dentry for a dead directory. */
1489         if (unlikely(IS_DEADDIR(dir))) {
1490                 dput(dentry);
1491                 return ERR_PTR(-ENOENT);
1492         }
1493
1494         old = dir->i_op->lookup(dir, dentry, flags);
1495         if (unlikely(old)) {
1496                 dput(dentry);
1497                 dentry = old;
1498         }
1499         return dentry;
1500 }
1501
1502 static struct dentry *__lookup_hash(const struct qstr *name,
1503                 struct dentry *base, unsigned int flags)
1504 {
1505         struct dentry *dentry = lookup_dcache(name, base, flags);
1506
1507         if (dentry)
1508                 return dentry;
1509
1510         dentry = d_alloc(base, name);
1511         if (unlikely(!dentry))
1512                 return ERR_PTR(-ENOMEM);
1513
1514         return lookup_real(base->d_inode, dentry, flags);
1515 }
1516
1517 static int lookup_fast(struct nameidata *nd,
1518                        struct path *path, struct inode **inode,
1519                        unsigned *seqp)
1520 {
1521         struct vfsmount *mnt = nd->path.mnt;
1522         struct dentry *dentry, *parent = nd->path.dentry;
1523         int status = 1;
1524         int err;
1525
1526         /*
1527          * Rename seqlock is not required here because in the off chance
1528          * of a false negative due to a concurrent rename, the caller is
1529          * going to fall back to non-racy lookup.
1530          */
1531         if (nd->flags & LOOKUP_RCU) {
1532                 unsigned seq;
1533                 bool negative;
1534                 dentry = __d_lookup_rcu(parent, &nd->last, &seq);
1535                 if (unlikely(!dentry)) {
1536                         if (unlazy_walk(nd, NULL, 0))
1537                                 return -ECHILD;
1538                         return 0;
1539                 }
1540
1541                 /*
1542                  * This sequence count validates that the inode matches
1543                  * the dentry name information from lookup.
1544                  */
1545                 *inode = d_backing_inode(dentry);
1546                 negative = d_is_negative(dentry);
1547                 if (unlikely(read_seqcount_retry(&dentry->d_seq, seq)))
1548                         return -ECHILD;
1549
1550                 /*
1551                  * This sequence count validates that the parent had no
1552                  * changes while we did the lookup of the dentry above.
1553                  *
1554                  * The memory barrier in read_seqcount_begin of child is
1555                  *  enough, we can use __read_seqcount_retry here.
1556                  */
1557                 if (unlikely(__read_seqcount_retry(&parent->d_seq, nd->seq)))
1558                         return -ECHILD;
1559
1560                 *seqp = seq;
1561                 if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
1562                         status = d_revalidate(dentry, nd->flags);
1563                 if (unlikely(status <= 0)) {
1564                         if (unlazy_walk(nd, dentry, seq))
1565                                 return -ECHILD;
1566                         if (status == -ECHILD)
1567                                 status = d_revalidate(dentry, nd->flags);
1568                 } else {
1569                         /*
1570                          * Note: do negative dentry check after revalidation in
1571                          * case that drops it.
1572                          */
1573                         if (unlikely(negative))
1574                                 return -ENOENT;
1575                         path->mnt = mnt;
1576                         path->dentry = dentry;
1577                         if (likely(__follow_mount_rcu(nd, path, inode, seqp)))
1578                                 return 1;
1579                         if (unlazy_walk(nd, dentry, seq))
1580                                 return -ECHILD;
1581                 }
1582         } else {
1583                 dentry = __d_lookup(parent, &nd->last);
1584                 if (unlikely(!dentry))
1585                         return 0;
1586                 if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
1587                         status = d_revalidate(dentry, nd->flags);
1588         }
1589         if (unlikely(status <= 0)) {
1590                 if (!status)
1591                         d_invalidate(dentry);
1592                 dput(dentry);
1593                 return status;
1594         }
1595         if (unlikely(d_is_negative(dentry))) {
1596                 dput(dentry);
1597                 return -ENOENT;
1598         }
1599
1600         path->mnt = mnt;
1601         path->dentry = dentry;
1602         err = follow_managed(path, nd);
1603         if (likely(err > 0))
1604                 *inode = d_backing_inode(path->dentry);
1605         return err;
1606 }
1607
1608 /* Fast lookup failed, do it the slow way */
1609 static struct dentry *lookup_slow(const struct qstr *name,
1610                                   struct dentry *dir,
1611                                   unsigned int flags)
1612 {
1613         struct dentry *dentry = ERR_PTR(-ENOENT), *old;
1614         struct inode *inode = dir->d_inode;
1615         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
1616
1617         inode_lock_shared(inode);
1618         /* Don't go there if it's already dead */
1619         if (unlikely(IS_DEADDIR(inode)))
1620                 goto out;
1621 again:
1622         dentry = d_alloc_parallel(dir, name, &wq);
1623         if (IS_ERR(dentry))
1624                 goto out;
1625         if (unlikely(!d_in_lookup(dentry))) {
1626                 if ((dentry->d_flags & DCACHE_OP_REVALIDATE) &&
1627                     !(flags & LOOKUP_NO_REVAL)) {
1628                         int error = d_revalidate(dentry, flags);
1629                         if (unlikely(error <= 0)) {
1630                                 if (!error) {
1631                                         d_invalidate(dentry);
1632                                         dput(dentry);
1633                                         goto again;
1634                                 }
1635                                 dput(dentry);
1636                                 dentry = ERR_PTR(error);
1637                         }
1638                 }
1639         } else {
1640                 old = inode->i_op->lookup(inode, dentry, flags);
1641                 d_lookup_done(dentry);
1642                 if (unlikely(old)) {
1643                         dput(dentry);
1644                         dentry = old;
1645                 }
1646         }
1647 out:
1648         inode_unlock_shared(inode);
1649         return dentry;
1650 }
1651
1652 static inline int may_lookup(struct nameidata *nd)
1653 {
1654         if (nd->flags & LOOKUP_RCU) {
1655                 int err = inode_permission(nd->inode, MAY_EXEC|MAY_NOT_BLOCK);
1656                 if (err != -ECHILD)
1657                         return err;
1658                 if (unlazy_walk(nd, NULL, 0))
1659                         return -ECHILD;
1660         }
1661         return inode_permission(nd->inode, MAY_EXEC);
1662 }
1663
1664 static inline int handle_dots(struct nameidata *nd, int type)
1665 {
1666         if (type == LAST_DOTDOT) {
1667                 if (!nd->root.mnt)
1668                         set_root(nd);
1669                 if (nd->flags & LOOKUP_RCU) {
1670                         return follow_dotdot_rcu(nd);
1671                 } else
1672                         return follow_dotdot(nd);
1673         }
1674         return 0;
1675 }
1676
1677 static int pick_link(struct nameidata *nd, struct path *link,
1678                      struct inode *inode, unsigned seq)
1679 {
1680         int error;
1681         struct saved *last;
1682         if (unlikely(nd->total_link_count++ >= MAXSYMLINKS)) {
1683                 path_to_nameidata(link, nd);
1684                 return -ELOOP;
1685         }
1686         if (!(nd->flags & LOOKUP_RCU)) {
1687                 if (link->mnt == nd->path.mnt)
1688                         mntget(link->mnt);
1689         }
1690         error = nd_alloc_stack(nd);
1691         if (unlikely(error)) {
1692                 if (error == -ECHILD) {
1693                         if (unlikely(unlazy_link(nd, link, seq)))
1694                                 return -ECHILD;
1695                         error = nd_alloc_stack(nd);
1696                 }
1697                 if (error) {
1698                         path_put(link);
1699                         return error;
1700                 }
1701         }
1702
1703         last = nd->stack + nd->depth++;
1704         last->link = *link;
1705         clear_delayed_call(&last->done);
1706         nd->link_inode = inode;
1707         last->seq = seq;
1708         return 1;
1709 }
1710
1711 /*
1712  * Do we need to follow links? We _really_ want to be able
1713  * to do this check without having to look at inode->i_op,
1714  * so we keep a cache of "no, this doesn't need follow_link"
1715  * for the common case.
1716  */
1717 static inline int should_follow_link(struct nameidata *nd, struct path *link,
1718                                      int follow,
1719                                      struct inode *inode, unsigned seq)
1720 {
1721         if (likely(!d_is_symlink(link->dentry)))
1722                 return 0;
1723         if (!follow)
1724                 return 0;
1725         /* make sure that d_is_symlink above matches inode */
1726         if (nd->flags & LOOKUP_RCU) {
1727                 if (read_seqcount_retry(&link->dentry->d_seq, seq))
1728                         return -ECHILD;
1729         }
1730         return pick_link(nd, link, inode, seq);
1731 }
1732
1733 enum {WALK_GET = 1, WALK_PUT = 2};
1734
1735 static int walk_component(struct nameidata *nd, int flags)
1736 {
1737         struct path path;
1738         struct inode *inode;
1739         unsigned seq;
1740         int err;
1741         /*
1742          * "." and ".." are special - ".." especially so because it has
1743          * to be able to know about the current root directory and
1744          * parent relationships.
1745          */
1746         if (unlikely(nd->last_type != LAST_NORM)) {
1747                 err = handle_dots(nd, nd->last_type);
1748                 if (flags & WALK_PUT)
1749                         put_link(nd);
1750                 return err;
1751         }
1752         err = lookup_fast(nd, &path, &inode, &seq);
1753         if (unlikely(err <= 0)) {
1754                 if (err < 0)
1755                         return err;
1756                 path.dentry = lookup_slow(&nd->last, nd->path.dentry,
1757                                           nd->flags);
1758                 if (IS_ERR(path.dentry))
1759                         return PTR_ERR(path.dentry);
1760
1761                 path.mnt = nd->path.mnt;
1762                 err = follow_managed(&path, nd);
1763                 if (unlikely(err < 0))
1764                         return err;
1765
1766                 if (unlikely(d_is_negative(path.dentry))) {
1767                         path_to_nameidata(&path, nd);
1768                         return -ENOENT;
1769                 }
1770
1771                 seq = 0;        /* we are already out of RCU mode */
1772                 inode = d_backing_inode(path.dentry);
1773         }
1774
1775         if (flags & WALK_PUT)
1776                 put_link(nd);
1777         err = should_follow_link(nd, &path, flags & WALK_GET, inode, seq);
1778         if (unlikely(err))
1779                 return err;
1780         path_to_nameidata(&path, nd);
1781         nd->inode = inode;
1782         nd->seq = seq;
1783         return 0;
1784 }
1785
1786 /*
1787  * We can do the critical dentry name comparison and hashing
1788  * operations one word at a time, but we are limited to:
1789  *
1790  * - Architectures with fast unaligned word accesses. We could
1791  *   do a "get_unaligned()" if this helps and is sufficiently
1792  *   fast.
1793  *
1794  * - non-CONFIG_DEBUG_PAGEALLOC configurations (so that we
1795  *   do not trap on the (extremely unlikely) case of a page
1796  *   crossing operation.
1797  *
1798  * - Furthermore, we need an efficient 64-bit compile for the
1799  *   64-bit case in order to generate the "number of bytes in
1800  *   the final mask". Again, that could be replaced with a
1801  *   efficient population count instruction or similar.
1802  */
1803 #ifdef CONFIG_DCACHE_WORD_ACCESS
1804
1805 #include <asm/word-at-a-time.h>
1806
1807 #ifdef HASH_MIX
1808
1809 /* Architecture provides HASH_MIX and fold_hash() in <asm/hash.h> */
1810
1811 #elif defined(CONFIG_64BIT)
1812 /*
1813  * Register pressure in the mixing function is an issue, particularly
1814  * on 32-bit x86, but almost any function requires one state value and
1815  * one temporary.  Instead, use a function designed for two state values
1816  * and no temporaries.
1817  *
1818  * This function cannot create a collision in only two iterations, so
1819  * we have two iterations to achieve avalanche.  In those two iterations,
1820  * we have six layers of mixing, which is enough to spread one bit's
1821  * influence out to 2^6 = 64 state bits.
1822  *
1823  * Rotate constants are scored by considering either 64 one-bit input
1824  * deltas or 64*63/2 = 2016 two-bit input deltas, and finding the
1825  * probability of that delta causing a change to each of the 128 output
1826  * bits, using a sample of random initial states.
1827  *
1828  * The Shannon entropy of the computed probabilities is then summed
1829  * to produce a score.  Ideally, any input change has a 50% chance of
1830  * toggling any given output bit.
1831  *
1832  * Mixing scores (in bits) for (12,45):
1833  * Input delta: 1-bit      2-bit
1834  * 1 round:     713.3    42542.6
1835  * 2 rounds:   2753.7   140389.8
1836  * 3 rounds:   5954.1   233458.2
1837  * 4 rounds:   7862.6   256672.2
1838  * Perfect:    8192     258048
1839  *            (64*128) (64*63/2 * 128)
1840  */
1841 #define HASH_MIX(x, y, a)       \
1842         (       x ^= (a),       \
1843         y ^= x, x = rol64(x,12),\
1844         x += y, y = rol64(y,45),\
1845         y *= 9                  )
1846
1847 /*
1848  * Fold two longs into one 32-bit hash value.  This must be fast, but
1849  * latency isn't quite as critical, as there is a fair bit of additional
1850  * work done before the hash value is used.
1851  */
1852 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1853 {
1854         y ^= x * GOLDEN_RATIO_64;
1855         y *= GOLDEN_RATIO_64;
1856         return y >> 32;
1857 }
1858
1859 #else   /* 32-bit case */
1860
1861 /*
1862  * Mixing scores (in bits) for (7,20):
1863  * Input delta: 1-bit      2-bit
1864  * 1 round:     330.3     9201.6
1865  * 2 rounds:   1246.4    25475.4
1866  * 3 rounds:   1907.1    31295.1
1867  * 4 rounds:   2042.3    31718.6
1868  * Perfect:    2048      31744
1869  *            (32*64)   (32*31/2 * 64)
1870  */
1871 #define HASH_MIX(x, y, a)       \
1872         (       x ^= (a),       \
1873         y ^= x, x = rol32(x, 7),\
1874         x += y, y = rol32(y,20),\
1875         y *= 9                  )
1876
1877 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1878 {
1879         /* Use arch-optimized multiply if one exists */
1880         return __hash_32(y ^ __hash_32(x));
1881 }
1882
1883 #endif
1884
1885 /*
1886  * Return the hash of a string of known length.  This is carfully
1887  * designed to match hash_name(), which is the more critical function.
1888  * In particular, we must end by hashing a final word containing 0..7
1889  * payload bytes, to match the way that hash_name() iterates until it
1890  * finds the delimiter after the name.
1891  */
1892 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1893 {
1894         unsigned long a, x = 0, y = (unsigned long)salt;
1895
1896         for (;;) {
1897                 if (!len)
1898                         goto done;
1899                 a = load_unaligned_zeropad(name);
1900                 if (len < sizeof(unsigned long))
1901                         break;
1902                 HASH_MIX(x, y, a);
1903                 name += sizeof(unsigned long);
1904                 len -= sizeof(unsigned long);
1905         }
1906         x ^= a & bytemask_from_count(len);
1907 done:
1908         return fold_hash(x, y);
1909 }
1910 EXPORT_SYMBOL(full_name_hash);
1911
1912 /* Return the "hash_len" (hash and length) of a null-terminated string */
1913 u64 hashlen_string(const void *salt, const char *name)
1914 {
1915         unsigned long a = 0, x = 0, y = (unsigned long)salt;
1916         unsigned long adata, mask, len;
1917         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1918
1919         len = 0;
1920         goto inside;
1921
1922         do {
1923                 HASH_MIX(x, y, a);
1924                 len += sizeof(unsigned long);
1925 inside:
1926                 a = load_unaligned_zeropad(name+len);
1927         } while (!has_zero(a, &adata, &constants));
1928
1929         adata = prep_zero_mask(a, adata, &constants);
1930         mask = create_zero_mask(adata);
1931         x ^= a & zero_bytemask(mask);
1932
1933         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1934 }
1935 EXPORT_SYMBOL(hashlen_string);
1936
1937 /*
1938  * Calculate the length and hash of the path component, and
1939  * return the "hash_len" as the result.
1940  */
1941 static inline u64 hash_name(const void *salt, const char *name)
1942 {
1943         unsigned long a = 0, b, x = 0, y = (unsigned long)salt;
1944         unsigned long adata, bdata, mask, len;
1945         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1946
1947         len = 0;
1948         goto inside;
1949
1950         do {
1951                 HASH_MIX(x, y, a);
1952                 len += sizeof(unsigned long);
1953 inside:
1954                 a = load_unaligned_zeropad(name+len);
1955                 b = a ^ REPEAT_BYTE('/');
1956         } while (!(has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)));
1957
1958         adata = prep_zero_mask(a, adata, &constants);
1959         bdata = prep_zero_mask(b, bdata, &constants);
1960         mask = create_zero_mask(adata | bdata);
1961         x ^= a & zero_bytemask(mask);
1962
1963         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1964 }
1965
1966 #else   /* !CONFIG_DCACHE_WORD_ACCESS: Slow, byte-at-a-time version */
1967
1968 /* Return the hash of a string of known length */
1969 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1970 {
1971         unsigned long hash = init_name_hash(salt);
1972         while (len--)
1973                 hash = partial_name_hash((unsigned char)*name++, hash);
1974         return end_name_hash(hash);
1975 }
1976 EXPORT_SYMBOL(full_name_hash);
1977
1978 /* Return the "hash_len" (hash and length) of a null-terminated string */
1979 u64 hashlen_string(const void *salt, const char *name)
1980 {
1981         unsigned long hash = init_name_hash(salt);
1982         unsigned long len = 0, c;
1983
1984         c = (unsigned char)*name;
1985         while (c) {
1986                 len++;
1987                 hash = partial_name_hash(c, hash);
1988                 c = (unsigned char)name[len];
1989         }
1990         return hashlen_create(end_name_hash(hash), len);
1991 }
1992 EXPORT_SYMBOL(hashlen_string);
1993
1994 /*
1995  * We know there's a real path component here of at least
1996  * one character.
1997  */
1998 static inline u64 hash_name(const void *salt, const char *name)
1999 {
2000         unsigned long hash = init_name_hash(salt);
2001         unsigned long len = 0, c;
2002
2003         c = (unsigned char)*name;
2004         do {
2005                 len++;
2006                 hash = partial_name_hash(c, hash);
2007                 c = (unsigned char)name[len];
2008         } while (c && c != '/');
2009         return hashlen_create(end_name_hash(hash), len);
2010 }
2011
2012 #endif
2013
2014 /*
2015  * Name resolution.
2016  * This is the basic name resolution function, turning a pathname into
2017  * the final dentry. We expect 'base' to be positive and a directory.
2018  *
2019  * Returns 0 and nd will have valid dentry and mnt on success.
2020  * Returns error and drops reference to input namei data on failure.
2021  */
2022 static int link_path_walk(const char *name, struct nameidata *nd)
2023 {
2024         int err;
2025
2026         while (*name=='/')
2027                 name++;
2028         if (!*name)
2029                 return 0;
2030
2031         /* At this point we know we have a real path component. */
2032         for(;;) {
2033                 u64 hash_len;
2034                 int type;
2035
2036                 err = may_lookup(nd);
2037                 if (err)
2038                         return err;
2039
2040                 hash_len = hash_name(nd->path.dentry, name);
2041
2042                 type = LAST_NORM;
2043                 if (name[0] == '.') switch (hashlen_len(hash_len)) {
2044                         case 2:
2045                                 if (name[1] == '.') {
2046                                         type = LAST_DOTDOT;
2047                                         nd->flags |= LOOKUP_JUMPED;
2048                                 }
2049                                 break;
2050                         case 1:
2051                                 type = LAST_DOT;
2052                 }
2053                 if (likely(type == LAST_NORM)) {
2054                         struct dentry *parent = nd->path.dentry;
2055                         nd->flags &= ~LOOKUP_JUMPED;
2056                         if (unlikely(parent->d_flags & DCACHE_OP_HASH)) {
2057                                 struct qstr this = { { .hash_len = hash_len }, .name = name };
2058                                 err = parent->d_op->d_hash(parent, &this);
2059                                 if (err < 0)
2060                                         return err;
2061                                 hash_len = this.hash_len;
2062                                 name = this.name;
2063                         }
2064                 }
2065
2066                 nd->last.hash_len = hash_len;
2067                 nd->last.name = name;
2068                 nd->last_type = type;
2069
2070                 name += hashlen_len(hash_len);
2071                 if (!*name)
2072                         goto OK;
2073                 /*
2074                  * If it wasn't NUL, we know it was '/'. Skip that
2075                  * slash, and continue until no more slashes.
2076                  */
2077                 do {
2078                         name++;
2079                 } while (unlikely(*name == '/'));
2080                 if (unlikely(!*name)) {
2081 OK:
2082                         /* pathname body, done */
2083                         if (!nd->depth)
2084                                 return 0;
2085                         name = nd->stack[nd->depth - 1].name;
2086                         /* trailing symlink, done */
2087                         if (!name)
2088                                 return 0;
2089                         /* last component of nested symlink */
2090                         err = walk_component(nd, WALK_GET | WALK_PUT);
2091                 } else {
2092                         err = walk_component(nd, WALK_GET);
2093                 }
2094                 if (err < 0)
2095                         return err;
2096
2097                 if (err) {
2098                         const char *s = get_link(nd);
2099
2100                         if (IS_ERR(s))
2101                                 return PTR_ERR(s);
2102                         err = 0;
2103                         if (unlikely(!s)) {
2104                                 /* jumped */
2105                                 put_link(nd);
2106                         } else {
2107                                 nd->stack[nd->depth - 1].name = name;
2108                                 name = s;
2109                                 continue;
2110                         }
2111                 }
2112                 if (unlikely(!d_can_lookup(nd->path.dentry))) {
2113                         if (nd->flags & LOOKUP_RCU) {
2114                                 if (unlazy_walk(nd, NULL, 0))
2115                                         return -ECHILD;
2116                         }
2117                         return -ENOTDIR;
2118                 }
2119         }
2120 }
2121
2122 static const char *path_init(struct nameidata *nd, unsigned flags)
2123 {
2124         int retval = 0;
2125         const char *s = nd->name->name;
2126
2127         nd->last_type = LAST_ROOT; /* if there are only slashes... */
2128         nd->flags = flags | LOOKUP_JUMPED | LOOKUP_PARENT;
2129         nd->depth = 0;
2130         if (flags & LOOKUP_ROOT) {
2131                 struct dentry *root = nd->root.dentry;
2132                 struct inode *inode = root->d_inode;
2133                 if (*s) {
2134                         if (!d_can_lookup(root))
2135                                 return ERR_PTR(-ENOTDIR);
2136                         retval = inode_permission(inode, MAY_EXEC);
2137                         if (retval)
2138                                 return ERR_PTR(retval);
2139                 }
2140                 nd->path = nd->root;
2141                 nd->inode = inode;
2142                 if (flags & LOOKUP_RCU) {
2143                         rcu_read_lock();
2144                         nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2145                         nd->root_seq = nd->seq;
2146                         nd->m_seq = read_seqbegin(&mount_lock);
2147                 } else {
2148                         path_get(&nd->path);
2149                 }
2150                 return s;
2151         }
2152
2153         nd->root.mnt = NULL;
2154         nd->path.mnt = NULL;
2155         nd->path.dentry = NULL;
2156
2157         nd->m_seq = read_seqbegin(&mount_lock);
2158         if (*s == '/') {
2159                 if (flags & LOOKUP_RCU)
2160                         rcu_read_lock();
2161                 set_root(nd);
2162                 if (likely(!nd_jump_root(nd)))
2163                         return s;
2164                 nd->root.mnt = NULL;
2165                 rcu_read_unlock();
2166                 return ERR_PTR(-ECHILD);
2167         } else if (nd->dfd == AT_FDCWD) {
2168                 if (flags & LOOKUP_RCU) {
2169                         struct fs_struct *fs = current->fs;
2170                         unsigned seq;
2171
2172                         rcu_read_lock();
2173
2174                         do {
2175                                 seq = read_seqcount_begin(&fs->seq);
2176                                 nd->path = fs->pwd;
2177                                 nd->inode = nd->path.dentry->d_inode;
2178                                 nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2179                         } while (read_seqcount_retry(&fs->seq, seq));
2180                 } else {
2181                         get_fs_pwd(current->fs, &nd->path);
2182                         nd->inode = nd->path.dentry->d_inode;
2183                 }
2184                 return s;
2185         } else {
2186                 /* Caller must check execute permissions on the starting path component */
2187                 struct fd f = fdget_raw(nd->dfd);
2188                 struct dentry *dentry;
2189
2190                 if (!f.file)
2191                         return ERR_PTR(-EBADF);
2192
2193                 dentry = f.file->f_path.dentry;
2194
2195                 if (*s) {
2196                         if (!d_can_lookup(dentry)) {
2197                                 fdput(f);
2198                                 return ERR_PTR(-ENOTDIR);
2199                         }
2200                 }
2201
2202                 nd->path = f.file->f_path;
2203                 if (flags & LOOKUP_RCU) {
2204                         rcu_read_lock();
2205                         nd->inode = nd->path.dentry->d_inode;
2206                         nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
2207                 } else {
2208                         path_get(&nd->path);
2209                         nd->inode = nd->path.dentry->d_inode;
2210                 }
2211                 fdput(f);
2212                 return s;
2213         }
2214 }
2215
2216 static const char *trailing_symlink(struct nameidata *nd)
2217 {
2218         const char *s;
2219         int error = may_follow_link(nd);
2220         if (unlikely(error))
2221                 return ERR_PTR(error);
2222         nd->flags |= LOOKUP_PARENT;
2223         nd->stack[0].name = NULL;
2224         s = get_link(nd);
2225         return s ? s : "";
2226 }
2227
2228 static inline int lookup_last(struct nameidata *nd)
2229 {
2230         if (nd->last_type == LAST_NORM && nd->last.name[nd->last.len])
2231                 nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
2232
2233         nd->flags &= ~LOOKUP_PARENT;
2234         return walk_component(nd,
2235                         nd->flags & LOOKUP_FOLLOW
2236                                 ? nd->depth
2237                                         ? WALK_PUT | WALK_GET
2238                                         : WALK_GET
2239                                 : 0);
2240 }
2241
2242 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2243 static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
2244 {
2245         const char *s = path_init(nd, flags);
2246         int err;
2247
2248         if (IS_ERR(s))
2249                 return PTR_ERR(s);
2250         while (!(err = link_path_walk(s, nd))
2251                 && ((err = lookup_last(nd)) > 0)) {
2252                 s = trailing_symlink(nd);
2253                 if (IS_ERR(s)) {
2254                         err = PTR_ERR(s);
2255                         break;
2256                 }
2257         }
2258         if (!err)
2259                 err = complete_walk(nd);
2260
2261         if (!err && nd->flags & LOOKUP_DIRECTORY)
2262                 if (!d_can_lookup(nd->path.dentry))
2263                         err = -ENOTDIR;
2264         if (!err) {
2265                 *path = nd->path;
2266                 nd->path.mnt = NULL;
2267                 nd->path.dentry = NULL;
2268         }
2269         terminate_walk(nd);
2270         return err;
2271 }
2272
2273 static int filename_lookup(int dfd, struct filename *name, unsigned flags,
2274                            struct path *path, struct path *root)
2275 {
2276         int retval;
2277         struct nameidata nd;
2278         if (IS_ERR(name))
2279                 return PTR_ERR(name);
2280         if (unlikely(root)) {
2281                 nd.root = *root;
2282                 flags |= LOOKUP_ROOT;
2283         }
2284         set_nameidata(&nd, dfd, name);
2285         retval = path_lookupat(&nd, flags | LOOKUP_RCU, path);
2286         if (unlikely(retval == -ECHILD))
2287                 retval = path_lookupat(&nd, flags, path);
2288         if (unlikely(retval == -ESTALE))
2289                 retval = path_lookupat(&nd, flags | LOOKUP_REVAL, path);
2290
2291         if (likely(!retval))
2292                 audit_inode(name, path->dentry, flags & LOOKUP_PARENT);
2293         restore_nameidata();
2294         putname(name);
2295         return retval;
2296 }
2297
2298 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2299 static int path_parentat(struct nameidata *nd, unsigned flags,
2300                                 struct path *parent)
2301 {
2302         const char *s = path_init(nd, flags);
2303         int err;
2304         if (IS_ERR(s))
2305                 return PTR_ERR(s);
2306         err = link_path_walk(s, nd);
2307         if (!err)
2308                 err = complete_walk(nd);
2309         if (!err) {
2310                 *parent = nd->path;
2311                 nd->path.mnt = NULL;
2312                 nd->path.dentry = NULL;
2313         }
2314         terminate_walk(nd);
2315         return err;
2316 }
2317
2318 static struct filename *filename_parentat(int dfd, struct filename *name,
2319                                 unsigned int flags, struct path *parent,
2320                                 struct qstr *last, int *type)
2321 {
2322         int retval;
2323         struct nameidata nd;
2324
2325         if (IS_ERR(name))
2326                 return name;
2327         set_nameidata(&nd, dfd, name);
2328         retval = path_parentat(&nd, flags | LOOKUP_RCU, parent);
2329         if (unlikely(retval == -ECHILD))
2330                 retval = path_parentat(&nd, flags, parent);
2331         if (unlikely(retval == -ESTALE))
2332                 retval = path_parentat(&nd, flags | LOOKUP_REVAL, parent);
2333         if (likely(!retval)) {
2334                 *last = nd.last;
2335                 *type = nd.last_type;
2336                 audit_inode(name, parent->dentry, LOOKUP_PARENT);
2337         } else {
2338                 putname(name);
2339                 name = ERR_PTR(retval);
2340         }
2341         restore_nameidata();
2342         return name;
2343 }
2344
2345 /* does lookup, returns the object with parent locked */
2346 struct dentry *kern_path_locked(const char *name, struct path *path)
2347 {
2348         struct filename *filename;
2349         struct dentry *d;
2350         struct qstr last;
2351         int type;
2352
2353         filename = filename_parentat(AT_FDCWD, getname_kernel(name), 0, path,
2354                                     &last, &type);
2355         if (IS_ERR(filename))
2356                 return ERR_CAST(filename);
2357         if (unlikely(type != LAST_NORM)) {
2358                 path_put(path);
2359                 putname(filename);
2360                 return ERR_PTR(-EINVAL);
2361         }
2362         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
2363         d = __lookup_hash(&last, path->dentry, 0);
2364         if (IS_ERR(d)) {
2365                 inode_unlock(path->dentry->d_inode);
2366                 path_put(path);
2367         }
2368         putname(filename);
2369         return d;
2370 }
2371
2372 int kern_path(const char *name, unsigned int flags, struct path *path)
2373 {
2374         return filename_lookup(AT_FDCWD, getname_kernel(name),
2375                                flags, path, NULL);
2376 }
2377 EXPORT_SYMBOL(kern_path);
2378
2379 /**
2380  * vfs_path_lookup - lookup a file path relative to a dentry-vfsmount pair
2381  * @dentry:  pointer to dentry of the base directory
2382  * @mnt: pointer to vfs mount of the base directory
2383  * @name: pointer to file name
2384  * @flags: lookup flags
2385  * @path: pointer to struct path to fill
2386  */
2387 int vfs_path_lookup(struct dentry *dentry, struct vfsmount *mnt,
2388                     const char *name, unsigned int flags,
2389                     struct path *path)
2390 {
2391         struct path root = {.mnt = mnt, .dentry = dentry};
2392         /* the first argument of filename_lookup() is ignored with root */
2393         return filename_lookup(AT_FDCWD, getname_kernel(name),
2394                                flags , path, &root);
2395 }
2396 EXPORT_SYMBOL(vfs_path_lookup);
2397
2398 /**
2399  * lookup_hash - lookup single pathname component on already hashed name
2400  * @name:       name and hash to lookup
2401  * @base:       base directory to lookup from
2402  *
2403  * The name must have been verified and hashed (see lookup_one_len()).  Using
2404  * this after just full_name_hash() is unsafe.
2405  *
2406  * This function also doesn't check for search permission on base directory.
2407  *
2408  * Use lookup_one_len_unlocked() instead, unless you really know what you are
2409  * doing.
2410  *
2411  * Do not hold i_mutex; this helper takes i_mutex if necessary.
2412  */
2413 struct dentry *lookup_hash(const struct qstr *name, struct dentry *base)
2414 {
2415         struct dentry *ret;
2416
2417         ret = lookup_dcache(name, base, 0);
2418         if (!ret)
2419                 ret = lookup_slow(name, base, 0);
2420
2421         return ret;
2422 }
2423 EXPORT_SYMBOL(lookup_hash);
2424
2425 /**
2426  * lookup_one_len - filesystem helper to lookup single pathname component
2427  * @name:       pathname component to lookup
2428  * @base:       base directory to lookup from
2429  * @len:        maximum length @len should be interpreted to
2430  *
2431  * Note that this routine is purely a helper for filesystem usage and should
2432  * not be called by generic code.
2433  *
2434  * The caller must hold base->i_mutex.
2435  */
2436 struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
2437 {
2438         struct qstr this;
2439         unsigned int c;
2440         int err;
2441
2442         WARN_ON_ONCE(!inode_is_locked(base->d_inode));
2443
2444         this.name = name;
2445         this.len = len;
2446         this.hash = full_name_hash(base, name, len);
2447         if (!len)
2448                 return ERR_PTR(-EACCES);
2449
2450         if (unlikely(name[0] == '.')) {
2451                 if (len < 2 || (len == 2 && name[1] == '.'))
2452                         return ERR_PTR(-EACCES);
2453         }
2454
2455         while (len--) {
2456                 c = *(const unsigned char *)name++;
2457                 if (c == '/' || c == '\0')
2458                         return ERR_PTR(-EACCES);
2459         }
2460         /*
2461          * See if the low-level filesystem might want
2462          * to use its own hash..
2463          */
2464         if (base->d_flags & DCACHE_OP_HASH) {
2465                 int err = base->d_op->d_hash(base, &this);
2466                 if (err < 0)
2467                         return ERR_PTR(err);
2468         }
2469
2470         err = inode_permission(base->d_inode, MAY_EXEC);
2471         if (err)
2472                 return ERR_PTR(err);
2473
2474         return __lookup_hash(&this, base, 0);
2475 }
2476 EXPORT_SYMBOL(lookup_one_len);
2477
2478 /**
2479  * lookup_one_len_unlocked - filesystem helper to lookup single pathname component
2480  * @name:       pathname component to lookup
2481  * @base:       base directory to lookup from
2482  * @len:        maximum length @len should be interpreted to
2483  *
2484  * Note that this routine is purely a helper for filesystem usage and should
2485  * not be called by generic code.
2486  *
2487  * Unlike lookup_one_len, it should be called without the parent
2488  * i_mutex held, and will take the i_mutex itself if necessary.
2489  */
2490 struct dentry *lookup_one_len_unlocked(const char *name,
2491                                        struct dentry *base, int len)
2492 {
2493         struct qstr this;
2494         unsigned int c;
2495         int err;
2496
2497         this.name = name;
2498         this.len = len;
2499         this.hash = full_name_hash(base, name, len);
2500         if (!len)
2501                 return ERR_PTR(-EACCES);
2502
2503         if (unlikely(name[0] == '.')) {
2504                 if (len < 2 || (len == 2 && name[1] == '.'))
2505                         return ERR_PTR(-EACCES);
2506         }
2507
2508         while (len--) {
2509                 c = *(const unsigned char *)name++;
2510                 if (c == '/' || c == '\0')
2511                         return ERR_PTR(-EACCES);
2512         }
2513         /*
2514          * See if the low-level filesystem might want
2515          * to use its own hash..
2516          */
2517         if (base->d_flags & DCACHE_OP_HASH) {
2518                 int err = base->d_op->d_hash(base, &this);
2519                 if (err < 0)
2520                         return ERR_PTR(err);
2521         }
2522
2523         err = inode_permission(base->d_inode, MAY_EXEC);
2524         if (err)
2525                 return ERR_PTR(err);
2526
2527         return lookup_hash(&this, base);
2528 }
2529 EXPORT_SYMBOL(lookup_one_len_unlocked);
2530
2531 #ifdef CONFIG_UNIX98_PTYS
2532 int path_pts(struct path *path)
2533 {
2534         /* Find something mounted on "pts" in the same directory as
2535          * the input path.
2536          */
2537         struct dentry *child, *parent;
2538         struct qstr this;
2539         int ret;
2540
2541         ret = path_parent_directory(path);
2542         if (ret)
2543                 return ret;
2544
2545         parent = path->dentry;
2546         this.name = "pts";
2547         this.len = 3;
2548         child = d_hash_and_lookup(parent, &this);
2549         if (!child)
2550                 return -ENOENT;
2551
2552         path->dentry = child;
2553         dput(parent);
2554         follow_mount(path);
2555         return 0;
2556 }
2557 #endif
2558
2559 int user_path_at_empty(int dfd, const char __user *name, unsigned flags,
2560                  struct path *path, int *empty)
2561 {
2562         return filename_lookup(dfd, getname_flags(name, flags, empty),
2563                                flags, path, NULL);
2564 }
2565 EXPORT_SYMBOL(user_path_at_empty);
2566
2567 /*
2568  * NB: most callers don't do anything directly with the reference to the
2569  *     to struct filename, but the nd->last pointer points into the name string
2570  *     allocated by getname. So we must hold the reference to it until all
2571  *     path-walking is complete.
2572  */
2573 static inline struct filename *
2574 user_path_parent(int dfd, const char __user *path,
2575                  struct path *parent,
2576                  struct qstr *last,
2577                  int *type,
2578                  unsigned int flags)
2579 {
2580         /* only LOOKUP_REVAL is allowed in extra flags */
2581         return filename_parentat(dfd, getname(path), flags & LOOKUP_REVAL,
2582                                  parent, last, type);
2583 }
2584
2585 /**
2586  * mountpoint_last - look up last component for umount
2587  * @nd:   pathwalk nameidata - currently pointing at parent directory of "last"
2588  * @path: pointer to container for result
2589  *
2590  * This is a special lookup_last function just for umount. In this case, we
2591  * need to resolve the path without doing any revalidation.
2592  *
2593  * The nameidata should be the result of doing a LOOKUP_PARENT pathwalk. Since
2594  * mountpoints are always pinned in the dcache, their ancestors are too. Thus,
2595  * in almost all cases, this lookup will be served out of the dcache. The only
2596  * cases where it won't are if nd->last refers to a symlink or the path is
2597  * bogus and it doesn't exist.
2598  *
2599  * Returns:
2600  * -error: if there was an error during lookup. This includes -ENOENT if the
2601  *         lookup found a negative dentry. The nd->path reference will also be
2602  *         put in this case.
2603  *
2604  * 0:      if we successfully resolved nd->path and found it to not to be a
2605  *         symlink that needs to be followed. "path" will also be populated.
2606  *         The nd->path reference will also be put.
2607  *
2608  * 1:      if we successfully resolved nd->last and found it to be a symlink
2609  *         that needs to be followed. "path" will be populated with the path
2610  *         to the link, and nd->path will *not* be put.
2611  */
2612 static int
2613 mountpoint_last(struct nameidata *nd, struct path *path)
2614 {
2615         int error = 0;
2616         struct dentry *dentry;
2617         struct dentry *dir = nd->path.dentry;
2618
2619         /* If we're in rcuwalk, drop out of it to handle last component */
2620         if (nd->flags & LOOKUP_RCU) {
2621                 if (unlazy_walk(nd, NULL, 0))
2622                         return -ECHILD;
2623         }
2624
2625         nd->flags &= ~LOOKUP_PARENT;
2626
2627         if (unlikely(nd->last_type != LAST_NORM)) {
2628                 error = handle_dots(nd, nd->last_type);
2629                 if (error)
2630                         return error;
2631                 dentry = dget(nd->path.dentry);
2632         } else {
2633                 dentry = d_lookup(dir, &nd->last);
2634                 if (!dentry) {
2635                         /*
2636                          * No cached dentry. Mounted dentries are pinned in the
2637                          * cache, so that means that this dentry is probably
2638                          * a symlink or the path doesn't actually point
2639                          * to a mounted dentry.
2640                          */
2641                         dentry = lookup_slow(&nd->last, dir,
2642                                              nd->flags | LOOKUP_NO_REVAL);
2643                         if (IS_ERR(dentry))
2644                                 return PTR_ERR(dentry);
2645                 }
2646         }
2647         if (d_is_negative(dentry)) {
2648                 dput(dentry);
2649                 return -ENOENT;
2650         }
2651         if (nd->depth)
2652                 put_link(nd);
2653         path->dentry = dentry;
2654         path->mnt = nd->path.mnt;
2655         error = should_follow_link(nd, path, nd->flags & LOOKUP_FOLLOW,
2656                                    d_backing_inode(dentry), 0);
2657         if (unlikely(error))
2658                 return error;
2659         mntget(path->mnt);
2660         follow_mount(path);
2661         return 0;
2662 }
2663
2664 /**
2665  * path_mountpoint - look up a path to be umounted
2666  * @nd:         lookup context
2667  * @flags:      lookup flags
2668  * @path:       pointer to container for result
2669  *
2670  * Look up the given name, but don't attempt to revalidate the last component.
2671  * Returns 0 and "path" will be valid on success; Returns error otherwise.
2672  */
2673 static int
2674 path_mountpoint(struct nameidata *nd, unsigned flags, struct path *path)
2675 {
2676         const char *s = path_init(nd, flags);
2677         int err;
2678         if (IS_ERR(s))
2679                 return PTR_ERR(s);
2680         while (!(err = link_path_walk(s, nd)) &&
2681                 (err = mountpoint_last(nd, path)) > 0) {
2682                 s = trailing_symlink(nd);
2683                 if (IS_ERR(s)) {
2684                         err = PTR_ERR(s);
2685                         break;
2686                 }
2687         }
2688         terminate_walk(nd);
2689         return err;
2690 }
2691
2692 static int
2693 filename_mountpoint(int dfd, struct filename *name, struct path *path,
2694                         unsigned int flags)
2695 {
2696         struct nameidata nd;
2697         int error;
2698         if (IS_ERR(name))
2699                 return PTR_ERR(name);
2700         set_nameidata(&nd, dfd, name);
2701         error = path_mountpoint(&nd, flags | LOOKUP_RCU, path);
2702         if (unlikely(error == -ECHILD))
2703                 error = path_mountpoint(&nd, flags, path);
2704         if (unlikely(error == -ESTALE))
2705                 error = path_mountpoint(&nd, flags | LOOKUP_REVAL, path);
2706         if (likely(!error))
2707                 audit_inode(name, path->dentry, 0);
2708         restore_nameidata();
2709         putname(name);
2710         return error;
2711 }
2712
2713 /**
2714  * user_path_mountpoint_at - lookup a path from userland in order to umount it
2715  * @dfd:        directory file descriptor
2716  * @name:       pathname from userland
2717  * @flags:      lookup flags
2718  * @path:       pointer to container to hold result
2719  *
2720  * A umount is a special case for path walking. We're not actually interested
2721  * in the inode in this situation, and ESTALE errors can be a problem. We
2722  * simply want track down the dentry and vfsmount attached at the mountpoint
2723  * and avoid revalidating the last component.
2724  *
2725  * Returns 0 and populates "path" on success.
2726  */
2727 int
2728 user_path_mountpoint_at(int dfd, const char __user *name, unsigned int flags,
2729                         struct path *path)
2730 {
2731         return filename_mountpoint(dfd, getname(name), path, flags);
2732 }
2733
2734 int
2735 kern_path_mountpoint(int dfd, const char *name, struct path *path,
2736                         unsigned int flags)
2737 {
2738         return filename_mountpoint(dfd, getname_kernel(name), path, flags);
2739 }
2740 EXPORT_SYMBOL(kern_path_mountpoint);
2741
2742 int __check_sticky(struct inode *dir, struct inode *inode)
2743 {
2744         kuid_t fsuid = current_fsuid();
2745
2746         if (uid_eq(inode->i_uid, fsuid))
2747                 return 0;
2748         if (uid_eq(dir->i_uid, fsuid))
2749                 return 0;
2750         return !capable_wrt_inode_uidgid(inode, CAP_FOWNER);
2751 }
2752 EXPORT_SYMBOL(__check_sticky);
2753
2754 /*
2755  *      Check whether we can remove a link victim from directory dir, check
2756  *  whether the type of victim is right.
2757  *  1. We can't do it if dir is read-only (done in permission())
2758  *  2. We should have write and exec permissions on dir
2759  *  3. We can't remove anything from append-only dir
2760  *  4. We can't do anything with immutable dir (done in permission())
2761  *  5. If the sticky bit on dir is set we should either
2762  *      a. be owner of dir, or
2763  *      b. be owner of victim, or
2764  *      c. have CAP_FOWNER capability
2765  *  6. If the victim is append-only or immutable we can't do antyhing with
2766  *     links pointing to it.
2767  *  7. If we were asked to remove a directory and victim isn't one - ENOTDIR.
2768  *  8. If we were asked to remove a non-directory and victim isn't one - EISDIR.
2769  *  9. We can't remove a root or mountpoint.
2770  * 10. We don't allow removal of NFS sillyrenamed files; it's handled by
2771  *     nfs_async_unlink().
2772  */
2773 static int may_delete(struct inode *dir, struct dentry *victim, bool isdir)
2774 {
2775         struct inode *inode = d_backing_inode(victim);
2776         int error;
2777
2778         if (d_is_negative(victim))
2779                 return -ENOENT;
2780         BUG_ON(!inode);
2781
2782         BUG_ON(victim->d_parent->d_inode != dir);
2783         audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);
2784
2785         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
2786         if (error)
2787                 return error;
2788         if (IS_APPEND(dir))
2789                 return -EPERM;
2790
2791         if (check_sticky(dir, inode) || IS_APPEND(inode) ||
2792             IS_IMMUTABLE(inode) || IS_SWAPFILE(inode))
2793                 return -EPERM;
2794         if (isdir) {
2795                 if (!d_is_dir(victim))
2796                         return -ENOTDIR;
2797                 if (IS_ROOT(victim))
2798                         return -EBUSY;
2799         } else if (d_is_dir(victim))
2800                 return -EISDIR;
2801         if (IS_DEADDIR(dir))
2802                 return -ENOENT;
2803         if (victim->d_flags & DCACHE_NFSFS_RENAMED)
2804                 return -EBUSY;
2805         return 0;
2806 }
2807
2808 /*      Check whether we can create an object with dentry child in directory
2809  *  dir.
2810  *  1. We can't do it if child already exists (open has special treatment for
2811  *     this case, but since we are inlined it's OK)
2812  *  2. We can't do it if dir is read-only (done in permission())
2813  *  3. We should have write and exec permissions on dir
2814  *  4. We can't do it if dir is immutable (done in permission())
2815  */
2816 static inline int may_create(struct inode *dir, struct dentry *child)
2817 {
2818         audit_inode_child(dir, child, AUDIT_TYPE_CHILD_CREATE);
2819         if (child->d_inode)
2820                 return -EEXIST;
2821         if (IS_DEADDIR(dir))
2822                 return -ENOENT;
2823         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
2824 }
2825
2826 /*
2827  * p1 and p2 should be directories on the same fs.
2828  */
2829 struct dentry *lock_rename(struct dentry *p1, struct dentry *p2)
2830 {
2831         struct dentry *p;
2832
2833         if (p1 == p2) {
2834                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2835                 return NULL;
2836         }
2837
2838         mutex_lock(&p1->d_sb->s_vfs_rename_mutex);
2839
2840         p = d_ancestor(p2, p1);
2841         if (p) {
2842                 inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
2843                 inode_lock_nested(p1->d_inode, I_MUTEX_CHILD);
2844                 return p;
2845         }
2846
2847         p = d_ancestor(p1, p2);
2848         if (p) {
2849                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2850                 inode_lock_nested(p2->d_inode, I_MUTEX_CHILD);
2851                 return p;
2852         }
2853
2854         inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2855         inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
2856         return NULL;
2857 }
2858 EXPORT_SYMBOL(lock_rename);
2859
2860 void unlock_rename(struct dentry *p1, struct dentry *p2)
2861 {
2862         inode_unlock(p1->d_inode);
2863         if (p1 != p2) {
2864                 inode_unlock(p2->d_inode);
2865                 mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
2866         }
2867 }
2868 EXPORT_SYMBOL(unlock_rename);
2869
2870 int vfs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2871                 bool want_excl)
2872 {
2873         int error = may_create(dir, dentry);
2874         if (error)
2875                 return error;
2876
2877         if (!dir->i_op->create)
2878                 return -EACCES; /* shouldn't it be ENOSYS? */
2879         mode &= S_IALLUGO;
2880         mode |= S_IFREG;
2881         error = security_inode_create(dir, dentry, mode);
2882         if (error)
2883                 return error;
2884         error = dir->i_op->create(dir, dentry, mode, want_excl);
2885         if (!error)
2886                 fsnotify_create(dir, dentry);
2887         return error;
2888 }
2889 EXPORT_SYMBOL(vfs_create);
2890
2891 static int may_open(struct path *path, int acc_mode, int flag)
2892 {
2893         struct dentry *dentry = path->dentry;
2894         struct inode *inode = dentry->d_inode;
2895         int error;
2896
2897         if (!inode)
2898                 return -ENOENT;
2899
2900         switch (inode->i_mode & S_IFMT) {
2901         case S_IFLNK:
2902                 return -ELOOP;
2903         case S_IFDIR:
2904                 if (acc_mode & MAY_WRITE)
2905                         return -EISDIR;
2906                 break;
2907         case S_IFBLK:
2908         case S_IFCHR:
2909                 if (path->mnt->mnt_flags & MNT_NODEV)
2910                         return -EACCES;
2911                 /*FALLTHRU*/
2912         case S_IFIFO:
2913         case S_IFSOCK:
2914                 flag &= ~O_TRUNC;
2915                 break;
2916         }
2917
2918         error = inode_permission(inode, MAY_OPEN | acc_mode);
2919         if (error)
2920                 return error;
2921
2922         /*
2923          * An append-only file must be opened in append mode for writing.
2924          */
2925         if (IS_APPEND(inode)) {
2926                 if  ((flag & O_ACCMODE) != O_RDONLY && !(flag & O_APPEND))
2927                         return -EPERM;
2928                 if (flag & O_TRUNC)
2929                         return -EPERM;
2930         }
2931
2932         /* O_NOATIME can only be set by the owner or superuser */
2933         if (flag & O_NOATIME && !inode_owner_or_capable(inode))
2934                 return -EPERM;
2935
2936         return 0;
2937 }
2938
2939 static int handle_truncate(struct file *filp)
2940 {
2941         struct path *path = &filp->f_path;
2942         struct inode *inode = path->dentry->d_inode;
2943         int error = get_write_access(inode);
2944         if (error)
2945                 return error;
2946         /*
2947          * Refuse to truncate files with mandatory locks held on them.
2948          */
2949         error = locks_verify_locked(filp);
2950         if (!error)
2951                 error = security_path_truncate(path);
2952         if (!error) {
2953                 error = do_truncate(path->dentry, 0,
2954                                     ATTR_MTIME|ATTR_CTIME|ATTR_OPEN,
2955                                     filp);
2956         }
2957         put_write_access(inode);
2958         return error;
2959 }
2960
2961 static inline int open_to_namei_flags(int flag)
2962 {
2963         if ((flag & O_ACCMODE) == 3)
2964                 flag--;
2965         return flag;
2966 }
2967
2968 static int may_o_create(const struct path *dir, struct dentry *dentry, umode_t mode)
2969 {
2970         int error = security_path_mknod(dir, dentry, mode, 0);
2971         if (error)
2972                 return error;
2973
2974         error = inode_permission(dir->dentry->d_inode, MAY_WRITE | MAY_EXEC);
2975         if (error)
2976                 return error;
2977
2978         return security_inode_create(dir->dentry->d_inode, dentry, mode);
2979 }
2980
2981 /*
2982  * Attempt to atomically look up, create and open a file from a negative
2983  * dentry.
2984  *
2985  * Returns 0 if successful.  The file will have been created and attached to
2986  * @file by the filesystem calling finish_open().
2987  *
2988  * Returns 1 if the file was looked up only or didn't need creating.  The
2989  * caller will need to perform the open themselves.  @path will have been
2990  * updated to point to the new dentry.  This may be negative.
2991  *
2992  * Returns an error code otherwise.
2993  */
2994 static int atomic_open(struct nameidata *nd, struct dentry *dentry,
2995                         struct path *path, struct file *file,
2996                         const struct open_flags *op,
2997                         int open_flag, umode_t mode,
2998                         int *opened)
2999 {
3000         struct dentry *const DENTRY_NOT_SET = (void *) -1UL;
3001         struct inode *dir =  nd->path.dentry->d_inode;
3002         int error;
3003
3004         if (!(~open_flag & (O_EXCL | O_CREAT))) /* both O_EXCL and O_CREAT */
3005                 open_flag &= ~O_TRUNC;
3006
3007         if (nd->flags & LOOKUP_DIRECTORY)
3008                 open_flag |= O_DIRECTORY;
3009
3010         file->f_path.dentry = DENTRY_NOT_SET;
3011         file->f_path.mnt = nd->path.mnt;
3012         error = dir->i_op->atomic_open(dir, dentry, file,
3013                                        open_to_namei_flags(open_flag),
3014                                        mode, opened);
3015         d_lookup_done(dentry);
3016         if (!error) {
3017                 /*
3018                  * We didn't have the inode before the open, so check open
3019                  * permission here.
3020                  */
3021                 int acc_mode = op->acc_mode;
3022                 if (*opened & FILE_CREATED) {
3023                         WARN_ON(!(open_flag & O_CREAT));
3024                         fsnotify_create(dir, dentry);
3025                         acc_mode = 0;
3026                 }
3027                 error = may_open(&file->f_path, acc_mode, open_flag);
3028                 if (WARN_ON(error > 0))
3029                         error = -EINVAL;
3030         } else if (error > 0) {
3031                 if (WARN_ON(file->f_path.dentry == DENTRY_NOT_SET)) {
3032                         error = -EIO;
3033                 } else {
3034                         if (file->f_path.dentry) {
3035                                 dput(dentry);
3036                                 dentry = file->f_path.dentry;
3037                         }
3038                         if (*opened & FILE_CREATED)
3039                                 fsnotify_create(dir, dentry);
3040                         if (unlikely(d_is_negative(dentry))) {
3041                                 error = -ENOENT;
3042                         } else {
3043                                 path->dentry = dentry;
3044                                 path->mnt = nd->path.mnt;
3045                                 return 1;
3046                         }
3047                 }
3048         }
3049         dput(dentry);
3050         return error;
3051 }
3052
3053 /*
3054  * Look up and maybe create and open the last component.
3055  *
3056  * Must be called with i_mutex held on parent.
3057  *
3058  * Returns 0 if the file was successfully atomically created (if necessary) and
3059  * opened.  In this case the file will be returned attached to @file.
3060  *
3061  * Returns 1 if the file was not completely opened at this time, though lookups
3062  * and creations will have been performed and the dentry returned in @path will
3063  * be positive upon return if O_CREAT was specified.  If O_CREAT wasn't
3064  * specified then a negative dentry may be returned.
3065  *
3066  * An error code is returned otherwise.
3067  *
3068  * FILE_CREATE will be set in @*opened if the dentry was created and will be
3069  * cleared otherwise prior to returning.
3070  */
3071 static int lookup_open(struct nameidata *nd, struct path *path,
3072                         struct file *file,
3073                         const struct open_flags *op,
3074                         bool got_write, int *opened)
3075 {
3076         struct dentry *dir = nd->path.dentry;
3077         struct inode *dir_inode = dir->d_inode;
3078         int open_flag = op->open_flag;
3079         struct dentry *dentry;
3080         int error, create_error = 0;
3081         umode_t mode = op->mode;
3082         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
3083
3084         if (unlikely(IS_DEADDIR(dir_inode)))
3085                 return -ENOENT;
3086
3087         *opened &= ~FILE_CREATED;
3088         dentry = d_lookup(dir, &nd->last);
3089         for (;;) {
3090                 if (!dentry) {
3091                         dentry = d_alloc_parallel(dir, &nd->last, &wq);
3092                         if (IS_ERR(dentry))
3093                                 return PTR_ERR(dentry);
3094                 }
3095                 if (d_in_lookup(dentry))
3096                         break;
3097
3098                 if (!(dentry->d_flags & DCACHE_OP_REVALIDATE))
3099                         break;
3100
3101                 error = d_revalidate(dentry, nd->flags);
3102                 if (likely(error > 0))
3103                         break;
3104                 if (error)
3105                         goto out_dput;
3106                 d_invalidate(dentry);
3107                 dput(dentry);
3108                 dentry = NULL;
3109         }
3110         if (dentry->d_inode) {
3111                 /* Cached positive dentry: will open in f_op->open */
3112                 goto out_no_open;
3113         }
3114
3115         /*
3116          * Checking write permission is tricky, bacuse we don't know if we are
3117          * going to actually need it: O_CREAT opens should work as long as the
3118          * file exists.  But checking existence breaks atomicity.  The trick is
3119          * to check access and if not granted clear O_CREAT from the flags.
3120          *
3121          * Another problem is returing the "right" error value (e.g. for an
3122          * O_EXCL open we want to return EEXIST not EROFS).
3123          */
3124         if (open_flag & O_CREAT) {
3125                 if (!IS_POSIXACL(dir->d_inode))
3126                         mode &= ~current_umask();
3127                 if (unlikely(!got_write)) {
3128                         create_error = -EROFS;
3129                         open_flag &= ~O_CREAT;
3130                         if (open_flag & (O_EXCL | O_TRUNC))
3131                                 goto no_open;
3132                         /* No side effects, safe to clear O_CREAT */
3133                 } else {
3134                         create_error = may_o_create(&nd->path, dentry, mode);
3135                         if (create_error) {
3136                                 open_flag &= ~O_CREAT;
3137                                 if (open_flag & O_EXCL)
3138                                         goto no_open;
3139                         }
3140                 }
3141         } else if ((open_flag & (O_TRUNC|O_WRONLY|O_RDWR)) &&
3142                    unlikely(!got_write)) {
3143                 /*
3144                  * No O_CREATE -> atomicity not a requirement -> fall
3145                  * back to lookup + open
3146                  */
3147                 goto no_open;
3148         }
3149
3150         if (dir_inode->i_op->atomic_open) {
3151                 error = atomic_open(nd, dentry, path, file, op, open_flag,
3152                                     mode, opened);
3153                 if (unlikely(error == -ENOENT) && create_error)
3154                         error = create_error;
3155                 return error;
3156         }
3157
3158 no_open:
3159         if (d_in_lookup(dentry)) {
3160                 struct dentry *res = dir_inode->i_op->lookup(dir_inode, dentry,
3161                                                              nd->flags);
3162                 d_lookup_done(dentry);
3163                 if (unlikely(res)) {
3164                         if (IS_ERR(res)) {
3165                                 error = PTR_ERR(res);
3166                                 goto out_dput;
3167                         }
3168                         dput(dentry);
3169                         dentry = res;
3170                 }
3171         }
3172
3173         /* Negative dentry, just create the file */
3174         if (!dentry->d_inode && (open_flag & O_CREAT)) {
3175                 *opened |= FILE_CREATED;
3176                 audit_inode_child(dir_inode, dentry, AUDIT_TYPE_CHILD_CREATE);
3177                 if (!dir_inode->i_op->create) {
3178                         error = -EACCES;
3179                         goto out_dput;
3180                 }
3181                 error = dir_inode->i_op->create(dir_inode, dentry, mode,
3182                                                 open_flag & O_EXCL);
3183                 if (error)
3184                         goto out_dput;
3185                 fsnotify_create(dir_inode, dentry);
3186         }
3187         if (unlikely(create_error) && !dentry->d_inode) {
3188                 error = create_error;
3189                 goto out_dput;
3190         }
3191 out_no_open:
3192         path->dentry = dentry;
3193         path->mnt = nd->path.mnt;
3194         return 1;
3195
3196 out_dput:
3197         dput(dentry);
3198         return error;
3199 }
3200
3201 /*
3202  * Handle the last step of open()
3203  */
3204 static int do_last(struct nameidata *nd,
3205                    struct file *file, const struct open_flags *op,
3206                    int *opened)
3207 {
3208         struct dentry *dir = nd->path.dentry;
3209         int open_flag = op->open_flag;
3210         bool will_truncate = (open_flag & O_TRUNC) != 0;
3211         bool got_write = false;
3212         int acc_mode = op->acc_mode;
3213         unsigned seq;
3214         struct inode *inode;
3215         struct path path;
3216         int error;
3217
3218         nd->flags &= ~LOOKUP_PARENT;
3219         nd->flags |= op->intent;
3220
3221         if (nd->last_type != LAST_NORM) {
3222                 error = handle_dots(nd, nd->last_type);
3223                 if (unlikely(error))
3224                         return error;
3225                 goto finish_open;
3226         }
3227
3228         if (!(open_flag & O_CREAT)) {
3229                 if (nd->last.name[nd->last.len])
3230                         nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
3231                 /* we _can_ be in RCU mode here */
3232                 error = lookup_fast(nd, &path, &inode, &seq);
3233                 if (likely(error > 0))
3234                         goto finish_lookup;
3235
3236                 if (error < 0)
3237                         return error;
3238
3239                 BUG_ON(nd->inode != dir->d_inode);
3240                 BUG_ON(nd->flags & LOOKUP_RCU);
3241         } else {
3242                 /* create side of things */
3243                 /*
3244                  * This will *only* deal with leaving RCU mode - LOOKUP_JUMPED
3245                  * has been cleared when we got to the last component we are
3246                  * about to look up
3247                  */
3248                 error = complete_walk(nd);
3249                 if (error)
3250                         return error;
3251
3252                 audit_inode(nd->name, dir, LOOKUP_PARENT);
3253                 /* trailing slashes? */
3254                 if (unlikely(nd->last.name[nd->last.len]))
3255                         return -EISDIR;
3256         }
3257
3258         if (open_flag & (O_CREAT | O_TRUNC | O_WRONLY | O_RDWR)) {
3259                 error = mnt_want_write(nd->path.mnt);
3260                 if (!error)
3261                         got_write = true;
3262                 /*
3263                  * do _not_ fail yet - we might not need that or fail with
3264                  * a different error; let lookup_open() decide; we'll be
3265                  * dropping this one anyway.
3266                  */
3267         }
3268         if (open_flag & O_CREAT)
3269                 inode_lock(dir->d_inode);
3270         else
3271                 inode_lock_shared(dir->d_inode);
3272         error = lookup_open(nd, &path, file, op, got_write, opened);
3273         if (open_flag & O_CREAT)
3274                 inode_unlock(dir->d_inode);
3275         else
3276                 inode_unlock_shared(dir->d_inode);
3277
3278         if (error <= 0) {
3279                 if (error)
3280                         goto out;
3281
3282                 if ((*opened & FILE_CREATED) ||
3283                     !S_ISREG(file_inode(file)->i_mode))
3284                         will_truncate = false;
3285
3286                 audit_inode(nd->name, file->f_path.dentry, 0);
3287                 goto opened;
3288         }
3289
3290         if (*opened & FILE_CREATED) {
3291                 /* Don't check for write permission, don't truncate */
3292                 open_flag &= ~O_TRUNC;
3293                 will_truncate = false;
3294                 acc_mode = 0;
3295                 path_to_nameidata(&path, nd);
3296                 goto finish_open_created;
3297         }
3298
3299         /*
3300          * If atomic_open() acquired write access it is dropped now due to
3301          * possible mount and symlink following (this might be optimized away if
3302          * necessary...)
3303          */
3304         if (got_write) {
3305                 mnt_drop_write(nd->path.mnt);
3306                 got_write = false;
3307         }
3308
3309         error = follow_managed(&path, nd);
3310         if (unlikely(error < 0))
3311                 return error;
3312
3313         if (unlikely(d_is_negative(path.dentry))) {
3314                 path_to_nameidata(&path, nd);
3315                 return -ENOENT;
3316         }
3317
3318         /*
3319          * create/update audit record if it already exists.
3320          */
3321         audit_inode(nd->name, path.dentry, 0);
3322
3323         if (unlikely((open_flag & (O_EXCL | O_CREAT)) == (O_EXCL | O_CREAT))) {
3324                 path_to_nameidata(&path, nd);
3325                 return -EEXIST;
3326         }
3327
3328         seq = 0;        /* out of RCU mode, so the value doesn't matter */
3329         inode = d_backing_inode(path.dentry);
3330 finish_lookup:
3331         if (nd->depth)
3332                 put_link(nd);
3333         error = should_follow_link(nd, &path, nd->flags & LOOKUP_FOLLOW,
3334                                    inode, seq);
3335         if (unlikely(error))
3336                 return error;
3337
3338         path_to_nameidata(&path, nd);
3339         nd->inode = inode;
3340         nd->seq = seq;
3341         /* Why this, you ask?  _Now_ we might have grown LOOKUP_JUMPED... */
3342 finish_open:
3343         error = complete_walk(nd);
3344         if (error)
3345                 return error;
3346         audit_inode(nd->name, nd->path.dentry, 0);
3347         error = -EISDIR;
3348         if ((open_flag & O_CREAT) && d_is_dir(nd->path.dentry))
3349                 goto out;
3350         error = -ENOTDIR;
3351         if ((nd->flags & LOOKUP_DIRECTORY) && !d_can_lookup(nd->path.dentry))
3352                 goto out;
3353         if (!d_is_reg(nd->path.dentry))
3354                 will_truncate = false;
3355
3356         if (will_truncate) {
3357                 error = mnt_want_write(nd->path.mnt);
3358                 if (error)
3359                         goto out;
3360                 got_write = true;
3361         }
3362 finish_open_created:
3363         error = may_open(&nd->path, acc_mode, open_flag);
3364         if (error)
3365                 goto out;
3366         BUG_ON(*opened & FILE_OPENED); /* once it's opened, it's opened */
3367         error = vfs_open(&nd->path, file, current_cred());
3368         if (error)
3369                 goto out;
3370         *opened |= FILE_OPENED;
3371 opened:
3372         error = open_check_o_direct(file);
3373         if (!error)
3374                 error = ima_file_check(file, op->acc_mode, *opened);
3375         if (!error && will_truncate)
3376                 error = handle_truncate(file);
3377 out:
3378         if (unlikely(error) && (*opened & FILE_OPENED))
3379                 fput(file);
3380         if (unlikely(error > 0)) {
3381                 WARN_ON(1);
3382                 error = -EINVAL;
3383         }
3384         if (got_write)
3385                 mnt_drop_write(nd->path.mnt);
3386         return error;
3387 }
3388
3389 static int do_tmpfile(struct nameidata *nd, unsigned flags,
3390                 const struct open_flags *op,
3391                 struct file *file, int *opened)
3392 {
3393         static const struct qstr name = QSTR_INIT("/", 1);
3394         struct dentry *child;
3395         struct inode *dir;
3396         struct path path;
3397         int error = path_lookupat(nd, flags | LOOKUP_DIRECTORY, &path);
3398         if (unlikely(error))
3399                 return error;
3400         error = mnt_want_write(path.mnt);
3401         if (unlikely(error))
3402                 goto out;
3403         dir = path.dentry->d_inode;
3404         /* we want directory to be writable */
3405         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
3406         if (error)
3407                 goto out2;
3408         if (!dir->i_op->tmpfile) {
3409                 error = -EOPNOTSUPP;
3410                 goto out2;
3411         }
3412         child = d_alloc(path.dentry, &name);
3413         if (unlikely(!child)) {
3414                 error = -ENOMEM;
3415                 goto out2;
3416         }
3417         dput(path.dentry);
3418         path.dentry = child;
3419         error = dir->i_op->tmpfile(dir, child, op->mode);
3420         if (error)
3421                 goto out2;
3422         audit_inode(nd->name, child, 0);
3423         /* Don't check for other permissions, the inode was just created */
3424         error = may_open(&path, 0, op->open_flag);
3425         if (error)
3426                 goto out2;
3427         file->f_path.mnt = path.mnt;
3428         error = finish_open(file, child, NULL, opened);
3429         if (error)
3430                 goto out2;
3431         error = open_check_o_direct(file);
3432         if (error) {
3433                 fput(file);
3434         } else if (!(op->open_flag & O_EXCL)) {
3435                 struct inode *inode = file_inode(file);
3436                 spin_lock(&inode->i_lock);
3437                 inode->i_state |= I_LINKABLE;
3438                 spin_unlock(&inode->i_lock);
3439         }
3440 out2:
3441         mnt_drop_write(path.mnt);
3442 out:
3443         path_put(&path);
3444         return error;
3445 }
3446
3447 static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
3448 {
3449         struct path path;
3450         int error = path_lookupat(nd, flags, &path);
3451         if (!error) {
3452                 audit_inode(nd->name, path.dentry, 0);
3453                 error = vfs_open(&path, file, current_cred());
3454                 path_put(&path);
3455         }
3456         return error;
3457 }
3458
3459 static struct file *path_openat(struct nameidata *nd,
3460                         const struct open_flags *op, unsigned flags)
3461 {
3462         const char *s;
3463         struct file *file;
3464         int opened = 0;
3465         int error;
3466
3467         file = get_empty_filp();
3468         if (IS_ERR(file))
3469                 return file;
3470
3471         file->f_flags = op->open_flag;
3472
3473         if (unlikely(file->f_flags & __O_TMPFILE)) {
3474                 error = do_tmpfile(nd, flags, op, file, &opened);
3475                 goto out2;
3476         }
3477
3478         if (unlikely(file->f_flags & O_PATH)) {
3479                 error = do_o_path(nd, flags, file);
3480                 if (!error)
3481                         opened |= FILE_OPENED;
3482                 goto out2;
3483         }
3484
3485         s = path_init(nd, flags);
3486         if (IS_ERR(s)) {
3487                 put_filp(file);
3488                 return ERR_CAST(s);
3489         }
3490         while (!(error = link_path_walk(s, nd)) &&
3491                 (error = do_last(nd, file, op, &opened)) > 0) {
3492                 nd->flags &= ~(LOOKUP_OPEN|LOOKUP_CREATE|LOOKUP_EXCL);
3493                 s = trailing_symlink(nd);
3494                 if (IS_ERR(s)) {
3495                         error = PTR_ERR(s);
3496                         break;
3497                 }
3498         }
3499         terminate_walk(nd);
3500 out2:
3501         if (!(opened & FILE_OPENED)) {
3502                 BUG_ON(!error);
3503                 put_filp(file);
3504         }
3505         if (unlikely(error)) {
3506                 if (error == -EOPENSTALE) {
3507                         if (flags & LOOKUP_RCU)
3508                                 error = -ECHILD;
3509                         else
3510                                 error = -ESTALE;
3511                 }
3512                 file = ERR_PTR(error);
3513         }
3514         return file;
3515 }
3516
3517 struct file *do_filp_open(int dfd, struct filename *pathname,
3518                 const struct open_flags *op)
3519 {
3520         struct nameidata nd;
3521         int flags = op->lookup_flags;
3522         struct file *filp;
3523
3524         set_nameidata(&nd, dfd, pathname);
3525         filp = path_openat(&nd, op, flags | LOOKUP_RCU);
3526         if (unlikely(filp == ERR_PTR(-ECHILD)))
3527                 filp = path_openat(&nd, op, flags);
3528         if (unlikely(filp == ERR_PTR(-ESTALE)))
3529                 filp = path_openat(&nd, op, flags | LOOKUP_REVAL);
3530         restore_nameidata();
3531         return filp;
3532 }
3533
3534 struct file *do_file_open_root(struct dentry *dentry, struct vfsmount *mnt,
3535                 const char *name, const struct open_flags *op)
3536 {
3537         struct nameidata nd;
3538         struct file *file;
3539         struct filename *filename;
3540         int flags = op->lookup_flags | LOOKUP_ROOT;
3541
3542         nd.root.mnt = mnt;
3543         nd.root.dentry = dentry;
3544
3545         if (d_is_symlink(dentry) && op->intent & LOOKUP_OPEN)
3546                 return ERR_PTR(-ELOOP);
3547
3548         filename = getname_kernel(name);
3549         if (IS_ERR(filename))
3550                 return ERR_CAST(filename);
3551
3552         set_nameidata(&nd, -1, filename);
3553         file = path_openat(&nd, op, flags | LOOKUP_RCU);
3554         if (unlikely(file == ERR_PTR(-ECHILD)))
3555                 file = path_openat(&nd, op, flags);
3556         if (unlikely(file == ERR_PTR(-ESTALE)))
3557                 file = path_openat(&nd, op, flags | LOOKUP_REVAL);
3558         restore_nameidata();
3559         putname(filename);
3560         return file;
3561 }
3562
3563 static struct dentry *filename_create(int dfd, struct filename *name,
3564                                 struct path *path, unsigned int lookup_flags)
3565 {
3566         struct dentry *dentry = ERR_PTR(-EEXIST);
3567         struct qstr last;
3568         int type;
3569         int err2;
3570         int error;
3571         bool is_dir = (lookup_flags & LOOKUP_DIRECTORY);
3572
3573         /*
3574          * Note that only LOOKUP_REVAL and LOOKUP_DIRECTORY matter here. Any
3575          * other flags passed in are ignored!
3576          */
3577         lookup_flags &= LOOKUP_REVAL;
3578
3579         name = filename_parentat(dfd, name, lookup_flags, path, &last, &type);
3580         if (IS_ERR(name))
3581                 return ERR_CAST(name);
3582
3583         /*
3584          * Yucky last component or no last component at all?
3585          * (foo/., foo/.., /////)
3586          */
3587         if (unlikely(type != LAST_NORM))
3588                 goto out;
3589
3590         /* don't fail immediately if it's r/o, at least try to report other errors */
3591         err2 = mnt_want_write(path->mnt);
3592         /*
3593          * Do the final lookup.
3594          */
3595         lookup_flags |= LOOKUP_CREATE | LOOKUP_EXCL;
3596         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
3597         dentry = __lookup_hash(&last, path->dentry, lookup_flags);
3598         if (IS_ERR(dentry))
3599                 goto unlock;
3600
3601         error = -EEXIST;
3602         if (d_is_positive(dentry))
3603                 goto fail;
3604
3605         /*
3606          * Special case - lookup gave negative, but... we had foo/bar/
3607          * From the vfs_mknod() POV we just have a negative dentry -
3608          * all is fine. Let's be bastards - you had / on the end, you've
3609          * been asking for (non-existent) directory. -ENOENT for you.
3610          */
3611         if (unlikely(!is_dir && last.name[last.len])) {
3612                 error = -ENOENT;
3613                 goto fail;
3614         }
3615         if (unlikely(err2)) {
3616                 error = err2;
3617                 goto fail;
3618         }
3619         putname(name);
3620         return dentry;
3621 fail:
3622         dput(dentry);
3623         dentry = ERR_PTR(error);
3624 unlock:
3625         inode_unlock(path->dentry->d_inode);
3626         if (!err2)
3627                 mnt_drop_write(path->mnt);
3628 out:
3629         path_put(path);
3630         putname(name);
3631         return dentry;
3632 }
3633
3634 struct dentry *kern_path_create(int dfd, const char *pathname,
3635                                 struct path *path, unsigned int lookup_flags)
3636 {
3637         return filename_create(dfd, getname_kernel(pathname),
3638                                 path, lookup_flags);
3639 }
3640 EXPORT_SYMBOL(kern_path_create);
3641
3642 void done_path_create(struct path *path, struct dentry *dentry)
3643 {
3644         dput(dentry);
3645         inode_unlock(path->dentry->d_inode);
3646         mnt_drop_write(path->mnt);
3647         path_put(path);
3648 }
3649 EXPORT_SYMBOL(done_path_create);
3650
3651 inline struct dentry *user_path_create(int dfd, const char __user *pathname,
3652                                 struct path *path, unsigned int lookup_flags)
3653 {
3654         return filename_create(dfd, getname(pathname), path, lookup_flags);
3655 }
3656 EXPORT_SYMBOL(user_path_create);
3657
3658 int vfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3659 {
3660         int error = may_create(dir, dentry);
3661
3662         if (error)
3663                 return error;
3664
3665         if ((S_ISCHR(mode) || S_ISBLK(mode)) && !capable(CAP_MKNOD))
3666                 return -EPERM;
3667
3668         if (!dir->i_op->mknod)
3669                 return -EPERM;
3670
3671         error = devcgroup_inode_mknod(mode, dev);
3672         if (error)
3673                 return error;
3674
3675         error = security_inode_mknod(dir, dentry, mode, dev);
3676         if (error)
3677                 return error;
3678
3679         error = dir->i_op->mknod(dir, dentry, mode, dev);
3680         if (!error)
3681                 fsnotify_create(dir, dentry);
3682         return error;
3683 }
3684 EXPORT_SYMBOL(vfs_mknod);
3685
3686 static int may_mknod(umode_t mode)
3687 {
3688         switch (mode & S_IFMT) {
3689         case S_IFREG:
3690         case S_IFCHR:
3691         case S_IFBLK:
3692         case S_IFIFO:
3693         case S_IFSOCK:
3694         case 0: /* zero mode translates to S_IFREG */
3695                 return 0;
3696         case S_IFDIR:
3697                 return -EPERM;
3698         default:
3699                 return -EINVAL;
3700         }
3701 }
3702
3703 SYSCALL_DEFINE4(mknodat, int, dfd, const char __user *, filename, umode_t, mode,
3704                 unsigned, dev)
3705 {
3706         struct dentry *dentry;
3707         struct path path;
3708         int error;
3709         unsigned int lookup_flags = 0;
3710
3711         error = may_mknod(mode);
3712         if (error)
3713                 return error;
3714 retry:
3715         dentry = user_path_create(dfd, filename, &path, lookup_flags);
3716         if (IS_ERR(dentry))
3717                 return PTR_ERR(dentry);
3718
3719         if (!IS_POSIXACL(path.dentry->d_inode))
3720                 mode &= ~current_umask();
3721         error = security_path_mknod(&path, dentry, mode, dev);
3722         if (error)
3723                 goto out;
3724         switch (mode & S_IFMT) {
3725                 case 0: case S_IFREG:
3726                         error = vfs_create(path.dentry->d_inode,dentry,mode,true);
3727                         if (!error)
3728                                 ima_post_path_mknod(dentry);
3729                         break;
3730                 case S_IFCHR: case S_IFBLK:
3731                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,
3732                                         new_decode_dev(dev));
3733                         break;
3734                 case S_IFIFO: case S_IFSOCK:
3735                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,0);
3736                         break;
3737         }
3738 out:
3739         done_path_create(&path, dentry);
3740         if (retry_estale(error, lookup_flags)) {
3741                 lookup_flags |= LOOKUP_REVAL;
3742                 goto retry;
3743         }
3744         return error;
3745 }
3746
3747 SYSCALL_DEFINE3(mknod, const char __user *, filename, umode_t, mode, unsigned, dev)
3748 {
3749         return sys_mknodat(AT_FDCWD, filename, mode, dev);
3750 }
3751
3752 int vfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3753 {
3754         int error = may_create(dir, dentry);
3755         unsigned max_links = dir->i_sb->s_max_links;
3756
3757         if (error)
3758                 return error;
3759
3760         if (!dir->i_op->mkdir)
3761                 return -EPERM;
3762
3763         mode &= (S_IRWXUGO|S_ISVTX);
3764         error = security_inode_mkdir(dir, dentry, mode);
3765         if (error)
3766                 return error;
3767
3768         if (max_links && dir->i_nlink >= max_links)
3769                 return -EMLINK;
3770
3771         error = dir->i_op->mkdir(dir, dentry, mode);
3772         if (!error)
3773                 fsnotify_mkdir(dir, dentry);
3774         return error;
3775 }
3776 EXPORT_SYMBOL(vfs_mkdir);
3777
3778 SYSCALL_DEFINE3(mkdirat, int, dfd, const char __user *, pathname, umode_t, mode)
3779 {
3780         struct dentry *dentry;
3781         struct path path;
3782         int error;
3783         unsigned int lookup_flags = LOOKUP_DIRECTORY;
3784
3785 retry:
3786         dentry = user_path_create(dfd, pathname, &path, lookup_flags);
3787         if (IS_ERR(dentry))
3788                 return PTR_ERR(dentry);
3789
3790         if (!IS_POSIXACL(path.dentry->d_inode))
3791                 mode &= ~current_umask();
3792         error = security_path_mkdir(&path, dentry, mode);
3793         if (!error)
3794                 error = vfs_mkdir(path.dentry->d_inode, dentry, mode);
3795         done_path_create(&path, dentry);
3796         if (retry_estale(error, lookup_flags)) {
3797                 lookup_flags |= LOOKUP_REVAL;
3798                 goto retry;
3799         }
3800         return error;
3801 }
3802
3803 SYSCALL_DEFINE2(mkdir, const char __user *, pathname, umode_t, mode)
3804 {
3805         return sys_mkdirat(AT_FDCWD, pathname, mode);
3806 }
3807
3808 int vfs_rmdir(struct inode *dir, struct dentry *dentry)
3809 {
3810         int error = may_delete(dir, dentry, 1);
3811
3812         if (error)
3813                 return error;
3814
3815         if (!dir->i_op->rmdir)
3816                 return -EPERM;
3817
3818         dget(dentry);
3819         inode_lock(dentry->d_inode);
3820
3821         error = -EBUSY;
3822         if (is_local_mountpoint(dentry))
3823                 goto out;
3824
3825         error = security_inode_rmdir(dir, dentry);
3826         if (error)
3827                 goto out;
3828
3829         shrink_dcache_parent(dentry);
3830         error = dir->i_op->rmdir(dir, dentry);
3831         if (error)
3832                 goto out;
3833
3834         dentry->d_inode->i_flags |= S_DEAD;
3835         dont_mount(dentry);
3836         detach_mounts(dentry);
3837
3838 out:
3839         inode_unlock(dentry->d_inode);
3840         dput(dentry);
3841         if (!error)
3842                 d_delete(dentry);
3843         return error;
3844 }
3845 EXPORT_SYMBOL(vfs_rmdir);
3846
3847 static long do_rmdir(int dfd, const char __user *pathname)
3848 {
3849         int error = 0;
3850         struct filename *name;
3851         struct dentry *dentry;
3852         struct path path;
3853         struct qstr last;
3854         int type;
3855         unsigned int lookup_flags = 0;
3856 retry:
3857         name = user_path_parent(dfd, pathname,
3858                                 &path, &last, &type, lookup_flags);
3859         if (IS_ERR(name))
3860                 return PTR_ERR(name);
3861
3862         switch (type) {
3863         case LAST_DOTDOT:
3864                 error = -ENOTEMPTY;
3865                 goto exit1;
3866         case LAST_DOT:
3867                 error = -EINVAL;
3868                 goto exit1;
3869         case LAST_ROOT:
3870                 error = -EBUSY;
3871                 goto exit1;
3872         }
3873
3874         error = mnt_want_write(path.mnt);
3875         if (error)
3876                 goto exit1;
3877
3878         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3879         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3880         error = PTR_ERR(dentry);
3881         if (IS_ERR(dentry))
3882                 goto exit2;
3883         if (!dentry->d_inode) {
3884                 error = -ENOENT;
3885                 goto exit3;
3886         }
3887         error = security_path_rmdir(&path, dentry);
3888         if (error)
3889                 goto exit3;
3890         error = vfs_rmdir(path.dentry->d_inode, dentry);
3891 exit3:
3892         dput(dentry);
3893 exit2:
3894         inode_unlock(path.dentry->d_inode);
3895         mnt_drop_write(path.mnt);
3896 exit1:
3897         path_put(&path);
3898         putname(name);
3899         if (retry_estale(error, lookup_flags)) {
3900                 lookup_flags |= LOOKUP_REVAL;
3901                 goto retry;
3902         }
3903         return error;
3904 }
3905
3906 SYSCALL_DEFINE1(rmdir, const char __user *, pathname)
3907 {
3908         return do_rmdir(AT_FDCWD, pathname);
3909 }
3910
3911 /**
3912  * vfs_unlink - unlink a filesystem object
3913  * @dir:        parent directory
3914  * @dentry:     victim
3915  * @delegated_inode: returns victim inode, if the inode is delegated.
3916  *
3917  * The caller must hold dir->i_mutex.
3918  *
3919  * If vfs_unlink discovers a delegation, it will return -EWOULDBLOCK and
3920  * return a reference to the inode in delegated_inode.  The caller
3921  * should then break the delegation on that inode and retry.  Because
3922  * breaking a delegation may take a long time, the caller should drop
3923  * dir->i_mutex before doing so.
3924  *
3925  * Alternatively, a caller may pass NULL for delegated_inode.  This may
3926  * be appropriate for callers that expect the underlying filesystem not
3927  * to be NFS exported.
3928  */
3929 int vfs_unlink(struct inode *dir, struct dentry *dentry, struct inode **delegated_inode)
3930 {
3931         struct inode *target = dentry->d_inode;
3932         int error = may_delete(dir, dentry, 0);
3933
3934         if (error)
3935                 return error;
3936
3937         if (!dir->i_op->unlink)
3938                 return -EPERM;
3939
3940         inode_lock(target);
3941         if (is_local_mountpoint(dentry))
3942                 error = -EBUSY;
3943         else {
3944                 error = security_inode_unlink(dir, dentry);
3945                 if (!error) {
3946                         error = try_break_deleg(target, delegated_inode);
3947                         if (error)
3948                                 goto out;
3949                         error = dir->i_op->unlink(dir, dentry);
3950                         if (!error) {
3951                                 dont_mount(dentry);
3952                                 detach_mounts(dentry);
3953                         }
3954                 }
3955         }
3956 out:
3957         inode_unlock(target);
3958
3959         /* We don't d_delete() NFS sillyrenamed files--they still exist. */
3960         if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
3961                 fsnotify_link_count(target);
3962                 d_delete(dentry);
3963         }
3964
3965         return error;
3966 }
3967 EXPORT_SYMBOL(vfs_unlink);
3968
3969 /*
3970  * Make sure that the actual truncation of the file will occur outside its
3971  * directory's i_mutex.  Truncate can take a long time if there is a lot of
3972  * writeout happening, and we don't want to prevent access to the directory
3973  * while waiting on the I/O.
3974  */
3975 static long do_unlinkat(int dfd, const char __user *pathname)
3976 {
3977         int error;
3978         struct filename *name;
3979         struct dentry *dentry;
3980         struct path path;
3981         struct qstr last;
3982         int type;
3983         struct inode *inode = NULL;
3984         struct inode *delegated_inode = NULL;
3985         unsigned int lookup_flags = 0;
3986 retry:
3987         name = user_path_parent(dfd, pathname,
3988                                 &path, &last, &type, lookup_flags);
3989         if (IS_ERR(name))
3990                 return PTR_ERR(name);
3991
3992         error = -EISDIR;
3993         if (type != LAST_NORM)
3994                 goto exit1;
3995
3996         error = mnt_want_write(path.mnt);
3997         if (error)
3998                 goto exit1;
3999 retry_deleg:
4000         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
4001         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
4002         error = PTR_ERR(dentry);
4003         if (!IS_ERR(dentry)) {
4004                 /* Why not before? Because we want correct error value */
4005                 if (last.name[last.len])
4006                         goto slashes;
4007                 inode = dentry->d_inode;
4008                 if (d_is_negative(dentry))
4009                         goto slashes;
4010                 ihold(inode);
4011                 error = security_path_unlink(&path, dentry);
4012                 if (error)
4013                         goto exit2;
4014                 error = vfs_unlink(path.dentry->d_inode, dentry, &delegated_inode);
4015 exit2:
4016                 dput(dentry);
4017         }
4018         inode_unlock(path.dentry->d_inode);
4019         if (inode)
4020                 iput(inode);    /* truncate the inode here */
4021         inode = NULL;
4022         if (delegated_inode) {
4023                 error = break_deleg_wait(&delegated_inode);
4024                 if (!error)
4025                         goto retry_deleg;
4026         }
4027         mnt_drop_write(path.mnt);
4028 exit1:
4029         path_put(&path);
4030         putname(name);
4031         if (retry_estale(error, lookup_flags)) {
4032                 lookup_flags |= LOOKUP_REVAL;
4033                 inode = NULL;
4034                 goto retry;
4035         }
4036         return error;
4037
4038 slashes:
4039         if (d_is_negative(dentry))
4040                 error = -ENOENT;
4041         else if (d_is_dir(dentry))
4042                 error = -EISDIR;
4043         else
4044                 error = -ENOTDIR;
4045         goto exit2;
4046 }
4047
4048 SYSCALL_DEFINE3(unlinkat, int, dfd, const char __user *, pathname, int, flag)
4049 {
4050         if ((flag & ~AT_REMOVEDIR) != 0)
4051                 return -EINVAL;
4052
4053         if (flag & AT_REMOVEDIR)
4054                 return do_rmdir(dfd, pathname);
4055
4056         return do_unlinkat(dfd, pathname);
4057 }
4058
4059 SYSCALL_DEFINE1(unlink, const char __user *, pathname)
4060 {
4061         return do_unlinkat(AT_FDCWD, pathname);
4062 }
4063
4064 int vfs_symlink(struct inode *dir, struct dentry *dentry, const char *oldname)
4065 {
4066         int error = may_create(dir, dentry);
4067
4068         if (error)
4069                 return error;
4070
4071         if (!dir->i_op->symlink)
4072                 return -EPERM;
4073
4074         error = security_inode_symlink(dir, dentry, oldname);
4075         if (error)
4076                 return error;
4077
4078         error = dir->i_op->symlink(dir, dentry, oldname);
4079         if (!error)
4080                 fsnotify_create(dir, dentry);
4081         return error;
4082 }
4083 EXPORT_SYMBOL(vfs_symlink);
4084
4085 SYSCALL_DEFINE3(symlinkat, const char __user *, oldname,
4086                 int, newdfd, const char __user *, newname)
4087 {
4088         int error;
4089         struct filename *from;
4090         struct dentry *dentry;
4091         struct path path;
4092         unsigned int lookup_flags = 0;
4093
4094         from = getname(oldname);
4095         if (IS_ERR(from))
4096                 return PTR_ERR(from);
4097 retry:
4098         dentry = user_path_create(newdfd, newname, &path, lookup_flags);
4099         error = PTR_ERR(dentry);
4100         if (IS_ERR(dentry))
4101                 goto out_putname;
4102
4103         error = security_path_symlink(&path, dentry, from->name);
4104         if (!error)
4105                 error = vfs_symlink(path.dentry->d_inode, dentry, from->name);
4106         done_path_create(&path, dentry);
4107         if (retry_estale(error, lookup_flags)) {
4108                 lookup_flags |= LOOKUP_REVAL;
4109                 goto retry;
4110         }
4111 out_putname:
4112         putname(from);
4113         return error;
4114 }
4115
4116 SYSCALL_DEFINE2(symlink, const char __user *, oldname, const char __user *, newname)
4117 {
4118         return sys_symlinkat(oldname, AT_FDCWD, newname);
4119 }
4120
4121 /**
4122  * vfs_link - create a new link
4123  * @old_dentry: object to be linked
4124  * @dir:        new parent
4125  * @new_dentry: where to create the new link
4126  * @delegated_inode: returns inode needing a delegation break
4127  *
4128  * The caller must hold dir->i_mutex
4129  *
4130  * If vfs_link discovers a delegation on the to-be-linked file in need
4131  * of breaking, it will return -EWOULDBLOCK and return a reference to the
4132  * inode in delegated_inode.  The caller should then break the delegation
4133  * and retry.  Because breaking a delegation may take a long time, the
4134  * caller should drop the i_mutex before doing so.
4135  *
4136  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4137  * be appropriate for callers that expect the underlying filesystem not
4138  * to be NFS exported.
4139  */
4140 int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_dentry, struct inode **delegated_inode)
4141 {
4142         struct inode *inode = old_dentry->d_inode;
4143         unsigned max_links = dir->i_sb->s_max_links;
4144         int error;
4145
4146         if (!inode)
4147                 return -ENOENT;
4148
4149         error = may_create(dir, new_dentry);
4150         if (error)
4151                 return error;
4152
4153         if (dir->i_sb != inode->i_sb)
4154                 return -EXDEV;
4155
4156         /*
4157          * A link to an append-only or immutable file cannot be created.
4158          */
4159         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4160                 return -EPERM;
4161         if (!dir->i_op->link)
4162                 return -EPERM;
4163         if (S_ISDIR(inode->i_mode))
4164                 return -EPERM;
4165
4166         error = security_inode_link(old_dentry, dir, new_dentry);
4167         if (error)
4168                 return error;
4169
4170         inode_lock(inode);
4171         /* Make sure we don't allow creating hardlink to an unlinked file */
4172         if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))
4173                 error =  -ENOENT;
4174         else if (max_links && inode->i_nlink >= max_links)
4175                 error = -EMLINK;
4176         else {
4177                 error = try_break_deleg(inode, delegated_inode);
4178                 if (!error)
4179                         error = dir->i_op->link(old_dentry, dir, new_dentry);
4180         }
4181
4182         if (!error && (inode->i_state & I_LINKABLE)) {
4183                 spin_lock(&inode->i_lock);
4184                 inode->i_state &= ~I_LINKABLE;
4185                 spin_unlock(&inode->i_lock);
4186         }
4187         inode_unlock(inode);
4188         if (!error)
4189                 fsnotify_link(dir, inode, new_dentry);
4190         return error;
4191 }
4192 EXPORT_SYMBOL(vfs_link);
4193
4194 /*
4195  * Hardlinks are often used in delicate situations.  We avoid
4196  * security-related surprises by not following symlinks on the
4197  * newname.  --KAB
4198  *
4199  * We don't follow them on the oldname either to be compatible
4200  * with linux 2.0, and to avoid hard-linking to directories
4201  * and other special files.  --ADM
4202  */
4203 SYSCALL_DEFINE5(linkat, int, olddfd, const char __user *, oldname,
4204                 int, newdfd, const char __user *, newname, int, flags)
4205 {
4206         struct dentry *new_dentry;
4207         struct path old_path, new_path;
4208         struct inode *delegated_inode = NULL;
4209         int how = 0;
4210         int error;
4211
4212         if ((flags & ~(AT_SYMLINK_FOLLOW | AT_EMPTY_PATH)) != 0)
4213                 return -EINVAL;
4214         /*
4215          * To use null names we require CAP_DAC_READ_SEARCH
4216          * This ensures that not everyone will be able to create
4217          * handlink using the passed filedescriptor.
4218          */
4219         if (flags & AT_EMPTY_PATH) {
4220                 if (!capable(CAP_DAC_READ_SEARCH))
4221                         return -ENOENT;
4222                 how = LOOKUP_EMPTY;
4223         }
4224
4225         if (flags & AT_SYMLINK_FOLLOW)
4226                 how |= LOOKUP_FOLLOW;
4227 retry:
4228         error = user_path_at(olddfd, oldname, how, &old_path);
4229         if (error)
4230                 return error;
4231
4232         new_dentry = user_path_create(newdfd, newname, &new_path,
4233                                         (how & LOOKUP_REVAL));
4234         error = PTR_ERR(new_dentry);
4235         if (IS_ERR(new_dentry))
4236                 goto out;
4237
4238         error = -EXDEV;
4239         if (old_path.mnt != new_path.mnt)
4240                 goto out_dput;
4241         error = may_linkat(&old_path);
4242         if (unlikely(error))
4243                 goto out_dput;
4244         error = security_path_link(old_path.dentry, &new_path, new_dentry);
4245         if (error)
4246                 goto out_dput;
4247         error = vfs_link(old_path.dentry, new_path.dentry->d_inode, new_dentry, &delegated_inode);
4248 out_dput:
4249         done_path_create(&new_path, new_dentry);
4250         if (delegated_inode) {
4251                 error = break_deleg_wait(&delegated_inode);
4252                 if (!error) {
4253                         path_put(&old_path);
4254                         goto retry;
4255                 }
4256         }
4257         if (retry_estale(error, how)) {
4258                 path_put(&old_path);
4259                 how |= LOOKUP_REVAL;
4260                 goto retry;
4261         }
4262 out:
4263         path_put(&old_path);
4264
4265         return error;
4266 }
4267
4268 SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname)
4269 {
4270         return sys_linkat(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4271 }
4272
4273 /**
4274  * vfs_rename - rename a filesystem object
4275  * @old_dir:    parent of source
4276  * @old_dentry: source
4277  * @new_dir:    parent of destination
4278  * @new_dentry: destination
4279  * @delegated_inode: returns an inode needing a delegation break
4280  * @flags:      rename flags
4281  *
4282  * The caller must hold multiple mutexes--see lock_rename()).
4283  *
4284  * If vfs_rename discovers a delegation in need of breaking at either
4285  * the source or destination, it will return -EWOULDBLOCK and return a
4286  * reference to the inode in delegated_inode.  The caller should then
4287  * break the delegation and retry.  Because breaking a delegation may
4288  * take a long time, the caller should drop all locks before doing
4289  * so.
4290  *
4291  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4292  * be appropriate for callers that expect the underlying filesystem not
4293  * to be NFS exported.
4294  *
4295  * The worst of all namespace operations - renaming directory. "Perverted"
4296  * doesn't even start to describe it. Somebody in UCB had a heck of a trip...
4297  * Problems:
4298  *      a) we can get into loop creation.
4299  *      b) race potential - two innocent renames can create a loop together.
4300  *         That's where 4.4 screws up. Current fix: serialization on
4301  *         sb->s_vfs_rename_mutex. We might be more accurate, but that's another
4302  *         story.
4303  *      c) we have to lock _four_ objects - parents and victim (if it exists),
4304  *         and source (if it is not a directory).
4305  *         And that - after we got ->i_mutex on parents (until then we don't know
4306  *         whether the target exists).  Solution: try to be smart with locking
4307  *         order for inodes.  We rely on the fact that tree topology may change
4308  *         only under ->s_vfs_rename_mutex _and_ that parent of the object we
4309  *         move will be locked.  Thus we can rank directories by the tree
4310  *         (ancestors first) and rank all non-directories after them.
4311  *         That works since everybody except rename does "lock parent, lookup,
4312  *         lock child" and rename is under ->s_vfs_rename_mutex.
4313  *         HOWEVER, it relies on the assumption that any object with ->lookup()
4314  *         has no more than 1 dentry.  If "hybrid" objects will ever appear,
4315  *         we'd better make sure that there's no link(2) for them.
4316  *      d) conversion from fhandle to dentry may come in the wrong moment - when
4317  *         we are removing the target. Solution: we will have to grab ->i_mutex
4318  *         in the fhandle_to_dentry code. [FIXME - current nfsfh.c relies on
4319  *         ->i_mutex on parents, which works but leads to some truly excessive
4320  *         locking].
4321  */
4322 int vfs_rename(struct inode *old_dir, struct dentry *old_dentry,
4323                struct inode *new_dir, struct dentry *new_dentry,
4324                struct inode **delegated_inode, unsigned int flags)
4325 {
4326         int error;
4327         bool is_dir = d_is_dir(old_dentry);
4328         const unsigned char *old_name;
4329         struct inode *source = old_dentry->d_inode;
4330         struct inode *target = new_dentry->d_inode;
4331         bool new_is_dir = false;
4332         unsigned max_links = new_dir->i_sb->s_max_links;
4333
4334         /*
4335          * Check source == target.
4336          * On overlayfs need to look at underlying inodes.
4337          */
4338         if (d_real_inode(old_dentry) == d_real_inode(new_dentry))
4339                 return 0;
4340
4341         error = may_delete(old_dir, old_dentry, is_dir);
4342         if (error)
4343                 return error;
4344
4345         if (!target) {
4346                 error = may_create(new_dir, new_dentry);
4347         } else {
4348                 new_is_dir = d_is_dir(new_dentry);
4349
4350                 if (!(flags & RENAME_EXCHANGE))
4351                         error = may_delete(new_dir, new_dentry, is_dir);
4352                 else
4353                         error = may_delete(new_dir, new_dentry, new_is_dir);
4354         }
4355         if (error)
4356                 return error;
4357
4358         if (!old_dir->i_op->rename && !old_dir->i_op->rename2)
4359                 return -EPERM;
4360
4361         if (flags && !old_dir->i_op->rename2)
4362                 return -EINVAL;
4363
4364         /*
4365          * If we are going to change the parent - check write permissions,
4366          * we'll need to flip '..'.
4367          */
4368         if (new_dir != old_dir) {
4369                 if (is_dir) {
4370                         error = inode_permission(source, MAY_WRITE);
4371                         if (error)
4372                                 return error;
4373                 }
4374                 if ((flags & RENAME_EXCHANGE) && new_is_dir) {
4375                         error = inode_permission(target, MAY_WRITE);
4376                         if (error)
4377                                 return error;
4378                 }
4379         }
4380
4381         error = security_inode_rename(old_dir, old_dentry, new_dir, new_dentry,
4382                                       flags);
4383         if (error)
4384                 return error;
4385
4386         old_name = fsnotify_oldname_init(old_dentry->d_name.name);
4387         dget(new_dentry);
4388         if (!is_dir || (flags & RENAME_EXCHANGE))
4389                 lock_two_nondirectories(source, target);
4390         else if (target)
4391                 inode_lock(target);
4392
4393         error = -EBUSY;
4394         if (is_local_mountpoint(old_dentry) || is_local_mountpoint(new_dentry))
4395                 goto out;
4396
4397         if (max_links && new_dir != old_dir) {
4398                 error = -EMLINK;
4399                 if (is_dir && !new_is_dir && new_dir->i_nlink >= max_links)
4400                         goto out;
4401                 if ((flags & RENAME_EXCHANGE) && !is_dir && new_is_dir &&
4402                     old_dir->i_nlink >= max_links)
4403                         goto out;
4404         }
4405         if (is_dir && !(flags & RENAME_EXCHANGE) && target)
4406                 shrink_dcache_parent(new_dentry);
4407         if (!is_dir) {
4408                 error = try_break_deleg(source, delegated_inode);
4409                 if (error)
4410                         goto out;
4411         }
4412         if (target && !new_is_dir) {
4413                 error = try_break_deleg(target, delegated_inode);
4414                 if (error)
4415                         goto out;
4416         }
4417         if (!old_dir->i_op->rename2) {
4418                 error = old_dir->i_op->rename(old_dir, old_dentry,
4419                                               new_dir, new_dentry);
4420         } else {
4421                 WARN_ON(old_dir->i_op->rename != NULL);
4422                 error = old_dir->i_op->rename2(old_dir, old_dentry,
4423                                                new_dir, new_dentry, flags);
4424         }
4425         if (error)
4426                 goto out;
4427
4428         if (!(flags & RENAME_EXCHANGE) && target) {
4429                 if (is_dir)
4430                         target->i_flags |= S_DEAD;
4431                 dont_mount(new_dentry);
4432                 detach_mounts(new_dentry);
4433         }
4434         if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE)) {
4435                 if (!(flags & RENAME_EXCHANGE))
4436                         d_move(old_dentry, new_dentry);
4437                 else
4438                         d_exchange(old_dentry, new_dentry);
4439         }
4440 out:
4441         if (!is_dir || (flags & RENAME_EXCHANGE))
4442                 unlock_two_nondirectories(source, target);
4443         else if (target)
4444                 inode_unlock(target);
4445         dput(new_dentry);
4446         if (!error) {
4447                 fsnotify_move(old_dir, new_dir, old_name, is_dir,
4448                               !(flags & RENAME_EXCHANGE) ? target : NULL, old_dentry);
4449                 if (flags & RENAME_EXCHANGE) {
4450                         fsnotify_move(new_dir, old_dir, old_dentry->d_name.name,
4451                                       new_is_dir, NULL, new_dentry);
4452                 }
4453         }
4454         fsnotify_oldname_free(old_name);
4455
4456         return error;
4457 }
4458 EXPORT_SYMBOL(vfs_rename);
4459
4460 SYSCALL_DEFINE5(renameat2, int, olddfd, const char __user *, oldname,
4461                 int, newdfd, const char __user *, newname, unsigned int, flags)
4462 {
4463         struct dentry *old_dentry, *new_dentry;
4464         struct dentry *trap;
4465         struct path old_path, new_path;
4466         struct qstr old_last, new_last;
4467         int old_type, new_type;
4468         struct inode *delegated_inode = NULL;
4469         struct filename *from;
4470         struct filename *to;
4471         unsigned int lookup_flags = 0, target_flags = LOOKUP_RENAME_TARGET;
4472         bool should_retry = false;
4473         int error;
4474
4475         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
4476                 return -EINVAL;
4477
4478         if ((flags & (RENAME_NOREPLACE | RENAME_WHITEOUT)) &&
4479             (flags & RENAME_EXCHANGE))
4480                 return -EINVAL;
4481
4482         if ((flags & RENAME_WHITEOUT) && !capable(CAP_MKNOD))
4483                 return -EPERM;
4484
4485         if (flags & RENAME_EXCHANGE)
4486                 target_flags = 0;
4487
4488 retry:
4489         from = user_path_parent(olddfd, oldname,
4490                                 &old_path, &old_last, &old_type, lookup_flags);
4491         if (IS_ERR(from)) {
4492                 error = PTR_ERR(from);
4493                 goto exit;
4494         }
4495
4496         to = user_path_parent(newdfd, newname,
4497                                 &new_path, &new_last, &new_type, lookup_flags);
4498         if (IS_ERR(to)) {
4499                 error = PTR_ERR(to);
4500                 goto exit1;
4501         }
4502
4503         error = -EXDEV;
4504         if (old_path.mnt != new_path.mnt)
4505                 goto exit2;
4506
4507         error = -EBUSY;
4508         if (old_type != LAST_NORM)
4509                 goto exit2;
4510
4511         if (flags & RENAME_NOREPLACE)
4512                 error = -EEXIST;
4513         if (new_type != LAST_NORM)
4514                 goto exit2;
4515
4516         error = mnt_want_write(old_path.mnt);
4517         if (error)
4518                 goto exit2;
4519
4520 retry_deleg:
4521         trap = lock_rename(new_path.dentry, old_path.dentry);
4522
4523         old_dentry = __lookup_hash(&old_last, old_path.dentry, lookup_flags);
4524         error = PTR_ERR(old_dentry);
4525         if (IS_ERR(old_dentry))
4526                 goto exit3;
4527         /* source must exist */
4528         error = -ENOENT;
4529         if (d_is_negative(old_dentry))
4530                 goto exit4;
4531         new_dentry = __lookup_hash(&new_last, new_path.dentry, lookup_flags | target_flags);
4532         error = PTR_ERR(new_dentry);
4533         if (IS_ERR(new_dentry))
4534                 goto exit4;
4535         error = -EEXIST;
4536         if ((flags & RENAME_NOREPLACE) && d_is_positive(new_dentry))
4537                 goto exit5;
4538         if (flags & RENAME_EXCHANGE) {
4539                 error = -ENOENT;
4540                 if (d_is_negative(new_dentry))
4541                         goto exit5;
4542
4543                 if (!d_is_dir(new_dentry)) {
4544                         error = -ENOTDIR;
4545                         if (new_last.name[new_last.len])
4546                                 goto exit5;
4547                 }
4548         }
4549         /* unless the source is a directory trailing slashes give -ENOTDIR */
4550         if (!d_is_dir(old_dentry)) {
4551                 error = -ENOTDIR;
4552                 if (old_last.name[old_last.len])
4553                         goto exit5;
4554                 if (!(flags & RENAME_EXCHANGE) && new_last.name[new_last.len])
4555                         goto exit5;
4556         }
4557         /* source should not be ancestor of target */
4558         error = -EINVAL;
4559         if (old_dentry == trap)
4560                 goto exit5;
4561         /* target should not be an ancestor of source */
4562         if (!(flags & RENAME_EXCHANGE))
4563                 error = -ENOTEMPTY;
4564         if (new_dentry == trap)
4565                 goto exit5;
4566
4567         error = security_path_rename(&old_path, old_dentry,
4568                                      &new_path, new_dentry, flags);
4569         if (error)
4570                 goto exit5;
4571         error = vfs_rename(old_path.dentry->d_inode, old_dentry,
4572                            new_path.dentry->d_inode, new_dentry,
4573                            &delegated_inode, flags);
4574 exit5:
4575         dput(new_dentry);
4576 exit4:
4577         dput(old_dentry);
4578 exit3:
4579         unlock_rename(new_path.dentry, old_path.dentry);
4580         if (delegated_inode) {
4581                 error = break_deleg_wait(&delegated_inode);
4582                 if (!error)
4583                         goto retry_deleg;
4584         }
4585         mnt_drop_write(old_path.mnt);
4586 exit2:
4587         if (retry_estale(error, lookup_flags))
4588                 should_retry = true;
4589         path_put(&new_path);
4590         putname(to);
4591 exit1:
4592         path_put(&old_path);
4593         putname(from);
4594         if (should_retry) {
4595                 should_retry = false;
4596                 lookup_flags |= LOOKUP_REVAL;
4597                 goto retry;
4598         }
4599 exit:
4600         return error;
4601 }
4602
4603 SYSCALL_DEFINE4(renameat, int, olddfd, const char __user *, oldname,
4604                 int, newdfd, const char __user *, newname)
4605 {
4606         return sys_renameat2(olddfd, oldname, newdfd, newname, 0);
4607 }
4608
4609 SYSCALL_DEFINE2(rename, const char __user *, oldname, const char __user *, newname)
4610 {
4611         return sys_renameat2(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4612 }
4613
4614 int vfs_whiteout(struct inode *dir, struct dentry *dentry)
4615 {
4616         int error = may_create(dir, dentry);
4617         if (error)
4618                 return error;
4619
4620         if (!dir->i_op->mknod)
4621                 return -EPERM;
4622
4623         return dir->i_op->mknod(dir, dentry,
4624                                 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
4625 }
4626 EXPORT_SYMBOL(vfs_whiteout);
4627
4628 int readlink_copy(char __user *buffer, int buflen, const char *link)
4629 {
4630         int len = PTR_ERR(link);
4631         if (IS_ERR(link))
4632                 goto out;
4633
4634         len = strlen(link);
4635         if (len > (unsigned) buflen)
4636                 len = buflen;
4637         if (copy_to_user(buffer, link, len))
4638                 len = -EFAULT;
4639 out:
4640         return len;
4641 }
4642
4643 /*
4644  * A helper for ->readlink().  This should be used *ONLY* for symlinks that
4645  * have ->get_link() not calling nd_jump_link().  Using (or not using) it
4646  * for any given inode is up to filesystem.
4647  */
4648 int generic_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4649 {
4650         DEFINE_DELAYED_CALL(done);
4651         struct inode *inode = d_inode(dentry);
4652         const char *link = inode->i_link;
4653         int res;
4654
4655         if (!link) {
4656                 link = inode->i_op->get_link(dentry, inode, &done);
4657                 if (IS_ERR(link))
4658                         return PTR_ERR(link);
4659         }
4660         res = readlink_copy(buffer, buflen, link);
4661         do_delayed_call(&done);
4662         return res;
4663 }
4664 EXPORT_SYMBOL(generic_readlink);
4665
4666 /* get the link contents into pagecache */
4667 const char *page_get_link(struct dentry *dentry, struct inode *inode,
4668                           struct delayed_call *callback)
4669 {
4670         char *kaddr;
4671         struct page *page;
4672         struct address_space *mapping = inode->i_mapping;
4673
4674         if (!dentry) {
4675                 page = find_get_page(mapping, 0);
4676                 if (!page)
4677                         return ERR_PTR(-ECHILD);
4678                 if (!PageUptodate(page)) {
4679                         put_page(page);
4680                         return ERR_PTR(-ECHILD);
4681                 }
4682         } else {
4683                 page = read_mapping_page(mapping, 0, NULL);
4684                 if (IS_ERR(page))
4685                         return (char*)page;
4686         }
4687         set_delayed_call(callback, page_put_link, page);
4688         BUG_ON(mapping_gfp_mask(mapping) & __GFP_HIGHMEM);
4689         kaddr = page_address(page);
4690         nd_terminate_link(kaddr, inode->i_size, PAGE_SIZE - 1);
4691         return kaddr;
4692 }
4693
4694 EXPORT_SYMBOL(page_get_link);
4695
4696 void page_put_link(void *arg)
4697 {
4698         put_page(arg);
4699 }
4700 EXPORT_SYMBOL(page_put_link);
4701
4702 int page_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4703 {
4704         DEFINE_DELAYED_CALL(done);
4705         int res = readlink_copy(buffer, buflen,
4706                                 page_get_link(dentry, d_inode(dentry),
4707                                               &done));
4708         do_delayed_call(&done);
4709         return res;
4710 }
4711 EXPORT_SYMBOL(page_readlink);
4712
4713 /*
4714  * The nofs argument instructs pagecache_write_begin to pass AOP_FLAG_NOFS
4715  */
4716 int __page_symlink(struct inode *inode, const char *symname, int len, int nofs)
4717 {
4718         struct address_space *mapping = inode->i_mapping;
4719         struct page *page;
4720         void *fsdata;
4721         int err;
4722         unsigned int flags = AOP_FLAG_UNINTERRUPTIBLE;
4723         if (nofs)
4724                 flags |= AOP_FLAG_NOFS;
4725
4726 retry:
4727         err = pagecache_write_begin(NULL, mapping, 0, len-1,
4728                                 flags, &page, &fsdata);
4729         if (err)
4730                 goto fail;
4731
4732         memcpy(page_address(page), symname, len-1);
4733
4734         err = pagecache_write_end(NULL, mapping, 0, len-1, len-1,
4735                                                         page, fsdata);
4736         if (err < 0)
4737                 goto fail;
4738         if (err < len-1)
4739                 goto retry;
4740
4741         mark_inode_dirty(inode);
4742         return 0;
4743 fail:
4744         return err;
4745 }
4746 EXPORT_SYMBOL(__page_symlink);
4747
4748 int page_symlink(struct inode *inode, const char *symname, int len)
4749 {
4750         return __page_symlink(inode, symname, len,
4751                         !mapping_gfp_constraint(inode->i_mapping, __GFP_FS));
4752 }
4753 EXPORT_SYMBOL(page_symlink);
4754
4755 const struct inode_operations page_symlink_inode_operations = {
4756         .readlink       = generic_readlink,
4757         .get_link       = page_get_link,
4758 };
4759 EXPORT_SYMBOL(page_symlink_inode_operations);