Merge branch 'parisc-4.10-1' of git://git.kernel.org/pub/scm/linux/kernel/git/deller...
[sfrench/cifs-2.6.git] / fs / namei.c
1 /*
2  *  linux/fs/namei.c
3  *
4  *  Copyright (C) 1991, 1992  Linus Torvalds
5  */
6
7 /*
8  * Some corrections by tytso.
9  */
10
11 /* [Feb 1997 T. Schoebel-Theuer] Complete rewrite of the pathname
12  * lookup logic.
13  */
14 /* [Feb-Apr 2000, AV] Rewrite to the new namespace architecture.
15  */
16
17 #include <linux/init.h>
18 #include <linux/export.h>
19 #include <linux/kernel.h>
20 #include <linux/slab.h>
21 #include <linux/fs.h>
22 #include <linux/namei.h>
23 #include <linux/pagemap.h>
24 #include <linux/fsnotify.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/ima.h>
28 #include <linux/syscalls.h>
29 #include <linux/mount.h>
30 #include <linux/audit.h>
31 #include <linux/capability.h>
32 #include <linux/file.h>
33 #include <linux/fcntl.h>
34 #include <linux/device_cgroup.h>
35 #include <linux/fs_struct.h>
36 #include <linux/posix_acl.h>
37 #include <linux/hash.h>
38 #include <linux/bitops.h>
39 #include <linux/init_task.h>
40 #include <asm/uaccess.h>
41
42 #include "internal.h"
43 #include "mount.h"
44
45 /* [Feb-1997 T. Schoebel-Theuer]
46  * Fundamental changes in the pathname lookup mechanisms (namei)
47  * were necessary because of omirr.  The reason is that omirr needs
48  * to know the _real_ pathname, not the user-supplied one, in case
49  * of symlinks (and also when transname replacements occur).
50  *
51  * The new code replaces the old recursive symlink resolution with
52  * an iterative one (in case of non-nested symlink chains).  It does
53  * this with calls to <fs>_follow_link().
54  * As a side effect, dir_namei(), _namei() and follow_link() are now 
55  * replaced with a single function lookup_dentry() that can handle all 
56  * the special cases of the former code.
57  *
58  * With the new dcache, the pathname is stored at each inode, at least as
59  * long as the refcount of the inode is positive.  As a side effect, the
60  * size of the dcache depends on the inode cache and thus is dynamic.
61  *
62  * [29-Apr-1998 C. Scott Ananian] Updated above description of symlink
63  * resolution to correspond with current state of the code.
64  *
65  * Note that the symlink resolution is not *completely* iterative.
66  * There is still a significant amount of tail- and mid- recursion in
67  * the algorithm.  Also, note that <fs>_readlink() is not used in
68  * lookup_dentry(): lookup_dentry() on the result of <fs>_readlink()
69  * may return different results than <fs>_follow_link().  Many virtual
70  * filesystems (including /proc) exhibit this behavior.
71  */
72
73 /* [24-Feb-97 T. Schoebel-Theuer] Side effects caused by new implementation:
74  * New symlink semantics: when open() is called with flags O_CREAT | O_EXCL
75  * and the name already exists in form of a symlink, try to create the new
76  * name indicated by the symlink. The old code always complained that the
77  * name already exists, due to not following the symlink even if its target
78  * is nonexistent.  The new semantics affects also mknod() and link() when
79  * the name is a symlink pointing to a non-existent name.
80  *
81  * I don't know which semantics is the right one, since I have no access
82  * to standards. But I found by trial that HP-UX 9.0 has the full "new"
83  * semantics implemented, while SunOS 4.1.1 and Solaris (SunOS 5.4) have the
84  * "old" one. Personally, I think the new semantics is much more logical.
85  * Note that "ln old new" where "new" is a symlink pointing to a non-existing
86  * file does succeed in both HP-UX and SunOs, but not in Solaris
87  * and in the old Linux semantics.
88  */
89
90 /* [16-Dec-97 Kevin Buhr] For security reasons, we change some symlink
91  * semantics.  See the comments in "open_namei" and "do_link" below.
92  *
93  * [10-Sep-98 Alan Modra] Another symlink change.
94  */
95
96 /* [Feb-Apr 2000 AV] Complete rewrite. Rules for symlinks:
97  *      inside the path - always follow.
98  *      in the last component in creation/removal/renaming - never follow.
99  *      if LOOKUP_FOLLOW passed - follow.
100  *      if the pathname has trailing slashes - follow.
101  *      otherwise - don't follow.
102  * (applied in that order).
103  *
104  * [Jun 2000 AV] Inconsistent behaviour of open() in case if flags==O_CREAT
105  * restored for 2.4. This is the last surviving part of old 4.2BSD bug.
106  * During the 2.4 we need to fix the userland stuff depending on it -
107  * hopefully we will be able to get rid of that wart in 2.5. So far only
108  * XEmacs seems to be relying on it...
109  */
110 /*
111  * [Sep 2001 AV] Single-semaphore locking scheme (kudos to David Holland)
112  * implemented.  Let's see if raised priority of ->s_vfs_rename_mutex gives
113  * any extra contention...
114  */
115
116 /* In order to reduce some races, while at the same time doing additional
117  * checking and hopefully speeding things up, we copy filenames to the
118  * kernel data space before using them..
119  *
120  * POSIX.1 2.4: an empty pathname is invalid (ENOENT).
121  * PATH_MAX includes the nul terminator --RR.
122  */
123
124 #define EMBEDDED_NAME_MAX       (PATH_MAX - offsetof(struct filename, iname))
125
126 struct filename *
127 getname_flags(const char __user *filename, int flags, int *empty)
128 {
129         struct filename *result;
130         char *kname;
131         int len;
132
133         result = audit_reusename(filename);
134         if (result)
135                 return result;
136
137         result = __getname();
138         if (unlikely(!result))
139                 return ERR_PTR(-ENOMEM);
140
141         /*
142          * First, try to embed the struct filename inside the names_cache
143          * allocation
144          */
145         kname = (char *)result->iname;
146         result->name = kname;
147
148         len = strncpy_from_user(kname, filename, EMBEDDED_NAME_MAX);
149         if (unlikely(len < 0)) {
150                 __putname(result);
151                 return ERR_PTR(len);
152         }
153
154         /*
155          * Uh-oh. We have a name that's approaching PATH_MAX. Allocate a
156          * separate struct filename so we can dedicate the entire
157          * names_cache allocation for the pathname, and re-do the copy from
158          * userland.
159          */
160         if (unlikely(len == EMBEDDED_NAME_MAX)) {
161                 const size_t size = offsetof(struct filename, iname[1]);
162                 kname = (char *)result;
163
164                 /*
165                  * size is chosen that way we to guarantee that
166                  * result->iname[0] is within the same object and that
167                  * kname can't be equal to result->iname, no matter what.
168                  */
169                 result = kzalloc(size, GFP_KERNEL);
170                 if (unlikely(!result)) {
171                         __putname(kname);
172                         return ERR_PTR(-ENOMEM);
173                 }
174                 result->name = kname;
175                 len = strncpy_from_user(kname, filename, PATH_MAX);
176                 if (unlikely(len < 0)) {
177                         __putname(kname);
178                         kfree(result);
179                         return ERR_PTR(len);
180                 }
181                 if (unlikely(len == PATH_MAX)) {
182                         __putname(kname);
183                         kfree(result);
184                         return ERR_PTR(-ENAMETOOLONG);
185                 }
186         }
187
188         result->refcnt = 1;
189         /* The empty path is special. */
190         if (unlikely(!len)) {
191                 if (empty)
192                         *empty = 1;
193                 if (!(flags & LOOKUP_EMPTY)) {
194                         putname(result);
195                         return ERR_PTR(-ENOENT);
196                 }
197         }
198
199         result->uptr = filename;
200         result->aname = NULL;
201         audit_getname(result);
202         return result;
203 }
204
205 struct filename *
206 getname(const char __user * filename)
207 {
208         return getname_flags(filename, 0, NULL);
209 }
210
211 struct filename *
212 getname_kernel(const char * filename)
213 {
214         struct filename *result;
215         int len = strlen(filename) + 1;
216
217         result = __getname();
218         if (unlikely(!result))
219                 return ERR_PTR(-ENOMEM);
220
221         if (len <= EMBEDDED_NAME_MAX) {
222                 result->name = (char *)result->iname;
223         } else if (len <= PATH_MAX) {
224                 struct filename *tmp;
225
226                 tmp = kmalloc(sizeof(*tmp), GFP_KERNEL);
227                 if (unlikely(!tmp)) {
228                         __putname(result);
229                         return ERR_PTR(-ENOMEM);
230                 }
231                 tmp->name = (char *)result;
232                 result = tmp;
233         } else {
234                 __putname(result);
235                 return ERR_PTR(-ENAMETOOLONG);
236         }
237         memcpy((char *)result->name, filename, len);
238         result->uptr = NULL;
239         result->aname = NULL;
240         result->refcnt = 1;
241         audit_getname(result);
242
243         return result;
244 }
245
246 void putname(struct filename *name)
247 {
248         BUG_ON(name->refcnt <= 0);
249
250         if (--name->refcnt > 0)
251                 return;
252
253         if (name->name != name->iname) {
254                 __putname(name->name);
255                 kfree(name);
256         } else
257                 __putname(name);
258 }
259
260 static int check_acl(struct inode *inode, int mask)
261 {
262 #ifdef CONFIG_FS_POSIX_ACL
263         struct posix_acl *acl;
264
265         if (mask & MAY_NOT_BLOCK) {
266                 acl = get_cached_acl_rcu(inode, ACL_TYPE_ACCESS);
267                 if (!acl)
268                         return -EAGAIN;
269                 /* no ->get_acl() calls in RCU mode... */
270                 if (is_uncached_acl(acl))
271                         return -ECHILD;
272                 return posix_acl_permission(inode, acl, mask & ~MAY_NOT_BLOCK);
273         }
274
275         acl = get_acl(inode, ACL_TYPE_ACCESS);
276         if (IS_ERR(acl))
277                 return PTR_ERR(acl);
278         if (acl) {
279                 int error = posix_acl_permission(inode, acl, mask);
280                 posix_acl_release(acl);
281                 return error;
282         }
283 #endif
284
285         return -EAGAIN;
286 }
287
288 /*
289  * This does the basic permission checking
290  */
291 static int acl_permission_check(struct inode *inode, int mask)
292 {
293         unsigned int mode = inode->i_mode;
294
295         if (likely(uid_eq(current_fsuid(), inode->i_uid)))
296                 mode >>= 6;
297         else {
298                 if (IS_POSIXACL(inode) && (mode & S_IRWXG)) {
299                         int error = check_acl(inode, mask);
300                         if (error != -EAGAIN)
301                                 return error;
302                 }
303
304                 if (in_group_p(inode->i_gid))
305                         mode >>= 3;
306         }
307
308         /*
309          * If the DACs are ok we don't need any capability check.
310          */
311         if ((mask & ~mode & (MAY_READ | MAY_WRITE | MAY_EXEC)) == 0)
312                 return 0;
313         return -EACCES;
314 }
315
316 /**
317  * generic_permission -  check for access rights on a Posix-like filesystem
318  * @inode:      inode to check access rights for
319  * @mask:       right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC, ...)
320  *
321  * Used to check for read/write/execute permissions on a file.
322  * We use "fsuid" for this, letting us set arbitrary permissions
323  * for filesystem access without changing the "normal" uids which
324  * are used for other things.
325  *
326  * generic_permission is rcu-walk aware. It returns -ECHILD in case an rcu-walk
327  * request cannot be satisfied (eg. requires blocking or too much complexity).
328  * It would then be called again in ref-walk mode.
329  */
330 int generic_permission(struct inode *inode, int mask)
331 {
332         int ret;
333
334         /*
335          * Do the basic permission checks.
336          */
337         ret = acl_permission_check(inode, mask);
338         if (ret != -EACCES)
339                 return ret;
340
341         if (S_ISDIR(inode->i_mode)) {
342                 /* DACs are overridable for directories */
343                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
344                         return 0;
345                 if (!(mask & MAY_WRITE))
346                         if (capable_wrt_inode_uidgid(inode,
347                                                      CAP_DAC_READ_SEARCH))
348                                 return 0;
349                 return -EACCES;
350         }
351         /*
352          * Read/write DACs are always overridable.
353          * Executable DACs are overridable when there is
354          * at least one exec bit set.
355          */
356         if (!(mask & MAY_EXEC) || (inode->i_mode & S_IXUGO))
357                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
358                         return 0;
359
360         /*
361          * Searching includes executable on directories, else just read.
362          */
363         mask &= MAY_READ | MAY_WRITE | MAY_EXEC;
364         if (mask == MAY_READ)
365                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_READ_SEARCH))
366                         return 0;
367
368         return -EACCES;
369 }
370 EXPORT_SYMBOL(generic_permission);
371
372 /*
373  * We _really_ want to just do "generic_permission()" without
374  * even looking at the inode->i_op values. So we keep a cache
375  * flag in inode->i_opflags, that says "this has not special
376  * permission function, use the fast case".
377  */
378 static inline int do_inode_permission(struct inode *inode, int mask)
379 {
380         if (unlikely(!(inode->i_opflags & IOP_FASTPERM))) {
381                 if (likely(inode->i_op->permission))
382                         return inode->i_op->permission(inode, mask);
383
384                 /* This gets set once for the inode lifetime */
385                 spin_lock(&inode->i_lock);
386                 inode->i_opflags |= IOP_FASTPERM;
387                 spin_unlock(&inode->i_lock);
388         }
389         return generic_permission(inode, mask);
390 }
391
392 /**
393  * __inode_permission - Check for access rights to a given inode
394  * @inode: Inode to check permission on
395  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
396  *
397  * Check for read/write/execute permissions on an inode.
398  *
399  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
400  *
401  * This does not check for a read-only file system.  You probably want
402  * inode_permission().
403  */
404 int __inode_permission(struct inode *inode, int mask)
405 {
406         int retval;
407
408         if (unlikely(mask & MAY_WRITE)) {
409                 /*
410                  * Nobody gets write access to an immutable file.
411                  */
412                 if (IS_IMMUTABLE(inode))
413                         return -EPERM;
414
415                 /*
416                  * Updating mtime will likely cause i_uid and i_gid to be
417                  * written back improperly if their true value is unknown
418                  * to the vfs.
419                  */
420                 if (HAS_UNMAPPED_ID(inode))
421                         return -EACCES;
422         }
423
424         retval = do_inode_permission(inode, mask);
425         if (retval)
426                 return retval;
427
428         retval = devcgroup_inode_permission(inode, mask);
429         if (retval)
430                 return retval;
431
432         return security_inode_permission(inode, mask);
433 }
434 EXPORT_SYMBOL(__inode_permission);
435
436 /**
437  * sb_permission - Check superblock-level permissions
438  * @sb: Superblock of inode to check permission on
439  * @inode: Inode to check permission on
440  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
441  *
442  * Separate out file-system wide checks from inode-specific permission checks.
443  */
444 static int sb_permission(struct super_block *sb, struct inode *inode, int mask)
445 {
446         if (unlikely(mask & MAY_WRITE)) {
447                 umode_t mode = inode->i_mode;
448
449                 /* Nobody gets write access to a read-only fs. */
450                 if ((sb->s_flags & MS_RDONLY) &&
451                     (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
452                         return -EROFS;
453         }
454         return 0;
455 }
456
457 /**
458  * inode_permission - Check for access rights to a given inode
459  * @inode: Inode to check permission on
460  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
461  *
462  * Check for read/write/execute permissions on an inode.  We use fs[ug]id for
463  * this, letting us set arbitrary permissions for filesystem access without
464  * changing the "normal" UIDs which are used for other things.
465  *
466  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
467  */
468 int inode_permission(struct inode *inode, int mask)
469 {
470         int retval;
471
472         retval = sb_permission(inode->i_sb, inode, mask);
473         if (retval)
474                 return retval;
475         return __inode_permission(inode, mask);
476 }
477 EXPORT_SYMBOL(inode_permission);
478
479 /**
480  * path_get - get a reference to a path
481  * @path: path to get the reference to
482  *
483  * Given a path increment the reference count to the dentry and the vfsmount.
484  */
485 void path_get(const struct path *path)
486 {
487         mntget(path->mnt);
488         dget(path->dentry);
489 }
490 EXPORT_SYMBOL(path_get);
491
492 /**
493  * path_put - put a reference to a path
494  * @path: path to put the reference to
495  *
496  * Given a path decrement the reference count to the dentry and the vfsmount.
497  */
498 void path_put(const struct path *path)
499 {
500         dput(path->dentry);
501         mntput(path->mnt);
502 }
503 EXPORT_SYMBOL(path_put);
504
505 #define EMBEDDED_LEVELS 2
506 struct nameidata {
507         struct path     path;
508         struct qstr     last;
509         struct path     root;
510         struct inode    *inode; /* path.dentry.d_inode */
511         unsigned int    flags;
512         unsigned        seq, m_seq;
513         int             last_type;
514         unsigned        depth;
515         int             total_link_count;
516         struct saved {
517                 struct path link;
518                 struct delayed_call done;
519                 const char *name;
520                 unsigned seq;
521         } *stack, internal[EMBEDDED_LEVELS];
522         struct filename *name;
523         struct nameidata *saved;
524         struct inode    *link_inode;
525         unsigned        root_seq;
526         int             dfd;
527 };
528
529 static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
530 {
531         struct nameidata *old = current->nameidata;
532         p->stack = p->internal;
533         p->dfd = dfd;
534         p->name = name;
535         p->total_link_count = old ? old->total_link_count : 0;
536         p->saved = old;
537         current->nameidata = p;
538 }
539
540 static void restore_nameidata(void)
541 {
542         struct nameidata *now = current->nameidata, *old = now->saved;
543
544         current->nameidata = old;
545         if (old)
546                 old->total_link_count = now->total_link_count;
547         if (now->stack != now->internal)
548                 kfree(now->stack);
549 }
550
551 static int __nd_alloc_stack(struct nameidata *nd)
552 {
553         struct saved *p;
554
555         if (nd->flags & LOOKUP_RCU) {
556                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
557                                   GFP_ATOMIC);
558                 if (unlikely(!p))
559                         return -ECHILD;
560         } else {
561                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
562                                   GFP_KERNEL);
563                 if (unlikely(!p))
564                         return -ENOMEM;
565         }
566         memcpy(p, nd->internal, sizeof(nd->internal));
567         nd->stack = p;
568         return 0;
569 }
570
571 /**
572  * path_connected - Verify that a path->dentry is below path->mnt.mnt_root
573  * @path: nameidate to verify
574  *
575  * Rename can sometimes move a file or directory outside of a bind
576  * mount, path_connected allows those cases to be detected.
577  */
578 static bool path_connected(const struct path *path)
579 {
580         struct vfsmount *mnt = path->mnt;
581
582         /* Only bind mounts can have disconnected paths */
583         if (mnt->mnt_root == mnt->mnt_sb->s_root)
584                 return true;
585
586         return is_subdir(path->dentry, mnt->mnt_root);
587 }
588
589 static inline int nd_alloc_stack(struct nameidata *nd)
590 {
591         if (likely(nd->depth != EMBEDDED_LEVELS))
592                 return 0;
593         if (likely(nd->stack != nd->internal))
594                 return 0;
595         return __nd_alloc_stack(nd);
596 }
597
598 static void drop_links(struct nameidata *nd)
599 {
600         int i = nd->depth;
601         while (i--) {
602                 struct saved *last = nd->stack + i;
603                 do_delayed_call(&last->done);
604                 clear_delayed_call(&last->done);
605         }
606 }
607
608 static void terminate_walk(struct nameidata *nd)
609 {
610         drop_links(nd);
611         if (!(nd->flags & LOOKUP_RCU)) {
612                 int i;
613                 path_put(&nd->path);
614                 for (i = 0; i < nd->depth; i++)
615                         path_put(&nd->stack[i].link);
616                 if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
617                         path_put(&nd->root);
618                         nd->root.mnt = NULL;
619                 }
620         } else {
621                 nd->flags &= ~LOOKUP_RCU;
622                 if (!(nd->flags & LOOKUP_ROOT))
623                         nd->root.mnt = NULL;
624                 rcu_read_unlock();
625         }
626         nd->depth = 0;
627 }
628
629 /* path_put is needed afterwards regardless of success or failure */
630 static bool legitimize_path(struct nameidata *nd,
631                             struct path *path, unsigned seq)
632 {
633         int res = __legitimize_mnt(path->mnt, nd->m_seq);
634         if (unlikely(res)) {
635                 if (res > 0)
636                         path->mnt = NULL;
637                 path->dentry = NULL;
638                 return false;
639         }
640         if (unlikely(!lockref_get_not_dead(&path->dentry->d_lockref))) {
641                 path->dentry = NULL;
642                 return false;
643         }
644         return !read_seqcount_retry(&path->dentry->d_seq, seq);
645 }
646
647 static bool legitimize_links(struct nameidata *nd)
648 {
649         int i;
650         for (i = 0; i < nd->depth; i++) {
651                 struct saved *last = nd->stack + i;
652                 if (unlikely(!legitimize_path(nd, &last->link, last->seq))) {
653                         drop_links(nd);
654                         nd->depth = i + 1;
655                         return false;
656                 }
657         }
658         return true;
659 }
660
661 /*
662  * Path walking has 2 modes, rcu-walk and ref-walk (see
663  * Documentation/filesystems/path-lookup.txt).  In situations when we can't
664  * continue in RCU mode, we attempt to drop out of rcu-walk mode and grab
665  * normal reference counts on dentries and vfsmounts to transition to ref-walk
666  * mode.  Refcounts are grabbed at the last known good point before rcu-walk
667  * got stuck, so ref-walk may continue from there. If this is not successful
668  * (eg. a seqcount has changed), then failure is returned and it's up to caller
669  * to restart the path walk from the beginning in ref-walk mode.
670  */
671
672 /**
673  * unlazy_walk - try to switch to ref-walk mode.
674  * @nd: nameidata pathwalk data
675  * @dentry: child of nd->path.dentry or NULL
676  * @seq: seq number to check dentry against
677  * Returns: 0 on success, -ECHILD on failure
678  *
679  * unlazy_walk attempts to legitimize the current nd->path, nd->root and dentry
680  * for ref-walk mode.  @dentry must be a path found by a do_lookup call on
681  * @nd or NULL.  Must be called from rcu-walk context.
682  * Nothing should touch nameidata between unlazy_walk() failure and
683  * terminate_walk().
684  */
685 static int unlazy_walk(struct nameidata *nd, struct dentry *dentry, unsigned seq)
686 {
687         struct dentry *parent = nd->path.dentry;
688
689         BUG_ON(!(nd->flags & LOOKUP_RCU));
690
691         nd->flags &= ~LOOKUP_RCU;
692         if (unlikely(!legitimize_links(nd)))
693                 goto out2;
694         if (unlikely(!legitimize_mnt(nd->path.mnt, nd->m_seq)))
695                 goto out2;
696         if (unlikely(!lockref_get_not_dead(&parent->d_lockref)))
697                 goto out1;
698
699         /*
700          * For a negative lookup, the lookup sequence point is the parents
701          * sequence point, and it only needs to revalidate the parent dentry.
702          *
703          * For a positive lookup, we need to move both the parent and the
704          * dentry from the RCU domain to be properly refcounted. And the
705          * sequence number in the dentry validates *both* dentry counters,
706          * since we checked the sequence number of the parent after we got
707          * the child sequence number. So we know the parent must still
708          * be valid if the child sequence number is still valid.
709          */
710         if (!dentry) {
711                 if (read_seqcount_retry(&parent->d_seq, nd->seq))
712                         goto out;
713                 BUG_ON(nd->inode != parent->d_inode);
714         } else {
715                 if (!lockref_get_not_dead(&dentry->d_lockref))
716                         goto out;
717                 if (read_seqcount_retry(&dentry->d_seq, seq))
718                         goto drop_dentry;
719         }
720
721         /*
722          * Sequence counts matched. Now make sure that the root is
723          * still valid and get it if required.
724          */
725         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
726                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq))) {
727                         rcu_read_unlock();
728                         dput(dentry);
729                         return -ECHILD;
730                 }
731         }
732
733         rcu_read_unlock();
734         return 0;
735
736 drop_dentry:
737         rcu_read_unlock();
738         dput(dentry);
739         goto drop_root_mnt;
740 out2:
741         nd->path.mnt = NULL;
742 out1:
743         nd->path.dentry = NULL;
744 out:
745         rcu_read_unlock();
746 drop_root_mnt:
747         if (!(nd->flags & LOOKUP_ROOT))
748                 nd->root.mnt = NULL;
749         return -ECHILD;
750 }
751
752 static int unlazy_link(struct nameidata *nd, struct path *link, unsigned seq)
753 {
754         if (unlikely(!legitimize_path(nd, link, seq))) {
755                 drop_links(nd);
756                 nd->depth = 0;
757                 nd->flags &= ~LOOKUP_RCU;
758                 nd->path.mnt = NULL;
759                 nd->path.dentry = NULL;
760                 if (!(nd->flags & LOOKUP_ROOT))
761                         nd->root.mnt = NULL;
762                 rcu_read_unlock();
763         } else if (likely(unlazy_walk(nd, NULL, 0)) == 0) {
764                 return 0;
765         }
766         path_put(link);
767         return -ECHILD;
768 }
769
770 static inline int d_revalidate(struct dentry *dentry, unsigned int flags)
771 {
772         return dentry->d_op->d_revalidate(dentry, flags);
773 }
774
775 /**
776  * complete_walk - successful completion of path walk
777  * @nd:  pointer nameidata
778  *
779  * If we had been in RCU mode, drop out of it and legitimize nd->path.
780  * Revalidate the final result, unless we'd already done that during
781  * the path walk or the filesystem doesn't ask for it.  Return 0 on
782  * success, -error on failure.  In case of failure caller does not
783  * need to drop nd->path.
784  */
785 static int complete_walk(struct nameidata *nd)
786 {
787         struct dentry *dentry = nd->path.dentry;
788         int status;
789
790         if (nd->flags & LOOKUP_RCU) {
791                 if (!(nd->flags & LOOKUP_ROOT))
792                         nd->root.mnt = NULL;
793                 if (unlikely(unlazy_walk(nd, NULL, 0)))
794                         return -ECHILD;
795         }
796
797         if (likely(!(nd->flags & LOOKUP_JUMPED)))
798                 return 0;
799
800         if (likely(!(dentry->d_flags & DCACHE_OP_WEAK_REVALIDATE)))
801                 return 0;
802
803         status = dentry->d_op->d_weak_revalidate(dentry, nd->flags);
804         if (status > 0)
805                 return 0;
806
807         if (!status)
808                 status = -ESTALE;
809
810         return status;
811 }
812
813 static void set_root(struct nameidata *nd)
814 {
815         struct fs_struct *fs = current->fs;
816
817         if (nd->flags & LOOKUP_RCU) {
818                 unsigned seq;
819
820                 do {
821                         seq = read_seqcount_begin(&fs->seq);
822                         nd->root = fs->root;
823                         nd->root_seq = __read_seqcount_begin(&nd->root.dentry->d_seq);
824                 } while (read_seqcount_retry(&fs->seq, seq));
825         } else {
826                 get_fs_root(fs, &nd->root);
827         }
828 }
829
830 static void path_put_conditional(struct path *path, struct nameidata *nd)
831 {
832         dput(path->dentry);
833         if (path->mnt != nd->path.mnt)
834                 mntput(path->mnt);
835 }
836
837 static inline void path_to_nameidata(const struct path *path,
838                                         struct nameidata *nd)
839 {
840         if (!(nd->flags & LOOKUP_RCU)) {
841                 dput(nd->path.dentry);
842                 if (nd->path.mnt != path->mnt)
843                         mntput(nd->path.mnt);
844         }
845         nd->path.mnt = path->mnt;
846         nd->path.dentry = path->dentry;
847 }
848
849 static int nd_jump_root(struct nameidata *nd)
850 {
851         if (nd->flags & LOOKUP_RCU) {
852                 struct dentry *d;
853                 nd->path = nd->root;
854                 d = nd->path.dentry;
855                 nd->inode = d->d_inode;
856                 nd->seq = nd->root_seq;
857                 if (unlikely(read_seqcount_retry(&d->d_seq, nd->seq)))
858                         return -ECHILD;
859         } else {
860                 path_put(&nd->path);
861                 nd->path = nd->root;
862                 path_get(&nd->path);
863                 nd->inode = nd->path.dentry->d_inode;
864         }
865         nd->flags |= LOOKUP_JUMPED;
866         return 0;
867 }
868
869 /*
870  * Helper to directly jump to a known parsed path from ->get_link,
871  * caller must have taken a reference to path beforehand.
872  */
873 void nd_jump_link(struct path *path)
874 {
875         struct nameidata *nd = current->nameidata;
876         path_put(&nd->path);
877
878         nd->path = *path;
879         nd->inode = nd->path.dentry->d_inode;
880         nd->flags |= LOOKUP_JUMPED;
881 }
882
883 static inline void put_link(struct nameidata *nd)
884 {
885         struct saved *last = nd->stack + --nd->depth;
886         do_delayed_call(&last->done);
887         if (!(nd->flags & LOOKUP_RCU))
888                 path_put(&last->link);
889 }
890
891 int sysctl_protected_symlinks __read_mostly = 0;
892 int sysctl_protected_hardlinks __read_mostly = 0;
893
894 /**
895  * may_follow_link - Check symlink following for unsafe situations
896  * @nd: nameidata pathwalk data
897  *
898  * In the case of the sysctl_protected_symlinks sysctl being enabled,
899  * CAP_DAC_OVERRIDE needs to be specifically ignored if the symlink is
900  * in a sticky world-writable directory. This is to protect privileged
901  * processes from failing races against path names that may change out
902  * from under them by way of other users creating malicious symlinks.
903  * It will permit symlinks to be followed only when outside a sticky
904  * world-writable directory, or when the uid of the symlink and follower
905  * match, or when the directory owner matches the symlink's owner.
906  *
907  * Returns 0 if following the symlink is allowed, -ve on error.
908  */
909 static inline int may_follow_link(struct nameidata *nd)
910 {
911         const struct inode *inode;
912         const struct inode *parent;
913         kuid_t puid;
914
915         if (!sysctl_protected_symlinks)
916                 return 0;
917
918         /* Allowed if owner and follower match. */
919         inode = nd->link_inode;
920         if (uid_eq(current_cred()->fsuid, inode->i_uid))
921                 return 0;
922
923         /* Allowed if parent directory not sticky and world-writable. */
924         parent = nd->inode;
925         if ((parent->i_mode & (S_ISVTX|S_IWOTH)) != (S_ISVTX|S_IWOTH))
926                 return 0;
927
928         /* Allowed if parent directory and link owner match. */
929         puid = parent->i_uid;
930         if (uid_valid(puid) && uid_eq(puid, inode->i_uid))
931                 return 0;
932
933         if (nd->flags & LOOKUP_RCU)
934                 return -ECHILD;
935
936         audit_log_link_denied("follow_link", &nd->stack[0].link);
937         return -EACCES;
938 }
939
940 /**
941  * safe_hardlink_source - Check for safe hardlink conditions
942  * @inode: the source inode to hardlink from
943  *
944  * Return false if at least one of the following conditions:
945  *    - inode is not a regular file
946  *    - inode is setuid
947  *    - inode is setgid and group-exec
948  *    - access failure for read and write
949  *
950  * Otherwise returns true.
951  */
952 static bool safe_hardlink_source(struct inode *inode)
953 {
954         umode_t mode = inode->i_mode;
955
956         /* Special files should not get pinned to the filesystem. */
957         if (!S_ISREG(mode))
958                 return false;
959
960         /* Setuid files should not get pinned to the filesystem. */
961         if (mode & S_ISUID)
962                 return false;
963
964         /* Executable setgid files should not get pinned to the filesystem. */
965         if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP))
966                 return false;
967
968         /* Hardlinking to unreadable or unwritable sources is dangerous. */
969         if (inode_permission(inode, MAY_READ | MAY_WRITE))
970                 return false;
971
972         return true;
973 }
974
975 /**
976  * may_linkat - Check permissions for creating a hardlink
977  * @link: the source to hardlink from
978  *
979  * Block hardlink when all of:
980  *  - sysctl_protected_hardlinks enabled
981  *  - fsuid does not match inode
982  *  - hardlink source is unsafe (see safe_hardlink_source() above)
983  *  - not CAP_FOWNER in a namespace with the inode owner uid mapped
984  *
985  * Returns 0 if successful, -ve on error.
986  */
987 static int may_linkat(struct path *link)
988 {
989         struct inode *inode;
990
991         if (!sysctl_protected_hardlinks)
992                 return 0;
993
994         inode = link->dentry->d_inode;
995
996         /* Source inode owner (or CAP_FOWNER) can hardlink all they like,
997          * otherwise, it must be a safe source.
998          */
999         if (inode_owner_or_capable(inode) || safe_hardlink_source(inode))
1000                 return 0;
1001
1002         audit_log_link_denied("linkat", link);
1003         return -EPERM;
1004 }
1005
1006 static __always_inline
1007 const char *get_link(struct nameidata *nd)
1008 {
1009         struct saved *last = nd->stack + nd->depth - 1;
1010         struct dentry *dentry = last->link.dentry;
1011         struct inode *inode = nd->link_inode;
1012         int error;
1013         const char *res;
1014
1015         if (!(nd->flags & LOOKUP_RCU)) {
1016                 touch_atime(&last->link);
1017                 cond_resched();
1018         } else if (atime_needs_update_rcu(&last->link, inode)) {
1019                 if (unlikely(unlazy_walk(nd, NULL, 0)))
1020                         return ERR_PTR(-ECHILD);
1021                 touch_atime(&last->link);
1022         }
1023
1024         error = security_inode_follow_link(dentry, inode,
1025                                            nd->flags & LOOKUP_RCU);
1026         if (unlikely(error))
1027                 return ERR_PTR(error);
1028
1029         nd->last_type = LAST_BIND;
1030         res = inode->i_link;
1031         if (!res) {
1032                 const char * (*get)(struct dentry *, struct inode *,
1033                                 struct delayed_call *);
1034                 get = inode->i_op->get_link;
1035                 if (nd->flags & LOOKUP_RCU) {
1036                         res = get(NULL, inode, &last->done);
1037                         if (res == ERR_PTR(-ECHILD)) {
1038                                 if (unlikely(unlazy_walk(nd, NULL, 0)))
1039                                         return ERR_PTR(-ECHILD);
1040                                 res = get(dentry, inode, &last->done);
1041                         }
1042                 } else {
1043                         res = get(dentry, inode, &last->done);
1044                 }
1045                 if (IS_ERR_OR_NULL(res))
1046                         return res;
1047         }
1048         if (*res == '/') {
1049                 if (!nd->root.mnt)
1050                         set_root(nd);
1051                 if (unlikely(nd_jump_root(nd)))
1052                         return ERR_PTR(-ECHILD);
1053                 while (unlikely(*++res == '/'))
1054                         ;
1055         }
1056         if (!*res)
1057                 res = NULL;
1058         return res;
1059 }
1060
1061 /*
1062  * follow_up - Find the mountpoint of path's vfsmount
1063  *
1064  * Given a path, find the mountpoint of its source file system.
1065  * Replace @path with the path of the mountpoint in the parent mount.
1066  * Up is towards /.
1067  *
1068  * Return 1 if we went up a level and 0 if we were already at the
1069  * root.
1070  */
1071 int follow_up(struct path *path)
1072 {
1073         struct mount *mnt = real_mount(path->mnt);
1074         struct mount *parent;
1075         struct dentry *mountpoint;
1076
1077         read_seqlock_excl(&mount_lock);
1078         parent = mnt->mnt_parent;
1079         if (parent == mnt) {
1080                 read_sequnlock_excl(&mount_lock);
1081                 return 0;
1082         }
1083         mntget(&parent->mnt);
1084         mountpoint = dget(mnt->mnt_mountpoint);
1085         read_sequnlock_excl(&mount_lock);
1086         dput(path->dentry);
1087         path->dentry = mountpoint;
1088         mntput(path->mnt);
1089         path->mnt = &parent->mnt;
1090         return 1;
1091 }
1092 EXPORT_SYMBOL(follow_up);
1093
1094 /*
1095  * Perform an automount
1096  * - return -EISDIR to tell follow_managed() to stop and return the path we
1097  *   were called with.
1098  */
1099 static int follow_automount(struct path *path, struct nameidata *nd,
1100                             bool *need_mntput)
1101 {
1102         struct vfsmount *mnt;
1103         const struct cred *old_cred;
1104         int err;
1105
1106         if (!path->dentry->d_op || !path->dentry->d_op->d_automount)
1107                 return -EREMOTE;
1108
1109         /* We don't want to mount if someone's just doing a stat -
1110          * unless they're stat'ing a directory and appended a '/' to
1111          * the name.
1112          *
1113          * We do, however, want to mount if someone wants to open or
1114          * create a file of any type under the mountpoint, wants to
1115          * traverse through the mountpoint or wants to open the
1116          * mounted directory.  Also, autofs may mark negative dentries
1117          * as being automount points.  These will need the attentions
1118          * of the daemon to instantiate them before they can be used.
1119          */
1120         if (!(nd->flags & (LOOKUP_PARENT | LOOKUP_DIRECTORY |
1121                            LOOKUP_OPEN | LOOKUP_CREATE | LOOKUP_AUTOMOUNT)) &&
1122             path->dentry->d_inode)
1123                 return -EISDIR;
1124
1125         if (path->dentry->d_sb->s_user_ns != &init_user_ns)
1126                 return -EACCES;
1127
1128         nd->total_link_count++;
1129         if (nd->total_link_count >= 40)
1130                 return -ELOOP;
1131
1132         old_cred = override_creds(&init_cred);
1133         mnt = path->dentry->d_op->d_automount(path);
1134         revert_creds(old_cred);
1135         if (IS_ERR(mnt)) {
1136                 /*
1137                  * The filesystem is allowed to return -EISDIR here to indicate
1138                  * it doesn't want to automount.  For instance, autofs would do
1139                  * this so that its userspace daemon can mount on this dentry.
1140                  *
1141                  * However, we can only permit this if it's a terminal point in
1142                  * the path being looked up; if it wasn't then the remainder of
1143                  * the path is inaccessible and we should say so.
1144                  */
1145                 if (PTR_ERR(mnt) == -EISDIR && (nd->flags & LOOKUP_PARENT))
1146                         return -EREMOTE;
1147                 return PTR_ERR(mnt);
1148         }
1149
1150         if (!mnt) /* mount collision */
1151                 return 0;
1152
1153         if (!*need_mntput) {
1154                 /* lock_mount() may release path->mnt on error */
1155                 mntget(path->mnt);
1156                 *need_mntput = true;
1157         }
1158         err = finish_automount(mnt, path);
1159
1160         switch (err) {
1161         case -EBUSY:
1162                 /* Someone else made a mount here whilst we were busy */
1163                 return 0;
1164         case 0:
1165                 path_put(path);
1166                 path->mnt = mnt;
1167                 path->dentry = dget(mnt->mnt_root);
1168                 return 0;
1169         default:
1170                 return err;
1171         }
1172
1173 }
1174
1175 /*
1176  * Handle a dentry that is managed in some way.
1177  * - Flagged for transit management (autofs)
1178  * - Flagged as mountpoint
1179  * - Flagged as automount point
1180  *
1181  * This may only be called in refwalk mode.
1182  *
1183  * Serialization is taken care of in namespace.c
1184  */
1185 static int follow_managed(struct path *path, struct nameidata *nd)
1186 {
1187         struct vfsmount *mnt = path->mnt; /* held by caller, must be left alone */
1188         unsigned managed;
1189         bool need_mntput = false;
1190         int ret = 0;
1191
1192         /* Given that we're not holding a lock here, we retain the value in a
1193          * local variable for each dentry as we look at it so that we don't see
1194          * the components of that value change under us */
1195         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1196                managed &= DCACHE_MANAGED_DENTRY,
1197                unlikely(managed != 0)) {
1198                 /* Allow the filesystem to manage the transit without i_mutex
1199                  * being held. */
1200                 if (managed & DCACHE_MANAGE_TRANSIT) {
1201                         BUG_ON(!path->dentry->d_op);
1202                         BUG_ON(!path->dentry->d_op->d_manage);
1203                         ret = path->dentry->d_op->d_manage(path, false);
1204                         if (ret < 0)
1205                                 break;
1206                 }
1207
1208                 /* Transit to a mounted filesystem. */
1209                 if (managed & DCACHE_MOUNTED) {
1210                         struct vfsmount *mounted = lookup_mnt(path);
1211                         if (mounted) {
1212                                 dput(path->dentry);
1213                                 if (need_mntput)
1214                                         mntput(path->mnt);
1215                                 path->mnt = mounted;
1216                                 path->dentry = dget(mounted->mnt_root);
1217                                 need_mntput = true;
1218                                 continue;
1219                         }
1220
1221                         /* Something is mounted on this dentry in another
1222                          * namespace and/or whatever was mounted there in this
1223                          * namespace got unmounted before lookup_mnt() could
1224                          * get it */
1225                 }
1226
1227                 /* Handle an automount point */
1228                 if (managed & DCACHE_NEED_AUTOMOUNT) {
1229                         ret = follow_automount(path, nd, &need_mntput);
1230                         if (ret < 0)
1231                                 break;
1232                         continue;
1233                 }
1234
1235                 /* We didn't change the current path point */
1236                 break;
1237         }
1238
1239         if (need_mntput && path->mnt == mnt)
1240                 mntput(path->mnt);
1241         if (ret == -EISDIR || !ret)
1242                 ret = 1;
1243         if (need_mntput)
1244                 nd->flags |= LOOKUP_JUMPED;
1245         if (unlikely(ret < 0))
1246                 path_put_conditional(path, nd);
1247         return ret;
1248 }
1249
1250 int follow_down_one(struct path *path)
1251 {
1252         struct vfsmount *mounted;
1253
1254         mounted = lookup_mnt(path);
1255         if (mounted) {
1256                 dput(path->dentry);
1257                 mntput(path->mnt);
1258                 path->mnt = mounted;
1259                 path->dentry = dget(mounted->mnt_root);
1260                 return 1;
1261         }
1262         return 0;
1263 }
1264 EXPORT_SYMBOL(follow_down_one);
1265
1266 static inline int managed_dentry_rcu(const struct path *path)
1267 {
1268         return (path->dentry->d_flags & DCACHE_MANAGE_TRANSIT) ?
1269                 path->dentry->d_op->d_manage(path, true) : 0;
1270 }
1271
1272 /*
1273  * Try to skip to top of mountpoint pile in rcuwalk mode.  Fail if
1274  * we meet a managed dentry that would need blocking.
1275  */
1276 static bool __follow_mount_rcu(struct nameidata *nd, struct path *path,
1277                                struct inode **inode, unsigned *seqp)
1278 {
1279         for (;;) {
1280                 struct mount *mounted;
1281                 /*
1282                  * Don't forget we might have a non-mountpoint managed dentry
1283                  * that wants to block transit.
1284                  */
1285                 switch (managed_dentry_rcu(path)) {
1286                 case -ECHILD:
1287                 default:
1288                         return false;
1289                 case -EISDIR:
1290                         return true;
1291                 case 0:
1292                         break;
1293                 }
1294
1295                 if (!d_mountpoint(path->dentry))
1296                         return !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1297
1298                 mounted = __lookup_mnt(path->mnt, path->dentry);
1299                 if (!mounted)
1300                         break;
1301                 path->mnt = &mounted->mnt;
1302                 path->dentry = mounted->mnt.mnt_root;
1303                 nd->flags |= LOOKUP_JUMPED;
1304                 *seqp = read_seqcount_begin(&path->dentry->d_seq);
1305                 /*
1306                  * Update the inode too. We don't need to re-check the
1307                  * dentry sequence number here after this d_inode read,
1308                  * because a mount-point is always pinned.
1309                  */
1310                 *inode = path->dentry->d_inode;
1311         }
1312         return !read_seqretry(&mount_lock, nd->m_seq) &&
1313                 !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1314 }
1315
1316 static int follow_dotdot_rcu(struct nameidata *nd)
1317 {
1318         struct inode *inode = nd->inode;
1319
1320         while (1) {
1321                 if (path_equal(&nd->path, &nd->root))
1322                         break;
1323                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1324                         struct dentry *old = nd->path.dentry;
1325                         struct dentry *parent = old->d_parent;
1326                         unsigned seq;
1327
1328                         inode = parent->d_inode;
1329                         seq = read_seqcount_begin(&parent->d_seq);
1330                         if (unlikely(read_seqcount_retry(&old->d_seq, nd->seq)))
1331                                 return -ECHILD;
1332                         nd->path.dentry = parent;
1333                         nd->seq = seq;
1334                         if (unlikely(!path_connected(&nd->path)))
1335                                 return -ENOENT;
1336                         break;
1337                 } else {
1338                         struct mount *mnt = real_mount(nd->path.mnt);
1339                         struct mount *mparent = mnt->mnt_parent;
1340                         struct dentry *mountpoint = mnt->mnt_mountpoint;
1341                         struct inode *inode2 = mountpoint->d_inode;
1342                         unsigned seq = read_seqcount_begin(&mountpoint->d_seq);
1343                         if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1344                                 return -ECHILD;
1345                         if (&mparent->mnt == nd->path.mnt)
1346                                 break;
1347                         /* we know that mountpoint was pinned */
1348                         nd->path.dentry = mountpoint;
1349                         nd->path.mnt = &mparent->mnt;
1350                         inode = inode2;
1351                         nd->seq = seq;
1352                 }
1353         }
1354         while (unlikely(d_mountpoint(nd->path.dentry))) {
1355                 struct mount *mounted;
1356                 mounted = __lookup_mnt(nd->path.mnt, nd->path.dentry);
1357                 if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1358                         return -ECHILD;
1359                 if (!mounted)
1360                         break;
1361                 nd->path.mnt = &mounted->mnt;
1362                 nd->path.dentry = mounted->mnt.mnt_root;
1363                 inode = nd->path.dentry->d_inode;
1364                 nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
1365         }
1366         nd->inode = inode;
1367         return 0;
1368 }
1369
1370 /*
1371  * Follow down to the covering mount currently visible to userspace.  At each
1372  * point, the filesystem owning that dentry may be queried as to whether the
1373  * caller is permitted to proceed or not.
1374  */
1375 int follow_down(struct path *path)
1376 {
1377         unsigned managed;
1378         int ret;
1379
1380         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1381                unlikely(managed & DCACHE_MANAGED_DENTRY)) {
1382                 /* Allow the filesystem to manage the transit without i_mutex
1383                  * being held.
1384                  *
1385                  * We indicate to the filesystem if someone is trying to mount
1386                  * something here.  This gives autofs the chance to deny anyone
1387                  * other than its daemon the right to mount on its
1388                  * superstructure.
1389                  *
1390                  * The filesystem may sleep at this point.
1391                  */
1392                 if (managed & DCACHE_MANAGE_TRANSIT) {
1393                         BUG_ON(!path->dentry->d_op);
1394                         BUG_ON(!path->dentry->d_op->d_manage);
1395                         ret = path->dentry->d_op->d_manage(path, false);
1396                         if (ret < 0)
1397                                 return ret == -EISDIR ? 0 : ret;
1398                 }
1399
1400                 /* Transit to a mounted filesystem. */
1401                 if (managed & DCACHE_MOUNTED) {
1402                         struct vfsmount *mounted = lookup_mnt(path);
1403                         if (!mounted)
1404                                 break;
1405                         dput(path->dentry);
1406                         mntput(path->mnt);
1407                         path->mnt = mounted;
1408                         path->dentry = dget(mounted->mnt_root);
1409                         continue;
1410                 }
1411
1412                 /* Don't handle automount points here */
1413                 break;
1414         }
1415         return 0;
1416 }
1417 EXPORT_SYMBOL(follow_down);
1418
1419 /*
1420  * Skip to top of mountpoint pile in refwalk mode for follow_dotdot()
1421  */
1422 static void follow_mount(struct path *path)
1423 {
1424         while (d_mountpoint(path->dentry)) {
1425                 struct vfsmount *mounted = lookup_mnt(path);
1426                 if (!mounted)
1427                         break;
1428                 dput(path->dentry);
1429                 mntput(path->mnt);
1430                 path->mnt = mounted;
1431                 path->dentry = dget(mounted->mnt_root);
1432         }
1433 }
1434
1435 static int path_parent_directory(struct path *path)
1436 {
1437         struct dentry *old = path->dentry;
1438         /* rare case of legitimate dget_parent()... */
1439         path->dentry = dget_parent(path->dentry);
1440         dput(old);
1441         if (unlikely(!path_connected(path)))
1442                 return -ENOENT;
1443         return 0;
1444 }
1445
1446 static int follow_dotdot(struct nameidata *nd)
1447 {
1448         while(1) {
1449                 if (nd->path.dentry == nd->root.dentry &&
1450                     nd->path.mnt == nd->root.mnt) {
1451                         break;
1452                 }
1453                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1454                         int ret = path_parent_directory(&nd->path);
1455                         if (ret)
1456                                 return ret;
1457                         break;
1458                 }
1459                 if (!follow_up(&nd->path))
1460                         break;
1461         }
1462         follow_mount(&nd->path);
1463         nd->inode = nd->path.dentry->d_inode;
1464         return 0;
1465 }
1466
1467 /*
1468  * This looks up the name in dcache and possibly revalidates the found dentry.
1469  * NULL is returned if the dentry does not exist in the cache.
1470  */
1471 static struct dentry *lookup_dcache(const struct qstr *name,
1472                                     struct dentry *dir,
1473                                     unsigned int flags)
1474 {
1475         struct dentry *dentry;
1476         int error;
1477
1478         dentry = d_lookup(dir, name);
1479         if (dentry) {
1480                 if (dentry->d_flags & DCACHE_OP_REVALIDATE) {
1481                         error = d_revalidate(dentry, flags);
1482                         if (unlikely(error <= 0)) {
1483                                 if (!error)
1484                                         d_invalidate(dentry);
1485                                 dput(dentry);
1486                                 return ERR_PTR(error);
1487                         }
1488                 }
1489         }
1490         return dentry;
1491 }
1492
1493 /*
1494  * Call i_op->lookup on the dentry.  The dentry must be negative and
1495  * unhashed.
1496  *
1497  * dir->d_inode->i_mutex must be held
1498  */
1499 static struct dentry *lookup_real(struct inode *dir, struct dentry *dentry,
1500                                   unsigned int flags)
1501 {
1502         struct dentry *old;
1503
1504         /* Don't create child dentry for a dead directory. */
1505         if (unlikely(IS_DEADDIR(dir))) {
1506                 dput(dentry);
1507                 return ERR_PTR(-ENOENT);
1508         }
1509
1510         old = dir->i_op->lookup(dir, dentry, flags);
1511         if (unlikely(old)) {
1512                 dput(dentry);
1513                 dentry = old;
1514         }
1515         return dentry;
1516 }
1517
1518 static struct dentry *__lookup_hash(const struct qstr *name,
1519                 struct dentry *base, unsigned int flags)
1520 {
1521         struct dentry *dentry = lookup_dcache(name, base, flags);
1522
1523         if (dentry)
1524                 return dentry;
1525
1526         dentry = d_alloc(base, name);
1527         if (unlikely(!dentry))
1528                 return ERR_PTR(-ENOMEM);
1529
1530         return lookup_real(base->d_inode, dentry, flags);
1531 }
1532
1533 static int lookup_fast(struct nameidata *nd,
1534                        struct path *path, struct inode **inode,
1535                        unsigned *seqp)
1536 {
1537         struct vfsmount *mnt = nd->path.mnt;
1538         struct dentry *dentry, *parent = nd->path.dentry;
1539         int status = 1;
1540         int err;
1541
1542         /*
1543          * Rename seqlock is not required here because in the off chance
1544          * of a false negative due to a concurrent rename, the caller is
1545          * going to fall back to non-racy lookup.
1546          */
1547         if (nd->flags & LOOKUP_RCU) {
1548                 unsigned seq;
1549                 bool negative;
1550                 dentry = __d_lookup_rcu(parent, &nd->last, &seq);
1551                 if (unlikely(!dentry)) {
1552                         if (unlazy_walk(nd, NULL, 0))
1553                                 return -ECHILD;
1554                         return 0;
1555                 }
1556
1557                 /*
1558                  * This sequence count validates that the inode matches
1559                  * the dentry name information from lookup.
1560                  */
1561                 *inode = d_backing_inode(dentry);
1562                 negative = d_is_negative(dentry);
1563                 if (unlikely(read_seqcount_retry(&dentry->d_seq, seq)))
1564                         return -ECHILD;
1565
1566                 /*
1567                  * This sequence count validates that the parent had no
1568                  * changes while we did the lookup of the dentry above.
1569                  *
1570                  * The memory barrier in read_seqcount_begin of child is
1571                  *  enough, we can use __read_seqcount_retry here.
1572                  */
1573                 if (unlikely(__read_seqcount_retry(&parent->d_seq, nd->seq)))
1574                         return -ECHILD;
1575
1576                 *seqp = seq;
1577                 if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
1578                         status = d_revalidate(dentry, nd->flags);
1579                 if (unlikely(status <= 0)) {
1580                         if (unlazy_walk(nd, dentry, seq))
1581                                 return -ECHILD;
1582                         if (status == -ECHILD)
1583                                 status = d_revalidate(dentry, nd->flags);
1584                 } else {
1585                         /*
1586                          * Note: do negative dentry check after revalidation in
1587                          * case that drops it.
1588                          */
1589                         if (unlikely(negative))
1590                                 return -ENOENT;
1591                         path->mnt = mnt;
1592                         path->dentry = dentry;
1593                         if (likely(__follow_mount_rcu(nd, path, inode, seqp)))
1594                                 return 1;
1595                         if (unlazy_walk(nd, dentry, seq))
1596                                 return -ECHILD;
1597                 }
1598         } else {
1599                 dentry = __d_lookup(parent, &nd->last);
1600                 if (unlikely(!dentry))
1601                         return 0;
1602                 if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
1603                         status = d_revalidate(dentry, nd->flags);
1604         }
1605         if (unlikely(status <= 0)) {
1606                 if (!status)
1607                         d_invalidate(dentry);
1608                 dput(dentry);
1609                 return status;
1610         }
1611         if (unlikely(d_is_negative(dentry))) {
1612                 dput(dentry);
1613                 return -ENOENT;
1614         }
1615
1616         path->mnt = mnt;
1617         path->dentry = dentry;
1618         err = follow_managed(path, nd);
1619         if (likely(err > 0))
1620                 *inode = d_backing_inode(path->dentry);
1621         return err;
1622 }
1623
1624 /* Fast lookup failed, do it the slow way */
1625 static struct dentry *lookup_slow(const struct qstr *name,
1626                                   struct dentry *dir,
1627                                   unsigned int flags)
1628 {
1629         struct dentry *dentry = ERR_PTR(-ENOENT), *old;
1630         struct inode *inode = dir->d_inode;
1631         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
1632
1633         inode_lock_shared(inode);
1634         /* Don't go there if it's already dead */
1635         if (unlikely(IS_DEADDIR(inode)))
1636                 goto out;
1637 again:
1638         dentry = d_alloc_parallel(dir, name, &wq);
1639         if (IS_ERR(dentry))
1640                 goto out;
1641         if (unlikely(!d_in_lookup(dentry))) {
1642                 if ((dentry->d_flags & DCACHE_OP_REVALIDATE) &&
1643                     !(flags & LOOKUP_NO_REVAL)) {
1644                         int error = d_revalidate(dentry, flags);
1645                         if (unlikely(error <= 0)) {
1646                                 if (!error) {
1647                                         d_invalidate(dentry);
1648                                         dput(dentry);
1649                                         goto again;
1650                                 }
1651                                 dput(dentry);
1652                                 dentry = ERR_PTR(error);
1653                         }
1654                 }
1655         } else {
1656                 old = inode->i_op->lookup(inode, dentry, flags);
1657                 d_lookup_done(dentry);
1658                 if (unlikely(old)) {
1659                         dput(dentry);
1660                         dentry = old;
1661                 }
1662         }
1663 out:
1664         inode_unlock_shared(inode);
1665         return dentry;
1666 }
1667
1668 static inline int may_lookup(struct nameidata *nd)
1669 {
1670         if (nd->flags & LOOKUP_RCU) {
1671                 int err = inode_permission(nd->inode, MAY_EXEC|MAY_NOT_BLOCK);
1672                 if (err != -ECHILD)
1673                         return err;
1674                 if (unlazy_walk(nd, NULL, 0))
1675                         return -ECHILD;
1676         }
1677         return inode_permission(nd->inode, MAY_EXEC);
1678 }
1679
1680 static inline int handle_dots(struct nameidata *nd, int type)
1681 {
1682         if (type == LAST_DOTDOT) {
1683                 if (!nd->root.mnt)
1684                         set_root(nd);
1685                 if (nd->flags & LOOKUP_RCU) {
1686                         return follow_dotdot_rcu(nd);
1687                 } else
1688                         return follow_dotdot(nd);
1689         }
1690         return 0;
1691 }
1692
1693 static int pick_link(struct nameidata *nd, struct path *link,
1694                      struct inode *inode, unsigned seq)
1695 {
1696         int error;
1697         struct saved *last;
1698         if (unlikely(nd->total_link_count++ >= MAXSYMLINKS)) {
1699                 path_to_nameidata(link, nd);
1700                 return -ELOOP;
1701         }
1702         if (!(nd->flags & LOOKUP_RCU)) {
1703                 if (link->mnt == nd->path.mnt)
1704                         mntget(link->mnt);
1705         }
1706         error = nd_alloc_stack(nd);
1707         if (unlikely(error)) {
1708                 if (error == -ECHILD) {
1709                         if (unlikely(unlazy_link(nd, link, seq)))
1710                                 return -ECHILD;
1711                         error = nd_alloc_stack(nd);
1712                 }
1713                 if (error) {
1714                         path_put(link);
1715                         return error;
1716                 }
1717         }
1718
1719         last = nd->stack + nd->depth++;
1720         last->link = *link;
1721         clear_delayed_call(&last->done);
1722         nd->link_inode = inode;
1723         last->seq = seq;
1724         return 1;
1725 }
1726
1727 enum {WALK_FOLLOW = 1, WALK_MORE = 2};
1728
1729 /*
1730  * Do we need to follow links? We _really_ want to be able
1731  * to do this check without having to look at inode->i_op,
1732  * so we keep a cache of "no, this doesn't need follow_link"
1733  * for the common case.
1734  */
1735 static inline int step_into(struct nameidata *nd, struct path *path,
1736                             int flags, struct inode *inode, unsigned seq)
1737 {
1738         if (!(flags & WALK_MORE) && nd->depth)
1739                 put_link(nd);
1740         if (likely(!d_is_symlink(path->dentry)) ||
1741            !(flags & WALK_FOLLOW || nd->flags & LOOKUP_FOLLOW)) {
1742                 /* not a symlink or should not follow */
1743                 path_to_nameidata(path, nd);
1744                 nd->inode = inode;
1745                 nd->seq = seq;
1746                 return 0;
1747         }
1748         /* make sure that d_is_symlink above matches inode */
1749         if (nd->flags & LOOKUP_RCU) {
1750                 if (read_seqcount_retry(&path->dentry->d_seq, seq))
1751                         return -ECHILD;
1752         }
1753         return pick_link(nd, path, inode, seq);
1754 }
1755
1756 static int walk_component(struct nameidata *nd, int flags)
1757 {
1758         struct path path;
1759         struct inode *inode;
1760         unsigned seq;
1761         int err;
1762         /*
1763          * "." and ".." are special - ".." especially so because it has
1764          * to be able to know about the current root directory and
1765          * parent relationships.
1766          */
1767         if (unlikely(nd->last_type != LAST_NORM)) {
1768                 err = handle_dots(nd, nd->last_type);
1769                 if (!(flags & WALK_MORE) && nd->depth)
1770                         put_link(nd);
1771                 return err;
1772         }
1773         err = lookup_fast(nd, &path, &inode, &seq);
1774         if (unlikely(err <= 0)) {
1775                 if (err < 0)
1776                         return err;
1777                 path.dentry = lookup_slow(&nd->last, nd->path.dentry,
1778                                           nd->flags);
1779                 if (IS_ERR(path.dentry))
1780                         return PTR_ERR(path.dentry);
1781
1782                 path.mnt = nd->path.mnt;
1783                 err = follow_managed(&path, nd);
1784                 if (unlikely(err < 0))
1785                         return err;
1786
1787                 if (unlikely(d_is_negative(path.dentry))) {
1788                         path_to_nameidata(&path, nd);
1789                         return -ENOENT;
1790                 }
1791
1792                 seq = 0;        /* we are already out of RCU mode */
1793                 inode = d_backing_inode(path.dentry);
1794         }
1795
1796         return step_into(nd, &path, flags, inode, seq);
1797 }
1798
1799 /*
1800  * We can do the critical dentry name comparison and hashing
1801  * operations one word at a time, but we are limited to:
1802  *
1803  * - Architectures with fast unaligned word accesses. We could
1804  *   do a "get_unaligned()" if this helps and is sufficiently
1805  *   fast.
1806  *
1807  * - non-CONFIG_DEBUG_PAGEALLOC configurations (so that we
1808  *   do not trap on the (extremely unlikely) case of a page
1809  *   crossing operation.
1810  *
1811  * - Furthermore, we need an efficient 64-bit compile for the
1812  *   64-bit case in order to generate the "number of bytes in
1813  *   the final mask". Again, that could be replaced with a
1814  *   efficient population count instruction or similar.
1815  */
1816 #ifdef CONFIG_DCACHE_WORD_ACCESS
1817
1818 #include <asm/word-at-a-time.h>
1819
1820 #ifdef HASH_MIX
1821
1822 /* Architecture provides HASH_MIX and fold_hash() in <asm/hash.h> */
1823
1824 #elif defined(CONFIG_64BIT)
1825 /*
1826  * Register pressure in the mixing function is an issue, particularly
1827  * on 32-bit x86, but almost any function requires one state value and
1828  * one temporary.  Instead, use a function designed for two state values
1829  * and no temporaries.
1830  *
1831  * This function cannot create a collision in only two iterations, so
1832  * we have two iterations to achieve avalanche.  In those two iterations,
1833  * we have six layers of mixing, which is enough to spread one bit's
1834  * influence out to 2^6 = 64 state bits.
1835  *
1836  * Rotate constants are scored by considering either 64 one-bit input
1837  * deltas or 64*63/2 = 2016 two-bit input deltas, and finding the
1838  * probability of that delta causing a change to each of the 128 output
1839  * bits, using a sample of random initial states.
1840  *
1841  * The Shannon entropy of the computed probabilities is then summed
1842  * to produce a score.  Ideally, any input change has a 50% chance of
1843  * toggling any given output bit.
1844  *
1845  * Mixing scores (in bits) for (12,45):
1846  * Input delta: 1-bit      2-bit
1847  * 1 round:     713.3    42542.6
1848  * 2 rounds:   2753.7   140389.8
1849  * 3 rounds:   5954.1   233458.2
1850  * 4 rounds:   7862.6   256672.2
1851  * Perfect:    8192     258048
1852  *            (64*128) (64*63/2 * 128)
1853  */
1854 #define HASH_MIX(x, y, a)       \
1855         (       x ^= (a),       \
1856         y ^= x, x = rol64(x,12),\
1857         x += y, y = rol64(y,45),\
1858         y *= 9                  )
1859
1860 /*
1861  * Fold two longs into one 32-bit hash value.  This must be fast, but
1862  * latency isn't quite as critical, as there is a fair bit of additional
1863  * work done before the hash value is used.
1864  */
1865 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1866 {
1867         y ^= x * GOLDEN_RATIO_64;
1868         y *= GOLDEN_RATIO_64;
1869         return y >> 32;
1870 }
1871
1872 #else   /* 32-bit case */
1873
1874 /*
1875  * Mixing scores (in bits) for (7,20):
1876  * Input delta: 1-bit      2-bit
1877  * 1 round:     330.3     9201.6
1878  * 2 rounds:   1246.4    25475.4
1879  * 3 rounds:   1907.1    31295.1
1880  * 4 rounds:   2042.3    31718.6
1881  * Perfect:    2048      31744
1882  *            (32*64)   (32*31/2 * 64)
1883  */
1884 #define HASH_MIX(x, y, a)       \
1885         (       x ^= (a),       \
1886         y ^= x, x = rol32(x, 7),\
1887         x += y, y = rol32(y,20),\
1888         y *= 9                  )
1889
1890 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1891 {
1892         /* Use arch-optimized multiply if one exists */
1893         return __hash_32(y ^ __hash_32(x));
1894 }
1895
1896 #endif
1897
1898 /*
1899  * Return the hash of a string of known length.  This is carfully
1900  * designed to match hash_name(), which is the more critical function.
1901  * In particular, we must end by hashing a final word containing 0..7
1902  * payload bytes, to match the way that hash_name() iterates until it
1903  * finds the delimiter after the name.
1904  */
1905 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1906 {
1907         unsigned long a, x = 0, y = (unsigned long)salt;
1908
1909         for (;;) {
1910                 if (!len)
1911                         goto done;
1912                 a = load_unaligned_zeropad(name);
1913                 if (len < sizeof(unsigned long))
1914                         break;
1915                 HASH_MIX(x, y, a);
1916                 name += sizeof(unsigned long);
1917                 len -= sizeof(unsigned long);
1918         }
1919         x ^= a & bytemask_from_count(len);
1920 done:
1921         return fold_hash(x, y);
1922 }
1923 EXPORT_SYMBOL(full_name_hash);
1924
1925 /* Return the "hash_len" (hash and length) of a null-terminated string */
1926 u64 hashlen_string(const void *salt, const char *name)
1927 {
1928         unsigned long a = 0, x = 0, y = (unsigned long)salt;
1929         unsigned long adata, mask, len;
1930         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1931
1932         len = 0;
1933         goto inside;
1934
1935         do {
1936                 HASH_MIX(x, y, a);
1937                 len += sizeof(unsigned long);
1938 inside:
1939                 a = load_unaligned_zeropad(name+len);
1940         } while (!has_zero(a, &adata, &constants));
1941
1942         adata = prep_zero_mask(a, adata, &constants);
1943         mask = create_zero_mask(adata);
1944         x ^= a & zero_bytemask(mask);
1945
1946         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1947 }
1948 EXPORT_SYMBOL(hashlen_string);
1949
1950 /*
1951  * Calculate the length and hash of the path component, and
1952  * return the "hash_len" as the result.
1953  */
1954 static inline u64 hash_name(const void *salt, const char *name)
1955 {
1956         unsigned long a = 0, b, x = 0, y = (unsigned long)salt;
1957         unsigned long adata, bdata, mask, len;
1958         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1959
1960         len = 0;
1961         goto inside;
1962
1963         do {
1964                 HASH_MIX(x, y, a);
1965                 len += sizeof(unsigned long);
1966 inside:
1967                 a = load_unaligned_zeropad(name+len);
1968                 b = a ^ REPEAT_BYTE('/');
1969         } while (!(has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)));
1970
1971         adata = prep_zero_mask(a, adata, &constants);
1972         bdata = prep_zero_mask(b, bdata, &constants);
1973         mask = create_zero_mask(adata | bdata);
1974         x ^= a & zero_bytemask(mask);
1975
1976         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1977 }
1978
1979 #else   /* !CONFIG_DCACHE_WORD_ACCESS: Slow, byte-at-a-time version */
1980
1981 /* Return the hash of a string of known length */
1982 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1983 {
1984         unsigned long hash = init_name_hash(salt);
1985         while (len--)
1986                 hash = partial_name_hash((unsigned char)*name++, hash);
1987         return end_name_hash(hash);
1988 }
1989 EXPORT_SYMBOL(full_name_hash);
1990
1991 /* Return the "hash_len" (hash and length) of a null-terminated string */
1992 u64 hashlen_string(const void *salt, const char *name)
1993 {
1994         unsigned long hash = init_name_hash(salt);
1995         unsigned long len = 0, c;
1996
1997         c = (unsigned char)*name;
1998         while (c) {
1999                 len++;
2000                 hash = partial_name_hash(c, hash);
2001                 c = (unsigned char)name[len];
2002         }
2003         return hashlen_create(end_name_hash(hash), len);
2004 }
2005 EXPORT_SYMBOL(hashlen_string);
2006
2007 /*
2008  * We know there's a real path component here of at least
2009  * one character.
2010  */
2011 static inline u64 hash_name(const void *salt, const char *name)
2012 {
2013         unsigned long hash = init_name_hash(salt);
2014         unsigned long len = 0, c;
2015
2016         c = (unsigned char)*name;
2017         do {
2018                 len++;
2019                 hash = partial_name_hash(c, hash);
2020                 c = (unsigned char)name[len];
2021         } while (c && c != '/');
2022         return hashlen_create(end_name_hash(hash), len);
2023 }
2024
2025 #endif
2026
2027 /*
2028  * Name resolution.
2029  * This is the basic name resolution function, turning a pathname into
2030  * the final dentry. We expect 'base' to be positive and a directory.
2031  *
2032  * Returns 0 and nd will have valid dentry and mnt on success.
2033  * Returns error and drops reference to input namei data on failure.
2034  */
2035 static int link_path_walk(const char *name, struct nameidata *nd)
2036 {
2037         int err;
2038
2039         while (*name=='/')
2040                 name++;
2041         if (!*name)
2042                 return 0;
2043
2044         /* At this point we know we have a real path component. */
2045         for(;;) {
2046                 u64 hash_len;
2047                 int type;
2048
2049                 err = may_lookup(nd);
2050                 if (err)
2051                         return err;
2052
2053                 hash_len = hash_name(nd->path.dentry, name);
2054
2055                 type = LAST_NORM;
2056                 if (name[0] == '.') switch (hashlen_len(hash_len)) {
2057                         case 2:
2058                                 if (name[1] == '.') {
2059                                         type = LAST_DOTDOT;
2060                                         nd->flags |= LOOKUP_JUMPED;
2061                                 }
2062                                 break;
2063                         case 1:
2064                                 type = LAST_DOT;
2065                 }
2066                 if (likely(type == LAST_NORM)) {
2067                         struct dentry *parent = nd->path.dentry;
2068                         nd->flags &= ~LOOKUP_JUMPED;
2069                         if (unlikely(parent->d_flags & DCACHE_OP_HASH)) {
2070                                 struct qstr this = { { .hash_len = hash_len }, .name = name };
2071                                 err = parent->d_op->d_hash(parent, &this);
2072                                 if (err < 0)
2073                                         return err;
2074                                 hash_len = this.hash_len;
2075                                 name = this.name;
2076                         }
2077                 }
2078
2079                 nd->last.hash_len = hash_len;
2080                 nd->last.name = name;
2081                 nd->last_type = type;
2082
2083                 name += hashlen_len(hash_len);
2084                 if (!*name)
2085                         goto OK;
2086                 /*
2087                  * If it wasn't NUL, we know it was '/'. Skip that
2088                  * slash, and continue until no more slashes.
2089                  */
2090                 do {
2091                         name++;
2092                 } while (unlikely(*name == '/'));
2093                 if (unlikely(!*name)) {
2094 OK:
2095                         /* pathname body, done */
2096                         if (!nd->depth)
2097                                 return 0;
2098                         name = nd->stack[nd->depth - 1].name;
2099                         /* trailing symlink, done */
2100                         if (!name)
2101                                 return 0;
2102                         /* last component of nested symlink */
2103                         err = walk_component(nd, WALK_FOLLOW);
2104                 } else {
2105                         /* not the last component */
2106                         err = walk_component(nd, WALK_FOLLOW | WALK_MORE);
2107                 }
2108                 if (err < 0)
2109                         return err;
2110
2111                 if (err) {
2112                         const char *s = get_link(nd);
2113
2114                         if (IS_ERR(s))
2115                                 return PTR_ERR(s);
2116                         err = 0;
2117                         if (unlikely(!s)) {
2118                                 /* jumped */
2119                                 put_link(nd);
2120                         } else {
2121                                 nd->stack[nd->depth - 1].name = name;
2122                                 name = s;
2123                                 continue;
2124                         }
2125                 }
2126                 if (unlikely(!d_can_lookup(nd->path.dentry))) {
2127                         if (nd->flags & LOOKUP_RCU) {
2128                                 if (unlazy_walk(nd, NULL, 0))
2129                                         return -ECHILD;
2130                         }
2131                         return -ENOTDIR;
2132                 }
2133         }
2134 }
2135
2136 static const char *path_init(struct nameidata *nd, unsigned flags)
2137 {
2138         int retval = 0;
2139         const char *s = nd->name->name;
2140
2141         nd->last_type = LAST_ROOT; /* if there are only slashes... */
2142         nd->flags = flags | LOOKUP_JUMPED | LOOKUP_PARENT;
2143         nd->depth = 0;
2144         if (flags & LOOKUP_ROOT) {
2145                 struct dentry *root = nd->root.dentry;
2146                 struct inode *inode = root->d_inode;
2147                 if (*s) {
2148                         if (!d_can_lookup(root))
2149                                 return ERR_PTR(-ENOTDIR);
2150                         retval = inode_permission(inode, MAY_EXEC);
2151                         if (retval)
2152                                 return ERR_PTR(retval);
2153                 }
2154                 nd->path = nd->root;
2155                 nd->inode = inode;
2156                 if (flags & LOOKUP_RCU) {
2157                         rcu_read_lock();
2158                         nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2159                         nd->root_seq = nd->seq;
2160                         nd->m_seq = read_seqbegin(&mount_lock);
2161                 } else {
2162                         path_get(&nd->path);
2163                 }
2164                 return s;
2165         }
2166
2167         nd->root.mnt = NULL;
2168         nd->path.mnt = NULL;
2169         nd->path.dentry = NULL;
2170
2171         nd->m_seq = read_seqbegin(&mount_lock);
2172         if (*s == '/') {
2173                 if (flags & LOOKUP_RCU)
2174                         rcu_read_lock();
2175                 set_root(nd);
2176                 if (likely(!nd_jump_root(nd)))
2177                         return s;
2178                 nd->root.mnt = NULL;
2179                 rcu_read_unlock();
2180                 return ERR_PTR(-ECHILD);
2181         } else if (nd->dfd == AT_FDCWD) {
2182                 if (flags & LOOKUP_RCU) {
2183                         struct fs_struct *fs = current->fs;
2184                         unsigned seq;
2185
2186                         rcu_read_lock();
2187
2188                         do {
2189                                 seq = read_seqcount_begin(&fs->seq);
2190                                 nd->path = fs->pwd;
2191                                 nd->inode = nd->path.dentry->d_inode;
2192                                 nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2193                         } while (read_seqcount_retry(&fs->seq, seq));
2194                 } else {
2195                         get_fs_pwd(current->fs, &nd->path);
2196                         nd->inode = nd->path.dentry->d_inode;
2197                 }
2198                 return s;
2199         } else {
2200                 /* Caller must check execute permissions on the starting path component */
2201                 struct fd f = fdget_raw(nd->dfd);
2202                 struct dentry *dentry;
2203
2204                 if (!f.file)
2205                         return ERR_PTR(-EBADF);
2206
2207                 dentry = f.file->f_path.dentry;
2208
2209                 if (*s) {
2210                         if (!d_can_lookup(dentry)) {
2211                                 fdput(f);
2212                                 return ERR_PTR(-ENOTDIR);
2213                         }
2214                 }
2215
2216                 nd->path = f.file->f_path;
2217                 if (flags & LOOKUP_RCU) {
2218                         rcu_read_lock();
2219                         nd->inode = nd->path.dentry->d_inode;
2220                         nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
2221                 } else {
2222                         path_get(&nd->path);
2223                         nd->inode = nd->path.dentry->d_inode;
2224                 }
2225                 fdput(f);
2226                 return s;
2227         }
2228 }
2229
2230 static const char *trailing_symlink(struct nameidata *nd)
2231 {
2232         const char *s;
2233         int error = may_follow_link(nd);
2234         if (unlikely(error))
2235                 return ERR_PTR(error);
2236         nd->flags |= LOOKUP_PARENT;
2237         nd->stack[0].name = NULL;
2238         s = get_link(nd);
2239         return s ? s : "";
2240 }
2241
2242 static inline int lookup_last(struct nameidata *nd)
2243 {
2244         if (nd->last_type == LAST_NORM && nd->last.name[nd->last.len])
2245                 nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
2246
2247         nd->flags &= ~LOOKUP_PARENT;
2248         return walk_component(nd, 0);
2249 }
2250
2251 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2252 static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
2253 {
2254         const char *s = path_init(nd, flags);
2255         int err;
2256
2257         if (IS_ERR(s))
2258                 return PTR_ERR(s);
2259         while (!(err = link_path_walk(s, nd))
2260                 && ((err = lookup_last(nd)) > 0)) {
2261                 s = trailing_symlink(nd);
2262                 if (IS_ERR(s)) {
2263                         err = PTR_ERR(s);
2264                         break;
2265                 }
2266         }
2267         if (!err)
2268                 err = complete_walk(nd);
2269
2270         if (!err && nd->flags & LOOKUP_DIRECTORY)
2271                 if (!d_can_lookup(nd->path.dentry))
2272                         err = -ENOTDIR;
2273         if (!err) {
2274                 *path = nd->path;
2275                 nd->path.mnt = NULL;
2276                 nd->path.dentry = NULL;
2277         }
2278         terminate_walk(nd);
2279         return err;
2280 }
2281
2282 static int filename_lookup(int dfd, struct filename *name, unsigned flags,
2283                            struct path *path, struct path *root)
2284 {
2285         int retval;
2286         struct nameidata nd;
2287         if (IS_ERR(name))
2288                 return PTR_ERR(name);
2289         if (unlikely(root)) {
2290                 nd.root = *root;
2291                 flags |= LOOKUP_ROOT;
2292         }
2293         set_nameidata(&nd, dfd, name);
2294         retval = path_lookupat(&nd, flags | LOOKUP_RCU, path);
2295         if (unlikely(retval == -ECHILD))
2296                 retval = path_lookupat(&nd, flags, path);
2297         if (unlikely(retval == -ESTALE))
2298                 retval = path_lookupat(&nd, flags | LOOKUP_REVAL, path);
2299
2300         if (likely(!retval))
2301                 audit_inode(name, path->dentry, flags & LOOKUP_PARENT);
2302         restore_nameidata();
2303         putname(name);
2304         return retval;
2305 }
2306
2307 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2308 static int path_parentat(struct nameidata *nd, unsigned flags,
2309                                 struct path *parent)
2310 {
2311         const char *s = path_init(nd, flags);
2312         int err;
2313         if (IS_ERR(s))
2314                 return PTR_ERR(s);
2315         err = link_path_walk(s, nd);
2316         if (!err)
2317                 err = complete_walk(nd);
2318         if (!err) {
2319                 *parent = nd->path;
2320                 nd->path.mnt = NULL;
2321                 nd->path.dentry = NULL;
2322         }
2323         terminate_walk(nd);
2324         return err;
2325 }
2326
2327 static struct filename *filename_parentat(int dfd, struct filename *name,
2328                                 unsigned int flags, struct path *parent,
2329                                 struct qstr *last, int *type)
2330 {
2331         int retval;
2332         struct nameidata nd;
2333
2334         if (IS_ERR(name))
2335                 return name;
2336         set_nameidata(&nd, dfd, name);
2337         retval = path_parentat(&nd, flags | LOOKUP_RCU, parent);
2338         if (unlikely(retval == -ECHILD))
2339                 retval = path_parentat(&nd, flags, parent);
2340         if (unlikely(retval == -ESTALE))
2341                 retval = path_parentat(&nd, flags | LOOKUP_REVAL, parent);
2342         if (likely(!retval)) {
2343                 *last = nd.last;
2344                 *type = nd.last_type;
2345                 audit_inode(name, parent->dentry, LOOKUP_PARENT);
2346         } else {
2347                 putname(name);
2348                 name = ERR_PTR(retval);
2349         }
2350         restore_nameidata();
2351         return name;
2352 }
2353
2354 /* does lookup, returns the object with parent locked */
2355 struct dentry *kern_path_locked(const char *name, struct path *path)
2356 {
2357         struct filename *filename;
2358         struct dentry *d;
2359         struct qstr last;
2360         int type;
2361
2362         filename = filename_parentat(AT_FDCWD, getname_kernel(name), 0, path,
2363                                     &last, &type);
2364         if (IS_ERR(filename))
2365                 return ERR_CAST(filename);
2366         if (unlikely(type != LAST_NORM)) {
2367                 path_put(path);
2368                 putname(filename);
2369                 return ERR_PTR(-EINVAL);
2370         }
2371         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
2372         d = __lookup_hash(&last, path->dentry, 0);
2373         if (IS_ERR(d)) {
2374                 inode_unlock(path->dentry->d_inode);
2375                 path_put(path);
2376         }
2377         putname(filename);
2378         return d;
2379 }
2380
2381 int kern_path(const char *name, unsigned int flags, struct path *path)
2382 {
2383         return filename_lookup(AT_FDCWD, getname_kernel(name),
2384                                flags, path, NULL);
2385 }
2386 EXPORT_SYMBOL(kern_path);
2387
2388 /**
2389  * vfs_path_lookup - lookup a file path relative to a dentry-vfsmount pair
2390  * @dentry:  pointer to dentry of the base directory
2391  * @mnt: pointer to vfs mount of the base directory
2392  * @name: pointer to file name
2393  * @flags: lookup flags
2394  * @path: pointer to struct path to fill
2395  */
2396 int vfs_path_lookup(struct dentry *dentry, struct vfsmount *mnt,
2397                     const char *name, unsigned int flags,
2398                     struct path *path)
2399 {
2400         struct path root = {.mnt = mnt, .dentry = dentry};
2401         /* the first argument of filename_lookup() is ignored with root */
2402         return filename_lookup(AT_FDCWD, getname_kernel(name),
2403                                flags , path, &root);
2404 }
2405 EXPORT_SYMBOL(vfs_path_lookup);
2406
2407 /**
2408  * lookup_one_len - filesystem helper to lookup single pathname component
2409  * @name:       pathname component to lookup
2410  * @base:       base directory to lookup from
2411  * @len:        maximum length @len should be interpreted to
2412  *
2413  * Note that this routine is purely a helper for filesystem usage and should
2414  * not be called by generic code.
2415  *
2416  * The caller must hold base->i_mutex.
2417  */
2418 struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
2419 {
2420         struct qstr this;
2421         unsigned int c;
2422         int err;
2423
2424         WARN_ON_ONCE(!inode_is_locked(base->d_inode));
2425
2426         this.name = name;
2427         this.len = len;
2428         this.hash = full_name_hash(base, name, len);
2429         if (!len)
2430                 return ERR_PTR(-EACCES);
2431
2432         if (unlikely(name[0] == '.')) {
2433                 if (len < 2 || (len == 2 && name[1] == '.'))
2434                         return ERR_PTR(-EACCES);
2435         }
2436
2437         while (len--) {
2438                 c = *(const unsigned char *)name++;
2439                 if (c == '/' || c == '\0')
2440                         return ERR_PTR(-EACCES);
2441         }
2442         /*
2443          * See if the low-level filesystem might want
2444          * to use its own hash..
2445          */
2446         if (base->d_flags & DCACHE_OP_HASH) {
2447                 int err = base->d_op->d_hash(base, &this);
2448                 if (err < 0)
2449                         return ERR_PTR(err);
2450         }
2451
2452         err = inode_permission(base->d_inode, MAY_EXEC);
2453         if (err)
2454                 return ERR_PTR(err);
2455
2456         return __lookup_hash(&this, base, 0);
2457 }
2458 EXPORT_SYMBOL(lookup_one_len);
2459
2460 /**
2461  * lookup_one_len_unlocked - filesystem helper to lookup single pathname component
2462  * @name:       pathname component to lookup
2463  * @base:       base directory to lookup from
2464  * @len:        maximum length @len should be interpreted to
2465  *
2466  * Note that this routine is purely a helper for filesystem usage and should
2467  * not be called by generic code.
2468  *
2469  * Unlike lookup_one_len, it should be called without the parent
2470  * i_mutex held, and will take the i_mutex itself if necessary.
2471  */
2472 struct dentry *lookup_one_len_unlocked(const char *name,
2473                                        struct dentry *base, int len)
2474 {
2475         struct qstr this;
2476         unsigned int c;
2477         int err;
2478         struct dentry *ret;
2479
2480         this.name = name;
2481         this.len = len;
2482         this.hash = full_name_hash(base, name, len);
2483         if (!len)
2484                 return ERR_PTR(-EACCES);
2485
2486         if (unlikely(name[0] == '.')) {
2487                 if (len < 2 || (len == 2 && name[1] == '.'))
2488                         return ERR_PTR(-EACCES);
2489         }
2490
2491         while (len--) {
2492                 c = *(const unsigned char *)name++;
2493                 if (c == '/' || c == '\0')
2494                         return ERR_PTR(-EACCES);
2495         }
2496         /*
2497          * See if the low-level filesystem might want
2498          * to use its own hash..
2499          */
2500         if (base->d_flags & DCACHE_OP_HASH) {
2501                 int err = base->d_op->d_hash(base, &this);
2502                 if (err < 0)
2503                         return ERR_PTR(err);
2504         }
2505
2506         err = inode_permission(base->d_inode, MAY_EXEC);
2507         if (err)
2508                 return ERR_PTR(err);
2509
2510         ret = lookup_dcache(&this, base, 0);
2511         if (!ret)
2512                 ret = lookup_slow(&this, base, 0);
2513         return ret;
2514 }
2515 EXPORT_SYMBOL(lookup_one_len_unlocked);
2516
2517 #ifdef CONFIG_UNIX98_PTYS
2518 int path_pts(struct path *path)
2519 {
2520         /* Find something mounted on "pts" in the same directory as
2521          * the input path.
2522          */
2523         struct dentry *child, *parent;
2524         struct qstr this;
2525         int ret;
2526
2527         ret = path_parent_directory(path);
2528         if (ret)
2529                 return ret;
2530
2531         parent = path->dentry;
2532         this.name = "pts";
2533         this.len = 3;
2534         child = d_hash_and_lookup(parent, &this);
2535         if (!child)
2536                 return -ENOENT;
2537
2538         path->dentry = child;
2539         dput(parent);
2540         follow_mount(path);
2541         return 0;
2542 }
2543 #endif
2544
2545 int user_path_at_empty(int dfd, const char __user *name, unsigned flags,
2546                  struct path *path, int *empty)
2547 {
2548         return filename_lookup(dfd, getname_flags(name, flags, empty),
2549                                flags, path, NULL);
2550 }
2551 EXPORT_SYMBOL(user_path_at_empty);
2552
2553 /**
2554  * mountpoint_last - look up last component for umount
2555  * @nd:   pathwalk nameidata - currently pointing at parent directory of "last"
2556  *
2557  * This is a special lookup_last function just for umount. In this case, we
2558  * need to resolve the path without doing any revalidation.
2559  *
2560  * The nameidata should be the result of doing a LOOKUP_PARENT pathwalk. Since
2561  * mountpoints are always pinned in the dcache, their ancestors are too. Thus,
2562  * in almost all cases, this lookup will be served out of the dcache. The only
2563  * cases where it won't are if nd->last refers to a symlink or the path is
2564  * bogus and it doesn't exist.
2565  *
2566  * Returns:
2567  * -error: if there was an error during lookup. This includes -ENOENT if the
2568  *         lookup found a negative dentry.
2569  *
2570  * 0:      if we successfully resolved nd->last and found it to not to be a
2571  *         symlink that needs to be followed.
2572  *
2573  * 1:      if we successfully resolved nd->last and found it to be a symlink
2574  *         that needs to be followed.
2575  */
2576 static int
2577 mountpoint_last(struct nameidata *nd)
2578 {
2579         int error = 0;
2580         struct dentry *dir = nd->path.dentry;
2581         struct path path;
2582
2583         /* If we're in rcuwalk, drop out of it to handle last component */
2584         if (nd->flags & LOOKUP_RCU) {
2585                 if (unlazy_walk(nd, NULL, 0))
2586                         return -ECHILD;
2587         }
2588
2589         nd->flags &= ~LOOKUP_PARENT;
2590
2591         if (unlikely(nd->last_type != LAST_NORM)) {
2592                 error = handle_dots(nd, nd->last_type);
2593                 if (error)
2594                         return error;
2595                 path.dentry = dget(nd->path.dentry);
2596         } else {
2597                 path.dentry = d_lookup(dir, &nd->last);
2598                 if (!path.dentry) {
2599                         /*
2600                          * No cached dentry. Mounted dentries are pinned in the
2601                          * cache, so that means that this dentry is probably
2602                          * a symlink or the path doesn't actually point
2603                          * to a mounted dentry.
2604                          */
2605                         path.dentry = lookup_slow(&nd->last, dir,
2606                                              nd->flags | LOOKUP_NO_REVAL);
2607                         if (IS_ERR(path.dentry))
2608                                 return PTR_ERR(path.dentry);
2609                 }
2610         }
2611         if (d_is_negative(path.dentry)) {
2612                 dput(path.dentry);
2613                 return -ENOENT;
2614         }
2615         path.mnt = nd->path.mnt;
2616         return step_into(nd, &path, 0, d_backing_inode(path.dentry), 0);
2617 }
2618
2619 /**
2620  * path_mountpoint - look up a path to be umounted
2621  * @nd:         lookup context
2622  * @flags:      lookup flags
2623  * @path:       pointer to container for result
2624  *
2625  * Look up the given name, but don't attempt to revalidate the last component.
2626  * Returns 0 and "path" will be valid on success; Returns error otherwise.
2627  */
2628 static int
2629 path_mountpoint(struct nameidata *nd, unsigned flags, struct path *path)
2630 {
2631         const char *s = path_init(nd, flags);
2632         int err;
2633         if (IS_ERR(s))
2634                 return PTR_ERR(s);
2635         while (!(err = link_path_walk(s, nd)) &&
2636                 (err = mountpoint_last(nd)) > 0) {
2637                 s = trailing_symlink(nd);
2638                 if (IS_ERR(s)) {
2639                         err = PTR_ERR(s);
2640                         break;
2641                 }
2642         }
2643         if (!err) {
2644                 *path = nd->path;
2645                 nd->path.mnt = NULL;
2646                 nd->path.dentry = NULL;
2647                 follow_mount(path);
2648         }
2649         terminate_walk(nd);
2650         return err;
2651 }
2652
2653 static int
2654 filename_mountpoint(int dfd, struct filename *name, struct path *path,
2655                         unsigned int flags)
2656 {
2657         struct nameidata nd;
2658         int error;
2659         if (IS_ERR(name))
2660                 return PTR_ERR(name);
2661         set_nameidata(&nd, dfd, name);
2662         error = path_mountpoint(&nd, flags | LOOKUP_RCU, path);
2663         if (unlikely(error == -ECHILD))
2664                 error = path_mountpoint(&nd, flags, path);
2665         if (unlikely(error == -ESTALE))
2666                 error = path_mountpoint(&nd, flags | LOOKUP_REVAL, path);
2667         if (likely(!error))
2668                 audit_inode(name, path->dentry, 0);
2669         restore_nameidata();
2670         putname(name);
2671         return error;
2672 }
2673
2674 /**
2675  * user_path_mountpoint_at - lookup a path from userland in order to umount it
2676  * @dfd:        directory file descriptor
2677  * @name:       pathname from userland
2678  * @flags:      lookup flags
2679  * @path:       pointer to container to hold result
2680  *
2681  * A umount is a special case for path walking. We're not actually interested
2682  * in the inode in this situation, and ESTALE errors can be a problem. We
2683  * simply want track down the dentry and vfsmount attached at the mountpoint
2684  * and avoid revalidating the last component.
2685  *
2686  * Returns 0 and populates "path" on success.
2687  */
2688 int
2689 user_path_mountpoint_at(int dfd, const char __user *name, unsigned int flags,
2690                         struct path *path)
2691 {
2692         return filename_mountpoint(dfd, getname(name), path, flags);
2693 }
2694
2695 int
2696 kern_path_mountpoint(int dfd, const char *name, struct path *path,
2697                         unsigned int flags)
2698 {
2699         return filename_mountpoint(dfd, getname_kernel(name), path, flags);
2700 }
2701 EXPORT_SYMBOL(kern_path_mountpoint);
2702
2703 int __check_sticky(struct inode *dir, struct inode *inode)
2704 {
2705         kuid_t fsuid = current_fsuid();
2706
2707         if (uid_eq(inode->i_uid, fsuid))
2708                 return 0;
2709         if (uid_eq(dir->i_uid, fsuid))
2710                 return 0;
2711         return !capable_wrt_inode_uidgid(inode, CAP_FOWNER);
2712 }
2713 EXPORT_SYMBOL(__check_sticky);
2714
2715 /*
2716  *      Check whether we can remove a link victim from directory dir, check
2717  *  whether the type of victim is right.
2718  *  1. We can't do it if dir is read-only (done in permission())
2719  *  2. We should have write and exec permissions on dir
2720  *  3. We can't remove anything from append-only dir
2721  *  4. We can't do anything with immutable dir (done in permission())
2722  *  5. If the sticky bit on dir is set we should either
2723  *      a. be owner of dir, or
2724  *      b. be owner of victim, or
2725  *      c. have CAP_FOWNER capability
2726  *  6. If the victim is append-only or immutable we can't do antyhing with
2727  *     links pointing to it.
2728  *  7. If the victim has an unknown uid or gid we can't change the inode.
2729  *  8. If we were asked to remove a directory and victim isn't one - ENOTDIR.
2730  *  9. If we were asked to remove a non-directory and victim isn't one - EISDIR.
2731  * 10. We can't remove a root or mountpoint.
2732  * 11. We don't allow removal of NFS sillyrenamed files; it's handled by
2733  *     nfs_async_unlink().
2734  */
2735 static int may_delete(struct inode *dir, struct dentry *victim, bool isdir)
2736 {
2737         struct inode *inode = d_backing_inode(victim);
2738         int error;
2739
2740         if (d_is_negative(victim))
2741                 return -ENOENT;
2742         BUG_ON(!inode);
2743
2744         BUG_ON(victim->d_parent->d_inode != dir);
2745         audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);
2746
2747         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
2748         if (error)
2749                 return error;
2750         if (IS_APPEND(dir))
2751                 return -EPERM;
2752
2753         if (check_sticky(dir, inode) || IS_APPEND(inode) ||
2754             IS_IMMUTABLE(inode) || IS_SWAPFILE(inode) || HAS_UNMAPPED_ID(inode))
2755                 return -EPERM;
2756         if (isdir) {
2757                 if (!d_is_dir(victim))
2758                         return -ENOTDIR;
2759                 if (IS_ROOT(victim))
2760                         return -EBUSY;
2761         } else if (d_is_dir(victim))
2762                 return -EISDIR;
2763         if (IS_DEADDIR(dir))
2764                 return -ENOENT;
2765         if (victim->d_flags & DCACHE_NFSFS_RENAMED)
2766                 return -EBUSY;
2767         return 0;
2768 }
2769
2770 /*      Check whether we can create an object with dentry child in directory
2771  *  dir.
2772  *  1. We can't do it if child already exists (open has special treatment for
2773  *     this case, but since we are inlined it's OK)
2774  *  2. We can't do it if dir is read-only (done in permission())
2775  *  3. We can't do it if the fs can't represent the fsuid or fsgid.
2776  *  4. We should have write and exec permissions on dir
2777  *  5. We can't do it if dir is immutable (done in permission())
2778  */
2779 static inline int may_create(struct inode *dir, struct dentry *child)
2780 {
2781         struct user_namespace *s_user_ns;
2782         audit_inode_child(dir, child, AUDIT_TYPE_CHILD_CREATE);
2783         if (child->d_inode)
2784                 return -EEXIST;
2785         if (IS_DEADDIR(dir))
2786                 return -ENOENT;
2787         s_user_ns = dir->i_sb->s_user_ns;
2788         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
2789             !kgid_has_mapping(s_user_ns, current_fsgid()))
2790                 return -EOVERFLOW;
2791         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
2792 }
2793
2794 /*
2795  * p1 and p2 should be directories on the same fs.
2796  */
2797 struct dentry *lock_rename(struct dentry *p1, struct dentry *p2)
2798 {
2799         struct dentry *p;
2800
2801         if (p1 == p2) {
2802                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2803                 return NULL;
2804         }
2805
2806         mutex_lock(&p1->d_sb->s_vfs_rename_mutex);
2807
2808         p = d_ancestor(p2, p1);
2809         if (p) {
2810                 inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
2811                 inode_lock_nested(p1->d_inode, I_MUTEX_CHILD);
2812                 return p;
2813         }
2814
2815         p = d_ancestor(p1, p2);
2816         if (p) {
2817                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2818                 inode_lock_nested(p2->d_inode, I_MUTEX_CHILD);
2819                 return p;
2820         }
2821
2822         inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2823         inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
2824         return NULL;
2825 }
2826 EXPORT_SYMBOL(lock_rename);
2827
2828 void unlock_rename(struct dentry *p1, struct dentry *p2)
2829 {
2830         inode_unlock(p1->d_inode);
2831         if (p1 != p2) {
2832                 inode_unlock(p2->d_inode);
2833                 mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
2834         }
2835 }
2836 EXPORT_SYMBOL(unlock_rename);
2837
2838 int vfs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2839                 bool want_excl)
2840 {
2841         int error = may_create(dir, dentry);
2842         if (error)
2843                 return error;
2844
2845         if (!dir->i_op->create)
2846                 return -EACCES; /* shouldn't it be ENOSYS? */
2847         mode &= S_IALLUGO;
2848         mode |= S_IFREG;
2849         error = security_inode_create(dir, dentry, mode);
2850         if (error)
2851                 return error;
2852         error = dir->i_op->create(dir, dentry, mode, want_excl);
2853         if (!error)
2854                 fsnotify_create(dir, dentry);
2855         return error;
2856 }
2857 EXPORT_SYMBOL(vfs_create);
2858
2859 bool may_open_dev(const struct path *path)
2860 {
2861         return !(path->mnt->mnt_flags & MNT_NODEV) &&
2862                 !(path->mnt->mnt_sb->s_iflags & SB_I_NODEV);
2863 }
2864
2865 static int may_open(const struct path *path, int acc_mode, int flag)
2866 {
2867         struct dentry *dentry = path->dentry;
2868         struct inode *inode = dentry->d_inode;
2869         int error;
2870
2871         if (!inode)
2872                 return -ENOENT;
2873
2874         switch (inode->i_mode & S_IFMT) {
2875         case S_IFLNK:
2876                 return -ELOOP;
2877         case S_IFDIR:
2878                 if (acc_mode & MAY_WRITE)
2879                         return -EISDIR;
2880                 break;
2881         case S_IFBLK:
2882         case S_IFCHR:
2883                 if (!may_open_dev(path))
2884                         return -EACCES;
2885                 /*FALLTHRU*/
2886         case S_IFIFO:
2887         case S_IFSOCK:
2888                 flag &= ~O_TRUNC;
2889                 break;
2890         }
2891
2892         error = inode_permission(inode, MAY_OPEN | acc_mode);
2893         if (error)
2894                 return error;
2895
2896         /*
2897          * An append-only file must be opened in append mode for writing.
2898          */
2899         if (IS_APPEND(inode)) {
2900                 if  ((flag & O_ACCMODE) != O_RDONLY && !(flag & O_APPEND))
2901                         return -EPERM;
2902                 if (flag & O_TRUNC)
2903                         return -EPERM;
2904         }
2905
2906         /* O_NOATIME can only be set by the owner or superuser */
2907         if (flag & O_NOATIME && !inode_owner_or_capable(inode))
2908                 return -EPERM;
2909
2910         return 0;
2911 }
2912
2913 static int handle_truncate(struct file *filp)
2914 {
2915         const struct path *path = &filp->f_path;
2916         struct inode *inode = path->dentry->d_inode;
2917         int error = get_write_access(inode);
2918         if (error)
2919                 return error;
2920         /*
2921          * Refuse to truncate files with mandatory locks held on them.
2922          */
2923         error = locks_verify_locked(filp);
2924         if (!error)
2925                 error = security_path_truncate(path);
2926         if (!error) {
2927                 error = do_truncate(path->dentry, 0,
2928                                     ATTR_MTIME|ATTR_CTIME|ATTR_OPEN,
2929                                     filp);
2930         }
2931         put_write_access(inode);
2932         return error;
2933 }
2934
2935 static inline int open_to_namei_flags(int flag)
2936 {
2937         if ((flag & O_ACCMODE) == 3)
2938                 flag--;
2939         return flag;
2940 }
2941
2942 static int may_o_create(const struct path *dir, struct dentry *dentry, umode_t mode)
2943 {
2944         int error = security_path_mknod(dir, dentry, mode, 0);
2945         if (error)
2946                 return error;
2947
2948         error = inode_permission(dir->dentry->d_inode, MAY_WRITE | MAY_EXEC);
2949         if (error)
2950                 return error;
2951
2952         return security_inode_create(dir->dentry->d_inode, dentry, mode);
2953 }
2954
2955 /*
2956  * Attempt to atomically look up, create and open a file from a negative
2957  * dentry.
2958  *
2959  * Returns 0 if successful.  The file will have been created and attached to
2960  * @file by the filesystem calling finish_open().
2961  *
2962  * Returns 1 if the file was looked up only or didn't need creating.  The
2963  * caller will need to perform the open themselves.  @path will have been
2964  * updated to point to the new dentry.  This may be negative.
2965  *
2966  * Returns an error code otherwise.
2967  */
2968 static int atomic_open(struct nameidata *nd, struct dentry *dentry,
2969                         struct path *path, struct file *file,
2970                         const struct open_flags *op,
2971                         int open_flag, umode_t mode,
2972                         int *opened)
2973 {
2974         struct dentry *const DENTRY_NOT_SET = (void *) -1UL;
2975         struct inode *dir =  nd->path.dentry->d_inode;
2976         int error;
2977
2978         if (!(~open_flag & (O_EXCL | O_CREAT))) /* both O_EXCL and O_CREAT */
2979                 open_flag &= ~O_TRUNC;
2980
2981         if (nd->flags & LOOKUP_DIRECTORY)
2982                 open_flag |= O_DIRECTORY;
2983
2984         file->f_path.dentry = DENTRY_NOT_SET;
2985         file->f_path.mnt = nd->path.mnt;
2986         error = dir->i_op->atomic_open(dir, dentry, file,
2987                                        open_to_namei_flags(open_flag),
2988                                        mode, opened);
2989         d_lookup_done(dentry);
2990         if (!error) {
2991                 /*
2992                  * We didn't have the inode before the open, so check open
2993                  * permission here.
2994                  */
2995                 int acc_mode = op->acc_mode;
2996                 if (*opened & FILE_CREATED) {
2997                         WARN_ON(!(open_flag & O_CREAT));
2998                         fsnotify_create(dir, dentry);
2999                         acc_mode = 0;
3000                 }
3001                 error = may_open(&file->f_path, acc_mode, open_flag);
3002                 if (WARN_ON(error > 0))
3003                         error = -EINVAL;
3004         } else if (error > 0) {
3005                 if (WARN_ON(file->f_path.dentry == DENTRY_NOT_SET)) {
3006                         error = -EIO;
3007                 } else {
3008                         if (file->f_path.dentry) {
3009                                 dput(dentry);
3010                                 dentry = file->f_path.dentry;
3011                         }
3012                         if (*opened & FILE_CREATED)
3013                                 fsnotify_create(dir, dentry);
3014                         if (unlikely(d_is_negative(dentry))) {
3015                                 error = -ENOENT;
3016                         } else {
3017                                 path->dentry = dentry;
3018                                 path->mnt = nd->path.mnt;
3019                                 return 1;
3020                         }
3021                 }
3022         }
3023         dput(dentry);
3024         return error;
3025 }
3026
3027 /*
3028  * Look up and maybe create and open the last component.
3029  *
3030  * Must be called with i_mutex held on parent.
3031  *
3032  * Returns 0 if the file was successfully atomically created (if necessary) and
3033  * opened.  In this case the file will be returned attached to @file.
3034  *
3035  * Returns 1 if the file was not completely opened at this time, though lookups
3036  * and creations will have been performed and the dentry returned in @path will
3037  * be positive upon return if O_CREAT was specified.  If O_CREAT wasn't
3038  * specified then a negative dentry may be returned.
3039  *
3040  * An error code is returned otherwise.
3041  *
3042  * FILE_CREATE will be set in @*opened if the dentry was created and will be
3043  * cleared otherwise prior to returning.
3044  */
3045 static int lookup_open(struct nameidata *nd, struct path *path,
3046                         struct file *file,
3047                         const struct open_flags *op,
3048                         bool got_write, int *opened)
3049 {
3050         struct dentry *dir = nd->path.dentry;
3051         struct inode *dir_inode = dir->d_inode;
3052         int open_flag = op->open_flag;
3053         struct dentry *dentry;
3054         int error, create_error = 0;
3055         umode_t mode = op->mode;
3056         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
3057
3058         if (unlikely(IS_DEADDIR(dir_inode)))
3059                 return -ENOENT;
3060
3061         *opened &= ~FILE_CREATED;
3062         dentry = d_lookup(dir, &nd->last);
3063         for (;;) {
3064                 if (!dentry) {
3065                         dentry = d_alloc_parallel(dir, &nd->last, &wq);
3066                         if (IS_ERR(dentry))
3067                                 return PTR_ERR(dentry);
3068                 }
3069                 if (d_in_lookup(dentry))
3070                         break;
3071
3072                 if (!(dentry->d_flags & DCACHE_OP_REVALIDATE))
3073                         break;
3074
3075                 error = d_revalidate(dentry, nd->flags);
3076                 if (likely(error > 0))
3077                         break;
3078                 if (error)
3079                         goto out_dput;
3080                 d_invalidate(dentry);
3081                 dput(dentry);
3082                 dentry = NULL;
3083         }
3084         if (dentry->d_inode) {
3085                 /* Cached positive dentry: will open in f_op->open */
3086                 goto out_no_open;
3087         }
3088
3089         /*
3090          * Checking write permission is tricky, bacuse we don't know if we are
3091          * going to actually need it: O_CREAT opens should work as long as the
3092          * file exists.  But checking existence breaks atomicity.  The trick is
3093          * to check access and if not granted clear O_CREAT from the flags.
3094          *
3095          * Another problem is returing the "right" error value (e.g. for an
3096          * O_EXCL open we want to return EEXIST not EROFS).
3097          */
3098         if (open_flag & O_CREAT) {
3099                 if (!IS_POSIXACL(dir->d_inode))
3100                         mode &= ~current_umask();
3101                 if (unlikely(!got_write)) {
3102                         create_error = -EROFS;
3103                         open_flag &= ~O_CREAT;
3104                         if (open_flag & (O_EXCL | O_TRUNC))
3105                                 goto no_open;
3106                         /* No side effects, safe to clear O_CREAT */
3107                 } else {
3108                         create_error = may_o_create(&nd->path, dentry, mode);
3109                         if (create_error) {
3110                                 open_flag &= ~O_CREAT;
3111                                 if (open_flag & O_EXCL)
3112                                         goto no_open;
3113                         }
3114                 }
3115         } else if ((open_flag & (O_TRUNC|O_WRONLY|O_RDWR)) &&
3116                    unlikely(!got_write)) {
3117                 /*
3118                  * No O_CREATE -> atomicity not a requirement -> fall
3119                  * back to lookup + open
3120                  */
3121                 goto no_open;
3122         }
3123
3124         if (dir_inode->i_op->atomic_open) {
3125                 error = atomic_open(nd, dentry, path, file, op, open_flag,
3126                                     mode, opened);
3127                 if (unlikely(error == -ENOENT) && create_error)
3128                         error = create_error;
3129                 return error;
3130         }
3131
3132 no_open:
3133         if (d_in_lookup(dentry)) {
3134                 struct dentry *res = dir_inode->i_op->lookup(dir_inode, dentry,
3135                                                              nd->flags);
3136                 d_lookup_done(dentry);
3137                 if (unlikely(res)) {
3138                         if (IS_ERR(res)) {
3139                                 error = PTR_ERR(res);
3140                                 goto out_dput;
3141                         }
3142                         dput(dentry);
3143                         dentry = res;
3144                 }
3145         }
3146
3147         /* Negative dentry, just create the file */
3148         if (!dentry->d_inode && (open_flag & O_CREAT)) {
3149                 *opened |= FILE_CREATED;
3150                 audit_inode_child(dir_inode, dentry, AUDIT_TYPE_CHILD_CREATE);
3151                 if (!dir_inode->i_op->create) {
3152                         error = -EACCES;
3153                         goto out_dput;
3154                 }
3155                 error = dir_inode->i_op->create(dir_inode, dentry, mode,
3156                                                 open_flag & O_EXCL);
3157                 if (error)
3158                         goto out_dput;
3159                 fsnotify_create(dir_inode, dentry);
3160         }
3161         if (unlikely(create_error) && !dentry->d_inode) {
3162                 error = create_error;
3163                 goto out_dput;
3164         }
3165 out_no_open:
3166         path->dentry = dentry;
3167         path->mnt = nd->path.mnt;
3168         return 1;
3169
3170 out_dput:
3171         dput(dentry);
3172         return error;
3173 }
3174
3175 /*
3176  * Handle the last step of open()
3177  */
3178 static int do_last(struct nameidata *nd,
3179                    struct file *file, const struct open_flags *op,
3180                    int *opened)
3181 {
3182         struct dentry *dir = nd->path.dentry;
3183         int open_flag = op->open_flag;
3184         bool will_truncate = (open_flag & O_TRUNC) != 0;
3185         bool got_write = false;
3186         int acc_mode = op->acc_mode;
3187         unsigned seq;
3188         struct inode *inode;
3189         struct path path;
3190         int error;
3191
3192         nd->flags &= ~LOOKUP_PARENT;
3193         nd->flags |= op->intent;
3194
3195         if (nd->last_type != LAST_NORM) {
3196                 error = handle_dots(nd, nd->last_type);
3197                 if (unlikely(error))
3198                         return error;
3199                 goto finish_open;
3200         }
3201
3202         if (!(open_flag & O_CREAT)) {
3203                 if (nd->last.name[nd->last.len])
3204                         nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
3205                 /* we _can_ be in RCU mode here */
3206                 error = lookup_fast(nd, &path, &inode, &seq);
3207                 if (likely(error > 0))
3208                         goto finish_lookup;
3209
3210                 if (error < 0)
3211                         return error;
3212
3213                 BUG_ON(nd->inode != dir->d_inode);
3214                 BUG_ON(nd->flags & LOOKUP_RCU);
3215         } else {
3216                 /* create side of things */
3217                 /*
3218                  * This will *only* deal with leaving RCU mode - LOOKUP_JUMPED
3219                  * has been cleared when we got to the last component we are
3220                  * about to look up
3221                  */
3222                 error = complete_walk(nd);
3223                 if (error)
3224                         return error;
3225
3226                 audit_inode(nd->name, dir, LOOKUP_PARENT);
3227                 /* trailing slashes? */
3228                 if (unlikely(nd->last.name[nd->last.len]))
3229                         return -EISDIR;
3230         }
3231
3232         if (open_flag & (O_CREAT | O_TRUNC | O_WRONLY | O_RDWR)) {
3233                 error = mnt_want_write(nd->path.mnt);
3234                 if (!error)
3235                         got_write = true;
3236                 /*
3237                  * do _not_ fail yet - we might not need that or fail with
3238                  * a different error; let lookup_open() decide; we'll be
3239                  * dropping this one anyway.
3240                  */
3241         }
3242         if (open_flag & O_CREAT)
3243                 inode_lock(dir->d_inode);
3244         else
3245                 inode_lock_shared(dir->d_inode);
3246         error = lookup_open(nd, &path, file, op, got_write, opened);
3247         if (open_flag & O_CREAT)
3248                 inode_unlock(dir->d_inode);
3249         else
3250                 inode_unlock_shared(dir->d_inode);
3251
3252         if (error <= 0) {
3253                 if (error)
3254                         goto out;
3255
3256                 if ((*opened & FILE_CREATED) ||
3257                     !S_ISREG(file_inode(file)->i_mode))
3258                         will_truncate = false;
3259
3260                 audit_inode(nd->name, file->f_path.dentry, 0);
3261                 goto opened;
3262         }
3263
3264         if (*opened & FILE_CREATED) {
3265                 /* Don't check for write permission, don't truncate */
3266                 open_flag &= ~O_TRUNC;
3267                 will_truncate = false;
3268                 acc_mode = 0;
3269                 path_to_nameidata(&path, nd);
3270                 goto finish_open_created;
3271         }
3272
3273         /*
3274          * If atomic_open() acquired write access it is dropped now due to
3275          * possible mount and symlink following (this might be optimized away if
3276          * necessary...)
3277          */
3278         if (got_write) {
3279                 mnt_drop_write(nd->path.mnt);
3280                 got_write = false;
3281         }
3282
3283         error = follow_managed(&path, nd);
3284         if (unlikely(error < 0))
3285                 return error;
3286
3287         if (unlikely(d_is_negative(path.dentry))) {
3288                 path_to_nameidata(&path, nd);
3289                 return -ENOENT;
3290         }
3291
3292         /*
3293          * create/update audit record if it already exists.
3294          */
3295         audit_inode(nd->name, path.dentry, 0);
3296
3297         if (unlikely((open_flag & (O_EXCL | O_CREAT)) == (O_EXCL | O_CREAT))) {
3298                 path_to_nameidata(&path, nd);
3299                 return -EEXIST;
3300         }
3301
3302         seq = 0;        /* out of RCU mode, so the value doesn't matter */
3303         inode = d_backing_inode(path.dentry);
3304 finish_lookup:
3305         error = step_into(nd, &path, 0, inode, seq);
3306         if (unlikely(error))
3307                 return error;
3308 finish_open:
3309         /* Why this, you ask?  _Now_ we might have grown LOOKUP_JUMPED... */
3310         error = complete_walk(nd);
3311         if (error)
3312                 return error;
3313         audit_inode(nd->name, nd->path.dentry, 0);
3314         error = -EISDIR;
3315         if ((open_flag & O_CREAT) && d_is_dir(nd->path.dentry))
3316                 goto out;
3317         error = -ENOTDIR;
3318         if ((nd->flags & LOOKUP_DIRECTORY) && !d_can_lookup(nd->path.dentry))
3319                 goto out;
3320         if (!d_is_reg(nd->path.dentry))
3321                 will_truncate = false;
3322
3323         if (will_truncate) {
3324                 error = mnt_want_write(nd->path.mnt);
3325                 if (error)
3326                         goto out;
3327                 got_write = true;
3328         }
3329 finish_open_created:
3330         error = may_open(&nd->path, acc_mode, open_flag);
3331         if (error)
3332                 goto out;
3333         BUG_ON(*opened & FILE_OPENED); /* once it's opened, it's opened */
3334         error = vfs_open(&nd->path, file, current_cred());
3335         if (error)
3336                 goto out;
3337         *opened |= FILE_OPENED;
3338 opened:
3339         error = open_check_o_direct(file);
3340         if (!error)
3341                 error = ima_file_check(file, op->acc_mode, *opened);
3342         if (!error && will_truncate)
3343                 error = handle_truncate(file);
3344 out:
3345         if (unlikely(error) && (*opened & FILE_OPENED))
3346                 fput(file);
3347         if (unlikely(error > 0)) {
3348                 WARN_ON(1);
3349                 error = -EINVAL;
3350         }
3351         if (got_write)
3352                 mnt_drop_write(nd->path.mnt);
3353         return error;
3354 }
3355
3356 static int do_tmpfile(struct nameidata *nd, unsigned flags,
3357                 const struct open_flags *op,
3358                 struct file *file, int *opened)
3359 {
3360         static const struct qstr name = QSTR_INIT("/", 1);
3361         struct dentry *child;
3362         struct inode *dir;
3363         struct path path;
3364         int error = path_lookupat(nd, flags | LOOKUP_DIRECTORY, &path);
3365         if (unlikely(error))
3366                 return error;
3367         error = mnt_want_write(path.mnt);
3368         if (unlikely(error))
3369                 goto out;
3370         dir = path.dentry->d_inode;
3371         /* we want directory to be writable */
3372         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
3373         if (error)
3374                 goto out2;
3375         if (!dir->i_op->tmpfile) {
3376                 error = -EOPNOTSUPP;
3377                 goto out2;
3378         }
3379         child = d_alloc(path.dentry, &name);
3380         if (unlikely(!child)) {
3381                 error = -ENOMEM;
3382                 goto out2;
3383         }
3384         dput(path.dentry);
3385         path.dentry = child;
3386         error = dir->i_op->tmpfile(dir, child, op->mode);
3387         if (error)
3388                 goto out2;
3389         audit_inode(nd->name, child, 0);
3390         /* Don't check for other permissions, the inode was just created */
3391         error = may_open(&path, 0, op->open_flag);
3392         if (error)
3393                 goto out2;
3394         file->f_path.mnt = path.mnt;
3395         error = finish_open(file, child, NULL, opened);
3396         if (error)
3397                 goto out2;
3398         error = open_check_o_direct(file);
3399         if (error) {
3400                 fput(file);
3401         } else if (!(op->open_flag & O_EXCL)) {
3402                 struct inode *inode = file_inode(file);
3403                 spin_lock(&inode->i_lock);
3404                 inode->i_state |= I_LINKABLE;
3405                 spin_unlock(&inode->i_lock);
3406         }
3407 out2:
3408         mnt_drop_write(path.mnt);
3409 out:
3410         path_put(&path);
3411         return error;
3412 }
3413
3414 static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
3415 {
3416         struct path path;
3417         int error = path_lookupat(nd, flags, &path);
3418         if (!error) {
3419                 audit_inode(nd->name, path.dentry, 0);
3420                 error = vfs_open(&path, file, current_cred());
3421                 path_put(&path);
3422         }
3423         return error;
3424 }
3425
3426 static struct file *path_openat(struct nameidata *nd,
3427                         const struct open_flags *op, unsigned flags)
3428 {
3429         const char *s;
3430         struct file *file;
3431         int opened = 0;
3432         int error;
3433
3434         file = get_empty_filp();
3435         if (IS_ERR(file))
3436                 return file;
3437
3438         file->f_flags = op->open_flag;
3439
3440         if (unlikely(file->f_flags & __O_TMPFILE)) {
3441                 error = do_tmpfile(nd, flags, op, file, &opened);
3442                 goto out2;
3443         }
3444
3445         if (unlikely(file->f_flags & O_PATH)) {
3446                 error = do_o_path(nd, flags, file);
3447                 if (!error)
3448                         opened |= FILE_OPENED;
3449                 goto out2;
3450         }
3451
3452         s = path_init(nd, flags);
3453         if (IS_ERR(s)) {
3454                 put_filp(file);
3455                 return ERR_CAST(s);
3456         }
3457         while (!(error = link_path_walk(s, nd)) &&
3458                 (error = do_last(nd, file, op, &opened)) > 0) {
3459                 nd->flags &= ~(LOOKUP_OPEN|LOOKUP_CREATE|LOOKUP_EXCL);
3460                 s = trailing_symlink(nd);
3461                 if (IS_ERR(s)) {
3462                         error = PTR_ERR(s);
3463                         break;
3464                 }
3465         }
3466         terminate_walk(nd);
3467 out2:
3468         if (!(opened & FILE_OPENED)) {
3469                 BUG_ON(!error);
3470                 put_filp(file);
3471         }
3472         if (unlikely(error)) {
3473                 if (error == -EOPENSTALE) {
3474                         if (flags & LOOKUP_RCU)
3475                                 error = -ECHILD;
3476                         else
3477                                 error = -ESTALE;
3478                 }
3479                 file = ERR_PTR(error);
3480         }
3481         return file;
3482 }
3483
3484 struct file *do_filp_open(int dfd, struct filename *pathname,
3485                 const struct open_flags *op)
3486 {
3487         struct nameidata nd;
3488         int flags = op->lookup_flags;
3489         struct file *filp;
3490
3491         set_nameidata(&nd, dfd, pathname);
3492         filp = path_openat(&nd, op, flags | LOOKUP_RCU);
3493         if (unlikely(filp == ERR_PTR(-ECHILD)))
3494                 filp = path_openat(&nd, op, flags);
3495         if (unlikely(filp == ERR_PTR(-ESTALE)))
3496                 filp = path_openat(&nd, op, flags | LOOKUP_REVAL);
3497         restore_nameidata();
3498         return filp;
3499 }
3500
3501 struct file *do_file_open_root(struct dentry *dentry, struct vfsmount *mnt,
3502                 const char *name, const struct open_flags *op)
3503 {
3504         struct nameidata nd;
3505         struct file *file;
3506         struct filename *filename;
3507         int flags = op->lookup_flags | LOOKUP_ROOT;
3508
3509         nd.root.mnt = mnt;
3510         nd.root.dentry = dentry;
3511
3512         if (d_is_symlink(dentry) && op->intent & LOOKUP_OPEN)
3513                 return ERR_PTR(-ELOOP);
3514
3515         filename = getname_kernel(name);
3516         if (IS_ERR(filename))
3517                 return ERR_CAST(filename);
3518
3519         set_nameidata(&nd, -1, filename);
3520         file = path_openat(&nd, op, flags | LOOKUP_RCU);
3521         if (unlikely(file == ERR_PTR(-ECHILD)))
3522                 file = path_openat(&nd, op, flags);
3523         if (unlikely(file == ERR_PTR(-ESTALE)))
3524                 file = path_openat(&nd, op, flags | LOOKUP_REVAL);
3525         restore_nameidata();
3526         putname(filename);
3527         return file;
3528 }
3529
3530 static struct dentry *filename_create(int dfd, struct filename *name,
3531                                 struct path *path, unsigned int lookup_flags)
3532 {
3533         struct dentry *dentry = ERR_PTR(-EEXIST);
3534         struct qstr last;
3535         int type;
3536         int err2;
3537         int error;
3538         bool is_dir = (lookup_flags & LOOKUP_DIRECTORY);
3539
3540         /*
3541          * Note that only LOOKUP_REVAL and LOOKUP_DIRECTORY matter here. Any
3542          * other flags passed in are ignored!
3543          */
3544         lookup_flags &= LOOKUP_REVAL;
3545
3546         name = filename_parentat(dfd, name, lookup_flags, path, &last, &type);
3547         if (IS_ERR(name))
3548                 return ERR_CAST(name);
3549
3550         /*
3551          * Yucky last component or no last component at all?
3552          * (foo/., foo/.., /////)
3553          */
3554         if (unlikely(type != LAST_NORM))
3555                 goto out;
3556
3557         /* don't fail immediately if it's r/o, at least try to report other errors */
3558         err2 = mnt_want_write(path->mnt);
3559         /*
3560          * Do the final lookup.
3561          */
3562         lookup_flags |= LOOKUP_CREATE | LOOKUP_EXCL;
3563         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
3564         dentry = __lookup_hash(&last, path->dentry, lookup_flags);
3565         if (IS_ERR(dentry))
3566                 goto unlock;
3567
3568         error = -EEXIST;
3569         if (d_is_positive(dentry))
3570                 goto fail;
3571
3572         /*
3573          * Special case - lookup gave negative, but... we had foo/bar/
3574          * From the vfs_mknod() POV we just have a negative dentry -
3575          * all is fine. Let's be bastards - you had / on the end, you've
3576          * been asking for (non-existent) directory. -ENOENT for you.
3577          */
3578         if (unlikely(!is_dir && last.name[last.len])) {
3579                 error = -ENOENT;
3580                 goto fail;
3581         }
3582         if (unlikely(err2)) {
3583                 error = err2;
3584                 goto fail;
3585         }
3586         putname(name);
3587         return dentry;
3588 fail:
3589         dput(dentry);
3590         dentry = ERR_PTR(error);
3591 unlock:
3592         inode_unlock(path->dentry->d_inode);
3593         if (!err2)
3594                 mnt_drop_write(path->mnt);
3595 out:
3596         path_put(path);
3597         putname(name);
3598         return dentry;
3599 }
3600
3601 struct dentry *kern_path_create(int dfd, const char *pathname,
3602                                 struct path *path, unsigned int lookup_flags)
3603 {
3604         return filename_create(dfd, getname_kernel(pathname),
3605                                 path, lookup_flags);
3606 }
3607 EXPORT_SYMBOL(kern_path_create);
3608
3609 void done_path_create(struct path *path, struct dentry *dentry)
3610 {
3611         dput(dentry);
3612         inode_unlock(path->dentry->d_inode);
3613         mnt_drop_write(path->mnt);
3614         path_put(path);
3615 }
3616 EXPORT_SYMBOL(done_path_create);
3617
3618 inline struct dentry *user_path_create(int dfd, const char __user *pathname,
3619                                 struct path *path, unsigned int lookup_flags)
3620 {
3621         return filename_create(dfd, getname(pathname), path, lookup_flags);
3622 }
3623 EXPORT_SYMBOL(user_path_create);
3624
3625 int vfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3626 {
3627         int error = may_create(dir, dentry);
3628
3629         if (error)
3630                 return error;
3631
3632         if ((S_ISCHR(mode) || S_ISBLK(mode)) && !capable(CAP_MKNOD))
3633                 return -EPERM;
3634
3635         if (!dir->i_op->mknod)
3636                 return -EPERM;
3637
3638         error = devcgroup_inode_mknod(mode, dev);
3639         if (error)
3640                 return error;
3641
3642         error = security_inode_mknod(dir, dentry, mode, dev);
3643         if (error)
3644                 return error;
3645
3646         error = dir->i_op->mknod(dir, dentry, mode, dev);
3647         if (!error)
3648                 fsnotify_create(dir, dentry);
3649         return error;
3650 }
3651 EXPORT_SYMBOL(vfs_mknod);
3652
3653 static int may_mknod(umode_t mode)
3654 {
3655         switch (mode & S_IFMT) {
3656         case S_IFREG:
3657         case S_IFCHR:
3658         case S_IFBLK:
3659         case S_IFIFO:
3660         case S_IFSOCK:
3661         case 0: /* zero mode translates to S_IFREG */
3662                 return 0;
3663         case S_IFDIR:
3664                 return -EPERM;
3665         default:
3666                 return -EINVAL;
3667         }
3668 }
3669
3670 SYSCALL_DEFINE4(mknodat, int, dfd, const char __user *, filename, umode_t, mode,
3671                 unsigned, dev)
3672 {
3673         struct dentry *dentry;
3674         struct path path;
3675         int error;
3676         unsigned int lookup_flags = 0;
3677
3678         error = may_mknod(mode);
3679         if (error)
3680                 return error;
3681 retry:
3682         dentry = user_path_create(dfd, filename, &path, lookup_flags);
3683         if (IS_ERR(dentry))
3684                 return PTR_ERR(dentry);
3685
3686         if (!IS_POSIXACL(path.dentry->d_inode))
3687                 mode &= ~current_umask();
3688         error = security_path_mknod(&path, dentry, mode, dev);
3689         if (error)
3690                 goto out;
3691         switch (mode & S_IFMT) {
3692                 case 0: case S_IFREG:
3693                         error = vfs_create(path.dentry->d_inode,dentry,mode,true);
3694                         if (!error)
3695                                 ima_post_path_mknod(dentry);
3696                         break;
3697                 case S_IFCHR: case S_IFBLK:
3698                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,
3699                                         new_decode_dev(dev));
3700                         break;
3701                 case S_IFIFO: case S_IFSOCK:
3702                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,0);
3703                         break;
3704         }
3705 out:
3706         done_path_create(&path, dentry);
3707         if (retry_estale(error, lookup_flags)) {
3708                 lookup_flags |= LOOKUP_REVAL;
3709                 goto retry;
3710         }
3711         return error;
3712 }
3713
3714 SYSCALL_DEFINE3(mknod, const char __user *, filename, umode_t, mode, unsigned, dev)
3715 {
3716         return sys_mknodat(AT_FDCWD, filename, mode, dev);
3717 }
3718
3719 int vfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3720 {
3721         int error = may_create(dir, dentry);
3722         unsigned max_links = dir->i_sb->s_max_links;
3723
3724         if (error)
3725                 return error;
3726
3727         if (!dir->i_op->mkdir)
3728                 return -EPERM;
3729
3730         mode &= (S_IRWXUGO|S_ISVTX);
3731         error = security_inode_mkdir(dir, dentry, mode);
3732         if (error)
3733                 return error;
3734
3735         if (max_links && dir->i_nlink >= max_links)
3736                 return -EMLINK;
3737
3738         error = dir->i_op->mkdir(dir, dentry, mode);
3739         if (!error)
3740                 fsnotify_mkdir(dir, dentry);
3741         return error;
3742 }
3743 EXPORT_SYMBOL(vfs_mkdir);
3744
3745 SYSCALL_DEFINE3(mkdirat, int, dfd, const char __user *, pathname, umode_t, mode)
3746 {
3747         struct dentry *dentry;
3748         struct path path;
3749         int error;
3750         unsigned int lookup_flags = LOOKUP_DIRECTORY;
3751
3752 retry:
3753         dentry = user_path_create(dfd, pathname, &path, lookup_flags);
3754         if (IS_ERR(dentry))
3755                 return PTR_ERR(dentry);
3756
3757         if (!IS_POSIXACL(path.dentry->d_inode))
3758                 mode &= ~current_umask();
3759         error = security_path_mkdir(&path, dentry, mode);
3760         if (!error)
3761                 error = vfs_mkdir(path.dentry->d_inode, dentry, mode);
3762         done_path_create(&path, dentry);
3763         if (retry_estale(error, lookup_flags)) {
3764                 lookup_flags |= LOOKUP_REVAL;
3765                 goto retry;
3766         }
3767         return error;
3768 }
3769
3770 SYSCALL_DEFINE2(mkdir, const char __user *, pathname, umode_t, mode)
3771 {
3772         return sys_mkdirat(AT_FDCWD, pathname, mode);
3773 }
3774
3775 int vfs_rmdir(struct inode *dir, struct dentry *dentry)
3776 {
3777         int error = may_delete(dir, dentry, 1);
3778
3779         if (error)
3780                 return error;
3781
3782         if (!dir->i_op->rmdir)
3783                 return -EPERM;
3784
3785         dget(dentry);
3786         inode_lock(dentry->d_inode);
3787
3788         error = -EBUSY;
3789         if (is_local_mountpoint(dentry))
3790                 goto out;
3791
3792         error = security_inode_rmdir(dir, dentry);
3793         if (error)
3794                 goto out;
3795
3796         shrink_dcache_parent(dentry);
3797         error = dir->i_op->rmdir(dir, dentry);
3798         if (error)
3799                 goto out;
3800
3801         dentry->d_inode->i_flags |= S_DEAD;
3802         dont_mount(dentry);
3803         detach_mounts(dentry);
3804
3805 out:
3806         inode_unlock(dentry->d_inode);
3807         dput(dentry);
3808         if (!error)
3809                 d_delete(dentry);
3810         return error;
3811 }
3812 EXPORT_SYMBOL(vfs_rmdir);
3813
3814 static long do_rmdir(int dfd, const char __user *pathname)
3815 {
3816         int error = 0;
3817         struct filename *name;
3818         struct dentry *dentry;
3819         struct path path;
3820         struct qstr last;
3821         int type;
3822         unsigned int lookup_flags = 0;
3823 retry:
3824         name = filename_parentat(dfd, getname(pathname), lookup_flags,
3825                                 &path, &last, &type);
3826         if (IS_ERR(name))
3827                 return PTR_ERR(name);
3828
3829         switch (type) {
3830         case LAST_DOTDOT:
3831                 error = -ENOTEMPTY;
3832                 goto exit1;
3833         case LAST_DOT:
3834                 error = -EINVAL;
3835                 goto exit1;
3836         case LAST_ROOT:
3837                 error = -EBUSY;
3838                 goto exit1;
3839         }
3840
3841         error = mnt_want_write(path.mnt);
3842         if (error)
3843                 goto exit1;
3844
3845         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3846         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3847         error = PTR_ERR(dentry);
3848         if (IS_ERR(dentry))
3849                 goto exit2;
3850         if (!dentry->d_inode) {
3851                 error = -ENOENT;
3852                 goto exit3;
3853         }
3854         error = security_path_rmdir(&path, dentry);
3855         if (error)
3856                 goto exit3;
3857         error = vfs_rmdir(path.dentry->d_inode, dentry);
3858 exit3:
3859         dput(dentry);
3860 exit2:
3861         inode_unlock(path.dentry->d_inode);
3862         mnt_drop_write(path.mnt);
3863 exit1:
3864         path_put(&path);
3865         putname(name);
3866         if (retry_estale(error, lookup_flags)) {
3867                 lookup_flags |= LOOKUP_REVAL;
3868                 goto retry;
3869         }
3870         return error;
3871 }
3872
3873 SYSCALL_DEFINE1(rmdir, const char __user *, pathname)
3874 {
3875         return do_rmdir(AT_FDCWD, pathname);
3876 }
3877
3878 /**
3879  * vfs_unlink - unlink a filesystem object
3880  * @dir:        parent directory
3881  * @dentry:     victim
3882  * @delegated_inode: returns victim inode, if the inode is delegated.
3883  *
3884  * The caller must hold dir->i_mutex.
3885  *
3886  * If vfs_unlink discovers a delegation, it will return -EWOULDBLOCK and
3887  * return a reference to the inode in delegated_inode.  The caller
3888  * should then break the delegation on that inode and retry.  Because
3889  * breaking a delegation may take a long time, the caller should drop
3890  * dir->i_mutex before doing so.
3891  *
3892  * Alternatively, a caller may pass NULL for delegated_inode.  This may
3893  * be appropriate for callers that expect the underlying filesystem not
3894  * to be NFS exported.
3895  */
3896 int vfs_unlink(struct inode *dir, struct dentry *dentry, struct inode **delegated_inode)
3897 {
3898         struct inode *target = dentry->d_inode;
3899         int error = may_delete(dir, dentry, 0);
3900
3901         if (error)
3902                 return error;
3903
3904         if (!dir->i_op->unlink)
3905                 return -EPERM;
3906
3907         inode_lock(target);
3908         if (is_local_mountpoint(dentry))
3909                 error = -EBUSY;
3910         else {
3911                 error = security_inode_unlink(dir, dentry);
3912                 if (!error) {
3913                         error = try_break_deleg(target, delegated_inode);
3914                         if (error)
3915                                 goto out;
3916                         error = dir->i_op->unlink(dir, dentry);
3917                         if (!error) {
3918                                 dont_mount(dentry);
3919                                 detach_mounts(dentry);
3920                         }
3921                 }
3922         }
3923 out:
3924         inode_unlock(target);
3925
3926         /* We don't d_delete() NFS sillyrenamed files--they still exist. */
3927         if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
3928                 fsnotify_link_count(target);
3929                 d_delete(dentry);
3930         }
3931
3932         return error;
3933 }
3934 EXPORT_SYMBOL(vfs_unlink);
3935
3936 /*
3937  * Make sure that the actual truncation of the file will occur outside its
3938  * directory's i_mutex.  Truncate can take a long time if there is a lot of
3939  * writeout happening, and we don't want to prevent access to the directory
3940  * while waiting on the I/O.
3941  */
3942 static long do_unlinkat(int dfd, const char __user *pathname)
3943 {
3944         int error;
3945         struct filename *name;
3946         struct dentry *dentry;
3947         struct path path;
3948         struct qstr last;
3949         int type;
3950         struct inode *inode = NULL;
3951         struct inode *delegated_inode = NULL;
3952         unsigned int lookup_flags = 0;
3953 retry:
3954         name = filename_parentat(dfd, getname(pathname), lookup_flags,
3955                                 &path, &last, &type);
3956         if (IS_ERR(name))
3957                 return PTR_ERR(name);
3958
3959         error = -EISDIR;
3960         if (type != LAST_NORM)
3961                 goto exit1;
3962
3963         error = mnt_want_write(path.mnt);
3964         if (error)
3965                 goto exit1;
3966 retry_deleg:
3967         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3968         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3969         error = PTR_ERR(dentry);
3970         if (!IS_ERR(dentry)) {
3971                 /* Why not before? Because we want correct error value */
3972                 if (last.name[last.len])
3973                         goto slashes;
3974                 inode = dentry->d_inode;
3975                 if (d_is_negative(dentry))
3976                         goto slashes;
3977                 ihold(inode);
3978                 error = security_path_unlink(&path, dentry);
3979                 if (error)
3980                         goto exit2;
3981                 error = vfs_unlink(path.dentry->d_inode, dentry, &delegated_inode);
3982 exit2:
3983                 dput(dentry);
3984         }
3985         inode_unlock(path.dentry->d_inode);
3986         if (inode)
3987                 iput(inode);    /* truncate the inode here */
3988         inode = NULL;
3989         if (delegated_inode) {
3990                 error = break_deleg_wait(&delegated_inode);
3991                 if (!error)
3992                         goto retry_deleg;
3993         }
3994         mnt_drop_write(path.mnt);
3995 exit1:
3996         path_put(&path);
3997         putname(name);
3998         if (retry_estale(error, lookup_flags)) {
3999                 lookup_flags |= LOOKUP_REVAL;
4000                 inode = NULL;
4001                 goto retry;
4002         }
4003         return error;
4004
4005 slashes:
4006         if (d_is_negative(dentry))
4007                 error = -ENOENT;
4008         else if (d_is_dir(dentry))
4009                 error = -EISDIR;
4010         else
4011                 error = -ENOTDIR;
4012         goto exit2;
4013 }
4014
4015 SYSCALL_DEFINE3(unlinkat, int, dfd, const char __user *, pathname, int, flag)
4016 {
4017         if ((flag & ~AT_REMOVEDIR) != 0)
4018                 return -EINVAL;
4019
4020         if (flag & AT_REMOVEDIR)
4021                 return do_rmdir(dfd, pathname);
4022
4023         return do_unlinkat(dfd, pathname);
4024 }
4025
4026 SYSCALL_DEFINE1(unlink, const char __user *, pathname)
4027 {
4028         return do_unlinkat(AT_FDCWD, pathname);
4029 }
4030
4031 int vfs_symlink(struct inode *dir, struct dentry *dentry, const char *oldname)
4032 {
4033         int error = may_create(dir, dentry);
4034
4035         if (error)
4036                 return error;
4037
4038         if (!dir->i_op->symlink)
4039                 return -EPERM;
4040
4041         error = security_inode_symlink(dir, dentry, oldname);
4042         if (error)
4043                 return error;
4044
4045         error = dir->i_op->symlink(dir, dentry, oldname);
4046         if (!error)
4047                 fsnotify_create(dir, dentry);
4048         return error;
4049 }
4050 EXPORT_SYMBOL(vfs_symlink);
4051
4052 SYSCALL_DEFINE3(symlinkat, const char __user *, oldname,
4053                 int, newdfd, const char __user *, newname)
4054 {
4055         int error;
4056         struct filename *from;
4057         struct dentry *dentry;
4058         struct path path;
4059         unsigned int lookup_flags = 0;
4060
4061         from = getname(oldname);
4062         if (IS_ERR(from))
4063                 return PTR_ERR(from);
4064 retry:
4065         dentry = user_path_create(newdfd, newname, &path, lookup_flags);
4066         error = PTR_ERR(dentry);
4067         if (IS_ERR(dentry))
4068                 goto out_putname;
4069
4070         error = security_path_symlink(&path, dentry, from->name);
4071         if (!error)
4072                 error = vfs_symlink(path.dentry->d_inode, dentry, from->name);
4073         done_path_create(&path, dentry);
4074         if (retry_estale(error, lookup_flags)) {
4075                 lookup_flags |= LOOKUP_REVAL;
4076                 goto retry;
4077         }
4078 out_putname:
4079         putname(from);
4080         return error;
4081 }
4082
4083 SYSCALL_DEFINE2(symlink, const char __user *, oldname, const char __user *, newname)
4084 {
4085         return sys_symlinkat(oldname, AT_FDCWD, newname);
4086 }
4087
4088 /**
4089  * vfs_link - create a new link
4090  * @old_dentry: object to be linked
4091  * @dir:        new parent
4092  * @new_dentry: where to create the new link
4093  * @delegated_inode: returns inode needing a delegation break
4094  *
4095  * The caller must hold dir->i_mutex
4096  *
4097  * If vfs_link discovers a delegation on the to-be-linked file in need
4098  * of breaking, it will return -EWOULDBLOCK and return a reference to the
4099  * inode in delegated_inode.  The caller should then break the delegation
4100  * and retry.  Because breaking a delegation may take a long time, the
4101  * caller should drop the i_mutex before doing so.
4102  *
4103  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4104  * be appropriate for callers that expect the underlying filesystem not
4105  * to be NFS exported.
4106  */
4107 int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_dentry, struct inode **delegated_inode)
4108 {
4109         struct inode *inode = old_dentry->d_inode;
4110         unsigned max_links = dir->i_sb->s_max_links;
4111         int error;
4112
4113         if (!inode)
4114                 return -ENOENT;
4115
4116         error = may_create(dir, new_dentry);
4117         if (error)
4118                 return error;
4119
4120         if (dir->i_sb != inode->i_sb)
4121                 return -EXDEV;
4122
4123         /*
4124          * A link to an append-only or immutable file cannot be created.
4125          */
4126         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4127                 return -EPERM;
4128         /*
4129          * Updating the link count will likely cause i_uid and i_gid to
4130          * be writen back improperly if their true value is unknown to
4131          * the vfs.
4132          */
4133         if (HAS_UNMAPPED_ID(inode))
4134                 return -EPERM;
4135         if (!dir->i_op->link)
4136                 return -EPERM;
4137         if (S_ISDIR(inode->i_mode))
4138                 return -EPERM;
4139
4140         error = security_inode_link(old_dentry, dir, new_dentry);
4141         if (error)
4142                 return error;
4143
4144         inode_lock(inode);
4145         /* Make sure we don't allow creating hardlink to an unlinked file */
4146         if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))
4147                 error =  -ENOENT;
4148         else if (max_links && inode->i_nlink >= max_links)
4149                 error = -EMLINK;
4150         else {
4151                 error = try_break_deleg(inode, delegated_inode);
4152                 if (!error)
4153                         error = dir->i_op->link(old_dentry, dir, new_dentry);
4154         }
4155
4156         if (!error && (inode->i_state & I_LINKABLE)) {
4157                 spin_lock(&inode->i_lock);
4158                 inode->i_state &= ~I_LINKABLE;
4159                 spin_unlock(&inode->i_lock);
4160         }
4161         inode_unlock(inode);
4162         if (!error)
4163                 fsnotify_link(dir, inode, new_dentry);
4164         return error;
4165 }
4166 EXPORT_SYMBOL(vfs_link);
4167
4168 /*
4169  * Hardlinks are often used in delicate situations.  We avoid
4170  * security-related surprises by not following symlinks on the
4171  * newname.  --KAB
4172  *
4173  * We don't follow them on the oldname either to be compatible
4174  * with linux 2.0, and to avoid hard-linking to directories
4175  * and other special files.  --ADM
4176  */
4177 SYSCALL_DEFINE5(linkat, int, olddfd, const char __user *, oldname,
4178                 int, newdfd, const char __user *, newname, int, flags)
4179 {
4180         struct dentry *new_dentry;
4181         struct path old_path, new_path;
4182         struct inode *delegated_inode = NULL;
4183         int how = 0;
4184         int error;
4185
4186         if ((flags & ~(AT_SYMLINK_FOLLOW | AT_EMPTY_PATH)) != 0)
4187                 return -EINVAL;
4188         /*
4189          * To use null names we require CAP_DAC_READ_SEARCH
4190          * This ensures that not everyone will be able to create
4191          * handlink using the passed filedescriptor.
4192          */
4193         if (flags & AT_EMPTY_PATH) {
4194                 if (!capable(CAP_DAC_READ_SEARCH))
4195                         return -ENOENT;
4196                 how = LOOKUP_EMPTY;
4197         }
4198
4199         if (flags & AT_SYMLINK_FOLLOW)
4200                 how |= LOOKUP_FOLLOW;
4201 retry:
4202         error = user_path_at(olddfd, oldname, how, &old_path);
4203         if (error)
4204                 return error;
4205
4206         new_dentry = user_path_create(newdfd, newname, &new_path,
4207                                         (how & LOOKUP_REVAL));
4208         error = PTR_ERR(new_dentry);
4209         if (IS_ERR(new_dentry))
4210                 goto out;
4211
4212         error = -EXDEV;
4213         if (old_path.mnt != new_path.mnt)
4214                 goto out_dput;
4215         error = may_linkat(&old_path);
4216         if (unlikely(error))
4217                 goto out_dput;
4218         error = security_path_link(old_path.dentry, &new_path, new_dentry);
4219         if (error)
4220                 goto out_dput;
4221         error = vfs_link(old_path.dentry, new_path.dentry->d_inode, new_dentry, &delegated_inode);
4222 out_dput:
4223         done_path_create(&new_path, new_dentry);
4224         if (delegated_inode) {
4225                 error = break_deleg_wait(&delegated_inode);
4226                 if (!error) {
4227                         path_put(&old_path);
4228                         goto retry;
4229                 }
4230         }
4231         if (retry_estale(error, how)) {
4232                 path_put(&old_path);
4233                 how |= LOOKUP_REVAL;
4234                 goto retry;
4235         }
4236 out:
4237         path_put(&old_path);
4238
4239         return error;
4240 }
4241
4242 SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname)
4243 {
4244         return sys_linkat(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4245 }
4246
4247 /**
4248  * vfs_rename - rename a filesystem object
4249  * @old_dir:    parent of source
4250  * @old_dentry: source
4251  * @new_dir:    parent of destination
4252  * @new_dentry: destination
4253  * @delegated_inode: returns an inode needing a delegation break
4254  * @flags:      rename flags
4255  *
4256  * The caller must hold multiple mutexes--see lock_rename()).
4257  *
4258  * If vfs_rename discovers a delegation in need of breaking at either
4259  * the source or destination, it will return -EWOULDBLOCK and return a
4260  * reference to the inode in delegated_inode.  The caller should then
4261  * break the delegation and retry.  Because breaking a delegation may
4262  * take a long time, the caller should drop all locks before doing
4263  * so.
4264  *
4265  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4266  * be appropriate for callers that expect the underlying filesystem not
4267  * to be NFS exported.
4268  *
4269  * The worst of all namespace operations - renaming directory. "Perverted"
4270  * doesn't even start to describe it. Somebody in UCB had a heck of a trip...
4271  * Problems:
4272  *      a) we can get into loop creation.
4273  *      b) race potential - two innocent renames can create a loop together.
4274  *         That's where 4.4 screws up. Current fix: serialization on
4275  *         sb->s_vfs_rename_mutex. We might be more accurate, but that's another
4276  *         story.
4277  *      c) we have to lock _four_ objects - parents and victim (if it exists),
4278  *         and source (if it is not a directory).
4279  *         And that - after we got ->i_mutex on parents (until then we don't know
4280  *         whether the target exists).  Solution: try to be smart with locking
4281  *         order for inodes.  We rely on the fact that tree topology may change
4282  *         only under ->s_vfs_rename_mutex _and_ that parent of the object we
4283  *         move will be locked.  Thus we can rank directories by the tree
4284  *         (ancestors first) and rank all non-directories after them.
4285  *         That works since everybody except rename does "lock parent, lookup,
4286  *         lock child" and rename is under ->s_vfs_rename_mutex.
4287  *         HOWEVER, it relies on the assumption that any object with ->lookup()
4288  *         has no more than 1 dentry.  If "hybrid" objects will ever appear,
4289  *         we'd better make sure that there's no link(2) for them.
4290  *      d) conversion from fhandle to dentry may come in the wrong moment - when
4291  *         we are removing the target. Solution: we will have to grab ->i_mutex
4292  *         in the fhandle_to_dentry code. [FIXME - current nfsfh.c relies on
4293  *         ->i_mutex on parents, which works but leads to some truly excessive
4294  *         locking].
4295  */
4296 int vfs_rename(struct inode *old_dir, struct dentry *old_dentry,
4297                struct inode *new_dir, struct dentry *new_dentry,
4298                struct inode **delegated_inode, unsigned int flags)
4299 {
4300         int error;
4301         bool is_dir = d_is_dir(old_dentry);
4302         const unsigned char *old_name;
4303         struct inode *source = old_dentry->d_inode;
4304         struct inode *target = new_dentry->d_inode;
4305         bool new_is_dir = false;
4306         unsigned max_links = new_dir->i_sb->s_max_links;
4307
4308         if (source == target)
4309                 return 0;
4310
4311         error = may_delete(old_dir, old_dentry, is_dir);
4312         if (error)
4313                 return error;
4314
4315         if (!target) {
4316                 error = may_create(new_dir, new_dentry);
4317         } else {
4318                 new_is_dir = d_is_dir(new_dentry);
4319
4320                 if (!(flags & RENAME_EXCHANGE))
4321                         error = may_delete(new_dir, new_dentry, is_dir);
4322                 else
4323                         error = may_delete(new_dir, new_dentry, new_is_dir);
4324         }
4325         if (error)
4326                 return error;
4327
4328         if (!old_dir->i_op->rename)
4329                 return -EPERM;
4330
4331         /*
4332          * If we are going to change the parent - check write permissions,
4333          * we'll need to flip '..'.
4334          */
4335         if (new_dir != old_dir) {
4336                 if (is_dir) {
4337                         error = inode_permission(source, MAY_WRITE);
4338                         if (error)
4339                                 return error;
4340                 }
4341                 if ((flags & RENAME_EXCHANGE) && new_is_dir) {
4342                         error = inode_permission(target, MAY_WRITE);
4343                         if (error)
4344                                 return error;
4345                 }
4346         }
4347
4348         error = security_inode_rename(old_dir, old_dentry, new_dir, new_dentry,
4349                                       flags);
4350         if (error)
4351                 return error;
4352
4353         old_name = fsnotify_oldname_init(old_dentry->d_name.name);
4354         dget(new_dentry);
4355         if (!is_dir || (flags & RENAME_EXCHANGE))
4356                 lock_two_nondirectories(source, target);
4357         else if (target)
4358                 inode_lock(target);
4359
4360         error = -EBUSY;
4361         if (is_local_mountpoint(old_dentry) || is_local_mountpoint(new_dentry))
4362                 goto out;
4363
4364         if (max_links && new_dir != old_dir) {
4365                 error = -EMLINK;
4366                 if (is_dir && !new_is_dir && new_dir->i_nlink >= max_links)
4367                         goto out;
4368                 if ((flags & RENAME_EXCHANGE) && !is_dir && new_is_dir &&
4369                     old_dir->i_nlink >= max_links)
4370                         goto out;
4371         }
4372         if (is_dir && !(flags & RENAME_EXCHANGE) && target)
4373                 shrink_dcache_parent(new_dentry);
4374         if (!is_dir) {
4375                 error = try_break_deleg(source, delegated_inode);
4376                 if (error)
4377                         goto out;
4378         }
4379         if (target && !new_is_dir) {
4380                 error = try_break_deleg(target, delegated_inode);
4381                 if (error)
4382                         goto out;
4383         }
4384         error = old_dir->i_op->rename(old_dir, old_dentry,
4385                                        new_dir, new_dentry, flags);
4386         if (error)
4387                 goto out;
4388
4389         if (!(flags & RENAME_EXCHANGE) && target) {
4390                 if (is_dir)
4391                         target->i_flags |= S_DEAD;
4392                 dont_mount(new_dentry);
4393                 detach_mounts(new_dentry);
4394         }
4395         if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE)) {
4396                 if (!(flags & RENAME_EXCHANGE))
4397                         d_move(old_dentry, new_dentry);
4398                 else
4399                         d_exchange(old_dentry, new_dentry);
4400         }
4401 out:
4402         if (!is_dir || (flags & RENAME_EXCHANGE))
4403                 unlock_two_nondirectories(source, target);
4404         else if (target)
4405                 inode_unlock(target);
4406         dput(new_dentry);
4407         if (!error) {
4408                 fsnotify_move(old_dir, new_dir, old_name, is_dir,
4409                               !(flags & RENAME_EXCHANGE) ? target : NULL, old_dentry);
4410                 if (flags & RENAME_EXCHANGE) {
4411                         fsnotify_move(new_dir, old_dir, old_dentry->d_name.name,
4412                                       new_is_dir, NULL, new_dentry);
4413                 }
4414         }
4415         fsnotify_oldname_free(old_name);
4416
4417         return error;
4418 }
4419 EXPORT_SYMBOL(vfs_rename);
4420
4421 SYSCALL_DEFINE5(renameat2, int, olddfd, const char __user *, oldname,
4422                 int, newdfd, const char __user *, newname, unsigned int, flags)
4423 {
4424         struct dentry *old_dentry, *new_dentry;
4425         struct dentry *trap;
4426         struct path old_path, new_path;
4427         struct qstr old_last, new_last;
4428         int old_type, new_type;
4429         struct inode *delegated_inode = NULL;
4430         struct filename *from;
4431         struct filename *to;
4432         unsigned int lookup_flags = 0, target_flags = LOOKUP_RENAME_TARGET;
4433         bool should_retry = false;
4434         int error;
4435
4436         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
4437                 return -EINVAL;
4438
4439         if ((flags & (RENAME_NOREPLACE | RENAME_WHITEOUT)) &&
4440             (flags & RENAME_EXCHANGE))
4441                 return -EINVAL;
4442
4443         if ((flags & RENAME_WHITEOUT) && !capable(CAP_MKNOD))
4444                 return -EPERM;
4445
4446         if (flags & RENAME_EXCHANGE)
4447                 target_flags = 0;
4448
4449 retry:
4450         from = filename_parentat(olddfd, getname(oldname), lookup_flags,
4451                                 &old_path, &old_last, &old_type);
4452         if (IS_ERR(from)) {
4453                 error = PTR_ERR(from);
4454                 goto exit;
4455         }
4456
4457         to = filename_parentat(newdfd, getname(newname), lookup_flags,
4458                                 &new_path, &new_last, &new_type);
4459         if (IS_ERR(to)) {
4460                 error = PTR_ERR(to);
4461                 goto exit1;
4462         }
4463
4464         error = -EXDEV;
4465         if (old_path.mnt != new_path.mnt)
4466                 goto exit2;
4467
4468         error = -EBUSY;
4469         if (old_type != LAST_NORM)
4470                 goto exit2;
4471
4472         if (flags & RENAME_NOREPLACE)
4473                 error = -EEXIST;
4474         if (new_type != LAST_NORM)
4475                 goto exit2;
4476
4477         error = mnt_want_write(old_path.mnt);
4478         if (error)
4479                 goto exit2;
4480
4481 retry_deleg:
4482         trap = lock_rename(new_path.dentry, old_path.dentry);
4483
4484         old_dentry = __lookup_hash(&old_last, old_path.dentry, lookup_flags);
4485         error = PTR_ERR(old_dentry);
4486         if (IS_ERR(old_dentry))
4487                 goto exit3;
4488         /* source must exist */
4489         error = -ENOENT;
4490         if (d_is_negative(old_dentry))
4491                 goto exit4;
4492         new_dentry = __lookup_hash(&new_last, new_path.dentry, lookup_flags | target_flags);
4493         error = PTR_ERR(new_dentry);
4494         if (IS_ERR(new_dentry))
4495                 goto exit4;
4496         error = -EEXIST;
4497         if ((flags & RENAME_NOREPLACE) && d_is_positive(new_dentry))
4498                 goto exit5;
4499         if (flags & RENAME_EXCHANGE) {
4500                 error = -ENOENT;
4501                 if (d_is_negative(new_dentry))
4502                         goto exit5;
4503
4504                 if (!d_is_dir(new_dentry)) {
4505                         error = -ENOTDIR;
4506                         if (new_last.name[new_last.len])
4507                                 goto exit5;
4508                 }
4509         }
4510         /* unless the source is a directory trailing slashes give -ENOTDIR */
4511         if (!d_is_dir(old_dentry)) {
4512                 error = -ENOTDIR;
4513                 if (old_last.name[old_last.len])
4514                         goto exit5;
4515                 if (!(flags & RENAME_EXCHANGE) && new_last.name[new_last.len])
4516                         goto exit5;
4517         }
4518         /* source should not be ancestor of target */
4519         error = -EINVAL;
4520         if (old_dentry == trap)
4521                 goto exit5;
4522         /* target should not be an ancestor of source */
4523         if (!(flags & RENAME_EXCHANGE))
4524                 error = -ENOTEMPTY;
4525         if (new_dentry == trap)
4526                 goto exit5;
4527
4528         error = security_path_rename(&old_path, old_dentry,
4529                                      &new_path, new_dentry, flags);
4530         if (error)
4531                 goto exit5;
4532         error = vfs_rename(old_path.dentry->d_inode, old_dentry,
4533                            new_path.dentry->d_inode, new_dentry,
4534                            &delegated_inode, flags);
4535 exit5:
4536         dput(new_dentry);
4537 exit4:
4538         dput(old_dentry);
4539 exit3:
4540         unlock_rename(new_path.dentry, old_path.dentry);
4541         if (delegated_inode) {
4542                 error = break_deleg_wait(&delegated_inode);
4543                 if (!error)
4544                         goto retry_deleg;
4545         }
4546         mnt_drop_write(old_path.mnt);
4547 exit2:
4548         if (retry_estale(error, lookup_flags))
4549                 should_retry = true;
4550         path_put(&new_path);
4551         putname(to);
4552 exit1:
4553         path_put(&old_path);
4554         putname(from);
4555         if (should_retry) {
4556                 should_retry = false;
4557                 lookup_flags |= LOOKUP_REVAL;
4558                 goto retry;
4559         }
4560 exit:
4561         return error;
4562 }
4563
4564 SYSCALL_DEFINE4(renameat, int, olddfd, const char __user *, oldname,
4565                 int, newdfd, const char __user *, newname)
4566 {
4567         return sys_renameat2(olddfd, oldname, newdfd, newname, 0);
4568 }
4569
4570 SYSCALL_DEFINE2(rename, const char __user *, oldname, const char __user *, newname)
4571 {
4572         return sys_renameat2(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4573 }
4574
4575 int vfs_whiteout(struct inode *dir, struct dentry *dentry)
4576 {
4577         int error = may_create(dir, dentry);
4578         if (error)
4579                 return error;
4580
4581         if (!dir->i_op->mknod)
4582                 return -EPERM;
4583
4584         return dir->i_op->mknod(dir, dentry,
4585                                 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
4586 }
4587 EXPORT_SYMBOL(vfs_whiteout);
4588
4589 int readlink_copy(char __user *buffer, int buflen, const char *link)
4590 {
4591         int len = PTR_ERR(link);
4592         if (IS_ERR(link))
4593                 goto out;
4594
4595         len = strlen(link);
4596         if (len > (unsigned) buflen)
4597                 len = buflen;
4598         if (copy_to_user(buffer, link, len))
4599                 len = -EFAULT;
4600 out:
4601         return len;
4602 }
4603
4604 /*
4605  * A helper for ->readlink().  This should be used *ONLY* for symlinks that
4606  * have ->get_link() not calling nd_jump_link().  Using (or not using) it
4607  * for any given inode is up to filesystem.
4608  */
4609 static int generic_readlink(struct dentry *dentry, char __user *buffer,
4610                             int buflen)
4611 {
4612         DEFINE_DELAYED_CALL(done);
4613         struct inode *inode = d_inode(dentry);
4614         const char *link = inode->i_link;
4615         int res;
4616
4617         if (!link) {
4618                 link = inode->i_op->get_link(dentry, inode, &done);
4619                 if (IS_ERR(link))
4620                         return PTR_ERR(link);
4621         }
4622         res = readlink_copy(buffer, buflen, link);
4623         do_delayed_call(&done);
4624         return res;
4625 }
4626
4627 /**
4628  * vfs_readlink - copy symlink body into userspace buffer
4629  * @dentry: dentry on which to get symbolic link
4630  * @buffer: user memory pointer
4631  * @buflen: size of buffer
4632  *
4633  * Does not touch atime.  That's up to the caller if necessary
4634  *
4635  * Does not call security hook.
4636  */
4637 int vfs_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4638 {
4639         struct inode *inode = d_inode(dentry);
4640
4641         if (unlikely(!(inode->i_opflags & IOP_DEFAULT_READLINK))) {
4642                 if (unlikely(inode->i_op->readlink))
4643                         return inode->i_op->readlink(dentry, buffer, buflen);
4644
4645                 if (!d_is_symlink(dentry))
4646                         return -EINVAL;
4647
4648                 spin_lock(&inode->i_lock);
4649                 inode->i_opflags |= IOP_DEFAULT_READLINK;
4650                 spin_unlock(&inode->i_lock);
4651         }
4652
4653         return generic_readlink(dentry, buffer, buflen);
4654 }
4655 EXPORT_SYMBOL(vfs_readlink);
4656
4657 /**
4658  * vfs_get_link - get symlink body
4659  * @dentry: dentry on which to get symbolic link
4660  * @done: caller needs to free returned data with this
4661  *
4662  * Calls security hook and i_op->get_link() on the supplied inode.
4663  *
4664  * It does not touch atime.  That's up to the caller if necessary.
4665  *
4666  * Does not work on "special" symlinks like /proc/$$/fd/N
4667  */
4668 const char *vfs_get_link(struct dentry *dentry, struct delayed_call *done)
4669 {
4670         const char *res = ERR_PTR(-EINVAL);
4671         struct inode *inode = d_inode(dentry);
4672
4673         if (d_is_symlink(dentry)) {
4674                 res = ERR_PTR(security_inode_readlink(dentry));
4675                 if (!res)
4676                         res = inode->i_op->get_link(dentry, inode, done);
4677         }
4678         return res;
4679 }
4680 EXPORT_SYMBOL(vfs_get_link);
4681
4682 /* get the link contents into pagecache */
4683 const char *page_get_link(struct dentry *dentry, struct inode *inode,
4684                           struct delayed_call *callback)
4685 {
4686         char *kaddr;
4687         struct page *page;
4688         struct address_space *mapping = inode->i_mapping;
4689
4690         if (!dentry) {
4691                 page = find_get_page(mapping, 0);
4692                 if (!page)
4693                         return ERR_PTR(-ECHILD);
4694                 if (!PageUptodate(page)) {
4695                         put_page(page);
4696                         return ERR_PTR(-ECHILD);
4697                 }
4698         } else {
4699                 page = read_mapping_page(mapping, 0, NULL);
4700                 if (IS_ERR(page))
4701                         return (char*)page;
4702         }
4703         set_delayed_call(callback, page_put_link, page);
4704         BUG_ON(mapping_gfp_mask(mapping) & __GFP_HIGHMEM);
4705         kaddr = page_address(page);
4706         nd_terminate_link(kaddr, inode->i_size, PAGE_SIZE - 1);
4707         return kaddr;
4708 }
4709
4710 EXPORT_SYMBOL(page_get_link);
4711
4712 void page_put_link(void *arg)
4713 {
4714         put_page(arg);
4715 }
4716 EXPORT_SYMBOL(page_put_link);
4717
4718 int page_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4719 {
4720         DEFINE_DELAYED_CALL(done);
4721         int res = readlink_copy(buffer, buflen,
4722                                 page_get_link(dentry, d_inode(dentry),
4723                                               &done));
4724         do_delayed_call(&done);
4725         return res;
4726 }
4727 EXPORT_SYMBOL(page_readlink);
4728
4729 /*
4730  * The nofs argument instructs pagecache_write_begin to pass AOP_FLAG_NOFS
4731  */
4732 int __page_symlink(struct inode *inode, const char *symname, int len, int nofs)
4733 {
4734         struct address_space *mapping = inode->i_mapping;
4735         struct page *page;
4736         void *fsdata;
4737         int err;
4738         unsigned int flags = AOP_FLAG_UNINTERRUPTIBLE;
4739         if (nofs)
4740                 flags |= AOP_FLAG_NOFS;
4741
4742 retry:
4743         err = pagecache_write_begin(NULL, mapping, 0, len-1,
4744                                 flags, &page, &fsdata);
4745         if (err)
4746                 goto fail;
4747
4748         memcpy(page_address(page), symname, len-1);
4749
4750         err = pagecache_write_end(NULL, mapping, 0, len-1, len-1,
4751                                                         page, fsdata);
4752         if (err < 0)
4753                 goto fail;
4754         if (err < len-1)
4755                 goto retry;
4756
4757         mark_inode_dirty(inode);
4758         return 0;
4759 fail:
4760         return err;
4761 }
4762 EXPORT_SYMBOL(__page_symlink);
4763
4764 int page_symlink(struct inode *inode, const char *symname, int len)
4765 {
4766         return __page_symlink(inode, symname, len,
4767                         !mapping_gfp_constraint(inode->i_mapping, __GFP_FS));
4768 }
4769 EXPORT_SYMBOL(page_symlink);
4770
4771 const struct inode_operations page_symlink_inode_operations = {
4772         .get_link       = page_get_link,
4773 };
4774 EXPORT_SYMBOL(page_symlink_inode_operations);