fs/namei.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  linux/fs/namei.c
   4  *
   5  *  Copyright (C) 1991, 1992  Linus Torvalds
   6  */
   7
   8 /*
   9  * Some corrections by tytso.
  10  */
  11
  12 /* [Feb 1997 T. Schoebel-Theuer] Complete rewrite of the pathname
  13  * lookup logic.
  14  */
  15 /* [Feb-Apr 2000, AV] Rewrite to the new namespace architecture.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/export.h>
  20 #include <linux/kernel.h>
  21 #include <linux/slab.h>
  22 #include <linux/fs.h>
  23 #include <linux/namei.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/fsnotify.h>
  26 #include <linux/personality.h>
  27 #include <linux/security.h>
  28 #include <linux/ima.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/mount.h>
  31 #include <linux/audit.h>
  32 #include <linux/capability.h>
  33 #include <linux/file.h>
  34 #include <linux/fcntl.h>
  35 #include <linux/device_cgroup.h>
  36 #include <linux/fs_struct.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/hash.h>
  39 #include <linux/bitops.h>
  40 #include <linux/init_task.h>
  41 #include <linux/uaccess.h>
  42
  43 #include "internal.h"
  44 #include "mount.h"
  45
  46 /* [Feb-1997 T. Schoebel-Theuer]
  47  * Fundamental changes in the pathname lookup mechanisms (namei)
  48  * were necessary because of omirr.  The reason is that omirr needs
  49  * to know the _real_ pathname, not the user-supplied one, in case
  50  * of symlinks (and also when transname replacements occur).
  51  *
  52  * The new code replaces the old recursive symlink resolution with
  53  * an iterative one (in case of non-nested symlink chains).  It does
  54  * this with calls to <fs>_follow_link().
  55  * As a side effect, dir_namei(), _namei() and follow_link() are now
  56  * replaced with a single function lookup_dentry() that can handle all
  57  * the special cases of the former code.
  58  *
  59  * With the new dcache, the pathname is stored at each inode, at least as
  60  * long as the refcount of the inode is positive.  As a side effect, the
  61  * size of the dcache depends on the inode cache and thus is dynamic.
  62  *
  63  * [29-Apr-1998 C. Scott Ananian] Updated above description of symlink
  64  * resolution to correspond with current state of the code.
  65  *
  66  * Note that the symlink resolution is not *completely* iterative.
  67  * There is still a significant amount of tail- and mid- recursion in
  68  * the algorithm.  Also, note that <fs>_readlink() is not used in
  69  * lookup_dentry(): lookup_dentry() on the result of <fs>_readlink()
  70  * may return different results than <fs>_follow_link().  Many virtual
  71  * filesystems (including /proc) exhibit this behavior.
  72  */
  73
  74 /* [24-Feb-97 T. Schoebel-Theuer] Side effects caused by new implementation:
  75  * New symlink semantics: when open() is called with flags O_CREAT | O_EXCL
  76  * and the name already exists in form of a symlink, try to create the new
  77  * name indicated by the symlink. The old code always complained that the
  78  * name already exists, due to not following the symlink even if its target
  79  * is nonexistent.  The new semantics affects also mknod() and link() when
  80  * the name is a symlink pointing to a non-existent name.
  81  *
  82  * I don't know which semantics is the right one, since I have no access
  83  * to standards. But I found by trial that HP-UX 9.0 has the full "new"
  84  * semantics implemented, while SunOS 4.1.1 and Solaris (SunOS 5.4) have the
  85  * "old" one. Personally, I think the new semantics is much more logical.
  86  * Note that "ln old new" where "new" is a symlink pointing to a non-existing
  87  * file does succeed in both HP-UX and SunOs, but not in Solaris
  88  * and in the old Linux semantics.
  89  */
  90
  91 /* [16-Dec-97 Kevin Buhr] For security reasons, we change some symlink
  92  * semantics.  See the comments in "open_namei" and "do_link" below.
  93  *
  94  * [10-Sep-98 Alan Modra] Another symlink change.
  95  */
  96
  97 /* [Feb-Apr 2000 AV] Complete rewrite. Rules for symlinks:
  98  *      inside the path - always follow.
  99  *      in the last component in creation/removal/renaming - never follow.
 100  *      if LOOKUP_FOLLOW passed - follow.
 101  *      if the pathname has trailing slashes - follow.
 102  *      otherwise - don't follow.
 103  * (applied in that order).
 104  *
 105  * [Jun 2000 AV] Inconsistent behaviour of open() in case if flags==O_CREAT
 106  * restored for 2.4. This is the last surviving part of old 4.2BSD bug.
 107  * During the 2.4 we need to fix the userland stuff depending on it -
 108  * hopefully we will be able to get rid of that wart in 2.5. So far only
 109  * XEmacs seems to be relying on it...
 110  */
 111 /*
 112  * [Sep 2001 AV] Single-semaphore locking scheme (kudos to David Holland)
 113  * implemented.  Let's see if raised priority of ->s_vfs_rename_mutex gives
 114  * any extra contention...
 115  */
 116
 117 /* In order to reduce some races, while at the same time doing additional
 118  * checking and hopefully speeding things up, we copy filenames to the
 119  * kernel data space before using them..
 120  *
 121  * POSIX.1 2.4: an empty pathname is invalid (ENOENT).
 122  * PATH_MAX includes the nul terminator --RR.
 123  */
 124
 125 #define EMBEDDED_NAME_MAX       (PATH_MAX - offsetof(struct filename, iname))
 126
 127 struct filename *
 128 getname_flags(const char __user *filename, int flags, int *empty)
 129 {
 130         struct filename *result;
 131         char *kname;
 132         int len;
 133
 134         result = audit_reusename(filename);
 135         if (result)
 136                 return result;
 137
 138         result = __getname();
 139         if (unlikely(!result))
 140                 return ERR_PTR(-ENOMEM);
 141
 142         /*
 143          * First, try to embed the struct filename inside the names_cache
 144          * allocation
 145          */
 146         kname = (char *)result->iname;
 147         result->name = kname;
 148
 149         len = strncpy_from_user(kname, filename, EMBEDDED_NAME_MAX);
 150         if (unlikely(len < 0)) {
 151                 __putname(result);
 152                 return ERR_PTR(len);
 153         }
 154
 155         /*
 156          * Uh-oh. We have a name that's approaching PATH_MAX. Allocate a
 157          * separate struct filename so we can dedicate the entire
 158          * names_cache allocation for the pathname, and re-do the copy from
 159          * userland.
 160          */
 161         if (unlikely(len == EMBEDDED_NAME_MAX)) {
 162                 const size_t size = offsetof(struct filename, iname[1]);
 163                 kname = (char *)result;
 164
 165                 /*
 166                  * size is chosen that way we to guarantee that
 167                  * result->iname[0] is within the same object and that
 168                  * kname can't be equal to result->iname, no matter what.
 169                  */
 170                 result = kzalloc(size, GFP_KERNEL);
 171                 if (unlikely(!result)) {
 172                         __putname(kname);
 173                         return ERR_PTR(-ENOMEM);
 174                 }
 175                 result->name = kname;
 176                 len = strncpy_from_user(kname, filename, PATH_MAX);
 177                 if (unlikely(len < 0)) {
 178                         __putname(kname);
 179                         kfree(result);
 180                         return ERR_PTR(len);
 181                 }
 182                 if (unlikely(len == PATH_MAX)) {
 183                         __putname(kname);
 184                         kfree(result);
 185                         return ERR_PTR(-ENAMETOOLONG);
 186                 }
 187         }
 188
 189         result->refcnt = 1;
 190         /* The empty path is special. */
 191         if (unlikely(!len)) {
 192                 if (empty)
 193                         *empty = 1;
 194                 if (!(flags & LOOKUP_EMPTY)) {
 195                         putname(result);
 196                         return ERR_PTR(-ENOENT);
 197                 }
 198         }
 199
 200         result->uptr = filename;
 201         result->aname = NULL;
 202         audit_getname(result);
 203         return result;
 204 }
 205
 206 struct filename *
 207 getname(const char __user * filename)
 208 {
 209         return getname_flags(filename, 0, NULL);
 210 }
 211
 212 struct filename *
 213 getname_kernel(const char * filename)
 214 {
 215         struct filename *result;
 216         int len = strlen(filename) + 1;
 217
 218         result = __getname();
 219         if (unlikely(!result))
 220                 return ERR_PTR(-ENOMEM);
 221
 222         if (len <= EMBEDDED_NAME_MAX) {
 223                 result->name = (char *)result->iname;
 224         } else if (len <= PATH_MAX) {
 225                 struct filename *tmp;
 226
 227                 tmp = kmalloc(sizeof(*tmp), GFP_KERNEL);
 228                 if (unlikely(!tmp)) {
 229                         __putname(result);
 230                         return ERR_PTR(-ENOMEM);
 231                 }
 232                 tmp->name = (char *)result;
 233                 result = tmp;
 234         } else {
 235                 __putname(result);
 236                 return ERR_PTR(-ENAMETOOLONG);
 237         }
 238         memcpy((char *)result->name, filename, len);
 239         result->uptr = NULL;
 240         result->aname = NULL;
 241         result->refcnt = 1;
 242         audit_getname(result);
 243
 244         return result;
 245 }
 246
 247 void putname(struct filename *name)
 248 {
 249         BUG_ON(name->refcnt <= 0);
 250
 251         if (--name->refcnt > 0)
 252                 return;
 253
 254         if (name->name != name->iname) {
 255                 __putname(name->name);
 256                 kfree(name);
 257         } else
 258                 __putname(name);
 259 }
 260
 261 static int check_acl(struct inode *inode, int mask)
 262 {
 263 #ifdef CONFIG_FS_POSIX_ACL
 264         struct posix_acl *acl;
 265
 266         if (mask & MAY_NOT_BLOCK) {
 267                 acl = get_cached_acl_rcu(inode, ACL_TYPE_ACCESS);
 268                 if (!acl)
 269                         return -EAGAIN;
 270                 /* no ->get_acl() calls in RCU mode... */
 271                 if (is_uncached_acl(acl))
 272                         return -ECHILD;
 273                 return posix_acl_permission(inode, acl, mask & ~MAY_NOT_BLOCK);
 274         }
 275
 276         acl = get_acl(inode, ACL_TYPE_ACCESS);
 277         if (IS_ERR(acl))
 278                 return PTR_ERR(acl);
 279         if (acl) {
 280                 int error = posix_acl_permission(inode, acl, mask);
 281                 posix_acl_release(acl);
 282                 return error;
 283         }
 284 #endif
 285
 286         return -EAGAIN;
 287 }
 288
 289 /*
 290  * This does the basic permission checking
 291  */
 292 static int acl_permission_check(struct inode *inode, int mask)
 293 {
 294         unsigned int mode = inode->i_mode;
 295
 296         if (likely(uid_eq(current_fsuid(), inode->i_uid)))
 297                 mode >>= 6;
 298         else {
 299                 if (IS_POSIXACL(inode) && (mode & S_IRWXG)) {
 300                         int error = check_acl(inode, mask);
 301                         if (error != -EAGAIN)
 302                                 return error;
 303                 }
 304
 305                 if (in_group_p(inode->i_gid))
 306                         mode >>= 3;
 307         }
 308
 309         /*
 310          * If the DACs are ok we don't need any capability check.
 311          */
 312         if ((mask & ~mode & (MAY_READ | MAY_WRITE | MAY_EXEC)) == 0)
 313                 return 0;
 314         return -EACCES;
 315 }
 316
 317 /**
 318  * generic_permission -  check for access rights on a Posix-like filesystem
 319  * @inode:      inode to check access rights for
 320  * @mask:       right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC, ...)
 321  *
 322  * Used to check for read/write/execute permissions on a file.
 323  * We use "fsuid" for this, letting us set arbitrary permissions
 324  * for filesystem access without changing the "normal" uids which
 325  * are used for other things.
 326  *
 327  * generic_permission is rcu-walk aware. It returns -ECHILD in case an rcu-walk
 328  * request cannot be satisfied (eg. requires blocking or too much complexity).
 329  * It would then be called again in ref-walk mode.
 330  */
 331 int generic_permission(struct inode *inode, int mask)
 332 {
 333         int ret;
 334
 335         /*
 336          * Do the basic permission checks.
 337          */
 338         ret = acl_permission_check(inode, mask);
 339         if (ret != -EACCES)
 340                 return ret;
 341
 342         if (S_ISDIR(inode->i_mode)) {
 343                 /* DACs are overridable for directories */
 344                 if (!(mask & MAY_WRITE))
 345                         if (capable_wrt_inode_uidgid(inode,
 346                                                      CAP_DAC_READ_SEARCH))
 347                                 return 0;
 348                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
 349                         return 0;
 350                 return -EACCES;
 351         }
 352
 353         /*
 354          * Searching includes executable on directories, else just read.
 355          */
 356         mask &= MAY_READ | MAY_WRITE | MAY_EXEC;
 357         if (mask == MAY_READ)
 358                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_READ_SEARCH))
 359                         return 0;
 360         /*
 361          * Read/write DACs are always overridable.
 362          * Executable DACs are overridable when there is
 363          * at least one exec bit set.
 364          */
 365         if (!(mask & MAY_EXEC) || (inode->i_mode & S_IXUGO))
 366                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
 367                         return 0;
 368
 369         return -EACCES;
 370 }
 371 EXPORT_SYMBOL(generic_permission);
 372
 373 /*
 374  * We _really_ want to just do "generic_permission()" without
 375  * even looking at the inode->i_op values. So we keep a cache
 376  * flag in inode->i_opflags, that says "this has not special
 377  * permission function, use the fast case".
 378  */
 379 static inline int do_inode_permission(struct inode *inode, int mask)
 380 {
 381         if (unlikely(!(inode->i_opflags & IOP_FASTPERM))) {
 382                 if (likely(inode->i_op->permission))
 383                         return inode->i_op->permission(inode, mask);
 384
 385                 /* This gets set once for the inode lifetime */
 386                 spin_lock(&inode->i_lock);
 387                 inode->i_opflags |= IOP_FASTPERM;
 388                 spin_unlock(&inode->i_lock);
 389         }
 390         return generic_permission(inode, mask);
 391 }
 392
 393 /**
 394  * sb_permission - Check superblock-level permissions
 395  * @sb: Superblock of inode to check permission on
 396  * @inode: Inode to check permission on
 397  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 398  *
 399  * Separate out file-system wide checks from inode-specific permission checks.
 400  */
 401 static int sb_permission(struct super_block *sb, struct inode *inode, int mask)
 402 {
 403         if (unlikely(mask & MAY_WRITE)) {
 404                 umode_t mode = inode->i_mode;
 405
 406                 /* Nobody gets write access to a read-only fs. */
 407                 if (sb_rdonly(sb) && (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
 408                         return -EROFS;
 409         }
 410         return 0;
 411 }
 412
 413 /**
 414  * inode_permission - Check for access rights to a given inode
 415  * @inode: Inode to check permission on
 416  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 417  *
 418  * Check for read/write/execute permissions on an inode.  We use fs[ug]id for
 419  * this, letting us set arbitrary permissions for filesystem access without
 420  * changing the "normal" UIDs which are used for other things.
 421  *
 422  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
 423  */
 424 int inode_permission(struct inode *inode, int mask)
 425 {
 426         int retval;
 427
 428         retval = sb_permission(inode->i_sb, inode, mask);
 429         if (retval)
 430                 return retval;
 431
 432         if (unlikely(mask & MAY_WRITE)) {
 433                 /*
 434                  * Nobody gets write access to an immutable file.
 435                  */
 436                 if (IS_IMMUTABLE(inode))
 437                         return -EPERM;
 438
 439                 /*
 440                  * Updating mtime will likely cause i_uid and i_gid to be
 441                  * written back improperly if their true value is unknown
 442                  * to the vfs.
 443                  */
 444                 if (HAS_UNMAPPED_ID(inode))
 445                         return -EACCES;
 446         }
 447
 448         retval = do_inode_permission(inode, mask);
 449         if (retval)
 450                 return retval;
 451
 452         retval = devcgroup_inode_permission(inode, mask);
 453         if (retval)
 454                 return retval;
 455
 456         return security_inode_permission(inode, mask);
 457 }
 458 EXPORT_SYMBOL(inode_permission);
 459
 460 /**
 461  * path_get - get a reference to a path
 462  * @path: path to get the reference to
 463  *
 464  * Given a path increment the reference count to the dentry and the vfsmount.
 465  */
 466 void path_get(const struct path *path)
 467 {
 468         mntget(path->mnt);
 469         dget(path->dentry);
 470 }
 471 EXPORT_SYMBOL(path_get);
 472
 473 /**
 474  * path_put - put a reference to a path
 475  * @path: path to put the reference to
 476  *
 477  * Given a path decrement the reference count to the dentry and the vfsmount.
 478  */
 479 void path_put(const struct path *path)
 480 {
 481         dput(path->dentry);
 482         mntput(path->mnt);
 483 }
 484 EXPORT_SYMBOL(path_put);
 485
 486 #define EMBEDDED_LEVELS 2
 487 struct nameidata {
 488         struct path     path;
 489         struct qstr     last;
 490         struct path     root;
 491         struct inode    *inode; /* path.dentry.d_inode */
 492         unsigned int    flags;
 493         unsigned        seq, m_seq;
 494         int             last_type;
 495         unsigned        depth;
 496         int             total_link_count;
 497         struct saved {
 498                 struct path link;
 499                 struct delayed_call done;
 500                 const char *name;
 501                 unsigned seq;
 502         } *stack, internal[EMBEDDED_LEVELS];
 503         struct filename *name;
 504         struct nameidata *saved;
 505         struct inode    *link_inode;
 506         unsigned        root_seq;
 507         int             dfd;
 508 } __randomize_layout;
 509
 510 static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
 511 {
 512         struct nameidata *old = current->nameidata;
 513         p->stack = p->internal;
 514         p->dfd = dfd;
 515         p->name = name;
 516         p->total_link_count = old ? old->total_link_count : 0;
 517         p->saved = old;
 518         current->nameidata = p;
 519 }
 520
 521 static void restore_nameidata(void)
 522 {
 523         struct nameidata *now = current->nameidata, *old = now->saved;
 524
 525         current->nameidata = old;
 526         if (old)
 527                 old->total_link_count = now->total_link_count;
 528         if (now->stack != now->internal)
 529                 kfree(now->stack);
 530 }
 531
 532 static int __nd_alloc_stack(struct nameidata *nd)
 533 {
 534         struct saved *p;
 535
 536         if (nd->flags & LOOKUP_RCU) {
 537                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
 538                                   GFP_ATOMIC);
 539                 if (unlikely(!p))
 540                         return -ECHILD;
 541         } else {
 542                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
 543                                   GFP_KERNEL);
 544                 if (unlikely(!p))
 545                         return -ENOMEM;
 546         }
 547         memcpy(p, nd->internal, sizeof(nd->internal));
 548         nd->stack = p;
 549         return 0;
 550 }
 551
 552 /**
 553  * path_connected - Verify that a path->dentry is below path->mnt.mnt_root
 554  * @path: nameidate to verify
 555  *
 556  * Rename can sometimes move a file or directory outside of a bind
 557  * mount, path_connected allows those cases to be detected.
 558  */
 559 static bool path_connected(const struct path *path)
 560 {
 561         struct vfsmount *mnt = path->mnt;
 562         struct super_block *sb = mnt->mnt_sb;
 563
 564         /* Bind mounts and multi-root filesystems can have disconnected paths */
 565         if (!(sb->s_iflags & SB_I_MULTIROOT) && (mnt->mnt_root == sb->s_root))
 566                 return true;
 567
 568         return is_subdir(path->dentry, mnt->mnt_root);
 569 }
 570
 571 static inline int nd_alloc_stack(struct nameidata *nd)
 572 {
 573         if (likely(nd->depth != EMBEDDED_LEVELS))
 574                 return 0;
 575         if (likely(nd->stack != nd->internal))
 576                 return 0;
 577         return __nd_alloc_stack(nd);
 578 }
 579
 580 static void drop_links(struct nameidata *nd)
 581 {
 582         int i = nd->depth;
 583         while (i--) {
 584                 struct saved *last = nd->stack + i;
 585                 do_delayed_call(&last->done);
 586                 clear_delayed_call(&last->done);
 587         }
 588 }
 589
 590 static void terminate_walk(struct nameidata *nd)
 591 {
 592         drop_links(nd);
 593         if (!(nd->flags & LOOKUP_RCU)) {
 594                 int i;
 595                 path_put(&nd->path);
 596                 for (i = 0; i < nd->depth; i++)
 597                         path_put(&nd->stack[i].link);
 598                 if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 599                         path_put(&nd->root);
 600                         nd->root.mnt = NULL;
 601                 }
 602         } else {
 603                 nd->flags &= ~LOOKUP_RCU;
 604                 if (!(nd->flags & LOOKUP_ROOT))
 605                         nd->root.mnt = NULL;
 606                 rcu_read_unlock();
 607         }
 608         nd->depth = 0;
 609 }
 610
 611 /* path_put is needed afterwards regardless of success or failure */
 612 static bool legitimize_path(struct nameidata *nd,
 613                             struct path *path, unsigned seq)
 614 {
 615         int res = __legitimize_mnt(path->mnt, nd->m_seq);
 616         if (unlikely(res)) {
 617                 if (res > 0)
 618                         path->mnt = NULL;
 619                 path->dentry = NULL;
 620                 return false;
 621         }
 622         if (unlikely(!lockref_get_not_dead(&path->dentry->d_lockref))) {
 623                 path->dentry = NULL;
 624                 return false;
 625         }
 626         return !read_seqcount_retry(&path->dentry->d_seq, seq);
 627 }
 628
 629 static bool legitimize_links(struct nameidata *nd)
 630 {
 631         int i;
 632         for (i = 0; i < nd->depth; i++) {
 633                 struct saved *last = nd->stack + i;
 634                 if (unlikely(!legitimize_path(nd, &last->link, last->seq))) {
 635                         drop_links(nd);
 636                         nd->depth = i + 1;
 637                         return false;
 638                 }
 639         }
 640         return true;
 641 }
 642
 643 /*
 644  * Path walking has 2 modes, rcu-walk and ref-walk (see
 645  * Documentation/filesystems/path-lookup.txt).  In situations when we can't
 646  * continue in RCU mode, we attempt to drop out of rcu-walk mode and grab
 647  * normal reference counts on dentries and vfsmounts to transition to ref-walk
 648  * mode.  Refcounts are grabbed at the last known good point before rcu-walk
 649  * got stuck, so ref-walk may continue from there. If this is not successful
 650  * (eg. a seqcount has changed), then failure is returned and it's up to caller
 651  * to restart the path walk from the beginning in ref-walk mode.
 652  */
 653
 654 /**
 655  * unlazy_walk - try to switch to ref-walk mode.
 656  * @nd: nameidata pathwalk data
 657  * Returns: 0 on success, -ECHILD on failure
 658  *
 659  * unlazy_walk attempts to legitimize the current nd->path and nd->root
 660  * for ref-walk mode.
 661  * Must be called from rcu-walk context.
 662  * Nothing should touch nameidata between unlazy_walk() failure and
 663  * terminate_walk().
 664  */
 665 static int unlazy_walk(struct nameidata *nd)
 666 {
 667         struct dentry *parent = nd->path.dentry;
 668
 669         BUG_ON(!(nd->flags & LOOKUP_RCU));
 670
 671         nd->flags &= ~LOOKUP_RCU;
 672         if (unlikely(!legitimize_links(nd)))
 673                 goto out2;
 674         if (unlikely(!legitimize_path(nd, &nd->path, nd->seq)))
 675                 goto out1;
 676         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 677                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq)))
 678                         goto out;
 679         }
 680         rcu_read_unlock();
 681         BUG_ON(nd->inode != parent->d_inode);
 682         return 0;
 683
 684 out2:
 685         nd->path.mnt = NULL;
 686         nd->path.dentry = NULL;
 687 out1:
 688         if (!(nd->flags & LOOKUP_ROOT))
 689                 nd->root.mnt = NULL;
 690 out:
 691         rcu_read_unlock();
 692         return -ECHILD;
 693 }
 694
 695 /**
 696  * unlazy_child - try to switch to ref-walk mode.
 697  * @nd: nameidata pathwalk data
 698  * @dentry: child of nd->path.dentry
 699  * @seq: seq number to check dentry against
 700  * Returns: 0 on success, -ECHILD on failure
 701  *
 702  * unlazy_child attempts to legitimize the current nd->path, nd->root and dentry
 703  * for ref-walk mode.  @dentry must be a path found by a do_lookup call on
 704  * @nd.  Must be called from rcu-walk context.
 705  * Nothing should touch nameidata between unlazy_child() failure and
 706  * terminate_walk().
 707  */
 708 static int unlazy_child(struct nameidata *nd, struct dentry *dentry, unsigned seq)
 709 {
 710         BUG_ON(!(nd->flags & LOOKUP_RCU));
 711
 712         nd->flags &= ~LOOKUP_RCU;
 713         if (unlikely(!legitimize_links(nd)))
 714                 goto out2;
 715         if (unlikely(!legitimize_mnt(nd->path.mnt, nd->m_seq)))
 716                 goto out2;
 717         if (unlikely(!lockref_get_not_dead(&nd->path.dentry->d_lockref)))
 718                 goto out1;
 719
 720         /*
 721          * We need to move both the parent and the dentry from the RCU domain
 722          * to be properly refcounted. And the sequence number in the dentry
 723          * validates *both* dentry counters, since we checked the sequence
 724          * number of the parent after we got the child sequence number. So we
 725          * know the parent must still be valid if the child sequence number is
 726          */
 727         if (unlikely(!lockref_get_not_dead(&dentry->d_lockref)))
 728                 goto out;
 729         if (unlikely(read_seqcount_retry(&dentry->d_seq, seq))) {
 730                 rcu_read_unlock();
 731                 dput(dentry);
 732                 goto drop_root_mnt;
 733         }
 734         /*
 735          * Sequence counts matched. Now make sure that the root is
 736          * still valid and get it if required.
 737          */
 738         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 739                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq))) {
 740                         rcu_read_unlock();
 741                         dput(dentry);
 742                         return -ECHILD;
 743                 }
 744         }
 745
 746         rcu_read_unlock();
 747         return 0;
 748
 749 out2:
 750         nd->path.mnt = NULL;
 751 out1:
 752         nd->path.dentry = NULL;
 753 out:
 754         rcu_read_unlock();
 755 drop_root_mnt:
 756         if (!(nd->flags & LOOKUP_ROOT))
 757                 nd->root.mnt = NULL;
 758         return -ECHILD;
 759 }
 760
 761 static inline int d_revalidate(struct dentry *dentry, unsigned int flags)
 762 {
 763         if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
 764                 return dentry->d_op->d_revalidate(dentry, flags);
 765         else
 766                 return 1;
 767 }
 768
 769 /**
 770  * complete_walk - successful completion of path walk
 771  * @nd:  pointer nameidata
 772  *
 773  * If we had been in RCU mode, drop out of it and legitimize nd->path.
 774  * Revalidate the final result, unless we'd already done that during
 775  * the path walk or the filesystem doesn't ask for it.  Return 0 on
 776  * success, -error on failure.  In case of failure caller does not
 777  * need to drop nd->path.
 778  */
 779 static int complete_walk(struct nameidata *nd)
 780 {
 781         struct dentry *dentry = nd->path.dentry;
 782         int status;
 783
 784         if (nd->flags & LOOKUP_RCU) {
 785                 if (!(nd->flags & LOOKUP_ROOT))
 786                         nd->root.mnt = NULL;
 787                 if (unlikely(unlazy_walk(nd)))
 788                         return -ECHILD;
 789         }
 790
 791         if (likely(!(nd->flags & LOOKUP_JUMPED)))
 792                 return 0;
 793
 794         if (likely(!(dentry->d_flags & DCACHE_OP_WEAK_REVALIDATE)))
 795                 return 0;
 796
 797         status = dentry->d_op->d_weak_revalidate(dentry, nd->flags);
 798         if (status > 0)
 799                 return 0;
 800
 801         if (!status)
 802                 status = -ESTALE;
 803
 804         return status;
 805 }
 806
 807 static void set_root(struct nameidata *nd)
 808 {
 809         struct fs_struct *fs = current->fs;
 810
 811         if (nd->flags & LOOKUP_RCU) {
 812                 unsigned seq;
 813
 814                 do {
 815                         seq = read_seqcount_begin(&fs->seq);
 816                         nd->root = fs->root;
 817                         nd->root_seq = __read_seqcount_begin(&nd->root.dentry->d_seq);
 818                 } while (read_seqcount_retry(&fs->seq, seq));
 819         } else {
 820                 get_fs_root(fs, &nd->root);
 821         }
 822 }
 823
 824 static void path_put_conditional(struct path *path, struct nameidata *nd)
 825 {
 826         dput(path->dentry);
 827         if (path->mnt != nd->path.mnt)
 828                 mntput(path->mnt);
 829 }
 830
 831 static inline void path_to_nameidata(const struct path *path,
 832                                         struct nameidata *nd)
 833 {
 834         if (!(nd->flags & LOOKUP_RCU)) {
 835                 dput(nd->path.dentry);
 836                 if (nd->path.mnt != path->mnt)
 837                         mntput(nd->path.mnt);
 838         }
 839         nd->path.mnt = path->mnt;
 840         nd->path.dentry = path->dentry;
 841 }
 842
 843 static int nd_jump_root(struct nameidata *nd)
 844 {
 845         if (nd->flags & LOOKUP_RCU) {
 846                 struct dentry *d;
 847                 nd->path = nd->root;
 848                 d = nd->path.dentry;
 849                 nd->inode = d->d_inode;
 850                 nd->seq = nd->root_seq;
 851                 if (unlikely(read_seqcount_retry(&d->d_seq, nd->seq)))
 852                         return -ECHILD;
 853         } else {
 854                 path_put(&nd->path);
 855                 nd->path = nd->root;
 856                 path_get(&nd->path);
 857                 nd->inode = nd->path.dentry->d_inode;
 858         }
 859         nd->flags |= LOOKUP_JUMPED;
 860         return 0;
 861 }
 862
 863 /*
 864  * Helper to directly jump to a known parsed path from ->get_link,
 865  * caller must have taken a reference to path beforehand.
 866  */
 867 void nd_jump_link(struct path *path)
 868 {
 869         struct nameidata *nd = current->nameidata;
 870         path_put(&nd->path);
 871
 872         nd->path = *path;
 873         nd->inode = nd->path.dentry->d_inode;
 874         nd->flags |= LOOKUP_JUMPED;
 875 }
 876
 877 static inline void put_link(struct nameidata *nd)
 878 {
 879         struct saved *last = nd->stack + --nd->depth;
 880         do_delayed_call(&last->done);
 881         if (!(nd->flags & LOOKUP_RCU))
 882                 path_put(&last->link);
 883 }
 884
 885 int sysctl_protected_symlinks __read_mostly = 0;
 886 int sysctl_protected_hardlinks __read_mostly = 0;
 887
 888 /**
 889  * may_follow_link - Check symlink following for unsafe situations
 890  * @nd: nameidata pathwalk data
 891  *
 892  * In the case of the sysctl_protected_symlinks sysctl being enabled,
 893  * CAP_DAC_OVERRIDE needs to be specifically ignored if the symlink is
 894  * in a sticky world-writable directory. This is to protect privileged
 895  * processes from failing races against path names that may change out
 896  * from under them by way of other users creating malicious symlinks.
 897  * It will permit symlinks to be followed only when outside a sticky
 898  * world-writable directory, or when the uid of the symlink and follower
 899  * match, or when the directory owner matches the symlink's owner.
 900  *
 901  * Returns 0 if following the symlink is allowed, -ve on error.
 902  */
 903 static inline int may_follow_link(struct nameidata *nd)
 904 {
 905         const struct inode *inode;
 906         const struct inode *parent;
 907         kuid_t puid;
 908
 909         if (!sysctl_protected_symlinks)
 910                 return 0;
 911
 912         /* Allowed if owner and follower match. */
 913         inode = nd->link_inode;
 914         if (uid_eq(current_cred()->fsuid, inode->i_uid))
 915                 return 0;
 916
 917         /* Allowed if parent directory not sticky and world-writable. */
 918         parent = nd->inode;
 919         if ((parent->i_mode & (S_ISVTX|S_IWOTH)) != (S_ISVTX|S_IWOTH))
 920                 return 0;
 921
 922         /* Allowed if parent directory and link owner match. */
 923         puid = parent->i_uid;
 924         if (uid_valid(puid) && uid_eq(puid, inode->i_uid))
 925                 return 0;
 926
 927         if (nd->flags & LOOKUP_RCU)
 928                 return -ECHILD;
 929
 930         audit_log_link_denied("follow_link", &nd->stack[0].link);
 931         return -EACCES;
 932 }
 933
 934 /**
 935  * safe_hardlink_source - Check for safe hardlink conditions
 936  * @inode: the source inode to hardlink from
 937  *
 938  * Return false if at least one of the following conditions:
 939  *    - inode is not a regular file
 940  *    - inode is setuid
 941  *    - inode is setgid and group-exec
 942  *    - access failure for read and write
 943  *
 944  * Otherwise returns true.
 945  */
 946 static bool safe_hardlink_source(struct inode *inode)
 947 {
 948         umode_t mode = inode->i_mode;
 949
 950         /* Special files should not get pinned to the filesystem. */
 951         if (!S_ISREG(mode))
 952                 return false;
 953
 954         /* Setuid files should not get pinned to the filesystem. */
 955         if (mode & S_ISUID)
 956                 return false;
 957
 958         /* Executable setgid files should not get pinned to the filesystem. */
 959         if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP))
 960                 return false;
 961
 962         /* Hardlinking to unreadable or unwritable sources is dangerous. */
 963         if (inode_permission(inode, MAY_READ | MAY_WRITE))
 964                 return false;
 965
 966         return true;
 967 }
 968
 969 /**
 970  * may_linkat - Check permissions for creating a hardlink
 971  * @link: the source to hardlink from
 972  *
 973  * Block hardlink when all of:
 974  *  - sysctl_protected_hardlinks enabled
 975  *  - fsuid does not match inode
 976  *  - hardlink source is unsafe (see safe_hardlink_source() above)
 977  *  - not CAP_FOWNER in a namespace with the inode owner uid mapped
 978  *
 979  * Returns 0 if successful, -ve on error.
 980  */
 981 static int may_linkat(struct path *link)
 982 {
 983         struct inode *inode;
 984
 985         if (!sysctl_protected_hardlinks)
 986                 return 0;
 987
 988         inode = link->dentry->d_inode;
 989
 990         /* Source inode owner (or CAP_FOWNER) can hardlink all they like,
 991          * otherwise, it must be a safe source.
 992          */
 993         if (safe_hardlink_source(inode) || inode_owner_or_capable(inode))
 994                 return 0;
 995
 996         audit_log_link_denied("linkat", link);
 997         return -EPERM;
 998 }
 999
1000 static __always_inline
1001 const char *get_link(struct nameidata *nd)
1002 {
1003         struct saved *last = nd->stack + nd->depth - 1;
1004         struct dentry *dentry = last->link.dentry;
1005         struct inode *inode = nd->link_inode;
1006         int error;
1007         const char *res;
1008
1009         if (!(nd->flags & LOOKUP_RCU)) {
1010                 touch_atime(&last->link);
1011                 cond_resched();
1012         } else if (atime_needs_update_rcu(&last->link, inode)) {
1013                 if (unlikely(unlazy_walk(nd)))
1014                         return ERR_PTR(-ECHILD);
1015                 touch_atime(&last->link);
1016         }
1017
1018         error = security_inode_follow_link(dentry, inode,
1019                                            nd->flags & LOOKUP_RCU);
1020         if (unlikely(error))
1021                 return ERR_PTR(error);
1022
1023         nd->last_type = LAST_BIND;
1024         res = inode->i_link;
1025         if (!res) {
1026                 const char * (*get)(struct dentry *, struct inode *,
1027                                 struct delayed_call *);
1028                 get = inode->i_op->get_link;
1029                 if (nd->flags & LOOKUP_RCU) {
1030                         res = get(NULL, inode, &last->done);
1031                         if (res == ERR_PTR(-ECHILD)) {
1032                                 if (unlikely(unlazy_walk(nd)))
1033                                         return ERR_PTR(-ECHILD);
1034                                 res = get(dentry, inode, &last->done);
1035                         }
1036                 } else {
1037                         res = get(dentry, inode, &last->done);
1038                 }
1039                 if (IS_ERR_OR_NULL(res))
1040                         return res;
1041         }
1042         if (*res == '/') {
1043                 if (!nd->root.mnt)
1044                         set_root(nd);
1045                 if (unlikely(nd_jump_root(nd)))
1046                         return ERR_PTR(-ECHILD);
1047                 while (unlikely(*++res == '/'))
1048                         ;
1049         }
1050         if (!*res)
1051                 res = NULL;
1052         return res;
1053 }
1054
1055 /*
1056  * follow_up - Find the mountpoint of path's vfsmount
1057  *
1058  * Given a path, find the mountpoint of its source file system.
1059  * Replace @path with the path of the mountpoint in the parent mount.
1060  * Up is towards /.
1061  *
1062  * Return 1 if we went up a level and 0 if we were already at the
1063  * root.
1064  */
1065 int follow_up(struct path *path)
1066 {
1067         struct mount *mnt = real_mount(path->mnt);
1068         struct mount *parent;
1069         struct dentry *mountpoint;
1070
1071         read_seqlock_excl(&mount_lock);
1072         parent = mnt->mnt_parent;
1073         if (parent == mnt) {
1074                 read_sequnlock_excl(&mount_lock);
1075                 return 0;
1076         }
1077         mntget(&parent->mnt);
1078         mountpoint = dget(mnt->mnt_mountpoint);
1079         read_sequnlock_excl(&mount_lock);
1080         dput(path->dentry);
1081         path->dentry = mountpoint;
1082         mntput(path->mnt);
1083         path->mnt = &parent->mnt;
1084         return 1;
1085 }
1086 EXPORT_SYMBOL(follow_up);
1087
1088 /*
1089  * Perform an automount
1090  * - return -EISDIR to tell follow_managed() to stop and return the path we
1091  *   were called with.
1092  */
1093 static int follow_automount(struct path *path, struct nameidata *nd,
1094                             bool *need_mntput)
1095 {
1096         struct vfsmount *mnt;
1097         int err;
1098
1099         if (!path->dentry->d_op || !path->dentry->d_op->d_automount)
1100                 return -EREMOTE;
1101
1102         /* We don't want to mount if someone's just doing a stat -
1103          * unless they're stat'ing a directory and appended a '/' to
1104          * the name.
1105          *
1106          * We do, however, want to mount if someone wants to open or
1107          * create a file of any type under the mountpoint, wants to
1108          * traverse through the mountpoint or wants to open the
1109          * mounted directory.  Also, autofs may mark negative dentries
1110          * as being automount points.  These will need the attentions
1111          * of the daemon to instantiate them before they can be used.
1112          */
1113         if (!(nd->flags & (LOOKUP_PARENT | LOOKUP_DIRECTORY |
1114                            LOOKUP_OPEN | LOOKUP_CREATE | LOOKUP_AUTOMOUNT)) &&
1115             path->dentry->d_inode)
1116                 return -EISDIR;
1117
1118         nd->total_link_count++;
1119         if (nd->total_link_count >= 40)
1120                 return -ELOOP;
1121
1122         mnt = path->dentry->d_op->d_automount(path);
1123         if (IS_ERR(mnt)) {
1124                 /*
1125                  * The filesystem is allowed to return -EISDIR here to indicate
1126                  * it doesn't want to automount.  For instance, autofs would do
1127                  * this so that its userspace daemon can mount on this dentry.
1128                  *
1129                  * However, we can only permit this if it's a terminal point in
1130                  * the path being looked up; if it wasn't then the remainder of
1131                  * the path is inaccessible and we should say so.
1132                  */
1133                 if (PTR_ERR(mnt) == -EISDIR && (nd->flags & LOOKUP_PARENT))
1134                         return -EREMOTE;
1135                 return PTR_ERR(mnt);
1136         }
1137
1138         if (!mnt) /* mount collision */
1139                 return 0;
1140
1141         if (!*need_mntput) {
1142                 /* lock_mount() may release path->mnt on error */
1143                 mntget(path->mnt);
1144                 *need_mntput = true;
1145         }
1146         err = finish_automount(mnt, path);
1147
1148         switch (err) {
1149         case -EBUSY:
1150                 /* Someone else made a mount here whilst we were busy */
1151                 return 0;
1152         case 0:
1153                 path_put(path);
1154                 path->mnt = mnt;
1155                 path->dentry = dget(mnt->mnt_root);
1156                 return 0;
1157         default:
1158                 return err;
1159         }
1160
1161 }
1162
1163 /*
1164  * Handle a dentry that is managed in some way.
1165  * - Flagged for transit management (autofs)
1166  * - Flagged as mountpoint
1167  * - Flagged as automount point
1168  *
1169  * This may only be called in refwalk mode.
1170  *
1171  * Serialization is taken care of in namespace.c
1172  */
1173 static int follow_managed(struct path *path, struct nameidata *nd)
1174 {
1175         struct vfsmount *mnt = path->mnt; /* held by caller, must be left alone */
1176         unsigned managed;
1177         bool need_mntput = false;
1178         int ret = 0;
1179
1180         /* Given that we're not holding a lock here, we retain the value in a
1181          * local variable for each dentry as we look at it so that we don't see
1182          * the components of that value change under us */
1183         while (managed = READ_ONCE(path->dentry->d_flags),
1184                managed &= DCACHE_MANAGED_DENTRY,
1185                unlikely(managed != 0)) {
1186                 /* Allow the filesystem to manage the transit without i_mutex
1187                  * being held. */
1188                 if (managed & DCACHE_MANAGE_TRANSIT) {
1189                         BUG_ON(!path->dentry->d_op);
1190                         BUG_ON(!path->dentry->d_op->d_manage);
1191                         ret = path->dentry->d_op->d_manage(path, false);
1192                         if (ret < 0)
1193                                 break;
1194                 }
1195
1196                 /* Transit to a mounted filesystem. */
1197                 if (managed & DCACHE_MOUNTED) {
1198                         struct vfsmount *mounted = lookup_mnt(path);
1199                         if (mounted) {
1200                                 dput(path->dentry);
1201                                 if (need_mntput)
1202                                         mntput(path->mnt);
1203                                 path->mnt = mounted;
1204                                 path->dentry = dget(mounted->mnt_root);
1205                                 need_mntput = true;
1206                                 continue;
1207                         }
1208
1209                         /* Something is mounted on this dentry in another
1210                          * namespace and/or whatever was mounted there in this
1211                          * namespace got unmounted before lookup_mnt() could
1212                          * get it */
1213                 }
1214
1215                 /* Handle an automount point */
1216                 if (managed & DCACHE_NEED_AUTOMOUNT) {
1217                         ret = follow_automount(path, nd, &need_mntput);
1218                         if (ret < 0)
1219                                 break;
1220                         continue;
1221                 }
1222
1223                 /* We didn't change the current path point */
1224                 break;
1225         }
1226
1227         if (need_mntput && path->mnt == mnt)
1228                 mntput(path->mnt);
1229         if (ret == -EISDIR || !ret)
1230                 ret = 1;
1231         if (need_mntput)
1232                 nd->flags |= LOOKUP_JUMPED;
1233         if (unlikely(ret < 0))
1234                 path_put_conditional(path, nd);
1235         return ret;
1236 }
1237
1238 int follow_down_one(struct path *path)
1239 {
1240         struct vfsmount *mounted;
1241
1242         mounted = lookup_mnt(path);
1243         if (mounted) {
1244                 dput(path->dentry);
1245                 mntput(path->mnt);
1246                 path->mnt = mounted;
1247                 path->dentry = dget(mounted->mnt_root);
1248                 return 1;
1249         }
1250         return 0;
1251 }
1252 EXPORT_SYMBOL(follow_down_one);
1253
1254 static inline int managed_dentry_rcu(const struct path *path)
1255 {
1256         return (path->dentry->d_flags & DCACHE_MANAGE_TRANSIT) ?
1257                 path->dentry->d_op->d_manage(path, true) : 0;
1258 }
1259
1260 /*
1261  * Try to skip to top of mountpoint pile in rcuwalk mode.  Fail if
1262  * we meet a managed dentry that would need blocking.
1263  */
1264 static bool __follow_mount_rcu(struct nameidata *nd, struct path *path,
1265                                struct inode **inode, unsigned *seqp)
1266 {
1267         for (;;) {
1268                 struct mount *mounted;
1269                 /*
1270                  * Don't forget we might have a non-mountpoint managed dentry
1271                  * that wants to block transit.
1272                  */
1273                 switch (managed_dentry_rcu(path)) {
1274                 case -ECHILD:
1275                 default:
1276                         return false;
1277                 case -EISDIR:
1278                         return true;
1279                 case 0:
1280                         break;
1281                 }
1282
1283                 if (!d_mountpoint(path->dentry))
1284                         return !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1285
1286                 mounted = __lookup_mnt(path->mnt, path->dentry);
1287                 if (!mounted)
1288                         break;
1289                 path->mnt = &mounted->mnt;
1290                 path->dentry = mounted->mnt.mnt_root;
1291                 nd->flags |= LOOKUP_JUMPED;
1292                 *seqp = read_seqcount_begin(&path->dentry->d_seq);
1293                 /*
1294                  * Update the inode too. We don't need to re-check the
1295                  * dentry sequence number here after this d_inode read,
1296                  * because a mount-point is always pinned.
1297                  */
1298                 *inode = path->dentry->d_inode;
1299         }
1300         return !read_seqretry(&mount_lock, nd->m_seq) &&
1301                 !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1302 }
1303
1304 static int follow_dotdot_rcu(struct nameidata *nd)
1305 {
1306         struct inode *inode = nd->inode;
1307
1308         while (1) {
1309                 if (path_equal(&nd->path, &nd->root))
1310                         break;
1311                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1312                         struct dentry *old = nd->path.dentry;
1313                         struct dentry *parent = old->d_parent;
1314                         unsigned seq;
1315
1316                         inode = parent->d_inode;
1317                         seq = read_seqcount_begin(&parent->d_seq);
1318                         if (unlikely(read_seqcount_retry(&old->d_seq, nd->seq)))
1319                                 return -ECHILD;
1320                         nd->path.dentry = parent;
1321                         nd->seq = seq;
1322                         if (unlikely(!path_connected(&nd->path)))
1323                                 return -ENOENT;
1324                         break;
1325                 } else {
1326                         struct mount *mnt = real_mount(nd->path.mnt);
1327                         struct mount *mparent = mnt->mnt_parent;
1328                         struct dentry *mountpoint = mnt->mnt_mountpoint;
1329                         struct inode *inode2 = mountpoint->d_inode;
1330                         unsigned seq = read_seqcount_begin(&mountpoint->d_seq);
1331                         if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1332                                 return -ECHILD;
1333                         if (&mparent->mnt == nd->path.mnt)
1334                                 break;
1335                         /* we know that mountpoint was pinned */
1336                         nd->path.dentry = mountpoint;
1337                         nd->path.mnt = &mparent->mnt;
1338                         inode = inode2;
1339                         nd->seq = seq;
1340                 }
1341         }
1342         while (unlikely(d_mountpoint(nd->path.dentry))) {
1343                 struct mount *mounted;
1344                 mounted = __lookup_mnt(nd->path.mnt, nd->path.dentry);
1345                 if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1346                         return -ECHILD;
1347                 if (!mounted)
1348                         break;
1349                 nd->path.mnt = &mounted->mnt;
1350                 nd->path.dentry = mounted->mnt.mnt_root;
1351                 inode = nd->path.dentry->d_inode;
1352                 nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
1353         }
1354         nd->inode = inode;
1355         return 0;
1356 }
1357
1358 /*
1359  * Follow down to the covering mount currently visible to userspace.  At each
1360  * point, the filesystem owning that dentry may be queried as to whether the
1361  * caller is permitted to proceed or not.
1362  */
1363 int follow_down(struct path *path)
1364 {
1365         unsigned managed;
1366         int ret;
1367
1368         while (managed = READ_ONCE(path->dentry->d_flags),
1369                unlikely(managed & DCACHE_MANAGED_DENTRY)) {
1370                 /* Allow the filesystem to manage the transit without i_mutex
1371                  * being held.
1372                  *
1373                  * We indicate to the filesystem if someone is trying to mount
1374                  * something here.  This gives autofs the chance to deny anyone
1375                  * other than its daemon the right to mount on its
1376                  * superstructure.
1377                  *
1378                  * The filesystem may sleep at this point.
1379                  */
1380                 if (managed & DCACHE_MANAGE_TRANSIT) {
1381                         BUG_ON(!path->dentry->d_op);
1382                         BUG_ON(!path->dentry->d_op->d_manage);
1383                         ret = path->dentry->d_op->d_manage(path, false);
1384                         if (ret < 0)
1385                                 return ret == -EISDIR ? 0 : ret;
1386                 }
1387
1388                 /* Transit to a mounted filesystem. */
1389                 if (managed & DCACHE_MOUNTED) {
1390                         struct vfsmount *mounted = lookup_mnt(path);
1391                         if (!mounted)
1392                                 break;
1393                         dput(path->dentry);
1394                         mntput(path->mnt);
1395                         path->mnt = mounted;
1396                         path->dentry = dget(mounted->mnt_root);
1397                         continue;
1398                 }
1399
1400                 /* Don't handle automount points here */
1401                 break;
1402         }
1403         return 0;
1404 }
1405 EXPORT_SYMBOL(follow_down);
1406
1407 /*
1408  * Skip to top of mountpoint pile in refwalk mode for follow_dotdot()
1409  */
1410 static void follow_mount(struct path *path)
1411 {
1412         while (d_mountpoint(path->dentry)) {
1413                 struct vfsmount *mounted = lookup_mnt(path);
1414                 if (!mounted)
1415                         break;
1416                 dput(path->dentry);
1417                 mntput(path->mnt);
1418                 path->mnt = mounted;
1419                 path->dentry = dget(mounted->mnt_root);
1420         }
1421 }
1422
1423 static int path_parent_directory(struct path *path)
1424 {
1425         struct dentry *old = path->dentry;
1426         /* rare case of legitimate dget_parent()... */
1427         path->dentry = dget_parent(path->dentry);
1428         dput(old);
1429         if (unlikely(!path_connected(path)))
1430                 return -ENOENT;
1431         return 0;
1432 }
1433
1434 static int follow_dotdot(struct nameidata *nd)
1435 {
1436         while(1) {
1437                 if (nd->path.dentry == nd->root.dentry &&
1438                     nd->path.mnt == nd->root.mnt) {
1439                         break;
1440                 }
1441                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1442                         int ret = path_parent_directory(&nd->path);
1443                         if (ret)
1444                                 return ret;
1445                         break;
1446                 }
1447                 if (!follow_up(&nd->path))
1448                         break;
1449         }
1450         follow_mount(&nd->path);
1451         nd->inode = nd->path.dentry->d_inode;
1452         return 0;
1453 }
1454
1455 /*
1456  * This looks up the name in dcache and possibly revalidates the found dentry.
1457  * NULL is returned if the dentry does not exist in the cache.
1458  */
1459 static struct dentry *lookup_dcache(const struct qstr *name,
1460                                     struct dentry *dir,
1461                                     unsigned int flags)
1462 {
1463         struct dentry *dentry = d_lookup(dir, name);
1464         if (dentry) {
1465                 int error = d_revalidate(dentry, flags);
1466                 if (unlikely(error <= 0)) {
1467                         if (!error)
1468                                 d_invalidate(dentry);
1469                         dput(dentry);
1470                         return ERR_PTR(error);
1471                 }
1472         }
1473         return dentry;
1474 }
1475
1476 /*
1477  * Parent directory has inode locked exclusive.  This is one
1478  * and only case when ->lookup() gets called on non in-lookup
1479  * dentries - as the matter of fact, this only gets called
1480  * when directory is guaranteed to have no in-lookup children
1481  * at all.
1482  */
1483 static struct dentry *__lookup_hash(const struct qstr *name,
1484                 struct dentry *base, unsigned int flags)
1485 {
1486         struct dentry *dentry = lookup_dcache(name, base, flags);
1487         struct dentry *old;
1488         struct inode *dir = base->d_inode;
1489
1490         if (dentry)
1491                 return dentry;
1492
1493         /* Don't create child dentry for a dead directory. */
1494         if (unlikely(IS_DEADDIR(dir)))
1495                 return ERR_PTR(-ENOENT);
1496
1497         dentry = d_alloc(base, name);
1498         if (unlikely(!dentry))
1499                 return ERR_PTR(-ENOMEM);
1500
1501         old = dir->i_op->lookup(dir, dentry, flags);
1502         if (unlikely(old)) {
1503                 dput(dentry);
1504                 dentry = old;
1505         }
1506         return dentry;
1507 }
1508
1509 static int lookup_fast(struct nameidata *nd,
1510                        struct path *path, struct inode **inode,
1511                        unsigned *seqp)
1512 {
1513         struct vfsmount *mnt = nd->path.mnt;
1514         struct dentry *dentry, *parent = nd->path.dentry;
1515         int status = 1;
1516         int err;
1517
1518         /*
1519          * Rename seqlock is not required here because in the off chance
1520          * of a false negative due to a concurrent rename, the caller is
1521          * going to fall back to non-racy lookup.
1522          */
1523         if (nd->flags & LOOKUP_RCU) {
1524                 unsigned seq;
1525                 bool negative;
1526                 dentry = __d_lookup_rcu(parent, &nd->last, &seq);
1527                 if (unlikely(!dentry)) {
1528                         if (unlazy_walk(nd))
1529                                 return -ECHILD;
1530                         return 0;
1531                 }
1532
1533                 /*
1534                  * This sequence count validates that the inode matches
1535                  * the dentry name information from lookup.
1536                  */
1537                 *inode = d_backing_inode(dentry);
1538                 negative = d_is_negative(dentry);
1539                 if (unlikely(read_seqcount_retry(&dentry->d_seq, seq)))
1540                         return -ECHILD;
1541
1542                 /*
1543                  * This sequence count validates that the parent had no
1544                  * changes while we did the lookup of the dentry above.
1545                  *
1546                  * The memory barrier in read_seqcount_begin of child is
1547                  *  enough, we can use __read_seqcount_retry here.
1548                  */
1549                 if (unlikely(__read_seqcount_retry(&parent->d_seq, nd->seq)))
1550                         return -ECHILD;
1551
1552                 *seqp = seq;
1553                 status = d_revalidate(dentry, nd->flags);
1554                 if (likely(status > 0)) {
1555                         /*
1556                          * Note: do negative dentry check after revalidation in
1557                          * case that drops it.
1558                          */
1559                         if (unlikely(negative))
1560                                 return -ENOENT;
1561                         path->mnt = mnt;
1562                         path->dentry = dentry;
1563                         if (likely(__follow_mount_rcu(nd, path, inode, seqp)))
1564                                 return 1;
1565                 }
1566                 if (unlazy_child(nd, dentry, seq))
1567                         return -ECHILD;
1568                 if (unlikely(status == -ECHILD))
1569                         /* we'd been told to redo it in non-rcu mode */
1570                         status = d_revalidate(dentry, nd->flags);
1571         } else {
1572                 dentry = __d_lookup(parent, &nd->last);
1573                 if (unlikely(!dentry))
1574                         return 0;
1575                 status = d_revalidate(dentry, nd->flags);
1576         }
1577         if (unlikely(status <= 0)) {
1578                 if (!status)
1579                         d_invalidate(dentry);
1580                 dput(dentry);
1581                 return status;
1582         }
1583         if (unlikely(d_is_negative(dentry))) {
1584                 dput(dentry);
1585                 return -ENOENT;
1586         }
1587
1588         path->mnt = mnt;
1589         path->dentry = dentry;
1590         err = follow_managed(path, nd);
1591         if (likely(err > 0))
1592                 *inode = d_backing_inode(path->dentry);
1593         return err;
1594 }
1595
1596 /* Fast lookup failed, do it the slow way */
1597 static struct dentry *lookup_slow(const struct qstr *name,
1598                                   struct dentry *dir,
1599                                   unsigned int flags)
1600 {
1601         struct dentry *dentry = ERR_PTR(-ENOENT), *old;
1602         struct inode *inode = dir->d_inode;
1603         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
1604
1605         inode_lock_shared(inode);
1606         /* Don't go there if it's already dead */
1607         if (unlikely(IS_DEADDIR(inode)))
1608                 goto out;
1609 again:
1610         dentry = d_alloc_parallel(dir, name, &wq);
1611         if (IS_ERR(dentry))
1612                 goto out;
1613         if (unlikely(!d_in_lookup(dentry))) {
1614                 if (!(flags & LOOKUP_NO_REVAL)) {
1615                         int error = d_revalidate(dentry, flags);
1616                         if (unlikely(error <= 0)) {
1617                                 if (!error) {
1618                                         d_invalidate(dentry);
1619                                         dput(dentry);
1620                                         goto again;
1621                                 }
1622                                 dput(dentry);
1623                                 dentry = ERR_PTR(error);
1624                         }
1625                 }
1626         } else {
1627                 old = inode->i_op->lookup(inode, dentry, flags);
1628                 d_lookup_done(dentry);
1629                 if (unlikely(old)) {
1630                         dput(dentry);
1631                         dentry = old;
1632                 }
1633         }
1634 out:
1635         inode_unlock_shared(inode);
1636         return dentry;
1637 }
1638
1639 static inline int may_lookup(struct nameidata *nd)
1640 {
1641         if (nd->flags & LOOKUP_RCU) {
1642                 int err = inode_permission(nd->inode, MAY_EXEC|MAY_NOT_BLOCK);
1643                 if (err != -ECHILD)
1644                         return err;
1645                 if (unlazy_walk(nd))
1646                         return -ECHILD;
1647         }
1648         return inode_permission(nd->inode, MAY_EXEC);
1649 }
1650
1651 static inline int handle_dots(struct nameidata *nd, int type)
1652 {
1653         if (type == LAST_DOTDOT) {
1654                 if (!nd->root.mnt)
1655                         set_root(nd);
1656                 if (nd->flags & LOOKUP_RCU) {
1657                         return follow_dotdot_rcu(nd);
1658                 } else
1659                         return follow_dotdot(nd);
1660         }
1661         return 0;
1662 }
1663
1664 static int pick_link(struct nameidata *nd, struct path *link,
1665                      struct inode *inode, unsigned seq)
1666 {
1667         int error;
1668         struct saved *last;
1669         if (unlikely(nd->total_link_count++ >= MAXSYMLINKS)) {
1670                 path_to_nameidata(link, nd);
1671                 return -ELOOP;
1672         }
1673         if (!(nd->flags & LOOKUP_RCU)) {
1674                 if (link->mnt == nd->path.mnt)
1675                         mntget(link->mnt);
1676         }
1677         error = nd_alloc_stack(nd);
1678         if (unlikely(error)) {
1679                 if (error == -ECHILD) {
1680                         if (unlikely(!legitimize_path(nd, link, seq))) {
1681                                 drop_links(nd);
1682                                 nd->depth = 0;
1683                                 nd->flags &= ~LOOKUP_RCU;
1684                                 nd->path.mnt = NULL;
1685                                 nd->path.dentry = NULL;
1686                                 if (!(nd->flags & LOOKUP_ROOT))
1687                                         nd->root.mnt = NULL;
1688                                 rcu_read_unlock();
1689                         } else if (likely(unlazy_walk(nd)) == 0)
1690                                 error = nd_alloc_stack(nd);
1691                 }
1692                 if (error) {
1693                         path_put(link);
1694                         return error;
1695                 }
1696         }
1697
1698         last = nd->stack + nd->depth++;
1699         last->link = *link;
1700         clear_delayed_call(&last->done);
1701         nd->link_inode = inode;
1702         last->seq = seq;
1703         return 1;
1704 }
1705
1706 enum {WALK_FOLLOW = 1, WALK_MORE = 2};
1707
1708 /*
1709  * Do we need to follow links? We _really_ want to be able
1710  * to do this check without having to look at inode->i_op,
1711  * so we keep a cache of "no, this doesn't need follow_link"
1712  * for the common case.
1713  */
1714 static inline int step_into(struct nameidata *nd, struct path *path,
1715                             int flags, struct inode *inode, unsigned seq)
1716 {
1717         if (!(flags & WALK_MORE) && nd->depth)
1718                 put_link(nd);
1719         if (likely(!d_is_symlink(path->dentry)) ||
1720            !(flags & WALK_FOLLOW || nd->flags & LOOKUP_FOLLOW)) {
1721                 /* not a symlink or should not follow */
1722                 path_to_nameidata(path, nd);
1723                 nd->inode = inode;
1724                 nd->seq = seq;
1725                 return 0;
1726         }
1727         /* make sure that d_is_symlink above matches inode */
1728         if (nd->flags & LOOKUP_RCU) {
1729                 if (read_seqcount_retry(&path->dentry->d_seq, seq))
1730                         return -ECHILD;
1731         }
1732         return pick_link(nd, path, inode, seq);
1733 }
1734
1735 static int walk_component(struct nameidata *nd, int flags)
1736 {
1737         struct path path;
1738         struct inode *inode;
1739         unsigned seq;
1740         int err;
1741         /*
1742          * "." and ".." are special - ".." especially so because it has
1743          * to be able to know about the current root directory and
1744          * parent relationships.
1745          */
1746         if (unlikely(nd->last_type != LAST_NORM)) {
1747                 err = handle_dots(nd, nd->last_type);
1748                 if (!(flags & WALK_MORE) && nd->depth)
1749                         put_link(nd);
1750                 return err;
1751         }
1752         err = lookup_fast(nd, &path, &inode, &seq);
1753         if (unlikely(err <= 0)) {
1754                 if (err < 0)
1755                         return err;
1756                 path.dentry = lookup_slow(&nd->last, nd->path.dentry,
1757                                           nd->flags);
1758                 if (IS_ERR(path.dentry))
1759                         return PTR_ERR(path.dentry);
1760
1761                 path.mnt = nd->path.mnt;
1762                 err = follow_managed(&path, nd);
1763                 if (unlikely(err < 0))
1764                         return err;
1765
1766                 if (unlikely(d_is_negative(path.dentry))) {
1767                         path_to_nameidata(&path, nd);
1768                         return -ENOENT;
1769                 }
1770
1771                 seq = 0;        /* we are already out of RCU mode */
1772                 inode = d_backing_inode(path.dentry);
1773         }
1774
1775         return step_into(nd, &path, flags, inode, seq);
1776 }
1777
1778 /*
1779  * We can do the critical dentry name comparison and hashing
1780  * operations one word at a time, but we are limited to:
1781  *
1782  * - Architectures with fast unaligned word accesses. We could
1783  *   do a "get_unaligned()" if this helps and is sufficiently
1784  *   fast.
1785  *
1786  * - non-CONFIG_DEBUG_PAGEALLOC configurations (so that we
1787  *   do not trap on the (extremely unlikely) case of a page
1788  *   crossing operation.
1789  *
1790  * - Furthermore, we need an efficient 64-bit compile for the
1791  *   64-bit case in order to generate the "number of bytes in
1792  *   the final mask". Again, that could be replaced with a
1793  *   efficient population count instruction or similar.
1794  */
1795 #ifdef CONFIG_DCACHE_WORD_ACCESS
1796
1797 #include <asm/word-at-a-time.h>
1798
1799 #ifdef HASH_MIX
1800
1801 /* Architecture provides HASH_MIX and fold_hash() in <asm/hash.h> */
1802
1803 #elif defined(CONFIG_64BIT)
1804 /*
1805  * Register pressure in the mixing function is an issue, particularly
1806  * on 32-bit x86, but almost any function requires one state value and
1807  * one temporary.  Instead, use a function designed for two state values
1808  * and no temporaries.
1809  *
1810  * This function cannot create a collision in only two iterations, so
1811  * we have two iterations to achieve avalanche.  In those two iterations,
1812  * we have six layers of mixing, which is enough to spread one bit's
1813  * influence out to 2^6 = 64 state bits.
1814  *
1815  * Rotate constants are scored by considering either 64 one-bit input
1816  * deltas or 64*63/2 = 2016 two-bit input deltas, and finding the
1817  * probability of that delta causing a change to each of the 128 output
1818  * bits, using a sample of random initial states.
1819  *
1820  * The Shannon entropy of the computed probabilities is then summed
1821  * to produce a score.  Ideally, any input change has a 50% chance of
1822  * toggling any given output bit.
1823  *
1824  * Mixing scores (in bits) for (12,45):
1825  * Input delta: 1-bit      2-bit
1826  * 1 round:     713.3    42542.6
1827  * 2 rounds:   2753.7   140389.8
1828  * 3 rounds:   5954.1   233458.2
1829  * 4 rounds:   7862.6   256672.2
1830  * Perfect:    8192     258048
1831  *            (64*128) (64*63/2 * 128)
1832  */
1833 #define HASH_MIX(x, y, a)       \
1834         (       x ^= (a),       \
1835         y ^= x, x = rol64(x,12),\
1836         x += y, y = rol64(y,45),\
1837         y *= 9                  )
1838
1839 /*
1840  * Fold two longs into one 32-bit hash value.  This must be fast, but
1841  * latency isn't quite as critical, as there is a fair bit of additional
1842  * work done before the hash value is used.
1843  */
1844 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1845 {
1846         y ^= x * GOLDEN_RATIO_64;
1847         y *= GOLDEN_RATIO_64;
1848         return y >> 32;
1849 }
1850
1851 #else   /* 32-bit case */
1852
1853 /*
1854  * Mixing scores (in bits) for (7,20):
1855  * Input delta: 1-bit      2-bit
1856  * 1 round:     330.3     9201.6
1857  * 2 rounds:   1246.4    25475.4
1858  * 3 rounds:   1907.1    31295.1
1859  * 4 rounds:   2042.3    31718.6
1860  * Perfect:    2048      31744
1861  *            (32*64)   (32*31/2 * 64)
1862  */
1863 #define HASH_MIX(x, y, a)       \
1864         (       x ^= (a),       \
1865         y ^= x, x = rol32(x, 7),\
1866         x += y, y = rol32(y,20),\
1867         y *= 9                  )
1868
1869 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1870 {
1871         /* Use arch-optimized multiply if one exists */
1872         return __hash_32(y ^ __hash_32(x));
1873 }
1874
1875 #endif
1876
1877 /*
1878  * Return the hash of a string of known length.  This is carfully
1879  * designed to match hash_name(), which is the more critical function.
1880  * In particular, we must end by hashing a final word containing 0..7
1881  * payload bytes, to match the way that hash_name() iterates until it
1882  * finds the delimiter after the name.
1883  */
1884 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1885 {
1886         unsigned long a, x = 0, y = (unsigned long)salt;
1887
1888         for (;;) {
1889                 if (!len)
1890                         goto done;
1891                 a = load_unaligned_zeropad(name);
1892                 if (len < sizeof(unsigned long))
1893                         break;
1894                 HASH_MIX(x, y, a);
1895                 name += sizeof(unsigned long);
1896                 len -= sizeof(unsigned long);
1897         }
1898         x ^= a & bytemask_from_count(len);
1899 done:
1900         return fold_hash(x, y);
1901 }
1902 EXPORT_SYMBOL(full_name_hash);
1903
1904 /* Return the "hash_len" (hash and length) of a null-terminated string */
1905 u64 hashlen_string(const void *salt, const char *name)
1906 {
1907         unsigned long a = 0, x = 0, y = (unsigned long)salt;
1908         unsigned long adata, mask, len;
1909         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1910
1911         len = 0;
1912         goto inside;
1913
1914         do {
1915                 HASH_MIX(x, y, a);
1916                 len += sizeof(unsigned long);
1917 inside:
1918                 a = load_unaligned_zeropad(name+len);
1919         } while (!has_zero(a, &adata, &constants));
1920
1921         adata = prep_zero_mask(a, adata, &constants);
1922         mask = create_zero_mask(adata);
1923         x ^= a & zero_bytemask(mask);
1924
1925         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1926 }
1927 EXPORT_SYMBOL(hashlen_string);
1928
1929 /*
1930  * Calculate the length and hash of the path component, and
1931  * return the "hash_len" as the result.
1932  */
1933 static inline u64 hash_name(const void *salt, const char *name)
1934 {
1935         unsigned long a = 0, b, x = 0, y = (unsigned long)salt;
1936         unsigned long adata, bdata, mask, len;
1937         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1938
1939         len = 0;
1940         goto inside;
1941
1942         do {
1943                 HASH_MIX(x, y, a);
1944                 len += sizeof(unsigned long);
1945 inside:
1946                 a = load_unaligned_zeropad(name+len);
1947                 b = a ^ REPEAT_BYTE('/');
1948         } while (!(has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)));
1949
1950         adata = prep_zero_mask(a, adata, &constants);
1951         bdata = prep_zero_mask(b, bdata, &constants);
1952         mask = create_zero_mask(adata | bdata);
1953         x ^= a & zero_bytemask(mask);
1954
1955         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1956 }
1957
1958 #else   /* !CONFIG_DCACHE_WORD_ACCESS: Slow, byte-at-a-time version */
1959
1960 /* Return the hash of a string of known length */
1961 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1962 {
1963         unsigned long hash = init_name_hash(salt);
1964         while (len--)
1965                 hash = partial_name_hash((unsigned char)*name++, hash);
1966         return end_name_hash(hash);
1967 }
1968 EXPORT_SYMBOL(full_name_hash);
1969
1970 /* Return the "hash_len" (hash and length) of a null-terminated string */
1971 u64 hashlen_string(const void *salt, const char *name)
1972 {
1973         unsigned long hash = init_name_hash(salt);
1974         unsigned long len = 0, c;
1975
1976         c = (unsigned char)*name;
1977         while (c) {
1978                 len++;
1979                 hash = partial_name_hash(c, hash);
1980                 c = (unsigned char)name[len];
1981         }
1982         return hashlen_create(end_name_hash(hash), len);
1983 }
1984 EXPORT_SYMBOL(hashlen_string);
1985
1986 /*
1987  * We know there's a real path component here of at least
1988  * one character.
1989  */
1990 static inline u64 hash_name(const void *salt, const char *name)
1991 {
1992         unsigned long hash = init_name_hash(salt);
1993         unsigned long len = 0, c;
1994
1995         c = (unsigned char)*name;
1996         do {
1997                 len++;
1998                 hash = partial_name_hash(c, hash);
1999                 c = (unsigned char)name[len];
2000         } while (c && c != '/');
2001         return hashlen_create(end_name_hash(hash), len);
2002 }
2003
2004 #endif
2005
2006 /*
2007  * Name resolution.
2008  * This is the basic name resolution function, turning a pathname into
2009  * the final dentry. We expect 'base' to be positive and a directory.
2010  *
2011  * Returns 0 and nd will have valid dentry and mnt on success.
2012  * Returns error and drops reference to input namei data on failure.
2013  */
2014 static int link_path_walk(const char *name, struct nameidata *nd)
2015 {
2016         int err;
2017
2018         while (*name=='/')
2019                 name++;
2020         if (!*name)
2021                 return 0;
2022
2023         /* At this point we know we have a real path component. */
2024         for(;;) {
2025                 u64 hash_len;
2026                 int type;
2027
2028                 err = may_lookup(nd);
2029                 if (err)
2030                         return err;
2031
2032                 hash_len = hash_name(nd->path.dentry, name);
2033
2034                 type = LAST_NORM;
2035                 if (name[0] == '.') switch (hashlen_len(hash_len)) {
2036                         case 2:
2037                                 if (name[1] == '.') {
2038                                         type = LAST_DOTDOT;
2039                                         nd->flags |= LOOKUP_JUMPED;
2040                                 }
2041                                 break;
2042                         case 1:
2043                                 type = LAST_DOT;
2044                 }
2045                 if (likely(type == LAST_NORM)) {
2046                         struct dentry *parent = nd->path.dentry;
2047                         nd->flags &= ~LOOKUP_JUMPED;
2048                         if (unlikely(parent->d_flags & DCACHE_OP_HASH)) {
2049                                 struct qstr this = { { .hash_len = hash_len }, .name = name };
2050                                 err = parent->d_op->d_hash(parent, &this);
2051                                 if (err < 0)
2052                                         return err;
2053                                 hash_len = this.hash_len;
2054                                 name = this.name;
2055                         }
2056                 }
2057
2058                 nd->last.hash_len = hash_len;
2059                 nd->last.name = name;
2060                 nd->last_type = type;
2061
2062                 name += hashlen_len(hash_len);
2063                 if (!*name)
2064                         goto OK;
2065                 /*
2066                  * If it wasn't NUL, we know it was '/'. Skip that
2067                  * slash, and continue until no more slashes.
2068                  */
2069                 do {
2070                         name++;
2071                 } while (unlikely(*name == '/'));
2072                 if (unlikely(!*name)) {
2073 OK:
2074                         /* pathname body, done */
2075                         if (!nd->depth)
2076                                 return 0;
2077                         name = nd->stack[nd->depth - 1].name;
2078                         /* trailing symlink, done */
2079                         if (!name)
2080                                 return 0;
2081                         /* last component of nested symlink */
2082                         err = walk_component(nd, WALK_FOLLOW);
2083                 } else {
2084                         /* not the last component */
2085                         err = walk_component(nd, WALK_FOLLOW | WALK_MORE);
2086                 }
2087                 if (err < 0)
2088                         return err;
2089
2090                 if (err) {
2091                         const char *s = get_link(nd);
2092
2093                         if (IS_ERR(s))
2094                                 return PTR_ERR(s);
2095                         err = 0;
2096                         if (unlikely(!s)) {
2097                                 /* jumped */
2098                                 put_link(nd);
2099                         } else {
2100                                 nd->stack[nd->depth - 1].name = name;
2101                                 name = s;
2102                                 continue;
2103                         }
2104                 }
2105                 if (unlikely(!d_can_lookup(nd->path.dentry))) {
2106                         if (nd->flags & LOOKUP_RCU) {
2107                                 if (unlazy_walk(nd))
2108                                         return -ECHILD;
2109                         }
2110                         return -ENOTDIR;
2111                 }
2112         }
2113 }
2114
2115 static const char *path_init(struct nameidata *nd, unsigned flags)
2116 {
2117         const char *s = nd->name->name;
2118
2119         if (!*s)
2120                 flags &= ~LOOKUP_RCU;
2121
2122         nd->last_type = LAST_ROOT; /* if there are only slashes... */
2123         nd->flags = flags | LOOKUP_JUMPED | LOOKUP_PARENT;
2124         nd->depth = 0;
2125         if (flags & LOOKUP_ROOT) {
2126                 struct dentry *root = nd->root.dentry;
2127                 struct inode *inode = root->d_inode;
2128                 if (*s && unlikely(!d_can_lookup(root)))
2129                         return ERR_PTR(-ENOTDIR);
2130                 nd->path = nd->root;
2131                 nd->inode = inode;
2132                 if (flags & LOOKUP_RCU) {
2133                         rcu_read_lock();
2134                         nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2135                         nd->root_seq = nd->seq;
2136                         nd->m_seq = read_seqbegin(&mount_lock);
2137                 } else {
2138                         path_get(&nd->path);
2139                 }
2140                 return s;
2141         }
2142
2143         nd->root.mnt = NULL;
2144         nd->path.mnt = NULL;
2145         nd->path.dentry = NULL;
2146
2147         nd->m_seq = read_seqbegin(&mount_lock);
2148         if (*s == '/') {
2149                 if (flags & LOOKUP_RCU)
2150                         rcu_read_lock();
2151                 set_root(nd);
2152                 if (likely(!nd_jump_root(nd)))
2153                         return s;
2154                 nd->root.mnt = NULL;
2155                 rcu_read_unlock();
2156                 return ERR_PTR(-ECHILD);
2157         } else if (nd->dfd == AT_FDCWD) {
2158                 if (flags & LOOKUP_RCU) {
2159                         struct fs_struct *fs = current->fs;
2160                         unsigned seq;
2161
2162                         rcu_read_lock();
2163
2164                         do {
2165                                 seq = read_seqcount_begin(&fs->seq);
2166                                 nd->path = fs->pwd;
2167                                 nd->inode = nd->path.dentry->d_inode;
2168                                 nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2169                         } while (read_seqcount_retry(&fs->seq, seq));
2170                 } else {
2171                         get_fs_pwd(current->fs, &nd->path);
2172                         nd->inode = nd->path.dentry->d_inode;
2173                 }
2174                 return s;
2175         } else {
2176                 /* Caller must check execute permissions on the starting path component */
2177                 struct fd f = fdget_raw(nd->dfd);
2178                 struct dentry *dentry;
2179
2180                 if (!f.file)
2181                         return ERR_PTR(-EBADF);
2182
2183                 dentry = f.file->f_path.dentry;
2184
2185                 if (*s) {
2186                         if (!d_can_lookup(dentry)) {
2187                                 fdput(f);
2188                                 return ERR_PTR(-ENOTDIR);
2189                         }
2190                 }
2191
2192                 nd->path = f.file->f_path;
2193                 if (flags & LOOKUP_RCU) {
2194                         rcu_read_lock();
2195                         nd->inode = nd->path.dentry->d_inode;
2196                         nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
2197                 } else {
2198                         path_get(&nd->path);
2199                         nd->inode = nd->path.dentry->d_inode;
2200                 }
2201                 fdput(f);
2202                 return s;
2203         }
2204 }
2205
2206 static const char *trailing_symlink(struct nameidata *nd)
2207 {
2208         const char *s;
2209         int error = may_follow_link(nd);
2210         if (unlikely(error))
2211                 return ERR_PTR(error);
2212         nd->flags |= LOOKUP_PARENT;
2213         nd->stack[0].name = NULL;
2214         s = get_link(nd);
2215         return s ? s : "";
2216 }
2217
2218 static inline int lookup_last(struct nameidata *nd)
2219 {
2220         if (nd->last_type == LAST_NORM && nd->last.name[nd->last.len])
2221                 nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
2222
2223         nd->flags &= ~LOOKUP_PARENT;
2224         return walk_component(nd, 0);
2225 }
2226
2227 static int handle_lookup_down(struct nameidata *nd)
2228 {
2229         struct path path = nd->path;
2230         struct inode *inode = nd->inode;
2231         unsigned seq = nd->seq;
2232         int err;
2233
2234         if (nd->flags & LOOKUP_RCU) {
2235                 /*
2236                  * don't bother with unlazy_walk on failure - we are
2237                  * at the very beginning of walk, so we lose nothing
2238                  * if we simply redo everything in non-RCU mode
2239                  */
2240                 if (unlikely(!__follow_mount_rcu(nd, &path, &inode, &seq)))
2241                         return -ECHILD;
2242         } else {
2243                 dget(path.dentry);
2244                 err = follow_managed(&path, nd);
2245                 if (unlikely(err < 0))
2246                         return err;
2247                 inode = d_backing_inode(path.dentry);
2248                 seq = 0;
2249         }
2250         path_to_nameidata(&path, nd);
2251         nd->inode = inode;
2252         nd->seq = seq;
2253         return 0;
2254 }
2255
2256 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2257 static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
2258 {
2259         const char *s = path_init(nd, flags);
2260         int err;
2261
2262         if (IS_ERR(s))
2263                 return PTR_ERR(s);
2264
2265         if (unlikely(flags & LOOKUP_DOWN)) {
2266                 err = handle_lookup_down(nd);
2267                 if (unlikely(err < 0)) {
2268                         terminate_walk(nd);
2269                         return err;
2270                 }
2271         }
2272
2273         while (!(err = link_path_walk(s, nd))
2274                 && ((err = lookup_last(nd)) > 0)) {
2275                 s = trailing_symlink(nd);
2276                 if (IS_ERR(s)) {
2277                         err = PTR_ERR(s);
2278                         break;
2279                 }
2280         }
2281         if (!err)
2282                 err = complete_walk(nd);
2283
2284         if (!err && nd->flags & LOOKUP_DIRECTORY)
2285                 if (!d_can_lookup(nd->path.dentry))
2286                         err = -ENOTDIR;
2287         if (!err) {
2288                 *path = nd->path;
2289                 nd->path.mnt = NULL;
2290                 nd->path.dentry = NULL;
2291         }
2292         terminate_walk(nd);
2293         return err;
2294 }
2295
2296 static int filename_lookup(int dfd, struct filename *name, unsigned flags,
2297                            struct path *path, struct path *root)
2298 {
2299         int retval;
2300         struct nameidata nd;
2301         if (IS_ERR(name))
2302                 return PTR_ERR(name);
2303         if (unlikely(root)) {
2304                 nd.root = *root;
2305                 flags |= LOOKUP_ROOT;
2306         }
2307         set_nameidata(&nd, dfd, name);
2308         retval = path_lookupat(&nd, flags | LOOKUP_RCU, path);
2309         if (unlikely(retval == -ECHILD))
2310                 retval = path_lookupat(&nd, flags, path);
2311         if (unlikely(retval == -ESTALE))
2312                 retval = path_lookupat(&nd, flags | LOOKUP_REVAL, path);
2313
2314         if (likely(!retval))
2315                 audit_inode(name, path->dentry, flags & LOOKUP_PARENT);
2316         restore_nameidata();
2317         putname(name);
2318         return retval;
2319 }
2320
2321 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2322 static int path_parentat(struct nameidata *nd, unsigned flags,
2323                                 struct path *parent)
2324 {
2325         const char *s = path_init(nd, flags);
2326         int err;
2327         if (IS_ERR(s))
2328                 return PTR_ERR(s);
2329         err = link_path_walk(s, nd);
2330         if (!err)
2331                 err = complete_walk(nd);
2332         if (!err) {
2333                 *parent = nd->path;
2334                 nd->path.mnt = NULL;
2335                 nd->path.dentry = NULL;
2336         }
2337         terminate_walk(nd);
2338         return err;
2339 }
2340
2341 static struct filename *filename_parentat(int dfd, struct filename *name,
2342                                 unsigned int flags, struct path *parent,
2343                                 struct qstr *last, int *type)
2344 {
2345         int retval;
2346         struct nameidata nd;
2347
2348         if (IS_ERR(name))
2349                 return name;
2350         set_nameidata(&nd, dfd, name);
2351         retval = path_parentat(&nd, flags | LOOKUP_RCU, parent);
2352         if (unlikely(retval == -ECHILD))
2353                 retval = path_parentat(&nd, flags, parent);
2354         if (unlikely(retval == -ESTALE))
2355                 retval = path_parentat(&nd, flags | LOOKUP_REVAL, parent);
2356         if (likely(!retval)) {
2357                 *last = nd.last;
2358                 *type = nd.last_type;
2359                 audit_inode(name, parent->dentry, LOOKUP_PARENT);
2360         } else {
2361                 putname(name);
2362                 name = ERR_PTR(retval);
2363         }
2364         restore_nameidata();
2365         return name;
2366 }
2367
2368 /* does lookup, returns the object with parent locked */
2369 struct dentry *kern_path_locked(const char *name, struct path *path)
2370 {
2371         struct filename *filename;
2372         struct dentry *d;
2373         struct qstr last;
2374         int type;
2375
2376         filename = filename_parentat(AT_FDCWD, getname_kernel(name), 0, path,
2377                                     &last, &type);
2378         if (IS_ERR(filename))
2379                 return ERR_CAST(filename);
2380         if (unlikely(type != LAST_NORM)) {
2381                 path_put(path);
2382                 putname(filename);
2383                 return ERR_PTR(-EINVAL);
2384         }
2385         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
2386         d = __lookup_hash(&last, path->dentry, 0);
2387         if (IS_ERR(d)) {
2388                 inode_unlock(path->dentry->d_inode);
2389                 path_put(path);
2390         }
2391         putname(filename);
2392         return d;
2393 }
2394
2395 int kern_path(const char *name, unsigned int flags, struct path *path)
2396 {
2397         return filename_lookup(AT_FDCWD, getname_kernel(name),
2398                                flags, path, NULL);
2399 }
2400 EXPORT_SYMBOL(kern_path);
2401
2402 /**
2403  * vfs_path_lookup - lookup a file path relative to a dentry-vfsmount pair
2404  * @dentry:  pointer to dentry of the base directory
2405  * @mnt: pointer to vfs mount of the base directory
2406  * @name: pointer to file name
2407  * @flags: lookup flags
2408  * @path: pointer to struct path to fill
2409  */
2410 int vfs_path_lookup(struct dentry *dentry, struct vfsmount *mnt,
2411                     const char *name, unsigned int flags,
2412                     struct path *path)
2413 {
2414         struct path root = {.mnt = mnt, .dentry = dentry};
2415         /* the first argument of filename_lookup() is ignored with root */
2416         return filename_lookup(AT_FDCWD, getname_kernel(name),
2417                                flags , path, &root);
2418 }
2419 EXPORT_SYMBOL(vfs_path_lookup);
2420
2421 /**
2422  * lookup_one_len - filesystem helper to lookup single pathname component
2423  * @name:       pathname component to lookup
2424  * @base:       base directory to lookup from
2425  * @len:        maximum length @len should be interpreted to
2426  *
2427  * Note that this routine is purely a helper for filesystem usage and should
2428  * not be called by generic code.
2429  *
2430  * The caller must hold base->i_mutex.
2431  */
2432 struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
2433 {
2434         struct qstr this;
2435         unsigned int c;
2436         int err;
2437
2438         WARN_ON_ONCE(!inode_is_locked(base->d_inode));
2439
2440         this.name = name;
2441         this.len = len;
2442         this.hash = full_name_hash(base, name, len);
2443         if (!len)
2444                 return ERR_PTR(-EACCES);
2445
2446         if (unlikely(name[0] == '.')) {
2447                 if (len < 2 || (len == 2 && name[1] == '.'))
2448                         return ERR_PTR(-EACCES);
2449         }
2450
2451         while (len--) {
2452                 c = *(const unsigned char *)name++;
2453                 if (c == '/' || c == '\0')
2454                         return ERR_PTR(-EACCES);
2455         }
2456         /*
2457          * See if the low-level filesystem might want
2458          * to use its own hash..
2459          */
2460         if (base->d_flags & DCACHE_OP_HASH) {
2461                 int err = base->d_op->d_hash(base, &this);
2462                 if (err < 0)
2463                         return ERR_PTR(err);
2464         }
2465
2466         err = inode_permission(base->d_inode, MAY_EXEC);
2467         if (err)
2468                 return ERR_PTR(err);
2469
2470         return __lookup_hash(&this, base, 0);
2471 }
2472 EXPORT_SYMBOL(lookup_one_len);
2473
2474 /**
2475  * lookup_one_len_unlocked - filesystem helper to lookup single pathname component
2476  * @name:       pathname component to lookup
2477  * @base:       base directory to lookup from
2478  * @len:        maximum length @len should be interpreted to
2479  *
2480  * Note that this routine is purely a helper for filesystem usage and should
2481  * not be called by generic code.
2482  *
2483  * Unlike lookup_one_len, it should be called without the parent
2484  * i_mutex held, and will take the i_mutex itself if necessary.
2485  */
2486 struct dentry *lookup_one_len_unlocked(const char *name,
2487                                        struct dentry *base, int len)
2488 {
2489         struct qstr this;
2490         unsigned int c;
2491         int err;
2492         struct dentry *ret;
2493
2494         this.name = name;
2495         this.len = len;
2496         this.hash = full_name_hash(base, name, len);
2497         if (!len)
2498                 return ERR_PTR(-EACCES);
2499
2500         if (unlikely(name[0] == '.')) {
2501                 if (len < 2 || (len == 2 && name[1] == '.'))
2502                         return ERR_PTR(-EACCES);
2503         }
2504
2505         while (len--) {
2506                 c = *(const unsigned char *)name++;
2507                 if (c == '/' || c == '\0')
2508                         return ERR_PTR(-EACCES);
2509         }
2510         /*
2511          * See if the low-level filesystem might want
2512          * to use its own hash..
2513          */
2514         if (base->d_flags & DCACHE_OP_HASH) {
2515                 int err = base->d_op->d_hash(base, &this);
2516                 if (err < 0)
2517                         return ERR_PTR(err);
2518         }
2519
2520         err = inode_permission(base->d_inode, MAY_EXEC);
2521         if (err)
2522                 return ERR_PTR(err);
2523
2524         ret = lookup_dcache(&this, base, 0);
2525         if (!ret)
2526                 ret = lookup_slow(&this, base, 0);
2527         return ret;
2528 }
2529 EXPORT_SYMBOL(lookup_one_len_unlocked);
2530
2531 #ifdef CONFIG_UNIX98_PTYS
2532 int path_pts(struct path *path)
2533 {
2534         /* Find something mounted on "pts" in the same directory as
2535          * the input path.
2536          */
2537         struct dentry *child, *parent;
2538         struct qstr this;
2539         int ret;
2540
2541         ret = path_parent_directory(path);
2542         if (ret)
2543                 return ret;
2544
2545         parent = path->dentry;
2546         this.name = "pts";
2547         this.len = 3;
2548         child = d_hash_and_lookup(parent, &this);
2549         if (!child)
2550                 return -ENOENT;
2551
2552         path->dentry = child;
2553         dput(parent);
2554         follow_mount(path);
2555         return 0;
2556 }
2557 #endif
2558
2559 int user_path_at_empty(int dfd, const char __user *name, unsigned flags,
2560                  struct path *path, int *empty)
2561 {
2562         return filename_lookup(dfd, getname_flags(name, flags, empty),
2563                                flags, path, NULL);
2564 }
2565 EXPORT_SYMBOL(user_path_at_empty);
2566
2567 /**
2568  * mountpoint_last - look up last component for umount
2569  * @nd:   pathwalk nameidata - currently pointing at parent directory of "last"
2570  *
2571  * This is a special lookup_last function just for umount. In this case, we
2572  * need to resolve the path without doing any revalidation.
2573  *
2574  * The nameidata should be the result of doing a LOOKUP_PARENT pathwalk. Since
2575  * mountpoints are always pinned in the dcache, their ancestors are too. Thus,
2576  * in almost all cases, this lookup will be served out of the dcache. The only
2577  * cases where it won't are if nd->last refers to a symlink or the path is
2578  * bogus and it doesn't exist.
2579  *
2580  * Returns:
2581  * -error: if there was an error during lookup. This includes -ENOENT if the
2582  *         lookup found a negative dentry.
2583  *
2584  * 0:      if we successfully resolved nd->last and found it to not to be a
2585  *         symlink that needs to be followed.
2586  *
2587  * 1:      if we successfully resolved nd->last and found it to be a symlink
2588  *         that needs to be followed.
2589  */
2590 static int
2591 mountpoint_last(struct nameidata *nd)
2592 {
2593         int error = 0;
2594         struct dentry *dir = nd->path.dentry;
2595         struct path path;
2596
2597         /* If we're in rcuwalk, drop out of it to handle last component */
2598         if (nd->flags & LOOKUP_RCU) {
2599                 if (unlazy_walk(nd))
2600                         return -ECHILD;
2601         }
2602
2603         nd->flags &= ~LOOKUP_PARENT;
2604
2605         if (unlikely(nd->last_type != LAST_NORM)) {
2606                 error = handle_dots(nd, nd->last_type);
2607                 if (error)
2608                         return error;
2609                 path.dentry = dget(nd->path.dentry);
2610         } else {
2611                 path.dentry = d_lookup(dir, &nd->last);
2612                 if (!path.dentry) {
2613                         /*
2614                          * No cached dentry. Mounted dentries are pinned in the
2615                          * cache, so that means that this dentry is probably
2616                          * a symlink or the path doesn't actually point
2617                          * to a mounted dentry.
2618                          */
2619                         path.dentry = lookup_slow(&nd->last, dir,
2620                                              nd->flags | LOOKUP_NO_REVAL);
2621                         if (IS_ERR(path.dentry))
2622                                 return PTR_ERR(path.dentry);
2623                 }
2624         }
2625         if (d_is_negative(path.dentry)) {
2626                 dput(path.dentry);
2627                 return -ENOENT;
2628         }
2629         path.mnt = nd->path.mnt;
2630         return step_into(nd, &path, 0, d_backing_inode(path.dentry), 0);
2631 }
2632
2633 /**
2634  * path_mountpoint - look up a path to be umounted
2635  * @nd:         lookup context
2636  * @flags:      lookup flags
2637  * @path:       pointer to container for result
2638  *
2639  * Look up the given name, but don't attempt to revalidate the last component.
2640  * Returns 0 and "path" will be valid on success; Returns error otherwise.
2641  */
2642 static int
2643 path_mountpoint(struct nameidata *nd, unsigned flags, struct path *path)
2644 {
2645         const char *s = path_init(nd, flags);
2646         int err;
2647         if (IS_ERR(s))
2648                 return PTR_ERR(s);
2649         while (!(err = link_path_walk(s, nd)) &&
2650                 (err = mountpoint_last(nd)) > 0) {
2651                 s = trailing_symlink(nd);
2652                 if (IS_ERR(s)) {
2653                         err = PTR_ERR(s);
2654                         break;
2655                 }
2656         }
2657         if (!err) {
2658                 *path = nd->path;
2659                 nd->path.mnt = NULL;
2660                 nd->path.dentry = NULL;
2661                 follow_mount(path);
2662         }
2663         terminate_walk(nd);
2664         return err;
2665 }
2666
2667 static int
2668 filename_mountpoint(int dfd, struct filename *name, struct path *path,
2669                         unsigned int flags)
2670 {
2671         struct nameidata nd;
2672         int error;
2673         if (IS_ERR(name))
2674                 return PTR_ERR(name);
2675         set_nameidata(&nd, dfd, name);
2676         error = path_mountpoint(&nd, flags | LOOKUP_RCU, path);
2677         if (unlikely(error == -ECHILD))
2678                 error = path_mountpoint(&nd, flags, path);
2679         if (unlikely(error == -ESTALE))
2680                 error = path_mountpoint(&nd, flags | LOOKUP_REVAL, path);
2681         if (likely(!error))
2682                 audit_inode(name, path->dentry, 0);
2683         restore_nameidata();
2684         putname(name);
2685         return error;
2686 }
2687
2688 /**
2689  * user_path_mountpoint_at - lookup a path from userland in order to umount it
2690  * @dfd:        directory file descriptor
2691  * @name:       pathname from userland
2692  * @flags:      lookup flags
2693  * @path:       pointer to container to hold result
2694  *
2695  * A umount is a special case for path walking. We're not actually interested
2696  * in the inode in this situation, and ESTALE errors can be a problem. We
2697  * simply want track down the dentry and vfsmount attached at the mountpoint
2698  * and avoid revalidating the last component.
2699  *
2700  * Returns 0 and populates "path" on success.
2701  */
2702 int
2703 user_path_mountpoint_at(int dfd, const char __user *name, unsigned int flags,
2704                         struct path *path)
2705 {
2706         return filename_mountpoint(dfd, getname(name), path, flags);
2707 }
2708
2709 int
2710 kern_path_mountpoint(int dfd, const char *name, struct path *path,
2711                         unsigned int flags)
2712 {
2713         return filename_mountpoint(dfd, getname_kernel(name), path, flags);
2714 }
2715 EXPORT_SYMBOL(kern_path_mountpoint);
2716
2717 int __check_sticky(struct inode *dir, struct inode *inode)
2718 {
2719         kuid_t fsuid = current_fsuid();
2720
2721         if (uid_eq(inode->i_uid, fsuid))
2722                 return 0;
2723         if (uid_eq(dir->i_uid, fsuid))
2724                 return 0;
2725         return !capable_wrt_inode_uidgid(inode, CAP_FOWNER);
2726 }
2727 EXPORT_SYMBOL(__check_sticky);
2728
2729 /*
2730  *      Check whether we can remove a link victim from directory dir, check
2731  *  whether the type of victim is right.
2732  *  1. We can't do it if dir is read-only (done in permission())
2733  *  2. We should have write and exec permissions on dir
2734  *  3. We can't remove anything from append-only dir
2735  *  4. We can't do anything with immutable dir (done in permission())
2736  *  5. If the sticky bit on dir is set we should either
2737  *      a. be owner of dir, or
2738  *      b. be owner of victim, or
2739  *      c. have CAP_FOWNER capability
2740  *  6. If the victim is append-only or immutable we can't do antyhing with
2741  *     links pointing to it.
2742  *  7. If the victim has an unknown uid or gid we can't change the inode.
2743  *  8. If we were asked to remove a directory and victim isn't one - ENOTDIR.
2744  *  9. If we were asked to remove a non-directory and victim isn't one - EISDIR.
2745  * 10. We can't remove a root or mountpoint.
2746  * 11. We don't allow removal of NFS sillyrenamed files; it's handled by
2747  *     nfs_async_unlink().
2748  */
2749 static int may_delete(struct inode *dir, struct dentry *victim, bool isdir)
2750 {
2751         struct inode *inode = d_backing_inode(victim);
2752         int error;
2753
2754         if (d_is_negative(victim))
2755                 return -ENOENT;
2756         BUG_ON(!inode);
2757
2758         BUG_ON(victim->d_parent->d_inode != dir);
2759         audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);
2760
2761         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
2762         if (error)
2763                 return error;
2764         if (IS_APPEND(dir))
2765                 return -EPERM;
2766
2767         if (check_sticky(dir, inode) || IS_APPEND(inode) ||
2768             IS_IMMUTABLE(inode) || IS_SWAPFILE(inode) || HAS_UNMAPPED_ID(inode))
2769                 return -EPERM;
2770         if (isdir) {
2771                 if (!d_is_dir(victim))
2772                         return -ENOTDIR;
2773                 if (IS_ROOT(victim))
2774                         return -EBUSY;
2775         } else if (d_is_dir(victim))
2776                 return -EISDIR;
2777         if (IS_DEADDIR(dir))
2778                 return -ENOENT;
2779         if (victim->d_flags & DCACHE_NFSFS_RENAMED)
2780                 return -EBUSY;
2781         return 0;
2782 }
2783
2784 /*      Check whether we can create an object with dentry child in directory
2785  *  dir.
2786  *  1. We can't do it if child already exists (open has special treatment for
2787  *     this case, but since we are inlined it's OK)
2788  *  2. We can't do it if dir is read-only (done in permission())
2789  *  3. We can't do it if the fs can't represent the fsuid or fsgid.
2790  *  4. We should have write and exec permissions on dir
2791  *  5. We can't do it if dir is immutable (done in permission())
2792  */
2793 static inline int may_create(struct inode *dir, struct dentry *child)
2794 {
2795         struct user_namespace *s_user_ns;
2796         audit_inode_child(dir, child, AUDIT_TYPE_CHILD_CREATE);
2797         if (child->d_inode)
2798                 return -EEXIST;
2799         if (IS_DEADDIR(dir))
2800                 return -ENOENT;
2801         s_user_ns = dir->i_sb->s_user_ns;
2802         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
2803             !kgid_has_mapping(s_user_ns, current_fsgid()))
2804                 return -EOVERFLOW;
2805         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
2806 }
2807
2808 /*
2809  * p1 and p2 should be directories on the same fs.
2810  */
2811 struct dentry *lock_rename(struct dentry *p1, struct dentry *p2)
2812 {
2813         struct dentry *p;
2814
2815         if (p1 == p2) {
2816                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2817                 return NULL;
2818         }
2819
2820         mutex_lock(&p1->d_sb->s_vfs_rename_mutex);
2821
2822         p = d_ancestor(p2, p1);
2823         if (p) {
2824                 inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
2825                 inode_lock_nested(p1->d_inode, I_MUTEX_CHILD);
2826                 return p;
2827         }
2828
2829         p = d_ancestor(p1, p2);
2830         if (p) {
2831                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2832                 inode_lock_nested(p2->d_inode, I_MUTEX_CHILD);
2833                 return p;
2834         }
2835
2836         inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2837         inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
2838         return NULL;
2839 }
2840 EXPORT_SYMBOL(lock_rename);
2841
2842 void unlock_rename(struct dentry *p1, struct dentry *p2)
2843 {
2844         inode_unlock(p1->d_inode);
2845         if (p1 != p2) {
2846                 inode_unlock(p2->d_inode);
2847                 mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
2848         }
2849 }
2850 EXPORT_SYMBOL(unlock_rename);
2851
2852 int vfs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2853                 bool want_excl)
2854 {
2855         int error = may_create(dir, dentry);
2856         if (error)
2857                 return error;
2858
2859         if (!dir->i_op->create)
2860                 return -EACCES; /* shouldn't it be ENOSYS? */
2861         mode &= S_IALLUGO;
2862         mode |= S_IFREG;
2863         error = security_inode_create(dir, dentry, mode);
2864         if (error)
2865                 return error;
2866         error = dir->i_op->create(dir, dentry, mode, want_excl);
2867         if (!error)
2868                 fsnotify_create(dir, dentry);
2869         return error;
2870 }
2871 EXPORT_SYMBOL(vfs_create);
2872
2873 int vfs_mkobj(struct dentry *dentry, umode_t mode,
2874                 int (*f)(struct dentry *, umode_t, void *),
2875                 void *arg)
2876 {
2877         struct inode *dir = dentry->d_parent->d_inode;
2878         int error = may_create(dir, dentry);
2879         if (error)
2880                 return error;
2881
2882         mode &= S_IALLUGO;
2883         mode |= S_IFREG;
2884         error = security_inode_create(dir, dentry, mode);
2885         if (error)
2886                 return error;
2887         error = f(dentry, mode, arg);
2888         if (!error)
2889                 fsnotify_create(dir, dentry);
2890         return error;
2891 }
2892 EXPORT_SYMBOL(vfs_mkobj);
2893
2894 bool may_open_dev(const struct path *path)
2895 {
2896         return !(path->mnt->mnt_flags & MNT_NODEV) &&
2897                 !(path->mnt->mnt_sb->s_iflags & SB_I_NODEV);
2898 }
2899
2900 static int may_open(const struct path *path, int acc_mode, int flag)
2901 {
2902         struct dentry *dentry = path->dentry;
2903         struct inode *inode = dentry->d_inode;
2904         int error;
2905
2906         if (!inode)
2907                 return -ENOENT;
2908
2909         switch (inode->i_mode & S_IFMT) {
2910         case S_IFLNK:
2911                 return -ELOOP;
2912         case S_IFDIR:
2913                 if (acc_mode & MAY_WRITE)
2914                         return -EISDIR;
2915                 break;
2916         case S_IFBLK:
2917         case S_IFCHR:
2918                 if (!may_open_dev(path))
2919                         return -EACCES;
2920                 /*FALLTHRU*/
2921         case S_IFIFO:
2922         case S_IFSOCK:
2923                 flag &= ~O_TRUNC;
2924                 break;
2925         }
2926
2927         error = inode_permission(inode, MAY_OPEN | acc_mode);
2928         if (error)
2929                 return error;
2930
2931         /*
2932          * An append-only file must be opened in append mode for writing.
2933          */
2934         if (IS_APPEND(inode)) {
2935                 if  ((flag & O_ACCMODE) != O_RDONLY && !(flag & O_APPEND))
2936                         return -EPERM;
2937                 if (flag & O_TRUNC)
2938                         return -EPERM;
2939         }
2940
2941         /* O_NOATIME can only be set by the owner or superuser */
2942         if (flag & O_NOATIME && !inode_owner_or_capable(inode))
2943                 return -EPERM;
2944
2945         return 0;
2946 }
2947
2948 static int handle_truncate(struct file *filp)
2949 {
2950         const struct path *path = &filp->f_path;
2951         struct inode *inode = path->dentry->d_inode;
2952         int error = get_write_access(inode);
2953         if (error)
2954                 return error;
2955         /*
2956          * Refuse to truncate files with mandatory locks held on them.
2957          */
2958         error = locks_verify_locked(filp);
2959         if (!error)
2960                 error = security_path_truncate(path);
2961         if (!error) {
2962                 error = do_truncate(path->dentry, 0,
2963                                     ATTR_MTIME|ATTR_CTIME|ATTR_OPEN,
2964                                     filp);
2965         }
2966         put_write_access(inode);
2967         return error;
2968 }
2969
2970 static inline int open_to_namei_flags(int flag)
2971 {
2972         if ((flag & O_ACCMODE) == 3)
2973                 flag--;
2974         return flag;
2975 }
2976
2977 static int may_o_create(const struct path *dir, struct dentry *dentry, umode_t mode)
2978 {
2979         struct user_namespace *s_user_ns;
2980         int error = security_path_mknod(dir, dentry, mode, 0);
2981         if (error)
2982                 return error;
2983
2984         s_user_ns = dir->dentry->d_sb->s_user_ns;
2985         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
2986             !kgid_has_mapping(s_user_ns, current_fsgid()))
2987                 return -EOVERFLOW;
2988
2989         error = inode_permission(dir->dentry->d_inode, MAY_WRITE | MAY_EXEC);
2990         if (error)
2991                 return error;
2992
2993         return security_inode_create(dir->dentry->d_inode, dentry, mode);
2994 }
2995
2996 /*
2997  * Attempt to atomically look up, create and open a file from a negative
2998  * dentry.
2999  *
3000  * Returns 0 if successful.  The file will have been created and attached to
3001  * @file by the filesystem calling finish_open().
3002  *
3003  * Returns 1 if the file was looked up only or didn't need creating.  The
3004  * caller will need to perform the open themselves.  @path will have been
3005  * updated to point to the new dentry.  This may be negative.
3006  *
3007  * Returns an error code otherwise.
3008  */
3009 static int atomic_open(struct nameidata *nd, struct dentry *dentry,
3010                         struct path *path, struct file *file,
3011                         const struct open_flags *op,
3012                         int open_flag, umode_t mode,
3013                         int *opened)
3014 {
3015         struct dentry *const DENTRY_NOT_SET = (void *) -1UL;
3016         struct inode *dir =  nd->path.dentry->d_inode;
3017         int error;
3018
3019         if (!(~open_flag & (O_EXCL | O_CREAT))) /* both O_EXCL and O_CREAT */
3020                 open_flag &= ~O_TRUNC;
3021
3022         if (nd->flags & LOOKUP_DIRECTORY)
3023                 open_flag |= O_DIRECTORY;
3024
3025         file->f_path.dentry = DENTRY_NOT_SET;
3026         file->f_path.mnt = nd->path.mnt;
3027         error = dir->i_op->atomic_open(dir, dentry, file,
3028                                        open_to_namei_flags(open_flag),
3029                                        mode, opened);
3030         d_lookup_done(dentry);
3031         if (!error) {
3032                 /*
3033                  * We didn't have the inode before the open, so check open
3034                  * permission here.
3035                  */
3036                 int acc_mode = op->acc_mode;
3037                 if (*opened & FILE_CREATED) {
3038                         WARN_ON(!(open_flag & O_CREAT));
3039                         fsnotify_create(dir, dentry);
3040                         acc_mode = 0;
3041                 }
3042                 error = may_open(&file->f_path, acc_mode, open_flag);
3043                 if (WARN_ON(error > 0))
3044                         error = -EINVAL;
3045         } else if (error > 0) {
3046                 if (WARN_ON(file->f_path.dentry == DENTRY_NOT_SET)) {
3047                         error = -EIO;
3048                 } else {
3049                         if (file->f_path.dentry) {
3050                                 dput(dentry);
3051                                 dentry = file->f_path.dentry;
3052                         }
3053                         if (*opened & FILE_CREATED)
3054                                 fsnotify_create(dir, dentry);
3055                         if (unlikely(d_is_negative(dentry))) {
3056                                 error = -ENOENT;
3057                         } else {
3058                                 path->dentry = dentry;
3059                                 path->mnt = nd->path.mnt;
3060                                 return 1;
3061                         }
3062                 }
3063         }
3064         dput(dentry);
3065         return error;
3066 }
3067
3068 /*
3069  * Look up and maybe create and open the last component.
3070  *
3071  * Must be called with i_mutex held on parent.
3072  *
3073  * Returns 0 if the file was successfully atomically created (if necessary) and
3074  * opened.  In this case the file will be returned attached to @file.
3075  *
3076  * Returns 1 if the file was not completely opened at this time, though lookups
3077  * and creations will have been performed and the dentry returned in @path will
3078  * be positive upon return if O_CREAT was specified.  If O_CREAT wasn't
3079  * specified then a negative dentry may be returned.
3080  *
3081  * An error code is returned otherwise.
3082  *
3083  * FILE_CREATE will be set in @*opened if the dentry was created and will be
3084  * cleared otherwise prior to returning.
3085  */
3086 static int lookup_open(struct nameidata *nd, struct path *path,
3087                         struct file *file,
3088                         const struct open_flags *op,
3089                         bool got_write, int *opened)
3090 {
3091         struct dentry *dir = nd->path.dentry;
3092         struct inode *dir_inode = dir->d_inode;
3093         int open_flag = op->open_flag;
3094         struct dentry *dentry;
3095         int error, create_error = 0;
3096         umode_t mode = op->mode;
3097         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
3098
3099         if (unlikely(IS_DEADDIR(dir_inode)))
3100                 return -ENOENT;
3101
3102         *opened &= ~FILE_CREATED;
3103         dentry = d_lookup(dir, &nd->last);
3104         for (;;) {
3105                 if (!dentry) {
3106                         dentry = d_alloc_parallel(dir, &nd->last, &wq);
3107                         if (IS_ERR(dentry))
3108                                 return PTR_ERR(dentry);
3109                 }
3110                 if (d_in_lookup(dentry))
3111                         break;
3112
3113                 error = d_revalidate(dentry, nd->flags);
3114                 if (likely(error > 0))
3115                         break;
3116                 if (error)
3117                         goto out_dput;
3118                 d_invalidate(dentry);
3119                 dput(dentry);
3120                 dentry = NULL;
3121         }
3122         if (dentry->d_inode) {
3123                 /* Cached positive dentry: will open in f_op->open */
3124                 goto out_no_open;
3125         }
3126
3127         /*
3128          * Checking write permission is tricky, bacuse we don't know if we are
3129          * going to actually need it: O_CREAT opens should work as long as the
3130          * file exists.  But checking existence breaks atomicity.  The trick is
3131          * to check access and if not granted clear O_CREAT from the flags.
3132          *
3133          * Another problem is returing the "right" error value (e.g. for an
3134          * O_EXCL open we want to return EEXIST not EROFS).
3135          */
3136         if (open_flag & O_CREAT) {
3137                 if (!IS_POSIXACL(dir->d_inode))
3138                         mode &= ~current_umask();
3139                 if (unlikely(!got_write)) {
3140                         create_error = -EROFS;
3141                         open_flag &= ~O_CREAT;
3142                         if (open_flag & (O_EXCL | O_TRUNC))
3143                                 goto no_open;
3144                         /* No side effects, safe to clear O_CREAT */
3145                 } else {
3146                         create_error = may_o_create(&nd->path, dentry, mode);
3147                         if (create_error) {
3148                                 open_flag &= ~O_CREAT;
3149                                 if (open_flag & O_EXCL)
3150                                         goto no_open;
3151                         }
3152                 }
3153         } else if ((open_flag & (O_TRUNC|O_WRONLY|O_RDWR)) &&
3154                    unlikely(!got_write)) {
3155                 /*
3156                  * No O_CREATE -> atomicity not a requirement -> fall
3157                  * back to lookup + open
3158                  */
3159                 goto no_open;
3160         }
3161
3162         if (dir_inode->i_op->atomic_open) {
3163                 error = atomic_open(nd, dentry, path, file, op, open_flag,
3164                                     mode, opened);
3165                 if (unlikely(error == -ENOENT) && create_error)
3166                         error = create_error;
3167                 return error;
3168         }
3169
3170 no_open:
3171         if (d_in_lookup(dentry)) {
3172                 struct dentry *res = dir_inode->i_op->lookup(dir_inode, dentry,
3173                                                              nd->flags);
3174                 d_lookup_done(dentry);
3175                 if (unlikely(res)) {
3176                         if (IS_ERR(res)) {
3177                                 error = PTR_ERR(res);
3178                                 goto out_dput;
3179                         }
3180                         dput(dentry);
3181                         dentry = res;
3182                 }
3183         }
3184
3185         /* Negative dentry, just create the file */
3186         if (!dentry->d_inode && (open_flag & O_CREAT)) {
3187                 *opened |= FILE_CREATED;
3188                 audit_inode_child(dir_inode, dentry, AUDIT_TYPE_CHILD_CREATE);
3189                 if (!dir_inode->i_op->create) {
3190                         error = -EACCES;
3191                         goto out_dput;
3192                 }
3193                 error = dir_inode->i_op->create(dir_inode, dentry, mode,
3194                                                 open_flag & O_EXCL);
3195                 if (error)
3196                         goto out_dput;
3197                 fsnotify_create(dir_inode, dentry);
3198         }
3199         if (unlikely(create_error) && !dentry->d_inode) {
3200                 error = create_error;
3201                 goto out_dput;
3202         }
3203 out_no_open:
3204         path->dentry = dentry;
3205         path->mnt = nd->path.mnt;
3206         return 1;
3207
3208 out_dput:
3209         dput(dentry);
3210         return error;
3211 }
3212
3213 /*
3214  * Handle the last step of open()
3215  */
3216 static int do_last(struct nameidata *nd,
3217                    struct file *file, const struct open_flags *op,
3218                    int *opened)
3219 {
3220         struct dentry *dir = nd->path.dentry;
3221         int open_flag = op->open_flag;
3222         bool will_truncate = (open_flag & O_TRUNC) != 0;
3223         bool got_write = false;
3224         int acc_mode = op->acc_mode;
3225         unsigned seq;
3226         struct inode *inode;
3227         struct path path;
3228         int error;
3229
3230         nd->flags &= ~LOOKUP_PARENT;
3231         nd->flags |= op->intent;
3232
3233         if (nd->last_type != LAST_NORM) {
3234                 error = handle_dots(nd, nd->last_type);
3235                 if (unlikely(error))
3236                         return error;
3237                 goto finish_open;
3238         }
3239
3240         if (!(open_flag & O_CREAT)) {
3241                 if (nd->last.name[nd->last.len])
3242                         nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
3243                 /* we _can_ be in RCU mode here */
3244                 error = lookup_fast(nd, &path, &inode, &seq);
3245                 if (likely(error > 0))
3246                         goto finish_lookup;
3247
3248                 if (error < 0)
3249                         return error;
3250
3251                 BUG_ON(nd->inode != dir->d_inode);
3252                 BUG_ON(nd->flags & LOOKUP_RCU);
3253         } else {
3254                 /* create side of things */
3255                 /*
3256                  * This will *only* deal with leaving RCU mode - LOOKUP_JUMPED
3257                  * has been cleared when we got to the last component we are
3258                  * about to look up
3259                  */
3260                 error = complete_walk(nd);
3261                 if (error)
3262                         return error;
3263
3264                 audit_inode(nd->name, dir, LOOKUP_PARENT);
3265                 /* trailing slashes? */
3266                 if (unlikely(nd->last.name[nd->last.len]))
3267                         return -EISDIR;
3268         }
3269
3270         if (open_flag & (O_CREAT | O_TRUNC | O_WRONLY | O_RDWR)) {
3271                 error = mnt_want_write(nd->path.mnt);
3272                 if (!error)
3273                         got_write = true;
3274                 /*
3275                  * do _not_ fail yet - we might not need that or fail with
3276                  * a different error; let lookup_open() decide; we'll be
3277                  * dropping this one anyway.
3278                  */
3279         }
3280         if (open_flag & O_CREAT)
3281                 inode_lock(dir->d_inode);
3282         else
3283                 inode_lock_shared(dir->d_inode);
3284         error = lookup_open(nd, &path, file, op, got_write, opened);
3285         if (open_flag & O_CREAT)
3286                 inode_unlock(dir->d_inode);
3287         else
3288                 inode_unlock_shared(dir->d_inode);
3289
3290         if (error <= 0) {
3291                 if (error)
3292                         goto out;
3293
3294                 if ((*opened & FILE_CREATED) ||
3295                     !S_ISREG(file_inode(file)->i_mode))
3296                         will_truncate = false;
3297
3298                 audit_inode(nd->name, file->f_path.dentry, 0);
3299                 goto opened;
3300         }
3301
3302         if (*opened & FILE_CREATED) {
3303                 /* Don't check for write permission, don't truncate */
3304                 open_flag &= ~O_TRUNC;
3305                 will_truncate = false;
3306                 acc_mode = 0;
3307                 path_to_nameidata(&path, nd);
3308                 goto finish_open_created;
3309         }
3310
3311         /*
3312          * If atomic_open() acquired write access it is dropped now due to
3313          * possible mount and symlink following (this might be optimized away if
3314          * necessary...)
3315          */
3316         if (got_write) {
3317                 mnt_drop_write(nd->path.mnt);
3318                 got_write = false;
3319         }
3320
3321         error = follow_managed(&path, nd);
3322         if (unlikely(error < 0))
3323                 return error;
3324
3325         if (unlikely(d_is_negative(path.dentry))) {
3326                 path_to_nameidata(&path, nd);
3327                 return -ENOENT;
3328         }
3329
3330         /*
3331          * create/update audit record if it already exists.
3332          */
3333         audit_inode(nd->name, path.dentry, 0);
3334
3335         if (unlikely((open_flag & (O_EXCL | O_CREAT)) == (O_EXCL | O_CREAT))) {
3336                 path_to_nameidata(&path, nd);
3337                 return -EEXIST;
3338         }
3339
3340         seq = 0;        /* out of RCU mode, so the value doesn't matter */
3341         inode = d_backing_inode(path.dentry);
3342 finish_lookup:
3343         error = step_into(nd, &path, 0, inode, seq);
3344         if (unlikely(error))
3345                 return error;
3346 finish_open:
3347         /* Why this, you ask?  _Now_ we might have grown LOOKUP_JUMPED... */
3348         error = complete_walk(nd);
3349         if (error)
3350                 return error;
3351         audit_inode(nd->name, nd->path.dentry, 0);
3352         error = -EISDIR;
3353         if ((open_flag & O_CREAT) && d_is_dir(nd->path.dentry))
3354                 goto out;
3355         error = -ENOTDIR;
3356         if ((nd->flags & LOOKUP_DIRECTORY) && !d_can_lookup(nd->path.dentry))
3357                 goto out;
3358         if (!d_is_reg(nd->path.dentry))
3359                 will_truncate = false;
3360
3361         if (will_truncate) {
3362                 error = mnt_want_write(nd->path.mnt);
3363                 if (error)
3364                         goto out;
3365                 got_write = true;
3366         }
3367 finish_open_created:
3368         error = may_open(&nd->path, acc_mode, open_flag);
3369         if (error)
3370                 goto out;
3371         BUG_ON(*opened & FILE_OPENED); /* once it's opened, it's opened */
3372         error = vfs_open(&nd->path, file, current_cred());
3373         if (error)
3374                 goto out;
3375         *opened |= FILE_OPENED;
3376 opened:
3377         error = open_check_o_direct(file);
3378         if (!error)
3379                 error = ima_file_check(file, op->acc_mode, *opened);
3380         if (!error && will_truncate)
3381                 error = handle_truncate(file);
3382 out:
3383         if (unlikely(error) && (*opened & FILE_OPENED))
3384                 fput(file);
3385         if (unlikely(error > 0)) {
3386                 WARN_ON(1);
3387                 error = -EINVAL;
3388         }
3389         if (got_write)
3390                 mnt_drop_write(nd->path.mnt);
3391         return error;
3392 }
3393
3394 struct dentry *vfs_tmpfile(struct dentry *dentry, umode_t mode, int open_flag)
3395 {
3396         struct dentry *child = NULL;
3397         struct inode *dir = dentry->d_inode;
3398         struct inode *inode;
3399         int error;
3400
3401         /* we want directory to be writable */
3402         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
3403         if (error)
3404                 goto out_err;
3405         error = -EOPNOTSUPP;
3406         if (!dir->i_op->tmpfile)
3407                 goto out_err;
3408         error = -ENOMEM;
3409         child = d_alloc(dentry, &slash_name);
3410         if (unlikely(!child))
3411                 goto out_err;
3412         error = dir->i_op->tmpfile(dir, child, mode);
3413         if (error)
3414                 goto out_err;
3415         error = -ENOENT;
3416         inode = child->d_inode;
3417         if (unlikely(!inode))
3418                 goto out_err;
3419         if (!(open_flag & O_EXCL)) {
3420                 spin_lock(&inode->i_lock);
3421                 inode->i_state |= I_LINKABLE;
3422                 spin_unlock(&inode->i_lock);
3423         }
3424         return child;
3425
3426 out_err:
3427         dput(child);
3428         return ERR_PTR(error);
3429 }
3430 EXPORT_SYMBOL(vfs_tmpfile);
3431
3432 static int do_tmpfile(struct nameidata *nd, unsigned flags,
3433                 const struct open_flags *op,
3434                 struct file *file, int *opened)
3435 {
3436         struct dentry *child;
3437         struct path path;
3438         int error = path_lookupat(nd, flags | LOOKUP_DIRECTORY, &path);
3439         if (unlikely(error))
3440                 return error;
3441         error = mnt_want_write(path.mnt);
3442         if (unlikely(error))
3443                 goto out;
3444         child = vfs_tmpfile(path.dentry, op->mode, op->open_flag);
3445         error = PTR_ERR(child);
3446         if (IS_ERR(child))
3447                 goto out2;
3448         dput(path.dentry);
3449         path.dentry = child;
3450         audit_inode(nd->name, child, 0);
3451         /* Don't check for other permissions, the inode was just created */
3452         error = may_open(&path, 0, op->open_flag);
3453         if (error)
3454                 goto out2;
3455         file->f_path.mnt = path.mnt;
3456         error = finish_open(file, child, NULL, opened);
3457         if (error)
3458                 goto out2;
3459         error = open_check_o_direct(file);
3460         if (error)
3461                 fput(file);
3462 out2:
3463         mnt_drop_write(path.mnt);
3464 out:
3465         path_put(&path);
3466         return error;
3467 }
3468
3469 static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
3470 {
3471         struct path path;
3472         int error = path_lookupat(nd, flags, &path);
3473         if (!error) {
3474                 audit_inode(nd->name, path.dentry, 0);
3475                 error = vfs_open(&path, file, current_cred());
3476                 path_put(&path);
3477         }
3478         return error;
3479 }
3480
3481 static struct file *path_openat(struct nameidata *nd,
3482                         const struct open_flags *op, unsigned flags)
3483 {
3484         const char *s;
3485         struct file *file;
3486         int opened = 0;
3487         int error;
3488
3489         file = get_empty_filp();
3490         if (IS_ERR(file))
3491                 return file;
3492
3493         file->f_flags = op->open_flag;
3494
3495         if (unlikely(file->f_flags & __O_TMPFILE)) {
3496                 error = do_tmpfile(nd, flags, op, file, &opened);
3497                 goto out2;
3498         }
3499
3500         if (unlikely(file->f_flags & O_PATH)) {
3501                 error = do_o_path(nd, flags, file);
3502                 if (!error)
3503                         opened |= FILE_OPENED;
3504                 goto out2;
3505         }
3506
3507         s = path_init(nd, flags);
3508         if (IS_ERR(s)) {
3509                 put_filp(file);
3510                 return ERR_CAST(s);
3511         }
3512         while (!(error = link_path_walk(s, nd)) &&
3513                 (error = do_last(nd, file, op, &opened)) > 0) {
3514                 nd->flags &= ~(LOOKUP_OPEN|LOOKUP_CREATE|LOOKUP_EXCL);
3515                 s = trailing_symlink(nd);
3516                 if (IS_ERR(s)) {
3517                         error = PTR_ERR(s);
3518                         break;
3519                 }
3520         }
3521         terminate_walk(nd);
3522 out2:
3523         if (!(opened & FILE_OPENED)) {
3524                 BUG_ON(!error);
3525                 put_filp(file);
3526         }
3527         if (unlikely(error)) {
3528                 if (error == -EOPENSTALE) {
3529                         if (flags & LOOKUP_RCU)
3530                                 error = -ECHILD;
3531                         else
3532                                 error = -ESTALE;
3533                 }
3534                 file = ERR_PTR(error);
3535         }
3536         return file;
3537 }
3538
3539 struct file *do_filp_open(int dfd, struct filename *pathname,
3540                 const struct open_flags *op)
3541 {
3542         struct nameidata nd;
3543         int flags = op->lookup_flags;
3544         struct file *filp;
3545
3546         set_nameidata(&nd, dfd, pathname);
3547         filp = path_openat(&nd, op, flags | LOOKUP_RCU);
3548         if (unlikely(filp == ERR_PTR(-ECHILD)))
3549                 filp = path_openat(&nd, op, flags);
3550         if (unlikely(filp == ERR_PTR(-ESTALE)))
3551                 filp = path_openat(&nd, op, flags | LOOKUP_REVAL);
3552         restore_nameidata();
3553         return filp;
3554 }
3555
3556 struct file *do_file_open_root(struct dentry *dentry, struct vfsmount *mnt,
3557                 const char *name, const struct open_flags *op)
3558 {
3559         struct nameidata nd;
3560         struct file *file;
3561         struct filename *filename;
3562         int flags = op->lookup_flags | LOOKUP_ROOT;
3563
3564         nd.root.mnt = mnt;
3565         nd.root.dentry = dentry;
3566
3567         if (d_is_symlink(dentry) && op->intent & LOOKUP_OPEN)
3568                 return ERR_PTR(-ELOOP);
3569
3570         filename = getname_kernel(name);
3571         if (IS_ERR(filename))
3572                 return ERR_CAST(filename);
3573
3574         set_nameidata(&nd, -1, filename);
3575         file = path_openat(&nd, op, flags | LOOKUP_RCU);
3576         if (unlikely(file == ERR_PTR(-ECHILD)))
3577                 file = path_openat(&nd, op, flags);
3578         if (unlikely(file == ERR_PTR(-ESTALE)))
3579                 file = path_openat(&nd, op, flags | LOOKUP_REVAL);
3580         restore_nameidata();
3581         putname(filename);
3582         return file;
3583 }
3584
3585 static struct dentry *filename_create(int dfd, struct filename *name,
3586                                 struct path *path, unsigned int lookup_flags)
3587 {
3588         struct dentry *dentry = ERR_PTR(-EEXIST);
3589         struct qstr last;
3590         int type;
3591         int err2;
3592         int error;
3593         bool is_dir = (lookup_flags & LOOKUP_DIRECTORY);
3594
3595         /*
3596          * Note that only LOOKUP_REVAL and LOOKUP_DIRECTORY matter here. Any
3597          * other flags passed in are ignored!
3598          */
3599         lookup_flags &= LOOKUP_REVAL;
3600
3601         name = filename_parentat(dfd, name, lookup_flags, path, &last, &type);
3602         if (IS_ERR(name))
3603                 return ERR_CAST(name);
3604
3605         /*
3606          * Yucky last component or no last component at all?
3607          * (foo/., foo/.., /////)
3608          */
3609         if (unlikely(type != LAST_NORM))
3610                 goto out;
3611
3612         /* don't fail immediately if it's r/o, at least try to report other errors */
3613         err2 = mnt_want_write(path->mnt);
3614         /*
3615          * Do the final lookup.
3616          */
3617         lookup_flags |= LOOKUP_CREATE | LOOKUP_EXCL;
3618         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
3619         dentry = __lookup_hash(&last, path->dentry, lookup_flags);
3620         if (IS_ERR(dentry))
3621                 goto unlock;
3622
3623         error = -EEXIST;
3624         if (d_is_positive(dentry))
3625                 goto fail;
3626
3627         /*
3628          * Special case - lookup gave negative, but... we had foo/bar/
3629          * From the vfs_mknod() POV we just have a negative dentry -
3630          * all is fine. Let's be bastards - you had / on the end, you've
3631          * been asking for (non-existent) directory. -ENOENT for you.
3632          */
3633         if (unlikely(!is_dir && last.name[last.len])) {
3634                 error = -ENOENT;
3635                 goto fail;
3636         }
3637         if (unlikely(err2)) {
3638                 error = err2;
3639                 goto fail;
3640         }
3641         putname(name);
3642         return dentry;
3643 fail:
3644         dput(dentry);
3645         dentry = ERR_PTR(error);
3646 unlock:
3647         inode_unlock(path->dentry->d_inode);
3648         if (!err2)
3649                 mnt_drop_write(path->mnt);
3650 out:
3651         path_put(path);
3652         putname(name);
3653         return dentry;
3654 }
3655
3656 struct dentry *kern_path_create(int dfd, const char *pathname,
3657                                 struct path *path, unsigned int lookup_flags)
3658 {
3659         return filename_create(dfd, getname_kernel(pathname),
3660                                 path, lookup_flags);
3661 }
3662 EXPORT_SYMBOL(kern_path_create);
3663
3664 void done_path_create(struct path *path, struct dentry *dentry)
3665 {
3666         dput(dentry);
3667         inode_unlock(path->dentry->d_inode);
3668         mnt_drop_write(path->mnt);
3669         path_put(path);
3670 }
3671 EXPORT_SYMBOL(done_path_create);
3672
3673 inline struct dentry *user_path_create(int dfd, const char __user *pathname,
3674                                 struct path *path, unsigned int lookup_flags)
3675 {
3676         return filename_create(dfd, getname(pathname), path, lookup_flags);
3677 }
3678 EXPORT_SYMBOL(user_path_create);
3679
3680 int vfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3681 {
3682         int error = may_create(dir, dentry);
3683
3684         if (error)
3685                 return error;
3686
3687         if ((S_ISCHR(mode) || S_ISBLK(mode)) && !capable(CAP_MKNOD))
3688                 return -EPERM;
3689
3690         if (!dir->i_op->mknod)
3691                 return -EPERM;
3692
3693         error = devcgroup_inode_mknod(mode, dev);
3694         if (error)
3695                 return error;
3696
3697         error = security_inode_mknod(dir, dentry, mode, dev);
3698         if (error)
3699                 return error;
3700
3701         error = dir->i_op->mknod(dir, dentry, mode, dev);
3702         if (!error)
3703                 fsnotify_create(dir, dentry);
3704         return error;
3705 }
3706 EXPORT_SYMBOL(vfs_mknod);
3707
3708 static int may_mknod(umode_t mode)
3709 {
3710         switch (mode & S_IFMT) {
3711         case S_IFREG:
3712         case S_IFCHR:
3713         case S_IFBLK:
3714         case S_IFIFO:
3715         case S_IFSOCK:
3716         case 0: /* zero mode translates to S_IFREG */
3717                 return 0;
3718         case S_IFDIR:
3719                 return -EPERM;
3720         default:
3721                 return -EINVAL;
3722         }
3723 }
3724
3725 long do_mknodat(int dfd, const char __user *filename, umode_t mode,
3726                 unsigned int dev)
3727 {
3728         struct dentry *dentry;
3729         struct path path;
3730         int error;
3731         unsigned int lookup_flags = 0;
3732
3733         error = may_mknod(mode);
3734         if (error)
3735                 return error;
3736 retry:
3737         dentry = user_path_create(dfd, filename, &path, lookup_flags);
3738         if (IS_ERR(dentry))
3739                 return PTR_ERR(dentry);
3740
3741         if (!IS_POSIXACL(path.dentry->d_inode))
3742                 mode &= ~current_umask();
3743         error = security_path_mknod(&path, dentry, mode, dev);
3744         if (error)
3745                 goto out;
3746         switch (mode & S_IFMT) {
3747                 case 0: case S_IFREG:
3748                         error = vfs_create(path.dentry->d_inode,dentry,mode,true);
3749                         if (!error)
3750                                 ima_post_path_mknod(dentry);
3751                         break;
3752                 case S_IFCHR: case S_IFBLK:
3753                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,
3754                                         new_decode_dev(dev));
3755                         break;
3756                 case S_IFIFO: case S_IFSOCK:
3757                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,0);
3758                         break;
3759         }
3760 out:
3761         done_path_create(&path, dentry);
3762         if (retry_estale(error, lookup_flags)) {
3763                 lookup_flags |= LOOKUP_REVAL;
3764                 goto retry;
3765         }
3766         return error;
3767 }
3768
3769 SYSCALL_DEFINE4(mknodat, int, dfd, const char __user *, filename, umode_t, mode,
3770                 unsigned int, dev)
3771 {
3772         return do_mknodat(dfd, filename, mode, dev);
3773 }
3774
3775 SYSCALL_DEFINE3(mknod, const char __user *, filename, umode_t, mode, unsigned, dev)
3776 {
3777         return do_mknodat(AT_FDCWD, filename, mode, dev);
3778 }
3779
3780 int vfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3781 {
3782         int error = may_create(dir, dentry);
3783         unsigned max_links = dir->i_sb->s_max_links;
3784
3785         if (error)
3786                 return error;
3787
3788         if (!dir->i_op->mkdir)
3789                 return -EPERM;
3790
3791         mode &= (S_IRWXUGO|S_ISVTX);
3792         error = security_inode_mkdir(dir, dentry, mode);
3793         if (error)
3794                 return error;
3795
3796         if (max_links && dir->i_nlink >= max_links)
3797                 return -EMLINK;
3798
3799         error = dir->i_op->mkdir(dir, dentry, mode);
3800         if (!error)
3801                 fsnotify_mkdir(dir, dentry);
3802         return error;
3803 }
3804 EXPORT_SYMBOL(vfs_mkdir);
3805
3806 long do_mkdirat(int dfd, const char __user *pathname, umode_t mode)
3807 {
3808         struct dentry *dentry;
3809         struct path path;
3810         int error;
3811         unsigned int lookup_flags = LOOKUP_DIRECTORY;
3812
3813 retry:
3814         dentry = user_path_create(dfd, pathname, &path, lookup_flags);
3815         if (IS_ERR(dentry))
3816                 return PTR_ERR(dentry);
3817
3818         if (!IS_POSIXACL(path.dentry->d_inode))
3819                 mode &= ~current_umask();
3820         error = security_path_mkdir(&path, dentry, mode);
3821         if (!error)
3822                 error = vfs_mkdir(path.dentry->d_inode, dentry, mode);
3823         done_path_create(&path, dentry);
3824         if (retry_estale(error, lookup_flags)) {
3825                 lookup_flags |= LOOKUP_REVAL;
3826                 goto retry;
3827         }
3828         return error;
3829 }
3830
3831 SYSCALL_DEFINE3(mkdirat, int, dfd, const char __user *, pathname, umode_t, mode)
3832 {
3833         return do_mkdirat(dfd, pathname, mode);
3834 }
3835
3836 SYSCALL_DEFINE2(mkdir, const char __user *, pathname, umode_t, mode)
3837 {
3838         return do_mkdirat(AT_FDCWD, pathname, mode);
3839 }
3840
3841 int vfs_rmdir(struct inode *dir, struct dentry *dentry)
3842 {
3843         int error = may_delete(dir, dentry, 1);
3844
3845         if (error)
3846                 return error;
3847
3848         if (!dir->i_op->rmdir)
3849                 return -EPERM;
3850
3851         dget(dentry);
3852         inode_lock(dentry->d_inode);
3853
3854         error = -EBUSY;
3855         if (is_local_mountpoint(dentry))
3856                 goto out;
3857
3858         error = security_inode_rmdir(dir, dentry);
3859         if (error)
3860                 goto out;
3861
3862         shrink_dcache_parent(dentry);
3863         error = dir->i_op->rmdir(dir, dentry);
3864         if (error)
3865                 goto out;
3866
3867         dentry->d_inode->i_flags |= S_DEAD;
3868         dont_mount(dentry);
3869         detach_mounts(dentry);
3870
3871 out:
3872         inode_unlock(dentry->d_inode);
3873         dput(dentry);
3874         if (!error)
3875                 d_delete(dentry);
3876         return error;
3877 }
3878 EXPORT_SYMBOL(vfs_rmdir);
3879
3880 long do_rmdir(int dfd, const char __user *pathname)
3881 {
3882         int error = 0;
3883         struct filename *name;
3884         struct dentry *dentry;
3885         struct path path;
3886         struct qstr last;
3887         int type;
3888         unsigned int lookup_flags = 0;
3889 retry:
3890         name = filename_parentat(dfd, getname(pathname), lookup_flags,
3891                                 &path, &last, &type);
3892         if (IS_ERR(name))
3893                 return PTR_ERR(name);
3894
3895         switch (type) {
3896         case LAST_DOTDOT:
3897                 error = -ENOTEMPTY;
3898                 goto exit1;
3899         case LAST_DOT:
3900                 error = -EINVAL;
3901                 goto exit1;
3902         case LAST_ROOT:
3903                 error = -EBUSY;
3904                 goto exit1;
3905         }
3906
3907         error = mnt_want_write(path.mnt);
3908         if (error)
3909                 goto exit1;
3910
3911         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3912         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3913         error = PTR_ERR(dentry);
3914         if (IS_ERR(dentry))
3915                 goto exit2;
3916         if (!dentry->d_inode) {
3917                 error = -ENOENT;
3918                 goto exit3;
3919         }
3920         error = security_path_rmdir(&path, dentry);
3921         if (error)
3922                 goto exit3;
3923         error = vfs_rmdir(path.dentry->d_inode, dentry);
3924 exit3:
3925         dput(dentry);
3926 exit2:
3927         inode_unlock(path.dentry->d_inode);
3928         mnt_drop_write(path.mnt);
3929 exit1:
3930         path_put(&path);
3931         putname(name);
3932         if (retry_estale(error, lookup_flags)) {
3933                 lookup_flags |= LOOKUP_REVAL;
3934                 goto retry;
3935         }
3936         return error;
3937 }
3938
3939 SYSCALL_DEFINE1(rmdir, const char __user *, pathname)
3940 {
3941         return do_rmdir(AT_FDCWD, pathname);
3942 }
3943
3944 /**
3945  * vfs_unlink - unlink a filesystem object
3946  * @dir:        parent directory
3947  * @dentry:     victim
3948  * @delegated_inode: returns victim inode, if the inode is delegated.
3949  *
3950  * The caller must hold dir->i_mutex.
3951  *
3952  * If vfs_unlink discovers a delegation, it will return -EWOULDBLOCK and
3953  * return a reference to the inode in delegated_inode.  The caller
3954  * should then break the delegation on that inode and retry.  Because
3955  * breaking a delegation may take a long time, the caller should drop
3956  * dir->i_mutex before doing so.
3957  *
3958  * Alternatively, a caller may pass NULL for delegated_inode.  This may
3959  * be appropriate for callers that expect the underlying filesystem not
3960  * to be NFS exported.
3961  */
3962 int vfs_unlink(struct inode *dir, struct dentry *dentry, struct inode **delegated_inode)
3963 {
3964         struct inode *target = dentry->d_inode;
3965         int error = may_delete(dir, dentry, 0);
3966
3967         if (error)
3968                 return error;
3969
3970         if (!dir->i_op->unlink)
3971                 return -EPERM;
3972
3973         inode_lock(target);
3974         if (is_local_mountpoint(dentry))
3975                 error = -EBUSY;
3976         else {
3977                 error = security_inode_unlink(dir, dentry);
3978                 if (!error) {
3979                         error = try_break_deleg(target, delegated_inode);
3980                         if (error)
3981                                 goto out;
3982                         error = dir->i_op->unlink(dir, dentry);
3983                         if (!error) {
3984                                 dont_mount(dentry);
3985                                 detach_mounts(dentry);
3986                         }
3987                 }
3988         }
3989 out:
3990         inode_unlock(target);
3991
3992         /* We don't d_delete() NFS sillyrenamed files--they still exist. */
3993         if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
3994                 fsnotify_link_count(target);
3995                 d_delete(dentry);
3996         }
3997
3998         return error;
3999 }
4000 EXPORT_SYMBOL(vfs_unlink);
4001
4002 /*
4003  * Make sure that the actual truncation of the file will occur outside its
4004  * directory's i_mutex.  Truncate can take a long time if there is a lot of
4005  * writeout happening, and we don't want to prevent access to the directory
4006  * while waiting on the I/O.
4007  */
4008 long do_unlinkat(int dfd, struct filename *name)
4009 {
4010         int error;
4011         struct dentry *dentry;
4012         struct path path;
4013         struct qstr last;
4014         int type;
4015         struct inode *inode = NULL;
4016         struct inode *delegated_inode = NULL;
4017         unsigned int lookup_flags = 0;
4018 retry:
4019         name = filename_parentat(dfd, name, lookup_flags, &path, &last, &type);
4020         if (IS_ERR(name))
4021                 return PTR_ERR(name);
4022
4023         error = -EISDIR;
4024         if (type != LAST_NORM)
4025                 goto exit1;
4026
4027         error = mnt_want_write(path.mnt);
4028         if (error)
4029                 goto exit1;
4030 retry_deleg:
4031         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
4032         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
4033         error = PTR_ERR(dentry);
4034         if (!IS_ERR(dentry)) {
4035                 /* Why not before? Because we want correct error value */
4036                 if (last.name[last.len])
4037                         goto slashes;
4038                 inode = dentry->d_inode;
4039                 if (d_is_negative(dentry))
4040                         goto slashes;
4041                 ihold(inode);
4042                 error = security_path_unlink(&path, dentry);
4043                 if (error)
4044                         goto exit2;
4045                 error = vfs_unlink(path.dentry->d_inode, dentry, &delegated_inode);
4046 exit2:
4047                 dput(dentry);
4048         }
4049         inode_unlock(path.dentry->d_inode);
4050         if (inode)
4051                 iput(inode);    /* truncate the inode here */
4052         inode = NULL;
4053         if (delegated_inode) {
4054                 error = break_deleg_wait(&delegated_inode);
4055                 if (!error)
4056                         goto retry_deleg;
4057         }
4058         mnt_drop_write(path.mnt);
4059 exit1:
4060         path_put(&path);
4061         if (retry_estale(error, lookup_flags)) {
4062                 lookup_flags |= LOOKUP_REVAL;
4063                 inode = NULL;
4064                 goto retry;
4065         }
4066         putname(name);
4067         return error;
4068
4069 slashes:
4070         if (d_is_negative(dentry))
4071                 error = -ENOENT;
4072         else if (d_is_dir(dentry))
4073                 error = -EISDIR;
4074         else
4075                 error = -ENOTDIR;
4076         goto exit2;
4077 }
4078
4079 SYSCALL_DEFINE3(unlinkat, int, dfd, const char __user *, pathname, int, flag)
4080 {
4081         if ((flag & ~AT_REMOVEDIR) != 0)
4082                 return -EINVAL;
4083
4084         if (flag & AT_REMOVEDIR)
4085                 return do_rmdir(dfd, pathname);
4086
4087         return do_unlinkat(dfd, getname(pathname));
4088 }
4089
4090 SYSCALL_DEFINE1(unlink, const char __user *, pathname)
4091 {
4092         return do_unlinkat(AT_FDCWD, getname(pathname));
4093 }
4094
4095 int vfs_symlink(struct inode *dir, struct dentry *dentry, const char *oldname)
4096 {
4097         int error = may_create(dir, dentry);
4098
4099         if (error)
4100                 return error;
4101
4102         if (!dir->i_op->symlink)
4103                 return -EPERM;
4104
4105         error = security_inode_symlink(dir, dentry, oldname);
4106         if (error)
4107                 return error;
4108
4109         error = dir->i_op->symlink(dir, dentry, oldname);
4110         if (!error)
4111                 fsnotify_create(dir, dentry);
4112         return error;
4113 }
4114 EXPORT_SYMBOL(vfs_symlink);
4115
4116 long do_symlinkat(const char __user *oldname, int newdfd,
4117                   const char __user *newname)
4118 {
4119         int error;
4120         struct filename *from;
4121         struct dentry *dentry;
4122         struct path path;
4123         unsigned int lookup_flags = 0;
4124
4125         from = getname(oldname);
4126         if (IS_ERR(from))
4127                 return PTR_ERR(from);
4128 retry:
4129         dentry = user_path_create(newdfd, newname, &path, lookup_flags);
4130         error = PTR_ERR(dentry);
4131         if (IS_ERR(dentry))
4132                 goto out_putname;
4133
4134         error = security_path_symlink(&path, dentry, from->name);
4135         if (!error)
4136                 error = vfs_symlink(path.dentry->d_inode, dentry, from->name);
4137         done_path_create(&path, dentry);
4138         if (retry_estale(error, lookup_flags)) {
4139                 lookup_flags |= LOOKUP_REVAL;
4140                 goto retry;
4141         }
4142 out_putname:
4143         putname(from);
4144         return error;
4145 }
4146
4147 SYSCALL_DEFINE3(symlinkat, const char __user *, oldname,
4148                 int, newdfd, const char __user *, newname)
4149 {
4150         return do_symlinkat(oldname, newdfd, newname);
4151 }
4152
4153 SYSCALL_DEFINE2(symlink, const char __user *, oldname, const char __user *, newname)
4154 {
4155         return do_symlinkat(oldname, AT_FDCWD, newname);
4156 }
4157
4158 /**
4159  * vfs_link - create a new link
4160  * @old_dentry: object to be linked
4161  * @dir:        new parent
4162  * @new_dentry: where to create the new link
4163  * @delegated_inode: returns inode needing a delegation break
4164  *
4165  * The caller must hold dir->i_mutex
4166  *
4167  * If vfs_link discovers a delegation on the to-be-linked file in need
4168  * of breaking, it will return -EWOULDBLOCK and return a reference to the
4169  * inode in delegated_inode.  The caller should then break the delegation
4170  * and retry.  Because breaking a delegation may take a long time, the
4171  * caller should drop the i_mutex before doing so.
4172  *
4173  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4174  * be appropriate for callers that expect the underlying filesystem not
4175  * to be NFS exported.
4176  */
4177 int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_dentry, struct inode **delegated_inode)
4178 {
4179         struct inode *inode = old_dentry->d_inode;
4180         unsigned max_links = dir->i_sb->s_max_links;
4181         int error;
4182
4183         if (!inode)
4184                 return -ENOENT;
4185
4186         error = may_create(dir, new_dentry);
4187         if (error)
4188                 return error;
4189
4190         if (dir->i_sb != inode->i_sb)
4191                 return -EXDEV;
4192
4193         /*
4194          * A link to an append-only or immutable file cannot be created.
4195          */
4196         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4197                 return -EPERM;
4198         /*
4199          * Updating the link count will likely cause i_uid and i_gid to
4200          * be writen back improperly if their true value is unknown to
4201          * the vfs.
4202          */
4203         if (HAS_UNMAPPED_ID(inode))
4204                 return -EPERM;
4205         if (!dir->i_op->link)
4206                 return -EPERM;
4207         if (S_ISDIR(inode->i_mode))
4208                 return -EPERM;
4209
4210         error = security_inode_link(old_dentry, dir, new_dentry);
4211         if (error)
4212                 return error;
4213
4214         inode_lock(inode);
4215         /* Make sure we don't allow creating hardlink to an unlinked file */
4216         if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))
4217                 error =  -ENOENT;
4218         else if (max_links && inode->i_nlink >= max_links)
4219                 error = -EMLINK;
4220         else {
4221                 error = try_break_deleg(inode, delegated_inode);
4222                 if (!error)
4223                         error = dir->i_op->link(old_dentry, dir, new_dentry);
4224         }
4225
4226         if (!error && (inode->i_state & I_LINKABLE)) {
4227                 spin_lock(&inode->i_lock);
4228                 inode->i_state &= ~I_LINKABLE;
4229                 spin_unlock(&inode->i_lock);
4230         }
4231         inode_unlock(inode);
4232         if (!error)
4233                 fsnotify_link(dir, inode, new_dentry);
4234         return error;
4235 }
4236 EXPORT_SYMBOL(vfs_link);
4237
4238 /*
4239  * Hardlinks are often used in delicate situations.  We avoid
4240  * security-related surprises by not following symlinks on the
4241  * newname.  --KAB
4242  *
4243  * We don't follow them on the oldname either to be compatible
4244  * with linux 2.0, and to avoid hard-linking to directories
4245  * and other special files.  --ADM
4246  */
4247 int do_linkat(int olddfd, const char __user *oldname, int newdfd,
4248               const char __user *newname, int flags)
4249 {
4250         struct dentry *new_dentry;
4251         struct path old_path, new_path;
4252         struct inode *delegated_inode = NULL;
4253         int how = 0;
4254         int error;
4255
4256         if ((flags & ~(AT_SYMLINK_FOLLOW | AT_EMPTY_PATH)) != 0)
4257                 return -EINVAL;
4258         /*
4259          * To use null names we require CAP_DAC_READ_SEARCH
4260          * This ensures that not everyone will be able to create
4261          * handlink using the passed filedescriptor.
4262          */
4263         if (flags & AT_EMPTY_PATH) {
4264                 if (!capable(CAP_DAC_READ_SEARCH))
4265                         return -ENOENT;
4266                 how = LOOKUP_EMPTY;
4267         }
4268
4269         if (flags & AT_SYMLINK_FOLLOW)
4270                 how |= LOOKUP_FOLLOW;
4271 retry:
4272         error = user_path_at(olddfd, oldname, how, &old_path);
4273         if (error)
4274                 return error;
4275
4276         new_dentry = user_path_create(newdfd, newname, &new_path,
4277                                         (how & LOOKUP_REVAL));
4278         error = PTR_ERR(new_dentry);
4279         if (IS_ERR(new_dentry))
4280                 goto out;
4281
4282         error = -EXDEV;
4283         if (old_path.mnt != new_path.mnt)
4284                 goto out_dput;
4285         error = may_linkat(&old_path);
4286         if (unlikely(error))
4287                 goto out_dput;
4288         error = security_path_link(old_path.dentry, &new_path, new_dentry);
4289         if (error)
4290                 goto out_dput;
4291         error = vfs_link(old_path.dentry, new_path.dentry->d_inode, new_dentry, &delegated_inode);
4292 out_dput:
4293         done_path_create(&new_path, new_dentry);
4294         if (delegated_inode) {
4295                 error = break_deleg_wait(&delegated_inode);
4296                 if (!error) {
4297                         path_put(&old_path);
4298                         goto retry;
4299                 }
4300         }
4301         if (retry_estale(error, how)) {
4302                 path_put(&old_path);
4303                 how |= LOOKUP_REVAL;
4304                 goto retry;
4305         }
4306 out:
4307         path_put(&old_path);
4308
4309         return error;
4310 }
4311
4312 SYSCALL_DEFINE5(linkat, int, olddfd, const char __user *, oldname,
4313                 int, newdfd, const char __user *, newname, int, flags)
4314 {
4315         return do_linkat(olddfd, oldname, newdfd, newname, flags);
4316 }
4317
4318 SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname)
4319 {
4320         return do_linkat(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4321 }
4322
4323 /**
4324  * vfs_rename - rename a filesystem object
4325  * @old_dir:    parent of source
4326  * @old_dentry: source
4327  * @new_dir:    parent of destination
4328  * @new_dentry: destination
4329  * @delegated_inode: returns an inode needing a delegation break
4330  * @flags:      rename flags
4331  *
4332  * The caller must hold multiple mutexes--see lock_rename()).
4333  *
4334  * If vfs_rename discovers a delegation in need of breaking at either
4335  * the source or destination, it will return -EWOULDBLOCK and return a
4336  * reference to the inode in delegated_inode.  The caller should then
4337  * break the delegation and retry.  Because breaking a delegation may
4338  * take a long time, the caller should drop all locks before doing
4339  * so.
4340  *
4341  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4342  * be appropriate for callers that expect the underlying filesystem not
4343  * to be NFS exported.
4344  *
4345  * The worst of all namespace operations - renaming directory. "Perverted"
4346  * doesn't even start to describe it. Somebody in UCB had a heck of a trip...
4347  * Problems:
4348  *
4349  *      a) we can get into loop creation.
4350  *      b) race potential - two innocent renames can create a loop together.
4351  *         That's where 4.4 screws up. Current fix: serialization on
4352  *         sb->s_vfs_rename_mutex. We might be more accurate, but that's another
4353  *         story.
4354  *      c) we have to lock _four_ objects - parents and victim (if it exists),
4355  *         and source (if it is not a directory).
4356  *         And that - after we got ->i_mutex on parents (until then we don't know
4357  *         whether the target exists).  Solution: try to be smart with locking
4358  *         order for inodes.  We rely on the fact that tree topology may change
4359  *         only under ->s_vfs_rename_mutex _and_ that parent of the object we
4360  *         move will be locked.  Thus we can rank directories by the tree
4361  *         (ancestors first) and rank all non-directories after them.
4362  *         That works since everybody except rename does "lock parent, lookup,
4363  *         lock child" and rename is under ->s_vfs_rename_mutex.
4364  *         HOWEVER, it relies on the assumption that any object with ->lookup()
4365  *         has no more than 1 dentry.  If "hybrid" objects will ever appear,
4366  *         we'd better make sure that there's no link(2) for them.
4367  *      d) conversion from fhandle to dentry may come in the wrong moment - when
4368  *         we are removing the target. Solution: we will have to grab ->i_mutex
4369  *         in the fhandle_to_dentry code. [FIXME - current nfsfh.c relies on
4370  *         ->i_mutex on parents, which works but leads to some truly excessive
4371  *         locking].
4372  */
4373 int vfs_rename(struct inode *old_dir, struct dentry *old_dentry,
4374                struct inode *new_dir, struct dentry *new_dentry,
4375                struct inode **delegated_inode, unsigned int flags)
4376 {
4377         int error;
4378         bool is_dir = d_is_dir(old_dentry);
4379         struct inode *source = old_dentry->d_inode;
4380         struct inode *target = new_dentry->d_inode;
4381         bool new_is_dir = false;
4382         unsigned max_links = new_dir->i_sb->s_max_links;
4383         struct name_snapshot old_name;
4384
4385         if (source == target)
4386                 return 0;
4387
4388         error = may_delete(old_dir, old_dentry, is_dir);
4389         if (error)
4390                 return error;
4391
4392         if (!target) {
4393                 error = may_create(new_dir, new_dentry);
4394         } else {
4395                 new_is_dir = d_is_dir(new_dentry);
4396
4397                 if (!(flags & RENAME_EXCHANGE))
4398                         error = may_delete(new_dir, new_dentry, is_dir);
4399                 else
4400                         error = may_delete(new_dir, new_dentry, new_is_dir);
4401         }
4402         if (error)
4403                 return error;
4404
4405         if (!old_dir->i_op->rename)
4406                 return -EPERM;
4407
4408         /*
4409          * If we are going to change the parent - check write permissions,
4410          * we'll need to flip '..'.
4411          */
4412         if (new_dir != old_dir) {
4413                 if (is_dir) {
4414                         error = inode_permission(source, MAY_WRITE);
4415                         if (error)
4416                                 return error;
4417                 }
4418                 if ((flags & RENAME_EXCHANGE) && new_is_dir) {
4419                         error = inode_permission(target, MAY_WRITE);
4420                         if (error)
4421                                 return error;
4422                 }
4423         }
4424
4425         error = security_inode_rename(old_dir, old_dentry, new_dir, new_dentry,
4426                                       flags);
4427         if (error)
4428                 return error;
4429
4430         take_dentry_name_snapshot(&old_name, old_dentry);
4431         dget(new_dentry);
4432         if (!is_dir || (flags & RENAME_EXCHANGE))
4433                 lock_two_nondirectories(source, target);
4434         else if (target)
4435                 inode_lock(target);
4436
4437         error = -EBUSY;
4438         if (is_local_mountpoint(old_dentry) || is_local_mountpoint(new_dentry))
4439                 goto out;
4440
4441         if (max_links && new_dir != old_dir) {
4442                 error = -EMLINK;
4443                 if (is_dir && !new_is_dir && new_dir->i_nlink >= max_links)
4444                         goto out;
4445                 if ((flags & RENAME_EXCHANGE) && !is_dir && new_is_dir &&
4446                     old_dir->i_nlink >= max_links)
4447                         goto out;
4448         }
4449         if (is_dir && !(flags & RENAME_EXCHANGE) && target)
4450                 shrink_dcache_parent(new_dentry);
4451         if (!is_dir) {
4452                 error = try_break_deleg(source, delegated_inode);
4453                 if (error)
4454                         goto out;
4455         }
4456         if (target && !new_is_dir) {
4457                 error = try_break_deleg(target, delegated_inode);
4458                 if (error)
4459                         goto out;
4460         }
4461         error = old_dir->i_op->rename(old_dir, old_dentry,
4462                                        new_dir, new_dentry, flags);
4463         if (error)
4464                 goto out;
4465
4466         if (!(flags & RENAME_EXCHANGE) && target) {
4467                 if (is_dir)
4468                         target->i_flags |= S_DEAD;
4469                 dont_mount(new_dentry);
4470                 detach_mounts(new_dentry);
4471         }
4472         if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE)) {
4473                 if (!(flags & RENAME_EXCHANGE))
4474                         d_move(old_dentry, new_dentry);
4475                 else
4476                         d_exchange(old_dentry, new_dentry);
4477         }
4478 out:
4479         if (!is_dir || (flags & RENAME_EXCHANGE))
4480                 unlock_two_nondirectories(source, target);
4481         else if (target)
4482                 inode_unlock(target);
4483         dput(new_dentry);
4484         if (!error) {
4485                 fsnotify_move(old_dir, new_dir, old_name.name, is_dir,
4486                               !(flags & RENAME_EXCHANGE) ? target : NULL, old_dentry);
4487                 if (flags & RENAME_EXCHANGE) {
4488                         fsnotify_move(new_dir, old_dir, old_dentry->d_name.name,
4489                                       new_is_dir, NULL, new_dentry);
4490                 }
4491         }
4492         release_dentry_name_snapshot(&old_name);
4493
4494         return error;
4495 }
4496 EXPORT_SYMBOL(vfs_rename);
4497
4498 static int do_renameat2(int olddfd, const char __user *oldname, int newdfd,
4499                         const char __user *newname, unsigned int flags)
4500 {
4501         struct dentry *old_dentry, *new_dentry;
4502         struct dentry *trap;
4503         struct path old_path, new_path;
4504         struct qstr old_last, new_last;
4505         int old_type, new_type;
4506         struct inode *delegated_inode = NULL;
4507         struct filename *from;
4508         struct filename *to;
4509         unsigned int lookup_flags = 0, target_flags = LOOKUP_RENAME_TARGET;
4510         bool should_retry = false;
4511         int error;
4512
4513         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
4514                 return -EINVAL;
4515
4516         if ((flags & (RENAME_NOREPLACE | RENAME_WHITEOUT)) &&
4517             (flags & RENAME_EXCHANGE))
4518                 return -EINVAL;
4519
4520         if ((flags & RENAME_WHITEOUT) && !capable(CAP_MKNOD))
4521                 return -EPERM;
4522
4523         if (flags & RENAME_EXCHANGE)
4524                 target_flags = 0;
4525
4526 retry:
4527         from = filename_parentat(olddfd, getname(oldname), lookup_flags,
4528                                 &old_path, &old_last, &old_type);
4529         if (IS_ERR(from)) {
4530                 error = PTR_ERR(from);
4531                 goto exit;
4532         }
4533
4534         to = filename_parentat(newdfd, getname(newname), lookup_flags,
4535                                 &new_path, &new_last, &new_type);
4536         if (IS_ERR(to)) {
4537                 error = PTR_ERR(to);
4538                 goto exit1;
4539         }
4540
4541         error = -EXDEV;
4542         if (old_path.mnt != new_path.mnt)
4543                 goto exit2;
4544
4545         error = -EBUSY;
4546         if (old_type != LAST_NORM)
4547                 goto exit2;
4548
4549         if (flags & RENAME_NOREPLACE)
4550                 error = -EEXIST;
4551         if (new_type != LAST_NORM)
4552                 goto exit2;
4553
4554         error = mnt_want_write(old_path.mnt);
4555         if (error)
4556                 goto exit2;
4557
4558 retry_deleg:
4559         trap = lock_rename(new_path.dentry, old_path.dentry);
4560
4561         old_dentry = __lookup_hash(&old_last, old_path.dentry, lookup_flags);
4562         error = PTR_ERR(old_dentry);
4563         if (IS_ERR(old_dentry))
4564                 goto exit3;
4565         /* source must exist */
4566         error = -ENOENT;
4567         if (d_is_negative(old_dentry))
4568                 goto exit4;
4569         new_dentry = __lookup_hash(&new_last, new_path.dentry, lookup_flags | target_flags);
4570         error = PTR_ERR(new_dentry);
4571         if (IS_ERR(new_dentry))
4572                 goto exit4;
4573         error = -EEXIST;
4574         if ((flags & RENAME_NOREPLACE) && d_is_positive(new_dentry))
4575                 goto exit5;
4576         if (flags & RENAME_EXCHANGE) {
4577                 error = -ENOENT;
4578                 if (d_is_negative(new_dentry))
4579                         goto exit5;
4580
4581                 if (!d_is_dir(new_dentry)) {
4582                         error = -ENOTDIR;
4583                         if (new_last.name[new_last.len])
4584                                 goto exit5;
4585                 }
4586         }
4587         /* unless the source is a directory trailing slashes give -ENOTDIR */
4588         if (!d_is_dir(old_dentry)) {
4589                 error = -ENOTDIR;
4590                 if (old_last.name[old_last.len])
4591                         goto exit5;
4592                 if (!(flags & RENAME_EXCHANGE) && new_last.name[new_last.len])
4593                         goto exit5;
4594         }
4595         /* source should not be ancestor of target */
4596         error = -EINVAL;
4597         if (old_dentry == trap)
4598                 goto exit5;
4599         /* target should not be an ancestor of source */
4600         if (!(flags & RENAME_EXCHANGE))
4601                 error = -ENOTEMPTY;
4602         if (new_dentry == trap)
4603                 goto exit5;
4604
4605         error = security_path_rename(&old_path, old_dentry,
4606                                      &new_path, new_dentry, flags);
4607         if (error)
4608                 goto exit5;
4609         error = vfs_rename(old_path.dentry->d_inode, old_dentry,
4610                            new_path.dentry->d_inode, new_dentry,
4611                            &delegated_inode, flags);
4612 exit5:
4613         dput(new_dentry);
4614 exit4:
4615         dput(old_dentry);
4616 exit3:
4617         unlock_rename(new_path.dentry, old_path.dentry);
4618         if (delegated_inode) {
4619                 error = break_deleg_wait(&delegated_inode);
4620                 if (!error)
4621                         goto retry_deleg;
4622         }
4623         mnt_drop_write(old_path.mnt);
4624 exit2:
4625         if (retry_estale(error, lookup_flags))
4626                 should_retry = true;
4627         path_put(&new_path);
4628         putname(to);
4629 exit1:
4630         path_put(&old_path);
4631         putname(from);
4632         if (should_retry) {
4633                 should_retry = false;
4634                 lookup_flags |= LOOKUP_REVAL;
4635                 goto retry;
4636         }
4637 exit:
4638         return error;
4639 }
4640
4641 SYSCALL_DEFINE5(renameat2, int, olddfd, const char __user *, oldname,
4642                 int, newdfd, const char __user *, newname, unsigned int, flags)
4643 {
4644         return do_renameat2(olddfd, oldname, newdfd, newname, flags);
4645 }
4646
4647 SYSCALL_DEFINE4(renameat, int, olddfd, const char __user *, oldname,
4648                 int, newdfd, const char __user *, newname)
4649 {
4650         return do_renameat2(olddfd, oldname, newdfd, newname, 0);
4651 }
4652
4653 SYSCALL_DEFINE2(rename, const char __user *, oldname, const char __user *, newname)
4654 {
4655         return do_renameat2(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4656 }
4657
4658 int vfs_whiteout(struct inode *dir, struct dentry *dentry)
4659 {
4660         int error = may_create(dir, dentry);
4661         if (error)
4662                 return error;
4663
4664         if (!dir->i_op->mknod)
4665                 return -EPERM;
4666
4667         return dir->i_op->mknod(dir, dentry,
4668                                 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
4669 }
4670 EXPORT_SYMBOL(vfs_whiteout);
4671
4672 int readlink_copy(char __user *buffer, int buflen, const char *link)
4673 {
4674         int len = PTR_ERR(link);
4675         if (IS_ERR(link))
4676                 goto out;
4677
4678         len = strlen(link);
4679         if (len > (unsigned) buflen)
4680                 len = buflen;
4681         if (copy_to_user(buffer, link, len))
4682                 len = -EFAULT;
4683 out:
4684         return len;
4685 }
4686
4687 /*
4688  * A helper for ->readlink().  This should be used *ONLY* for symlinks that
4689  * have ->get_link() not calling nd_jump_link().  Using (or not using) it
4690  * for any given inode is up to filesystem.
4691  */
4692 static int generic_readlink(struct dentry *dentry, char __user *buffer,
4693                             int buflen)
4694 {
4695         DEFINE_DELAYED_CALL(done);
4696         struct inode *inode = d_inode(dentry);
4697         const char *link = inode->i_link;
4698         int res;
4699
4700         if (!link) {
4701                 link = inode->i_op->get_link(dentry, inode, &done);
4702                 if (IS_ERR(link))
4703                         return PTR_ERR(link);
4704         }
4705         res = readlink_copy(buffer, buflen, link);
4706         do_delayed_call(&done);
4707         return res;
4708 }
4709
4710 /**
4711  * vfs_readlink - copy symlink body into userspace buffer
4712  * @dentry: dentry on which to get symbolic link
4713  * @buffer: user memory pointer
4714  * @buflen: size of buffer
4715  *
4716  * Does not touch atime.  That's up to the caller if necessary
4717  *
4718  * Does not call security hook.
4719  */
4720 int vfs_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4721 {
4722         struct inode *inode = d_inode(dentry);
4723
4724         if (unlikely(!(inode->i_opflags & IOP_DEFAULT_READLINK))) {
4725                 if (unlikely(inode->i_op->readlink))
4726                         return inode->i_op->readlink(dentry, buffer, buflen);
4727
4728                 if (!d_is_symlink(dentry))
4729                         return -EINVAL;
4730
4731                 spin_lock(&inode->i_lock);
4732                 inode->i_opflags |= IOP_DEFAULT_READLINK;
4733                 spin_unlock(&inode->i_lock);
4734         }
4735
4736         return generic_readlink(dentry, buffer, buflen);
4737 }
4738 EXPORT_SYMBOL(vfs_readlink);
4739
4740 /**
4741  * vfs_get_link - get symlink body
4742  * @dentry: dentry on which to get symbolic link
4743  * @done: caller needs to free returned data with this
4744  *
4745  * Calls security hook and i_op->get_link() on the supplied inode.
4746  *
4747  * It does not touch atime.  That's up to the caller if necessary.
4748  *
4749  * Does not work on "special" symlinks like /proc/$$/fd/N
4750  */
4751 const char *vfs_get_link(struct dentry *dentry, struct delayed_call *done)
4752 {
4753         const char *res = ERR_PTR(-EINVAL);
4754         struct inode *inode = d_inode(dentry);
4755
4756         if (d_is_symlink(dentry)) {
4757                 res = ERR_PTR(security_inode_readlink(dentry));
4758                 if (!res)
4759                         res = inode->i_op->get_link(dentry, inode, done);
4760         }
4761         return res;
4762 }
4763 EXPORT_SYMBOL(vfs_get_link);
4764
4765 /* get the link contents into pagecache */
4766 const char *page_get_link(struct dentry *dentry, struct inode *inode,
4767                           struct delayed_call *callback)
4768 {
4769         char *kaddr;
4770         struct page *page;
4771         struct address_space *mapping = inode->i_mapping;
4772
4773         if (!dentry) {
4774                 page = find_get_page(mapping, 0);
4775                 if (!page)
4776                         return ERR_PTR(-ECHILD);
4777                 if (!PageUptodate(page)) {
4778                         put_page(page);
4779                         return ERR_PTR(-ECHILD);
4780                 }
4781         } else {
4782                 page = read_mapping_page(mapping, 0, NULL);
4783                 if (IS_ERR(page))
4784                         return (char*)page;
4785         }
4786         set_delayed_call(callback, page_put_link, page);
4787         BUG_ON(mapping_gfp_mask(mapping) & __GFP_HIGHMEM);
4788         kaddr = page_address(page);
4789         nd_terminate_link(kaddr, inode->i_size, PAGE_SIZE - 1);
4790         return kaddr;
4791 }
4792
4793 EXPORT_SYMBOL(page_get_link);
4794
4795 void page_put_link(void *arg)
4796 {
4797         put_page(arg);
4798 }
4799 EXPORT_SYMBOL(page_put_link);
4800
4801 int page_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4802 {
4803         DEFINE_DELAYED_CALL(done);
4804         int res = readlink_copy(buffer, buflen,
4805                                 page_get_link(dentry, d_inode(dentry),
4806                                               &done));
4807         do_delayed_call(&done);
4808         return res;
4809 }
4810 EXPORT_SYMBOL(page_readlink);
4811
4812 /*
4813  * The nofs argument instructs pagecache_write_begin to pass AOP_FLAG_NOFS
4814  */
4815 int __page_symlink(struct inode *inode, const char *symname, int len, int nofs)
4816 {
4817         struct address_space *mapping = inode->i_mapping;
4818         struct page *page;
4819         void *fsdata;
4820         int err;
4821         unsigned int flags = 0;
4822         if (nofs)
4823                 flags |= AOP_FLAG_NOFS;
4824
4825 retry:
4826         err = pagecache_write_begin(NULL, mapping, 0, len-1,
4827                                 flags, &page, &fsdata);
4828         if (err)
4829                 goto fail;
4830
4831         memcpy(page_address(page), symname, len-1);
4832
4833         err = pagecache_write_end(NULL, mapping, 0, len-1, len-1,
4834                                                         page, fsdata);
4835         if (err < 0)
4836                 goto fail;
4837         if (err < len-1)
4838                 goto retry;
4839
4840         mark_inode_dirty(inode);
4841         return 0;
4842 fail:
4843         return err;
4844 }
4845 EXPORT_SYMBOL(__page_symlink);
4846
4847 int page_symlink(struct inode *inode, const char *symname, int len)
4848 {
4849         return __page_symlink(inode, symname, len,
4850                         !mapping_gfp_constraint(inode->i_mapping, __GFP_FS));
4851 }
4852 EXPORT_SYMBOL(page_symlink);
4853
4854 const struct inode_operations page_symlink_inode_operations = {
4855         .get_link       = page_get_link,
4856 };
4857 EXPORT_SYMBOL(page_symlink_inode_operations);