fs/namei.c

   1 /*
   2  *  linux/fs/namei.c
   3  *
   4  *  Copyright (C) 1991, 1992  Linus Torvalds
   5  */
   6
   7 /*
   8  * Some corrections by tytso.
   9  */
  10
  11 /* [Feb 1997 T. Schoebel-Theuer] Complete rewrite of the pathname
  12  * lookup logic.
  13  */
  14 /* [Feb-Apr 2000, AV] Rewrite to the new namespace architecture.
  15  */
  16
  17 #include <linux/init.h>
  18 #include <linux/export.h>
  19 #include <linux/kernel.h>
  20 #include <linux/slab.h>
  21 #include <linux/fs.h>
  22 #include <linux/namei.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/fsnotify.h>
  25 #include <linux/personality.h>
  26 #include <linux/security.h>
  27 #include <linux/ima.h>
  28 #include <linux/syscalls.h>
  29 #include <linux/mount.h>
  30 #include <linux/audit.h>
  31 #include <linux/capability.h>
  32 #include <linux/file.h>
  33 #include <linux/fcntl.h>
  34 #include <linux/device_cgroup.h>
  35 #include <linux/fs_struct.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/hash.h>
  38 #include <linux/bitops.h>
  39 #include <linux/init_task.h>
  40 #include <linux/uaccess.h>
  41
  42 #include "internal.h"
  43 #include "mount.h"
  44
  45 /* [Feb-1997 T. Schoebel-Theuer]
  46  * Fundamental changes in the pathname lookup mechanisms (namei)
  47  * were necessary because of omirr.  The reason is that omirr needs
  48  * to know the _real_ pathname, not the user-supplied one, in case
  49  * of symlinks (and also when transname replacements occur).
  50  *
  51  * The new code replaces the old recursive symlink resolution with
  52  * an iterative one (in case of non-nested symlink chains).  It does
  53  * this with calls to <fs>_follow_link().
  54  * As a side effect, dir_namei(), _namei() and follow_link() are now
  55  * replaced with a single function lookup_dentry() that can handle all
  56  * the special cases of the former code.
  57  *
  58  * With the new dcache, the pathname is stored at each inode, at least as
  59  * long as the refcount of the inode is positive.  As a side effect, the
  60  * size of the dcache depends on the inode cache and thus is dynamic.
  61  *
  62  * [29-Apr-1998 C. Scott Ananian] Updated above description of symlink
  63  * resolution to correspond with current state of the code.
  64  *
  65  * Note that the symlink resolution is not *completely* iterative.
  66  * There is still a significant amount of tail- and mid- recursion in
  67  * the algorithm.  Also, note that <fs>_readlink() is not used in
  68  * lookup_dentry(): lookup_dentry() on the result of <fs>_readlink()
  69  * may return different results than <fs>_follow_link().  Many virtual
  70  * filesystems (including /proc) exhibit this behavior.
  71  */
  72
  73 /* [24-Feb-97 T. Schoebel-Theuer] Side effects caused by new implementation:
  74  * New symlink semantics: when open() is called with flags O_CREAT | O_EXCL
  75  * and the name already exists in form of a symlink, try to create the new
  76  * name indicated by the symlink. The old code always complained that the
  77  * name already exists, due to not following the symlink even if its target
  78  * is nonexistent.  The new semantics affects also mknod() and link() when
  79  * the name is a symlink pointing to a non-existent name.
  80  *
  81  * I don't know which semantics is the right one, since I have no access
  82  * to standards. But I found by trial that HP-UX 9.0 has the full "new"
  83  * semantics implemented, while SunOS 4.1.1 and Solaris (SunOS 5.4) have the
  84  * "old" one. Personally, I think the new semantics is much more logical.
  85  * Note that "ln old new" where "new" is a symlink pointing to a non-existing
  86  * file does succeed in both HP-UX and SunOs, but not in Solaris
  87  * and in the old Linux semantics.
  88  */
  89
  90 /* [16-Dec-97 Kevin Buhr] For security reasons, we change some symlink
  91  * semantics.  See the comments in "open_namei" and "do_link" below.
  92  *
  93  * [10-Sep-98 Alan Modra] Another symlink change.
  94  */
  95
  96 /* [Feb-Apr 2000 AV] Complete rewrite. Rules for symlinks:
  97  *      inside the path - always follow.
  98  *      in the last component in creation/removal/renaming - never follow.
  99  *      if LOOKUP_FOLLOW passed - follow.
 100  *      if the pathname has trailing slashes - follow.
 101  *      otherwise - don't follow.
 102  * (applied in that order).
 103  *
 104  * [Jun 2000 AV] Inconsistent behaviour of open() in case if flags==O_CREAT
 105  * restored for 2.4. This is the last surviving part of old 4.2BSD bug.
 106  * During the 2.4 we need to fix the userland stuff depending on it -
 107  * hopefully we will be able to get rid of that wart in 2.5. So far only
 108  * XEmacs seems to be relying on it...
 109  */
 110 /*
 111  * [Sep 2001 AV] Single-semaphore locking scheme (kudos to David Holland)
 112  * implemented.  Let's see if raised priority of ->s_vfs_rename_mutex gives
 113  * any extra contention...
 114  */
 115
 116 /* In order to reduce some races, while at the same time doing additional
 117  * checking and hopefully speeding things up, we copy filenames to the
 118  * kernel data space before using them..
 119  *
 120  * POSIX.1 2.4: an empty pathname is invalid (ENOENT).
 121  * PATH_MAX includes the nul terminator --RR.
 122  */
 123
 124 #define EMBEDDED_NAME_MAX       (PATH_MAX - offsetof(struct filename, iname))
 125
 126 struct filename *
 127 getname_flags(const char __user *filename, int flags, int *empty)
 128 {
 129         struct filename *result;
 130         char *kname;
 131         int len;
 132
 133         result = audit_reusename(filename);
 134         if (result)
 135                 return result;
 136
 137         result = __getname();
 138         if (unlikely(!result))
 139                 return ERR_PTR(-ENOMEM);
 140
 141         /*
 142          * First, try to embed the struct filename inside the names_cache
 143          * allocation
 144          */
 145         kname = (char *)result->iname;
 146         result->name = kname;
 147
 148         len = strncpy_from_user(kname, filename, EMBEDDED_NAME_MAX);
 149         if (unlikely(len < 0)) {
 150                 __putname(result);
 151                 return ERR_PTR(len);
 152         }
 153
 154         /*
 155          * Uh-oh. We have a name that's approaching PATH_MAX. Allocate a
 156          * separate struct filename so we can dedicate the entire
 157          * names_cache allocation for the pathname, and re-do the copy from
 158          * userland.
 159          */
 160         if (unlikely(len == EMBEDDED_NAME_MAX)) {
 161                 const size_t size = offsetof(struct filename, iname[1]);
 162                 kname = (char *)result;
 163
 164                 /*
 165                  * size is chosen that way we to guarantee that
 166                  * result->iname[0] is within the same object and that
 167                  * kname can't be equal to result->iname, no matter what.
 168                  */
 169                 result = kzalloc(size, GFP_KERNEL);
 170                 if (unlikely(!result)) {
 171                         __putname(kname);
 172                         return ERR_PTR(-ENOMEM);
 173                 }
 174                 result->name = kname;
 175                 len = strncpy_from_user(kname, filename, PATH_MAX);
 176                 if (unlikely(len < 0)) {
 177                         __putname(kname);
 178                         kfree(result);
 179                         return ERR_PTR(len);
 180                 }
 181                 if (unlikely(len == PATH_MAX)) {
 182                         __putname(kname);
 183                         kfree(result);
 184                         return ERR_PTR(-ENAMETOOLONG);
 185                 }
 186         }
 187
 188         result->refcnt = 1;
 189         /* The empty path is special. */
 190         if (unlikely(!len)) {
 191                 if (empty)
 192                         *empty = 1;
 193                 if (!(flags & LOOKUP_EMPTY)) {
 194                         putname(result);
 195                         return ERR_PTR(-ENOENT);
 196                 }
 197         }
 198
 199         result->uptr = filename;
 200         result->aname = NULL;
 201         audit_getname(result);
 202         return result;
 203 }
 204
 205 struct filename *
 206 getname(const char __user * filename)
 207 {
 208         return getname_flags(filename, 0, NULL);
 209 }
 210
 211 struct filename *
 212 getname_kernel(const char * filename)
 213 {
 214         struct filename *result;
 215         int len = strlen(filename) + 1;
 216
 217         result = __getname();
 218         if (unlikely(!result))
 219                 return ERR_PTR(-ENOMEM);
 220
 221         if (len <= EMBEDDED_NAME_MAX) {
 222                 result->name = (char *)result->iname;
 223         } else if (len <= PATH_MAX) {
 224                 struct filename *tmp;
 225
 226                 tmp = kmalloc(sizeof(*tmp), GFP_KERNEL);
 227                 if (unlikely(!tmp)) {
 228                         __putname(result);
 229                         return ERR_PTR(-ENOMEM);
 230                 }
 231                 tmp->name = (char *)result;
 232                 result = tmp;
 233         } else {
 234                 __putname(result);
 235                 return ERR_PTR(-ENAMETOOLONG);
 236         }
 237         memcpy((char *)result->name, filename, len);
 238         result->uptr = NULL;
 239         result->aname = NULL;
 240         result->refcnt = 1;
 241         audit_getname(result);
 242
 243         return result;
 244 }
 245
 246 void putname(struct filename *name)
 247 {
 248         BUG_ON(name->refcnt <= 0);
 249
 250         if (--name->refcnt > 0)
 251                 return;
 252
 253         if (name->name != name->iname) {
 254                 __putname(name->name);
 255                 kfree(name);
 256         } else
 257                 __putname(name);
 258 }
 259
 260 static int check_acl(struct inode *inode, int mask)
 261 {
 262 #ifdef CONFIG_FS_POSIX_ACL
 263         struct posix_acl *acl;
 264
 265         if (mask & MAY_NOT_BLOCK) {
 266                 acl = get_cached_acl_rcu(inode, ACL_TYPE_ACCESS);
 267                 if (!acl)
 268                         return -EAGAIN;
 269                 /* no ->get_acl() calls in RCU mode... */
 270                 if (is_uncached_acl(acl))
 271                         return -ECHILD;
 272                 return posix_acl_permission(inode, acl, mask & ~MAY_NOT_BLOCK);
 273         }
 274
 275         acl = get_acl(inode, ACL_TYPE_ACCESS);
 276         if (IS_ERR(acl))
 277                 return PTR_ERR(acl);
 278         if (acl) {
 279                 int error = posix_acl_permission(inode, acl, mask);
 280                 posix_acl_release(acl);
 281                 return error;
 282         }
 283 #endif
 284
 285         return -EAGAIN;
 286 }
 287
 288 /*
 289  * This does the basic permission checking
 290  */
 291 static int acl_permission_check(struct inode *inode, int mask)
 292 {
 293         unsigned int mode = inode->i_mode;
 294
 295         if (likely(uid_eq(current_fsuid(), inode->i_uid)))
 296                 mode >>= 6;
 297         else {
 298                 if (IS_POSIXACL(inode) && (mode & S_IRWXG)) {
 299                         int error = check_acl(inode, mask);
 300                         if (error != -EAGAIN)
 301                                 return error;
 302                 }
 303
 304                 if (in_group_p(inode->i_gid))
 305                         mode >>= 3;
 306         }
 307
 308         /*
 309          * If the DACs are ok we don't need any capability check.
 310          */
 311         if ((mask & ~mode & (MAY_READ | MAY_WRITE | MAY_EXEC)) == 0)
 312                 return 0;
 313         return -EACCES;
 314 }
 315
 316 /**
 317  * generic_permission -  check for access rights on a Posix-like filesystem
 318  * @inode:      inode to check access rights for
 319  * @mask:       right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC, ...)
 320  *
 321  * Used to check for read/write/execute permissions on a file.
 322  * We use "fsuid" for this, letting us set arbitrary permissions
 323  * for filesystem access without changing the "normal" uids which
 324  * are used for other things.
 325  *
 326  * generic_permission is rcu-walk aware. It returns -ECHILD in case an rcu-walk
 327  * request cannot be satisfied (eg. requires blocking or too much complexity).
 328  * It would then be called again in ref-walk mode.
 329  */
 330 int generic_permission(struct inode *inode, int mask)
 331 {
 332         int ret;
 333
 334         /*
 335          * Do the basic permission checks.
 336          */
 337         ret = acl_permission_check(inode, mask);
 338         if (ret != -EACCES)
 339                 return ret;
 340
 341         if (S_ISDIR(inode->i_mode)) {
 342                 /* DACs are overridable for directories */
 343                 if (!(mask & MAY_WRITE))
 344                         if (capable_wrt_inode_uidgid(inode,
 345                                                      CAP_DAC_READ_SEARCH))
 346                                 return 0;
 347                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
 348                         return 0;
 349                 return -EACCES;
 350         }
 351
 352         /*
 353          * Searching includes executable on directories, else just read.
 354          */
 355         mask &= MAY_READ | MAY_WRITE | MAY_EXEC;
 356         if (mask == MAY_READ)
 357                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_READ_SEARCH))
 358                         return 0;
 359         /*
 360          * Read/write DACs are always overridable.
 361          * Executable DACs are overridable when there is
 362          * at least one exec bit set.
 363          */
 364         if (!(mask & MAY_EXEC) || (inode->i_mode & S_IXUGO))
 365                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
 366                         return 0;
 367
 368         return -EACCES;
 369 }
 370 EXPORT_SYMBOL(generic_permission);
 371
 372 /*
 373  * We _really_ want to just do "generic_permission()" without
 374  * even looking at the inode->i_op values. So we keep a cache
 375  * flag in inode->i_opflags, that says "this has not special
 376  * permission function, use the fast case".
 377  */
 378 static inline int do_inode_permission(struct inode *inode, int mask)
 379 {
 380         if (unlikely(!(inode->i_opflags & IOP_FASTPERM))) {
 381                 if (likely(inode->i_op->permission))
 382                         return inode->i_op->permission(inode, mask);
 383
 384                 /* This gets set once for the inode lifetime */
 385                 spin_lock(&inode->i_lock);
 386                 inode->i_opflags |= IOP_FASTPERM;
 387                 spin_unlock(&inode->i_lock);
 388         }
 389         return generic_permission(inode, mask);
 390 }
 391
 392 /**
 393  * __inode_permission - Check for access rights to a given inode
 394  * @inode: Inode to check permission on
 395  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 396  *
 397  * Check for read/write/execute permissions on an inode.
 398  *
 399  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
 400  *
 401  * This does not check for a read-only file system.  You probably want
 402  * inode_permission().
 403  */
 404 int __inode_permission(struct inode *inode, int mask)
 405 {
 406         int retval;
 407
 408         if (unlikely(mask & MAY_WRITE)) {
 409                 /*
 410                  * Nobody gets write access to an immutable file.
 411                  */
 412                 if (IS_IMMUTABLE(inode))
 413                         return -EPERM;
 414
 415                 /*
 416                  * Updating mtime will likely cause i_uid and i_gid to be
 417                  * written back improperly if their true value is unknown
 418                  * to the vfs.
 419                  */
 420                 if (HAS_UNMAPPED_ID(inode))
 421                         return -EACCES;
 422         }
 423
 424         retval = do_inode_permission(inode, mask);
 425         if (retval)
 426                 return retval;
 427
 428         retval = devcgroup_inode_permission(inode, mask);
 429         if (retval)
 430                 return retval;
 431
 432         return security_inode_permission(inode, mask);
 433 }
 434 EXPORT_SYMBOL(__inode_permission);
 435
 436 /**
 437  * sb_permission - Check superblock-level permissions
 438  * @sb: Superblock of inode to check permission on
 439  * @inode: Inode to check permission on
 440  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 441  *
 442  * Separate out file-system wide checks from inode-specific permission checks.
 443  */
 444 static int sb_permission(struct super_block *sb, struct inode *inode, int mask)
 445 {
 446         if (unlikely(mask & MAY_WRITE)) {
 447                 umode_t mode = inode->i_mode;
 448
 449                 /* Nobody gets write access to a read-only fs. */
 450                 if (sb_rdonly(sb) && (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
 451                         return -EROFS;
 452         }
 453         return 0;
 454 }
 455
 456 /**
 457  * inode_permission - Check for access rights to a given inode
 458  * @inode: Inode to check permission on
 459  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 460  *
 461  * Check for read/write/execute permissions on an inode.  We use fs[ug]id for
 462  * this, letting us set arbitrary permissions for filesystem access without
 463  * changing the "normal" UIDs which are used for other things.
 464  *
 465  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
 466  */
 467 int inode_permission(struct inode *inode, int mask)
 468 {
 469         int retval;
 470
 471         retval = sb_permission(inode->i_sb, inode, mask);
 472         if (retval)
 473                 return retval;
 474         return __inode_permission(inode, mask);
 475 }
 476 EXPORT_SYMBOL(inode_permission);
 477
 478 /**
 479  * path_get - get a reference to a path
 480  * @path: path to get the reference to
 481  *
 482  * Given a path increment the reference count to the dentry and the vfsmount.
 483  */
 484 void path_get(const struct path *path)
 485 {
 486         mntget(path->mnt);
 487         dget(path->dentry);
 488 }
 489 EXPORT_SYMBOL(path_get);
 490
 491 /**
 492  * path_put - put a reference to a path
 493  * @path: path to put the reference to
 494  *
 495  * Given a path decrement the reference count to the dentry and the vfsmount.
 496  */
 497 void path_put(const struct path *path)
 498 {
 499         dput(path->dentry);
 500         mntput(path->mnt);
 501 }
 502 EXPORT_SYMBOL(path_put);
 503
 504 #define EMBEDDED_LEVELS 2
 505 struct nameidata {
 506         struct path     path;
 507         struct qstr     last;
 508         struct path     root;
 509         struct inode    *inode; /* path.dentry.d_inode */
 510         unsigned int    flags;
 511         unsigned        seq, m_seq;
 512         int             last_type;
 513         unsigned        depth;
 514         int             total_link_count;
 515         struct saved {
 516                 struct path link;
 517                 struct delayed_call done;
 518                 const char *name;
 519                 unsigned seq;
 520         } *stack, internal[EMBEDDED_LEVELS];
 521         struct filename *name;
 522         struct nameidata *saved;
 523         struct inode    *link_inode;
 524         unsigned        root_seq;
 525         int             dfd;
 526 } __randomize_layout;
 527
 528 static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
 529 {
 530         struct nameidata *old = current->nameidata;
 531         p->stack = p->internal;
 532         p->dfd = dfd;
 533         p->name = name;
 534         p->total_link_count = old ? old->total_link_count : 0;
 535         p->saved = old;
 536         current->nameidata = p;
 537 }
 538
 539 static void restore_nameidata(void)
 540 {
 541         struct nameidata *now = current->nameidata, *old = now->saved;
 542
 543         current->nameidata = old;
 544         if (old)
 545                 old->total_link_count = now->total_link_count;
 546         if (now->stack != now->internal)
 547                 kfree(now->stack);
 548 }
 549
 550 static int __nd_alloc_stack(struct nameidata *nd)
 551 {
 552         struct saved *p;
 553
 554         if (nd->flags & LOOKUP_RCU) {
 555                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
 556                                   GFP_ATOMIC);
 557                 if (unlikely(!p))
 558                         return -ECHILD;
 559         } else {
 560                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
 561                                   GFP_KERNEL);
 562                 if (unlikely(!p))
 563                         return -ENOMEM;
 564         }
 565         memcpy(p, nd->internal, sizeof(nd->internal));
 566         nd->stack = p;
 567         return 0;
 568 }
 569
 570 /**
 571  * path_connected - Verify that a path->dentry is below path->mnt.mnt_root
 572  * @path: nameidate to verify
 573  *
 574  * Rename can sometimes move a file or directory outside of a bind
 575  * mount, path_connected allows those cases to be detected.
 576  */
 577 static bool path_connected(const struct path *path)
 578 {
 579         struct vfsmount *mnt = path->mnt;
 580
 581         /* Only bind mounts can have disconnected paths */
 582         if (mnt->mnt_root == mnt->mnt_sb->s_root)
 583                 return true;
 584
 585         return is_subdir(path->dentry, mnt->mnt_root);
 586 }
 587
 588 static inline int nd_alloc_stack(struct nameidata *nd)
 589 {
 590         if (likely(nd->depth != EMBEDDED_LEVELS))
 591                 return 0;
 592         if (likely(nd->stack != nd->internal))
 593                 return 0;
 594         return __nd_alloc_stack(nd);
 595 }
 596
 597 static void drop_links(struct nameidata *nd)
 598 {
 599         int i = nd->depth;
 600         while (i--) {
 601                 struct saved *last = nd->stack + i;
 602                 do_delayed_call(&last->done);
 603                 clear_delayed_call(&last->done);
 604         }
 605 }
 606
 607 static void terminate_walk(struct nameidata *nd)
 608 {
 609         drop_links(nd);
 610         if (!(nd->flags & LOOKUP_RCU)) {
 611                 int i;
 612                 path_put(&nd->path);
 613                 for (i = 0; i < nd->depth; i++)
 614                         path_put(&nd->stack[i].link);
 615                 if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 616                         path_put(&nd->root);
 617                         nd->root.mnt = NULL;
 618                 }
 619         } else {
 620                 nd->flags &= ~LOOKUP_RCU;
 621                 if (!(nd->flags & LOOKUP_ROOT))
 622                         nd->root.mnt = NULL;
 623                 rcu_read_unlock();
 624         }
 625         nd->depth = 0;
 626 }
 627
 628 /* path_put is needed afterwards regardless of success or failure */
 629 static bool legitimize_path(struct nameidata *nd,
 630                             struct path *path, unsigned seq)
 631 {
 632         int res = __legitimize_mnt(path->mnt, nd->m_seq);
 633         if (unlikely(res)) {
 634                 if (res > 0)
 635                         path->mnt = NULL;
 636                 path->dentry = NULL;
 637                 return false;
 638         }
 639         if (unlikely(!lockref_get_not_dead(&path->dentry->d_lockref))) {
 640                 path->dentry = NULL;
 641                 return false;
 642         }
 643         return !read_seqcount_retry(&path->dentry->d_seq, seq);
 644 }
 645
 646 static bool legitimize_links(struct nameidata *nd)
 647 {
 648         int i;
 649         for (i = 0; i < nd->depth; i++) {
 650                 struct saved *last = nd->stack + i;
 651                 if (unlikely(!legitimize_path(nd, &last->link, last->seq))) {
 652                         drop_links(nd);
 653                         nd->depth = i + 1;
 654                         return false;
 655                 }
 656         }
 657         return true;
 658 }
 659
 660 /*
 661  * Path walking has 2 modes, rcu-walk and ref-walk (see
 662  * Documentation/filesystems/path-lookup.txt).  In situations when we can't
 663  * continue in RCU mode, we attempt to drop out of rcu-walk mode and grab
 664  * normal reference counts on dentries and vfsmounts to transition to ref-walk
 665  * mode.  Refcounts are grabbed at the last known good point before rcu-walk
 666  * got stuck, so ref-walk may continue from there. If this is not successful
 667  * (eg. a seqcount has changed), then failure is returned and it's up to caller
 668  * to restart the path walk from the beginning in ref-walk mode.
 669  */
 670
 671 /**
 672  * unlazy_walk - try to switch to ref-walk mode.
 673  * @nd: nameidata pathwalk data
 674  * Returns: 0 on success, -ECHILD on failure
 675  *
 676  * unlazy_walk attempts to legitimize the current nd->path and nd->root
 677  * for ref-walk mode.
 678  * Must be called from rcu-walk context.
 679  * Nothing should touch nameidata between unlazy_walk() failure and
 680  * terminate_walk().
 681  */
 682 static int unlazy_walk(struct nameidata *nd)
 683 {
 684         struct dentry *parent = nd->path.dentry;
 685
 686         BUG_ON(!(nd->flags & LOOKUP_RCU));
 687
 688         nd->flags &= ~LOOKUP_RCU;
 689         if (unlikely(!legitimize_links(nd)))
 690                 goto out2;
 691         if (unlikely(!legitimize_path(nd, &nd->path, nd->seq)))
 692                 goto out1;
 693         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 694                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq)))
 695                         goto out;
 696         }
 697         rcu_read_unlock();
 698         BUG_ON(nd->inode != parent->d_inode);
 699         return 0;
 700
 701 out2:
 702         nd->path.mnt = NULL;
 703         nd->path.dentry = NULL;
 704 out1:
 705         if (!(nd->flags & LOOKUP_ROOT))
 706                 nd->root.mnt = NULL;
 707 out:
 708         rcu_read_unlock();
 709         return -ECHILD;
 710 }
 711
 712 /**
 713  * unlazy_child - try to switch to ref-walk mode.
 714  * @nd: nameidata pathwalk data
 715  * @dentry: child of nd->path.dentry
 716  * @seq: seq number to check dentry against
 717  * Returns: 0 on success, -ECHILD on failure
 718  *
 719  * unlazy_child attempts to legitimize the current nd->path, nd->root and dentry
 720  * for ref-walk mode.  @dentry must be a path found by a do_lookup call on
 721  * @nd.  Must be called from rcu-walk context.
 722  * Nothing should touch nameidata between unlazy_child() failure and
 723  * terminate_walk().
 724  */
 725 static int unlazy_child(struct nameidata *nd, struct dentry *dentry, unsigned seq)
 726 {
 727         BUG_ON(!(nd->flags & LOOKUP_RCU));
 728
 729         nd->flags &= ~LOOKUP_RCU;
 730         if (unlikely(!legitimize_links(nd)))
 731                 goto out2;
 732         if (unlikely(!legitimize_mnt(nd->path.mnt, nd->m_seq)))
 733                 goto out2;
 734         if (unlikely(!lockref_get_not_dead(&nd->path.dentry->d_lockref)))
 735                 goto out1;
 736
 737         /*
 738          * We need to move both the parent and the dentry from the RCU domain
 739          * to be properly refcounted. And the sequence number in the dentry
 740          * validates *both* dentry counters, since we checked the sequence
 741          * number of the parent after we got the child sequence number. So we
 742          * know the parent must still be valid if the child sequence number is
 743          */
 744         if (unlikely(!lockref_get_not_dead(&dentry->d_lockref)))
 745                 goto out;
 746         if (unlikely(read_seqcount_retry(&dentry->d_seq, seq))) {
 747                 rcu_read_unlock();
 748                 dput(dentry);
 749                 goto drop_root_mnt;
 750         }
 751         /*
 752          * Sequence counts matched. Now make sure that the root is
 753          * still valid and get it if required.
 754          */
 755         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 756                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq))) {
 757                         rcu_read_unlock();
 758                         dput(dentry);
 759                         return -ECHILD;
 760                 }
 761         }
 762
 763         rcu_read_unlock();
 764         return 0;
 765
 766 out2:
 767         nd->path.mnt = NULL;
 768 out1:
 769         nd->path.dentry = NULL;
 770 out:
 771         rcu_read_unlock();
 772 drop_root_mnt:
 773         if (!(nd->flags & LOOKUP_ROOT))
 774                 nd->root.mnt = NULL;
 775         return -ECHILD;
 776 }
 777
 778 static inline int d_revalidate(struct dentry *dentry, unsigned int flags)
 779 {
 780         if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
 781                 return dentry->d_op->d_revalidate(dentry, flags);
 782         else
 783                 return 1;
 784 }
 785
 786 /**
 787  * complete_walk - successful completion of path walk
 788  * @nd:  pointer nameidata
 789  *
 790  * If we had been in RCU mode, drop out of it and legitimize nd->path.
 791  * Revalidate the final result, unless we'd already done that during
 792  * the path walk or the filesystem doesn't ask for it.  Return 0 on
 793  * success, -error on failure.  In case of failure caller does not
 794  * need to drop nd->path.
 795  */
 796 static int complete_walk(struct nameidata *nd)
 797 {
 798         struct dentry *dentry = nd->path.dentry;
 799         int status;
 800
 801         if (nd->flags & LOOKUP_RCU) {
 802                 if (!(nd->flags & LOOKUP_ROOT))
 803                         nd->root.mnt = NULL;
 804                 if (unlikely(unlazy_walk(nd)))
 805                         return -ECHILD;
 806         }
 807
 808         if (likely(!(nd->flags & LOOKUP_JUMPED)))
 809                 return 0;
 810
 811         if (likely(!(dentry->d_flags & DCACHE_OP_WEAK_REVALIDATE)))
 812                 return 0;
 813
 814         status = dentry->d_op->d_weak_revalidate(dentry, nd->flags);
 815         if (status > 0)
 816                 return 0;
 817
 818         if (!status)
 819                 status = -ESTALE;
 820
 821         return status;
 822 }
 823
 824 static void set_root(struct nameidata *nd)
 825 {
 826         struct fs_struct *fs = current->fs;
 827
 828         if (nd->flags & LOOKUP_RCU) {
 829                 unsigned seq;
 830
 831                 do {
 832                         seq = read_seqcount_begin(&fs->seq);
 833                         nd->root = fs->root;
 834                         nd->root_seq = __read_seqcount_begin(&nd->root.dentry->d_seq);
 835                 } while (read_seqcount_retry(&fs->seq, seq));
 836         } else {
 837                 get_fs_root(fs, &nd->root);
 838         }
 839 }
 840
 841 static void path_put_conditional(struct path *path, struct nameidata *nd)
 842 {
 843         dput(path->dentry);
 844         if (path->mnt != nd->path.mnt)
 845                 mntput(path->mnt);
 846 }
 847
 848 static inline void path_to_nameidata(const struct path *path,
 849                                         struct nameidata *nd)
 850 {
 851         if (!(nd->flags & LOOKUP_RCU)) {
 852                 dput(nd->path.dentry);
 853                 if (nd->path.mnt != path->mnt)
 854                         mntput(nd->path.mnt);
 855         }
 856         nd->path.mnt = path->mnt;
 857         nd->path.dentry = path->dentry;
 858 }
 859
 860 static int nd_jump_root(struct nameidata *nd)
 861 {
 862         if (nd->flags & LOOKUP_RCU) {
 863                 struct dentry *d;
 864                 nd->path = nd->root;
 865                 d = nd->path.dentry;
 866                 nd->inode = d->d_inode;
 867                 nd->seq = nd->root_seq;
 868                 if (unlikely(read_seqcount_retry(&d->d_seq, nd->seq)))
 869                         return -ECHILD;
 870         } else {
 871                 path_put(&nd->path);
 872                 nd->path = nd->root;
 873                 path_get(&nd->path);
 874                 nd->inode = nd->path.dentry->d_inode;
 875         }
 876         nd->flags |= LOOKUP_JUMPED;
 877         return 0;
 878 }
 879
 880 /*
 881  * Helper to directly jump to a known parsed path from ->get_link,
 882  * caller must have taken a reference to path beforehand.
 883  */
 884 void nd_jump_link(struct path *path)
 885 {
 886         struct nameidata *nd = current->nameidata;
 887         path_put(&nd->path);
 888
 889         nd->path = *path;
 890         nd->inode = nd->path.dentry->d_inode;
 891         nd->flags |= LOOKUP_JUMPED;
 892 }
 893
 894 static inline void put_link(struct nameidata *nd)
 895 {
 896         struct saved *last = nd->stack + --nd->depth;
 897         do_delayed_call(&last->done);
 898         if (!(nd->flags & LOOKUP_RCU))
 899                 path_put(&last->link);
 900 }
 901
 902 int sysctl_protected_symlinks __read_mostly = 0;
 903 int sysctl_protected_hardlinks __read_mostly = 0;
 904
 905 /**
 906  * may_follow_link - Check symlink following for unsafe situations
 907  * @nd: nameidata pathwalk data
 908  *
 909  * In the case of the sysctl_protected_symlinks sysctl being enabled,
 910  * CAP_DAC_OVERRIDE needs to be specifically ignored if the symlink is
 911  * in a sticky world-writable directory. This is to protect privileged
 912  * processes from failing races against path names that may change out
 913  * from under them by way of other users creating malicious symlinks.
 914  * It will permit symlinks to be followed only when outside a sticky
 915  * world-writable directory, or when the uid of the symlink and follower
 916  * match, or when the directory owner matches the symlink's owner.
 917  *
 918  * Returns 0 if following the symlink is allowed, -ve on error.
 919  */
 920 static inline int may_follow_link(struct nameidata *nd)
 921 {
 922         const struct inode *inode;
 923         const struct inode *parent;
 924         kuid_t puid;
 925
 926         if (!sysctl_protected_symlinks)
 927                 return 0;
 928
 929         /* Allowed if owner and follower match. */
 930         inode = nd->link_inode;
 931         if (uid_eq(current_cred()->fsuid, inode->i_uid))
 932                 return 0;
 933
 934         /* Allowed if parent directory not sticky and world-writable. */
 935         parent = nd->inode;
 936         if ((parent->i_mode & (S_ISVTX|S_IWOTH)) != (S_ISVTX|S_IWOTH))
 937                 return 0;
 938
 939         /* Allowed if parent directory and link owner match. */
 940         puid = parent->i_uid;
 941         if (uid_valid(puid) && uid_eq(puid, inode->i_uid))
 942                 return 0;
 943
 944         if (nd->flags & LOOKUP_RCU)
 945                 return -ECHILD;
 946
 947         audit_log_link_denied("follow_link", &nd->stack[0].link);
 948         return -EACCES;
 949 }
 950
 951 /**
 952  * safe_hardlink_source - Check for safe hardlink conditions
 953  * @inode: the source inode to hardlink from
 954  *
 955  * Return false if at least one of the following conditions:
 956  *    - inode is not a regular file
 957  *    - inode is setuid
 958  *    - inode is setgid and group-exec
 959  *    - access failure for read and write
 960  *
 961  * Otherwise returns true.
 962  */
 963 static bool safe_hardlink_source(struct inode *inode)
 964 {
 965         umode_t mode = inode->i_mode;
 966
 967         /* Special files should not get pinned to the filesystem. */
 968         if (!S_ISREG(mode))
 969                 return false;
 970
 971         /* Setuid files should not get pinned to the filesystem. */
 972         if (mode & S_ISUID)
 973                 return false;
 974
 975         /* Executable setgid files should not get pinned to the filesystem. */
 976         if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP))
 977                 return false;
 978
 979         /* Hardlinking to unreadable or unwritable sources is dangerous. */
 980         if (inode_permission(inode, MAY_READ | MAY_WRITE))
 981                 return false;
 982
 983         return true;
 984 }
 985
 986 /**
 987  * may_linkat - Check permissions for creating a hardlink
 988  * @link: the source to hardlink from
 989  *
 990  * Block hardlink when all of:
 991  *  - sysctl_protected_hardlinks enabled
 992  *  - fsuid does not match inode
 993  *  - hardlink source is unsafe (see safe_hardlink_source() above)
 994  *  - not CAP_FOWNER in a namespace with the inode owner uid mapped
 995  *
 996  * Returns 0 if successful, -ve on error.
 997  */
 998 static int may_linkat(struct path *link)
 999 {
1000         struct inode *inode;
1001
1002         if (!sysctl_protected_hardlinks)
1003                 return 0;
1004
1005         inode = link->dentry->d_inode;
1006
1007         /* Source inode owner (or CAP_FOWNER) can hardlink all they like,
1008          * otherwise, it must be a safe source.
1009          */
1010         if (safe_hardlink_source(inode) || inode_owner_or_capable(inode))
1011                 return 0;
1012
1013         audit_log_link_denied("linkat", link);
1014         return -EPERM;
1015 }
1016
1017 static __always_inline
1018 const char *get_link(struct nameidata *nd)
1019 {
1020         struct saved *last = nd->stack + nd->depth - 1;
1021         struct dentry *dentry = last->link.dentry;
1022         struct inode *inode = nd->link_inode;
1023         int error;
1024         const char *res;
1025
1026         if (!(nd->flags & LOOKUP_RCU)) {
1027                 touch_atime(&last->link);
1028                 cond_resched();
1029         } else if (atime_needs_update_rcu(&last->link, inode)) {
1030                 if (unlikely(unlazy_walk(nd)))
1031                         return ERR_PTR(-ECHILD);
1032                 touch_atime(&last->link);
1033         }
1034
1035         error = security_inode_follow_link(dentry, inode,
1036                                            nd->flags & LOOKUP_RCU);
1037         if (unlikely(error))
1038                 return ERR_PTR(error);
1039
1040         nd->last_type = LAST_BIND;
1041         res = inode->i_link;
1042         if (!res) {
1043                 const char * (*get)(struct dentry *, struct inode *,
1044                                 struct delayed_call *);
1045                 get = inode->i_op->get_link;
1046                 if (nd->flags & LOOKUP_RCU) {
1047                         res = get(NULL, inode, &last->done);
1048                         if (res == ERR_PTR(-ECHILD)) {
1049                                 if (unlikely(unlazy_walk(nd)))
1050                                         return ERR_PTR(-ECHILD);
1051                                 res = get(dentry, inode, &last->done);
1052                         }
1053                 } else {
1054                         res = get(dentry, inode, &last->done);
1055                 }
1056                 if (IS_ERR_OR_NULL(res))
1057                         return res;
1058         }
1059         if (*res == '/') {
1060                 if (!nd->root.mnt)
1061                         set_root(nd);
1062                 if (unlikely(nd_jump_root(nd)))
1063                         return ERR_PTR(-ECHILD);
1064                 while (unlikely(*++res == '/'))
1065                         ;
1066         }
1067         if (!*res)
1068                 res = NULL;
1069         return res;
1070 }
1071
1072 /*
1073  * follow_up - Find the mountpoint of path's vfsmount
1074  *
1075  * Given a path, find the mountpoint of its source file system.
1076  * Replace @path with the path of the mountpoint in the parent mount.
1077  * Up is towards /.
1078  *
1079  * Return 1 if we went up a level and 0 if we were already at the
1080  * root.
1081  */
1082 int follow_up(struct path *path)
1083 {
1084         struct mount *mnt = real_mount(path->mnt);
1085         struct mount *parent;
1086         struct dentry *mountpoint;
1087
1088         read_seqlock_excl(&mount_lock);
1089         parent = mnt->mnt_parent;
1090         if (parent == mnt) {
1091                 read_sequnlock_excl(&mount_lock);
1092                 return 0;
1093         }
1094         mntget(&parent->mnt);
1095         mountpoint = dget(mnt->mnt_mountpoint);
1096         read_sequnlock_excl(&mount_lock);
1097         dput(path->dentry);
1098         path->dentry = mountpoint;
1099         mntput(path->mnt);
1100         path->mnt = &parent->mnt;
1101         return 1;
1102 }
1103 EXPORT_SYMBOL(follow_up);
1104
1105 /*
1106  * Perform an automount
1107  * - return -EISDIR to tell follow_managed() to stop and return the path we
1108  *   were called with.
1109  */
1110 static int follow_automount(struct path *path, struct nameidata *nd,
1111                             bool *need_mntput)
1112 {
1113         struct vfsmount *mnt;
1114         int err;
1115
1116         if (!path->dentry->d_op || !path->dentry->d_op->d_automount)
1117                 return -EREMOTE;
1118
1119         /* We don't want to mount if someone's just doing a stat -
1120          * unless they're stat'ing a directory and appended a '/' to
1121          * the name.
1122          *
1123          * We do, however, want to mount if someone wants to open or
1124          * create a file of any type under the mountpoint, wants to
1125          * traverse through the mountpoint or wants to open the
1126          * mounted directory.  Also, autofs may mark negative dentries
1127          * as being automount points.  These will need the attentions
1128          * of the daemon to instantiate them before they can be used.
1129          */
1130         if (!(nd->flags & (LOOKUP_PARENT | LOOKUP_DIRECTORY |
1131                            LOOKUP_OPEN | LOOKUP_CREATE |
1132                            LOOKUP_AUTOMOUNT))) {
1133                 /* Positive dentry that isn't meant to trigger an
1134                  * automount, EISDIR will allow it to be used,
1135                  * otherwise there's no mount here "now" so return
1136                  * ENOENT.
1137                  */
1138                 if (path->dentry->d_inode)
1139                         return -EISDIR;
1140                 else
1141                         return -ENOENT;
1142         }
1143
1144         if (path->dentry->d_sb->s_user_ns != &init_user_ns)
1145                 return -EACCES;
1146
1147         nd->total_link_count++;
1148         if (nd->total_link_count >= 40)
1149                 return -ELOOP;
1150
1151         mnt = path->dentry->d_op->d_automount(path);
1152         if (IS_ERR(mnt)) {
1153                 /*
1154                  * The filesystem is allowed to return -EISDIR here to indicate
1155                  * it doesn't want to automount.  For instance, autofs would do
1156                  * this so that its userspace daemon can mount on this dentry.
1157                  *
1158                  * However, we can only permit this if it's a terminal point in
1159                  * the path being looked up; if it wasn't then the remainder of
1160                  * the path is inaccessible and we should say so.
1161                  */
1162                 if (PTR_ERR(mnt) == -EISDIR && (nd->flags & LOOKUP_PARENT))
1163                         return -EREMOTE;
1164                 return PTR_ERR(mnt);
1165         }
1166
1167         if (!mnt) /* mount collision */
1168                 return 0;
1169
1170         if (!*need_mntput) {
1171                 /* lock_mount() may release path->mnt on error */
1172                 mntget(path->mnt);
1173                 *need_mntput = true;
1174         }
1175         err = finish_automount(mnt, path);
1176
1177         switch (err) {
1178         case -EBUSY:
1179                 /* Someone else made a mount here whilst we were busy */
1180                 return 0;
1181         case 0:
1182                 path_put(path);
1183                 path->mnt = mnt;
1184                 path->dentry = dget(mnt->mnt_root);
1185                 return 0;
1186         default:
1187                 return err;
1188         }
1189
1190 }
1191
1192 /*
1193  * Handle a dentry that is managed in some way.
1194  * - Flagged for transit management (autofs)
1195  * - Flagged as mountpoint
1196  * - Flagged as automount point
1197  *
1198  * This may only be called in refwalk mode.
1199  *
1200  * Serialization is taken care of in namespace.c
1201  */
1202 static int follow_managed(struct path *path, struct nameidata *nd)
1203 {
1204         struct vfsmount *mnt = path->mnt; /* held by caller, must be left alone */
1205         unsigned managed;
1206         bool need_mntput = false;
1207         int ret = 0;
1208
1209         /* Given that we're not holding a lock here, we retain the value in a
1210          * local variable for each dentry as we look at it so that we don't see
1211          * the components of that value change under us */
1212         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1213                managed &= DCACHE_MANAGED_DENTRY,
1214                unlikely(managed != 0)) {
1215                 /* Allow the filesystem to manage the transit without i_mutex
1216                  * being held. */
1217                 if (managed & DCACHE_MANAGE_TRANSIT) {
1218                         BUG_ON(!path->dentry->d_op);
1219                         BUG_ON(!path->dentry->d_op->d_manage);
1220                         ret = path->dentry->d_op->d_manage(path, false);
1221                         if (ret < 0)
1222                                 break;
1223                 }
1224
1225                 /* Transit to a mounted filesystem. */
1226                 if (managed & DCACHE_MOUNTED) {
1227                         struct vfsmount *mounted = lookup_mnt(path);
1228                         if (mounted) {
1229                                 dput(path->dentry);
1230                                 if (need_mntput)
1231                                         mntput(path->mnt);
1232                                 path->mnt = mounted;
1233                                 path->dentry = dget(mounted->mnt_root);
1234                                 need_mntput = true;
1235                                 continue;
1236                         }
1237
1238                         /* Something is mounted on this dentry in another
1239                          * namespace and/or whatever was mounted there in this
1240                          * namespace got unmounted before lookup_mnt() could
1241                          * get it */
1242                 }
1243
1244                 /* Handle an automount point */
1245                 if (managed & DCACHE_NEED_AUTOMOUNT) {
1246                         ret = follow_automount(path, nd, &need_mntput);
1247                         if (ret < 0)
1248                                 break;
1249                         continue;
1250                 }
1251
1252                 /* We didn't change the current path point */
1253                 break;
1254         }
1255
1256         if (need_mntput && path->mnt == mnt)
1257                 mntput(path->mnt);
1258         if (ret == -EISDIR || !ret)
1259                 ret = 1;
1260         if (need_mntput)
1261                 nd->flags |= LOOKUP_JUMPED;
1262         if (unlikely(ret < 0))
1263                 path_put_conditional(path, nd);
1264         return ret;
1265 }
1266
1267 int follow_down_one(struct path *path)
1268 {
1269         struct vfsmount *mounted;
1270
1271         mounted = lookup_mnt(path);
1272         if (mounted) {
1273                 dput(path->dentry);
1274                 mntput(path->mnt);
1275                 path->mnt = mounted;
1276                 path->dentry = dget(mounted->mnt_root);
1277                 return 1;
1278         }
1279         return 0;
1280 }
1281 EXPORT_SYMBOL(follow_down_one);
1282
1283 static inline int managed_dentry_rcu(const struct path *path)
1284 {
1285         return (path->dentry->d_flags & DCACHE_MANAGE_TRANSIT) ?
1286                 path->dentry->d_op->d_manage(path, true) : 0;
1287 }
1288
1289 /*
1290  * Try to skip to top of mountpoint pile in rcuwalk mode.  Fail if
1291  * we meet a managed dentry that would need blocking.
1292  */
1293 static bool __follow_mount_rcu(struct nameidata *nd, struct path *path,
1294                                struct inode **inode, unsigned *seqp)
1295 {
1296         for (;;) {
1297                 struct mount *mounted;
1298                 /*
1299                  * Don't forget we might have a non-mountpoint managed dentry
1300                  * that wants to block transit.
1301                  */
1302                 switch (managed_dentry_rcu(path)) {
1303                 case -ECHILD:
1304                 default:
1305                         return false;
1306                 case -EISDIR:
1307                         return true;
1308                 case 0:
1309                         break;
1310                 }
1311
1312                 if (!d_mountpoint(path->dentry))
1313                         return !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1314
1315                 mounted = __lookup_mnt(path->mnt, path->dentry);
1316                 if (!mounted)
1317                         break;
1318                 path->mnt = &mounted->mnt;
1319                 path->dentry = mounted->mnt.mnt_root;
1320                 nd->flags |= LOOKUP_JUMPED;
1321                 *seqp = read_seqcount_begin(&path->dentry->d_seq);
1322                 /*
1323                  * Update the inode too. We don't need to re-check the
1324                  * dentry sequence number here after this d_inode read,
1325                  * because a mount-point is always pinned.
1326                  */
1327                 *inode = path->dentry->d_inode;
1328         }
1329         return !read_seqretry(&mount_lock, nd->m_seq) &&
1330                 !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1331 }
1332
1333 static int follow_dotdot_rcu(struct nameidata *nd)
1334 {
1335         struct inode *inode = nd->inode;
1336
1337         while (1) {
1338                 if (path_equal(&nd->path, &nd->root))
1339                         break;
1340                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1341                         struct dentry *old = nd->path.dentry;
1342                         struct dentry *parent = old->d_parent;
1343                         unsigned seq;
1344
1345                         inode = parent->d_inode;
1346                         seq = read_seqcount_begin(&parent->d_seq);
1347                         if (unlikely(read_seqcount_retry(&old->d_seq, nd->seq)))
1348                                 return -ECHILD;
1349                         nd->path.dentry = parent;
1350                         nd->seq = seq;
1351                         if (unlikely(!path_connected(&nd->path)))
1352                                 return -ENOENT;
1353                         break;
1354                 } else {
1355                         struct mount *mnt = real_mount(nd->path.mnt);
1356                         struct mount *mparent = mnt->mnt_parent;
1357                         struct dentry *mountpoint = mnt->mnt_mountpoint;
1358                         struct inode *inode2 = mountpoint->d_inode;
1359                         unsigned seq = read_seqcount_begin(&mountpoint->d_seq);
1360                         if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1361                                 return -ECHILD;
1362                         if (&mparent->mnt == nd->path.mnt)
1363                                 break;
1364                         /* we know that mountpoint was pinned */
1365                         nd->path.dentry = mountpoint;
1366                         nd->path.mnt = &mparent->mnt;
1367                         inode = inode2;
1368                         nd->seq = seq;
1369                 }
1370         }
1371         while (unlikely(d_mountpoint(nd->path.dentry))) {
1372                 struct mount *mounted;
1373                 mounted = __lookup_mnt(nd->path.mnt, nd->path.dentry);
1374                 if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1375                         return -ECHILD;
1376                 if (!mounted)
1377                         break;
1378                 nd->path.mnt = &mounted->mnt;
1379                 nd->path.dentry = mounted->mnt.mnt_root;
1380                 inode = nd->path.dentry->d_inode;
1381                 nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
1382         }
1383         nd->inode = inode;
1384         return 0;
1385 }
1386
1387 /*
1388  * Follow down to the covering mount currently visible to userspace.  At each
1389  * point, the filesystem owning that dentry may be queried as to whether the
1390  * caller is permitted to proceed or not.
1391  */
1392 int follow_down(struct path *path)
1393 {
1394         unsigned managed;
1395         int ret;
1396
1397         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1398                unlikely(managed & DCACHE_MANAGED_DENTRY)) {
1399                 /* Allow the filesystem to manage the transit without i_mutex
1400                  * being held.
1401                  *
1402                  * We indicate to the filesystem if someone is trying to mount
1403                  * something here.  This gives autofs the chance to deny anyone
1404                  * other than its daemon the right to mount on its
1405                  * superstructure.
1406                  *
1407                  * The filesystem may sleep at this point.
1408                  */
1409                 if (managed & DCACHE_MANAGE_TRANSIT) {
1410                         BUG_ON(!path->dentry->d_op);
1411                         BUG_ON(!path->dentry->d_op->d_manage);
1412                         ret = path->dentry->d_op->d_manage(path, false);
1413                         if (ret < 0)
1414                                 return ret == -EISDIR ? 0 : ret;
1415                 }
1416
1417                 /* Transit to a mounted filesystem. */
1418                 if (managed & DCACHE_MOUNTED) {
1419                         struct vfsmount *mounted = lookup_mnt(path);
1420                         if (!mounted)
1421                                 break;
1422                         dput(path->dentry);
1423                         mntput(path->mnt);
1424                         path->mnt = mounted;
1425                         path->dentry = dget(mounted->mnt_root);
1426                         continue;
1427                 }
1428
1429                 /* Don't handle automount points here */
1430                 break;
1431         }
1432         return 0;
1433 }
1434 EXPORT_SYMBOL(follow_down);
1435
1436 /*
1437  * Skip to top of mountpoint pile in refwalk mode for follow_dotdot()
1438  */
1439 static void follow_mount(struct path *path)
1440 {
1441         while (d_mountpoint(path->dentry)) {
1442                 struct vfsmount *mounted = lookup_mnt(path);
1443                 if (!mounted)
1444                         break;
1445                 dput(path->dentry);
1446                 mntput(path->mnt);
1447                 path->mnt = mounted;
1448                 path->dentry = dget(mounted->mnt_root);
1449         }
1450 }
1451
1452 static int path_parent_directory(struct path *path)
1453 {
1454         struct dentry *old = path->dentry;
1455         /* rare case of legitimate dget_parent()... */
1456         path->dentry = dget_parent(path->dentry);
1457         dput(old);
1458         if (unlikely(!path_connected(path)))
1459                 return -ENOENT;
1460         return 0;
1461 }
1462
1463 static int follow_dotdot(struct nameidata *nd)
1464 {
1465         while(1) {
1466                 if (nd->path.dentry == nd->root.dentry &&
1467                     nd->path.mnt == nd->root.mnt) {
1468                         break;
1469                 }
1470                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1471                         int ret = path_parent_directory(&nd->path);
1472                         if (ret)
1473                                 return ret;
1474                         break;
1475                 }
1476                 if (!follow_up(&nd->path))
1477                         break;
1478         }
1479         follow_mount(&nd->path);
1480         nd->inode = nd->path.dentry->d_inode;
1481         return 0;
1482 }
1483
1484 /*
1485  * This looks up the name in dcache and possibly revalidates the found dentry.
1486  * NULL is returned if the dentry does not exist in the cache.
1487  */
1488 static struct dentry *lookup_dcache(const struct qstr *name,
1489                                     struct dentry *dir,
1490                                     unsigned int flags)
1491 {
1492         struct dentry *dentry = d_lookup(dir, name);
1493         if (dentry) {
1494                 int error = d_revalidate(dentry, flags);
1495                 if (unlikely(error <= 0)) {
1496                         if (!error)
1497                                 d_invalidate(dentry);
1498                         dput(dentry);
1499                         return ERR_PTR(error);
1500                 }
1501         }
1502         return dentry;
1503 }
1504
1505 /*
1506  * Call i_op->lookup on the dentry.  The dentry must be negative and
1507  * unhashed.
1508  *
1509  * dir->d_inode->i_mutex must be held
1510  */
1511 static struct dentry *lookup_real(struct inode *dir, struct dentry *dentry,
1512                                   unsigned int flags)
1513 {
1514         struct dentry *old;
1515
1516         /* Don't create child dentry for a dead directory. */
1517         if (unlikely(IS_DEADDIR(dir))) {
1518                 dput(dentry);
1519                 return ERR_PTR(-ENOENT);
1520         }
1521
1522         old = dir->i_op->lookup(dir, dentry, flags);
1523         if (unlikely(old)) {
1524                 dput(dentry);
1525                 dentry = old;
1526         }
1527         return dentry;
1528 }
1529
1530 static struct dentry *__lookup_hash(const struct qstr *name,
1531                 struct dentry *base, unsigned int flags)
1532 {
1533         struct dentry *dentry = lookup_dcache(name, base, flags);
1534
1535         if (dentry)
1536                 return dentry;
1537
1538         dentry = d_alloc(base, name);
1539         if (unlikely(!dentry))
1540                 return ERR_PTR(-ENOMEM);
1541
1542         return lookup_real(base->d_inode, dentry, flags);
1543 }
1544
1545 static int lookup_fast(struct nameidata *nd,
1546                        struct path *path, struct inode **inode,
1547                        unsigned *seqp)
1548 {
1549         struct vfsmount *mnt = nd->path.mnt;
1550         struct dentry *dentry, *parent = nd->path.dentry;
1551         int status = 1;
1552         int err;
1553
1554         /*
1555          * Rename seqlock is not required here because in the off chance
1556          * of a false negative due to a concurrent rename, the caller is
1557          * going to fall back to non-racy lookup.
1558          */
1559         if (nd->flags & LOOKUP_RCU) {
1560                 unsigned seq;
1561                 bool negative;
1562                 dentry = __d_lookup_rcu(parent, &nd->last, &seq);
1563                 if (unlikely(!dentry)) {
1564                         if (unlazy_walk(nd))
1565                                 return -ECHILD;
1566                         return 0;
1567                 }
1568
1569                 /*
1570                  * This sequence count validates that the inode matches
1571                  * the dentry name information from lookup.
1572                  */
1573                 *inode = d_backing_inode(dentry);
1574                 negative = d_is_negative(dentry);
1575                 if (unlikely(read_seqcount_retry(&dentry->d_seq, seq)))
1576                         return -ECHILD;
1577
1578                 /*
1579                  * This sequence count validates that the parent had no
1580                  * changes while we did the lookup of the dentry above.
1581                  *
1582                  * The memory barrier in read_seqcount_begin of child is
1583                  *  enough, we can use __read_seqcount_retry here.
1584                  */
1585                 if (unlikely(__read_seqcount_retry(&parent->d_seq, nd->seq)))
1586                         return -ECHILD;
1587
1588                 *seqp = seq;
1589                 status = d_revalidate(dentry, nd->flags);
1590                 if (likely(status > 0)) {
1591                         /*
1592                          * Note: do negative dentry check after revalidation in
1593                          * case that drops it.
1594                          */
1595                         if (unlikely(negative))
1596                                 return -ENOENT;
1597                         path->mnt = mnt;
1598                         path->dentry = dentry;
1599                         if (likely(__follow_mount_rcu(nd, path, inode, seqp)))
1600                                 return 1;
1601                 }
1602                 if (unlazy_child(nd, dentry, seq))
1603                         return -ECHILD;
1604                 if (unlikely(status == -ECHILD))
1605                         /* we'd been told to redo it in non-rcu mode */
1606                         status = d_revalidate(dentry, nd->flags);
1607         } else {
1608                 dentry = __d_lookup(parent, &nd->last);
1609                 if (unlikely(!dentry))
1610                         return 0;
1611                 status = d_revalidate(dentry, nd->flags);
1612         }
1613         if (unlikely(status <= 0)) {
1614                 if (!status)
1615                         d_invalidate(dentry);
1616                 dput(dentry);
1617                 return status;
1618         }
1619         if (unlikely(d_is_negative(dentry))) {
1620                 dput(dentry);
1621                 return -ENOENT;
1622         }
1623
1624         path->mnt = mnt;
1625         path->dentry = dentry;
1626         err = follow_managed(path, nd);
1627         if (likely(err > 0))
1628                 *inode = d_backing_inode(path->dentry);
1629         return err;
1630 }
1631
1632 /* Fast lookup failed, do it the slow way */
1633 static struct dentry *lookup_slow(const struct qstr *name,
1634                                   struct dentry *dir,
1635                                   unsigned int flags)
1636 {
1637         struct dentry *dentry = ERR_PTR(-ENOENT), *old;
1638         struct inode *inode = dir->d_inode;
1639         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
1640
1641         inode_lock_shared(inode);
1642         /* Don't go there if it's already dead */
1643         if (unlikely(IS_DEADDIR(inode)))
1644                 goto out;
1645 again:
1646         dentry = d_alloc_parallel(dir, name, &wq);
1647         if (IS_ERR(dentry))
1648                 goto out;
1649         if (unlikely(!d_in_lookup(dentry))) {
1650                 if (!(flags & LOOKUP_NO_REVAL)) {
1651                         int error = d_revalidate(dentry, flags);
1652                         if (unlikely(error <= 0)) {
1653                                 if (!error) {
1654                                         d_invalidate(dentry);
1655                                         dput(dentry);
1656                                         goto again;
1657                                 }
1658                                 dput(dentry);
1659                                 dentry = ERR_PTR(error);
1660                         }
1661                 }
1662         } else {
1663                 old = inode->i_op->lookup(inode, dentry, flags);
1664                 d_lookup_done(dentry);
1665                 if (unlikely(old)) {
1666                         dput(dentry);
1667                         dentry = old;
1668                 }
1669         }
1670 out:
1671         inode_unlock_shared(inode);
1672         return dentry;
1673 }
1674
1675 static inline int may_lookup(struct nameidata *nd)
1676 {
1677         if (nd->flags & LOOKUP_RCU) {
1678                 int err = inode_permission(nd->inode, MAY_EXEC|MAY_NOT_BLOCK);
1679                 if (err != -ECHILD)
1680                         return err;
1681                 if (unlazy_walk(nd))
1682                         return -ECHILD;
1683         }
1684         return inode_permission(nd->inode, MAY_EXEC);
1685 }
1686
1687 static inline int handle_dots(struct nameidata *nd, int type)
1688 {
1689         if (type == LAST_DOTDOT) {
1690                 if (!nd->root.mnt)
1691                         set_root(nd);
1692                 if (nd->flags & LOOKUP_RCU) {
1693                         return follow_dotdot_rcu(nd);
1694                 } else
1695                         return follow_dotdot(nd);
1696         }
1697         return 0;
1698 }
1699
1700 static int pick_link(struct nameidata *nd, struct path *link,
1701                      struct inode *inode, unsigned seq)
1702 {
1703         int error;
1704         struct saved *last;
1705         if (unlikely(nd->total_link_count++ >= MAXSYMLINKS)) {
1706                 path_to_nameidata(link, nd);
1707                 return -ELOOP;
1708         }
1709         if (!(nd->flags & LOOKUP_RCU)) {
1710                 if (link->mnt == nd->path.mnt)
1711                         mntget(link->mnt);
1712         }
1713         error = nd_alloc_stack(nd);
1714         if (unlikely(error)) {
1715                 if (error == -ECHILD) {
1716                         if (unlikely(!legitimize_path(nd, link, seq))) {
1717                                 drop_links(nd);
1718                                 nd->depth = 0;
1719                                 nd->flags &= ~LOOKUP_RCU;
1720                                 nd->path.mnt = NULL;
1721                                 nd->path.dentry = NULL;
1722                                 if (!(nd->flags & LOOKUP_ROOT))
1723                                         nd->root.mnt = NULL;
1724                                 rcu_read_unlock();
1725                         } else if (likely(unlazy_walk(nd)) == 0)
1726                                 error = nd_alloc_stack(nd);
1727                 }
1728                 if (error) {
1729                         path_put(link);
1730                         return error;
1731                 }
1732         }
1733
1734         last = nd->stack + nd->depth++;
1735         last->link = *link;
1736         clear_delayed_call(&last->done);
1737         nd->link_inode = inode;
1738         last->seq = seq;
1739         return 1;
1740 }
1741
1742 enum {WALK_FOLLOW = 1, WALK_MORE = 2};
1743
1744 /*
1745  * Do we need to follow links? We _really_ want to be able
1746  * to do this check without having to look at inode->i_op,
1747  * so we keep a cache of "no, this doesn't need follow_link"
1748  * for the common case.
1749  */
1750 static inline int step_into(struct nameidata *nd, struct path *path,
1751                             int flags, struct inode *inode, unsigned seq)
1752 {
1753         if (!(flags & WALK_MORE) && nd->depth)
1754                 put_link(nd);
1755         if (likely(!d_is_symlink(path->dentry)) ||
1756            !(flags & WALK_FOLLOW || nd->flags & LOOKUP_FOLLOW)) {
1757                 /* not a symlink or should not follow */
1758                 path_to_nameidata(path, nd);
1759                 nd->inode = inode;
1760                 nd->seq = seq;
1761                 return 0;
1762         }
1763         /* make sure that d_is_symlink above matches inode */
1764         if (nd->flags & LOOKUP_RCU) {
1765                 if (read_seqcount_retry(&path->dentry->d_seq, seq))
1766                         return -ECHILD;
1767         }
1768         return pick_link(nd, path, inode, seq);
1769 }
1770
1771 static int walk_component(struct nameidata *nd, int flags)
1772 {
1773         struct path path;
1774         struct inode *inode;
1775         unsigned seq;
1776         int err;
1777         /*
1778          * "." and ".." are special - ".." especially so because it has
1779          * to be able to know about the current root directory and
1780          * parent relationships.
1781          */
1782         if (unlikely(nd->last_type != LAST_NORM)) {
1783                 err = handle_dots(nd, nd->last_type);
1784                 if (!(flags & WALK_MORE) && nd->depth)
1785                         put_link(nd);
1786                 return err;
1787         }
1788         err = lookup_fast(nd, &path, &inode, &seq);
1789         if (unlikely(err <= 0)) {
1790                 if (err < 0)
1791                         return err;
1792                 path.dentry = lookup_slow(&nd->last, nd->path.dentry,
1793                                           nd->flags);
1794                 if (IS_ERR(path.dentry))
1795                         return PTR_ERR(path.dentry);
1796
1797                 path.mnt = nd->path.mnt;
1798                 err = follow_managed(&path, nd);
1799                 if (unlikely(err < 0))
1800                         return err;
1801
1802                 if (unlikely(d_is_negative(path.dentry))) {
1803                         path_to_nameidata(&path, nd);
1804                         return -ENOENT;
1805                 }
1806
1807                 seq = 0;        /* we are already out of RCU mode */
1808                 inode = d_backing_inode(path.dentry);
1809         }
1810
1811         return step_into(nd, &path, flags, inode, seq);
1812 }
1813
1814 /*
1815  * We can do the critical dentry name comparison and hashing
1816  * operations one word at a time, but we are limited to:
1817  *
1818  * - Architectures with fast unaligned word accesses. We could
1819  *   do a "get_unaligned()" if this helps and is sufficiently
1820  *   fast.
1821  *
1822  * - non-CONFIG_DEBUG_PAGEALLOC configurations (so that we
1823  *   do not trap on the (extremely unlikely) case of a page
1824  *   crossing operation.
1825  *
1826  * - Furthermore, we need an efficient 64-bit compile for the
1827  *   64-bit case in order to generate the "number of bytes in
1828  *   the final mask". Again, that could be replaced with a
1829  *   efficient population count instruction or similar.
1830  */
1831 #ifdef CONFIG_DCACHE_WORD_ACCESS
1832
1833 #include <asm/word-at-a-time.h>
1834
1835 #ifdef HASH_MIX
1836
1837 /* Architecture provides HASH_MIX and fold_hash() in <asm/hash.h> */
1838
1839 #elif defined(CONFIG_64BIT)
1840 /*
1841  * Register pressure in the mixing function is an issue, particularly
1842  * on 32-bit x86, but almost any function requires one state value and
1843  * one temporary.  Instead, use a function designed for two state values
1844  * and no temporaries.
1845  *
1846  * This function cannot create a collision in only two iterations, so
1847  * we have two iterations to achieve avalanche.  In those two iterations,
1848  * we have six layers of mixing, which is enough to spread one bit's
1849  * influence out to 2^6 = 64 state bits.
1850  *
1851  * Rotate constants are scored by considering either 64 one-bit input
1852  * deltas or 64*63/2 = 2016 two-bit input deltas, and finding the
1853  * probability of that delta causing a change to each of the 128 output
1854  * bits, using a sample of random initial states.
1855  *
1856  * The Shannon entropy of the computed probabilities is then summed
1857  * to produce a score.  Ideally, any input change has a 50% chance of
1858  * toggling any given output bit.
1859  *
1860  * Mixing scores (in bits) for (12,45):
1861  * Input delta: 1-bit      2-bit
1862  * 1 round:     713.3    42542.6
1863  * 2 rounds:   2753.7   140389.8
1864  * 3 rounds:   5954.1   233458.2
1865  * 4 rounds:   7862.6   256672.2
1866  * Perfect:    8192     258048
1867  *            (64*128) (64*63/2 * 128)
1868  */
1869 #define HASH_MIX(x, y, a)       \
1870         (       x ^= (a),       \
1871         y ^= x, x = rol64(x,12),\
1872         x += y, y = rol64(y,45),\
1873         y *= 9                  )
1874
1875 /*
1876  * Fold two longs into one 32-bit hash value.  This must be fast, but
1877  * latency isn't quite as critical, as there is a fair bit of additional
1878  * work done before the hash value is used.
1879  */
1880 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1881 {
1882         y ^= x * GOLDEN_RATIO_64;
1883         y *= GOLDEN_RATIO_64;
1884         return y >> 32;
1885 }
1886
1887 #else   /* 32-bit case */
1888
1889 /*
1890  * Mixing scores (in bits) for (7,20):
1891  * Input delta: 1-bit      2-bit
1892  * 1 round:     330.3     9201.6
1893  * 2 rounds:   1246.4    25475.4
1894  * 3 rounds:   1907.1    31295.1
1895  * 4 rounds:   2042.3    31718.6
1896  * Perfect:    2048      31744
1897  *            (32*64)   (32*31/2 * 64)
1898  */
1899 #define HASH_MIX(x, y, a)       \
1900         (       x ^= (a),       \
1901         y ^= x, x = rol32(x, 7),\
1902         x += y, y = rol32(y,20),\
1903         y *= 9                  )
1904
1905 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1906 {
1907         /* Use arch-optimized multiply if one exists */
1908         return __hash_32(y ^ __hash_32(x));
1909 }
1910
1911 #endif
1912
1913 /*
1914  * Return the hash of a string of known length.  This is carfully
1915  * designed to match hash_name(), which is the more critical function.
1916  * In particular, we must end by hashing a final word containing 0..7
1917  * payload bytes, to match the way that hash_name() iterates until it
1918  * finds the delimiter after the name.
1919  */
1920 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1921 {
1922         unsigned long a, x = 0, y = (unsigned long)salt;
1923
1924         for (;;) {
1925                 if (!len)
1926                         goto done;
1927                 a = load_unaligned_zeropad(name);
1928                 if (len < sizeof(unsigned long))
1929                         break;
1930                 HASH_MIX(x, y, a);
1931                 name += sizeof(unsigned long);
1932                 len -= sizeof(unsigned long);
1933         }
1934         x ^= a & bytemask_from_count(len);
1935 done:
1936         return fold_hash(x, y);
1937 }
1938 EXPORT_SYMBOL(full_name_hash);
1939
1940 /* Return the "hash_len" (hash and length) of a null-terminated string */
1941 u64 hashlen_string(const void *salt, const char *name)
1942 {
1943         unsigned long a = 0, x = 0, y = (unsigned long)salt;
1944         unsigned long adata, mask, len;
1945         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1946
1947         len = 0;
1948         goto inside;
1949
1950         do {
1951                 HASH_MIX(x, y, a);
1952                 len += sizeof(unsigned long);
1953 inside:
1954                 a = load_unaligned_zeropad(name+len);
1955         } while (!has_zero(a, &adata, &constants));
1956
1957         adata = prep_zero_mask(a, adata, &constants);
1958         mask = create_zero_mask(adata);
1959         x ^= a & zero_bytemask(mask);
1960
1961         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1962 }
1963 EXPORT_SYMBOL(hashlen_string);
1964
1965 /*
1966  * Calculate the length and hash of the path component, and
1967  * return the "hash_len" as the result.
1968  */
1969 static inline u64 hash_name(const void *salt, const char *name)
1970 {
1971         unsigned long a = 0, b, x = 0, y = (unsigned long)salt;
1972         unsigned long adata, bdata, mask, len;
1973         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1974
1975         len = 0;
1976         goto inside;
1977
1978         do {
1979                 HASH_MIX(x, y, a);
1980                 len += sizeof(unsigned long);
1981 inside:
1982                 a = load_unaligned_zeropad(name+len);
1983                 b = a ^ REPEAT_BYTE('/');
1984         } while (!(has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)));
1985
1986         adata = prep_zero_mask(a, adata, &constants);
1987         bdata = prep_zero_mask(b, bdata, &constants);
1988         mask = create_zero_mask(adata | bdata);
1989         x ^= a & zero_bytemask(mask);
1990
1991         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1992 }
1993
1994 #else   /* !CONFIG_DCACHE_WORD_ACCESS: Slow, byte-at-a-time version */
1995
1996 /* Return the hash of a string of known length */
1997 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1998 {
1999         unsigned long hash = init_name_hash(salt);
2000         while (len--)
2001                 hash = partial_name_hash((unsigned char)*name++, hash);
2002         return end_name_hash(hash);
2003 }
2004 EXPORT_SYMBOL(full_name_hash);
2005
2006 /* Return the "hash_len" (hash and length) of a null-terminated string */
2007 u64 hashlen_string(const void *salt, const char *name)
2008 {
2009         unsigned long hash = init_name_hash(salt);
2010         unsigned long len = 0, c;
2011
2012         c = (unsigned char)*name;
2013         while (c) {
2014                 len++;
2015                 hash = partial_name_hash(c, hash);
2016                 c = (unsigned char)name[len];
2017         }
2018         return hashlen_create(end_name_hash(hash), len);
2019 }
2020 EXPORT_SYMBOL(hashlen_string);
2021
2022 /*
2023  * We know there's a real path component here of at least
2024  * one character.
2025  */
2026 static inline u64 hash_name(const void *salt, const char *name)
2027 {
2028         unsigned long hash = init_name_hash(salt);
2029         unsigned long len = 0, c;
2030
2031         c = (unsigned char)*name;
2032         do {
2033                 len++;
2034                 hash = partial_name_hash(c, hash);
2035                 c = (unsigned char)name[len];
2036         } while (c && c != '/');
2037         return hashlen_create(end_name_hash(hash), len);
2038 }
2039
2040 #endif
2041
2042 /*
2043  * Name resolution.
2044  * This is the basic name resolution function, turning a pathname into
2045  * the final dentry. We expect 'base' to be positive and a directory.
2046  *
2047  * Returns 0 and nd will have valid dentry and mnt on success.
2048  * Returns error and drops reference to input namei data on failure.
2049  */
2050 static int link_path_walk(const char *name, struct nameidata *nd)
2051 {
2052         int err;
2053
2054         while (*name=='/')
2055                 name++;
2056         if (!*name)
2057                 return 0;
2058
2059         /* At this point we know we have a real path component. */
2060         for(;;) {
2061                 u64 hash_len;
2062                 int type;
2063
2064                 err = may_lookup(nd);
2065                 if (err)
2066                         return err;
2067
2068                 hash_len = hash_name(nd->path.dentry, name);
2069
2070                 type = LAST_NORM;
2071                 if (name[0] == '.') switch (hashlen_len(hash_len)) {
2072                         case 2:
2073                                 if (name[1] == '.') {
2074                                         type = LAST_DOTDOT;
2075                                         nd->flags |= LOOKUP_JUMPED;
2076                                 }
2077                                 break;
2078                         case 1:
2079                                 type = LAST_DOT;
2080                 }
2081                 if (likely(type == LAST_NORM)) {
2082                         struct dentry *parent = nd->path.dentry;
2083                         nd->flags &= ~LOOKUP_JUMPED;
2084                         if (unlikely(parent->d_flags & DCACHE_OP_HASH)) {
2085                                 struct qstr this = { { .hash_len = hash_len }, .name = name };
2086                                 err = parent->d_op->d_hash(parent, &this);
2087                                 if (err < 0)
2088                                         return err;
2089                                 hash_len = this.hash_len;
2090                                 name = this.name;
2091                         }
2092                 }
2093
2094                 nd->last.hash_len = hash_len;
2095                 nd->last.name = name;
2096                 nd->last_type = type;
2097
2098                 name += hashlen_len(hash_len);
2099                 if (!*name)
2100                         goto OK;
2101                 /*
2102                  * If it wasn't NUL, we know it was '/'. Skip that
2103                  * slash, and continue until no more slashes.
2104                  */
2105                 do {
2106                         name++;
2107                 } while (unlikely(*name == '/'));
2108                 if (unlikely(!*name)) {
2109 OK:
2110                         /* pathname body, done */
2111                         if (!nd->depth)
2112                                 return 0;
2113                         name = nd->stack[nd->depth - 1].name;
2114                         /* trailing symlink, done */
2115                         if (!name)
2116                                 return 0;
2117                         /* last component of nested symlink */
2118                         err = walk_component(nd, WALK_FOLLOW);
2119                 } else {
2120                         /* not the last component */
2121                         err = walk_component(nd, WALK_FOLLOW | WALK_MORE);
2122                 }
2123                 if (err < 0)
2124                         return err;
2125
2126                 if (err) {
2127                         const char *s = get_link(nd);
2128
2129                         if (IS_ERR(s))
2130                                 return PTR_ERR(s);
2131                         err = 0;
2132                         if (unlikely(!s)) {
2133                                 /* jumped */
2134                                 put_link(nd);
2135                         } else {
2136                                 nd->stack[nd->depth - 1].name = name;
2137                                 name = s;
2138                                 continue;
2139                         }
2140                 }
2141                 if (unlikely(!d_can_lookup(nd->path.dentry))) {
2142                         if (nd->flags & LOOKUP_RCU) {
2143                                 if (unlazy_walk(nd))
2144                                         return -ECHILD;
2145                         }
2146                         return -ENOTDIR;
2147                 }
2148         }
2149 }
2150
2151 static const char *path_init(struct nameidata *nd, unsigned flags)
2152 {
2153         const char *s = nd->name->name;
2154
2155         if (!*s)
2156                 flags &= ~LOOKUP_RCU;
2157
2158         nd->last_type = LAST_ROOT; /* if there are only slashes... */
2159         nd->flags = flags | LOOKUP_JUMPED | LOOKUP_PARENT;
2160         nd->depth = 0;
2161         if (flags & LOOKUP_ROOT) {
2162                 struct dentry *root = nd->root.dentry;
2163                 struct inode *inode = root->d_inode;
2164                 if (*s && unlikely(!d_can_lookup(root)))
2165                         return ERR_PTR(-ENOTDIR);
2166                 nd->path = nd->root;
2167                 nd->inode = inode;
2168                 if (flags & LOOKUP_RCU) {
2169                         rcu_read_lock();
2170                         nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2171                         nd->root_seq = nd->seq;
2172                         nd->m_seq = read_seqbegin(&mount_lock);
2173                 } else {
2174                         path_get(&nd->path);
2175                 }
2176                 return s;
2177         }
2178
2179         nd->root.mnt = NULL;
2180         nd->path.mnt = NULL;
2181         nd->path.dentry = NULL;
2182
2183         nd->m_seq = read_seqbegin(&mount_lock);
2184         if (*s == '/') {
2185                 if (flags & LOOKUP_RCU)
2186                         rcu_read_lock();
2187                 set_root(nd);
2188                 if (likely(!nd_jump_root(nd)))
2189                         return s;
2190                 nd->root.mnt = NULL;
2191                 rcu_read_unlock();
2192                 return ERR_PTR(-ECHILD);
2193         } else if (nd->dfd == AT_FDCWD) {
2194                 if (flags & LOOKUP_RCU) {
2195                         struct fs_struct *fs = current->fs;
2196                         unsigned seq;
2197
2198                         rcu_read_lock();
2199
2200                         do {
2201                                 seq = read_seqcount_begin(&fs->seq);
2202                                 nd->path = fs->pwd;
2203                                 nd->inode = nd->path.dentry->d_inode;
2204                                 nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2205                         } while (read_seqcount_retry(&fs->seq, seq));
2206                 } else {
2207                         get_fs_pwd(current->fs, &nd->path);
2208                         nd->inode = nd->path.dentry->d_inode;
2209                 }
2210                 return s;
2211         } else {
2212                 /* Caller must check execute permissions on the starting path component */
2213                 struct fd f = fdget_raw(nd->dfd);
2214                 struct dentry *dentry;
2215
2216                 if (!f.file)
2217                         return ERR_PTR(-EBADF);
2218
2219                 dentry = f.file->f_path.dentry;
2220
2221                 if (*s) {
2222                         if (!d_can_lookup(dentry)) {
2223                                 fdput(f);
2224                                 return ERR_PTR(-ENOTDIR);
2225                         }
2226                 }
2227
2228                 nd->path = f.file->f_path;
2229                 if (flags & LOOKUP_RCU) {
2230                         rcu_read_lock();
2231                         nd->inode = nd->path.dentry->d_inode;
2232                         nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
2233                 } else {
2234                         path_get(&nd->path);
2235                         nd->inode = nd->path.dentry->d_inode;
2236                 }
2237                 fdput(f);
2238                 return s;
2239         }
2240 }
2241
2242 static const char *trailing_symlink(struct nameidata *nd)
2243 {
2244         const char *s;
2245         int error = may_follow_link(nd);
2246         if (unlikely(error))
2247                 return ERR_PTR(error);
2248         nd->flags |= LOOKUP_PARENT;
2249         nd->stack[0].name = NULL;
2250         s = get_link(nd);
2251         return s ? s : "";
2252 }
2253
2254 static inline int lookup_last(struct nameidata *nd)
2255 {
2256         if (nd->last_type == LAST_NORM && nd->last.name[nd->last.len])
2257                 nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
2258
2259         nd->flags &= ~LOOKUP_PARENT;
2260         return walk_component(nd, 0);
2261 }
2262
2263 static int handle_lookup_down(struct nameidata *nd)
2264 {
2265         struct path path = nd->path;
2266         struct inode *inode = nd->inode;
2267         unsigned seq = nd->seq;
2268         int err;
2269
2270         if (nd->flags & LOOKUP_RCU) {
2271                 /*
2272                  * don't bother with unlazy_walk on failure - we are
2273                  * at the very beginning of walk, so we lose nothing
2274                  * if we simply redo everything in non-RCU mode
2275                  */
2276                 if (unlikely(!__follow_mount_rcu(nd, &path, &inode, &seq)))
2277                         return -ECHILD;
2278         } else {
2279                 dget(path.dentry);
2280                 err = follow_managed(&path, nd);
2281                 if (unlikely(err < 0))
2282                         return err;
2283                 inode = d_backing_inode(path.dentry);
2284                 seq = 0;
2285         }
2286         path_to_nameidata(&path, nd);
2287         nd->inode = inode;
2288         nd->seq = seq;
2289         return 0;
2290 }
2291
2292 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2293 static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
2294 {
2295         const char *s = path_init(nd, flags);
2296         int err;
2297
2298         if (IS_ERR(s))
2299                 return PTR_ERR(s);
2300
2301         if (unlikely(flags & LOOKUP_DOWN)) {
2302                 err = handle_lookup_down(nd);
2303                 if (unlikely(err < 0)) {
2304                         terminate_walk(nd);
2305                         return err;
2306                 }
2307         }
2308
2309         while (!(err = link_path_walk(s, nd))
2310                 && ((err = lookup_last(nd)) > 0)) {
2311                 s = trailing_symlink(nd);
2312                 if (IS_ERR(s)) {
2313                         err = PTR_ERR(s);
2314                         break;
2315                 }
2316         }
2317         if (!err)
2318                 err = complete_walk(nd);
2319
2320         if (!err && nd->flags & LOOKUP_DIRECTORY)
2321                 if (!d_can_lookup(nd->path.dentry))
2322                         err = -ENOTDIR;
2323         if (!err) {
2324                 *path = nd->path;
2325                 nd->path.mnt = NULL;
2326                 nd->path.dentry = NULL;
2327         }
2328         terminate_walk(nd);
2329         return err;
2330 }
2331
2332 static int filename_lookup(int dfd, struct filename *name, unsigned flags,
2333                            struct path *path, struct path *root)
2334 {
2335         int retval;
2336         struct nameidata nd;
2337         if (IS_ERR(name))
2338                 return PTR_ERR(name);
2339         if (unlikely(root)) {
2340                 nd.root = *root;
2341                 flags |= LOOKUP_ROOT;
2342         }
2343         set_nameidata(&nd, dfd, name);
2344         retval = path_lookupat(&nd, flags | LOOKUP_RCU, path);
2345         if (unlikely(retval == -ECHILD))
2346                 retval = path_lookupat(&nd, flags, path);
2347         if (unlikely(retval == -ESTALE))
2348                 retval = path_lookupat(&nd, flags | LOOKUP_REVAL, path);
2349
2350         if (likely(!retval))
2351                 audit_inode(name, path->dentry, flags & LOOKUP_PARENT);
2352         restore_nameidata();
2353         putname(name);
2354         return retval;
2355 }
2356
2357 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2358 static int path_parentat(struct nameidata *nd, unsigned flags,
2359                                 struct path *parent)
2360 {
2361         const char *s = path_init(nd, flags);
2362         int err;
2363         if (IS_ERR(s))
2364                 return PTR_ERR(s);
2365         err = link_path_walk(s, nd);
2366         if (!err)
2367                 err = complete_walk(nd);
2368         if (!err) {
2369                 *parent = nd->path;
2370                 nd->path.mnt = NULL;
2371                 nd->path.dentry = NULL;
2372         }
2373         terminate_walk(nd);
2374         return err;
2375 }
2376
2377 static struct filename *filename_parentat(int dfd, struct filename *name,
2378                                 unsigned int flags, struct path *parent,
2379                                 struct qstr *last, int *type)
2380 {
2381         int retval;
2382         struct nameidata nd;
2383
2384         if (IS_ERR(name))
2385                 return name;
2386         set_nameidata(&nd, dfd, name);
2387         retval = path_parentat(&nd, flags | LOOKUP_RCU, parent);
2388         if (unlikely(retval == -ECHILD))
2389                 retval = path_parentat(&nd, flags, parent);
2390         if (unlikely(retval == -ESTALE))
2391                 retval = path_parentat(&nd, flags | LOOKUP_REVAL, parent);
2392         if (likely(!retval)) {
2393                 *last = nd.last;
2394                 *type = nd.last_type;
2395                 audit_inode(name, parent->dentry, LOOKUP_PARENT);
2396         } else {
2397                 putname(name);
2398                 name = ERR_PTR(retval);
2399         }
2400         restore_nameidata();
2401         return name;
2402 }
2403
2404 /* does lookup, returns the object with parent locked */
2405 struct dentry *kern_path_locked(const char *name, struct path *path)
2406 {
2407         struct filename *filename;
2408         struct dentry *d;
2409         struct qstr last;
2410         int type;
2411
2412         filename = filename_parentat(AT_FDCWD, getname_kernel(name), 0, path,
2413                                     &last, &type);
2414         if (IS_ERR(filename))
2415                 return ERR_CAST(filename);
2416         if (unlikely(type != LAST_NORM)) {
2417                 path_put(path);
2418                 putname(filename);
2419                 return ERR_PTR(-EINVAL);
2420         }
2421         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
2422         d = __lookup_hash(&last, path->dentry, 0);
2423         if (IS_ERR(d)) {
2424                 inode_unlock(path->dentry->d_inode);
2425                 path_put(path);
2426         }
2427         putname(filename);
2428         return d;
2429 }
2430
2431 int kern_path(const char *name, unsigned int flags, struct path *path)
2432 {
2433         return filename_lookup(AT_FDCWD, getname_kernel(name),
2434                                flags, path, NULL);
2435 }
2436 EXPORT_SYMBOL(kern_path);
2437
2438 /**
2439  * vfs_path_lookup - lookup a file path relative to a dentry-vfsmount pair
2440  * @dentry:  pointer to dentry of the base directory
2441  * @mnt: pointer to vfs mount of the base directory
2442  * @name: pointer to file name
2443  * @flags: lookup flags
2444  * @path: pointer to struct path to fill
2445  */
2446 int vfs_path_lookup(struct dentry *dentry, struct vfsmount *mnt,
2447                     const char *name, unsigned int flags,
2448                     struct path *path)
2449 {
2450         struct path root = {.mnt = mnt, .dentry = dentry};
2451         /* the first argument of filename_lookup() is ignored with root */
2452         return filename_lookup(AT_FDCWD, getname_kernel(name),
2453                                flags , path, &root);
2454 }
2455 EXPORT_SYMBOL(vfs_path_lookup);
2456
2457 /**
2458  * lookup_one_len - filesystem helper to lookup single pathname component
2459  * @name:       pathname component to lookup
2460  * @base:       base directory to lookup from
2461  * @len:        maximum length @len should be interpreted to
2462  *
2463  * Note that this routine is purely a helper for filesystem usage and should
2464  * not be called by generic code.
2465  *
2466  * The caller must hold base->i_mutex.
2467  */
2468 struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
2469 {
2470         struct qstr this;
2471         unsigned int c;
2472         int err;
2473
2474         WARN_ON_ONCE(!inode_is_locked(base->d_inode));
2475
2476         this.name = name;
2477         this.len = len;
2478         this.hash = full_name_hash(base, name, len);
2479         if (!len)
2480                 return ERR_PTR(-EACCES);
2481
2482         if (unlikely(name[0] == '.')) {
2483                 if (len < 2 || (len == 2 && name[1] == '.'))
2484                         return ERR_PTR(-EACCES);
2485         }
2486
2487         while (len--) {
2488                 c = *(const unsigned char *)name++;
2489                 if (c == '/' || c == '\0')
2490                         return ERR_PTR(-EACCES);
2491         }
2492         /*
2493          * See if the low-level filesystem might want
2494          * to use its own hash..
2495          */
2496         if (base->d_flags & DCACHE_OP_HASH) {
2497                 int err = base->d_op->d_hash(base, &this);
2498                 if (err < 0)
2499                         return ERR_PTR(err);
2500         }
2501
2502         err = inode_permission(base->d_inode, MAY_EXEC);
2503         if (err)
2504                 return ERR_PTR(err);
2505
2506         return __lookup_hash(&this, base, 0);
2507 }
2508 EXPORT_SYMBOL(lookup_one_len);
2509
2510 /**
2511  * lookup_one_len_unlocked - filesystem helper to lookup single pathname component
2512  * @name:       pathname component to lookup
2513  * @base:       base directory to lookup from
2514  * @len:        maximum length @len should be interpreted to
2515  *
2516  * Note that this routine is purely a helper for filesystem usage and should
2517  * not be called by generic code.
2518  *
2519  * Unlike lookup_one_len, it should be called without the parent
2520  * i_mutex held, and will take the i_mutex itself if necessary.
2521  */
2522 struct dentry *lookup_one_len_unlocked(const char *name,
2523                                        struct dentry *base, int len)
2524 {
2525         struct qstr this;
2526         unsigned int c;
2527         int err;
2528         struct dentry *ret;
2529
2530         this.name = name;
2531         this.len = len;
2532         this.hash = full_name_hash(base, name, len);
2533         if (!len)
2534                 return ERR_PTR(-EACCES);
2535
2536         if (unlikely(name[0] == '.')) {
2537                 if (len < 2 || (len == 2 && name[1] == '.'))
2538                         return ERR_PTR(-EACCES);
2539         }
2540
2541         while (len--) {
2542                 c = *(const unsigned char *)name++;
2543                 if (c == '/' || c == '\0')
2544                         return ERR_PTR(-EACCES);
2545         }
2546         /*
2547          * See if the low-level filesystem might want
2548          * to use its own hash..
2549          */
2550         if (base->d_flags & DCACHE_OP_HASH) {
2551                 int err = base->d_op->d_hash(base, &this);
2552                 if (err < 0)
2553                         return ERR_PTR(err);
2554         }
2555
2556         err = inode_permission(base->d_inode, MAY_EXEC);
2557         if (err)
2558                 return ERR_PTR(err);
2559
2560         ret = lookup_dcache(&this, base, 0);
2561         if (!ret)
2562                 ret = lookup_slow(&this, base, 0);
2563         return ret;
2564 }
2565 EXPORT_SYMBOL(lookup_one_len_unlocked);
2566
2567 #ifdef CONFIG_UNIX98_PTYS
2568 int path_pts(struct path *path)
2569 {
2570         /* Find something mounted on "pts" in the same directory as
2571          * the input path.
2572          */
2573         struct dentry *child, *parent;
2574         struct qstr this;
2575         int ret;
2576
2577         ret = path_parent_directory(path);
2578         if (ret)
2579                 return ret;
2580
2581         parent = path->dentry;
2582         this.name = "pts";
2583         this.len = 3;
2584         child = d_hash_and_lookup(parent, &this);
2585         if (!child)
2586                 return -ENOENT;
2587
2588         path->dentry = child;
2589         dput(parent);
2590         follow_mount(path);
2591         return 0;
2592 }
2593 #endif
2594
2595 int user_path_at_empty(int dfd, const char __user *name, unsigned flags,
2596                  struct path *path, int *empty)
2597 {
2598         return filename_lookup(dfd, getname_flags(name, flags, empty),
2599                                flags, path, NULL);
2600 }
2601 EXPORT_SYMBOL(user_path_at_empty);
2602
2603 /**
2604  * mountpoint_last - look up last component for umount
2605  * @nd:   pathwalk nameidata - currently pointing at parent directory of "last"
2606  *
2607  * This is a special lookup_last function just for umount. In this case, we
2608  * need to resolve the path without doing any revalidation.
2609  *
2610  * The nameidata should be the result of doing a LOOKUP_PARENT pathwalk. Since
2611  * mountpoints are always pinned in the dcache, their ancestors are too. Thus,
2612  * in almost all cases, this lookup will be served out of the dcache. The only
2613  * cases where it won't are if nd->last refers to a symlink or the path is
2614  * bogus and it doesn't exist.
2615  *
2616  * Returns:
2617  * -error: if there was an error during lookup. This includes -ENOENT if the
2618  *         lookup found a negative dentry.
2619  *
2620  * 0:      if we successfully resolved nd->last and found it to not to be a
2621  *         symlink that needs to be followed.
2622  *
2623  * 1:      if we successfully resolved nd->last and found it to be a symlink
2624  *         that needs to be followed.
2625  */
2626 static int
2627 mountpoint_last(struct nameidata *nd)
2628 {
2629         int error = 0;
2630         struct dentry *dir = nd->path.dentry;
2631         struct path path;
2632
2633         /* If we're in rcuwalk, drop out of it to handle last component */
2634         if (nd->flags & LOOKUP_RCU) {
2635                 if (unlazy_walk(nd))
2636                         return -ECHILD;
2637         }
2638
2639         nd->flags &= ~LOOKUP_PARENT;
2640
2641         if (unlikely(nd->last_type != LAST_NORM)) {
2642                 error = handle_dots(nd, nd->last_type);
2643                 if (error)
2644                         return error;
2645                 path.dentry = dget(nd->path.dentry);
2646         } else {
2647                 path.dentry = d_lookup(dir, &nd->last);
2648                 if (!path.dentry) {
2649                         /*
2650                          * No cached dentry. Mounted dentries are pinned in the
2651                          * cache, so that means that this dentry is probably
2652                          * a symlink or the path doesn't actually point
2653                          * to a mounted dentry.
2654                          */
2655                         path.dentry = lookup_slow(&nd->last, dir,
2656                                              nd->flags | LOOKUP_NO_REVAL);
2657                         if (IS_ERR(path.dentry))
2658                                 return PTR_ERR(path.dentry);
2659                 }
2660         }
2661         if (d_is_negative(path.dentry)) {
2662                 dput(path.dentry);
2663                 return -ENOENT;
2664         }
2665         path.mnt = nd->path.mnt;
2666         return step_into(nd, &path, 0, d_backing_inode(path.dentry), 0);
2667 }
2668
2669 /**
2670  * path_mountpoint - look up a path to be umounted
2671  * @nd:         lookup context
2672  * @flags:      lookup flags
2673  * @path:       pointer to container for result
2674  *
2675  * Look up the given name, but don't attempt to revalidate the last component.
2676  * Returns 0 and "path" will be valid on success; Returns error otherwise.
2677  */
2678 static int
2679 path_mountpoint(struct nameidata *nd, unsigned flags, struct path *path)
2680 {
2681         const char *s = path_init(nd, flags);
2682         int err;
2683         if (IS_ERR(s))
2684                 return PTR_ERR(s);
2685         while (!(err = link_path_walk(s, nd)) &&
2686                 (err = mountpoint_last(nd)) > 0) {
2687                 s = trailing_symlink(nd);
2688                 if (IS_ERR(s)) {
2689                         err = PTR_ERR(s);
2690                         break;
2691                 }
2692         }
2693         if (!err) {
2694                 *path = nd->path;
2695                 nd->path.mnt = NULL;
2696                 nd->path.dentry = NULL;
2697                 follow_mount(path);
2698         }
2699         terminate_walk(nd);
2700         return err;
2701 }
2702
2703 static int
2704 filename_mountpoint(int dfd, struct filename *name, struct path *path,
2705                         unsigned int flags)
2706 {
2707         struct nameidata nd;
2708         int error;
2709         if (IS_ERR(name))
2710                 return PTR_ERR(name);
2711         set_nameidata(&nd, dfd, name);
2712         error = path_mountpoint(&nd, flags | LOOKUP_RCU, path);
2713         if (unlikely(error == -ECHILD))
2714                 error = path_mountpoint(&nd, flags, path);
2715         if (unlikely(error == -ESTALE))
2716                 error = path_mountpoint(&nd, flags | LOOKUP_REVAL, path);
2717         if (likely(!error))
2718                 audit_inode(name, path->dentry, 0);
2719         restore_nameidata();
2720         putname(name);
2721         return error;
2722 }
2723
2724 /**
2725  * user_path_mountpoint_at - lookup a path from userland in order to umount it
2726  * @dfd:        directory file descriptor
2727  * @name:       pathname from userland
2728  * @flags:      lookup flags
2729  * @path:       pointer to container to hold result
2730  *
2731  * A umount is a special case for path walking. We're not actually interested
2732  * in the inode in this situation, and ESTALE errors can be a problem. We
2733  * simply want track down the dentry and vfsmount attached at the mountpoint
2734  * and avoid revalidating the last component.
2735  *
2736  * Returns 0 and populates "path" on success.
2737  */
2738 int
2739 user_path_mountpoint_at(int dfd, const char __user *name, unsigned int flags,
2740                         struct path *path)
2741 {
2742         return filename_mountpoint(dfd, getname(name), path, flags);
2743 }
2744
2745 int
2746 kern_path_mountpoint(int dfd, const char *name, struct path *path,
2747                         unsigned int flags)
2748 {
2749         return filename_mountpoint(dfd, getname_kernel(name), path, flags);
2750 }
2751 EXPORT_SYMBOL(kern_path_mountpoint);
2752
2753 int __check_sticky(struct inode *dir, struct inode *inode)
2754 {
2755         kuid_t fsuid = current_fsuid();
2756
2757         if (uid_eq(inode->i_uid, fsuid))
2758                 return 0;
2759         if (uid_eq(dir->i_uid, fsuid))
2760                 return 0;
2761         return !capable_wrt_inode_uidgid(inode, CAP_FOWNER);
2762 }
2763 EXPORT_SYMBOL(__check_sticky);
2764
2765 /*
2766  *      Check whether we can remove a link victim from directory dir, check
2767  *  whether the type of victim is right.
2768  *  1. We can't do it if dir is read-only (done in permission())
2769  *  2. We should have write and exec permissions on dir
2770  *  3. We can't remove anything from append-only dir
2771  *  4. We can't do anything with immutable dir (done in permission())
2772  *  5. If the sticky bit on dir is set we should either
2773  *      a. be owner of dir, or
2774  *      b. be owner of victim, or
2775  *      c. have CAP_FOWNER capability
2776  *  6. If the victim is append-only or immutable we can't do antyhing with
2777  *     links pointing to it.
2778  *  7. If the victim has an unknown uid or gid we can't change the inode.
2779  *  8. If we were asked to remove a directory and victim isn't one - ENOTDIR.
2780  *  9. If we were asked to remove a non-directory and victim isn't one - EISDIR.
2781  * 10. We can't remove a root or mountpoint.
2782  * 11. We don't allow removal of NFS sillyrenamed files; it's handled by
2783  *     nfs_async_unlink().
2784  */
2785 static int may_delete(struct inode *dir, struct dentry *victim, bool isdir)
2786 {
2787         struct inode *inode = d_backing_inode(victim);
2788         int error;
2789
2790         if (d_is_negative(victim))
2791                 return -ENOENT;
2792         BUG_ON(!inode);
2793
2794         BUG_ON(victim->d_parent->d_inode != dir);
2795         audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);
2796
2797         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
2798         if (error)
2799                 return error;
2800         if (IS_APPEND(dir))
2801                 return -EPERM;
2802
2803         if (check_sticky(dir, inode) || IS_APPEND(inode) ||
2804             IS_IMMUTABLE(inode) || IS_SWAPFILE(inode) || HAS_UNMAPPED_ID(inode))
2805                 return -EPERM;
2806         if (isdir) {
2807                 if (!d_is_dir(victim))
2808                         return -ENOTDIR;
2809                 if (IS_ROOT(victim))
2810                         return -EBUSY;
2811         } else if (d_is_dir(victim))
2812                 return -EISDIR;
2813         if (IS_DEADDIR(dir))
2814                 return -ENOENT;
2815         if (victim->d_flags & DCACHE_NFSFS_RENAMED)
2816                 return -EBUSY;
2817         return 0;
2818 }
2819
2820 /*      Check whether we can create an object with dentry child in directory
2821  *  dir.
2822  *  1. We can't do it if child already exists (open has special treatment for
2823  *     this case, but since we are inlined it's OK)
2824  *  2. We can't do it if dir is read-only (done in permission())
2825  *  3. We can't do it if the fs can't represent the fsuid or fsgid.
2826  *  4. We should have write and exec permissions on dir
2827  *  5. We can't do it if dir is immutable (done in permission())
2828  */
2829 static inline int may_create(struct inode *dir, struct dentry *child)
2830 {
2831         struct user_namespace *s_user_ns;
2832         audit_inode_child(dir, child, AUDIT_TYPE_CHILD_CREATE);
2833         if (child->d_inode)
2834                 return -EEXIST;
2835         if (IS_DEADDIR(dir))
2836                 return -ENOENT;
2837         s_user_ns = dir->i_sb->s_user_ns;
2838         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
2839             !kgid_has_mapping(s_user_ns, current_fsgid()))
2840                 return -EOVERFLOW;
2841         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
2842 }
2843
2844 /*
2845  * p1 and p2 should be directories on the same fs.
2846  */
2847 struct dentry *lock_rename(struct dentry *p1, struct dentry *p2)
2848 {
2849         struct dentry *p;
2850
2851         if (p1 == p2) {
2852                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2853                 return NULL;
2854         }
2855
2856         mutex_lock(&p1->d_sb->s_vfs_rename_mutex);
2857
2858         p = d_ancestor(p2, p1);
2859         if (p) {
2860                 inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
2861                 inode_lock_nested(p1->d_inode, I_MUTEX_CHILD);
2862                 return p;
2863         }
2864
2865         p = d_ancestor(p1, p2);
2866         if (p) {
2867                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2868                 inode_lock_nested(p2->d_inode, I_MUTEX_CHILD);
2869                 return p;
2870         }
2871
2872         inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2873         inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
2874         return NULL;
2875 }
2876 EXPORT_SYMBOL(lock_rename);
2877
2878 void unlock_rename(struct dentry *p1, struct dentry *p2)
2879 {
2880         inode_unlock(p1->d_inode);
2881         if (p1 != p2) {
2882                 inode_unlock(p2->d_inode);
2883                 mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
2884         }
2885 }
2886 EXPORT_SYMBOL(unlock_rename);
2887
2888 int vfs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2889                 bool want_excl)
2890 {
2891         int error = may_create(dir, dentry);
2892         if (error)
2893                 return error;
2894
2895         if (!dir->i_op->create)
2896                 return -EACCES; /* shouldn't it be ENOSYS? */
2897         mode &= S_IALLUGO;
2898         mode |= S_IFREG;
2899         error = security_inode_create(dir, dentry, mode);
2900         if (error)
2901                 return error;
2902         error = dir->i_op->create(dir, dentry, mode, want_excl);
2903         if (!error)
2904                 fsnotify_create(dir, dentry);
2905         return error;
2906 }
2907 EXPORT_SYMBOL(vfs_create);
2908
2909 bool may_open_dev(const struct path *path)
2910 {
2911         return !(path->mnt->mnt_flags & MNT_NODEV) &&
2912                 !(path->mnt->mnt_sb->s_iflags & SB_I_NODEV);
2913 }
2914
2915 static int may_open(const struct path *path, int acc_mode, int flag)
2916 {
2917         struct dentry *dentry = path->dentry;
2918         struct inode *inode = dentry->d_inode;
2919         int error;
2920
2921         if (!inode)
2922                 return -ENOENT;
2923
2924         switch (inode->i_mode & S_IFMT) {
2925         case S_IFLNK:
2926                 return -ELOOP;
2927         case S_IFDIR:
2928                 if (acc_mode & MAY_WRITE)
2929                         return -EISDIR;
2930                 break;
2931         case S_IFBLK:
2932         case S_IFCHR:
2933                 if (!may_open_dev(path))
2934                         return -EACCES;
2935                 /*FALLTHRU*/
2936         case S_IFIFO:
2937         case S_IFSOCK:
2938                 flag &= ~O_TRUNC;
2939                 break;
2940         }
2941
2942         error = inode_permission(inode, MAY_OPEN | acc_mode);
2943         if (error)
2944                 return error;
2945
2946         /*
2947          * An append-only file must be opened in append mode for writing.
2948          */
2949         if (IS_APPEND(inode)) {
2950                 if  ((flag & O_ACCMODE) != O_RDONLY && !(flag & O_APPEND))
2951                         return -EPERM;
2952                 if (flag & O_TRUNC)
2953                         return -EPERM;
2954         }
2955
2956         /* O_NOATIME can only be set by the owner or superuser */
2957         if (flag & O_NOATIME && !inode_owner_or_capable(inode))
2958                 return -EPERM;
2959
2960         return 0;
2961 }
2962
2963 static int handle_truncate(struct file *filp)
2964 {
2965         const struct path *path = &filp->f_path;
2966         struct inode *inode = path->dentry->d_inode;
2967         int error = get_write_access(inode);
2968         if (error)
2969                 return error;
2970         /*
2971          * Refuse to truncate files with mandatory locks held on them.
2972          */
2973         error = locks_verify_locked(filp);
2974         if (!error)
2975                 error = security_path_truncate(path);
2976         if (!error) {
2977                 error = do_truncate(path->dentry, 0,
2978                                     ATTR_MTIME|ATTR_CTIME|ATTR_OPEN,
2979                                     filp);
2980         }
2981         put_write_access(inode);
2982         return error;
2983 }
2984
2985 static inline int open_to_namei_flags(int flag)
2986 {
2987         if ((flag & O_ACCMODE) == 3)
2988                 flag--;
2989         return flag;
2990 }
2991
2992 static int may_o_create(const struct path *dir, struct dentry *dentry, umode_t mode)
2993 {
2994         struct user_namespace *s_user_ns;
2995         int error = security_path_mknod(dir, dentry, mode, 0);
2996         if (error)
2997                 return error;
2998
2999         s_user_ns = dir->dentry->d_sb->s_user_ns;
3000         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
3001             !kgid_has_mapping(s_user_ns, current_fsgid()))
3002                 return -EOVERFLOW;
3003
3004         error = inode_permission(dir->dentry->d_inode, MAY_WRITE | MAY_EXEC);
3005         if (error)
3006                 return error;
3007
3008         return security_inode_create(dir->dentry->d_inode, dentry, mode);
3009 }
3010
3011 /*
3012  * Attempt to atomically look up, create and open a file from a negative
3013  * dentry.
3014  *
3015  * Returns 0 if successful.  The file will have been created and attached to
3016  * @file by the filesystem calling finish_open().
3017  *
3018  * Returns 1 if the file was looked up only or didn't need creating.  The
3019  * caller will need to perform the open themselves.  @path will have been
3020  * updated to point to the new dentry.  This may be negative.
3021  *
3022  * Returns an error code otherwise.
3023  */
3024 static int atomic_open(struct nameidata *nd, struct dentry *dentry,
3025                         struct path *path, struct file *file,
3026                         const struct open_flags *op,
3027                         int open_flag, umode_t mode,
3028                         int *opened)
3029 {
3030         struct dentry *const DENTRY_NOT_SET = (void *) -1UL;
3031         struct inode *dir =  nd->path.dentry->d_inode;
3032         int error;
3033
3034         if (!(~open_flag & (O_EXCL | O_CREAT))) /* both O_EXCL and O_CREAT */
3035                 open_flag &= ~O_TRUNC;
3036
3037         if (nd->flags & LOOKUP_DIRECTORY)
3038                 open_flag |= O_DIRECTORY;
3039
3040         file->f_path.dentry = DENTRY_NOT_SET;
3041         file->f_path.mnt = nd->path.mnt;
3042         error = dir->i_op->atomic_open(dir, dentry, file,
3043                                        open_to_namei_flags(open_flag),
3044                                        mode, opened);
3045         d_lookup_done(dentry);
3046         if (!error) {
3047                 /*
3048                  * We didn't have the inode before the open, so check open
3049                  * permission here.
3050                  */
3051                 int acc_mode = op->acc_mode;
3052                 if (*opened & FILE_CREATED) {
3053                         WARN_ON(!(open_flag & O_CREAT));
3054                         fsnotify_create(dir, dentry);
3055                         acc_mode = 0;
3056                 }
3057                 error = may_open(&file->f_path, acc_mode, open_flag);
3058                 if (WARN_ON(error > 0))
3059                         error = -EINVAL;
3060         } else if (error > 0) {
3061                 if (WARN_ON(file->f_path.dentry == DENTRY_NOT_SET)) {
3062                         error = -EIO;
3063                 } else {
3064                         if (file->f_path.dentry) {
3065                                 dput(dentry);
3066                                 dentry = file->f_path.dentry;
3067                         }
3068                         if (*opened & FILE_CREATED)
3069                                 fsnotify_create(dir, dentry);
3070                         if (unlikely(d_is_negative(dentry))) {
3071                                 error = -ENOENT;
3072                         } else {
3073                                 path->dentry = dentry;
3074                                 path->mnt = nd->path.mnt;
3075                                 return 1;
3076                         }
3077                 }
3078         }
3079         dput(dentry);
3080         return error;
3081 }
3082
3083 /*
3084  * Look up and maybe create and open the last component.
3085  *
3086  * Must be called with i_mutex held on parent.
3087  *
3088  * Returns 0 if the file was successfully atomically created (if necessary) and
3089  * opened.  In this case the file will be returned attached to @file.
3090  *
3091  * Returns 1 if the file was not completely opened at this time, though lookups
3092  * and creations will have been performed and the dentry returned in @path will
3093  * be positive upon return if O_CREAT was specified.  If O_CREAT wasn't
3094  * specified then a negative dentry may be returned.
3095  *
3096  * An error code is returned otherwise.
3097  *
3098  * FILE_CREATE will be set in @*opened if the dentry was created and will be
3099  * cleared otherwise prior to returning.
3100  */
3101 static int lookup_open(struct nameidata *nd, struct path *path,
3102                         struct file *file,
3103                         const struct open_flags *op,
3104                         bool got_write, int *opened)
3105 {
3106         struct dentry *dir = nd->path.dentry;
3107         struct inode *dir_inode = dir->d_inode;
3108         int open_flag = op->open_flag;
3109         struct dentry *dentry;
3110         int error, create_error = 0;
3111         umode_t mode = op->mode;
3112         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
3113
3114         if (unlikely(IS_DEADDIR(dir_inode)))
3115                 return -ENOENT;
3116
3117         *opened &= ~FILE_CREATED;
3118         dentry = d_lookup(dir, &nd->last);
3119         for (;;) {
3120                 if (!dentry) {
3121                         dentry = d_alloc_parallel(dir, &nd->last, &wq);
3122                         if (IS_ERR(dentry))
3123                                 return PTR_ERR(dentry);
3124                 }
3125                 if (d_in_lookup(dentry))
3126                         break;
3127
3128                 error = d_revalidate(dentry, nd->flags);
3129                 if (likely(error > 0))
3130                         break;
3131                 if (error)
3132                         goto out_dput;
3133                 d_invalidate(dentry);
3134                 dput(dentry);
3135                 dentry = NULL;
3136         }
3137         if (dentry->d_inode) {
3138                 /* Cached positive dentry: will open in f_op->open */
3139                 goto out_no_open;
3140         }
3141
3142         /*
3143          * Checking write permission is tricky, bacuse we don't know if we are
3144          * going to actually need it: O_CREAT opens should work as long as the
3145          * file exists.  But checking existence breaks atomicity.  The trick is
3146          * to check access and if not granted clear O_CREAT from the flags.
3147          *
3148          * Another problem is returing the "right" error value (e.g. for an
3149          * O_EXCL open we want to return EEXIST not EROFS).
3150          */
3151         if (open_flag & O_CREAT) {
3152                 if (!IS_POSIXACL(dir->d_inode))
3153                         mode &= ~current_umask();
3154                 if (unlikely(!got_write)) {
3155                         create_error = -EROFS;
3156                         open_flag &= ~O_CREAT;
3157                         if (open_flag & (O_EXCL | O_TRUNC))
3158                                 goto no_open;
3159                         /* No side effects, safe to clear O_CREAT */
3160                 } else {
3161                         create_error = may_o_create(&nd->path, dentry, mode);
3162                         if (create_error) {
3163                                 open_flag &= ~O_CREAT;
3164                                 if (open_flag & O_EXCL)
3165                                         goto no_open;
3166                         }
3167                 }
3168         } else if ((open_flag & (O_TRUNC|O_WRONLY|O_RDWR)) &&
3169                    unlikely(!got_write)) {
3170                 /*
3171                  * No O_CREATE -> atomicity not a requirement -> fall
3172                  * back to lookup + open
3173                  */
3174                 goto no_open;
3175         }
3176
3177         if (dir_inode->i_op->atomic_open) {
3178                 error = atomic_open(nd, dentry, path, file, op, open_flag,
3179                                     mode, opened);
3180                 if (unlikely(error == -ENOENT) && create_error)
3181                         error = create_error;
3182                 return error;
3183         }
3184
3185 no_open:
3186         if (d_in_lookup(dentry)) {
3187                 struct dentry *res = dir_inode->i_op->lookup(dir_inode, dentry,
3188                                                              nd->flags);
3189                 d_lookup_done(dentry);
3190                 if (unlikely(res)) {
3191                         if (IS_ERR(res)) {
3192                                 error = PTR_ERR(res);
3193                                 goto out_dput;
3194                         }
3195                         dput(dentry);
3196                         dentry = res;
3197                 }
3198         }
3199
3200         /* Negative dentry, just create the file */
3201         if (!dentry->d_inode && (open_flag & O_CREAT)) {
3202                 *opened |= FILE_CREATED;
3203                 audit_inode_child(dir_inode, dentry, AUDIT_TYPE_CHILD_CREATE);
3204                 if (!dir_inode->i_op->create) {
3205                         error = -EACCES;
3206                         goto out_dput;
3207                 }
3208                 error = dir_inode->i_op->create(dir_inode, dentry, mode,
3209                                                 open_flag & O_EXCL);
3210                 if (error)
3211                         goto out_dput;
3212                 fsnotify_create(dir_inode, dentry);
3213         }
3214         if (unlikely(create_error) && !dentry->d_inode) {
3215                 error = create_error;
3216                 goto out_dput;
3217         }
3218 out_no_open:
3219         path->dentry = dentry;
3220         path->mnt = nd->path.mnt;
3221         return 1;
3222
3223 out_dput:
3224         dput(dentry);
3225         return error;
3226 }
3227
3228 /*
3229  * Handle the last step of open()
3230  */
3231 static int do_last(struct nameidata *nd,
3232                    struct file *file, const struct open_flags *op,
3233                    int *opened)
3234 {
3235         struct dentry *dir = nd->path.dentry;
3236         int open_flag = op->open_flag;
3237         bool will_truncate = (open_flag & O_TRUNC) != 0;
3238         bool got_write = false;
3239         int acc_mode = op->acc_mode;
3240         unsigned seq;
3241         struct inode *inode;
3242         struct path path;
3243         int error;
3244
3245         nd->flags &= ~LOOKUP_PARENT;
3246         nd->flags |= op->intent;
3247
3248         if (nd->last_type != LAST_NORM) {
3249                 error = handle_dots(nd, nd->last_type);
3250                 if (unlikely(error))
3251                         return error;
3252                 goto finish_open;
3253         }
3254
3255         if (!(open_flag & O_CREAT)) {
3256                 if (nd->last.name[nd->last.len])
3257                         nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
3258                 /* we _can_ be in RCU mode here */
3259                 error = lookup_fast(nd, &path, &inode, &seq);
3260                 if (likely(error > 0))
3261                         goto finish_lookup;
3262
3263                 if (error < 0)
3264                         return error;
3265
3266                 BUG_ON(nd->inode != dir->d_inode);
3267                 BUG_ON(nd->flags & LOOKUP_RCU);
3268         } else {
3269                 /* create side of things */
3270                 /*
3271                  * This will *only* deal with leaving RCU mode - LOOKUP_JUMPED
3272                  * has been cleared when we got to the last component we are
3273                  * about to look up
3274                  */
3275                 error = complete_walk(nd);
3276                 if (error)
3277                         return error;
3278
3279                 audit_inode(nd->name, dir, LOOKUP_PARENT);
3280                 /* trailing slashes? */
3281                 if (unlikely(nd->last.name[nd->last.len]))
3282                         return -EISDIR;
3283         }
3284
3285         if (open_flag & (O_CREAT | O_TRUNC | O_WRONLY | O_RDWR)) {
3286                 error = mnt_want_write(nd->path.mnt);
3287                 if (!error)
3288                         got_write = true;
3289                 /*
3290                  * do _not_ fail yet - we might not need that or fail with
3291                  * a different error; let lookup_open() decide; we'll be
3292                  * dropping this one anyway.
3293                  */
3294         }
3295         if (open_flag & O_CREAT)
3296                 inode_lock(dir->d_inode);
3297         else
3298                 inode_lock_shared(dir->d_inode);
3299         error = lookup_open(nd, &path, file, op, got_write, opened);
3300         if (open_flag & O_CREAT)
3301                 inode_unlock(dir->d_inode);
3302         else
3303                 inode_unlock_shared(dir->d_inode);
3304
3305         if (error <= 0) {
3306                 if (error)
3307                         goto out;
3308
3309                 if ((*opened & FILE_CREATED) ||
3310                     !S_ISREG(file_inode(file)->i_mode))
3311                         will_truncate = false;
3312
3313                 audit_inode(nd->name, file->f_path.dentry, 0);
3314                 goto opened;
3315         }
3316
3317         if (*opened & FILE_CREATED) {
3318                 /* Don't check for write permission, don't truncate */
3319                 open_flag &= ~O_TRUNC;
3320                 will_truncate = false;
3321                 acc_mode = 0;
3322                 path_to_nameidata(&path, nd);
3323                 goto finish_open_created;
3324         }
3325
3326         /*
3327          * If atomic_open() acquired write access it is dropped now due to
3328          * possible mount and symlink following (this might be optimized away if
3329          * necessary...)
3330          */
3331         if (got_write) {
3332                 mnt_drop_write(nd->path.mnt);
3333                 got_write = false;
3334         }
3335
3336         error = follow_managed(&path, nd);
3337         if (unlikely(error < 0))
3338                 return error;
3339
3340         if (unlikely(d_is_negative(path.dentry))) {
3341                 path_to_nameidata(&path, nd);
3342                 return -ENOENT;
3343         }
3344
3345         /*
3346          * create/update audit record if it already exists.
3347          */
3348         audit_inode(nd->name, path.dentry, 0);
3349
3350         if (unlikely((open_flag & (O_EXCL | O_CREAT)) == (O_EXCL | O_CREAT))) {
3351                 path_to_nameidata(&path, nd);
3352                 return -EEXIST;
3353         }
3354
3355         seq = 0;        /* out of RCU mode, so the value doesn't matter */
3356         inode = d_backing_inode(path.dentry);
3357 finish_lookup:
3358         error = step_into(nd, &path, 0, inode, seq);
3359         if (unlikely(error))
3360                 return error;
3361 finish_open:
3362         /* Why this, you ask?  _Now_ we might have grown LOOKUP_JUMPED... */
3363         error = complete_walk(nd);
3364         if (error)
3365                 return error;
3366         audit_inode(nd->name, nd->path.dentry, 0);
3367         error = -EISDIR;
3368         if ((open_flag & O_CREAT) && d_is_dir(nd->path.dentry))
3369                 goto out;
3370         error = -ENOTDIR;
3371         if ((nd->flags & LOOKUP_DIRECTORY) && !d_can_lookup(nd->path.dentry))
3372                 goto out;
3373         if (!d_is_reg(nd->path.dentry))
3374                 will_truncate = false;
3375
3376         if (will_truncate) {
3377                 error = mnt_want_write(nd->path.mnt);
3378                 if (error)
3379                         goto out;
3380                 got_write = true;
3381         }
3382 finish_open_created:
3383         error = may_open(&nd->path, acc_mode, open_flag);
3384         if (error)
3385                 goto out;
3386         BUG_ON(*opened & FILE_OPENED); /* once it's opened, it's opened */
3387         error = vfs_open(&nd->path, file, current_cred());
3388         if (error)
3389                 goto out;
3390         *opened |= FILE_OPENED;
3391 opened:
3392         error = open_check_o_direct(file);
3393         if (!error)
3394                 error = ima_file_check(file, op->acc_mode, *opened);
3395         if (!error && will_truncate)
3396                 error = handle_truncate(file);
3397 out:
3398         if (unlikely(error) && (*opened & FILE_OPENED))
3399                 fput(file);
3400         if (unlikely(error > 0)) {
3401                 WARN_ON(1);
3402                 error = -EINVAL;
3403         }
3404         if (got_write)
3405                 mnt_drop_write(nd->path.mnt);
3406         return error;
3407 }
3408
3409 struct dentry *vfs_tmpfile(struct dentry *dentry, umode_t mode, int open_flag)
3410 {
3411         struct dentry *child = NULL;
3412         struct inode *dir = dentry->d_inode;
3413         struct inode *inode;
3414         int error;
3415
3416         /* we want directory to be writable */
3417         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
3418         if (error)
3419                 goto out_err;
3420         error = -EOPNOTSUPP;
3421         if (!dir->i_op->tmpfile)
3422                 goto out_err;
3423         error = -ENOMEM;
3424         child = d_alloc(dentry, &slash_name);
3425         if (unlikely(!child))
3426                 goto out_err;
3427         error = dir->i_op->tmpfile(dir, child, mode);
3428         if (error)
3429                 goto out_err;
3430         error = -ENOENT;
3431         inode = child->d_inode;
3432         if (unlikely(!inode))
3433                 goto out_err;
3434         if (!(open_flag & O_EXCL)) {
3435                 spin_lock(&inode->i_lock);
3436                 inode->i_state |= I_LINKABLE;
3437                 spin_unlock(&inode->i_lock);
3438         }
3439         return child;
3440
3441 out_err:
3442         dput(child);
3443         return ERR_PTR(error);
3444 }
3445 EXPORT_SYMBOL(vfs_tmpfile);
3446
3447 static int do_tmpfile(struct nameidata *nd, unsigned flags,
3448                 const struct open_flags *op,
3449                 struct file *file, int *opened)
3450 {
3451         struct dentry *child;
3452         struct path path;
3453         int error = path_lookupat(nd, flags | LOOKUP_DIRECTORY, &path);
3454         if (unlikely(error))
3455                 return error;
3456         error = mnt_want_write(path.mnt);
3457         if (unlikely(error))
3458                 goto out;
3459         child = vfs_tmpfile(path.dentry, op->mode, op->open_flag);
3460         error = PTR_ERR(child);
3461         if (unlikely(IS_ERR(child)))
3462                 goto out2;
3463         dput(path.dentry);
3464         path.dentry = child;
3465         audit_inode(nd->name, child, 0);
3466         /* Don't check for other permissions, the inode was just created */
3467         error = may_open(&path, 0, op->open_flag);
3468         if (error)
3469                 goto out2;
3470         file->f_path.mnt = path.mnt;
3471         error = finish_open(file, child, NULL, opened);
3472         if (error)
3473                 goto out2;
3474         error = open_check_o_direct(file);
3475         if (error)
3476                 fput(file);
3477 out2:
3478         mnt_drop_write(path.mnt);
3479 out:
3480         path_put(&path);
3481         return error;
3482 }
3483
3484 static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
3485 {
3486         struct path path;
3487         int error = path_lookupat(nd, flags, &path);
3488         if (!error) {
3489                 audit_inode(nd->name, path.dentry, 0);
3490                 error = vfs_open(&path, file, current_cred());
3491                 path_put(&path);
3492         }
3493         return error;
3494 }
3495
3496 static struct file *path_openat(struct nameidata *nd,
3497                         const struct open_flags *op, unsigned flags)
3498 {
3499         const char *s;
3500         struct file *file;
3501         int opened = 0;
3502         int error;
3503
3504         file = get_empty_filp();
3505         if (IS_ERR(file))
3506                 return file;
3507
3508         file->f_flags = op->open_flag;
3509
3510         if (unlikely(file->f_flags & __O_TMPFILE)) {
3511                 error = do_tmpfile(nd, flags, op, file, &opened);
3512                 goto out2;
3513         }
3514
3515         if (unlikely(file->f_flags & O_PATH)) {
3516                 error = do_o_path(nd, flags, file);
3517                 if (!error)
3518                         opened |= FILE_OPENED;
3519                 goto out2;
3520         }
3521
3522         s = path_init(nd, flags);
3523         if (IS_ERR(s)) {
3524                 put_filp(file);
3525                 return ERR_CAST(s);
3526         }
3527         while (!(error = link_path_walk(s, nd)) &&
3528                 (error = do_last(nd, file, op, &opened)) > 0) {
3529                 nd->flags &= ~(LOOKUP_OPEN|LOOKUP_CREATE|LOOKUP_EXCL);
3530                 s = trailing_symlink(nd);
3531                 if (IS_ERR(s)) {
3532                         error = PTR_ERR(s);
3533                         break;
3534                 }
3535         }
3536         terminate_walk(nd);
3537 out2:
3538         if (!(opened & FILE_OPENED)) {
3539                 BUG_ON(!error);
3540                 put_filp(file);
3541         }
3542         if (unlikely(error)) {
3543                 if (error == -EOPENSTALE) {
3544                         if (flags & LOOKUP_RCU)
3545                                 error = -ECHILD;
3546                         else
3547                                 error = -ESTALE;
3548                 }
3549                 file = ERR_PTR(error);
3550         }
3551         return file;
3552 }
3553
3554 struct file *do_filp_open(int dfd, struct filename *pathname,
3555                 const struct open_flags *op)
3556 {
3557         struct nameidata nd;
3558         int flags = op->lookup_flags;
3559         struct file *filp;
3560
3561         set_nameidata(&nd, dfd, pathname);
3562         filp = path_openat(&nd, op, flags | LOOKUP_RCU);
3563         if (unlikely(filp == ERR_PTR(-ECHILD)))
3564                 filp = path_openat(&nd, op, flags);
3565         if (unlikely(filp == ERR_PTR(-ESTALE)))
3566                 filp = path_openat(&nd, op, flags | LOOKUP_REVAL);
3567         restore_nameidata();
3568         return filp;
3569 }
3570
3571 struct file *do_file_open_root(struct dentry *dentry, struct vfsmount *mnt,
3572                 const char *name, const struct open_flags *op)
3573 {
3574         struct nameidata nd;
3575         struct file *file;
3576         struct filename *filename;
3577         int flags = op->lookup_flags | LOOKUP_ROOT;
3578
3579         nd.root.mnt = mnt;
3580         nd.root.dentry = dentry;
3581
3582         if (d_is_symlink(dentry) && op->intent & LOOKUP_OPEN)
3583                 return ERR_PTR(-ELOOP);
3584
3585         filename = getname_kernel(name);
3586         if (IS_ERR(filename))
3587                 return ERR_CAST(filename);
3588
3589         set_nameidata(&nd, -1, filename);
3590         file = path_openat(&nd, op, flags | LOOKUP_RCU);
3591         if (unlikely(file == ERR_PTR(-ECHILD)))
3592                 file = path_openat(&nd, op, flags);
3593         if (unlikely(file == ERR_PTR(-ESTALE)))
3594                 file = path_openat(&nd, op, flags | LOOKUP_REVAL);
3595         restore_nameidata();
3596         putname(filename);
3597         return file;
3598 }
3599
3600 static struct dentry *filename_create(int dfd, struct filename *name,
3601                                 struct path *path, unsigned int lookup_flags)
3602 {
3603         struct dentry *dentry = ERR_PTR(-EEXIST);
3604         struct qstr last;
3605         int type;
3606         int err2;
3607         int error;
3608         bool is_dir = (lookup_flags & LOOKUP_DIRECTORY);
3609
3610         /*
3611          * Note that only LOOKUP_REVAL and LOOKUP_DIRECTORY matter here. Any
3612          * other flags passed in are ignored!
3613          */
3614         lookup_flags &= LOOKUP_REVAL;
3615
3616         name = filename_parentat(dfd, name, lookup_flags, path, &last, &type);
3617         if (IS_ERR(name))
3618                 return ERR_CAST(name);
3619
3620         /*
3621          * Yucky last component or no last component at all?
3622          * (foo/., foo/.., /////)
3623          */
3624         if (unlikely(type != LAST_NORM))
3625                 goto out;
3626
3627         /* don't fail immediately if it's r/o, at least try to report other errors */
3628         err2 = mnt_want_write(path->mnt);
3629         /*
3630          * Do the final lookup.
3631          */
3632         lookup_flags |= LOOKUP_CREATE | LOOKUP_EXCL;
3633         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
3634         dentry = __lookup_hash(&last, path->dentry, lookup_flags);
3635         if (IS_ERR(dentry))
3636                 goto unlock;
3637
3638         error = -EEXIST;
3639         if (d_is_positive(dentry))
3640                 goto fail;
3641
3642         /*
3643          * Special case - lookup gave negative, but... we had foo/bar/
3644          * From the vfs_mknod() POV we just have a negative dentry -
3645          * all is fine. Let's be bastards - you had / on the end, you've
3646          * been asking for (non-existent) directory. -ENOENT for you.
3647          */
3648         if (unlikely(!is_dir && last.name[last.len])) {
3649                 error = -ENOENT;
3650                 goto fail;
3651         }
3652         if (unlikely(err2)) {
3653                 error = err2;
3654                 goto fail;
3655         }
3656         putname(name);
3657         return dentry;
3658 fail:
3659         dput(dentry);
3660         dentry = ERR_PTR(error);
3661 unlock:
3662         inode_unlock(path->dentry->d_inode);
3663         if (!err2)
3664                 mnt_drop_write(path->mnt);
3665 out:
3666         path_put(path);
3667         putname(name);
3668         return dentry;
3669 }
3670
3671 struct dentry *kern_path_create(int dfd, const char *pathname,
3672                                 struct path *path, unsigned int lookup_flags)
3673 {
3674         return filename_create(dfd, getname_kernel(pathname),
3675                                 path, lookup_flags);
3676 }
3677 EXPORT_SYMBOL(kern_path_create);
3678
3679 void done_path_create(struct path *path, struct dentry *dentry)
3680 {
3681         dput(dentry);
3682         inode_unlock(path->dentry->d_inode);
3683         mnt_drop_write(path->mnt);
3684         path_put(path);
3685 }
3686 EXPORT_SYMBOL(done_path_create);
3687
3688 inline struct dentry *user_path_create(int dfd, const char __user *pathname,
3689                                 struct path *path, unsigned int lookup_flags)
3690 {
3691         return filename_create(dfd, getname(pathname), path, lookup_flags);
3692 }
3693 EXPORT_SYMBOL(user_path_create);
3694
3695 int vfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3696 {
3697         int error = may_create(dir, dentry);
3698
3699         if (error)
3700                 return error;
3701
3702         if ((S_ISCHR(mode) || S_ISBLK(mode)) && !capable(CAP_MKNOD))
3703                 return -EPERM;
3704
3705         if (!dir->i_op->mknod)
3706                 return -EPERM;
3707
3708         error = devcgroup_inode_mknod(mode, dev);
3709         if (error)
3710                 return error;
3711
3712         error = security_inode_mknod(dir, dentry, mode, dev);
3713         if (error)
3714                 return error;
3715
3716         error = dir->i_op->mknod(dir, dentry, mode, dev);
3717         if (!error)
3718                 fsnotify_create(dir, dentry);
3719         return error;
3720 }
3721 EXPORT_SYMBOL(vfs_mknod);
3722
3723 static int may_mknod(umode_t mode)
3724 {
3725         switch (mode & S_IFMT) {
3726         case S_IFREG:
3727         case S_IFCHR:
3728         case S_IFBLK:
3729         case S_IFIFO:
3730         case S_IFSOCK:
3731         case 0: /* zero mode translates to S_IFREG */
3732                 return 0;
3733         case S_IFDIR:
3734                 return -EPERM;
3735         default:
3736                 return -EINVAL;
3737         }
3738 }
3739
3740 SYSCALL_DEFINE4(mknodat, int, dfd, const char __user *, filename, umode_t, mode,
3741                 unsigned, dev)
3742 {
3743         struct dentry *dentry;
3744         struct path path;
3745         int error;
3746         unsigned int lookup_flags = 0;
3747
3748         error = may_mknod(mode);
3749         if (error)
3750                 return error;
3751 retry:
3752         dentry = user_path_create(dfd, filename, &path, lookup_flags);
3753         if (IS_ERR(dentry))
3754                 return PTR_ERR(dentry);
3755
3756         if (!IS_POSIXACL(path.dentry->d_inode))
3757                 mode &= ~current_umask();
3758         error = security_path_mknod(&path, dentry, mode, dev);
3759         if (error)
3760                 goto out;
3761         switch (mode & S_IFMT) {
3762                 case 0: case S_IFREG:
3763                         error = vfs_create(path.dentry->d_inode,dentry,mode,true);
3764                         if (!error)
3765                                 ima_post_path_mknod(dentry);
3766                         break;
3767                 case S_IFCHR: case S_IFBLK:
3768                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,
3769                                         new_decode_dev(dev));
3770                         break;
3771                 case S_IFIFO: case S_IFSOCK:
3772                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,0);
3773                         break;
3774         }
3775 out:
3776         done_path_create(&path, dentry);
3777         if (retry_estale(error, lookup_flags)) {
3778                 lookup_flags |= LOOKUP_REVAL;
3779                 goto retry;
3780         }
3781         return error;
3782 }
3783
3784 SYSCALL_DEFINE3(mknod, const char __user *, filename, umode_t, mode, unsigned, dev)
3785 {
3786         return sys_mknodat(AT_FDCWD, filename, mode, dev);
3787 }
3788
3789 int vfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3790 {
3791         int error = may_create(dir, dentry);
3792         unsigned max_links = dir->i_sb->s_max_links;
3793
3794         if (error)
3795                 return error;
3796
3797         if (!dir->i_op->mkdir)
3798                 return -EPERM;
3799
3800         mode &= (S_IRWXUGO|S_ISVTX);
3801         error = security_inode_mkdir(dir, dentry, mode);
3802         if (error)
3803                 return error;
3804
3805         if (max_links && dir->i_nlink >= max_links)
3806                 return -EMLINK;
3807
3808         error = dir->i_op->mkdir(dir, dentry, mode);
3809         if (!error)
3810                 fsnotify_mkdir(dir, dentry);
3811         return error;
3812 }
3813 EXPORT_SYMBOL(vfs_mkdir);
3814
3815 SYSCALL_DEFINE3(mkdirat, int, dfd, const char __user *, pathname, umode_t, mode)
3816 {
3817         struct dentry *dentry;
3818         struct path path;
3819         int error;
3820         unsigned int lookup_flags = LOOKUP_DIRECTORY;
3821
3822 retry:
3823         dentry = user_path_create(dfd, pathname, &path, lookup_flags);
3824         if (IS_ERR(dentry))
3825                 return PTR_ERR(dentry);
3826
3827         if (!IS_POSIXACL(path.dentry->d_inode))
3828                 mode &= ~current_umask();
3829         error = security_path_mkdir(&path, dentry, mode);
3830         if (!error)
3831                 error = vfs_mkdir(path.dentry->d_inode, dentry, mode);
3832         done_path_create(&path, dentry);
3833         if (retry_estale(error, lookup_flags)) {
3834                 lookup_flags |= LOOKUP_REVAL;
3835                 goto retry;
3836         }
3837         return error;
3838 }
3839
3840 SYSCALL_DEFINE2(mkdir, const char __user *, pathname, umode_t, mode)
3841 {
3842         return sys_mkdirat(AT_FDCWD, pathname, mode);
3843 }
3844
3845 int vfs_rmdir(struct inode *dir, struct dentry *dentry)
3846 {
3847         int error = may_delete(dir, dentry, 1);
3848
3849         if (error)
3850                 return error;
3851
3852         if (!dir->i_op->rmdir)
3853                 return -EPERM;
3854
3855         dget(dentry);
3856         inode_lock(dentry->d_inode);
3857
3858         error = -EBUSY;
3859         if (is_local_mountpoint(dentry))
3860                 goto out;
3861
3862         error = security_inode_rmdir(dir, dentry);
3863         if (error)
3864                 goto out;
3865
3866         shrink_dcache_parent(dentry);
3867         error = dir->i_op->rmdir(dir, dentry);
3868         if (error)
3869                 goto out;
3870
3871         dentry->d_inode->i_flags |= S_DEAD;
3872         dont_mount(dentry);
3873         detach_mounts(dentry);
3874
3875 out:
3876         inode_unlock(dentry->d_inode);
3877         dput(dentry);
3878         if (!error)
3879                 d_delete(dentry);
3880         return error;
3881 }
3882 EXPORT_SYMBOL(vfs_rmdir);
3883
3884 static long do_rmdir(int dfd, const char __user *pathname)
3885 {
3886         int error = 0;
3887         struct filename *name;
3888         struct dentry *dentry;
3889         struct path path;
3890         struct qstr last;
3891         int type;
3892         unsigned int lookup_flags = 0;
3893 retry:
3894         name = filename_parentat(dfd, getname(pathname), lookup_flags,
3895                                 &path, &last, &type);
3896         if (IS_ERR(name))
3897                 return PTR_ERR(name);
3898
3899         switch (type) {
3900         case LAST_DOTDOT:
3901                 error = -ENOTEMPTY;
3902                 goto exit1;
3903         case LAST_DOT:
3904                 error = -EINVAL;
3905                 goto exit1;
3906         case LAST_ROOT:
3907                 error = -EBUSY;
3908                 goto exit1;
3909         }
3910
3911         error = mnt_want_write(path.mnt);
3912         if (error)
3913                 goto exit1;
3914
3915         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3916         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3917         error = PTR_ERR(dentry);
3918         if (IS_ERR(dentry))
3919                 goto exit2;
3920         if (!dentry->d_inode) {
3921                 error = -ENOENT;
3922                 goto exit3;
3923         }
3924         error = security_path_rmdir(&path, dentry);
3925         if (error)
3926                 goto exit3;
3927         error = vfs_rmdir(path.dentry->d_inode, dentry);
3928 exit3:
3929         dput(dentry);
3930 exit2:
3931         inode_unlock(path.dentry->d_inode);
3932         mnt_drop_write(path.mnt);
3933 exit1:
3934         path_put(&path);
3935         putname(name);
3936         if (retry_estale(error, lookup_flags)) {
3937                 lookup_flags |= LOOKUP_REVAL;
3938                 goto retry;
3939         }
3940         return error;
3941 }
3942
3943 SYSCALL_DEFINE1(rmdir, const char __user *, pathname)
3944 {
3945         return do_rmdir(AT_FDCWD, pathname);
3946 }
3947
3948 /**
3949  * vfs_unlink - unlink a filesystem object
3950  * @dir:        parent directory
3951  * @dentry:     victim
3952  * @delegated_inode: returns victim inode, if the inode is delegated.
3953  *
3954  * The caller must hold dir->i_mutex.
3955  *
3956  * If vfs_unlink discovers a delegation, it will return -EWOULDBLOCK and
3957  * return a reference to the inode in delegated_inode.  The caller
3958  * should then break the delegation on that inode and retry.  Because
3959  * breaking a delegation may take a long time, the caller should drop
3960  * dir->i_mutex before doing so.
3961  *
3962  * Alternatively, a caller may pass NULL for delegated_inode.  This may
3963  * be appropriate for callers that expect the underlying filesystem not
3964  * to be NFS exported.
3965  */
3966 int vfs_unlink(struct inode *dir, struct dentry *dentry, struct inode **delegated_inode)
3967 {
3968         struct inode *target = dentry->d_inode;
3969         int error = may_delete(dir, dentry, 0);
3970
3971         if (error)
3972                 return error;
3973
3974         if (!dir->i_op->unlink)
3975                 return -EPERM;
3976
3977         inode_lock(target);
3978         if (is_local_mountpoint(dentry))
3979                 error = -EBUSY;
3980         else {
3981                 error = security_inode_unlink(dir, dentry);
3982                 if (!error) {
3983                         error = try_break_deleg(target, delegated_inode);
3984                         if (error)
3985                                 goto out;
3986                         error = dir->i_op->unlink(dir, dentry);
3987                         if (!error) {
3988                                 dont_mount(dentry);
3989                                 detach_mounts(dentry);
3990                         }
3991                 }
3992         }
3993 out:
3994         inode_unlock(target);
3995
3996         /* We don't d_delete() NFS sillyrenamed files--they still exist. */
3997         if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
3998                 fsnotify_link_count(target);
3999                 d_delete(dentry);
4000         }
4001
4002         return error;
4003 }
4004 EXPORT_SYMBOL(vfs_unlink);
4005
4006 /*
4007  * Make sure that the actual truncation of the file will occur outside its
4008  * directory's i_mutex.  Truncate can take a long time if there is a lot of
4009  * writeout happening, and we don't want to prevent access to the directory
4010  * while waiting on the I/O.
4011  */
4012 static long do_unlinkat(int dfd, const char __user *pathname)
4013 {
4014         int error;
4015         struct filename *name;
4016         struct dentry *dentry;
4017         struct path path;
4018         struct qstr last;
4019         int type;
4020         struct inode *inode = NULL;
4021         struct inode *delegated_inode = NULL;
4022         unsigned int lookup_flags = 0;
4023 retry:
4024         name = filename_parentat(dfd, getname(pathname), lookup_flags,
4025                                 &path, &last, &type);
4026         if (IS_ERR(name))
4027                 return PTR_ERR(name);
4028
4029         error = -EISDIR;
4030         if (type != LAST_NORM)
4031                 goto exit1;
4032
4033         error = mnt_want_write(path.mnt);
4034         if (error)
4035                 goto exit1;
4036 retry_deleg:
4037         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
4038         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
4039         error = PTR_ERR(dentry);
4040         if (!IS_ERR(dentry)) {
4041                 /* Why not before? Because we want correct error value */
4042                 if (last.name[last.len])
4043                         goto slashes;
4044                 inode = dentry->d_inode;
4045                 if (d_is_negative(dentry))
4046                         goto slashes;
4047                 ihold(inode);
4048                 error = security_path_unlink(&path, dentry);
4049                 if (error)
4050                         goto exit2;
4051                 error = vfs_unlink(path.dentry->d_inode, dentry, &delegated_inode);
4052 exit2:
4053                 dput(dentry);
4054         }
4055         inode_unlock(path.dentry->d_inode);
4056         if (inode)
4057                 iput(inode);    /* truncate the inode here */
4058         inode = NULL;
4059         if (delegated_inode) {
4060                 error = break_deleg_wait(&delegated_inode);
4061                 if (!error)
4062                         goto retry_deleg;
4063         }
4064         mnt_drop_write(path.mnt);
4065 exit1:
4066         path_put(&path);
4067         putname(name);
4068         if (retry_estale(error, lookup_flags)) {
4069                 lookup_flags |= LOOKUP_REVAL;
4070                 inode = NULL;
4071                 goto retry;
4072         }
4073         return error;
4074
4075 slashes:
4076         if (d_is_negative(dentry))
4077                 error = -ENOENT;
4078         else if (d_is_dir(dentry))
4079                 error = -EISDIR;
4080         else
4081                 error = -ENOTDIR;
4082         goto exit2;
4083 }
4084
4085 SYSCALL_DEFINE3(unlinkat, int, dfd, const char __user *, pathname, int, flag)
4086 {
4087         if ((flag & ~AT_REMOVEDIR) != 0)
4088                 return -EINVAL;
4089
4090         if (flag & AT_REMOVEDIR)
4091                 return do_rmdir(dfd, pathname);
4092
4093         return do_unlinkat(dfd, pathname);
4094 }
4095
4096 SYSCALL_DEFINE1(unlink, const char __user *, pathname)
4097 {
4098         return do_unlinkat(AT_FDCWD, pathname);
4099 }
4100
4101 int vfs_symlink(struct inode *dir, struct dentry *dentry, const char *oldname)
4102 {
4103         int error = may_create(dir, dentry);
4104
4105         if (error)
4106                 return error;
4107
4108         if (!dir->i_op->symlink)
4109                 return -EPERM;
4110
4111         error = security_inode_symlink(dir, dentry, oldname);
4112         if (error)
4113                 return error;
4114
4115         error = dir->i_op->symlink(dir, dentry, oldname);
4116         if (!error)
4117                 fsnotify_create(dir, dentry);
4118         return error;
4119 }
4120 EXPORT_SYMBOL(vfs_symlink);
4121
4122 SYSCALL_DEFINE3(symlinkat, const char __user *, oldname,
4123                 int, newdfd, const char __user *, newname)
4124 {
4125         int error;
4126         struct filename *from;
4127         struct dentry *dentry;
4128         struct path path;
4129         unsigned int lookup_flags = 0;
4130
4131         from = getname(oldname);
4132         if (IS_ERR(from))
4133                 return PTR_ERR(from);
4134 retry:
4135         dentry = user_path_create(newdfd, newname, &path, lookup_flags);
4136         error = PTR_ERR(dentry);
4137         if (IS_ERR(dentry))
4138                 goto out_putname;
4139
4140         error = security_path_symlink(&path, dentry, from->name);
4141         if (!error)
4142                 error = vfs_symlink(path.dentry->d_inode, dentry, from->name);
4143         done_path_create(&path, dentry);
4144         if (retry_estale(error, lookup_flags)) {
4145                 lookup_flags |= LOOKUP_REVAL;
4146                 goto retry;
4147         }
4148 out_putname:
4149         putname(from);
4150         return error;
4151 }
4152
4153 SYSCALL_DEFINE2(symlink, const char __user *, oldname, const char __user *, newname)
4154 {
4155         return sys_symlinkat(oldname, AT_FDCWD, newname);
4156 }
4157
4158 /**
4159  * vfs_link - create a new link
4160  * @old_dentry: object to be linked
4161  * @dir:        new parent
4162  * @new_dentry: where to create the new link
4163  * @delegated_inode: returns inode needing a delegation break
4164  *
4165  * The caller must hold dir->i_mutex
4166  *
4167  * If vfs_link discovers a delegation on the to-be-linked file in need
4168  * of breaking, it will return -EWOULDBLOCK and return a reference to the
4169  * inode in delegated_inode.  The caller should then break the delegation
4170  * and retry.  Because breaking a delegation may take a long time, the
4171  * caller should drop the i_mutex before doing so.
4172  *
4173  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4174  * be appropriate for callers that expect the underlying filesystem not
4175  * to be NFS exported.
4176  */
4177 int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_dentry, struct inode **delegated_inode)
4178 {
4179         struct inode *inode = old_dentry->d_inode;
4180         unsigned max_links = dir->i_sb->s_max_links;
4181         int error;
4182
4183         if (!inode)
4184                 return -ENOENT;
4185
4186         error = may_create(dir, new_dentry);
4187         if (error)
4188                 return error;
4189
4190         if (dir->i_sb != inode->i_sb)
4191                 return -EXDEV;
4192
4193         /*
4194          * A link to an append-only or immutable file cannot be created.
4195          */
4196         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4197                 return -EPERM;
4198         /*
4199          * Updating the link count will likely cause i_uid and i_gid to
4200          * be writen back improperly if their true value is unknown to
4201          * the vfs.
4202          */
4203         if (HAS_UNMAPPED_ID(inode))
4204                 return -EPERM;
4205         if (!dir->i_op->link)
4206                 return -EPERM;
4207         if (S_ISDIR(inode->i_mode))
4208                 return -EPERM;
4209
4210         error = security_inode_link(old_dentry, dir, new_dentry);
4211         if (error)
4212                 return error;
4213
4214         inode_lock(inode);
4215         /* Make sure we don't allow creating hardlink to an unlinked file */
4216         if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))
4217                 error =  -ENOENT;
4218         else if (max_links && inode->i_nlink >= max_links)
4219                 error = -EMLINK;
4220         else {
4221                 error = try_break_deleg(inode, delegated_inode);
4222                 if (!error)
4223                         error = dir->i_op->link(old_dentry, dir, new_dentry);
4224         }
4225
4226         if (!error && (inode->i_state & I_LINKABLE)) {
4227                 spin_lock(&inode->i_lock);
4228                 inode->i_state &= ~I_LINKABLE;
4229                 spin_unlock(&inode->i_lock);
4230         }
4231         inode_unlock(inode);
4232         if (!error)
4233                 fsnotify_link(dir, inode, new_dentry);
4234         return error;
4235 }
4236 EXPORT_SYMBOL(vfs_link);
4237
4238 /*
4239  * Hardlinks are often used in delicate situations.  We avoid
4240  * security-related surprises by not following symlinks on the
4241  * newname.  --KAB
4242  *
4243  * We don't follow them on the oldname either to be compatible
4244  * with linux 2.0, and to avoid hard-linking to directories
4245  * and other special files.  --ADM
4246  */
4247 SYSCALL_DEFINE5(linkat, int, olddfd, const char __user *, oldname,
4248                 int, newdfd, const char __user *, newname, int, flags)
4249 {
4250         struct dentry *new_dentry;
4251         struct path old_path, new_path;
4252         struct inode *delegated_inode = NULL;
4253         int how = 0;
4254         int error;
4255
4256         if ((flags & ~(AT_SYMLINK_FOLLOW | AT_EMPTY_PATH)) != 0)
4257                 return -EINVAL;
4258         /*
4259          * To use null names we require CAP_DAC_READ_SEARCH
4260          * This ensures that not everyone will be able to create
4261          * handlink using the passed filedescriptor.
4262          */
4263         if (flags & AT_EMPTY_PATH) {
4264                 if (!capable(CAP_DAC_READ_SEARCH))
4265                         return -ENOENT;
4266                 how = LOOKUP_EMPTY;
4267         }
4268
4269         if (flags & AT_SYMLINK_FOLLOW)
4270                 how |= LOOKUP_FOLLOW;
4271 retry:
4272         error = user_path_at(olddfd, oldname, how, &old_path);
4273         if (error)
4274                 return error;
4275
4276         new_dentry = user_path_create(newdfd, newname, &new_path,
4277                                         (how & LOOKUP_REVAL));
4278         error = PTR_ERR(new_dentry);
4279         if (IS_ERR(new_dentry))
4280                 goto out;
4281
4282         error = -EXDEV;
4283         if (old_path.mnt != new_path.mnt)
4284                 goto out_dput;
4285         error = may_linkat(&old_path);
4286         if (unlikely(error))
4287                 goto out_dput;
4288         error = security_path_link(old_path.dentry, &new_path, new_dentry);
4289         if (error)
4290                 goto out_dput;
4291         error = vfs_link(old_path.dentry, new_path.dentry->d_inode, new_dentry, &delegated_inode);
4292 out_dput:
4293         done_path_create(&new_path, new_dentry);
4294         if (delegated_inode) {
4295                 error = break_deleg_wait(&delegated_inode);
4296                 if (!error) {
4297                         path_put(&old_path);
4298                         goto retry;
4299                 }
4300         }
4301         if (retry_estale(error, how)) {
4302                 path_put(&old_path);
4303                 how |= LOOKUP_REVAL;
4304                 goto retry;
4305         }
4306 out:
4307         path_put(&old_path);
4308
4309         return error;
4310 }
4311
4312 SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname)
4313 {
4314         return sys_linkat(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4315 }
4316
4317 /**
4318  * vfs_rename - rename a filesystem object
4319  * @old_dir:    parent of source
4320  * @old_dentry: source
4321  * @new_dir:    parent of destination
4322  * @new_dentry: destination
4323  * @delegated_inode: returns an inode needing a delegation break
4324  * @flags:      rename flags
4325  *
4326  * The caller must hold multiple mutexes--see lock_rename()).
4327  *
4328  * If vfs_rename discovers a delegation in need of breaking at either
4329  * the source or destination, it will return -EWOULDBLOCK and return a
4330  * reference to the inode in delegated_inode.  The caller should then
4331  * break the delegation and retry.  Because breaking a delegation may
4332  * take a long time, the caller should drop all locks before doing
4333  * so.
4334  *
4335  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4336  * be appropriate for callers that expect the underlying filesystem not
4337  * to be NFS exported.
4338  *
4339  * The worst of all namespace operations - renaming directory. "Perverted"
4340  * doesn't even start to describe it. Somebody in UCB had a heck of a trip...
4341  * Problems:
4342  *
4343  *      a) we can get into loop creation.
4344  *      b) race potential - two innocent renames can create a loop together.
4345  *         That's where 4.4 screws up. Current fix: serialization on
4346  *         sb->s_vfs_rename_mutex. We might be more accurate, but that's another
4347  *         story.
4348  *      c) we have to lock _four_ objects - parents and victim (if it exists),
4349  *         and source (if it is not a directory).
4350  *         And that - after we got ->i_mutex on parents (until then we don't know
4351  *         whether the target exists).  Solution: try to be smart with locking
4352  *         order for inodes.  We rely on the fact that tree topology may change
4353  *         only under ->s_vfs_rename_mutex _and_ that parent of the object we
4354  *         move will be locked.  Thus we can rank directories by the tree
4355  *         (ancestors first) and rank all non-directories after them.
4356  *         That works since everybody except rename does "lock parent, lookup,
4357  *         lock child" and rename is under ->s_vfs_rename_mutex.
4358  *         HOWEVER, it relies on the assumption that any object with ->lookup()
4359  *         has no more than 1 dentry.  If "hybrid" objects will ever appear,
4360  *         we'd better make sure that there's no link(2) for them.
4361  *      d) conversion from fhandle to dentry may come in the wrong moment - when
4362  *         we are removing the target. Solution: we will have to grab ->i_mutex
4363  *         in the fhandle_to_dentry code. [FIXME - current nfsfh.c relies on
4364  *         ->i_mutex on parents, which works but leads to some truly excessive
4365  *         locking].
4366  */
4367 int vfs_rename(struct inode *old_dir, struct dentry *old_dentry,
4368                struct inode *new_dir, struct dentry *new_dentry,
4369                struct inode **delegated_inode, unsigned int flags)
4370 {
4371         int error;
4372         bool is_dir = d_is_dir(old_dentry);
4373         struct inode *source = old_dentry->d_inode;
4374         struct inode *target = new_dentry->d_inode;
4375         bool new_is_dir = false;
4376         unsigned max_links = new_dir->i_sb->s_max_links;
4377         struct name_snapshot old_name;
4378
4379         if (source == target)
4380                 return 0;
4381
4382         error = may_delete(old_dir, old_dentry, is_dir);
4383         if (error)
4384                 return error;
4385
4386         if (!target) {
4387                 error = may_create(new_dir, new_dentry);
4388         } else {
4389                 new_is_dir = d_is_dir(new_dentry);
4390
4391                 if (!(flags & RENAME_EXCHANGE))
4392                         error = may_delete(new_dir, new_dentry, is_dir);
4393                 else
4394                         error = may_delete(new_dir, new_dentry, new_is_dir);
4395         }
4396         if (error)
4397                 return error;
4398
4399         if (!old_dir->i_op->rename)
4400                 return -EPERM;
4401
4402         /*
4403          * If we are going to change the parent - check write permissions,
4404          * we'll need to flip '..'.
4405          */
4406         if (new_dir != old_dir) {
4407                 if (is_dir) {
4408                         error = inode_permission(source, MAY_WRITE);
4409                         if (error)
4410                                 return error;
4411                 }
4412                 if ((flags & RENAME_EXCHANGE) && new_is_dir) {
4413                         error = inode_permission(target, MAY_WRITE);
4414                         if (error)
4415                                 return error;
4416                 }
4417         }
4418
4419         error = security_inode_rename(old_dir, old_dentry, new_dir, new_dentry,
4420                                       flags);
4421         if (error)
4422                 return error;
4423
4424         take_dentry_name_snapshot(&old_name, old_dentry);
4425         dget(new_dentry);
4426         if (!is_dir || (flags & RENAME_EXCHANGE))
4427                 lock_two_nondirectories(source, target);
4428         else if (target)
4429                 inode_lock(target);
4430
4431         error = -EBUSY;
4432         if (is_local_mountpoint(old_dentry) || is_local_mountpoint(new_dentry))
4433                 goto out;
4434
4435         if (max_links && new_dir != old_dir) {
4436                 error = -EMLINK;
4437                 if (is_dir && !new_is_dir && new_dir->i_nlink >= max_links)
4438                         goto out;
4439                 if ((flags & RENAME_EXCHANGE) && !is_dir && new_is_dir &&
4440                     old_dir->i_nlink >= max_links)
4441                         goto out;
4442         }
4443         if (is_dir && !(flags & RENAME_EXCHANGE) && target)
4444                 shrink_dcache_parent(new_dentry);
4445         if (!is_dir) {
4446                 error = try_break_deleg(source, delegated_inode);
4447                 if (error)
4448                         goto out;
4449         }
4450         if (target && !new_is_dir) {
4451                 error = try_break_deleg(target, delegated_inode);
4452                 if (error)
4453                         goto out;
4454         }
4455         error = old_dir->i_op->rename(old_dir, old_dentry,
4456                                        new_dir, new_dentry, flags);
4457         if (error)
4458                 goto out;
4459
4460         if (!(flags & RENAME_EXCHANGE) && target) {
4461                 if (is_dir)
4462                         target->i_flags |= S_DEAD;
4463                 dont_mount(new_dentry);
4464                 detach_mounts(new_dentry);
4465         }
4466         if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE)) {
4467                 if (!(flags & RENAME_EXCHANGE))
4468                         d_move(old_dentry, new_dentry);
4469                 else
4470                         d_exchange(old_dentry, new_dentry);
4471         }
4472 out:
4473         if (!is_dir || (flags & RENAME_EXCHANGE))
4474                 unlock_two_nondirectories(source, target);
4475         else if (target)
4476                 inode_unlock(target);
4477         dput(new_dentry);
4478         if (!error) {
4479                 fsnotify_move(old_dir, new_dir, old_name.name, is_dir,
4480                               !(flags & RENAME_EXCHANGE) ? target : NULL, old_dentry);
4481                 if (flags & RENAME_EXCHANGE) {
4482                         fsnotify_move(new_dir, old_dir, old_dentry->d_name.name,
4483                                       new_is_dir, NULL, new_dentry);
4484                 }
4485         }
4486         release_dentry_name_snapshot(&old_name);
4487
4488         return error;
4489 }
4490 EXPORT_SYMBOL(vfs_rename);
4491
4492 SYSCALL_DEFINE5(renameat2, int, olddfd, const char __user *, oldname,
4493                 int, newdfd, const char __user *, newname, unsigned int, flags)
4494 {
4495         struct dentry *old_dentry, *new_dentry;
4496         struct dentry *trap;
4497         struct path old_path, new_path;
4498         struct qstr old_last, new_last;
4499         int old_type, new_type;
4500         struct inode *delegated_inode = NULL;
4501         struct filename *from;
4502         struct filename *to;
4503         unsigned int lookup_flags = 0, target_flags = LOOKUP_RENAME_TARGET;
4504         bool should_retry = false;
4505         int error;
4506
4507         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
4508                 return -EINVAL;
4509
4510         if ((flags & (RENAME_NOREPLACE | RENAME_WHITEOUT)) &&
4511             (flags & RENAME_EXCHANGE))
4512                 return -EINVAL;
4513
4514         if ((flags & RENAME_WHITEOUT) && !capable(CAP_MKNOD))
4515                 return -EPERM;
4516
4517         if (flags & RENAME_EXCHANGE)
4518                 target_flags = 0;
4519
4520 retry:
4521         from = filename_parentat(olddfd, getname(oldname), lookup_flags,
4522                                 &old_path, &old_last, &old_type);
4523         if (IS_ERR(from)) {
4524                 error = PTR_ERR(from);
4525                 goto exit;
4526         }
4527
4528         to = filename_parentat(newdfd, getname(newname), lookup_flags,
4529                                 &new_path, &new_last, &new_type);
4530         if (IS_ERR(to)) {
4531                 error = PTR_ERR(to);
4532                 goto exit1;
4533         }
4534
4535         error = -EXDEV;
4536         if (old_path.mnt != new_path.mnt)
4537                 goto exit2;
4538
4539         error = -EBUSY;
4540         if (old_type != LAST_NORM)
4541                 goto exit2;
4542
4543         if (flags & RENAME_NOREPLACE)
4544                 error = -EEXIST;
4545         if (new_type != LAST_NORM)
4546                 goto exit2;
4547
4548         error = mnt_want_write(old_path.mnt);
4549         if (error)
4550                 goto exit2;
4551
4552 retry_deleg:
4553         trap = lock_rename(new_path.dentry, old_path.dentry);
4554
4555         old_dentry = __lookup_hash(&old_last, old_path.dentry, lookup_flags);
4556         error = PTR_ERR(old_dentry);
4557         if (IS_ERR(old_dentry))
4558                 goto exit3;
4559         /* source must exist */
4560         error = -ENOENT;
4561         if (d_is_negative(old_dentry))
4562                 goto exit4;
4563         new_dentry = __lookup_hash(&new_last, new_path.dentry, lookup_flags | target_flags);
4564         error = PTR_ERR(new_dentry);
4565         if (IS_ERR(new_dentry))
4566                 goto exit4;
4567         error = -EEXIST;
4568         if ((flags & RENAME_NOREPLACE) && d_is_positive(new_dentry))
4569                 goto exit5;
4570         if (flags & RENAME_EXCHANGE) {
4571                 error = -ENOENT;
4572                 if (d_is_negative(new_dentry))
4573                         goto exit5;
4574
4575                 if (!d_is_dir(new_dentry)) {
4576                         error = -ENOTDIR;
4577                         if (new_last.name[new_last.len])
4578                                 goto exit5;
4579                 }
4580         }
4581         /* unless the source is a directory trailing slashes give -ENOTDIR */
4582         if (!d_is_dir(old_dentry)) {
4583                 error = -ENOTDIR;
4584                 if (old_last.name[old_last.len])
4585                         goto exit5;
4586                 if (!(flags & RENAME_EXCHANGE) && new_last.name[new_last.len])
4587                         goto exit5;
4588         }
4589         /* source should not be ancestor of target */
4590         error = -EINVAL;
4591         if (old_dentry == trap)
4592                 goto exit5;
4593         /* target should not be an ancestor of source */
4594         if (!(flags & RENAME_EXCHANGE))
4595                 error = -ENOTEMPTY;
4596         if (new_dentry == trap)
4597                 goto exit5;
4598
4599         error = security_path_rename(&old_path, old_dentry,
4600                                      &new_path, new_dentry, flags);
4601         if (error)
4602                 goto exit5;
4603         error = vfs_rename(old_path.dentry->d_inode, old_dentry,
4604                            new_path.dentry->d_inode, new_dentry,
4605                            &delegated_inode, flags);
4606 exit5:
4607         dput(new_dentry);
4608 exit4:
4609         dput(old_dentry);
4610 exit3:
4611         unlock_rename(new_path.dentry, old_path.dentry);
4612         if (delegated_inode) {
4613                 error = break_deleg_wait(&delegated_inode);
4614                 if (!error)
4615                         goto retry_deleg;
4616         }
4617         mnt_drop_write(old_path.mnt);
4618 exit2:
4619         if (retry_estale(error, lookup_flags))
4620                 should_retry = true;
4621         path_put(&new_path);
4622         putname(to);
4623 exit1:
4624         path_put(&old_path);
4625         putname(from);
4626         if (should_retry) {
4627                 should_retry = false;
4628                 lookup_flags |= LOOKUP_REVAL;
4629                 goto retry;
4630         }
4631 exit:
4632         return error;
4633 }
4634
4635 SYSCALL_DEFINE4(renameat, int, olddfd, const char __user *, oldname,
4636                 int, newdfd, const char __user *, newname)
4637 {
4638         return sys_renameat2(olddfd, oldname, newdfd, newname, 0);
4639 }
4640
4641 SYSCALL_DEFINE2(rename, const char __user *, oldname, const char __user *, newname)
4642 {
4643         return sys_renameat2(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4644 }
4645
4646 int vfs_whiteout(struct inode *dir, struct dentry *dentry)
4647 {
4648         int error = may_create(dir, dentry);
4649         if (error)
4650                 return error;
4651
4652         if (!dir->i_op->mknod)
4653                 return -EPERM;
4654
4655         return dir->i_op->mknod(dir, dentry,
4656                                 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
4657 }
4658 EXPORT_SYMBOL(vfs_whiteout);
4659
4660 int readlink_copy(char __user *buffer, int buflen, const char *link)
4661 {
4662         int len = PTR_ERR(link);
4663         if (IS_ERR(link))
4664                 goto out;
4665
4666         len = strlen(link);
4667         if (len > (unsigned) buflen)
4668                 len = buflen;
4669         if (copy_to_user(buffer, link, len))
4670                 len = -EFAULT;
4671 out:
4672         return len;
4673 }
4674
4675 /*
4676  * A helper for ->readlink().  This should be used *ONLY* for symlinks that
4677  * have ->get_link() not calling nd_jump_link().  Using (or not using) it
4678  * for any given inode is up to filesystem.
4679  */
4680 static int generic_readlink(struct dentry *dentry, char __user *buffer,
4681                             int buflen)
4682 {
4683         DEFINE_DELAYED_CALL(done);
4684         struct inode *inode = d_inode(dentry);
4685         const char *link = inode->i_link;
4686         int res;
4687
4688         if (!link) {
4689                 link = inode->i_op->get_link(dentry, inode, &done);
4690                 if (IS_ERR(link))
4691                         return PTR_ERR(link);
4692         }
4693         res = readlink_copy(buffer, buflen, link);
4694         do_delayed_call(&done);
4695         return res;
4696 }
4697
4698 /**
4699  * vfs_readlink - copy symlink body into userspace buffer
4700  * @dentry: dentry on which to get symbolic link
4701  * @buffer: user memory pointer
4702  * @buflen: size of buffer
4703  *
4704  * Does not touch atime.  That's up to the caller if necessary
4705  *
4706  * Does not call security hook.
4707  */
4708 int vfs_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4709 {
4710         struct inode *inode = d_inode(dentry);
4711
4712         if (unlikely(!(inode->i_opflags & IOP_DEFAULT_READLINK))) {
4713                 if (unlikely(inode->i_op->readlink))
4714                         return inode->i_op->readlink(dentry, buffer, buflen);
4715
4716                 if (!d_is_symlink(dentry))
4717                         return -EINVAL;
4718
4719                 spin_lock(&inode->i_lock);
4720                 inode->i_opflags |= IOP_DEFAULT_READLINK;
4721                 spin_unlock(&inode->i_lock);
4722         }
4723
4724         return generic_readlink(dentry, buffer, buflen);
4725 }
4726 EXPORT_SYMBOL(vfs_readlink);
4727
4728 /**
4729  * vfs_get_link - get symlink body
4730  * @dentry: dentry on which to get symbolic link
4731  * @done: caller needs to free returned data with this
4732  *
4733  * Calls security hook and i_op->get_link() on the supplied inode.
4734  *
4735  * It does not touch atime.  That's up to the caller if necessary.
4736  *
4737  * Does not work on "special" symlinks like /proc/$$/fd/N
4738  */
4739 const char *vfs_get_link(struct dentry *dentry, struct delayed_call *done)
4740 {
4741         const char *res = ERR_PTR(-EINVAL);
4742         struct inode *inode = d_inode(dentry);
4743
4744         if (d_is_symlink(dentry)) {
4745                 res = ERR_PTR(security_inode_readlink(dentry));
4746                 if (!res)
4747                         res = inode->i_op->get_link(dentry, inode, done);
4748         }
4749         return res;
4750 }
4751 EXPORT_SYMBOL(vfs_get_link);
4752
4753 /* get the link contents into pagecache */
4754 const char *page_get_link(struct dentry *dentry, struct inode *inode,
4755                           struct delayed_call *callback)
4756 {
4757         char *kaddr;
4758         struct page *page;
4759         struct address_space *mapping = inode->i_mapping;
4760
4761         if (!dentry) {
4762                 page = find_get_page(mapping, 0);
4763                 if (!page)
4764                         return ERR_PTR(-ECHILD);
4765                 if (!PageUptodate(page)) {
4766                         put_page(page);
4767                         return ERR_PTR(-ECHILD);
4768                 }
4769         } else {
4770                 page = read_mapping_page(mapping, 0, NULL);
4771                 if (IS_ERR(page))
4772                         return (char*)page;
4773         }
4774         set_delayed_call(callback, page_put_link, page);
4775         BUG_ON(mapping_gfp_mask(mapping) & __GFP_HIGHMEM);
4776         kaddr = page_address(page);
4777         nd_terminate_link(kaddr, inode->i_size, PAGE_SIZE - 1);
4778         return kaddr;
4779 }
4780
4781 EXPORT_SYMBOL(page_get_link);
4782
4783 void page_put_link(void *arg)
4784 {
4785         put_page(arg);
4786 }
4787 EXPORT_SYMBOL(page_put_link);
4788
4789 int page_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4790 {
4791         DEFINE_DELAYED_CALL(done);
4792         int res = readlink_copy(buffer, buflen,
4793                                 page_get_link(dentry, d_inode(dentry),
4794                                               &done));
4795         do_delayed_call(&done);
4796         return res;
4797 }
4798 EXPORT_SYMBOL(page_readlink);
4799
4800 /*
4801  * The nofs argument instructs pagecache_write_begin to pass AOP_FLAG_NOFS
4802  */
4803 int __page_symlink(struct inode *inode, const char *symname, int len, int nofs)
4804 {
4805         struct address_space *mapping = inode->i_mapping;
4806         struct page *page;
4807         void *fsdata;
4808         int err;
4809         unsigned int flags = 0;
4810         if (nofs)
4811                 flags |= AOP_FLAG_NOFS;
4812
4813 retry:
4814         err = pagecache_write_begin(NULL, mapping, 0, len-1,
4815                                 flags, &page, &fsdata);
4816         if (err)
4817                 goto fail;
4818
4819         memcpy(page_address(page), symname, len-1);
4820
4821         err = pagecache_write_end(NULL, mapping, 0, len-1, len-1,
4822                                                         page, fsdata);
4823         if (err < 0)
4824                 goto fail;
4825         if (err < len-1)
4826                 goto retry;
4827
4828         mark_inode_dirty(inode);
4829         return 0;
4830 fail:
4831         return err;
4832 }
4833 EXPORT_SYMBOL(__page_symlink);
4834
4835 int page_symlink(struct inode *inode, const char *symname, int len)
4836 {
4837         return __page_symlink(inode, symname, len,
4838                         !mapping_gfp_constraint(inode->i_mapping, __GFP_FS));
4839 }
4840 EXPORT_SYMBOL(page_symlink);
4841
4842 const struct inode_operations page_symlink_inode_operations = {
4843         .get_link       = page_get_link,
4844 };
4845 EXPORT_SYMBOL(page_symlink_inode_operations);