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