fs/namei.c

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