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  * Call i_op->lookup on the dentry.  The dentry must be negative and
1478  * unhashed.
1479  *
1480  * dir->d_inode->i_mutex must be held
1481  */
1482 static struct dentry *lookup_real(struct inode *dir, struct dentry *dentry,
1483                                   unsigned int flags)
1484 {
1485         struct dentry *old;
1486
1487         /* Don't create child dentry for a dead directory. */
1488         if (unlikely(IS_DEADDIR(dir))) {
1489                 dput(dentry);
1490                 return ERR_PTR(-ENOENT);
1491         }
1492
1493         old = dir->i_op->lookup(dir, dentry, flags);
1494         if (unlikely(old)) {
1495                 dput(dentry);
1496                 dentry = old;
1497         }
1498         return dentry;
1499 }
1500
1501 static struct dentry *__lookup_hash(const struct qstr *name,
1502                 struct dentry *base, unsigned int flags)
1503 {
1504         struct dentry *dentry = lookup_dcache(name, base, flags);
1505
1506         if (dentry)
1507                 return dentry;
1508
1509         dentry = d_alloc(base, name);
1510         if (unlikely(!dentry))
1511                 return ERR_PTR(-ENOMEM);
1512
1513         return lookup_real(base->d_inode, dentry, flags);
1514 }
1515
1516 static int lookup_fast(struct nameidata *nd,
1517                        struct path *path, struct inode **inode,
1518                        unsigned *seqp)
1519 {
1520         struct vfsmount *mnt = nd->path.mnt;
1521         struct dentry *dentry, *parent = nd->path.dentry;
1522         int status = 1;
1523         int err;
1524
1525         /*
1526          * Rename seqlock is not required here because in the off chance
1527          * of a false negative due to a concurrent rename, the caller is
1528          * going to fall back to non-racy lookup.
1529          */
1530         if (nd->flags & LOOKUP_RCU) {
1531                 unsigned seq;
1532                 bool negative;
1533                 dentry = __d_lookup_rcu(parent, &nd->last, &seq);
1534                 if (unlikely(!dentry)) {
1535                         if (unlazy_walk(nd))
1536                                 return -ECHILD;
1537                         return 0;
1538                 }
1539
1540                 /*
1541                  * This sequence count validates that the inode matches
1542                  * the dentry name information from lookup.
1543                  */
1544                 *inode = d_backing_inode(dentry);
1545                 negative = d_is_negative(dentry);
1546                 if (unlikely(read_seqcount_retry(&dentry->d_seq, seq)))
1547                         return -ECHILD;
1548
1549                 /*
1550                  * This sequence count validates that the parent had no
1551                  * changes while we did the lookup of the dentry above.
1552                  *
1553                  * The memory barrier in read_seqcount_begin of child is
1554                  *  enough, we can use __read_seqcount_retry here.
1555                  */
1556                 if (unlikely(__read_seqcount_retry(&parent->d_seq, nd->seq)))
1557                         return -ECHILD;
1558
1559                 *seqp = seq;
1560                 status = d_revalidate(dentry, nd->flags);
1561                 if (likely(status > 0)) {
1562                         /*
1563                          * Note: do negative dentry check after revalidation in
1564                          * case that drops it.
1565                          */
1566                         if (unlikely(negative))
1567                                 return -ENOENT;
1568                         path->mnt = mnt;
1569                         path->dentry = dentry;
1570                         if (likely(__follow_mount_rcu(nd, path, inode, seqp)))
1571                                 return 1;
1572                 }
1573                 if (unlazy_child(nd, dentry, seq))
1574                         return -ECHILD;
1575                 if (unlikely(status == -ECHILD))
1576                         /* we'd been told to redo it in non-rcu mode */
1577                         status = d_revalidate(dentry, nd->flags);
1578         } else {
1579                 dentry = __d_lookup(parent, &nd->last);
1580                 if (unlikely(!dentry))
1581                         return 0;
1582                 status = d_revalidate(dentry, nd->flags);
1583         }
1584         if (unlikely(status <= 0)) {
1585                 if (!status)
1586                         d_invalidate(dentry);
1587                 dput(dentry);
1588                 return status;
1589         }
1590         if (unlikely(d_is_negative(dentry))) {
1591                 dput(dentry);
1592                 return -ENOENT;
1593         }
1594
1595         path->mnt = mnt;
1596         path->dentry = dentry;
1597         err = follow_managed(path, nd);
1598         if (likely(err > 0))
1599                 *inode = d_backing_inode(path->dentry);
1600         return err;
1601 }
1602
1603 /* Fast lookup failed, do it the slow way */
1604 static struct dentry *lookup_slow(const struct qstr *name,
1605                                   struct dentry *dir,
1606                                   unsigned int flags)
1607 {
1608         struct dentry *dentry = ERR_PTR(-ENOENT), *old;
1609         struct inode *inode = dir->d_inode;
1610         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
1611
1612         inode_lock_shared(inode);
1613         /* Don't go there if it's already dead */
1614         if (unlikely(IS_DEADDIR(inode)))
1615                 goto out;
1616 again:
1617         dentry = d_alloc_parallel(dir, name, &wq);
1618         if (IS_ERR(dentry))
1619                 goto out;
1620         if (unlikely(!d_in_lookup(dentry))) {
1621                 if (!(flags & LOOKUP_NO_REVAL)) {
1622                         int error = d_revalidate(dentry, flags);
1623                         if (unlikely(error <= 0)) {
1624                                 if (!error) {
1625                                         d_invalidate(dentry);
1626                                         dput(dentry);
1627                                         goto again;
1628                                 }
1629                                 dput(dentry);
1630                                 dentry = ERR_PTR(error);
1631                         }
1632                 }
1633         } else {
1634                 old = inode->i_op->lookup(inode, dentry, flags);
1635                 d_lookup_done(dentry);
1636                 if (unlikely(old)) {
1637                         dput(dentry);
1638                         dentry = old;
1639                 }
1640         }
1641 out:
1642         inode_unlock_shared(inode);
1643         return dentry;
1644 }
1645
1646 static inline int may_lookup(struct nameidata *nd)
1647 {
1648         if (nd->flags & LOOKUP_RCU) {
1649                 int err = inode_permission(nd->inode, MAY_EXEC|MAY_NOT_BLOCK);
1650                 if (err != -ECHILD)
1651                         return err;
1652                 if (unlazy_walk(nd))
1653                         return -ECHILD;
1654         }
1655         return inode_permission(nd->inode, MAY_EXEC);
1656 }
1657
1658 static inline int handle_dots(struct nameidata *nd, int type)
1659 {
1660         if (type == LAST_DOTDOT) {
1661                 if (!nd->root.mnt)
1662                         set_root(nd);
1663                 if (nd->flags & LOOKUP_RCU) {
1664                         return follow_dotdot_rcu(nd);
1665                 } else
1666                         return follow_dotdot(nd);
1667         }
1668         return 0;
1669 }
1670
1671 static int pick_link(struct nameidata *nd, struct path *link,
1672                      struct inode *inode, unsigned seq)
1673 {
1674         int error;
1675         struct saved *last;
1676         if (unlikely(nd->total_link_count++ >= MAXSYMLINKS)) {
1677                 path_to_nameidata(link, nd);
1678                 return -ELOOP;
1679         }
1680         if (!(nd->flags & LOOKUP_RCU)) {
1681                 if (link->mnt == nd->path.mnt)
1682                         mntget(link->mnt);
1683         }
1684         error = nd_alloc_stack(nd);
1685         if (unlikely(error)) {
1686                 if (error == -ECHILD) {
1687                         if (unlikely(!legitimize_path(nd, link, seq))) {
1688                                 drop_links(nd);
1689                                 nd->depth = 0;
1690                                 nd->flags &= ~LOOKUP_RCU;
1691                                 nd->path.mnt = NULL;
1692                                 nd->path.dentry = NULL;
1693                                 if (!(nd->flags & LOOKUP_ROOT))
1694                                         nd->root.mnt = NULL;
1695                                 rcu_read_unlock();
1696                         } else if (likely(unlazy_walk(nd)) == 0)
1697                                 error = nd_alloc_stack(nd);
1698                 }
1699                 if (error) {
1700                         path_put(link);
1701                         return error;
1702                 }
1703         }
1704
1705         last = nd->stack + nd->depth++;
1706         last->link = *link;
1707         clear_delayed_call(&last->done);
1708         nd->link_inode = inode;
1709         last->seq = seq;
1710         return 1;
1711 }
1712
1713 enum {WALK_FOLLOW = 1, WALK_MORE = 2};
1714
1715 /*
1716  * Do we need to follow links? We _really_ want to be able
1717  * to do this check without having to look at inode->i_op,
1718  * so we keep a cache of "no, this doesn't need follow_link"
1719  * for the common case.
1720  */
1721 static inline int step_into(struct nameidata *nd, struct path *path,
1722                             int flags, struct inode *inode, unsigned seq)
1723 {
1724         if (!(flags & WALK_MORE) && nd->depth)
1725                 put_link(nd);
1726         if (likely(!d_is_symlink(path->dentry)) ||
1727            !(flags & WALK_FOLLOW || nd->flags & LOOKUP_FOLLOW)) {
1728                 /* not a symlink or should not follow */
1729                 path_to_nameidata(path, nd);
1730                 nd->inode = inode;
1731                 nd->seq = seq;
1732                 return 0;
1733         }
1734         /* make sure that d_is_symlink above matches inode */
1735         if (nd->flags & LOOKUP_RCU) {
1736                 if (read_seqcount_retry(&path->dentry->d_seq, seq))
1737                         return -ECHILD;
1738         }
1739         return pick_link(nd, path, inode, seq);
1740 }
1741
1742 static int walk_component(struct nameidata *nd, int flags)
1743 {
1744         struct path path;
1745         struct inode *inode;
1746         unsigned seq;
1747         int err;
1748         /*
1749          * "." and ".." are special - ".." especially so because it has
1750          * to be able to know about the current root directory and
1751          * parent relationships.
1752          */
1753         if (unlikely(nd->last_type != LAST_NORM)) {
1754                 err = handle_dots(nd, nd->last_type);
1755                 if (!(flags & WALK_MORE) && nd->depth)
1756                         put_link(nd);
1757                 return err;
1758         }
1759         err = lookup_fast(nd, &path, &inode, &seq);
1760         if (unlikely(err <= 0)) {
1761                 if (err < 0)
1762                         return err;
1763                 path.dentry = lookup_slow(&nd->last, nd->path.dentry,
1764                                           nd->flags);
1765                 if (IS_ERR(path.dentry))
1766                         return PTR_ERR(path.dentry);
1767
1768                 path.mnt = nd->path.mnt;
1769                 err = follow_managed(&path, nd);
1770                 if (unlikely(err < 0))
1771                         return err;
1772
1773                 if (unlikely(d_is_negative(path.dentry))) {
1774                         path_to_nameidata(&path, nd);
1775                         return -ENOENT;
1776                 }
1777
1778                 seq = 0;        /* we are already out of RCU mode */
1779                 inode = d_backing_inode(path.dentry);
1780         }
1781
1782         return step_into(nd, &path, flags, inode, seq);
1783 }
1784
1785 /*
1786  * We can do the critical dentry name comparison and hashing
1787  * operations one word at a time, but we are limited to:
1788  *
1789  * - Architectures with fast unaligned word accesses. We could
1790  *   do a "get_unaligned()" if this helps and is sufficiently
1791  *   fast.
1792  *
1793  * - non-CONFIG_DEBUG_PAGEALLOC configurations (so that we
1794  *   do not trap on the (extremely unlikely) case of a page
1795  *   crossing operation.
1796  *
1797  * - Furthermore, we need an efficient 64-bit compile for the
1798  *   64-bit case in order to generate the "number of bytes in
1799  *   the final mask". Again, that could be replaced with a
1800  *   efficient population count instruction or similar.
1801  */
1802 #ifdef CONFIG_DCACHE_WORD_ACCESS
1803
1804 #include <asm/word-at-a-time.h>
1805
1806 #ifdef HASH_MIX
1807
1808 /* Architecture provides HASH_MIX and fold_hash() in <asm/hash.h> */
1809
1810 #elif defined(CONFIG_64BIT)
1811 /*
1812  * Register pressure in the mixing function is an issue, particularly
1813  * on 32-bit x86, but almost any function requires one state value and
1814  * one temporary.  Instead, use a function designed for two state values
1815  * and no temporaries.
1816  *
1817  * This function cannot create a collision in only two iterations, so
1818  * we have two iterations to achieve avalanche.  In those two iterations,
1819  * we have six layers of mixing, which is enough to spread one bit's
1820  * influence out to 2^6 = 64 state bits.
1821  *
1822  * Rotate constants are scored by considering either 64 one-bit input
1823  * deltas or 64*63/2 = 2016 two-bit input deltas, and finding the
1824  * probability of that delta causing a change to each of the 128 output
1825  * bits, using a sample of random initial states.
1826  *
1827  * The Shannon entropy of the computed probabilities is then summed
1828  * to produce a score.  Ideally, any input change has a 50% chance of
1829  * toggling any given output bit.
1830  *
1831  * Mixing scores (in bits) for (12,45):
1832  * Input delta: 1-bit      2-bit
1833  * 1 round:     713.3    42542.6
1834  * 2 rounds:   2753.7   140389.8
1835  * 3 rounds:   5954.1   233458.2
1836  * 4 rounds:   7862.6   256672.2
1837  * Perfect:    8192     258048
1838  *            (64*128) (64*63/2 * 128)
1839  */
1840 #define HASH_MIX(x, y, a)       \
1841         (       x ^= (a),       \
1842         y ^= x, x = rol64(x,12),\
1843         x += y, y = rol64(y,45),\
1844         y *= 9                  )
1845
1846 /*
1847  * Fold two longs into one 32-bit hash value.  This must be fast, but
1848  * latency isn't quite as critical, as there is a fair bit of additional
1849  * work done before the hash value is used.
1850  */
1851 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1852 {
1853         y ^= x * GOLDEN_RATIO_64;
1854         y *= GOLDEN_RATIO_64;
1855         return y >> 32;
1856 }
1857
1858 #else   /* 32-bit case */
1859
1860 /*
1861  * Mixing scores (in bits) for (7,20):
1862  * Input delta: 1-bit      2-bit
1863  * 1 round:     330.3     9201.6
1864  * 2 rounds:   1246.4    25475.4
1865  * 3 rounds:   1907.1    31295.1
1866  * 4 rounds:   2042.3    31718.6
1867  * Perfect:    2048      31744
1868  *            (32*64)   (32*31/2 * 64)
1869  */
1870 #define HASH_MIX(x, y, a)       \
1871         (       x ^= (a),       \
1872         y ^= x, x = rol32(x, 7),\
1873         x += y, y = rol32(y,20),\
1874         y *= 9                  )
1875
1876 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1877 {
1878         /* Use arch-optimized multiply if one exists */
1879         return __hash_32(y ^ __hash_32(x));
1880 }
1881
1882 #endif
1883
1884 /*
1885  * Return the hash of a string of known length.  This is carfully
1886  * designed to match hash_name(), which is the more critical function.
1887  * In particular, we must end by hashing a final word containing 0..7
1888  * payload bytes, to match the way that hash_name() iterates until it
1889  * finds the delimiter after the name.
1890  */
1891 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1892 {
1893         unsigned long a, x = 0, y = (unsigned long)salt;
1894
1895         for (;;) {
1896                 if (!len)
1897                         goto done;
1898                 a = load_unaligned_zeropad(name);
1899                 if (len < sizeof(unsigned long))
1900                         break;
1901                 HASH_MIX(x, y, a);
1902                 name += sizeof(unsigned long);
1903                 len -= sizeof(unsigned long);
1904         }
1905         x ^= a & bytemask_from_count(len);
1906 done:
1907         return fold_hash(x, y);
1908 }
1909 EXPORT_SYMBOL(full_name_hash);
1910
1911 /* Return the "hash_len" (hash and length) of a null-terminated string */
1912 u64 hashlen_string(const void *salt, const char *name)
1913 {
1914         unsigned long a = 0, x = 0, y = (unsigned long)salt;
1915         unsigned long adata, mask, len;
1916         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1917
1918         len = 0;
1919         goto inside;
1920
1921         do {
1922                 HASH_MIX(x, y, a);
1923                 len += sizeof(unsigned long);
1924 inside:
1925                 a = load_unaligned_zeropad(name+len);
1926         } while (!has_zero(a, &adata, &constants));
1927
1928         adata = prep_zero_mask(a, adata, &constants);
1929         mask = create_zero_mask(adata);
1930         x ^= a & zero_bytemask(mask);
1931
1932         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1933 }
1934 EXPORT_SYMBOL(hashlen_string);
1935
1936 /*
1937  * Calculate the length and hash of the path component, and
1938  * return the "hash_len" as the result.
1939  */
1940 static inline u64 hash_name(const void *salt, const char *name)
1941 {
1942         unsigned long a = 0, b, x = 0, y = (unsigned long)salt;
1943         unsigned long adata, bdata, mask, len;
1944         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1945
1946         len = 0;
1947         goto inside;
1948
1949         do {
1950                 HASH_MIX(x, y, a);
1951                 len += sizeof(unsigned long);
1952 inside:
1953                 a = load_unaligned_zeropad(name+len);
1954                 b = a ^ REPEAT_BYTE('/');
1955         } while (!(has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)));
1956
1957         adata = prep_zero_mask(a, adata, &constants);
1958         bdata = prep_zero_mask(b, bdata, &constants);
1959         mask = create_zero_mask(adata | bdata);
1960         x ^= a & zero_bytemask(mask);
1961
1962         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1963 }
1964
1965 #else   /* !CONFIG_DCACHE_WORD_ACCESS: Slow, byte-at-a-time version */
1966
1967 /* Return the hash of a string of known length */
1968 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1969 {
1970         unsigned long hash = init_name_hash(salt);
1971         while (len--)
1972                 hash = partial_name_hash((unsigned char)*name++, hash);
1973         return end_name_hash(hash);
1974 }
1975 EXPORT_SYMBOL(full_name_hash);
1976
1977 /* Return the "hash_len" (hash and length) of a null-terminated string */
1978 u64 hashlen_string(const void *salt, const char *name)
1979 {
1980         unsigned long hash = init_name_hash(salt);
1981         unsigned long len = 0, c;
1982
1983         c = (unsigned char)*name;
1984         while (c) {
1985                 len++;
1986                 hash = partial_name_hash(c, hash);
1987                 c = (unsigned char)name[len];
1988         }
1989         return hashlen_create(end_name_hash(hash), len);
1990 }
1991 EXPORT_SYMBOL(hashlen_string);
1992
1993 /*
1994  * We know there's a real path component here of at least
1995  * one character.
1996  */
1997 static inline u64 hash_name(const void *salt, const char *name)
1998 {
1999         unsigned long hash = init_name_hash(salt);
2000         unsigned long len = 0, c;
2001
2002         c = (unsigned char)*name;
2003         do {
2004                 len++;
2005                 hash = partial_name_hash(c, hash);
2006                 c = (unsigned char)name[len];
2007         } while (c && c != '/');
2008         return hashlen_create(end_name_hash(hash), len);
2009 }
2010
2011 #endif
2012
2013 /*
2014  * Name resolution.
2015  * This is the basic name resolution function, turning a pathname into
2016  * the final dentry. We expect 'base' to be positive and a directory.
2017  *
2018  * Returns 0 and nd will have valid dentry and mnt on success.
2019  * Returns error and drops reference to input namei data on failure.
2020  */
2021 static int link_path_walk(const char *name, struct nameidata *nd)
2022 {
2023         int err;
2024
2025         while (*name=='/')
2026                 name++;
2027         if (!*name)
2028                 return 0;
2029
2030         /* At this point we know we have a real path component. */
2031         for(;;) {
2032                 u64 hash_len;
2033                 int type;
2034
2035                 err = may_lookup(nd);
2036                 if (err)
2037                         return err;
2038
2039                 hash_len = hash_name(nd->path.dentry, name);
2040
2041                 type = LAST_NORM;
2042                 if (name[0] == '.') switch (hashlen_len(hash_len)) {
2043                         case 2:
2044                                 if (name[1] == '.') {
2045                                         type = LAST_DOTDOT;
2046                                         nd->flags |= LOOKUP_JUMPED;
2047                                 }
2048                                 break;
2049                         case 1:
2050                                 type = LAST_DOT;
2051                 }
2052                 if (likely(type == LAST_NORM)) {
2053                         struct dentry *parent = nd->path.dentry;
2054                         nd->flags &= ~LOOKUP_JUMPED;
2055                         if (unlikely(parent->d_flags & DCACHE_OP_HASH)) {
2056                                 struct qstr this = { { .hash_len = hash_len }, .name = name };
2057                                 err = parent->d_op->d_hash(parent, &this);
2058                                 if (err < 0)
2059                                         return err;
2060                                 hash_len = this.hash_len;
2061                                 name = this.name;
2062                         }
2063                 }
2064
2065                 nd->last.hash_len = hash_len;
2066                 nd->last.name = name;
2067                 nd->last_type = type;
2068
2069                 name += hashlen_len(hash_len);
2070                 if (!*name)
2071                         goto OK;
2072                 /*
2073                  * If it wasn't NUL, we know it was '/'. Skip that
2074                  * slash, and continue until no more slashes.
2075                  */
2076                 do {
2077                         name++;
2078                 } while (unlikely(*name == '/'));
2079                 if (unlikely(!*name)) {
2080 OK:
2081                         /* pathname body, done */
2082                         if (!nd->depth)
2083                                 return 0;
2084                         name = nd->stack[nd->depth - 1].name;
2085                         /* trailing symlink, done */
2086                         if (!name)
2087                                 return 0;
2088                         /* last component of nested symlink */
2089                         err = walk_component(nd, WALK_FOLLOW);
2090                 } else {
2091                         /* not the last component */
2092                         err = walk_component(nd, WALK_FOLLOW | WALK_MORE);
2093                 }
2094                 if (err < 0)
2095                         return err;
2096
2097                 if (err) {
2098                         const char *s = get_link(nd);
2099
2100                         if (IS_ERR(s))
2101                                 return PTR_ERR(s);
2102                         err = 0;
2103                         if (unlikely(!s)) {
2104                                 /* jumped */
2105                                 put_link(nd);
2106                         } else {
2107                                 nd->stack[nd->depth - 1].name = name;
2108                                 name = s;
2109                                 continue;
2110                         }
2111                 }
2112                 if (unlikely(!d_can_lookup(nd->path.dentry))) {
2113                         if (nd->flags & LOOKUP_RCU) {
2114                                 if (unlazy_walk(nd))
2115                                         return -ECHILD;
2116                         }
2117                         return -ENOTDIR;
2118                 }
2119         }
2120 }
2121
2122 static const char *path_init(struct nameidata *nd, unsigned flags)
2123 {
2124         const char *s = nd->name->name;
2125
2126         if (!*s)
2127                 flags &= ~LOOKUP_RCU;
2128
2129         nd->last_type = LAST_ROOT; /* if there are only slashes... */
2130         nd->flags = flags | LOOKUP_JUMPED | LOOKUP_PARENT;
2131         nd->depth = 0;
2132         if (flags & LOOKUP_ROOT) {
2133                 struct dentry *root = nd->root.dentry;
2134                 struct inode *inode = root->d_inode;
2135                 if (*s && unlikely(!d_can_lookup(root)))
2136                         return ERR_PTR(-ENOTDIR);
2137                 nd->path = nd->root;
2138                 nd->inode = inode;
2139                 if (flags & LOOKUP_RCU) {
2140                         rcu_read_lock();
2141                         nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2142                         nd->root_seq = nd->seq;
2143                         nd->m_seq = read_seqbegin(&mount_lock);
2144                 } else {
2145                         path_get(&nd->path);
2146                 }
2147                 return s;
2148         }
2149
2150         nd->root.mnt = NULL;
2151         nd->path.mnt = NULL;
2152         nd->path.dentry = NULL;
2153
2154         nd->m_seq = read_seqbegin(&mount_lock);
2155         if (*s == '/') {
2156                 if (flags & LOOKUP_RCU)
2157                         rcu_read_lock();
2158                 set_root(nd);
2159                 if (likely(!nd_jump_root(nd)))
2160                         return s;
2161                 nd->root.mnt = NULL;
2162                 rcu_read_unlock();
2163                 return ERR_PTR(-ECHILD);
2164         } else if (nd->dfd == AT_FDCWD) {
2165                 if (flags & LOOKUP_RCU) {
2166                         struct fs_struct *fs = current->fs;
2167                         unsigned seq;
2168
2169                         rcu_read_lock();
2170
2171                         do {
2172                                 seq = read_seqcount_begin(&fs->seq);
2173                                 nd->path = fs->pwd;
2174                                 nd->inode = nd->path.dentry->d_inode;
2175                                 nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2176                         } while (read_seqcount_retry(&fs->seq, seq));
2177                 } else {
2178                         get_fs_pwd(current->fs, &nd->path);
2179                         nd->inode = nd->path.dentry->d_inode;
2180                 }
2181                 return s;
2182         } else {
2183                 /* Caller must check execute permissions on the starting path component */
2184                 struct fd f = fdget_raw(nd->dfd);
2185                 struct dentry *dentry;
2186
2187                 if (!f.file)
2188                         return ERR_PTR(-EBADF);
2189
2190                 dentry = f.file->f_path.dentry;
2191
2192                 if (*s) {
2193                         if (!d_can_lookup(dentry)) {
2194                                 fdput(f);
2195                                 return ERR_PTR(-ENOTDIR);
2196                         }
2197                 }
2198
2199                 nd->path = f.file->f_path;
2200                 if (flags & LOOKUP_RCU) {
2201                         rcu_read_lock();
2202                         nd->inode = nd->path.dentry->d_inode;
2203                         nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
2204                 } else {
2205                         path_get(&nd->path);
2206                         nd->inode = nd->path.dentry->d_inode;
2207                 }
2208                 fdput(f);
2209                 return s;
2210         }
2211 }
2212
2213 static const char *trailing_symlink(struct nameidata *nd)
2214 {
2215         const char *s;
2216         int error = may_follow_link(nd);
2217         if (unlikely(error))
2218                 return ERR_PTR(error);
2219         nd->flags |= LOOKUP_PARENT;
2220         nd->stack[0].name = NULL;
2221         s = get_link(nd);
2222         return s ? s : "";
2223 }
2224
2225 static inline int lookup_last(struct nameidata *nd)
2226 {
2227         if (nd->last_type == LAST_NORM && nd->last.name[nd->last.len])
2228                 nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
2229
2230         nd->flags &= ~LOOKUP_PARENT;
2231         return walk_component(nd, 0);
2232 }
2233
2234 static int handle_lookup_down(struct nameidata *nd)
2235 {
2236         struct path path = nd->path;
2237         struct inode *inode = nd->inode;
2238         unsigned seq = nd->seq;
2239         int err;
2240
2241         if (nd->flags & LOOKUP_RCU) {
2242                 /*
2243                  * don't bother with unlazy_walk on failure - we are
2244                  * at the very beginning of walk, so we lose nothing
2245                  * if we simply redo everything in non-RCU mode
2246                  */
2247                 if (unlikely(!__follow_mount_rcu(nd, &path, &inode, &seq)))
2248                         return -ECHILD;
2249         } else {
2250                 dget(path.dentry);
2251                 err = follow_managed(&path, nd);
2252                 if (unlikely(err < 0))
2253                         return err;
2254                 inode = d_backing_inode(path.dentry);
2255                 seq = 0;
2256         }
2257         path_to_nameidata(&path, nd);
2258         nd->inode = inode;
2259         nd->seq = seq;
2260         return 0;
2261 }
2262
2263 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2264 static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
2265 {
2266         const char *s = path_init(nd, flags);
2267         int err;
2268
2269         if (IS_ERR(s))
2270                 return PTR_ERR(s);
2271
2272         if (unlikely(flags & LOOKUP_DOWN)) {
2273                 err = handle_lookup_down(nd);
2274                 if (unlikely(err < 0)) {
2275                         terminate_walk(nd);
2276                         return err;
2277                 }
2278         }
2279
2280         while (!(err = link_path_walk(s, nd))
2281                 && ((err = lookup_last(nd)) > 0)) {
2282                 s = trailing_symlink(nd);
2283                 if (IS_ERR(s)) {
2284                         err = PTR_ERR(s);
2285                         break;
2286                 }
2287         }
2288         if (!err)
2289                 err = complete_walk(nd);
2290
2291         if (!err && nd->flags & LOOKUP_DIRECTORY)
2292                 if (!d_can_lookup(nd->path.dentry))
2293                         err = -ENOTDIR;
2294         if (!err) {
2295                 *path = nd->path;
2296                 nd->path.mnt = NULL;
2297                 nd->path.dentry = NULL;
2298         }
2299         terminate_walk(nd);
2300         return err;
2301 }
2302
2303 static int filename_lookup(int dfd, struct filename *name, unsigned flags,
2304                            struct path *path, struct path *root)
2305 {
2306         int retval;
2307         struct nameidata nd;
2308         if (IS_ERR(name))
2309                 return PTR_ERR(name);
2310         if (unlikely(root)) {
2311                 nd.root = *root;
2312                 flags |= LOOKUP_ROOT;
2313         }
2314         set_nameidata(&nd, dfd, name);
2315         retval = path_lookupat(&nd, flags | LOOKUP_RCU, path);
2316         if (unlikely(retval == -ECHILD))
2317                 retval = path_lookupat(&nd, flags, path);
2318         if (unlikely(retval == -ESTALE))
2319                 retval = path_lookupat(&nd, flags | LOOKUP_REVAL, path);
2320
2321         if (likely(!retval))
2322                 audit_inode(name, path->dentry, flags & LOOKUP_PARENT);
2323         restore_nameidata();
2324         putname(name);
2325         return retval;
2326 }
2327
2328 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2329 static int path_parentat(struct nameidata *nd, unsigned flags,
2330                                 struct path *parent)
2331 {
2332         const char *s = path_init(nd, flags);
2333         int err;
2334         if (IS_ERR(s))
2335                 return PTR_ERR(s);
2336         err = link_path_walk(s, nd);
2337         if (!err)
2338                 err = complete_walk(nd);
2339         if (!err) {
2340                 *parent = nd->path;
2341                 nd->path.mnt = NULL;
2342                 nd->path.dentry = NULL;
2343         }
2344         terminate_walk(nd);
2345         return err;
2346 }
2347
2348 static struct filename *filename_parentat(int dfd, struct filename *name,
2349                                 unsigned int flags, struct path *parent,
2350                                 struct qstr *last, int *type)
2351 {
2352         int retval;
2353         struct nameidata nd;
2354
2355         if (IS_ERR(name))
2356                 return name;
2357         set_nameidata(&nd, dfd, name);
2358         retval = path_parentat(&nd, flags | LOOKUP_RCU, parent);
2359         if (unlikely(retval == -ECHILD))
2360                 retval = path_parentat(&nd, flags, parent);
2361         if (unlikely(retval == -ESTALE))
2362                 retval = path_parentat(&nd, flags | LOOKUP_REVAL, parent);
2363         if (likely(!retval)) {
2364                 *last = nd.last;
2365                 *type = nd.last_type;
2366                 audit_inode(name, parent->dentry, LOOKUP_PARENT);
2367         } else {
2368                 putname(name);
2369                 name = ERR_PTR(retval);
2370         }
2371         restore_nameidata();
2372         return name;
2373 }
2374
2375 /* does lookup, returns the object with parent locked */
2376 struct dentry *kern_path_locked(const char *name, struct path *path)
2377 {
2378         struct filename *filename;
2379         struct dentry *d;
2380         struct qstr last;
2381         int type;
2382
2383         filename = filename_parentat(AT_FDCWD, getname_kernel(name), 0, path,
2384                                     &last, &type);
2385         if (IS_ERR(filename))
2386                 return ERR_CAST(filename);
2387         if (unlikely(type != LAST_NORM)) {
2388                 path_put(path);
2389                 putname(filename);
2390                 return ERR_PTR(-EINVAL);
2391         }
2392         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
2393         d = __lookup_hash(&last, path->dentry, 0);
2394         if (IS_ERR(d)) {
2395                 inode_unlock(path->dentry->d_inode);
2396                 path_put(path);
2397         }
2398         putname(filename);
2399         return d;
2400 }
2401
2402 int kern_path(const char *name, unsigned int flags, struct path *path)
2403 {
2404         return filename_lookup(AT_FDCWD, getname_kernel(name),
2405                                flags, path, NULL);
2406 }
2407 EXPORT_SYMBOL(kern_path);
2408
2409 /**
2410  * vfs_path_lookup - lookup a file path relative to a dentry-vfsmount pair
2411  * @dentry:  pointer to dentry of the base directory
2412  * @mnt: pointer to vfs mount of the base directory
2413  * @name: pointer to file name
2414  * @flags: lookup flags
2415  * @path: pointer to struct path to fill
2416  */
2417 int vfs_path_lookup(struct dentry *dentry, struct vfsmount *mnt,
2418                     const char *name, unsigned int flags,
2419                     struct path *path)
2420 {
2421         struct path root = {.mnt = mnt, .dentry = dentry};
2422         /* the first argument of filename_lookup() is ignored with root */
2423         return filename_lookup(AT_FDCWD, getname_kernel(name),
2424                                flags , path, &root);
2425 }
2426 EXPORT_SYMBOL(vfs_path_lookup);
2427
2428 /**
2429  * lookup_one_len - filesystem helper to lookup single pathname component
2430  * @name:       pathname component to lookup
2431  * @base:       base directory to lookup from
2432  * @len:        maximum length @len should be interpreted to
2433  *
2434  * Note that this routine is purely a helper for filesystem usage and should
2435  * not be called by generic code.
2436  *
2437  * The caller must hold base->i_mutex.
2438  */
2439 struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
2440 {
2441         struct qstr this;
2442         unsigned int c;
2443         int err;
2444
2445         WARN_ON_ONCE(!inode_is_locked(base->d_inode));
2446
2447         this.name = name;
2448         this.len = len;
2449         this.hash = full_name_hash(base, name, len);
2450         if (!len)
2451                 return ERR_PTR(-EACCES);
2452
2453         if (unlikely(name[0] == '.')) {
2454                 if (len < 2 || (len == 2 && name[1] == '.'))
2455                         return ERR_PTR(-EACCES);
2456         }
2457
2458         while (len--) {
2459                 c = *(const unsigned char *)name++;
2460                 if (c == '/' || c == '\0')
2461                         return ERR_PTR(-EACCES);
2462         }
2463         /*
2464          * See if the low-level filesystem might want
2465          * to use its own hash..
2466          */
2467         if (base->d_flags & DCACHE_OP_HASH) {
2468                 int err = base->d_op->d_hash(base, &this);
2469                 if (err < 0)
2470                         return ERR_PTR(err);
2471         }
2472
2473         err = inode_permission(base->d_inode, MAY_EXEC);
2474         if (err)
2475                 return ERR_PTR(err);
2476
2477         return __lookup_hash(&this, base, 0);
2478 }
2479 EXPORT_SYMBOL(lookup_one_len);
2480
2481 /**
2482  * lookup_one_len_unlocked - filesystem helper to lookup single pathname component
2483  * @name:       pathname component to lookup
2484  * @base:       base directory to lookup from
2485  * @len:        maximum length @len should be interpreted to
2486  *
2487  * Note that this routine is purely a helper for filesystem usage and should
2488  * not be called by generic code.
2489  *
2490  * Unlike lookup_one_len, it should be called without the parent
2491  * i_mutex held, and will take the i_mutex itself if necessary.
2492  */
2493 struct dentry *lookup_one_len_unlocked(const char *name,
2494                                        struct dentry *base, int len)
2495 {
2496         struct qstr this;
2497         unsigned int c;
2498         int err;
2499         struct dentry *ret;
2500
2501         this.name = name;
2502         this.len = len;
2503         this.hash = full_name_hash(base, name, len);
2504         if (!len)
2505                 return ERR_PTR(-EACCES);
2506
2507         if (unlikely(name[0] == '.')) {
2508                 if (len < 2 || (len == 2 && name[1] == '.'))
2509                         return ERR_PTR(-EACCES);
2510         }
2511
2512         while (len--) {
2513                 c = *(const unsigned char *)name++;
2514                 if (c == '/' || c == '\0')
2515                         return ERR_PTR(-EACCES);
2516         }
2517         /*
2518          * See if the low-level filesystem might want
2519          * to use its own hash..
2520          */
2521         if (base->d_flags & DCACHE_OP_HASH) {
2522                 int err = base->d_op->d_hash(base, &this);
2523                 if (err < 0)
2524                         return ERR_PTR(err);
2525         }
2526
2527         err = inode_permission(base->d_inode, MAY_EXEC);
2528         if (err)
2529                 return ERR_PTR(err);
2530
2531         ret = lookup_dcache(&this, base, 0);
2532         if (!ret)
2533                 ret = lookup_slow(&this, base, 0);
2534         return ret;
2535 }
2536 EXPORT_SYMBOL(lookup_one_len_unlocked);
2537
2538 #ifdef CONFIG_UNIX98_PTYS
2539 int path_pts(struct path *path)
2540 {
2541         /* Find something mounted on "pts" in the same directory as
2542          * the input path.
2543          */
2544         struct dentry *child, *parent;
2545         struct qstr this;
2546         int ret;
2547
2548         ret = path_parent_directory(path);
2549         if (ret)
2550                 return ret;
2551
2552         parent = path->dentry;
2553         this.name = "pts";
2554         this.len = 3;
2555         child = d_hash_and_lookup(parent, &this);
2556         if (!child)
2557                 return -ENOENT;
2558
2559         path->dentry = child;
2560         dput(parent);
2561         follow_mount(path);
2562         return 0;
2563 }
2564 #endif
2565
2566 int user_path_at_empty(int dfd, const char __user *name, unsigned flags,
2567                  struct path *path, int *empty)
2568 {
2569         return filename_lookup(dfd, getname_flags(name, flags, empty),
2570                                flags, path, NULL);
2571 }
2572 EXPORT_SYMBOL(user_path_at_empty);
2573
2574 /**
2575  * mountpoint_last - look up last component for umount
2576  * @nd:   pathwalk nameidata - currently pointing at parent directory of "last"
2577  *
2578  * This is a special lookup_last function just for umount. In this case, we
2579  * need to resolve the path without doing any revalidation.
2580  *
2581  * The nameidata should be the result of doing a LOOKUP_PARENT pathwalk. Since
2582  * mountpoints are always pinned in the dcache, their ancestors are too. Thus,
2583  * in almost all cases, this lookup will be served out of the dcache. The only
2584  * cases where it won't are if nd->last refers to a symlink or the path is
2585  * bogus and it doesn't exist.
2586  *
2587  * Returns:
2588  * -error: if there was an error during lookup. This includes -ENOENT if the
2589  *         lookup found a negative dentry.
2590  *
2591  * 0:      if we successfully resolved nd->last and found it to not to be a
2592  *         symlink that needs to be followed.
2593  *
2594  * 1:      if we successfully resolved nd->last and found it to be a symlink
2595  *         that needs to be followed.
2596  */
2597 static int
2598 mountpoint_last(struct nameidata *nd)
2599 {
2600         int error = 0;
2601         struct dentry *dir = nd->path.dentry;
2602         struct path path;
2603
2604         /* If we're in rcuwalk, drop out of it to handle last component */
2605         if (nd->flags & LOOKUP_RCU) {
2606                 if (unlazy_walk(nd))
2607                         return -ECHILD;
2608         }
2609
2610         nd->flags &= ~LOOKUP_PARENT;
2611
2612         if (unlikely(nd->last_type != LAST_NORM)) {
2613                 error = handle_dots(nd, nd->last_type);
2614                 if (error)
2615                         return error;
2616                 path.dentry = dget(nd->path.dentry);
2617         } else {
2618                 path.dentry = d_lookup(dir, &nd->last);
2619                 if (!path.dentry) {
2620                         /*
2621                          * No cached dentry. Mounted dentries are pinned in the
2622                          * cache, so that means that this dentry is probably
2623                          * a symlink or the path doesn't actually point
2624                          * to a mounted dentry.
2625                          */
2626                         path.dentry = lookup_slow(&nd->last, dir,
2627                                              nd->flags | LOOKUP_NO_REVAL);
2628                         if (IS_ERR(path.dentry))
2629                                 return PTR_ERR(path.dentry);
2630                 }
2631         }
2632         if (d_is_negative(path.dentry)) {
2633                 dput(path.dentry);
2634                 return -ENOENT;
2635         }
2636         path.mnt = nd->path.mnt;
2637         return step_into(nd, &path, 0, d_backing_inode(path.dentry), 0);
2638 }
2639
2640 /**
2641  * path_mountpoint - look up a path to be umounted
2642  * @nd:         lookup context
2643  * @flags:      lookup flags
2644  * @path:       pointer to container for result
2645  *
2646  * Look up the given name, but don't attempt to revalidate the last component.
2647  * Returns 0 and "path" will be valid on success; Returns error otherwise.
2648  */
2649 static int
2650 path_mountpoint(struct nameidata *nd, unsigned flags, struct path *path)
2651 {
2652         const char *s = path_init(nd, flags);
2653         int err;
2654         if (IS_ERR(s))
2655                 return PTR_ERR(s);
2656         while (!(err = link_path_walk(s, nd)) &&
2657                 (err = mountpoint_last(nd)) > 0) {
2658                 s = trailing_symlink(nd);
2659                 if (IS_ERR(s)) {
2660                         err = PTR_ERR(s);
2661                         break;
2662                 }
2663         }
2664         if (!err) {
2665                 *path = nd->path;
2666                 nd->path.mnt = NULL;
2667                 nd->path.dentry = NULL;
2668                 follow_mount(path);
2669         }
2670         terminate_walk(nd);
2671         return err;
2672 }
2673
2674 static int
2675 filename_mountpoint(int dfd, struct filename *name, struct path *path,
2676                         unsigned int flags)
2677 {
2678         struct nameidata nd;
2679         int error;
2680         if (IS_ERR(name))
2681                 return PTR_ERR(name);
2682         set_nameidata(&nd, dfd, name);
2683         error = path_mountpoint(&nd, flags | LOOKUP_RCU, path);
2684         if (unlikely(error == -ECHILD))
2685                 error = path_mountpoint(&nd, flags, path);
2686         if (unlikely(error == -ESTALE))
2687                 error = path_mountpoint(&nd, flags | LOOKUP_REVAL, path);
2688         if (likely(!error))
2689                 audit_inode(name, path->dentry, 0);
2690         restore_nameidata();
2691         putname(name);
2692         return error;
2693 }
2694
2695 /**
2696  * user_path_mountpoint_at - lookup a path from userland in order to umount it
2697  * @dfd:        directory file descriptor
2698  * @name:       pathname from userland
2699  * @flags:      lookup flags
2700  * @path:       pointer to container to hold result
2701  *
2702  * A umount is a special case for path walking. We're not actually interested
2703  * in the inode in this situation, and ESTALE errors can be a problem. We
2704  * simply want track down the dentry and vfsmount attached at the mountpoint
2705  * and avoid revalidating the last component.
2706  *
2707  * Returns 0 and populates "path" on success.
2708  */
2709 int
2710 user_path_mountpoint_at(int dfd, const char __user *name, unsigned int flags,
2711                         struct path *path)
2712 {
2713         return filename_mountpoint(dfd, getname(name), path, flags);
2714 }
2715
2716 int
2717 kern_path_mountpoint(int dfd, const char *name, struct path *path,
2718                         unsigned int flags)
2719 {
2720         return filename_mountpoint(dfd, getname_kernel(name), path, flags);
2721 }
2722 EXPORT_SYMBOL(kern_path_mountpoint);
2723
2724 int __check_sticky(struct inode *dir, struct inode *inode)
2725 {
2726         kuid_t fsuid = current_fsuid();
2727
2728         if (uid_eq(inode->i_uid, fsuid))
2729                 return 0;
2730         if (uid_eq(dir->i_uid, fsuid))
2731                 return 0;
2732         return !capable_wrt_inode_uidgid(inode, CAP_FOWNER);
2733 }
2734 EXPORT_SYMBOL(__check_sticky);
2735
2736 /*
2737  *      Check whether we can remove a link victim from directory dir, check
2738  *  whether the type of victim is right.
2739  *  1. We can't do it if dir is read-only (done in permission())
2740  *  2. We should have write and exec permissions on dir
2741  *  3. We can't remove anything from append-only dir
2742  *  4. We can't do anything with immutable dir (done in permission())
2743  *  5. If the sticky bit on dir is set we should either
2744  *      a. be owner of dir, or
2745  *      b. be owner of victim, or
2746  *      c. have CAP_FOWNER capability
2747  *  6. If the victim is append-only or immutable we can't do antyhing with
2748  *     links pointing to it.
2749  *  7. If the victim has an unknown uid or gid we can't change the inode.
2750  *  8. If we were asked to remove a directory and victim isn't one - ENOTDIR.
2751  *  9. If we were asked to remove a non-directory and victim isn't one - EISDIR.
2752  * 10. We can't remove a root or mountpoint.
2753  * 11. We don't allow removal of NFS sillyrenamed files; it's handled by
2754  *     nfs_async_unlink().
2755  */
2756 static int may_delete(struct inode *dir, struct dentry *victim, bool isdir)
2757 {
2758         struct inode *inode = d_backing_inode(victim);
2759         int error;
2760
2761         if (d_is_negative(victim))
2762                 return -ENOENT;
2763         BUG_ON(!inode);
2764
2765         BUG_ON(victim->d_parent->d_inode != dir);
2766         audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);
2767
2768         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
2769         if (error)
2770                 return error;
2771         if (IS_APPEND(dir))
2772                 return -EPERM;
2773
2774         if (check_sticky(dir, inode) || IS_APPEND(inode) ||
2775             IS_IMMUTABLE(inode) || IS_SWAPFILE(inode) || HAS_UNMAPPED_ID(inode))
2776                 return -EPERM;
2777         if (isdir) {
2778                 if (!d_is_dir(victim))
2779                         return -ENOTDIR;
2780                 if (IS_ROOT(victim))
2781                         return -EBUSY;
2782         } else if (d_is_dir(victim))
2783                 return -EISDIR;
2784         if (IS_DEADDIR(dir))
2785                 return -ENOENT;
2786         if (victim->d_flags & DCACHE_NFSFS_RENAMED)
2787                 return -EBUSY;
2788         return 0;
2789 }
2790
2791 /*      Check whether we can create an object with dentry child in directory
2792  *  dir.
2793  *  1. We can't do it if child already exists (open has special treatment for
2794  *     this case, but since we are inlined it's OK)
2795  *  2. We can't do it if dir is read-only (done in permission())
2796  *  3. We can't do it if the fs can't represent the fsuid or fsgid.
2797  *  4. We should have write and exec permissions on dir
2798  *  5. We can't do it if dir is immutable (done in permission())
2799  */
2800 static inline int may_create(struct inode *dir, struct dentry *child)
2801 {
2802         struct user_namespace *s_user_ns;
2803         audit_inode_child(dir, child, AUDIT_TYPE_CHILD_CREATE);
2804         if (child->d_inode)
2805                 return -EEXIST;
2806         if (IS_DEADDIR(dir))
2807                 return -ENOENT;
2808         s_user_ns = dir->i_sb->s_user_ns;
2809         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
2810             !kgid_has_mapping(s_user_ns, current_fsgid()))
2811                 return -EOVERFLOW;
2812         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
2813 }
2814
2815 /*
2816  * p1 and p2 should be directories on the same fs.
2817  */
2818 struct dentry *lock_rename(struct dentry *p1, struct dentry *p2)
2819 {
2820         struct dentry *p;
2821
2822         if (p1 == p2) {
2823                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2824                 return NULL;
2825         }
2826
2827         mutex_lock(&p1->d_sb->s_vfs_rename_mutex);
2828
2829         p = d_ancestor(p2, p1);
2830         if (p) {
2831                 inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
2832                 inode_lock_nested(p1->d_inode, I_MUTEX_CHILD);
2833                 return p;
2834         }
2835
2836         p = d_ancestor(p1, p2);
2837         if (p) {
2838                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2839                 inode_lock_nested(p2->d_inode, I_MUTEX_CHILD);
2840                 return p;
2841         }
2842
2843         inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2844         inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
2845         return NULL;
2846 }
2847 EXPORT_SYMBOL(lock_rename);
2848
2849 void unlock_rename(struct dentry *p1, struct dentry *p2)
2850 {
2851         inode_unlock(p1->d_inode);
2852         if (p1 != p2) {
2853                 inode_unlock(p2->d_inode);
2854                 mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
2855         }
2856 }
2857 EXPORT_SYMBOL(unlock_rename);
2858
2859 int vfs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2860                 bool want_excl)
2861 {
2862         int error = may_create(dir, dentry);
2863         if (error)
2864                 return error;
2865
2866         if (!dir->i_op->create)
2867                 return -EACCES; /* shouldn't it be ENOSYS? */
2868         mode &= S_IALLUGO;
2869         mode |= S_IFREG;
2870         error = security_inode_create(dir, dentry, mode);
2871         if (error)
2872                 return error;
2873         error = dir->i_op->create(dir, dentry, mode, want_excl);
2874         if (!error)
2875                 fsnotify_create(dir, dentry);
2876         return error;
2877 }
2878 EXPORT_SYMBOL(vfs_create);
2879
2880 int vfs_mkobj(struct dentry *dentry, umode_t mode,
2881                 int (*f)(struct dentry *, umode_t, void *),
2882                 void *arg)
2883 {
2884         struct inode *dir = dentry->d_parent->d_inode;
2885         int error = may_create(dir, dentry);
2886         if (error)
2887                 return error;
2888
2889         mode &= S_IALLUGO;
2890         mode |= S_IFREG;
2891         error = security_inode_create(dir, dentry, mode);
2892         if (error)
2893                 return error;
2894         error = f(dentry, mode, arg);
2895         if (!error)
2896                 fsnotify_create(dir, dentry);
2897         return error;
2898 }
2899 EXPORT_SYMBOL(vfs_mkobj);
2900
2901 bool may_open_dev(const struct path *path)
2902 {
2903         return !(path->mnt->mnt_flags & MNT_NODEV) &&
2904                 !(path->mnt->mnt_sb->s_iflags & SB_I_NODEV);
2905 }
2906
2907 static int may_open(const struct path *path, int acc_mode, int flag)
2908 {
2909         struct dentry *dentry = path->dentry;
2910         struct inode *inode = dentry->d_inode;
2911         int error;
2912
2913         if (!inode)
2914                 return -ENOENT;
2915
2916         switch (inode->i_mode & S_IFMT) {
2917         case S_IFLNK:
2918                 return -ELOOP;
2919         case S_IFDIR:
2920                 if (acc_mode & MAY_WRITE)
2921                         return -EISDIR;
2922                 break;
2923         case S_IFBLK:
2924         case S_IFCHR:
2925                 if (!may_open_dev(path))
2926                         return -EACCES;
2927                 /*FALLTHRU*/
2928         case S_IFIFO:
2929         case S_IFSOCK:
2930                 flag &= ~O_TRUNC;
2931                 break;
2932         }
2933
2934         error = inode_permission(inode, MAY_OPEN | acc_mode);
2935         if (error)
2936                 return error;
2937
2938         /*
2939          * An append-only file must be opened in append mode for writing.
2940          */
2941         if (IS_APPEND(inode)) {
2942                 if  ((flag & O_ACCMODE) != O_RDONLY && !(flag & O_APPEND))
2943                         return -EPERM;
2944                 if (flag & O_TRUNC)
2945                         return -EPERM;
2946         }
2947
2948         /* O_NOATIME can only be set by the owner or superuser */
2949         if (flag & O_NOATIME && !inode_owner_or_capable(inode))
2950                 return -EPERM;
2951
2952         return 0;
2953 }
2954
2955 static int handle_truncate(struct file *filp)
2956 {
2957         const struct path *path = &filp->f_path;
2958         struct inode *inode = path->dentry->d_inode;
2959         int error = get_write_access(inode);
2960         if (error)
2961                 return error;
2962         /*
2963          * Refuse to truncate files with mandatory locks held on them.
2964          */
2965         error = locks_verify_locked(filp);
2966         if (!error)
2967                 error = security_path_truncate(path);
2968         if (!error) {
2969                 error = do_truncate(path->dentry, 0,
2970                                     ATTR_MTIME|ATTR_CTIME|ATTR_OPEN,
2971                                     filp);
2972         }
2973         put_write_access(inode);
2974         return error;
2975 }
2976
2977 static inline int open_to_namei_flags(int flag)
2978 {
2979         if ((flag & O_ACCMODE) == 3)
2980                 flag--;
2981         return flag;
2982 }
2983
2984 static int may_o_create(const struct path *dir, struct dentry *dentry, umode_t mode)
2985 {
2986         struct user_namespace *s_user_ns;
2987         int error = security_path_mknod(dir, dentry, mode, 0);
2988         if (error)
2989                 return error;
2990
2991         s_user_ns = dir->dentry->d_sb->s_user_ns;
2992         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
2993             !kgid_has_mapping(s_user_ns, current_fsgid()))
2994                 return -EOVERFLOW;
2995
2996         error = inode_permission(dir->dentry->d_inode, MAY_WRITE | MAY_EXEC);
2997         if (error)
2998                 return error;
2999
3000         return security_inode_create(dir->dentry->d_inode, dentry, mode);
3001 }
3002
3003 /*
3004  * Attempt to atomically look up, create and open a file from a negative
3005  * dentry.
3006  *
3007  * Returns 0 if successful.  The file will have been created and attached to
3008  * @file by the filesystem calling finish_open().
3009  *
3010  * Returns 1 if the file was looked up only or didn't need creating.  The
3011  * caller will need to perform the open themselves.  @path will have been
3012  * updated to point to the new dentry.  This may be negative.
3013  *
3014  * Returns an error code otherwise.
3015  */
3016 static int atomic_open(struct nameidata *nd, struct dentry *dentry,
3017                         struct path *path, struct file *file,
3018                         const struct open_flags *op,
3019                         int open_flag, umode_t mode,
3020                         int *opened)
3021 {
3022         struct dentry *const DENTRY_NOT_SET = (void *) -1UL;
3023         struct inode *dir =  nd->path.dentry->d_inode;
3024         int error;
3025
3026         if (!(~open_flag & (O_EXCL | O_CREAT))) /* both O_EXCL and O_CREAT */
3027                 open_flag &= ~O_TRUNC;
3028
3029         if (nd->flags & LOOKUP_DIRECTORY)
3030                 open_flag |= O_DIRECTORY;
3031
3032         file->f_path.dentry = DENTRY_NOT_SET;
3033         file->f_path.mnt = nd->path.mnt;
3034         error = dir->i_op->atomic_open(dir, dentry, file,
3035                                        open_to_namei_flags(open_flag),
3036                                        mode, opened);
3037         d_lookup_done(dentry);
3038         if (!error) {
3039                 /*
3040                  * We didn't have the inode before the open, so check open
3041                  * permission here.
3042                  */
3043                 int acc_mode = op->acc_mode;
3044                 if (*opened & FILE_CREATED) {
3045                         WARN_ON(!(open_flag & O_CREAT));
3046                         fsnotify_create(dir, dentry);
3047                         acc_mode = 0;
3048                 }
3049                 error = may_open(&file->f_path, acc_mode, open_flag);
3050                 if (WARN_ON(error > 0))
3051                         error = -EINVAL;
3052         } else if (error > 0) {
3053                 if (WARN_ON(file->f_path.dentry == DENTRY_NOT_SET)) {
3054                         error = -EIO;
3055                 } else {
3056                         if (file->f_path.dentry) {
3057                                 dput(dentry);
3058                                 dentry = file->f_path.dentry;
3059                         }
3060                         if (*opened & FILE_CREATED)
3061                                 fsnotify_create(dir, dentry);
3062                         if (unlikely(d_is_negative(dentry))) {
3063                                 error = -ENOENT;
3064                         } else {
3065                                 path->dentry = dentry;
3066                                 path->mnt = nd->path.mnt;
3067                                 return 1;
3068                         }
3069                 }
3070         }
3071         dput(dentry);
3072         return error;
3073 }
3074
3075 /*
3076  * Look up and maybe create and open the last component.
3077  *
3078  * Must be called with i_mutex held on parent.
3079  *
3080  * Returns 0 if the file was successfully atomically created (if necessary) and
3081  * opened.  In this case the file will be returned attached to @file.
3082  *
3083  * Returns 1 if the file was not completely opened at this time, though lookups
3084  * and creations will have been performed and the dentry returned in @path will
3085  * be positive upon return if O_CREAT was specified.  If O_CREAT wasn't
3086  * specified then a negative dentry may be returned.
3087  *
3088  * An error code is returned otherwise.
3089  *
3090  * FILE_CREATE will be set in @*opened if the dentry was created and will be
3091  * cleared otherwise prior to returning.
3092  */
3093 static int lookup_open(struct nameidata *nd, struct path *path,
3094                         struct file *file,
3095                         const struct open_flags *op,
3096                         bool got_write, int *opened)
3097 {
3098         struct dentry *dir = nd->path.dentry;
3099         struct inode *dir_inode = dir->d_inode;
3100         int open_flag = op->open_flag;
3101         struct dentry *dentry;
3102         int error, create_error = 0;
3103         umode_t mode = op->mode;
3104         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
3105
3106         if (unlikely(IS_DEADDIR(dir_inode)))
3107                 return -ENOENT;
3108
3109         *opened &= ~FILE_CREATED;
3110         dentry = d_lookup(dir, &nd->last);
3111         for (;;) {
3112                 if (!dentry) {
3113                         dentry = d_alloc_parallel(dir, &nd->last, &wq);
3114                         if (IS_ERR(dentry))
3115                                 return PTR_ERR(dentry);
3116                 }
3117                 if (d_in_lookup(dentry))
3118                         break;
3119
3120                 error = d_revalidate(dentry, nd->flags);
3121                 if (likely(error > 0))
3122                         break;
3123                 if (error)
3124                         goto out_dput;
3125                 d_invalidate(dentry);
3126                 dput(dentry);
3127                 dentry = NULL;
3128         }
3129         if (dentry->d_inode) {
3130                 /* Cached positive dentry: will open in f_op->open */
3131                 goto out_no_open;
3132         }
3133
3134         /*
3135          * Checking write permission is tricky, bacuse we don't know if we are
3136          * going to actually need it: O_CREAT opens should work as long as the
3137          * file exists.  But checking existence breaks atomicity.  The trick is
3138          * to check access and if not granted clear O_CREAT from the flags.
3139          *
3140          * Another problem is returing the "right" error value (e.g. for an
3141          * O_EXCL open we want to return EEXIST not EROFS).
3142          */
3143         if (open_flag & O_CREAT) {
3144                 if (!IS_POSIXACL(dir->d_inode))
3145                         mode &= ~current_umask();
3146                 if (unlikely(!got_write)) {
3147                         create_error = -EROFS;
3148                         open_flag &= ~O_CREAT;
3149                         if (open_flag & (O_EXCL | O_TRUNC))
3150                                 goto no_open;
3151                         /* No side effects, safe to clear O_CREAT */
3152                 } else {
3153                         create_error = may_o_create(&nd->path, dentry, mode);
3154                         if (create_error) {
3155                                 open_flag &= ~O_CREAT;
3156                                 if (open_flag & O_EXCL)
3157                                         goto no_open;
3158                         }
3159                 }
3160         } else if ((open_flag & (O_TRUNC|O_WRONLY|O_RDWR)) &&
3161                    unlikely(!got_write)) {
3162                 /*
3163                  * No O_CREATE -> atomicity not a requirement -> fall
3164                  * back to lookup + open
3165                  */
3166                 goto no_open;
3167         }
3168
3169         if (dir_inode->i_op->atomic_open) {
3170                 error = atomic_open(nd, dentry, path, file, op, open_flag,
3171                                     mode, opened);
3172                 if (unlikely(error == -ENOENT) && create_error)
3173                         error = create_error;
3174                 return error;
3175         }
3176
3177 no_open:
3178         if (d_in_lookup(dentry)) {
3179                 struct dentry *res = dir_inode->i_op->lookup(dir_inode, dentry,
3180                                                              nd->flags);
3181                 d_lookup_done(dentry);
3182                 if (unlikely(res)) {
3183                         if (IS_ERR(res)) {
3184                                 error = PTR_ERR(res);
3185                                 goto out_dput;
3186                         }
3187                         dput(dentry);
3188                         dentry = res;
3189                 }
3190         }
3191
3192         /* Negative dentry, just create the file */
3193         if (!dentry->d_inode && (open_flag & O_CREAT)) {
3194                 *opened |= FILE_CREATED;
3195                 audit_inode_child(dir_inode, dentry, AUDIT_TYPE_CHILD_CREATE);
3196                 if (!dir_inode->i_op->create) {
3197                         error = -EACCES;
3198                         goto out_dput;
3199                 }
3200                 error = dir_inode->i_op->create(dir_inode, dentry, mode,
3201                                                 open_flag & O_EXCL);
3202                 if (error)
3203                         goto out_dput;
3204                 fsnotify_create(dir_inode, dentry);
3205         }
3206         if (unlikely(create_error) && !dentry->d_inode) {
3207                 error = create_error;
3208                 goto out_dput;
3209         }
3210 out_no_open:
3211         path->dentry = dentry;
3212         path->mnt = nd->path.mnt;
3213         return 1;
3214
3215 out_dput:
3216         dput(dentry);
3217         return error;
3218 }
3219
3220 /*
3221  * Handle the last step of open()
3222  */
3223 static int do_last(struct nameidata *nd,
3224                    struct file *file, const struct open_flags *op,
3225                    int *opened)
3226 {
3227         struct dentry *dir = nd->path.dentry;
3228         int open_flag = op->open_flag;
3229         bool will_truncate = (open_flag & O_TRUNC) != 0;
3230         bool got_write = false;
3231         int acc_mode = op->acc_mode;
3232         unsigned seq;
3233         struct inode *inode;
3234         struct path path;
3235         int error;
3236
3237         nd->flags &= ~LOOKUP_PARENT;
3238         nd->flags |= op->intent;
3239
3240         if (nd->last_type != LAST_NORM) {
3241                 error = handle_dots(nd, nd->last_type);
3242                 if (unlikely(error))
3243                         return error;
3244                 goto finish_open;
3245         }
3246
3247         if (!(open_flag & O_CREAT)) {
3248                 if (nd->last.name[nd->last.len])
3249                         nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
3250                 /* we _can_ be in RCU mode here */
3251                 error = lookup_fast(nd, &path, &inode, &seq);
3252                 if (likely(error > 0))
3253                         goto finish_lookup;
3254
3255                 if (error < 0)
3256                         return error;
3257
3258                 BUG_ON(nd->inode != dir->d_inode);
3259                 BUG_ON(nd->flags & LOOKUP_RCU);
3260         } else {
3261                 /* create side of things */
3262                 /*
3263                  * This will *only* deal with leaving RCU mode - LOOKUP_JUMPED
3264                  * has been cleared when we got to the last component we are
3265                  * about to look up
3266                  */
3267                 error = complete_walk(nd);
3268                 if (error)
3269                         return error;
3270
3271                 audit_inode(nd->name, dir, LOOKUP_PARENT);
3272                 /* trailing slashes? */
3273                 if (unlikely(nd->last.name[nd->last.len]))
3274                         return -EISDIR;
3275         }
3276
3277         if (open_flag & (O_CREAT | O_TRUNC | O_WRONLY | O_RDWR)) {
3278                 error = mnt_want_write(nd->path.mnt);
3279                 if (!error)
3280                         got_write = true;
3281                 /*
3282                  * do _not_ fail yet - we might not need that or fail with
3283                  * a different error; let lookup_open() decide; we'll be
3284                  * dropping this one anyway.
3285                  */
3286         }
3287         if (open_flag & O_CREAT)
3288                 inode_lock(dir->d_inode);
3289         else
3290                 inode_lock_shared(dir->d_inode);
3291         error = lookup_open(nd, &path, file, op, got_write, opened);
3292         if (open_flag & O_CREAT)
3293                 inode_unlock(dir->d_inode);
3294         else
3295                 inode_unlock_shared(dir->d_inode);
3296
3297         if (error <= 0) {
3298                 if (error)
3299                         goto out;
3300
3301                 if ((*opened & FILE_CREATED) ||
3302                     !S_ISREG(file_inode(file)->i_mode))
3303                         will_truncate = false;
3304
3305                 audit_inode(nd->name, file->f_path.dentry, 0);
3306                 goto opened;
3307         }
3308
3309         if (*opened & FILE_CREATED) {
3310                 /* Don't check for write permission, don't truncate */
3311                 open_flag &= ~O_TRUNC;
3312                 will_truncate = false;
3313                 acc_mode = 0;
3314                 path_to_nameidata(&path, nd);
3315                 goto finish_open_created;
3316         }
3317
3318         /*
3319          * If atomic_open() acquired write access it is dropped now due to
3320          * possible mount and symlink following (this might be optimized away if
3321          * necessary...)
3322          */
3323         if (got_write) {
3324                 mnt_drop_write(nd->path.mnt);
3325                 got_write = false;
3326         }
3327
3328         error = follow_managed(&path, nd);
3329         if (unlikely(error < 0))
3330                 return error;
3331
3332         if (unlikely(d_is_negative(path.dentry))) {
3333                 path_to_nameidata(&path, nd);
3334                 return -ENOENT;
3335         }
3336
3337         /*
3338          * create/update audit record if it already exists.
3339          */
3340         audit_inode(nd->name, path.dentry, 0);
3341
3342         if (unlikely((open_flag & (O_EXCL | O_CREAT)) == (O_EXCL | O_CREAT))) {
3343                 path_to_nameidata(&path, nd);
3344                 return -EEXIST;
3345         }
3346
3347         seq = 0;        /* out of RCU mode, so the value doesn't matter */
3348         inode = d_backing_inode(path.dentry);
3349 finish_lookup:
3350         error = step_into(nd, &path, 0, inode, seq);
3351         if (unlikely(error))
3352                 return error;
3353 finish_open:
3354         /* Why this, you ask?  _Now_ we might have grown LOOKUP_JUMPED... */
3355         error = complete_walk(nd);
3356         if (error)
3357                 return error;
3358         audit_inode(nd->name, nd->path.dentry, 0);
3359         error = -EISDIR;
3360         if ((open_flag & O_CREAT) && d_is_dir(nd->path.dentry))
3361                 goto out;
3362         error = -ENOTDIR;
3363         if ((nd->flags & LOOKUP_DIRECTORY) && !d_can_lookup(nd->path.dentry))
3364                 goto out;
3365         if (!d_is_reg(nd->path.dentry))
3366                 will_truncate = false;
3367
3368         if (will_truncate) {
3369                 error = mnt_want_write(nd->path.mnt);
3370                 if (error)
3371                         goto out;
3372                 got_write = true;
3373         }
3374 finish_open_created:
3375         error = may_open(&nd->path, acc_mode, open_flag);
3376         if (error)
3377                 goto out;
3378         BUG_ON(*opened & FILE_OPENED); /* once it's opened, it's opened */
3379         error = vfs_open(&nd->path, file, current_cred());
3380         if (error)
3381                 goto out;
3382         *opened |= FILE_OPENED;
3383 opened:
3384         error = open_check_o_direct(file);
3385         if (!error)
3386                 error = ima_file_check(file, op->acc_mode, *opened);
3387         if (!error && will_truncate)
3388                 error = handle_truncate(file);
3389 out:
3390         if (unlikely(error) && (*opened & FILE_OPENED))
3391                 fput(file);
3392         if (unlikely(error > 0)) {
3393                 WARN_ON(1);
3394                 error = -EINVAL;
3395         }
3396         if (got_write)
3397                 mnt_drop_write(nd->path.mnt);
3398         return error;
3399 }
3400
3401 struct dentry *vfs_tmpfile(struct dentry *dentry, umode_t mode, int open_flag)
3402 {
3403         struct dentry *child = NULL;
3404         struct inode *dir = dentry->d_inode;
3405         struct inode *inode;
3406         int error;
3407
3408         /* we want directory to be writable */
3409         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
3410         if (error)
3411                 goto out_err;
3412         error = -EOPNOTSUPP;
3413         if (!dir->i_op->tmpfile)
3414                 goto out_err;
3415         error = -ENOMEM;
3416         child = d_alloc(dentry, &slash_name);
3417         if (unlikely(!child))
3418                 goto out_err;
3419         error = dir->i_op->tmpfile(dir, child, mode);
3420         if (error)
3421                 goto out_err;
3422         error = -ENOENT;
3423         inode = child->d_inode;
3424         if (unlikely(!inode))
3425                 goto out_err;
3426         if (!(open_flag & O_EXCL)) {
3427                 spin_lock(&inode->i_lock);
3428                 inode->i_state |= I_LINKABLE;
3429                 spin_unlock(&inode->i_lock);
3430         }
3431         return child;
3432
3433 out_err:
3434         dput(child);
3435         return ERR_PTR(error);
3436 }
3437 EXPORT_SYMBOL(vfs_tmpfile);
3438
3439 static int do_tmpfile(struct nameidata *nd, unsigned flags,
3440                 const struct open_flags *op,
3441                 struct file *file, int *opened)
3442 {
3443         struct dentry *child;
3444         struct path path;
3445         int error = path_lookupat(nd, flags | LOOKUP_DIRECTORY, &path);
3446         if (unlikely(error))
3447                 return error;
3448         error = mnt_want_write(path.mnt);
3449         if (unlikely(error))
3450                 goto out;
3451         child = vfs_tmpfile(path.dentry, op->mode, op->open_flag);
3452         error = PTR_ERR(child);
3453         if (IS_ERR(child))
3454                 goto out2;
3455         dput(path.dentry);
3456         path.dentry = child;
3457         audit_inode(nd->name, child, 0);
3458         /* Don't check for other permissions, the inode was just created */
3459         error = may_open(&path, 0, op->open_flag);
3460         if (error)
3461                 goto out2;
3462         file->f_path.mnt = path.mnt;
3463         error = finish_open(file, child, NULL, opened);
3464         if (error)
3465                 goto out2;
3466         error = open_check_o_direct(file);
3467         if (error)
3468                 fput(file);
3469 out2:
3470         mnt_drop_write(path.mnt);
3471 out:
3472         path_put(&path);
3473         return error;
3474 }
3475
3476 static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
3477 {
3478         struct path path;
3479         int error = path_lookupat(nd, flags, &path);
3480         if (!error) {
3481                 audit_inode(nd->name, path.dentry, 0);
3482                 error = vfs_open(&path, file, current_cred());
3483                 path_put(&path);
3484         }
3485         return error;
3486 }
3487
3488 static struct file *path_openat(struct nameidata *nd,
3489                         const struct open_flags *op, unsigned flags)
3490 {
3491         const char *s;
3492         struct file *file;
3493         int opened = 0;
3494         int error;
3495
3496         file = get_empty_filp();
3497         if (IS_ERR(file))
3498                 return file;
3499
3500         file->f_flags = op->open_flag;
3501
3502         if (unlikely(file->f_flags & __O_TMPFILE)) {
3503                 error = do_tmpfile(nd, flags, op, file, &opened);
3504                 goto out2;
3505         }
3506
3507         if (unlikely(file->f_flags & O_PATH)) {
3508                 error = do_o_path(nd, flags, file);
3509                 if (!error)
3510                         opened |= FILE_OPENED;
3511                 goto out2;
3512         }
3513
3514         s = path_init(nd, flags);
3515         if (IS_ERR(s)) {
3516                 put_filp(file);
3517                 return ERR_CAST(s);
3518         }
3519         while (!(error = link_path_walk(s, nd)) &&
3520                 (error = do_last(nd, file, op, &opened)) > 0) {
3521                 nd->flags &= ~(LOOKUP_OPEN|LOOKUP_CREATE|LOOKUP_EXCL);
3522                 s = trailing_symlink(nd);
3523                 if (IS_ERR(s)) {
3524                         error = PTR_ERR(s);
3525                         break;
3526                 }
3527         }
3528         terminate_walk(nd);
3529 out2:
3530         if (!(opened & FILE_OPENED)) {
3531                 BUG_ON(!error);
3532                 put_filp(file);
3533         }
3534         if (unlikely(error)) {
3535                 if (error == -EOPENSTALE) {
3536                         if (flags & LOOKUP_RCU)
3537                                 error = -ECHILD;
3538                         else
3539                                 error = -ESTALE;
3540                 }
3541                 file = ERR_PTR(error);
3542         }
3543         return file;
3544 }
3545
3546 struct file *do_filp_open(int dfd, struct filename *pathname,
3547                 const struct open_flags *op)
3548 {
3549         struct nameidata nd;
3550         int flags = op->lookup_flags;
3551         struct file *filp;
3552
3553         set_nameidata(&nd, dfd, pathname);
3554         filp = path_openat(&nd, op, flags | LOOKUP_RCU);
3555         if (unlikely(filp == ERR_PTR(-ECHILD)))
3556                 filp = path_openat(&nd, op, flags);
3557         if (unlikely(filp == ERR_PTR(-ESTALE)))
3558                 filp = path_openat(&nd, op, flags | LOOKUP_REVAL);
3559         restore_nameidata();
3560         return filp;
3561 }
3562
3563 struct file *do_file_open_root(struct dentry *dentry, struct vfsmount *mnt,
3564                 const char *name, const struct open_flags *op)
3565 {
3566         struct nameidata nd;
3567         struct file *file;
3568         struct filename *filename;
3569         int flags = op->lookup_flags | LOOKUP_ROOT;
3570
3571         nd.root.mnt = mnt;
3572         nd.root.dentry = dentry;
3573
3574         if (d_is_symlink(dentry) && op->intent & LOOKUP_OPEN)
3575                 return ERR_PTR(-ELOOP);
3576
3577         filename = getname_kernel(name);
3578         if (IS_ERR(filename))
3579                 return ERR_CAST(filename);
3580
3581         set_nameidata(&nd, -1, filename);
3582         file = path_openat(&nd, op, flags | LOOKUP_RCU);
3583         if (unlikely(file == ERR_PTR(-ECHILD)))
3584                 file = path_openat(&nd, op, flags);
3585         if (unlikely(file == ERR_PTR(-ESTALE)))
3586                 file = path_openat(&nd, op, flags | LOOKUP_REVAL);
3587         restore_nameidata();
3588         putname(filename);
3589         return file;
3590 }
3591
3592 static struct dentry *filename_create(int dfd, struct filename *name,
3593                                 struct path *path, unsigned int lookup_flags)
3594 {
3595         struct dentry *dentry = ERR_PTR(-EEXIST);
3596         struct qstr last;
3597         int type;
3598         int err2;
3599         int error;
3600         bool is_dir = (lookup_flags & LOOKUP_DIRECTORY);
3601
3602         /*
3603          * Note that only LOOKUP_REVAL and LOOKUP_DIRECTORY matter here. Any
3604          * other flags passed in are ignored!
3605          */
3606         lookup_flags &= LOOKUP_REVAL;
3607
3608         name = filename_parentat(dfd, name, lookup_flags, path, &last, &type);
3609         if (IS_ERR(name))
3610                 return ERR_CAST(name);
3611
3612         /*
3613          * Yucky last component or no last component at all?
3614          * (foo/., foo/.., /////)
3615          */
3616         if (unlikely(type != LAST_NORM))
3617                 goto out;
3618
3619         /* don't fail immediately if it's r/o, at least try to report other errors */
3620         err2 = mnt_want_write(path->mnt);
3621         /*
3622          * Do the final lookup.
3623          */
3624         lookup_flags |= LOOKUP_CREATE | LOOKUP_EXCL;
3625         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
3626         dentry = __lookup_hash(&last, path->dentry, lookup_flags);
3627         if (IS_ERR(dentry))
3628                 goto unlock;
3629
3630         error = -EEXIST;
3631         if (d_is_positive(dentry))
3632                 goto fail;
3633
3634         /*
3635          * Special case - lookup gave negative, but... we had foo/bar/
3636          * From the vfs_mknod() POV we just have a negative dentry -
3637          * all is fine. Let's be bastards - you had / on the end, you've
3638          * been asking for (non-existent) directory. -ENOENT for you.
3639          */
3640         if (unlikely(!is_dir && last.name[last.len])) {
3641                 error = -ENOENT;
3642                 goto fail;
3643         }
3644         if (unlikely(err2)) {
3645                 error = err2;
3646                 goto fail;
3647         }
3648         putname(name);
3649         return dentry;
3650 fail:
3651         dput(dentry);
3652         dentry = ERR_PTR(error);
3653 unlock:
3654         inode_unlock(path->dentry->d_inode);
3655         if (!err2)
3656                 mnt_drop_write(path->mnt);
3657 out:
3658         path_put(path);
3659         putname(name);
3660         return dentry;
3661 }
3662
3663 struct dentry *kern_path_create(int dfd, const char *pathname,
3664                                 struct path *path, unsigned int lookup_flags)
3665 {
3666         return filename_create(dfd, getname_kernel(pathname),
3667                                 path, lookup_flags);
3668 }
3669 EXPORT_SYMBOL(kern_path_create);
3670
3671 void done_path_create(struct path *path, struct dentry *dentry)
3672 {
3673         dput(dentry);
3674         inode_unlock(path->dentry->d_inode);
3675         mnt_drop_write(path->mnt);
3676         path_put(path);
3677 }
3678 EXPORT_SYMBOL(done_path_create);
3679
3680 inline struct dentry *user_path_create(int dfd, const char __user *pathname,
3681                                 struct path *path, unsigned int lookup_flags)
3682 {
3683         return filename_create(dfd, getname(pathname), path, lookup_flags);
3684 }
3685 EXPORT_SYMBOL(user_path_create);
3686
3687 int vfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3688 {
3689         int error = may_create(dir, dentry);
3690
3691         if (error)
3692                 return error;
3693
3694         if ((S_ISCHR(mode) || S_ISBLK(mode)) && !capable(CAP_MKNOD))
3695                 return -EPERM;
3696
3697         if (!dir->i_op->mknod)
3698                 return -EPERM;
3699
3700         error = devcgroup_inode_mknod(mode, dev);
3701         if (error)
3702                 return error;
3703
3704         error = security_inode_mknod(dir, dentry, mode, dev);
3705         if (error)
3706                 return error;
3707
3708         error = dir->i_op->mknod(dir, dentry, mode, dev);
3709         if (!error)
3710                 fsnotify_create(dir, dentry);
3711         return error;
3712 }
3713 EXPORT_SYMBOL(vfs_mknod);
3714
3715 static int may_mknod(umode_t mode)
3716 {
3717         switch (mode & S_IFMT) {
3718         case S_IFREG:
3719         case S_IFCHR:
3720         case S_IFBLK:
3721         case S_IFIFO:
3722         case S_IFSOCK:
3723         case 0: /* zero mode translates to S_IFREG */
3724                 return 0;
3725         case S_IFDIR:
3726                 return -EPERM;
3727         default:
3728                 return -EINVAL;
3729         }
3730 }
3731
3732 SYSCALL_DEFINE4(mknodat, int, dfd, const char __user *, filename, umode_t, mode,
3733                 unsigned, dev)
3734 {
3735         struct dentry *dentry;
3736         struct path path;
3737         int error;
3738         unsigned int lookup_flags = 0;
3739
3740         error = may_mknod(mode);
3741         if (error)
3742                 return error;
3743 retry:
3744         dentry = user_path_create(dfd, filename, &path, lookup_flags);
3745         if (IS_ERR(dentry))
3746                 return PTR_ERR(dentry);
3747
3748         if (!IS_POSIXACL(path.dentry->d_inode))
3749                 mode &= ~current_umask();
3750         error = security_path_mknod(&path, dentry, mode, dev);
3751         if (error)
3752                 goto out;
3753         switch (mode & S_IFMT) {
3754                 case 0: case S_IFREG:
3755                         error = vfs_create(path.dentry->d_inode,dentry,mode,true);
3756                         if (!error)
3757                                 ima_post_path_mknod(dentry);
3758                         break;
3759                 case S_IFCHR: case S_IFBLK:
3760                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,
3761                                         new_decode_dev(dev));
3762                         break;
3763                 case S_IFIFO: case S_IFSOCK:
3764                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,0);
3765                         break;
3766         }
3767 out:
3768         done_path_create(&path, dentry);
3769         if (retry_estale(error, lookup_flags)) {
3770                 lookup_flags |= LOOKUP_REVAL;
3771                 goto retry;
3772         }
3773         return error;
3774 }
3775
3776 SYSCALL_DEFINE3(mknod, const char __user *, filename, umode_t, mode, unsigned, dev)
3777 {
3778         return sys_mknodat(AT_FDCWD, filename, mode, dev);
3779 }
3780
3781 int vfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3782 {
3783         int error = may_create(dir, dentry);
3784         unsigned max_links = dir->i_sb->s_max_links;
3785
3786         if (error)
3787                 return error;
3788
3789         if (!dir->i_op->mkdir)
3790                 return -EPERM;
3791
3792         mode &= (S_IRWXUGO|S_ISVTX);
3793         error = security_inode_mkdir(dir, dentry, mode);
3794         if (error)
3795                 return error;
3796
3797         if (max_links && dir->i_nlink >= max_links)
3798                 return -EMLINK;
3799
3800         error = dir->i_op->mkdir(dir, dentry, mode);
3801         if (!error)
3802                 fsnotify_mkdir(dir, dentry);
3803         return error;
3804 }
3805 EXPORT_SYMBOL(vfs_mkdir);
3806
3807 SYSCALL_DEFINE3(mkdirat, int, dfd, const char __user *, pathname, umode_t, mode)
3808 {
3809         struct dentry *dentry;
3810         struct path path;
3811         int error;
3812         unsigned int lookup_flags = LOOKUP_DIRECTORY;
3813
3814 retry:
3815         dentry = user_path_create(dfd, pathname, &path, lookup_flags);
3816         if (IS_ERR(dentry))
3817                 return PTR_ERR(dentry);
3818
3819         if (!IS_POSIXACL(path.dentry->d_inode))
3820                 mode &= ~current_umask();
3821         error = security_path_mkdir(&path, dentry, mode);
3822         if (!error)
3823                 error = vfs_mkdir(path.dentry->d_inode, dentry, mode);
3824         done_path_create(&path, dentry);
3825         if (retry_estale(error, lookup_flags)) {
3826                 lookup_flags |= LOOKUP_REVAL;
3827                 goto retry;
3828         }
3829         return error;
3830 }
3831
3832 SYSCALL_DEFINE2(mkdir, const char __user *, pathname, umode_t, mode)
3833 {
3834         return sys_mkdirat(AT_FDCWD, pathname, mode);
3835 }
3836
3837 int vfs_rmdir(struct inode *dir, struct dentry *dentry)
3838 {
3839         int error = may_delete(dir, dentry, 1);
3840
3841         if (error)
3842                 return error;
3843
3844         if (!dir->i_op->rmdir)
3845                 return -EPERM;
3846
3847         dget(dentry);
3848         inode_lock(dentry->d_inode);
3849
3850         error = -EBUSY;
3851         if (is_local_mountpoint(dentry))
3852                 goto out;
3853
3854         error = security_inode_rmdir(dir, dentry);
3855         if (error)
3856                 goto out;
3857
3858         shrink_dcache_parent(dentry);
3859         error = dir->i_op->rmdir(dir, dentry);
3860         if (error)
3861                 goto out;
3862
3863         dentry->d_inode->i_flags |= S_DEAD;
3864         dont_mount(dentry);
3865         detach_mounts(dentry);
3866
3867 out:
3868         inode_unlock(dentry->d_inode);
3869         dput(dentry);
3870         if (!error)
3871                 d_delete(dentry);
3872         return error;
3873 }
3874 EXPORT_SYMBOL(vfs_rmdir);
3875
3876 static long do_rmdir(int dfd, const char __user *pathname)
3877 {
3878         int error = 0;
3879         struct filename *name;
3880         struct dentry *dentry;
3881         struct path path;
3882         struct qstr last;
3883         int type;
3884         unsigned int lookup_flags = 0;
3885 retry:
3886         name = filename_parentat(dfd, getname(pathname), lookup_flags,
3887                                 &path, &last, &type);
3888         if (IS_ERR(name))
3889                 return PTR_ERR(name);
3890
3891         switch (type) {
3892         case LAST_DOTDOT:
3893                 error = -ENOTEMPTY;
3894                 goto exit1;
3895         case LAST_DOT:
3896                 error = -EINVAL;
3897                 goto exit1;
3898         case LAST_ROOT:
3899                 error = -EBUSY;
3900                 goto exit1;
3901         }
3902
3903         error = mnt_want_write(path.mnt);
3904         if (error)
3905                 goto exit1;
3906
3907         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3908         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3909         error = PTR_ERR(dentry);
3910         if (IS_ERR(dentry))
3911                 goto exit2;
3912         if (!dentry->d_inode) {
3913                 error = -ENOENT;
3914                 goto exit3;
3915         }
3916         error = security_path_rmdir(&path, dentry);
3917         if (error)
3918                 goto exit3;
3919         error = vfs_rmdir(path.dentry->d_inode, dentry);
3920 exit3:
3921         dput(dentry);
3922 exit2:
3923         inode_unlock(path.dentry->d_inode);
3924         mnt_drop_write(path.mnt);
3925 exit1:
3926         path_put(&path);
3927         putname(name);
3928         if (retry_estale(error, lookup_flags)) {
3929                 lookup_flags |= LOOKUP_REVAL;
3930                 goto retry;
3931         }
3932         return error;
3933 }
3934
3935 SYSCALL_DEFINE1(rmdir, const char __user *, pathname)
3936 {
3937         return do_rmdir(AT_FDCWD, pathname);
3938 }
3939
3940 /**
3941  * vfs_unlink - unlink a filesystem object
3942  * @dir:        parent directory
3943  * @dentry:     victim
3944  * @delegated_inode: returns victim inode, if the inode is delegated.
3945  *
3946  * The caller must hold dir->i_mutex.
3947  *
3948  * If vfs_unlink discovers a delegation, it will return -EWOULDBLOCK and
3949  * return a reference to the inode in delegated_inode.  The caller
3950  * should then break the delegation on that inode and retry.  Because
3951  * breaking a delegation may take a long time, the caller should drop
3952  * dir->i_mutex before doing so.
3953  *
3954  * Alternatively, a caller may pass NULL for delegated_inode.  This may
3955  * be appropriate for callers that expect the underlying filesystem not
3956  * to be NFS exported.
3957  */
3958 int vfs_unlink(struct inode *dir, struct dentry *dentry, struct inode **delegated_inode)
3959 {
3960         struct inode *target = dentry->d_inode;
3961         int error = may_delete(dir, dentry, 0);
3962
3963         if (error)
3964                 return error;
3965
3966         if (!dir->i_op->unlink)
3967                 return -EPERM;
3968
3969         inode_lock(target);
3970         if (is_local_mountpoint(dentry))
3971                 error = -EBUSY;
3972         else {
3973                 error = security_inode_unlink(dir, dentry);
3974                 if (!error) {
3975                         error = try_break_deleg(target, delegated_inode);
3976                         if (error)
3977                                 goto out;
3978                         error = dir->i_op->unlink(dir, dentry);
3979                         if (!error) {
3980                                 dont_mount(dentry);
3981                                 detach_mounts(dentry);
3982                         }
3983                 }
3984         }
3985 out:
3986         inode_unlock(target);
3987
3988         /* We don't d_delete() NFS sillyrenamed files--they still exist. */
3989         if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
3990                 fsnotify_link_count(target);
3991                 d_delete(dentry);
3992         }
3993
3994         return error;
3995 }
3996 EXPORT_SYMBOL(vfs_unlink);
3997
3998 /*
3999  * Make sure that the actual truncation of the file will occur outside its
4000  * directory's i_mutex.  Truncate can take a long time if there is a lot of
4001  * writeout happening, and we don't want to prevent access to the directory
4002  * while waiting on the I/O.
4003  */
4004 long do_unlinkat(int dfd, struct filename *name)
4005 {
4006         int error;
4007         struct dentry *dentry;
4008         struct path path;
4009         struct qstr last;
4010         int type;
4011         struct inode *inode = NULL;
4012         struct inode *delegated_inode = NULL;
4013         unsigned int lookup_flags = 0;
4014 retry:
4015         name = filename_parentat(dfd, name, lookup_flags, &path, &last, &type);
4016         if (IS_ERR(name))
4017                 return PTR_ERR(name);
4018
4019         error = -EISDIR;
4020         if (type != LAST_NORM)
4021                 goto exit1;
4022
4023         error = mnt_want_write(path.mnt);
4024         if (error)
4025                 goto exit1;
4026 retry_deleg:
4027         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
4028         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
4029         error = PTR_ERR(dentry);
4030         if (!IS_ERR(dentry)) {
4031                 /* Why not before? Because we want correct error value */
4032                 if (last.name[last.len])
4033                         goto slashes;
4034                 inode = dentry->d_inode;
4035                 if (d_is_negative(dentry))
4036                         goto slashes;
4037                 ihold(inode);
4038                 error = security_path_unlink(&path, dentry);
4039                 if (error)
4040                         goto exit2;
4041                 error = vfs_unlink(path.dentry->d_inode, dentry, &delegated_inode);
4042 exit2:
4043                 dput(dentry);
4044         }
4045         inode_unlock(path.dentry->d_inode);
4046         if (inode)
4047                 iput(inode);    /* truncate the inode here */
4048         inode = NULL;
4049         if (delegated_inode) {
4050                 error = break_deleg_wait(&delegated_inode);
4051                 if (!error)
4052                         goto retry_deleg;
4053         }
4054         mnt_drop_write(path.mnt);
4055 exit1:
4056         path_put(&path);
4057         if (retry_estale(error, lookup_flags)) {
4058                 lookup_flags |= LOOKUP_REVAL;
4059                 inode = NULL;
4060                 goto retry;
4061         }
4062         putname(name);
4063         return error;
4064
4065 slashes:
4066         if (d_is_negative(dentry))
4067                 error = -ENOENT;
4068         else if (d_is_dir(dentry))
4069                 error = -EISDIR;
4070         else
4071                 error = -ENOTDIR;
4072         goto exit2;
4073 }
4074
4075 SYSCALL_DEFINE3(unlinkat, int, dfd, const char __user *, pathname, int, flag)
4076 {
4077         if ((flag & ~AT_REMOVEDIR) != 0)
4078                 return -EINVAL;
4079
4080         if (flag & AT_REMOVEDIR)
4081                 return do_rmdir(dfd, pathname);
4082
4083         return do_unlinkat(dfd, getname(pathname));
4084 }
4085
4086 SYSCALL_DEFINE1(unlink, const char __user *, pathname)
4087 {
4088         return do_unlinkat(AT_FDCWD, getname(pathname));
4089 }
4090
4091 int vfs_symlink(struct inode *dir, struct dentry *dentry, const char *oldname)
4092 {
4093         int error = may_create(dir, dentry);
4094
4095         if (error)
4096                 return error;
4097
4098         if (!dir->i_op->symlink)
4099                 return -EPERM;
4100
4101         error = security_inode_symlink(dir, dentry, oldname);
4102         if (error)
4103                 return error;
4104
4105         error = dir->i_op->symlink(dir, dentry, oldname);
4106         if (!error)
4107                 fsnotify_create(dir, dentry);
4108         return error;
4109 }
4110 EXPORT_SYMBOL(vfs_symlink);
4111
4112 SYSCALL_DEFINE3(symlinkat, const char __user *, oldname,
4113                 int, newdfd, const char __user *, newname)
4114 {
4115         int error;
4116         struct filename *from;
4117         struct dentry *dentry;
4118         struct path path;
4119         unsigned int lookup_flags = 0;
4120
4121         from = getname(oldname);
4122         if (IS_ERR(from))
4123                 return PTR_ERR(from);
4124 retry:
4125         dentry = user_path_create(newdfd, newname, &path, lookup_flags);
4126         error = PTR_ERR(dentry);
4127         if (IS_ERR(dentry))
4128                 goto out_putname;
4129
4130         error = security_path_symlink(&path, dentry, from->name);
4131         if (!error)
4132                 error = vfs_symlink(path.dentry->d_inode, dentry, from->name);
4133         done_path_create(&path, dentry);
4134         if (retry_estale(error, lookup_flags)) {
4135                 lookup_flags |= LOOKUP_REVAL;
4136                 goto retry;
4137         }
4138 out_putname:
4139         putname(from);
4140         return error;
4141 }
4142
4143 SYSCALL_DEFINE2(symlink, const char __user *, oldname, const char __user *, newname)
4144 {
4145         return sys_symlinkat(oldname, AT_FDCWD, newname);
4146 }
4147
4148 /**
4149  * vfs_link - create a new link
4150  * @old_dentry: object to be linked
4151  * @dir:        new parent
4152  * @new_dentry: where to create the new link
4153  * @delegated_inode: returns inode needing a delegation break
4154  *
4155  * The caller must hold dir->i_mutex
4156  *
4157  * If vfs_link discovers a delegation on the to-be-linked file in need
4158  * of breaking, it will return -EWOULDBLOCK and return a reference to the
4159  * inode in delegated_inode.  The caller should then break the delegation
4160  * and retry.  Because breaking a delegation may take a long time, the
4161  * caller should drop the i_mutex before doing so.
4162  *
4163  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4164  * be appropriate for callers that expect the underlying filesystem not
4165  * to be NFS exported.
4166  */
4167 int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_dentry, struct inode **delegated_inode)
4168 {
4169         struct inode *inode = old_dentry->d_inode;
4170         unsigned max_links = dir->i_sb->s_max_links;
4171         int error;
4172
4173         if (!inode)
4174                 return -ENOENT;
4175
4176         error = may_create(dir, new_dentry);
4177         if (error)
4178                 return error;
4179
4180         if (dir->i_sb != inode->i_sb)
4181                 return -EXDEV;
4182
4183         /*
4184          * A link to an append-only or immutable file cannot be created.
4185          */
4186         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4187                 return -EPERM;
4188         /*
4189          * Updating the link count will likely cause i_uid and i_gid to
4190          * be writen back improperly if their true value is unknown to
4191          * the vfs.
4192          */
4193         if (HAS_UNMAPPED_ID(inode))
4194                 return -EPERM;
4195         if (!dir->i_op->link)
4196                 return -EPERM;
4197         if (S_ISDIR(inode->i_mode))
4198                 return -EPERM;
4199
4200         error = security_inode_link(old_dentry, dir, new_dentry);
4201         if (error)
4202                 return error;
4203
4204         inode_lock(inode);
4205         /* Make sure we don't allow creating hardlink to an unlinked file */
4206         if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))
4207                 error =  -ENOENT;
4208         else if (max_links && inode->i_nlink >= max_links)
4209                 error = -EMLINK;
4210         else {
4211                 error = try_break_deleg(inode, delegated_inode);
4212                 if (!error)
4213                         error = dir->i_op->link(old_dentry, dir, new_dentry);
4214         }
4215
4216         if (!error && (inode->i_state & I_LINKABLE)) {
4217                 spin_lock(&inode->i_lock);
4218                 inode->i_state &= ~I_LINKABLE;
4219                 spin_unlock(&inode->i_lock);
4220         }
4221         inode_unlock(inode);
4222         if (!error)
4223                 fsnotify_link(dir, inode, new_dentry);
4224         return error;
4225 }
4226 EXPORT_SYMBOL(vfs_link);
4227
4228 /*
4229  * Hardlinks are often used in delicate situations.  We avoid
4230  * security-related surprises by not following symlinks on the
4231  * newname.  --KAB
4232  *
4233  * We don't follow them on the oldname either to be compatible
4234  * with linux 2.0, and to avoid hard-linking to directories
4235  * and other special files.  --ADM
4236  */
4237 SYSCALL_DEFINE5(linkat, int, olddfd, const char __user *, oldname,
4238                 int, newdfd, const char __user *, newname, int, flags)
4239 {
4240         struct dentry *new_dentry;
4241         struct path old_path, new_path;
4242         struct inode *delegated_inode = NULL;
4243         int how = 0;
4244         int error;
4245
4246         if ((flags & ~(AT_SYMLINK_FOLLOW | AT_EMPTY_PATH)) != 0)
4247                 return -EINVAL;
4248         /*
4249          * To use null names we require CAP_DAC_READ_SEARCH
4250          * This ensures that not everyone will be able to create
4251          * handlink using the passed filedescriptor.
4252          */
4253         if (flags & AT_EMPTY_PATH) {
4254                 if (!capable(CAP_DAC_READ_SEARCH))
4255                         return -ENOENT;
4256                 how = LOOKUP_EMPTY;
4257         }
4258
4259         if (flags & AT_SYMLINK_FOLLOW)
4260                 how |= LOOKUP_FOLLOW;
4261 retry:
4262         error = user_path_at(olddfd, oldname, how, &old_path);
4263         if (error)
4264                 return error;
4265
4266         new_dentry = user_path_create(newdfd, newname, &new_path,
4267                                         (how & LOOKUP_REVAL));
4268         error = PTR_ERR(new_dentry);
4269         if (IS_ERR(new_dentry))
4270                 goto out;
4271
4272         error = -EXDEV;
4273         if (old_path.mnt != new_path.mnt)
4274                 goto out_dput;
4275         error = may_linkat(&old_path);
4276         if (unlikely(error))
4277                 goto out_dput;
4278         error = security_path_link(old_path.dentry, &new_path, new_dentry);
4279         if (error)
4280                 goto out_dput;
4281         error = vfs_link(old_path.dentry, new_path.dentry->d_inode, new_dentry, &delegated_inode);
4282 out_dput:
4283         done_path_create(&new_path, new_dentry);
4284         if (delegated_inode) {
4285                 error = break_deleg_wait(&delegated_inode);
4286                 if (!error) {
4287                         path_put(&old_path);
4288                         goto retry;
4289                 }
4290         }
4291         if (retry_estale(error, how)) {
4292                 path_put(&old_path);
4293                 how |= LOOKUP_REVAL;
4294                 goto retry;
4295         }
4296 out:
4297         path_put(&old_path);
4298
4299         return error;
4300 }
4301
4302 SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname)
4303 {
4304         return sys_linkat(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4305 }
4306
4307 /**
4308  * vfs_rename - rename a filesystem object
4309  * @old_dir:    parent of source
4310  * @old_dentry: source
4311  * @new_dir:    parent of destination
4312  * @new_dentry: destination
4313  * @delegated_inode: returns an inode needing a delegation break
4314  * @flags:      rename flags
4315  *
4316  * The caller must hold multiple mutexes--see lock_rename()).
4317  *
4318  * If vfs_rename discovers a delegation in need of breaking at either
4319  * the source or destination, it will return -EWOULDBLOCK and return a
4320  * reference to the inode in delegated_inode.  The caller should then
4321  * break the delegation and retry.  Because breaking a delegation may
4322  * take a long time, the caller should drop all locks before doing
4323  * so.
4324  *
4325  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4326  * be appropriate for callers that expect the underlying filesystem not
4327  * to be NFS exported.
4328  *
4329  * The worst of all namespace operations - renaming directory. "Perverted"
4330  * doesn't even start to describe it. Somebody in UCB had a heck of a trip...
4331  * Problems:
4332  *
4333  *      a) we can get into loop creation.
4334  *      b) race potential - two innocent renames can create a loop together.
4335  *         That's where 4.4 screws up. Current fix: serialization on
4336  *         sb->s_vfs_rename_mutex. We might be more accurate, but that's another
4337  *         story.
4338  *      c) we have to lock _four_ objects - parents and victim (if it exists),
4339  *         and source (if it is not a directory).
4340  *         And that - after we got ->i_mutex on parents (until then we don't know
4341  *         whether the target exists).  Solution: try to be smart with locking
4342  *         order for inodes.  We rely on the fact that tree topology may change
4343  *         only under ->s_vfs_rename_mutex _and_ that parent of the object we
4344  *         move will be locked.  Thus we can rank directories by the tree
4345  *         (ancestors first) and rank all non-directories after them.
4346  *         That works since everybody except rename does "lock parent, lookup,
4347  *         lock child" and rename is under ->s_vfs_rename_mutex.
4348  *         HOWEVER, it relies on the assumption that any object with ->lookup()
4349  *         has no more than 1 dentry.  If "hybrid" objects will ever appear,
4350  *         we'd better make sure that there's no link(2) for them.
4351  *      d) conversion from fhandle to dentry may come in the wrong moment - when
4352  *         we are removing the target. Solution: we will have to grab ->i_mutex
4353  *         in the fhandle_to_dentry code. [FIXME - current nfsfh.c relies on
4354  *         ->i_mutex on parents, which works but leads to some truly excessive
4355  *         locking].
4356  */
4357 int vfs_rename(struct inode *old_dir, struct dentry *old_dentry,
4358                struct inode *new_dir, struct dentry *new_dentry,
4359                struct inode **delegated_inode, unsigned int flags)
4360 {
4361         int error;
4362         bool is_dir = d_is_dir(old_dentry);
4363         struct inode *source = old_dentry->d_inode;
4364         struct inode *target = new_dentry->d_inode;
4365         bool new_is_dir = false;
4366         unsigned max_links = new_dir->i_sb->s_max_links;
4367         struct name_snapshot old_name;
4368
4369         if (source == target)
4370                 return 0;
4371
4372         error = may_delete(old_dir, old_dentry, is_dir);
4373         if (error)
4374                 return error;
4375
4376         if (!target) {
4377                 error = may_create(new_dir, new_dentry);
4378         } else {
4379                 new_is_dir = d_is_dir(new_dentry);
4380
4381                 if (!(flags & RENAME_EXCHANGE))
4382                         error = may_delete(new_dir, new_dentry, is_dir);
4383                 else
4384                         error = may_delete(new_dir, new_dentry, new_is_dir);
4385         }
4386         if (error)
4387                 return error;
4388
4389         if (!old_dir->i_op->rename)
4390                 return -EPERM;
4391
4392         /*
4393          * If we are going to change the parent - check write permissions,
4394          * we'll need to flip '..'.
4395          */
4396         if (new_dir != old_dir) {
4397                 if (is_dir) {
4398                         error = inode_permission(source, MAY_WRITE);
4399                         if (error)
4400                                 return error;
4401                 }
4402                 if ((flags & RENAME_EXCHANGE) && new_is_dir) {
4403                         error = inode_permission(target, MAY_WRITE);
4404                         if (error)
4405                                 return error;
4406                 }
4407         }
4408
4409         error = security_inode_rename(old_dir, old_dentry, new_dir, new_dentry,
4410                                       flags);
4411         if (error)
4412                 return error;
4413
4414         take_dentry_name_snapshot(&old_name, old_dentry);
4415         dget(new_dentry);
4416         if (!is_dir || (flags & RENAME_EXCHANGE))
4417                 lock_two_nondirectories(source, target);
4418         else if (target)
4419                 inode_lock(target);
4420
4421         error = -EBUSY;
4422         if (is_local_mountpoint(old_dentry) || is_local_mountpoint(new_dentry))
4423                 goto out;
4424
4425         if (max_links && new_dir != old_dir) {
4426                 error = -EMLINK;
4427                 if (is_dir && !new_is_dir && new_dir->i_nlink >= max_links)
4428                         goto out;
4429                 if ((flags & RENAME_EXCHANGE) && !is_dir && new_is_dir &&
4430                     old_dir->i_nlink >= max_links)
4431                         goto out;
4432         }
4433         if (is_dir && !(flags & RENAME_EXCHANGE) && target)
4434                 shrink_dcache_parent(new_dentry);
4435         if (!is_dir) {
4436                 error = try_break_deleg(source, delegated_inode);
4437                 if (error)
4438                         goto out;
4439         }
4440         if (target && !new_is_dir) {
4441                 error = try_break_deleg(target, delegated_inode);
4442                 if (error)
4443                         goto out;
4444         }
4445         error = old_dir->i_op->rename(old_dir, old_dentry,
4446                                        new_dir, new_dentry, flags);
4447         if (error)
4448                 goto out;
4449
4450         if (!(flags & RENAME_EXCHANGE) && target) {
4451                 if (is_dir)
4452                         target->i_flags |= S_DEAD;
4453                 dont_mount(new_dentry);
4454                 detach_mounts(new_dentry);
4455         }
4456         if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE)) {
4457                 if (!(flags & RENAME_EXCHANGE))
4458                         d_move(old_dentry, new_dentry);
4459                 else
4460                         d_exchange(old_dentry, new_dentry);
4461         }
4462 out:
4463         if (!is_dir || (flags & RENAME_EXCHANGE))
4464                 unlock_two_nondirectories(source, target);
4465         else if (target)
4466                 inode_unlock(target);
4467         dput(new_dentry);
4468         if (!error) {
4469                 fsnotify_move(old_dir, new_dir, old_name.name, is_dir,
4470                               !(flags & RENAME_EXCHANGE) ? target : NULL, old_dentry);
4471                 if (flags & RENAME_EXCHANGE) {
4472                         fsnotify_move(new_dir, old_dir, old_dentry->d_name.name,
4473                                       new_is_dir, NULL, new_dentry);
4474                 }
4475         }
4476         release_dentry_name_snapshot(&old_name);
4477
4478         return error;
4479 }
4480 EXPORT_SYMBOL(vfs_rename);
4481
4482 SYSCALL_DEFINE5(renameat2, int, olddfd, const char __user *, oldname,
4483                 int, newdfd, const char __user *, newname, unsigned int, flags)
4484 {
4485         struct dentry *old_dentry, *new_dentry;
4486         struct dentry *trap;
4487         struct path old_path, new_path;
4488         struct qstr old_last, new_last;
4489         int old_type, new_type;
4490         struct inode *delegated_inode = NULL;
4491         struct filename *from;
4492         struct filename *to;
4493         unsigned int lookup_flags = 0, target_flags = LOOKUP_RENAME_TARGET;
4494         bool should_retry = false;
4495         int error;
4496
4497         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
4498                 return -EINVAL;
4499
4500         if ((flags & (RENAME_NOREPLACE | RENAME_WHITEOUT)) &&
4501             (flags & RENAME_EXCHANGE))
4502                 return -EINVAL;
4503
4504         if ((flags & RENAME_WHITEOUT) && !capable(CAP_MKNOD))
4505                 return -EPERM;
4506
4507         if (flags & RENAME_EXCHANGE)
4508                 target_flags = 0;
4509
4510 retry:
4511         from = filename_parentat(olddfd, getname(oldname), lookup_flags,
4512                                 &old_path, &old_last, &old_type);
4513         if (IS_ERR(from)) {
4514                 error = PTR_ERR(from);
4515                 goto exit;
4516         }
4517
4518         to = filename_parentat(newdfd, getname(newname), lookup_flags,
4519                                 &new_path, &new_last, &new_type);
4520         if (IS_ERR(to)) {
4521                 error = PTR_ERR(to);
4522                 goto exit1;
4523         }
4524
4525         error = -EXDEV;
4526         if (old_path.mnt != new_path.mnt)
4527                 goto exit2;
4528
4529         error = -EBUSY;
4530         if (old_type != LAST_NORM)
4531                 goto exit2;
4532
4533         if (flags & RENAME_NOREPLACE)
4534                 error = -EEXIST;
4535         if (new_type != LAST_NORM)
4536                 goto exit2;
4537
4538         error = mnt_want_write(old_path.mnt);
4539         if (error)
4540                 goto exit2;
4541
4542 retry_deleg:
4543         trap = lock_rename(new_path.dentry, old_path.dentry);
4544
4545         old_dentry = __lookup_hash(&old_last, old_path.dentry, lookup_flags);
4546         error = PTR_ERR(old_dentry);
4547         if (IS_ERR(old_dentry))
4548                 goto exit3;
4549         /* source must exist */
4550         error = -ENOENT;
4551         if (d_is_negative(old_dentry))
4552                 goto exit4;
4553         new_dentry = __lookup_hash(&new_last, new_path.dentry, lookup_flags | target_flags);
4554         error = PTR_ERR(new_dentry);
4555         if (IS_ERR(new_dentry))
4556                 goto exit4;
4557         error = -EEXIST;
4558         if ((flags & RENAME_NOREPLACE) && d_is_positive(new_dentry))
4559                 goto exit5;
4560         if (flags & RENAME_EXCHANGE) {
4561                 error = -ENOENT;
4562                 if (d_is_negative(new_dentry))
4563                         goto exit5;
4564
4565                 if (!d_is_dir(new_dentry)) {
4566                         error = -ENOTDIR;
4567                         if (new_last.name[new_last.len])
4568                                 goto exit5;
4569                 }
4570         }
4571         /* unless the source is a directory trailing slashes give -ENOTDIR */
4572         if (!d_is_dir(old_dentry)) {
4573                 error = -ENOTDIR;
4574                 if (old_last.name[old_last.len])
4575                         goto exit5;
4576                 if (!(flags & RENAME_EXCHANGE) && new_last.name[new_last.len])
4577                         goto exit5;
4578         }
4579         /* source should not be ancestor of target */
4580         error = -EINVAL;
4581         if (old_dentry == trap)
4582                 goto exit5;
4583         /* target should not be an ancestor of source */
4584         if (!(flags & RENAME_EXCHANGE))
4585                 error = -ENOTEMPTY;
4586         if (new_dentry == trap)
4587                 goto exit5;
4588
4589         error = security_path_rename(&old_path, old_dentry,
4590                                      &new_path, new_dentry, flags);
4591         if (error)
4592                 goto exit5;
4593         error = vfs_rename(old_path.dentry->d_inode, old_dentry,
4594                            new_path.dentry->d_inode, new_dentry,
4595                            &delegated_inode, flags);
4596 exit5:
4597         dput(new_dentry);
4598 exit4:
4599         dput(old_dentry);
4600 exit3:
4601         unlock_rename(new_path.dentry, old_path.dentry);
4602         if (delegated_inode) {
4603                 error = break_deleg_wait(&delegated_inode);
4604                 if (!error)
4605                         goto retry_deleg;
4606         }
4607         mnt_drop_write(old_path.mnt);
4608 exit2:
4609         if (retry_estale(error, lookup_flags))
4610                 should_retry = true;
4611         path_put(&new_path);
4612         putname(to);
4613 exit1:
4614         path_put(&old_path);
4615         putname(from);
4616         if (should_retry) {
4617                 should_retry = false;
4618                 lookup_flags |= LOOKUP_REVAL;
4619                 goto retry;
4620         }
4621 exit:
4622         return error;
4623 }
4624
4625 SYSCALL_DEFINE4(renameat, int, olddfd, const char __user *, oldname,
4626                 int, newdfd, const char __user *, newname)
4627 {
4628         return sys_renameat2(olddfd, oldname, newdfd, newname, 0);
4629 }
4630
4631 SYSCALL_DEFINE2(rename, const char __user *, oldname, const char __user *, newname)
4632 {
4633         return sys_renameat2(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4634 }
4635
4636 int vfs_whiteout(struct inode *dir, struct dentry *dentry)
4637 {
4638         int error = may_create(dir, dentry);
4639         if (error)
4640                 return error;
4641
4642         if (!dir->i_op->mknod)
4643                 return -EPERM;
4644
4645         return dir->i_op->mknod(dir, dentry,
4646                                 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
4647 }
4648 EXPORT_SYMBOL(vfs_whiteout);
4649
4650 int readlink_copy(char __user *buffer, int buflen, const char *link)
4651 {
4652         int len = PTR_ERR(link);
4653         if (IS_ERR(link))
4654                 goto out;
4655
4656         len = strlen(link);
4657         if (len > (unsigned) buflen)
4658                 len = buflen;
4659         if (copy_to_user(buffer, link, len))
4660                 len = -EFAULT;
4661 out:
4662         return len;
4663 }
4664
4665 /*
4666  * A helper for ->readlink().  This should be used *ONLY* for symlinks that
4667  * have ->get_link() not calling nd_jump_link().  Using (or not using) it
4668  * for any given inode is up to filesystem.
4669  */
4670 static int generic_readlink(struct dentry *dentry, char __user *buffer,
4671                             int buflen)
4672 {
4673         DEFINE_DELAYED_CALL(done);
4674         struct inode *inode = d_inode(dentry);
4675         const char *link = inode->i_link;
4676         int res;
4677
4678         if (!link) {
4679                 link = inode->i_op->get_link(dentry, inode, &done);
4680                 if (IS_ERR(link))
4681                         return PTR_ERR(link);
4682         }
4683         res = readlink_copy(buffer, buflen, link);
4684         do_delayed_call(&done);
4685         return res;
4686 }
4687
4688 /**
4689  * vfs_readlink - copy symlink body into userspace buffer
4690  * @dentry: dentry on which to get symbolic link
4691  * @buffer: user memory pointer
4692  * @buflen: size of buffer
4693  *
4694  * Does not touch atime.  That's up to the caller if necessary
4695  *
4696  * Does not call security hook.
4697  */
4698 int vfs_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4699 {
4700         struct inode *inode = d_inode(dentry);
4701
4702         if (unlikely(!(inode->i_opflags & IOP_DEFAULT_READLINK))) {
4703                 if (unlikely(inode->i_op->readlink))
4704                         return inode->i_op->readlink(dentry, buffer, buflen);
4705
4706                 if (!d_is_symlink(dentry))
4707                         return -EINVAL;
4708
4709                 spin_lock(&inode->i_lock);
4710                 inode->i_opflags |= IOP_DEFAULT_READLINK;
4711                 spin_unlock(&inode->i_lock);
4712         }
4713
4714         return generic_readlink(dentry, buffer, buflen);
4715 }
4716 EXPORT_SYMBOL(vfs_readlink);
4717
4718 /**
4719  * vfs_get_link - get symlink body
4720  * @dentry: dentry on which to get symbolic link
4721  * @done: caller needs to free returned data with this
4722  *
4723  * Calls security hook and i_op->get_link() on the supplied inode.
4724  *
4725  * It does not touch atime.  That's up to the caller if necessary.
4726  *
4727  * Does not work on "special" symlinks like /proc/$$/fd/N
4728  */
4729 const char *vfs_get_link(struct dentry *dentry, struct delayed_call *done)
4730 {
4731         const char *res = ERR_PTR(-EINVAL);
4732         struct inode *inode = d_inode(dentry);
4733
4734         if (d_is_symlink(dentry)) {
4735                 res = ERR_PTR(security_inode_readlink(dentry));
4736                 if (!res)
4737                         res = inode->i_op->get_link(dentry, inode, done);
4738         }
4739         return res;
4740 }
4741 EXPORT_SYMBOL(vfs_get_link);
4742
4743 /* get the link contents into pagecache */
4744 const char *page_get_link(struct dentry *dentry, struct inode *inode,
4745                           struct delayed_call *callback)
4746 {
4747         char *kaddr;
4748         struct page *page;
4749         struct address_space *mapping = inode->i_mapping;
4750
4751         if (!dentry) {
4752                 page = find_get_page(mapping, 0);
4753                 if (!page)
4754                         return ERR_PTR(-ECHILD);
4755                 if (!PageUptodate(page)) {
4756                         put_page(page);
4757                         return ERR_PTR(-ECHILD);
4758                 }
4759         } else {
4760                 page = read_mapping_page(mapping, 0, NULL);
4761                 if (IS_ERR(page))
4762                         return (char*)page;
4763         }
4764         set_delayed_call(callback, page_put_link, page);
4765         BUG_ON(mapping_gfp_mask(mapping) & __GFP_HIGHMEM);
4766         kaddr = page_address(page);
4767         nd_terminate_link(kaddr, inode->i_size, PAGE_SIZE - 1);
4768         return kaddr;
4769 }
4770
4771 EXPORT_SYMBOL(page_get_link);
4772
4773 void page_put_link(void *arg)
4774 {
4775         put_page(arg);
4776 }
4777 EXPORT_SYMBOL(page_put_link);
4778
4779 int page_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4780 {
4781         DEFINE_DELAYED_CALL(done);
4782         int res = readlink_copy(buffer, buflen,
4783                                 page_get_link(dentry, d_inode(dentry),
4784                                               &done));
4785         do_delayed_call(&done);
4786         return res;
4787 }
4788 EXPORT_SYMBOL(page_readlink);
4789
4790 /*
4791  * The nofs argument instructs pagecache_write_begin to pass AOP_FLAG_NOFS
4792  */
4793 int __page_symlink(struct inode *inode, const char *symname, int len, int nofs)
4794 {
4795         struct address_space *mapping = inode->i_mapping;
4796         struct page *page;
4797         void *fsdata;
4798         int err;
4799         unsigned int flags = 0;
4800         if (nofs)
4801                 flags |= AOP_FLAG_NOFS;
4802
4803 retry:
4804         err = pagecache_write_begin(NULL, mapping, 0, len-1,
4805                                 flags, &page, &fsdata);
4806         if (err)
4807                 goto fail;
4808
4809         memcpy(page_address(page), symname, len-1);
4810
4811         err = pagecache_write_end(NULL, mapping, 0, len-1, len-1,
4812                                                         page, fsdata);
4813         if (err < 0)
4814                 goto fail;
4815         if (err < len-1)
4816                 goto retry;
4817
4818         mark_inode_dirty(inode);
4819         return 0;
4820 fail:
4821         return err;
4822 }
4823 EXPORT_SYMBOL(__page_symlink);
4824
4825 int page_symlink(struct inode *inode, const char *symname, int len)
4826 {
4827         return __page_symlink(inode, symname, len,
4828                         !mapping_gfp_constraint(inode->i_mapping, __GFP_FS));
4829 }
4830 EXPORT_SYMBOL(page_symlink);
4831
4832 const struct inode_operations page_symlink_inode_operations = {
4833         .get_link       = page_get_link,
4834 };
4835 EXPORT_SYMBOL(page_symlink_inode_operations);