fs/namei.c

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