fs/namei.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  linux/fs/namei.c
   4  *
   5  *  Copyright (C) 1991, 1992  Linus Torvalds
   6  */
   7
   8 /*
   9  * Some corrections by tytso.
  10  */
  11
  12 /* [Feb 1997 T. Schoebel-Theuer] Complete rewrite of the pathname
  13  * lookup logic.
  14  */
  15 /* [Feb-Apr 2000, AV] Rewrite to the new namespace architecture.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/export.h>
  20 #include <linux/kernel.h>
  21 #include <linux/slab.h>
  22 #include <linux/fs.h>
  23 #include <linux/namei.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/fsnotify.h>
  26 #include <linux/personality.h>
  27 #include <linux/security.h>
  28 #include <linux/ima.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/mount.h>
  31 #include <linux/audit.h>
  32 #include <linux/capability.h>
  33 #include <linux/file.h>
  34 #include <linux/fcntl.h>
  35 #include <linux/device_cgroup.h>
  36 #include <linux/fs_struct.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/hash.h>
  39 #include <linux/bitops.h>
  40 #include <linux/init_task.h>
  41 #include <linux/uaccess.h>
  42 #include <linux/build_bug.h>
  43
  44 #include "internal.h"
  45 #include "mount.h"
  46
  47 /* [Feb-1997 T. Schoebel-Theuer]
  48  * Fundamental changes in the pathname lookup mechanisms (namei)
  49  * were necessary because of omirr.  The reason is that omirr needs
  50  * to know the _real_ pathname, not the user-supplied one, in case
  51  * of symlinks (and also when transname replacements occur).
  52  *
  53  * The new code replaces the old recursive symlink resolution with
  54  * an iterative one (in case of non-nested symlink chains).  It does
  55  * this with calls to <fs>_follow_link().
  56  * As a side effect, dir_namei(), _namei() and follow_link() are now
  57  * replaced with a single function lookup_dentry() that can handle all
  58  * the special cases of the former code.
  59  *
  60  * With the new dcache, the pathname is stored at each inode, at least as
  61  * long as the refcount of the inode is positive.  As a side effect, the
  62  * size of the dcache depends on the inode cache and thus is dynamic.
  63  *
  64  * [29-Apr-1998 C. Scott Ananian] Updated above description of symlink
  65  * resolution to correspond with current state of the code.
  66  *
  67  * Note that the symlink resolution is not *completely* iterative.
  68  * There is still a significant amount of tail- and mid- recursion in
  69  * the algorithm.  Also, note that <fs>_readlink() is not used in
  70  * lookup_dentry(): lookup_dentry() on the result of <fs>_readlink()
  71  * may return different results than <fs>_follow_link().  Many virtual
  72  * filesystems (including /proc) exhibit this behavior.
  73  */
  74
  75 /* [24-Feb-97 T. Schoebel-Theuer] Side effects caused by new implementation:
  76  * New symlink semantics: when open() is called with flags O_CREAT | O_EXCL
  77  * and the name already exists in form of a symlink, try to create the new
  78  * name indicated by the symlink. The old code always complained that the
  79  * name already exists, due to not following the symlink even if its target
  80  * is nonexistent.  The new semantics affects also mknod() and link() when
  81  * the name is a symlink pointing to a non-existent name.
  82  *
  83  * I don't know which semantics is the right one, since I have no access
  84  * to standards. But I found by trial that HP-UX 9.0 has the full "new"
  85  * semantics implemented, while SunOS 4.1.1 and Solaris (SunOS 5.4) have the
  86  * "old" one. Personally, I think the new semantics is much more logical.
  87  * Note that "ln old new" where "new" is a symlink pointing to a non-existing
  88  * file does succeed in both HP-UX and SunOs, but not in Solaris
  89  * and in the old Linux semantics.
  90  */
  91
  92 /* [16-Dec-97 Kevin Buhr] For security reasons, we change some symlink
  93  * semantics.  See the comments in "open_namei" and "do_link" below.
  94  *
  95  * [10-Sep-98 Alan Modra] Another symlink change.
  96  */
  97
  98 /* [Feb-Apr 2000 AV] Complete rewrite. Rules for symlinks:
  99  *      inside the path - always follow.
 100  *      in the last component in creation/removal/renaming - never follow.
 101  *      if LOOKUP_FOLLOW passed - follow.
 102  *      if the pathname has trailing slashes - follow.
 103  *      otherwise - don't follow.
 104  * (applied in that order).
 105  *
 106  * [Jun 2000 AV] Inconsistent behaviour of open() in case if flags==O_CREAT
 107  * restored for 2.4. This is the last surviving part of old 4.2BSD bug.
 108  * During the 2.4 we need to fix the userland stuff depending on it -
 109  * hopefully we will be able to get rid of that wart in 2.5. So far only
 110  * XEmacs seems to be relying on it...
 111  */
 112 /*
 113  * [Sep 2001 AV] Single-semaphore locking scheme (kudos to David Holland)
 114  * implemented.  Let's see if raised priority of ->s_vfs_rename_mutex gives
 115  * any extra contention...
 116  */
 117
 118 /* In order to reduce some races, while at the same time doing additional
 119  * checking and hopefully speeding things up, we copy filenames to the
 120  * kernel data space before using them..
 121  *
 122  * POSIX.1 2.4: an empty pathname is invalid (ENOENT).
 123  * PATH_MAX includes the nul terminator --RR.
 124  */
 125
 126 #define EMBEDDED_NAME_MAX       (PATH_MAX - offsetof(struct filename, iname))
 127
 128 struct filename *
 129 getname_flags(const char __user *filename, int flags, int *empty)
 130 {
 131         struct filename *result;
 132         char *kname;
 133         int len;
 134         BUILD_BUG_ON(offsetof(struct filename, iname) % sizeof(long) != 0);
 135
 136         result = audit_reusename(filename);
 137         if (result)
 138                 return result;
 139
 140         result = __getname();
 141         if (unlikely(!result))
 142                 return ERR_PTR(-ENOMEM);
 143
 144         /*
 145          * First, try to embed the struct filename inside the names_cache
 146          * allocation
 147          */
 148         kname = (char *)result->iname;
 149         result->name = kname;
 150
 151         len = strncpy_from_user(kname, filename, EMBEDDED_NAME_MAX);
 152         if (unlikely(len < 0)) {
 153                 __putname(result);
 154                 return ERR_PTR(len);
 155         }
 156
 157         /*
 158          * Uh-oh. We have a name that's approaching PATH_MAX. Allocate a
 159          * separate struct filename so we can dedicate the entire
 160          * names_cache allocation for the pathname, and re-do the copy from
 161          * userland.
 162          */
 163         if (unlikely(len == EMBEDDED_NAME_MAX)) {
 164                 const size_t size = offsetof(struct filename, iname[1]);
 165                 kname = (char *)result;
 166
 167                 /*
 168                  * size is chosen that way we to guarantee that
 169                  * result->iname[0] is within the same object and that
 170                  * kname can't be equal to result->iname, no matter what.
 171                  */
 172                 result = kzalloc(size, GFP_KERNEL);
 173                 if (unlikely(!result)) {
 174                         __putname(kname);
 175                         return ERR_PTR(-ENOMEM);
 176                 }
 177                 result->name = kname;
 178                 len = strncpy_from_user(kname, filename, PATH_MAX);
 179                 if (unlikely(len < 0)) {
 180                         __putname(kname);
 181                         kfree(result);
 182                         return ERR_PTR(len);
 183                 }
 184                 if (unlikely(len == PATH_MAX)) {
 185                         __putname(kname);
 186                         kfree(result);
 187                         return ERR_PTR(-ENAMETOOLONG);
 188                 }
 189         }
 190
 191         result->refcnt = 1;
 192         /* The empty path is special. */
 193         if (unlikely(!len)) {
 194                 if (empty)
 195                         *empty = 1;
 196                 if (!(flags & LOOKUP_EMPTY)) {
 197                         putname(result);
 198                         return ERR_PTR(-ENOENT);
 199                 }
 200         }
 201
 202         result->uptr = filename;
 203         result->aname = NULL;
 204         audit_getname(result);
 205         return result;
 206 }
 207
 208 struct filename *
 209 getname(const char __user * filename)
 210 {
 211         return getname_flags(filename, 0, NULL);
 212 }
 213
 214 struct filename *
 215 getname_kernel(const char * filename)
 216 {
 217         struct filename *result;
 218         int len = strlen(filename) + 1;
 219
 220         result = __getname();
 221         if (unlikely(!result))
 222                 return ERR_PTR(-ENOMEM);
 223
 224         if (len <= EMBEDDED_NAME_MAX) {
 225                 result->name = (char *)result->iname;
 226         } else if (len <= PATH_MAX) {
 227                 const size_t size = offsetof(struct filename, iname[1]);
 228                 struct filename *tmp;
 229
 230                 tmp = kmalloc(size, GFP_KERNEL);
 231                 if (unlikely(!tmp)) {
 232                         __putname(result);
 233                         return ERR_PTR(-ENOMEM);
 234                 }
 235                 tmp->name = (char *)result;
 236                 result = tmp;
 237         } else {
 238                 __putname(result);
 239                 return ERR_PTR(-ENAMETOOLONG);
 240         }
 241         memcpy((char *)result->name, filename, len);
 242         result->uptr = NULL;
 243         result->aname = NULL;
 244         result->refcnt = 1;
 245         audit_getname(result);
 246
 247         return result;
 248 }
 249
 250 void putname(struct filename *name)
 251 {
 252         BUG_ON(name->refcnt <= 0);
 253
 254         if (--name->refcnt > 0)
 255                 return;
 256
 257         if (name->name != name->iname) {
 258                 __putname(name->name);
 259                 kfree(name);
 260         } else
 261                 __putname(name);
 262 }
 263
 264 static int check_acl(struct inode *inode, int mask)
 265 {
 266 #ifdef CONFIG_FS_POSIX_ACL
 267         struct posix_acl *acl;
 268
 269         if (mask & MAY_NOT_BLOCK) {
 270                 acl = get_cached_acl_rcu(inode, ACL_TYPE_ACCESS);
 271                 if (!acl)
 272                         return -EAGAIN;
 273                 /* no ->get_acl() calls in RCU mode... */
 274                 if (is_uncached_acl(acl))
 275                         return -ECHILD;
 276                 return posix_acl_permission(inode, acl, mask & ~MAY_NOT_BLOCK);
 277         }
 278
 279         acl = get_acl(inode, ACL_TYPE_ACCESS);
 280         if (IS_ERR(acl))
 281                 return PTR_ERR(acl);
 282         if (acl) {
 283                 int error = posix_acl_permission(inode, acl, mask);
 284                 posix_acl_release(acl);
 285                 return error;
 286         }
 287 #endif
 288
 289         return -EAGAIN;
 290 }
 291
 292 /*
 293  * This does the basic permission checking
 294  */
 295 static int acl_permission_check(struct inode *inode, int mask)
 296 {
 297         unsigned int mode = inode->i_mode;
 298
 299         if (likely(uid_eq(current_fsuid(), inode->i_uid)))
 300                 mode >>= 6;
 301         else {
 302                 if (IS_POSIXACL(inode) && (mode & S_IRWXG)) {
 303                         int error = check_acl(inode, mask);
 304                         if (error != -EAGAIN)
 305                                 return error;
 306                 }
 307
 308                 if (in_group_p(inode->i_gid))
 309                         mode >>= 3;
 310         }
 311
 312         /*
 313          * If the DACs are ok we don't need any capability check.
 314          */
 315         if ((mask & ~mode & (MAY_READ | MAY_WRITE | MAY_EXEC)) == 0)
 316                 return 0;
 317         return -EACCES;
 318 }
 319
 320 /**
 321  * generic_permission -  check for access rights on a Posix-like filesystem
 322  * @inode:      inode to check access rights for
 323  * @mask:       right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC, ...)
 324  *
 325  * Used to check for read/write/execute permissions on a file.
 326  * We use "fsuid" for this, letting us set arbitrary permissions
 327  * for filesystem access without changing the "normal" uids which
 328  * are used for other things.
 329  *
 330  * generic_permission is rcu-walk aware. It returns -ECHILD in case an rcu-walk
 331  * request cannot be satisfied (eg. requires blocking or too much complexity).
 332  * It would then be called again in ref-walk mode.
 333  */
 334 int generic_permission(struct inode *inode, int mask)
 335 {
 336         int ret;
 337
 338         /*
 339          * Do the basic permission checks.
 340          */
 341         ret = acl_permission_check(inode, mask);
 342         if (ret != -EACCES)
 343                 return ret;
 344
 345         if (S_ISDIR(inode->i_mode)) {
 346                 /* DACs are overridable for directories */
 347                 if (!(mask & MAY_WRITE))
 348                         if (capable_wrt_inode_uidgid(inode,
 349                                                      CAP_DAC_READ_SEARCH))
 350                                 return 0;
 351                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
 352                         return 0;
 353                 return -EACCES;
 354         }
 355
 356         /*
 357          * Searching includes executable on directories, else just read.
 358          */
 359         mask &= MAY_READ | MAY_WRITE | MAY_EXEC;
 360         if (mask == MAY_READ)
 361                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_READ_SEARCH))
 362                         return 0;
 363         /*
 364          * Read/write DACs are always overridable.
 365          * Executable DACs are overridable when there is
 366          * at least one exec bit set.
 367          */
 368         if (!(mask & MAY_EXEC) || (inode->i_mode & S_IXUGO))
 369                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
 370                         return 0;
 371
 372         return -EACCES;
 373 }
 374 EXPORT_SYMBOL(generic_permission);
 375
 376 /*
 377  * We _really_ want to just do "generic_permission()" without
 378  * even looking at the inode->i_op values. So we keep a cache
 379  * flag in inode->i_opflags, that says "this has not special
 380  * permission function, use the fast case".
 381  */
 382 static inline int do_inode_permission(struct inode *inode, int mask)
 383 {
 384         if (unlikely(!(inode->i_opflags & IOP_FASTPERM))) {
 385                 if (likely(inode->i_op->permission))
 386                         return inode->i_op->permission(inode, mask);
 387
 388                 /* This gets set once for the inode lifetime */
 389                 spin_lock(&inode->i_lock);
 390                 inode->i_opflags |= IOP_FASTPERM;
 391                 spin_unlock(&inode->i_lock);
 392         }
 393         return generic_permission(inode, mask);
 394 }
 395
 396 /**
 397  * sb_permission - Check superblock-level permissions
 398  * @sb: Superblock of inode to check permission on
 399  * @inode: Inode to check permission on
 400  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 401  *
 402  * Separate out file-system wide checks from inode-specific permission checks.
 403  */
 404 static int sb_permission(struct super_block *sb, struct inode *inode, int mask)
 405 {
 406         if (unlikely(mask & MAY_WRITE)) {
 407                 umode_t mode = inode->i_mode;
 408
 409                 /* Nobody gets write access to a read-only fs. */
 410                 if (sb_rdonly(sb) && (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
 411                         return -EROFS;
 412         }
 413         return 0;
 414 }
 415
 416 /**
 417  * inode_permission - Check for access rights to a given inode
 418  * @inode: Inode to check permission on
 419  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 420  *
 421  * Check for read/write/execute permissions on an inode.  We use fs[ug]id for
 422  * this, letting us set arbitrary permissions for filesystem access without
 423  * changing the "normal" UIDs which are used for other things.
 424  *
 425  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
 426  */
 427 int inode_permission(struct inode *inode, int mask)
 428 {
 429         int retval;
 430
 431         retval = sb_permission(inode->i_sb, inode, mask);
 432         if (retval)
 433                 return retval;
 434
 435         if (unlikely(mask & MAY_WRITE)) {
 436                 /*
 437                  * Nobody gets write access to an immutable file.
 438                  */
 439                 if (IS_IMMUTABLE(inode))
 440                         return -EPERM;
 441
 442                 /*
 443                  * Updating mtime will likely cause i_uid and i_gid to be
 444                  * written back improperly if their true value is unknown
 445                  * to the vfs.
 446                  */
 447                 if (HAS_UNMAPPED_ID(inode))
 448                         return -EACCES;
 449         }
 450
 451         retval = do_inode_permission(inode, mask);
 452         if (retval)
 453                 return retval;
 454
 455         retval = devcgroup_inode_permission(inode, mask);
 456         if (retval)
 457                 return retval;
 458
 459         return security_inode_permission(inode, mask);
 460 }
 461 EXPORT_SYMBOL(inode_permission);
 462
 463 /**
 464  * path_get - get a reference to a path
 465  * @path: path to get the reference to
 466  *
 467  * Given a path increment the reference count to the dentry and the vfsmount.
 468  */
 469 void path_get(const struct path *path)
 470 {
 471         mntget(path->mnt);
 472         dget(path->dentry);
 473 }
 474 EXPORT_SYMBOL(path_get);
 475
 476 /**
 477  * path_put - put a reference to a path
 478  * @path: path to put the reference to
 479  *
 480  * Given a path decrement the reference count to the dentry and the vfsmount.
 481  */
 482 void path_put(const struct path *path)
 483 {
 484         dput(path->dentry);
 485         mntput(path->mnt);
 486 }
 487 EXPORT_SYMBOL(path_put);
 488
 489 #define EMBEDDED_LEVELS 2
 490 struct nameidata {
 491         struct path     path;
 492         struct qstr     last;
 493         struct path     root;
 494         struct inode    *inode; /* path.dentry.d_inode */
 495         unsigned int    flags;
 496         unsigned        seq, m_seq;
 497         int             last_type;
 498         unsigned        depth;
 499         int             total_link_count;
 500         struct saved {
 501                 struct path link;
 502                 struct delayed_call done;
 503                 const char *name;
 504                 unsigned seq;
 505         } *stack, internal[EMBEDDED_LEVELS];
 506         struct filename *name;
 507         struct nameidata *saved;
 508         struct inode    *link_inode;
 509         unsigned        root_seq;
 510         int             dfd;
 511 } __randomize_layout;
 512
 513 static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
 514 {
 515         struct nameidata *old = current->nameidata;
 516         p->stack = p->internal;
 517         p->dfd = dfd;
 518         p->name = name;
 519         p->total_link_count = old ? old->total_link_count : 0;
 520         p->saved = old;
 521         current->nameidata = p;
 522 }
 523
 524 static void restore_nameidata(void)
 525 {
 526         struct nameidata *now = current->nameidata, *old = now->saved;
 527
 528         current->nameidata = old;
 529         if (old)
 530                 old->total_link_count = now->total_link_count;
 531         if (now->stack != now->internal)
 532                 kfree(now->stack);
 533 }
 534
 535 static int __nd_alloc_stack(struct nameidata *nd)
 536 {
 537         struct saved *p;
 538
 539         if (nd->flags & LOOKUP_RCU) {
 540                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
 541                                   GFP_ATOMIC);
 542                 if (unlikely(!p))
 543                         return -ECHILD;
 544         } else {
 545                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
 546                                   GFP_KERNEL);
 547                 if (unlikely(!p))
 548                         return -ENOMEM;
 549         }
 550         memcpy(p, nd->internal, sizeof(nd->internal));
 551         nd->stack = p;
 552         return 0;
 553 }
 554
 555 /**
 556  * path_connected - Verify that a path->dentry is below path->mnt.mnt_root
 557  * @path: nameidate to verify
 558  *
 559  * Rename can sometimes move a file or directory outside of a bind
 560  * mount, path_connected allows those cases to be detected.
 561  */
 562 static bool path_connected(const struct path *path)
 563 {
 564         struct vfsmount *mnt = path->mnt;
 565         struct super_block *sb = mnt->mnt_sb;
 566
 567         /* Bind mounts and multi-root filesystems can have disconnected paths */
 568         if (!(sb->s_iflags & SB_I_MULTIROOT) && (mnt->mnt_root == sb->s_root))
 569                 return true;
 570
 571         return is_subdir(path->dentry, mnt->mnt_root);
 572 }
 573
 574 static inline int nd_alloc_stack(struct nameidata *nd)
 575 {
 576         if (likely(nd->depth != EMBEDDED_LEVELS))
 577                 return 0;
 578         if (likely(nd->stack != nd->internal))
 579                 return 0;
 580         return __nd_alloc_stack(nd);
 581 }
 582
 583 static void drop_links(struct nameidata *nd)
 584 {
 585         int i = nd->depth;
 586         while (i--) {
 587                 struct saved *last = nd->stack + i;
 588                 do_delayed_call(&last->done);
 589                 clear_delayed_call(&last->done);
 590         }
 591 }
 592
 593 static void terminate_walk(struct nameidata *nd)
 594 {
 595         drop_links(nd);
 596         if (!(nd->flags & LOOKUP_RCU)) {
 597                 int i;
 598                 path_put(&nd->path);
 599                 for (i = 0; i < nd->depth; i++)
 600                         path_put(&nd->stack[i].link);
 601                 if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 602                         path_put(&nd->root);
 603                         nd->root.mnt = NULL;
 604                 }
 605         } else {
 606                 nd->flags &= ~LOOKUP_RCU;
 607                 if (!(nd->flags & LOOKUP_ROOT))
 608                         nd->root.mnt = NULL;
 609                 rcu_read_unlock();
 610         }
 611         nd->depth = 0;
 612 }
 613
 614 /* path_put is needed afterwards regardless of success or failure */
 615 static bool legitimize_path(struct nameidata *nd,
 616                             struct path *path, unsigned seq)
 617 {
 618         int res = __legitimize_mnt(path->mnt, nd->m_seq);
 619         if (unlikely(res)) {
 620                 if (res > 0)
 621                         path->mnt = NULL;
 622                 path->dentry = NULL;
 623                 return false;
 624         }
 625         if (unlikely(!lockref_get_not_dead(&path->dentry->d_lockref))) {
 626                 path->dentry = NULL;
 627                 return false;
 628         }
 629         return !read_seqcount_retry(&path->dentry->d_seq, seq);
 630 }
 631
 632 static bool legitimize_links(struct nameidata *nd)
 633 {
 634         int i;
 635         for (i = 0; i < nd->depth; i++) {
 636                 struct saved *last = nd->stack + i;
 637                 if (unlikely(!legitimize_path(nd, &last->link, last->seq))) {
 638                         drop_links(nd);
 639                         nd->depth = i + 1;
 640                         return false;
 641                 }
 642         }
 643         return true;
 644 }
 645
 646 /*
 647  * Path walking has 2 modes, rcu-walk and ref-walk (see
 648  * Documentation/filesystems/path-lookup.txt).  In situations when we can't
 649  * continue in RCU mode, we attempt to drop out of rcu-walk mode and grab
 650  * normal reference counts on dentries and vfsmounts to transition to ref-walk
 651  * mode.  Refcounts are grabbed at the last known good point before rcu-walk
 652  * got stuck, so ref-walk may continue from there. If this is not successful
 653  * (eg. a seqcount has changed), then failure is returned and it's up to caller
 654  * to restart the path walk from the beginning in ref-walk mode.
 655  */
 656
 657 /**
 658  * unlazy_walk - try to switch to ref-walk mode.
 659  * @nd: nameidata pathwalk data
 660  * Returns: 0 on success, -ECHILD on failure
 661  *
 662  * unlazy_walk attempts to legitimize the current nd->path and nd->root
 663  * for ref-walk mode.
 664  * Must be called from rcu-walk context.
 665  * Nothing should touch nameidata between unlazy_walk() failure and
 666  * terminate_walk().
 667  */
 668 static int unlazy_walk(struct nameidata *nd)
 669 {
 670         struct dentry *parent = nd->path.dentry;
 671
 672         BUG_ON(!(nd->flags & LOOKUP_RCU));
 673
 674         nd->flags &= ~LOOKUP_RCU;
 675         if (unlikely(!legitimize_links(nd)))
 676                 goto out2;
 677         if (unlikely(!legitimize_path(nd, &nd->path, nd->seq)))
 678                 goto out1;
 679         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 680                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq)))
 681                         goto out;
 682         }
 683         rcu_read_unlock();
 684         BUG_ON(nd->inode != parent->d_inode);
 685         return 0;
 686
 687 out2:
 688         nd->path.mnt = NULL;
 689         nd->path.dentry = NULL;
 690 out1:
 691         if (!(nd->flags & LOOKUP_ROOT))
 692                 nd->root.mnt = NULL;
 693 out:
 694         rcu_read_unlock();
 695         return -ECHILD;
 696 }
 697
 698 /**
 699  * unlazy_child - try to switch to ref-walk mode.
 700  * @nd: nameidata pathwalk data
 701  * @dentry: child of nd->path.dentry
 702  * @seq: seq number to check dentry against
 703  * Returns: 0 on success, -ECHILD on failure
 704  *
 705  * unlazy_child attempts to legitimize the current nd->path, nd->root and dentry
 706  * for ref-walk mode.  @dentry must be a path found by a do_lookup call on
 707  * @nd.  Must be called from rcu-walk context.
 708  * Nothing should touch nameidata between unlazy_child() failure and
 709  * terminate_walk().
 710  */
 711 static int unlazy_child(struct nameidata *nd, struct dentry *dentry, unsigned seq)
 712 {
 713         BUG_ON(!(nd->flags & LOOKUP_RCU));
 714
 715         nd->flags &= ~LOOKUP_RCU;
 716         if (unlikely(!legitimize_links(nd)))
 717                 goto out2;
 718         if (unlikely(!legitimize_mnt(nd->path.mnt, nd->m_seq)))
 719                 goto out2;
 720         if (unlikely(!lockref_get_not_dead(&nd->path.dentry->d_lockref)))
 721                 goto out1;
 722
 723         /*
 724          * We need to move both the parent and the dentry from the RCU domain
 725          * to be properly refcounted. And the sequence number in the dentry
 726          * validates *both* dentry counters, since we checked the sequence
 727          * number of the parent after we got the child sequence number. So we
 728          * know the parent must still be valid if the child sequence number is
 729          */
 730         if (unlikely(!lockref_get_not_dead(&dentry->d_lockref)))
 731                 goto out;
 732         if (unlikely(read_seqcount_retry(&dentry->d_seq, seq))) {
 733                 rcu_read_unlock();
 734                 dput(dentry);
 735                 goto drop_root_mnt;
 736         }
 737         /*
 738          * Sequence counts matched. Now make sure that the root is
 739          * still valid and get it if required.
 740          */
 741         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 742                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq))) {
 743                         rcu_read_unlock();
 744                         dput(dentry);
 745                         return -ECHILD;
 746                 }
 747         }
 748
 749         rcu_read_unlock();
 750         return 0;
 751
 752 out2:
 753         nd->path.mnt = NULL;
 754 out1:
 755         nd->path.dentry = NULL;
 756 out:
 757         rcu_read_unlock();
 758 drop_root_mnt:
 759         if (!(nd->flags & LOOKUP_ROOT))
 760                 nd->root.mnt = NULL;
 761         return -ECHILD;
 762 }
 763
 764 static inline int d_revalidate(struct dentry *dentry, unsigned int flags)
 765 {
 766         if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
 767                 return dentry->d_op->d_revalidate(dentry, flags);
 768         else
 769                 return 1;
 770 }
 771
 772 /**
 773  * complete_walk - successful completion of path walk
 774  * @nd:  pointer nameidata
 775  *
 776  * If we had been in RCU mode, drop out of it and legitimize nd->path.
 777  * Revalidate the final result, unless we'd already done that during
 778  * the path walk or the filesystem doesn't ask for it.  Return 0 on
 779  * success, -error on failure.  In case of failure caller does not
 780  * need to drop nd->path.
 781  */
 782 static int complete_walk(struct nameidata *nd)
 783 {
 784         struct dentry *dentry = nd->path.dentry;
 785         int status;
 786
 787         if (nd->flags & LOOKUP_RCU) {
 788                 if (!(nd->flags & LOOKUP_ROOT))
 789                         nd->root.mnt = NULL;
 790                 if (unlikely(unlazy_walk(nd)))
 791                         return -ECHILD;
 792         }
 793
 794         if (likely(!(nd->flags & LOOKUP_JUMPED)))
 795                 return 0;
 796
 797         if (likely(!(dentry->d_flags & DCACHE_OP_WEAK_REVALIDATE)))
 798                 return 0;
 799
 800         status = dentry->d_op->d_weak_revalidate(dentry, nd->flags);
 801         if (status > 0)
 802                 return 0;
 803
 804         if (!status)
 805                 status = -ESTALE;
 806
 807         return status;
 808 }
 809
 810 static void set_root(struct nameidata *nd)
 811 {
 812         struct fs_struct *fs = current->fs;
 813
 814         if (nd->flags & LOOKUP_RCU) {
 815                 unsigned seq;
 816
 817                 do {
 818                         seq = read_seqcount_begin(&fs->seq);
 819                         nd->root = fs->root;
 820                         nd->root_seq = __read_seqcount_begin(&nd->root.dentry->d_seq);
 821                 } while (read_seqcount_retry(&fs->seq, seq));
 822         } else {
 823                 get_fs_root(fs, &nd->root);
 824         }
 825 }
 826
 827 static void path_put_conditional(struct path *path, struct nameidata *nd)
 828 {
 829         dput(path->dentry);
 830         if (path->mnt != nd->path.mnt)
 831                 mntput(path->mnt);
 832 }
 833
 834 static inline void path_to_nameidata(const struct path *path,
 835                                         struct nameidata *nd)
 836 {
 837         if (!(nd->flags & LOOKUP_RCU)) {
 838                 dput(nd->path.dentry);
 839                 if (nd->path.mnt != path->mnt)
 840                         mntput(nd->path.mnt);
 841         }
 842         nd->path.mnt = path->mnt;
 843         nd->path.dentry = path->dentry;
 844 }
 845
 846 static int nd_jump_root(struct nameidata *nd)
 847 {
 848         if (nd->flags & LOOKUP_RCU) {
 849                 struct dentry *d;
 850                 nd->path = nd->root;
 851                 d = nd->path.dentry;
 852                 nd->inode = d->d_inode;
 853                 nd->seq = nd->root_seq;
 854                 if (unlikely(read_seqcount_retry(&d->d_seq, nd->seq)))
 855                         return -ECHILD;
 856         } else {
 857                 path_put(&nd->path);
 858                 nd->path = nd->root;
 859                 path_get(&nd->path);
 860                 nd->inode = nd->path.dentry->d_inode;
 861         }
 862         nd->flags |= LOOKUP_JUMPED;
 863         return 0;
 864 }
 865
 866 /*
 867  * Helper to directly jump to a known parsed path from ->get_link,
 868  * caller must have taken a reference to path beforehand.
 869  */
 870 void nd_jump_link(struct path *path)
 871 {
 872         struct nameidata *nd = current->nameidata;
 873         path_put(&nd->path);
 874
 875         nd->path = *path;
 876         nd->inode = nd->path.dentry->d_inode;
 877         nd->flags |= LOOKUP_JUMPED;
 878 }
 879
 880 static inline void put_link(struct nameidata *nd)
 881 {
 882         struct saved *last = nd->stack + --nd->depth;
 883         do_delayed_call(&last->done);
 884         if (!(nd->flags & LOOKUP_RCU))
 885                 path_put(&last->link);
 886 }
 887
 888 int sysctl_protected_symlinks __read_mostly = 0;
 889 int sysctl_protected_hardlinks __read_mostly = 0;
 890
 891 /**
 892  * may_follow_link - Check symlink following for unsafe situations
 893  * @nd: nameidata pathwalk data
 894  *
 895  * In the case of the sysctl_protected_symlinks sysctl being enabled,
 896  * CAP_DAC_OVERRIDE needs to be specifically ignored if the symlink is
 897  * in a sticky world-writable directory. This is to protect privileged
 898  * processes from failing races against path names that may change out
 899  * from under them by way of other users creating malicious symlinks.
 900  * It will permit symlinks to be followed only when outside a sticky
 901  * world-writable directory, or when the uid of the symlink and follower
 902  * match, or when the directory owner matches the symlink's owner.
 903  *
 904  * Returns 0 if following the symlink is allowed, -ve on error.
 905  */
 906 static inline int may_follow_link(struct nameidata *nd)
 907 {
 908         const struct inode *inode;
 909         const struct inode *parent;
 910         kuid_t puid;
 911
 912         if (!sysctl_protected_symlinks)
 913                 return 0;
 914
 915         /* Allowed if owner and follower match. */
 916         inode = nd->link_inode;
 917         if (uid_eq(current_cred()->fsuid, inode->i_uid))
 918                 return 0;
 919
 920         /* Allowed if parent directory not sticky and world-writable. */
 921         parent = nd->inode;
 922         if ((parent->i_mode & (S_ISVTX|S_IWOTH)) != (S_ISVTX|S_IWOTH))
 923                 return 0;
 924
 925         /* Allowed if parent directory and link owner match. */
 926         puid = parent->i_uid;
 927         if (uid_valid(puid) && uid_eq(puid, inode->i_uid))
 928                 return 0;
 929
 930         if (nd->flags & LOOKUP_RCU)
 931                 return -ECHILD;
 932
 933         audit_inode(nd->name, nd->stack[0].link.dentry, 0);
 934         audit_log_link_denied("follow_link");
 935         return -EACCES;
 936 }
 937
 938 /**
 939  * safe_hardlink_source - Check for safe hardlink conditions
 940  * @inode: the source inode to hardlink from
 941  *
 942  * Return false if at least one of the following conditions:
 943  *    - inode is not a regular file
 944  *    - inode is setuid
 945  *    - inode is setgid and group-exec
 946  *    - access failure for read and write
 947  *
 948  * Otherwise returns true.
 949  */
 950 static bool safe_hardlink_source(struct inode *inode)
 951 {
 952         umode_t mode = inode->i_mode;
 953
 954         /* Special files should not get pinned to the filesystem. */
 955         if (!S_ISREG(mode))
 956                 return false;
 957
 958         /* Setuid files should not get pinned to the filesystem. */
 959         if (mode & S_ISUID)
 960                 return false;
 961
 962         /* Executable setgid files should not get pinned to the filesystem. */
 963         if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP))
 964                 return false;
 965
 966         /* Hardlinking to unreadable or unwritable sources is dangerous. */
 967         if (inode_permission(inode, MAY_READ | MAY_WRITE))
 968                 return false;
 969
 970         return true;
 971 }
 972
 973 /**
 974  * may_linkat - Check permissions for creating a hardlink
 975  * @link: the source to hardlink from
 976  *
 977  * Block hardlink when all of:
 978  *  - sysctl_protected_hardlinks enabled
 979  *  - fsuid does not match inode
 980  *  - hardlink source is unsafe (see safe_hardlink_source() above)
 981  *  - not CAP_FOWNER in a namespace with the inode owner uid mapped
 982  *
 983  * Returns 0 if successful, -ve on error.
 984  */
 985 static int may_linkat(struct path *link)
 986 {
 987         struct inode *inode;
 988
 989         if (!sysctl_protected_hardlinks)
 990                 return 0;
 991
 992         inode = link->dentry->d_inode;
 993
 994         /* Source inode owner (or CAP_FOWNER) can hardlink all they like,
 995          * otherwise, it must be a safe source.
 996          */
 997         if (safe_hardlink_source(inode) || inode_owner_or_capable(inode))
 998                 return 0;
 999
1000         audit_log_link_denied("linkat");
1001         return -EPERM;
1002 }
1003
1004 static __always_inline
1005 const char *get_link(struct nameidata *nd)
1006 {
1007         struct saved *last = nd->stack + nd->depth - 1;
1008         struct dentry *dentry = last->link.dentry;
1009         struct inode *inode = nd->link_inode;
1010         int error;
1011         const char *res;
1012
1013         if (!(nd->flags & LOOKUP_RCU)) {
1014                 touch_atime(&last->link);
1015                 cond_resched();
1016         } else if (atime_needs_update_rcu(&last->link, inode)) {
1017                 if (unlikely(unlazy_walk(nd)))
1018                         return ERR_PTR(-ECHILD);
1019                 touch_atime(&last->link);
1020         }
1021
1022         error = security_inode_follow_link(dentry, inode,
1023                                            nd->flags & LOOKUP_RCU);
1024         if (unlikely(error))
1025                 return ERR_PTR(error);
1026
1027         nd->last_type = LAST_BIND;
1028         res = inode->i_link;
1029         if (!res) {
1030                 const char * (*get)(struct dentry *, struct inode *,
1031                                 struct delayed_call *);
1032                 get = inode->i_op->get_link;
1033                 if (nd->flags & LOOKUP_RCU) {
1034                         res = get(NULL, inode, &last->done);
1035                         if (res == ERR_PTR(-ECHILD)) {
1036                                 if (unlikely(unlazy_walk(nd)))
1037                                         return ERR_PTR(-ECHILD);
1038                                 res = get(dentry, inode, &last->done);
1039                         }
1040                 } else {
1041                         res = get(dentry, inode, &last->done);
1042                 }
1043                 if (IS_ERR_OR_NULL(res))
1044                         return res;
1045         }
1046         if (*res == '/') {
1047                 if (!nd->root.mnt)
1048                         set_root(nd);
1049                 if (unlikely(nd_jump_root(nd)))
1050                         return ERR_PTR(-ECHILD);
1051                 while (unlikely(*++res == '/'))
1052                         ;
1053         }
1054         if (!*res)
1055                 res = NULL;
1056         return res;
1057 }
1058
1059 /*
1060  * follow_up - Find the mountpoint of path's vfsmount
1061  *
1062  * Given a path, find the mountpoint of its source file system.
1063  * Replace @path with the path of the mountpoint in the parent mount.
1064  * Up is towards /.
1065  *
1066  * Return 1 if we went up a level and 0 if we were already at the
1067  * root.
1068  */
1069 int follow_up(struct path *path)
1070 {
1071         struct mount *mnt = real_mount(path->mnt);
1072         struct mount *parent;
1073         struct dentry *mountpoint;
1074
1075         read_seqlock_excl(&mount_lock);
1076         parent = mnt->mnt_parent;
1077         if (parent == mnt) {
1078                 read_sequnlock_excl(&mount_lock);
1079                 return 0;
1080         }
1081         mntget(&parent->mnt);
1082         mountpoint = dget(mnt->mnt_mountpoint);
1083         read_sequnlock_excl(&mount_lock);
1084         dput(path->dentry);
1085         path->dentry = mountpoint;
1086         mntput(path->mnt);
1087         path->mnt = &parent->mnt;
1088         return 1;
1089 }
1090 EXPORT_SYMBOL(follow_up);
1091
1092 /*
1093  * Perform an automount
1094  * - return -EISDIR to tell follow_managed() to stop and return the path we
1095  *   were called with.
1096  */
1097 static int follow_automount(struct path *path, struct nameidata *nd,
1098                             bool *need_mntput)
1099 {
1100         struct vfsmount *mnt;
1101         int err;
1102
1103         if (!path->dentry->d_op || !path->dentry->d_op->d_automount)
1104                 return -EREMOTE;
1105
1106         /* We don't want to mount if someone's just doing a stat -
1107          * unless they're stat'ing a directory and appended a '/' to
1108          * the name.
1109          *
1110          * We do, however, want to mount if someone wants to open or
1111          * create a file of any type under the mountpoint, wants to
1112          * traverse through the mountpoint or wants to open the
1113          * mounted directory.  Also, autofs may mark negative dentries
1114          * as being automount points.  These will need the attentions
1115          * of the daemon to instantiate them before they can be used.
1116          */
1117         if (!(nd->flags & (LOOKUP_PARENT | LOOKUP_DIRECTORY |
1118                            LOOKUP_OPEN | LOOKUP_CREATE | LOOKUP_AUTOMOUNT)) &&
1119             path->dentry->d_inode)
1120                 return -EISDIR;
1121
1122         nd->total_link_count++;
1123         if (nd->total_link_count >= 40)
1124                 return -ELOOP;
1125
1126         mnt = path->dentry->d_op->d_automount(path);
1127         if (IS_ERR(mnt)) {
1128                 /*
1129                  * The filesystem is allowed to return -EISDIR here to indicate
1130                  * it doesn't want to automount.  For instance, autofs would do
1131                  * this so that its userspace daemon can mount on this dentry.
1132                  *
1133                  * However, we can only permit this if it's a terminal point in
1134                  * the path being looked up; if it wasn't then the remainder of
1135                  * the path is inaccessible and we should say so.
1136                  */
1137                 if (PTR_ERR(mnt) == -EISDIR && (nd->flags & LOOKUP_PARENT))
1138                         return -EREMOTE;
1139                 return PTR_ERR(mnt);
1140         }
1141
1142         if (!mnt) /* mount collision */
1143                 return 0;
1144
1145         if (!*need_mntput) {
1146                 /* lock_mount() may release path->mnt on error */
1147                 mntget(path->mnt);
1148                 *need_mntput = true;
1149         }
1150         err = finish_automount(mnt, path);
1151
1152         switch (err) {
1153         case -EBUSY:
1154                 /* Someone else made a mount here whilst we were busy */
1155                 return 0;
1156         case 0:
1157                 path_put(path);
1158                 path->mnt = mnt;
1159                 path->dentry = dget(mnt->mnt_root);
1160                 return 0;
1161         default:
1162                 return err;
1163         }
1164
1165 }
1166
1167 /*
1168  * Handle a dentry that is managed in some way.
1169  * - Flagged for transit management (autofs)
1170  * - Flagged as mountpoint
1171  * - Flagged as automount point
1172  *
1173  * This may only be called in refwalk mode.
1174  *
1175  * Serialization is taken care of in namespace.c
1176  */
1177 static int follow_managed(struct path *path, struct nameidata *nd)
1178 {
1179         struct vfsmount *mnt = path->mnt; /* held by caller, must be left alone */
1180         unsigned managed;
1181         bool need_mntput = false;
1182         int ret = 0;
1183
1184         /* Given that we're not holding a lock here, we retain the value in a
1185          * local variable for each dentry as we look at it so that we don't see
1186          * the components of that value change under us */
1187         while (managed = READ_ONCE(path->dentry->d_flags),
1188                managed &= DCACHE_MANAGED_DENTRY,
1189                unlikely(managed != 0)) {
1190                 /* Allow the filesystem to manage the transit without i_mutex
1191                  * being held. */
1192                 if (managed & DCACHE_MANAGE_TRANSIT) {
1193                         BUG_ON(!path->dentry->d_op);
1194                         BUG_ON(!path->dentry->d_op->d_manage);
1195                         ret = path->dentry->d_op->d_manage(path, false);
1196                         if (ret < 0)
1197                                 break;
1198                 }
1199
1200                 /* Transit to a mounted filesystem. */
1201                 if (managed & DCACHE_MOUNTED) {
1202                         struct vfsmount *mounted = lookup_mnt(path);
1203                         if (mounted) {
1204                                 dput(path->dentry);
1205                                 if (need_mntput)
1206                                         mntput(path->mnt);
1207                                 path->mnt = mounted;
1208                                 path->dentry = dget(mounted->mnt_root);
1209                                 need_mntput = true;
1210                                 continue;
1211                         }
1212
1213                         /* Something is mounted on this dentry in another
1214                          * namespace and/or whatever was mounted there in this
1215                          * namespace got unmounted before lookup_mnt() could
1216                          * get it */
1217                 }
1218
1219                 /* Handle an automount point */
1220                 if (managed & DCACHE_NEED_AUTOMOUNT) {
1221                         ret = follow_automount(path, nd, &need_mntput);
1222                         if (ret < 0)
1223                                 break;
1224                         continue;
1225                 }
1226
1227                 /* We didn't change the current path point */
1228                 break;
1229         }
1230
1231         if (need_mntput && path->mnt == mnt)
1232                 mntput(path->mnt);
1233         if (ret == -EISDIR || !ret)
1234                 ret = 1;
1235         if (need_mntput)
1236                 nd->flags |= LOOKUP_JUMPED;
1237         if (unlikely(ret < 0))
1238                 path_put_conditional(path, nd);
1239         return ret;
1240 }
1241
1242 int follow_down_one(struct path *path)
1243 {
1244         struct vfsmount *mounted;
1245
1246         mounted = lookup_mnt(path);
1247         if (mounted) {
1248                 dput(path->dentry);
1249                 mntput(path->mnt);
1250                 path->mnt = mounted;
1251                 path->dentry = dget(mounted->mnt_root);
1252                 return 1;
1253         }
1254         return 0;
1255 }
1256 EXPORT_SYMBOL(follow_down_one);
1257
1258 static inline int managed_dentry_rcu(const struct path *path)
1259 {
1260         return (path->dentry->d_flags & DCACHE_MANAGE_TRANSIT) ?
1261                 path->dentry->d_op->d_manage(path, true) : 0;
1262 }
1263
1264 /*
1265  * Try to skip to top of mountpoint pile in rcuwalk mode.  Fail if
1266  * we meet a managed dentry that would need blocking.
1267  */
1268 static bool __follow_mount_rcu(struct nameidata *nd, struct path *path,
1269                                struct inode **inode, unsigned *seqp)
1270 {
1271         for (;;) {
1272                 struct mount *mounted;
1273                 /*
1274                  * Don't forget we might have a non-mountpoint managed dentry
1275                  * that wants to block transit.
1276                  */
1277                 switch (managed_dentry_rcu(path)) {
1278                 case -ECHILD:
1279                 default:
1280                         return false;
1281                 case -EISDIR:
1282                         return true;
1283                 case 0:
1284                         break;
1285                 }
1286
1287                 if (!d_mountpoint(path->dentry))
1288                         return !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1289
1290                 mounted = __lookup_mnt(path->mnt, path->dentry);
1291                 if (!mounted)
1292                         break;
1293                 path->mnt = &mounted->mnt;
1294                 path->dentry = mounted->mnt.mnt_root;
1295                 nd->flags |= LOOKUP_JUMPED;
1296                 *seqp = read_seqcount_begin(&path->dentry->d_seq);
1297                 /*
1298                  * Update the inode too. We don't need to re-check the
1299                  * dentry sequence number here after this d_inode read,
1300                  * because a mount-point is always pinned.
1301                  */
1302                 *inode = path->dentry->d_inode;
1303         }
1304         return !read_seqretry(&mount_lock, nd->m_seq) &&
1305                 !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1306 }
1307
1308 static int follow_dotdot_rcu(struct nameidata *nd)
1309 {
1310         struct inode *inode = nd->inode;
1311
1312         while (1) {
1313                 if (path_equal(&nd->path, &nd->root))
1314                         break;
1315                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1316                         struct dentry *old = nd->path.dentry;
1317                         struct dentry *parent = old->d_parent;
1318                         unsigned seq;
1319
1320                         inode = parent->d_inode;
1321                         seq = read_seqcount_begin(&parent->d_seq);
1322                         if (unlikely(read_seqcount_retry(&old->d_seq, nd->seq)))
1323                                 return -ECHILD;
1324                         nd->path.dentry = parent;
1325                         nd->seq = seq;
1326                         if (unlikely(!path_connected(&nd->path)))
1327                                 return -ENOENT;
1328                         break;
1329                 } else {
1330                         struct mount *mnt = real_mount(nd->path.mnt);
1331                         struct mount *mparent = mnt->mnt_parent;
1332                         struct dentry *mountpoint = mnt->mnt_mountpoint;
1333                         struct inode *inode2 = mountpoint->d_inode;
1334                         unsigned seq = read_seqcount_begin(&mountpoint->d_seq);
1335                         if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1336                                 return -ECHILD;
1337                         if (&mparent->mnt == nd->path.mnt)
1338                                 break;
1339                         /* we know that mountpoint was pinned */
1340                         nd->path.dentry = mountpoint;
1341                         nd->path.mnt = &mparent->mnt;
1342                         inode = inode2;
1343                         nd->seq = seq;
1344                 }
1345         }
1346         while (unlikely(d_mountpoint(nd->path.dentry))) {
1347                 struct mount *mounted;
1348                 mounted = __lookup_mnt(nd->path.mnt, nd->path.dentry);
1349                 if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1350                         return -ECHILD;
1351                 if (!mounted)
1352                         break;
1353                 nd->path.mnt = &mounted->mnt;
1354                 nd->path.dentry = mounted->mnt.mnt_root;
1355                 inode = nd->path.dentry->d_inode;
1356                 nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
1357         }
1358         nd->inode = inode;
1359         return 0;
1360 }
1361
1362 /*
1363  * Follow down to the covering mount currently visible to userspace.  At each
1364  * point, the filesystem owning that dentry may be queried as to whether the
1365  * caller is permitted to proceed or not.
1366  */
1367 int follow_down(struct path *path)
1368 {
1369         unsigned managed;
1370         int ret;
1371
1372         while (managed = READ_ONCE(path->dentry->d_flags),
1373                unlikely(managed & DCACHE_MANAGED_DENTRY)) {
1374                 /* Allow the filesystem to manage the transit without i_mutex
1375                  * being held.
1376                  *
1377                  * We indicate to the filesystem if someone is trying to mount
1378                  * something here.  This gives autofs the chance to deny anyone
1379                  * other than its daemon the right to mount on its
1380                  * superstructure.
1381                  *
1382                  * The filesystem may sleep at this point.
1383                  */
1384                 if (managed & DCACHE_MANAGE_TRANSIT) {
1385                         BUG_ON(!path->dentry->d_op);
1386                         BUG_ON(!path->dentry->d_op->d_manage);
1387                         ret = path->dentry->d_op->d_manage(path, false);
1388                         if (ret < 0)
1389                                 return ret == -EISDIR ? 0 : ret;
1390                 }
1391
1392                 /* Transit to a mounted filesystem. */
1393                 if (managed & DCACHE_MOUNTED) {
1394                         struct vfsmount *mounted = lookup_mnt(path);
1395                         if (!mounted)
1396                                 break;
1397                         dput(path->dentry);
1398                         mntput(path->mnt);
1399                         path->mnt = mounted;
1400                         path->dentry = dget(mounted->mnt_root);
1401                         continue;
1402                 }
1403
1404                 /* Don't handle automount points here */
1405                 break;
1406         }
1407         return 0;
1408 }
1409 EXPORT_SYMBOL(follow_down);
1410
1411 /*
1412  * Skip to top of mountpoint pile in refwalk mode for follow_dotdot()
1413  */
1414 static void follow_mount(struct path *path)
1415 {
1416         while (d_mountpoint(path->dentry)) {
1417                 struct vfsmount *mounted = lookup_mnt(path);
1418                 if (!mounted)
1419                         break;
1420                 dput(path->dentry);
1421                 mntput(path->mnt);
1422                 path->mnt = mounted;
1423                 path->dentry = dget(mounted->mnt_root);
1424         }
1425 }
1426
1427 static int path_parent_directory(struct path *path)
1428 {
1429         struct dentry *old = path->dentry;
1430         /* rare case of legitimate dget_parent()... */
1431         path->dentry = dget_parent(path->dentry);
1432         dput(old);
1433         if (unlikely(!path_connected(path)))
1434                 return -ENOENT;
1435         return 0;
1436 }
1437
1438 static int follow_dotdot(struct nameidata *nd)
1439 {
1440         while(1) {
1441                 if (nd->path.dentry == nd->root.dentry &&
1442                     nd->path.mnt == nd->root.mnt) {
1443                         break;
1444                 }
1445                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1446                         int ret = path_parent_directory(&nd->path);
1447                         if (ret)
1448                                 return ret;
1449                         break;
1450                 }
1451                 if (!follow_up(&nd->path))
1452                         break;
1453         }
1454         follow_mount(&nd->path);
1455         nd->inode = nd->path.dentry->d_inode;
1456         return 0;
1457 }
1458
1459 /*
1460  * This looks up the name in dcache and possibly revalidates the found dentry.
1461  * NULL is returned if the dentry does not exist in the cache.
1462  */
1463 static struct dentry *lookup_dcache(const struct qstr *name,
1464                                     struct dentry *dir,
1465                                     unsigned int flags)
1466 {
1467         struct dentry *dentry = d_lookup(dir, name);
1468         if (dentry) {
1469                 int error = d_revalidate(dentry, flags);
1470                 if (unlikely(error <= 0)) {
1471                         if (!error)
1472                                 d_invalidate(dentry);
1473                         dput(dentry);
1474                         return ERR_PTR(error);
1475                 }
1476         }
1477         return dentry;
1478 }
1479
1480 /*
1481  * Parent directory has inode locked exclusive.  This is one
1482  * and only case when ->lookup() gets called on non in-lookup
1483  * dentries - as the matter of fact, this only gets called
1484  * when directory is guaranteed to have no in-lookup children
1485  * at all.
1486  */
1487 static struct dentry *__lookup_hash(const struct qstr *name,
1488                 struct dentry *base, unsigned int flags)
1489 {
1490         struct dentry *dentry = lookup_dcache(name, base, flags);
1491         struct dentry *old;
1492         struct inode *dir = base->d_inode;
1493
1494         if (dentry)
1495                 return dentry;
1496
1497         /* Don't create child dentry for a dead directory. */
1498         if (unlikely(IS_DEADDIR(dir)))
1499                 return ERR_PTR(-ENOENT);
1500
1501         dentry = d_alloc(base, name);
1502         if (unlikely(!dentry))
1503                 return ERR_PTR(-ENOMEM);
1504
1505         old = dir->i_op->lookup(dir, dentry, flags);
1506         if (unlikely(old)) {
1507                 dput(dentry);
1508                 dentry = old;
1509         }
1510         return dentry;
1511 }
1512
1513 static int lookup_fast(struct nameidata *nd,
1514                        struct path *path, struct inode **inode,
1515                        unsigned *seqp)
1516 {
1517         struct vfsmount *mnt = nd->path.mnt;
1518         struct dentry *dentry, *parent = nd->path.dentry;
1519         int status = 1;
1520         int err;
1521
1522         /*
1523          * Rename seqlock is not required here because in the off chance
1524          * of a false negative due to a concurrent rename, the caller is
1525          * going to fall back to non-racy lookup.
1526          */
1527         if (nd->flags & LOOKUP_RCU) {
1528                 unsigned seq;
1529                 bool negative;
1530                 dentry = __d_lookup_rcu(parent, &nd->last, &seq);
1531                 if (unlikely(!dentry)) {
1532                         if (unlazy_walk(nd))
1533                                 return -ECHILD;
1534                         return 0;
1535                 }
1536
1537                 /*
1538                  * This sequence count validates that the inode matches
1539                  * the dentry name information from lookup.
1540                  */
1541                 *inode = d_backing_inode(dentry);
1542                 negative = d_is_negative(dentry);
1543                 if (unlikely(read_seqcount_retry(&dentry->d_seq, seq)))
1544                         return -ECHILD;
1545
1546                 /*
1547                  * This sequence count validates that the parent had no
1548                  * changes while we did the lookup of the dentry above.
1549                  *
1550                  * The memory barrier in read_seqcount_begin of child is
1551                  *  enough, we can use __read_seqcount_retry here.
1552                  */
1553                 if (unlikely(__read_seqcount_retry(&parent->d_seq, nd->seq)))
1554                         return -ECHILD;
1555
1556                 *seqp = seq;
1557                 status = d_revalidate(dentry, nd->flags);
1558                 if (likely(status > 0)) {
1559                         /*
1560                          * Note: do negative dentry check after revalidation in
1561                          * case that drops it.
1562                          */
1563                         if (unlikely(negative))
1564                                 return -ENOENT;
1565                         path->mnt = mnt;
1566                         path->dentry = dentry;
1567                         if (likely(__follow_mount_rcu(nd, path, inode, seqp)))
1568                                 return 1;
1569                 }
1570                 if (unlazy_child(nd, dentry, seq))
1571                         return -ECHILD;
1572                 if (unlikely(status == -ECHILD))
1573                         /* we'd been told to redo it in non-rcu mode */
1574                         status = d_revalidate(dentry, nd->flags);
1575         } else {
1576                 dentry = __d_lookup(parent, &nd->last);
1577                 if (unlikely(!dentry))
1578                         return 0;
1579                 status = d_revalidate(dentry, nd->flags);
1580         }
1581         if (unlikely(status <= 0)) {
1582                 if (!status)
1583                         d_invalidate(dentry);
1584                 dput(dentry);
1585                 return status;
1586         }
1587         if (unlikely(d_is_negative(dentry))) {
1588                 dput(dentry);
1589                 return -ENOENT;
1590         }
1591
1592         path->mnt = mnt;
1593         path->dentry = dentry;
1594         err = follow_managed(path, nd);
1595         if (likely(err > 0))
1596                 *inode = d_backing_inode(path->dentry);
1597         return err;
1598 }
1599
1600 /* Fast lookup failed, do it the slow way */
1601 static struct dentry *__lookup_slow(const struct qstr *name,
1602                                     struct dentry *dir,
1603                                     unsigned int flags)
1604 {
1605         struct dentry *dentry, *old;
1606         struct inode *inode = dir->d_inode;
1607         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
1608
1609         /* Don't go there if it's already dead */
1610         if (unlikely(IS_DEADDIR(inode)))
1611                 return ERR_PTR(-ENOENT);
1612 again:
1613         dentry = d_alloc_parallel(dir, name, &wq);
1614         if (IS_ERR(dentry))
1615                 return dentry;
1616         if (unlikely(!d_in_lookup(dentry))) {
1617                 if (!(flags & LOOKUP_NO_REVAL)) {
1618                         int error = d_revalidate(dentry, flags);
1619                         if (unlikely(error <= 0)) {
1620                                 if (!error) {
1621                                         d_invalidate(dentry);
1622                                         dput(dentry);
1623                                         goto again;
1624                                 }
1625                                 dput(dentry);
1626                                 dentry = ERR_PTR(error);
1627                         }
1628                 }
1629         } else {
1630                 old = inode->i_op->lookup(inode, dentry, flags);
1631                 d_lookup_done(dentry);
1632                 if (unlikely(old)) {
1633                         dput(dentry);
1634                         dentry = old;
1635                 }
1636         }
1637         return dentry;
1638 }
1639
1640 static struct dentry *lookup_slow(const struct qstr *name,
1641                                   struct dentry *dir,
1642                                   unsigned int flags)
1643 {
1644         struct inode *inode = dir->d_inode;
1645         struct dentry *res;
1646         inode_lock_shared(inode);
1647         res = __lookup_slow(name, dir, flags);
1648         inode_unlock_shared(inode);
1649         return res;
1650 }
1651
1652 static inline int may_lookup(struct nameidata *nd)
1653 {
1654         if (nd->flags & LOOKUP_RCU) {
1655                 int err = inode_permission(nd->inode, MAY_EXEC|MAY_NOT_BLOCK);
1656                 if (err != -ECHILD)
1657                         return err;
1658                 if (unlazy_walk(nd))
1659                         return -ECHILD;
1660         }
1661         return inode_permission(nd->inode, MAY_EXEC);
1662 }
1663
1664 static inline int handle_dots(struct nameidata *nd, int type)
1665 {
1666         if (type == LAST_DOTDOT) {
1667                 if (!nd->root.mnt)
1668                         set_root(nd);
1669                 if (nd->flags & LOOKUP_RCU) {
1670                         return follow_dotdot_rcu(nd);
1671                 } else
1672                         return follow_dotdot(nd);
1673         }
1674         return 0;
1675 }
1676
1677 static int pick_link(struct nameidata *nd, struct path *link,
1678                      struct inode *inode, unsigned seq)
1679 {
1680         int error;
1681         struct saved *last;
1682         if (unlikely(nd->total_link_count++ >= MAXSYMLINKS)) {
1683                 path_to_nameidata(link, nd);
1684                 return -ELOOP;
1685         }
1686         if (!(nd->flags & LOOKUP_RCU)) {
1687                 if (link->mnt == nd->path.mnt)
1688                         mntget(link->mnt);
1689         }
1690         error = nd_alloc_stack(nd);
1691         if (unlikely(error)) {
1692                 if (error == -ECHILD) {
1693                         if (unlikely(!legitimize_path(nd, link, seq))) {
1694                                 drop_links(nd);
1695                                 nd->depth = 0;
1696                                 nd->flags &= ~LOOKUP_RCU;
1697                                 nd->path.mnt = NULL;
1698                                 nd->path.dentry = NULL;
1699                                 if (!(nd->flags & LOOKUP_ROOT))
1700                                         nd->root.mnt = NULL;
1701                                 rcu_read_unlock();
1702                         } else if (likely(unlazy_walk(nd)) == 0)
1703                                 error = nd_alloc_stack(nd);
1704                 }
1705                 if (error) {
1706                         path_put(link);
1707                         return error;
1708                 }
1709         }
1710
1711         last = nd->stack + nd->depth++;
1712         last->link = *link;
1713         clear_delayed_call(&last->done);
1714         nd->link_inode = inode;
1715         last->seq = seq;
1716         return 1;
1717 }
1718
1719 enum {WALK_FOLLOW = 1, WALK_MORE = 2};
1720
1721 /*
1722  * Do we need to follow links? We _really_ want to be able
1723  * to do this check without having to look at inode->i_op,
1724  * so we keep a cache of "no, this doesn't need follow_link"
1725  * for the common case.
1726  */
1727 static inline int step_into(struct nameidata *nd, struct path *path,
1728                             int flags, struct inode *inode, unsigned seq)
1729 {
1730         if (!(flags & WALK_MORE) && nd->depth)
1731                 put_link(nd);
1732         if (likely(!d_is_symlink(path->dentry)) ||
1733            !(flags & WALK_FOLLOW || nd->flags & LOOKUP_FOLLOW)) {
1734                 /* not a symlink or should not follow */
1735                 path_to_nameidata(path, nd);
1736                 nd->inode = inode;
1737                 nd->seq = seq;
1738                 return 0;
1739         }
1740         /* make sure that d_is_symlink above matches inode */
1741         if (nd->flags & LOOKUP_RCU) {
1742                 if (read_seqcount_retry(&path->dentry->d_seq, seq))
1743                         return -ECHILD;
1744         }
1745         return pick_link(nd, path, inode, seq);
1746 }
1747
1748 static int walk_component(struct nameidata *nd, int flags)
1749 {
1750         struct path path;
1751         struct inode *inode;
1752         unsigned seq;
1753         int err;
1754         /*
1755          * "." and ".." are special - ".." especially so because it has
1756          * to be able to know about the current root directory and
1757          * parent relationships.
1758          */
1759         if (unlikely(nd->last_type != LAST_NORM)) {
1760                 err = handle_dots(nd, nd->last_type);
1761                 if (!(flags & WALK_MORE) && nd->depth)
1762                         put_link(nd);
1763                 return err;
1764         }
1765         err = lookup_fast(nd, &path, &inode, &seq);
1766         if (unlikely(err <= 0)) {
1767                 if (err < 0)
1768                         return err;
1769                 path.dentry = lookup_slow(&nd->last, nd->path.dentry,
1770                                           nd->flags);
1771                 if (IS_ERR(path.dentry))
1772                         return PTR_ERR(path.dentry);
1773
1774                 path.mnt = nd->path.mnt;
1775                 err = follow_managed(&path, nd);
1776                 if (unlikely(err < 0))
1777                         return err;
1778
1779                 if (unlikely(d_is_negative(path.dentry))) {
1780                         path_to_nameidata(&path, nd);
1781                         return -ENOENT;
1782                 }
1783
1784                 seq = 0;        /* we are already out of RCU mode */
1785                 inode = d_backing_inode(path.dentry);
1786         }
1787
1788         return step_into(nd, &path, flags, inode, seq);
1789 }
1790
1791 /*
1792  * We can do the critical dentry name comparison and hashing
1793  * operations one word at a time, but we are limited to:
1794  *
1795  * - Architectures with fast unaligned word accesses. We could
1796  *   do a "get_unaligned()" if this helps and is sufficiently
1797  *   fast.
1798  *
1799  * - non-CONFIG_DEBUG_PAGEALLOC configurations (so that we
1800  *   do not trap on the (extremely unlikely) case of a page
1801  *   crossing operation.
1802  *
1803  * - Furthermore, we need an efficient 64-bit compile for the
1804  *   64-bit case in order to generate the "number of bytes in
1805  *   the final mask". Again, that could be replaced with a
1806  *   efficient population count instruction or similar.
1807  */
1808 #ifdef CONFIG_DCACHE_WORD_ACCESS
1809
1810 #include <asm/word-at-a-time.h>
1811
1812 #ifdef HASH_MIX
1813
1814 /* Architecture provides HASH_MIX and fold_hash() in <asm/hash.h> */
1815
1816 #elif defined(CONFIG_64BIT)
1817 /*
1818  * Register pressure in the mixing function is an issue, particularly
1819  * on 32-bit x86, but almost any function requires one state value and
1820  * one temporary.  Instead, use a function designed for two state values
1821  * and no temporaries.
1822  *
1823  * This function cannot create a collision in only two iterations, so
1824  * we have two iterations to achieve avalanche.  In those two iterations,
1825  * we have six layers of mixing, which is enough to spread one bit's
1826  * influence out to 2^6 = 64 state bits.
1827  *
1828  * Rotate constants are scored by considering either 64 one-bit input
1829  * deltas or 64*63/2 = 2016 two-bit input deltas, and finding the
1830  * probability of that delta causing a change to each of the 128 output
1831  * bits, using a sample of random initial states.
1832  *
1833  * The Shannon entropy of the computed probabilities is then summed
1834  * to produce a score.  Ideally, any input change has a 50% chance of
1835  * toggling any given output bit.
1836  *
1837  * Mixing scores (in bits) for (12,45):
1838  * Input delta: 1-bit      2-bit
1839  * 1 round:     713.3    42542.6
1840  * 2 rounds:   2753.7   140389.8
1841  * 3 rounds:   5954.1   233458.2
1842  * 4 rounds:   7862.6   256672.2
1843  * Perfect:    8192     258048
1844  *            (64*128) (64*63/2 * 128)
1845  */
1846 #define HASH_MIX(x, y, a)       \
1847         (       x ^= (a),       \
1848         y ^= x, x = rol64(x,12),\
1849         x += y, y = rol64(y,45),\
1850         y *= 9                  )
1851
1852 /*
1853  * Fold two longs into one 32-bit hash value.  This must be fast, but
1854  * latency isn't quite as critical, as there is a fair bit of additional
1855  * work done before the hash value is used.
1856  */
1857 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1858 {
1859         y ^= x * GOLDEN_RATIO_64;
1860         y *= GOLDEN_RATIO_64;
1861         return y >> 32;
1862 }
1863
1864 #else   /* 32-bit case */
1865
1866 /*
1867  * Mixing scores (in bits) for (7,20):
1868  * Input delta: 1-bit      2-bit
1869  * 1 round:     330.3     9201.6
1870  * 2 rounds:   1246.4    25475.4
1871  * 3 rounds:   1907.1    31295.1
1872  * 4 rounds:   2042.3    31718.6
1873  * Perfect:    2048      31744
1874  *            (32*64)   (32*31/2 * 64)
1875  */
1876 #define HASH_MIX(x, y, a)       \
1877         (       x ^= (a),       \
1878         y ^= x, x = rol32(x, 7),\
1879         x += y, y = rol32(y,20),\
1880         y *= 9                  )
1881
1882 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1883 {
1884         /* Use arch-optimized multiply if one exists */
1885         return __hash_32(y ^ __hash_32(x));
1886 }
1887
1888 #endif
1889
1890 /*
1891  * Return the hash of a string of known length.  This is carfully
1892  * designed to match hash_name(), which is the more critical function.
1893  * In particular, we must end by hashing a final word containing 0..7
1894  * payload bytes, to match the way that hash_name() iterates until it
1895  * finds the delimiter after the name.
1896  */
1897 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1898 {
1899         unsigned long a, x = 0, y = (unsigned long)salt;
1900
1901         for (;;) {
1902                 if (!len)
1903                         goto done;
1904                 a = load_unaligned_zeropad(name);
1905                 if (len < sizeof(unsigned long))
1906                         break;
1907                 HASH_MIX(x, y, a);
1908                 name += sizeof(unsigned long);
1909                 len -= sizeof(unsigned long);
1910         }
1911         x ^= a & bytemask_from_count(len);
1912 done:
1913         return fold_hash(x, y);
1914 }
1915 EXPORT_SYMBOL(full_name_hash);
1916
1917 /* Return the "hash_len" (hash and length) of a null-terminated string */
1918 u64 hashlen_string(const void *salt, const char *name)
1919 {
1920         unsigned long a = 0, x = 0, y = (unsigned long)salt;
1921         unsigned long adata, mask, len;
1922         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1923
1924         len = 0;
1925         goto inside;
1926
1927         do {
1928                 HASH_MIX(x, y, a);
1929                 len += sizeof(unsigned long);
1930 inside:
1931                 a = load_unaligned_zeropad(name+len);
1932         } while (!has_zero(a, &adata, &constants));
1933
1934         adata = prep_zero_mask(a, adata, &constants);
1935         mask = create_zero_mask(adata);
1936         x ^= a & zero_bytemask(mask);
1937
1938         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1939 }
1940 EXPORT_SYMBOL(hashlen_string);
1941
1942 /*
1943  * Calculate the length and hash of the path component, and
1944  * return the "hash_len" as the result.
1945  */
1946 static inline u64 hash_name(const void *salt, const char *name)
1947 {
1948         unsigned long a = 0, b, x = 0, y = (unsigned long)salt;
1949         unsigned long adata, bdata, mask, len;
1950         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1951
1952         len = 0;
1953         goto inside;
1954
1955         do {
1956                 HASH_MIX(x, y, a);
1957                 len += sizeof(unsigned long);
1958 inside:
1959                 a = load_unaligned_zeropad(name+len);
1960                 b = a ^ REPEAT_BYTE('/');
1961         } while (!(has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)));
1962
1963         adata = prep_zero_mask(a, adata, &constants);
1964         bdata = prep_zero_mask(b, bdata, &constants);
1965         mask = create_zero_mask(adata | bdata);
1966         x ^= a & zero_bytemask(mask);
1967
1968         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1969 }
1970
1971 #else   /* !CONFIG_DCACHE_WORD_ACCESS: Slow, byte-at-a-time version */
1972
1973 /* Return the hash of a string of known length */
1974 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1975 {
1976         unsigned long hash = init_name_hash(salt);
1977         while (len--)
1978                 hash = partial_name_hash((unsigned char)*name++, hash);
1979         return end_name_hash(hash);
1980 }
1981 EXPORT_SYMBOL(full_name_hash);
1982
1983 /* Return the "hash_len" (hash and length) of a null-terminated string */
1984 u64 hashlen_string(const void *salt, const char *name)
1985 {
1986         unsigned long hash = init_name_hash(salt);
1987         unsigned long len = 0, c;
1988
1989         c = (unsigned char)*name;
1990         while (c) {
1991                 len++;
1992                 hash = partial_name_hash(c, hash);
1993                 c = (unsigned char)name[len];
1994         }
1995         return hashlen_create(end_name_hash(hash), len);
1996 }
1997 EXPORT_SYMBOL(hashlen_string);
1998
1999 /*
2000  * We know there's a real path component here of at least
2001  * one character.
2002  */
2003 static inline u64 hash_name(const void *salt, const char *name)
2004 {
2005         unsigned long hash = init_name_hash(salt);
2006         unsigned long len = 0, c;
2007
2008         c = (unsigned char)*name;
2009         do {
2010                 len++;
2011                 hash = partial_name_hash(c, hash);
2012                 c = (unsigned char)name[len];
2013         } while (c && c != '/');
2014         return hashlen_create(end_name_hash(hash), len);
2015 }
2016
2017 #endif
2018
2019 /*
2020  * Name resolution.
2021  * This is the basic name resolution function, turning a pathname into
2022  * the final dentry. We expect 'base' to be positive and a directory.
2023  *
2024  * Returns 0 and nd will have valid dentry and mnt on success.
2025  * Returns error and drops reference to input namei data on failure.
2026  */
2027 static int link_path_walk(const char *name, struct nameidata *nd)
2028 {
2029         int err;
2030
2031         while (*name=='/')
2032                 name++;
2033         if (!*name)
2034                 return 0;
2035
2036         /* At this point we know we have a real path component. */
2037         for(;;) {
2038                 u64 hash_len;
2039                 int type;
2040
2041                 err = may_lookup(nd);
2042                 if (err)
2043                         return err;
2044
2045                 hash_len = hash_name(nd->path.dentry, name);
2046
2047                 type = LAST_NORM;
2048                 if (name[0] == '.') switch (hashlen_len(hash_len)) {
2049                         case 2:
2050                                 if (name[1] == '.') {
2051                                         type = LAST_DOTDOT;
2052                                         nd->flags |= LOOKUP_JUMPED;
2053                                 }
2054                                 break;
2055                         case 1:
2056                                 type = LAST_DOT;
2057                 }
2058                 if (likely(type == LAST_NORM)) {
2059                         struct dentry *parent = nd->path.dentry;
2060                         nd->flags &= ~LOOKUP_JUMPED;
2061                         if (unlikely(parent->d_flags & DCACHE_OP_HASH)) {
2062                                 struct qstr this = { { .hash_len = hash_len }, .name = name };
2063                                 err = parent->d_op->d_hash(parent, &this);
2064                                 if (err < 0)
2065                                         return err;
2066                                 hash_len = this.hash_len;
2067                                 name = this.name;
2068                         }
2069                 }
2070
2071                 nd->last.hash_len = hash_len;
2072                 nd->last.name = name;
2073                 nd->last_type = type;
2074
2075                 name += hashlen_len(hash_len);
2076                 if (!*name)
2077                         goto OK;
2078                 /*
2079                  * If it wasn't NUL, we know it was '/'. Skip that
2080                  * slash, and continue until no more slashes.
2081                  */
2082                 do {
2083                         name++;
2084                 } while (unlikely(*name == '/'));
2085                 if (unlikely(!*name)) {
2086 OK:
2087                         /* pathname body, done */
2088                         if (!nd->depth)
2089                                 return 0;
2090                         name = nd->stack[nd->depth - 1].name;
2091                         /* trailing symlink, done */
2092                         if (!name)
2093                                 return 0;
2094                         /* last component of nested symlink */
2095                         err = walk_component(nd, WALK_FOLLOW);
2096                 } else {
2097                         /* not the last component */
2098                         err = walk_component(nd, WALK_FOLLOW | WALK_MORE);
2099                 }
2100                 if (err < 0)
2101                         return err;
2102
2103                 if (err) {
2104                         const char *s = get_link(nd);
2105
2106                         if (IS_ERR(s))
2107                                 return PTR_ERR(s);
2108                         err = 0;
2109                         if (unlikely(!s)) {
2110                                 /* jumped */
2111                                 put_link(nd);
2112                         } else {
2113                                 nd->stack[nd->depth - 1].name = name;
2114                                 name = s;
2115                                 continue;
2116                         }
2117                 }
2118                 if (unlikely(!d_can_lookup(nd->path.dentry))) {
2119                         if (nd->flags & LOOKUP_RCU) {
2120                                 if (unlazy_walk(nd))
2121                                         return -ECHILD;
2122                         }
2123                         return -ENOTDIR;
2124                 }
2125         }
2126 }
2127
2128 static const char *path_init(struct nameidata *nd, unsigned flags)
2129 {
2130         const char *s = nd->name->name;
2131
2132         if (!*s)
2133                 flags &= ~LOOKUP_RCU;
2134
2135         nd->last_type = LAST_ROOT; /* if there are only slashes... */
2136         nd->flags = flags | LOOKUP_JUMPED | LOOKUP_PARENT;
2137         nd->depth = 0;
2138         if (flags & LOOKUP_ROOT) {
2139                 struct dentry *root = nd->root.dentry;
2140                 struct inode *inode = root->d_inode;
2141                 if (*s && unlikely(!d_can_lookup(root)))
2142                         return ERR_PTR(-ENOTDIR);
2143                 nd->path = nd->root;
2144                 nd->inode = inode;
2145                 if (flags & LOOKUP_RCU) {
2146                         rcu_read_lock();
2147                         nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2148                         nd->root_seq = nd->seq;
2149                         nd->m_seq = read_seqbegin(&mount_lock);
2150                 } else {
2151                         path_get(&nd->path);
2152                 }
2153                 return s;
2154         }
2155
2156         nd->root.mnt = NULL;
2157         nd->path.mnt = NULL;
2158         nd->path.dentry = NULL;
2159
2160         nd->m_seq = read_seqbegin(&mount_lock);
2161         if (*s == '/') {
2162                 if (flags & LOOKUP_RCU)
2163                         rcu_read_lock();
2164                 set_root(nd);
2165                 if (likely(!nd_jump_root(nd)))
2166                         return s;
2167                 nd->root.mnt = NULL;
2168                 rcu_read_unlock();
2169                 return ERR_PTR(-ECHILD);
2170         } else if (nd->dfd == AT_FDCWD) {
2171                 if (flags & LOOKUP_RCU) {
2172                         struct fs_struct *fs = current->fs;
2173                         unsigned seq;
2174
2175                         rcu_read_lock();
2176
2177                         do {
2178                                 seq = read_seqcount_begin(&fs->seq);
2179                                 nd->path = fs->pwd;
2180                                 nd->inode = nd->path.dentry->d_inode;
2181                                 nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2182                         } while (read_seqcount_retry(&fs->seq, seq));
2183                 } else {
2184                         get_fs_pwd(current->fs, &nd->path);
2185                         nd->inode = nd->path.dentry->d_inode;
2186                 }
2187                 return s;
2188         } else {
2189                 /* Caller must check execute permissions on the starting path component */
2190                 struct fd f = fdget_raw(nd->dfd);
2191                 struct dentry *dentry;
2192
2193                 if (!f.file)
2194                         return ERR_PTR(-EBADF);
2195
2196                 dentry = f.file->f_path.dentry;
2197
2198                 if (*s) {
2199                         if (!d_can_lookup(dentry)) {
2200                                 fdput(f);
2201                                 return ERR_PTR(-ENOTDIR);
2202                         }
2203                 }
2204
2205                 nd->path = f.file->f_path;
2206                 if (flags & LOOKUP_RCU) {
2207                         rcu_read_lock();
2208                         nd->inode = nd->path.dentry->d_inode;
2209                         nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
2210                 } else {
2211                         path_get(&nd->path);
2212                         nd->inode = nd->path.dentry->d_inode;
2213                 }
2214                 fdput(f);
2215                 return s;
2216         }
2217 }
2218
2219 static const char *trailing_symlink(struct nameidata *nd)
2220 {
2221         const char *s;
2222         int error = may_follow_link(nd);
2223         if (unlikely(error))
2224                 return ERR_PTR(error);
2225         nd->flags |= LOOKUP_PARENT;
2226         nd->stack[0].name = NULL;
2227         s = get_link(nd);
2228         return s ? s : "";
2229 }
2230
2231 static inline int lookup_last(struct nameidata *nd)
2232 {
2233         if (nd->last_type == LAST_NORM && nd->last.name[nd->last.len])
2234                 nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
2235
2236         nd->flags &= ~LOOKUP_PARENT;
2237         return walk_component(nd, 0);
2238 }
2239
2240 static int handle_lookup_down(struct nameidata *nd)
2241 {
2242         struct path path = nd->path;
2243         struct inode *inode = nd->inode;
2244         unsigned seq = nd->seq;
2245         int err;
2246
2247         if (nd->flags & LOOKUP_RCU) {
2248                 /*
2249                  * don't bother with unlazy_walk on failure - we are
2250                  * at the very beginning of walk, so we lose nothing
2251                  * if we simply redo everything in non-RCU mode
2252                  */
2253                 if (unlikely(!__follow_mount_rcu(nd, &path, &inode, &seq)))
2254                         return -ECHILD;
2255         } else {
2256                 dget(path.dentry);
2257                 err = follow_managed(&path, nd);
2258                 if (unlikely(err < 0))
2259                         return err;
2260                 inode = d_backing_inode(path.dentry);
2261                 seq = 0;
2262         }
2263         path_to_nameidata(&path, nd);
2264         nd->inode = inode;
2265         nd->seq = seq;
2266         return 0;
2267 }
2268
2269 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2270 static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
2271 {
2272         const char *s = path_init(nd, flags);
2273         int err;
2274
2275         if (IS_ERR(s))
2276                 return PTR_ERR(s);
2277
2278         if (unlikely(flags & LOOKUP_DOWN)) {
2279                 err = handle_lookup_down(nd);
2280                 if (unlikely(err < 0)) {
2281                         terminate_walk(nd);
2282                         return err;
2283                 }
2284         }
2285
2286         while (!(err = link_path_walk(s, nd))
2287                 && ((err = lookup_last(nd)) > 0)) {
2288                 s = trailing_symlink(nd);
2289                 if (IS_ERR(s)) {
2290                         err = PTR_ERR(s);
2291                         break;
2292                 }
2293         }
2294         if (!err)
2295                 err = complete_walk(nd);
2296
2297         if (!err && nd->flags & LOOKUP_DIRECTORY)
2298                 if (!d_can_lookup(nd->path.dentry))
2299                         err = -ENOTDIR;
2300         if (!err) {
2301                 *path = nd->path;
2302                 nd->path.mnt = NULL;
2303                 nd->path.dentry = NULL;
2304         }
2305         terminate_walk(nd);
2306         return err;
2307 }
2308
2309 static int filename_lookup(int dfd, struct filename *name, unsigned flags,
2310                            struct path *path, struct path *root)
2311 {
2312         int retval;
2313         struct nameidata nd;
2314         if (IS_ERR(name))
2315                 return PTR_ERR(name);
2316         if (unlikely(root)) {
2317                 nd.root = *root;
2318                 flags |= LOOKUP_ROOT;
2319         }
2320         set_nameidata(&nd, dfd, name);
2321         retval = path_lookupat(&nd, flags | LOOKUP_RCU, path);
2322         if (unlikely(retval == -ECHILD))
2323                 retval = path_lookupat(&nd, flags, path);
2324         if (unlikely(retval == -ESTALE))
2325                 retval = path_lookupat(&nd, flags | LOOKUP_REVAL, path);
2326
2327         if (likely(!retval))
2328                 audit_inode(name, path->dentry, flags & LOOKUP_PARENT);
2329         restore_nameidata();
2330         putname(name);
2331         return retval;
2332 }
2333
2334 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2335 static int path_parentat(struct nameidata *nd, unsigned flags,
2336                                 struct path *parent)
2337 {
2338         const char *s = path_init(nd, flags);
2339         int err;
2340         if (IS_ERR(s))
2341                 return PTR_ERR(s);
2342         err = link_path_walk(s, nd);
2343         if (!err)
2344                 err = complete_walk(nd);
2345         if (!err) {
2346                 *parent = nd->path;
2347                 nd->path.mnt = NULL;
2348                 nd->path.dentry = NULL;
2349         }
2350         terminate_walk(nd);
2351         return err;
2352 }
2353
2354 static struct filename *filename_parentat(int dfd, struct filename *name,
2355                                 unsigned int flags, struct path *parent,
2356                                 struct qstr *last, int *type)
2357 {
2358         int retval;
2359         struct nameidata nd;
2360
2361         if (IS_ERR(name))
2362                 return name;
2363         set_nameidata(&nd, dfd, name);
2364         retval = path_parentat(&nd, flags | LOOKUP_RCU, parent);
2365         if (unlikely(retval == -ECHILD))
2366                 retval = path_parentat(&nd, flags, parent);
2367         if (unlikely(retval == -ESTALE))
2368                 retval = path_parentat(&nd, flags | LOOKUP_REVAL, parent);
2369         if (likely(!retval)) {
2370                 *last = nd.last;
2371                 *type = nd.last_type;
2372                 audit_inode(name, parent->dentry, LOOKUP_PARENT);
2373         } else {
2374                 putname(name);
2375                 name = ERR_PTR(retval);
2376         }
2377         restore_nameidata();
2378         return name;
2379 }
2380
2381 /* does lookup, returns the object with parent locked */
2382 struct dentry *kern_path_locked(const char *name, struct path *path)
2383 {
2384         struct filename *filename;
2385         struct dentry *d;
2386         struct qstr last;
2387         int type;
2388
2389         filename = filename_parentat(AT_FDCWD, getname_kernel(name), 0, path,
2390                                     &last, &type);
2391         if (IS_ERR(filename))
2392                 return ERR_CAST(filename);
2393         if (unlikely(type != LAST_NORM)) {
2394                 path_put(path);
2395                 putname(filename);
2396                 return ERR_PTR(-EINVAL);
2397         }
2398         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
2399         d = __lookup_hash(&last, path->dentry, 0);
2400         if (IS_ERR(d)) {
2401                 inode_unlock(path->dentry->d_inode);
2402                 path_put(path);
2403         }
2404         putname(filename);
2405         return d;
2406 }
2407
2408 int kern_path(const char *name, unsigned int flags, struct path *path)
2409 {
2410         return filename_lookup(AT_FDCWD, getname_kernel(name),
2411                                flags, path, NULL);
2412 }
2413 EXPORT_SYMBOL(kern_path);
2414
2415 /**
2416  * vfs_path_lookup - lookup a file path relative to a dentry-vfsmount pair
2417  * @dentry:  pointer to dentry of the base directory
2418  * @mnt: pointer to vfs mount of the base directory
2419  * @name: pointer to file name
2420  * @flags: lookup flags
2421  * @path: pointer to struct path to fill
2422  */
2423 int vfs_path_lookup(struct dentry *dentry, struct vfsmount *mnt,
2424                     const char *name, unsigned int flags,
2425                     struct path *path)
2426 {
2427         struct path root = {.mnt = mnt, .dentry = dentry};
2428         /* the first argument of filename_lookup() is ignored with root */
2429         return filename_lookup(AT_FDCWD, getname_kernel(name),
2430                                flags , path, &root);
2431 }
2432 EXPORT_SYMBOL(vfs_path_lookup);
2433
2434 static int lookup_one_len_common(const char *name, struct dentry *base,
2435                                  int len, struct qstr *this)
2436 {
2437         this->name = name;
2438         this->len = len;
2439         this->hash = full_name_hash(base, name, len);
2440         if (!len)
2441                 return -EACCES;
2442
2443         if (unlikely(name[0] == '.')) {
2444                 if (len < 2 || (len == 2 && name[1] == '.'))
2445                         return -EACCES;
2446         }
2447
2448         while (len--) {
2449                 unsigned int c = *(const unsigned char *)name++;
2450                 if (c == '/' || c == '\0')
2451                         return -EACCES;
2452         }
2453         /*
2454          * See if the low-level filesystem might want
2455          * to use its own hash..
2456          */
2457         if (base->d_flags & DCACHE_OP_HASH) {
2458                 int err = base->d_op->d_hash(base, this);
2459                 if (err < 0)
2460                         return err;
2461         }
2462
2463         return inode_permission(base->d_inode, MAY_EXEC);
2464 }
2465
2466 /**
2467  * lookup_one_len - filesystem helper to lookup single pathname component
2468  * @name:       pathname component to lookup
2469  * @base:       base directory to lookup from
2470  * @len:        maximum length @len should be interpreted to
2471  *
2472  * Note that this routine is purely a helper for filesystem usage and should
2473  * not be called by generic code.
2474  *
2475  * The caller must hold base->i_mutex.
2476  */
2477 struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
2478 {
2479         struct dentry *dentry;
2480         struct qstr this;
2481         int err;
2482
2483         WARN_ON_ONCE(!inode_is_locked(base->d_inode));
2484
2485         err = lookup_one_len_common(name, base, len, &this);
2486         if (err)
2487                 return ERR_PTR(err);
2488
2489         dentry = lookup_dcache(&this, base, 0);
2490         return dentry ? dentry : __lookup_slow(&this, base, 0);
2491 }
2492 EXPORT_SYMBOL(lookup_one_len);
2493
2494 /**
2495  * lookup_one_len_unlocked - filesystem helper to lookup single pathname component
2496  * @name:       pathname component to lookup
2497  * @base:       base directory to lookup from
2498  * @len:        maximum length @len should be interpreted to
2499  *
2500  * Note that this routine is purely a helper for filesystem usage and should
2501  * not be called by generic code.
2502  *
2503  * Unlike lookup_one_len, it should be called without the parent
2504  * i_mutex held, and will take the i_mutex itself if necessary.
2505  */
2506 struct dentry *lookup_one_len_unlocked(const char *name,
2507                                        struct dentry *base, int len)
2508 {
2509         struct qstr this;
2510         int err;
2511         struct dentry *ret;
2512
2513         err = lookup_one_len_common(name, base, len, &this);
2514         if (err)
2515                 return ERR_PTR(err);
2516
2517         ret = lookup_dcache(&this, base, 0);
2518         if (!ret)
2519                 ret = lookup_slow(&this, base, 0);
2520         return ret;
2521 }
2522 EXPORT_SYMBOL(lookup_one_len_unlocked);
2523
2524 #ifdef CONFIG_UNIX98_PTYS
2525 int path_pts(struct path *path)
2526 {
2527         /* Find something mounted on "pts" in the same directory as
2528          * the input path.
2529          */
2530         struct dentry *child, *parent;
2531         struct qstr this;
2532         int ret;
2533
2534         ret = path_parent_directory(path);
2535         if (ret)
2536                 return ret;
2537
2538         parent = path->dentry;
2539         this.name = "pts";
2540         this.len = 3;
2541         child = d_hash_and_lookup(parent, &this);
2542         if (!child)
2543                 return -ENOENT;
2544
2545         path->dentry = child;
2546         dput(parent);
2547         follow_mount(path);
2548         return 0;
2549 }
2550 #endif
2551
2552 int user_path_at_empty(int dfd, const char __user *name, unsigned flags,
2553                  struct path *path, int *empty)
2554 {
2555         return filename_lookup(dfd, getname_flags(name, flags, empty),
2556                                flags, path, NULL);
2557 }
2558 EXPORT_SYMBOL(user_path_at_empty);
2559
2560 /**
2561  * mountpoint_last - look up last component for umount
2562  * @nd:   pathwalk nameidata - currently pointing at parent directory of "last"
2563  *
2564  * This is a special lookup_last function just for umount. In this case, we
2565  * need to resolve the path without doing any revalidation.
2566  *
2567  * The nameidata should be the result of doing a LOOKUP_PARENT pathwalk. Since
2568  * mountpoints are always pinned in the dcache, their ancestors are too. Thus,
2569  * in almost all cases, this lookup will be served out of the dcache. The only
2570  * cases where it won't are if nd->last refers to a symlink or the path is
2571  * bogus and it doesn't exist.
2572  *
2573  * Returns:
2574  * -error: if there was an error during lookup. This includes -ENOENT if the
2575  *         lookup found a negative dentry.
2576  *
2577  * 0:      if we successfully resolved nd->last and found it to not to be a
2578  *         symlink that needs to be followed.
2579  *
2580  * 1:      if we successfully resolved nd->last and found it to be a symlink
2581  *         that needs to be followed.
2582  */
2583 static int
2584 mountpoint_last(struct nameidata *nd)
2585 {
2586         int error = 0;
2587         struct dentry *dir = nd->path.dentry;
2588         struct path path;
2589
2590         /* If we're in rcuwalk, drop out of it to handle last component */
2591         if (nd->flags & LOOKUP_RCU) {
2592                 if (unlazy_walk(nd))
2593                         return -ECHILD;
2594         }
2595
2596         nd->flags &= ~LOOKUP_PARENT;
2597
2598         if (unlikely(nd->last_type != LAST_NORM)) {
2599                 error = handle_dots(nd, nd->last_type);
2600                 if (error)
2601                         return error;
2602                 path.dentry = dget(nd->path.dentry);
2603         } else {
2604                 path.dentry = d_lookup(dir, &nd->last);
2605                 if (!path.dentry) {
2606                         /*
2607                          * No cached dentry. Mounted dentries are pinned in the
2608                          * cache, so that means that this dentry is probably
2609                          * a symlink or the path doesn't actually point
2610                          * to a mounted dentry.
2611                          */
2612                         path.dentry = lookup_slow(&nd->last, dir,
2613                                              nd->flags | LOOKUP_NO_REVAL);
2614                         if (IS_ERR(path.dentry))
2615                                 return PTR_ERR(path.dentry);
2616                 }
2617         }
2618         if (d_is_negative(path.dentry)) {
2619                 dput(path.dentry);
2620                 return -ENOENT;
2621         }
2622         path.mnt = nd->path.mnt;
2623         return step_into(nd, &path, 0, d_backing_inode(path.dentry), 0);
2624 }
2625
2626 /**
2627  * path_mountpoint - look up a path to be umounted
2628  * @nd:         lookup context
2629  * @flags:      lookup flags
2630  * @path:       pointer to container for result
2631  *
2632  * Look up the given name, but don't attempt to revalidate the last component.
2633  * Returns 0 and "path" will be valid on success; Returns error otherwise.
2634  */
2635 static int
2636 path_mountpoint(struct nameidata *nd, unsigned flags, struct path *path)
2637 {
2638         const char *s = path_init(nd, flags);
2639         int err;
2640         if (IS_ERR(s))
2641                 return PTR_ERR(s);
2642         while (!(err = link_path_walk(s, nd)) &&
2643                 (err = mountpoint_last(nd)) > 0) {
2644                 s = trailing_symlink(nd);
2645                 if (IS_ERR(s)) {
2646                         err = PTR_ERR(s);
2647                         break;
2648                 }
2649         }
2650         if (!err) {
2651                 *path = nd->path;
2652                 nd->path.mnt = NULL;
2653                 nd->path.dentry = NULL;
2654                 follow_mount(path);
2655         }
2656         terminate_walk(nd);
2657         return err;
2658 }
2659
2660 static int
2661 filename_mountpoint(int dfd, struct filename *name, struct path *path,
2662                         unsigned int flags)
2663 {
2664         struct nameidata nd;
2665         int error;
2666         if (IS_ERR(name))
2667                 return PTR_ERR(name);
2668         set_nameidata(&nd, dfd, name);
2669         error = path_mountpoint(&nd, flags | LOOKUP_RCU, path);
2670         if (unlikely(error == -ECHILD))
2671                 error = path_mountpoint(&nd, flags, path);
2672         if (unlikely(error == -ESTALE))
2673                 error = path_mountpoint(&nd, flags | LOOKUP_REVAL, path);
2674         if (likely(!error))
2675                 audit_inode(name, path->dentry, 0);
2676         restore_nameidata();
2677         putname(name);
2678         return error;
2679 }
2680
2681 /**
2682  * user_path_mountpoint_at - lookup a path from userland in order to umount it
2683  * @dfd:        directory file descriptor
2684  * @name:       pathname from userland
2685  * @flags:      lookup flags
2686  * @path:       pointer to container to hold result
2687  *
2688  * A umount is a special case for path walking. We're not actually interested
2689  * in the inode in this situation, and ESTALE errors can be a problem. We
2690  * simply want track down the dentry and vfsmount attached at the mountpoint
2691  * and avoid revalidating the last component.
2692  *
2693  * Returns 0 and populates "path" on success.
2694  */
2695 int
2696 user_path_mountpoint_at(int dfd, const char __user *name, unsigned int flags,
2697                         struct path *path)
2698 {
2699         return filename_mountpoint(dfd, getname(name), path, flags);
2700 }
2701
2702 int
2703 kern_path_mountpoint(int dfd, const char *name, struct path *path,
2704                         unsigned int flags)
2705 {
2706         return filename_mountpoint(dfd, getname_kernel(name), path, flags);
2707 }
2708 EXPORT_SYMBOL(kern_path_mountpoint);
2709
2710 int __check_sticky(struct inode *dir, struct inode *inode)
2711 {
2712         kuid_t fsuid = current_fsuid();
2713
2714         if (uid_eq(inode->i_uid, fsuid))
2715                 return 0;
2716         if (uid_eq(dir->i_uid, fsuid))
2717                 return 0;
2718         return !capable_wrt_inode_uidgid(inode, CAP_FOWNER);
2719 }
2720 EXPORT_SYMBOL(__check_sticky);
2721
2722 /*
2723  *      Check whether we can remove a link victim from directory dir, check
2724  *  whether the type of victim is right.
2725  *  1. We can't do it if dir is read-only (done in permission())
2726  *  2. We should have write and exec permissions on dir
2727  *  3. We can't remove anything from append-only dir
2728  *  4. We can't do anything with immutable dir (done in permission())
2729  *  5. If the sticky bit on dir is set we should either
2730  *      a. be owner of dir, or
2731  *      b. be owner of victim, or
2732  *      c. have CAP_FOWNER capability
2733  *  6. If the victim is append-only or immutable we can't do antyhing with
2734  *     links pointing to it.
2735  *  7. If the victim has an unknown uid or gid we can't change the inode.
2736  *  8. If we were asked to remove a directory and victim isn't one - ENOTDIR.
2737  *  9. If we were asked to remove a non-directory and victim isn't one - EISDIR.
2738  * 10. We can't remove a root or mountpoint.
2739  * 11. We don't allow removal of NFS sillyrenamed files; it's handled by
2740  *     nfs_async_unlink().
2741  */
2742 static int may_delete(struct inode *dir, struct dentry *victim, bool isdir)
2743 {
2744         struct inode *inode = d_backing_inode(victim);
2745         int error;
2746
2747         if (d_is_negative(victim))
2748                 return -ENOENT;
2749         BUG_ON(!inode);
2750
2751         BUG_ON(victim->d_parent->d_inode != dir);
2752         audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);
2753
2754         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
2755         if (error)
2756                 return error;
2757         if (IS_APPEND(dir))
2758                 return -EPERM;
2759
2760         if (check_sticky(dir, inode) || IS_APPEND(inode) ||
2761             IS_IMMUTABLE(inode) || IS_SWAPFILE(inode) || HAS_UNMAPPED_ID(inode))
2762                 return -EPERM;
2763         if (isdir) {
2764                 if (!d_is_dir(victim))
2765                         return -ENOTDIR;
2766                 if (IS_ROOT(victim))
2767                         return -EBUSY;
2768         } else if (d_is_dir(victim))
2769                 return -EISDIR;
2770         if (IS_DEADDIR(dir))
2771                 return -ENOENT;
2772         if (victim->d_flags & DCACHE_NFSFS_RENAMED)
2773                 return -EBUSY;
2774         return 0;
2775 }
2776
2777 /*      Check whether we can create an object with dentry child in directory
2778  *  dir.
2779  *  1. We can't do it if child already exists (open has special treatment for
2780  *     this case, but since we are inlined it's OK)
2781  *  2. We can't do it if dir is read-only (done in permission())
2782  *  3. We can't do it if the fs can't represent the fsuid or fsgid.
2783  *  4. We should have write and exec permissions on dir
2784  *  5. We can't do it if dir is immutable (done in permission())
2785  */
2786 static inline int may_create(struct inode *dir, struct dentry *child)
2787 {
2788         struct user_namespace *s_user_ns;
2789         audit_inode_child(dir, child, AUDIT_TYPE_CHILD_CREATE);
2790         if (child->d_inode)
2791                 return -EEXIST;
2792         if (IS_DEADDIR(dir))
2793                 return -ENOENT;
2794         s_user_ns = dir->i_sb->s_user_ns;
2795         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
2796             !kgid_has_mapping(s_user_ns, current_fsgid()))
2797                 return -EOVERFLOW;
2798         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
2799 }
2800
2801 /*
2802  * p1 and p2 should be directories on the same fs.
2803  */
2804 struct dentry *lock_rename(struct dentry *p1, struct dentry *p2)
2805 {
2806         struct dentry *p;
2807
2808         if (p1 == p2) {
2809                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2810                 return NULL;
2811         }
2812
2813         mutex_lock(&p1->d_sb->s_vfs_rename_mutex);
2814
2815         p = d_ancestor(p2, p1);
2816         if (p) {
2817                 inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
2818                 inode_lock_nested(p1->d_inode, I_MUTEX_CHILD);
2819                 return p;
2820         }
2821
2822         p = d_ancestor(p1, p2);
2823         if (p) {
2824                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2825                 inode_lock_nested(p2->d_inode, I_MUTEX_CHILD);
2826                 return p;
2827         }
2828
2829         inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2830         inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
2831         return NULL;
2832 }
2833 EXPORT_SYMBOL(lock_rename);
2834
2835 void unlock_rename(struct dentry *p1, struct dentry *p2)
2836 {
2837         inode_unlock(p1->d_inode);
2838         if (p1 != p2) {
2839                 inode_unlock(p2->d_inode);
2840                 mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
2841         }
2842 }
2843 EXPORT_SYMBOL(unlock_rename);
2844
2845 int vfs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2846                 bool want_excl)
2847 {
2848         int error = may_create(dir, dentry);
2849         if (error)
2850                 return error;
2851
2852         if (!dir->i_op->create)
2853                 return -EACCES; /* shouldn't it be ENOSYS? */
2854         mode &= S_IALLUGO;
2855         mode |= S_IFREG;
2856         error = security_inode_create(dir, dentry, mode);
2857         if (error)
2858                 return error;
2859         error = dir->i_op->create(dir, dentry, mode, want_excl);
2860         if (!error)
2861                 fsnotify_create(dir, dentry);
2862         return error;
2863 }
2864 EXPORT_SYMBOL(vfs_create);
2865
2866 int vfs_mkobj(struct dentry *dentry, umode_t mode,
2867                 int (*f)(struct dentry *, umode_t, void *),
2868                 void *arg)
2869 {
2870         struct inode *dir = dentry->d_parent->d_inode;
2871         int error = may_create(dir, dentry);
2872         if (error)
2873                 return error;
2874
2875         mode &= S_IALLUGO;
2876         mode |= S_IFREG;
2877         error = security_inode_create(dir, dentry, mode);
2878         if (error)
2879                 return error;
2880         error = f(dentry, mode, arg);
2881         if (!error)
2882                 fsnotify_create(dir, dentry);
2883         return error;
2884 }
2885 EXPORT_SYMBOL(vfs_mkobj);
2886
2887 bool may_open_dev(const struct path *path)
2888 {
2889         return !(path->mnt->mnt_flags & MNT_NODEV) &&
2890                 !(path->mnt->mnt_sb->s_iflags & SB_I_NODEV);
2891 }
2892
2893 static int may_open(const struct path *path, int acc_mode, int flag)
2894 {
2895         struct dentry *dentry = path->dentry;
2896         struct inode *inode = dentry->d_inode;
2897         int error;
2898
2899         if (!inode)
2900                 return -ENOENT;
2901
2902         switch (inode->i_mode & S_IFMT) {
2903         case S_IFLNK:
2904                 return -ELOOP;
2905         case S_IFDIR:
2906                 if (acc_mode & MAY_WRITE)
2907                         return -EISDIR;
2908                 break;
2909         case S_IFBLK:
2910         case S_IFCHR:
2911                 if (!may_open_dev(path))
2912                         return -EACCES;
2913                 /*FALLTHRU*/
2914         case S_IFIFO:
2915         case S_IFSOCK:
2916                 flag &= ~O_TRUNC;
2917                 break;
2918         }
2919
2920         error = inode_permission(inode, MAY_OPEN | acc_mode);
2921         if (error)
2922                 return error;
2923
2924         /*
2925          * An append-only file must be opened in append mode for writing.
2926          */
2927         if (IS_APPEND(inode)) {
2928                 if  ((flag & O_ACCMODE) != O_RDONLY && !(flag & O_APPEND))
2929                         return -EPERM;
2930                 if (flag & O_TRUNC)
2931                         return -EPERM;
2932         }
2933
2934         /* O_NOATIME can only be set by the owner or superuser */
2935         if (flag & O_NOATIME && !inode_owner_or_capable(inode))
2936                 return -EPERM;
2937
2938         return 0;
2939 }
2940
2941 static int handle_truncate(struct file *filp)
2942 {
2943         const struct path *path = &filp->f_path;
2944         struct inode *inode = path->dentry->d_inode;
2945         int error = get_write_access(inode);
2946         if (error)
2947                 return error;
2948         /*
2949          * Refuse to truncate files with mandatory locks held on them.
2950          */
2951         error = locks_verify_locked(filp);
2952         if (!error)
2953                 error = security_path_truncate(path);
2954         if (!error) {
2955                 error = do_truncate(path->dentry, 0,
2956                                     ATTR_MTIME|ATTR_CTIME|ATTR_OPEN,
2957                                     filp);
2958         }
2959         put_write_access(inode);
2960         return error;
2961 }
2962
2963 static inline int open_to_namei_flags(int flag)
2964 {
2965         if ((flag & O_ACCMODE) == 3)
2966                 flag--;
2967         return flag;
2968 }
2969
2970 static int may_o_create(const struct path *dir, struct dentry *dentry, umode_t mode)
2971 {
2972         struct user_namespace *s_user_ns;
2973         int error = security_path_mknod(dir, dentry, mode, 0);
2974         if (error)
2975                 return error;
2976
2977         s_user_ns = dir->dentry->d_sb->s_user_ns;
2978         if (!kuid_has_mapping(s_user_ns, current_fsuid()) ||
2979             !kgid_has_mapping(s_user_ns, current_fsgid()))
2980                 return -EOVERFLOW;
2981
2982         error = inode_permission(dir->dentry->d_inode, MAY_WRITE | MAY_EXEC);
2983         if (error)
2984                 return error;
2985
2986         return security_inode_create(dir->dentry->d_inode, dentry, mode);
2987 }
2988
2989 /*
2990  * Attempt to atomically look up, create and open a file from a negative
2991  * dentry.
2992  *
2993  * Returns 0 if successful.  The file will have been created and attached to
2994  * @file by the filesystem calling finish_open().
2995  *
2996  * Returns 1 if the file was looked up only or didn't need creating.  The
2997  * caller will need to perform the open themselves.  @path will have been
2998  * updated to point to the new dentry.  This may be negative.
2999  *
3000  * Returns an error code otherwise.
3001  */
3002 static int atomic_open(struct nameidata *nd, struct dentry *dentry,
3003                         struct path *path, struct file *file,
3004                         const struct open_flags *op,
3005                         int open_flag, umode_t mode,
3006                         int *opened)
3007 {
3008         struct dentry *const DENTRY_NOT_SET = (void *) -1UL;
3009         struct inode *dir =  nd->path.dentry->d_inode;
3010         int error;
3011
3012         if (!(~open_flag & (O_EXCL | O_CREAT))) /* both O_EXCL and O_CREAT */
3013                 open_flag &= ~O_TRUNC;
3014
3015         if (nd->flags & LOOKUP_DIRECTORY)
3016                 open_flag |= O_DIRECTORY;
3017
3018         file->f_path.dentry = DENTRY_NOT_SET;
3019         file->f_path.mnt = nd->path.mnt;
3020         error = dir->i_op->atomic_open(dir, dentry, file,
3021                                        open_to_namei_flags(open_flag),
3022                                        mode, opened);
3023         d_lookup_done(dentry);
3024         if (!error) {
3025                 /*
3026                  * We didn't have the inode before the open, so check open
3027                  * permission here.
3028                  */
3029                 int acc_mode = op->acc_mode;
3030                 if (*opened & FILE_CREATED) {
3031                         WARN_ON(!(open_flag & O_CREAT));
3032                         fsnotify_create(dir, dentry);
3033                         acc_mode = 0;
3034                 }
3035                 error = may_open(&file->f_path, acc_mode, open_flag);
3036                 if (WARN_ON(error > 0))
3037                         error = -EINVAL;
3038         } else if (error > 0) {
3039                 if (WARN_ON(file->f_path.dentry == DENTRY_NOT_SET)) {
3040                         error = -EIO;
3041                 } else {
3042                         if (file->f_path.dentry) {
3043                                 dput(dentry);
3044                                 dentry = file->f_path.dentry;
3045                         }
3046                         if (*opened & FILE_CREATED)
3047                                 fsnotify_create(dir, dentry);
3048                         if (unlikely(d_is_negative(dentry))) {
3049                                 error = -ENOENT;
3050                         } else {
3051                                 path->dentry = dentry;
3052                                 path->mnt = nd->path.mnt;
3053                                 return 1;
3054                         }
3055                 }
3056         }
3057         dput(dentry);
3058         return error;
3059 }
3060
3061 /*
3062  * Look up and maybe create and open the last component.
3063  *
3064  * Must be called with i_mutex held on parent.
3065  *
3066  * Returns 0 if the file was successfully atomically created (if necessary) and
3067  * opened.  In this case the file will be returned attached to @file.
3068  *
3069  * Returns 1 if the file was not completely opened at this time, though lookups
3070  * and creations will have been performed and the dentry returned in @path will
3071  * be positive upon return if O_CREAT was specified.  If O_CREAT wasn't
3072  * specified then a negative dentry may be returned.
3073  *
3074  * An error code is returned otherwise.
3075  *
3076  * FILE_CREATE will be set in @*opened if the dentry was created and will be
3077  * cleared otherwise prior to returning.
3078  */
3079 static int lookup_open(struct nameidata *nd, struct path *path,
3080                         struct file *file,
3081                         const struct open_flags *op,
3082                         bool got_write, int *opened)
3083 {
3084         struct dentry *dir = nd->path.dentry;
3085         struct inode *dir_inode = dir->d_inode;
3086         int open_flag = op->open_flag;
3087         struct dentry *dentry;
3088         int error, create_error = 0;
3089         umode_t mode = op->mode;
3090         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
3091
3092         if (unlikely(IS_DEADDIR(dir_inode)))
3093                 return -ENOENT;
3094
3095         *opened &= ~FILE_CREATED;
3096         dentry = d_lookup(dir, &nd->last);
3097         for (;;) {
3098                 if (!dentry) {
3099                         dentry = d_alloc_parallel(dir, &nd->last, &wq);
3100                         if (IS_ERR(dentry))
3101                                 return PTR_ERR(dentry);
3102                 }
3103                 if (d_in_lookup(dentry))
3104                         break;
3105
3106                 error = d_revalidate(dentry, nd->flags);
3107                 if (likely(error > 0))
3108                         break;
3109                 if (error)
3110                         goto out_dput;
3111                 d_invalidate(dentry);
3112                 dput(dentry);
3113                 dentry = NULL;
3114         }
3115         if (dentry->d_inode) {
3116                 /* Cached positive dentry: will open in f_op->open */
3117                 goto out_no_open;
3118         }
3119
3120         /*
3121          * Checking write permission is tricky, bacuse we don't know if we are
3122          * going to actually need it: O_CREAT opens should work as long as the
3123          * file exists.  But checking existence breaks atomicity.  The trick is
3124          * to check access and if not granted clear O_CREAT from the flags.
3125          *
3126          * Another problem is returing the "right" error value (e.g. for an
3127          * O_EXCL open we want to return EEXIST not EROFS).
3128          */
3129         if (open_flag & O_CREAT) {
3130                 if (!IS_POSIXACL(dir->d_inode))
3131                         mode &= ~current_umask();
3132                 if (unlikely(!got_write)) {
3133                         create_error = -EROFS;
3134                         open_flag &= ~O_CREAT;
3135                         if (open_flag & (O_EXCL | O_TRUNC))
3136                                 goto no_open;
3137                         /* No side effects, safe to clear O_CREAT */
3138                 } else {
3139                         create_error = may_o_create(&nd->path, dentry, mode);
3140                         if (create_error) {
3141                                 open_flag &= ~O_CREAT;
3142                                 if (open_flag & O_EXCL)
3143                                         goto no_open;
3144                         }
3145                 }
3146         } else if ((open_flag & (O_TRUNC|O_WRONLY|O_RDWR)) &&
3147                    unlikely(!got_write)) {
3148                 /*
3149                  * No O_CREATE -> atomicity not a requirement -> fall
3150                  * back to lookup + open
3151                  */
3152                 goto no_open;
3153         }
3154
3155         if (dir_inode->i_op->atomic_open) {
3156                 error = atomic_open(nd, dentry, path, file, op, open_flag,
3157                                     mode, opened);
3158                 if (unlikely(error == -ENOENT) && create_error)
3159                         error = create_error;
3160                 return error;
3161         }
3162
3163 no_open:
3164         if (d_in_lookup(dentry)) {
3165                 struct dentry *res = dir_inode->i_op->lookup(dir_inode, dentry,
3166                                                              nd->flags);
3167                 d_lookup_done(dentry);
3168                 if (unlikely(res)) {
3169                         if (IS_ERR(res)) {
3170                                 error = PTR_ERR(res);
3171                                 goto out_dput;
3172                         }
3173                         dput(dentry);
3174                         dentry = res;
3175                 }
3176         }
3177
3178         /* Negative dentry, just create the file */
3179         if (!dentry->d_inode && (open_flag & O_CREAT)) {
3180                 *opened |= FILE_CREATED;
3181                 audit_inode_child(dir_inode, dentry, AUDIT_TYPE_CHILD_CREATE);
3182                 if (!dir_inode->i_op->create) {
3183                         error = -EACCES;
3184                         goto out_dput;
3185                 }
3186                 error = dir_inode->i_op->create(dir_inode, dentry, mode,
3187                                                 open_flag & O_EXCL);
3188                 if (error)
3189                         goto out_dput;
3190                 fsnotify_create(dir_inode, dentry);
3191         }
3192         if (unlikely(create_error) && !dentry->d_inode) {
3193                 error = create_error;
3194                 goto out_dput;
3195         }
3196 out_no_open:
3197         path->dentry = dentry;
3198         path->mnt = nd->path.mnt;
3199         return 1;
3200
3201 out_dput:
3202         dput(dentry);
3203         return error;
3204 }
3205
3206 /*
3207  * Handle the last step of open()
3208  */
3209 static int do_last(struct nameidata *nd,
3210                    struct file *file, const struct open_flags *op,
3211                    int *opened)
3212 {
3213         struct dentry *dir = nd->path.dentry;
3214         int open_flag = op->open_flag;
3215         bool will_truncate = (open_flag & O_TRUNC) != 0;
3216         bool got_write = false;
3217         int acc_mode = op->acc_mode;
3218         unsigned seq;
3219         struct inode *inode;
3220         struct path path;
3221         int error;
3222
3223         nd->flags &= ~LOOKUP_PARENT;
3224         nd->flags |= op->intent;
3225
3226         if (nd->last_type != LAST_NORM) {
3227                 error = handle_dots(nd, nd->last_type);
3228                 if (unlikely(error))
3229                         return error;
3230                 goto finish_open;
3231         }
3232
3233         if (!(open_flag & O_CREAT)) {
3234                 if (nd->last.name[nd->last.len])
3235                         nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
3236                 /* we _can_ be in RCU mode here */
3237                 error = lookup_fast(nd, &path, &inode, &seq);
3238                 if (likely(error > 0))
3239                         goto finish_lookup;
3240
3241                 if (error < 0)
3242                         return error;
3243
3244                 BUG_ON(nd->inode != dir->d_inode);
3245                 BUG_ON(nd->flags & LOOKUP_RCU);
3246         } else {
3247                 /* create side of things */
3248                 /*
3249                  * This will *only* deal with leaving RCU mode - LOOKUP_JUMPED
3250                  * has been cleared when we got to the last component we are
3251                  * about to look up
3252                  */
3253                 error = complete_walk(nd);
3254                 if (error)
3255                         return error;
3256
3257                 audit_inode(nd->name, dir, LOOKUP_PARENT);
3258                 /* trailing slashes? */
3259                 if (unlikely(nd->last.name[nd->last.len]))
3260                         return -EISDIR;
3261         }
3262
3263         if (open_flag & (O_CREAT | O_TRUNC | O_WRONLY | O_RDWR)) {
3264                 error = mnt_want_write(nd->path.mnt);
3265                 if (!error)
3266                         got_write = true;
3267                 /*
3268                  * do _not_ fail yet - we might not need that or fail with
3269                  * a different error; let lookup_open() decide; we'll be
3270                  * dropping this one anyway.
3271                  */
3272         }
3273         if (open_flag & O_CREAT)
3274                 inode_lock(dir->d_inode);
3275         else
3276                 inode_lock_shared(dir->d_inode);
3277         error = lookup_open(nd, &path, file, op, got_write, opened);
3278         if (open_flag & O_CREAT)
3279                 inode_unlock(dir->d_inode);
3280         else
3281                 inode_unlock_shared(dir->d_inode);
3282
3283         if (error <= 0) {
3284                 if (error)
3285                         goto out;
3286
3287                 if ((*opened & FILE_CREATED) ||
3288                     !S_ISREG(file_inode(file)->i_mode))
3289                         will_truncate = false;
3290
3291                 audit_inode(nd->name, file->f_path.dentry, 0);
3292                 goto opened;
3293         }
3294
3295         if (*opened & FILE_CREATED) {
3296                 /* Don't check for write permission, don't truncate */
3297                 open_flag &= ~O_TRUNC;
3298                 will_truncate = false;
3299                 acc_mode = 0;
3300                 path_to_nameidata(&path, nd);
3301                 goto finish_open_created;
3302         }
3303
3304         /*
3305          * If atomic_open() acquired write access it is dropped now due to
3306          * possible mount and symlink following (this might be optimized away if
3307          * necessary...)
3308          */
3309         if (got_write) {
3310                 mnt_drop_write(nd->path.mnt);
3311                 got_write = false;
3312         }
3313
3314         error = follow_managed(&path, nd);
3315         if (unlikely(error < 0))
3316                 return error;
3317
3318         if (unlikely(d_is_negative(path.dentry))) {
3319                 path_to_nameidata(&path, nd);
3320                 return -ENOENT;
3321         }
3322
3323         /*
3324          * create/update audit record if it already exists.
3325          */
3326         audit_inode(nd->name, path.dentry, 0);
3327
3328         if (unlikely((open_flag & (O_EXCL | O_CREAT)) == (O_EXCL | O_CREAT))) {
3329                 path_to_nameidata(&path, nd);
3330                 return -EEXIST;
3331         }
3332
3333         seq = 0;        /* out of RCU mode, so the value doesn't matter */
3334         inode = d_backing_inode(path.dentry);
3335 finish_lookup:
3336         error = step_into(nd, &path, 0, inode, seq);
3337         if (unlikely(error))
3338                 return error;
3339 finish_open:
3340         /* Why this, you ask?  _Now_ we might have grown LOOKUP_JUMPED... */
3341         error = complete_walk(nd);
3342         if (error)
3343                 return error;
3344         audit_inode(nd->name, nd->path.dentry, 0);
3345         error = -EISDIR;
3346         if ((open_flag & O_CREAT) && d_is_dir(nd->path.dentry))
3347                 goto out;
3348         error = -ENOTDIR;
3349         if ((nd->flags & LOOKUP_DIRECTORY) && !d_can_lookup(nd->path.dentry))
3350                 goto out;
3351         if (!d_is_reg(nd->path.dentry))
3352                 will_truncate = false;
3353
3354         if (will_truncate) {
3355                 error = mnt_want_write(nd->path.mnt);
3356                 if (error)
3357                         goto out;
3358                 got_write = true;
3359         }
3360 finish_open_created:
3361         error = may_open(&nd->path, acc_mode, open_flag);
3362         if (error)
3363                 goto out;
3364         BUG_ON(*opened & FILE_OPENED); /* once it's opened, it's opened */
3365         error = vfs_open(&nd->path, file, current_cred());
3366         if (error)
3367                 goto out;
3368         *opened |= FILE_OPENED;
3369 opened:
3370         error = ima_file_check(file, op->acc_mode, *opened);
3371         if (!error && will_truncate)
3372                 error = handle_truncate(file);
3373 out:
3374         if (unlikely(error) && (*opened & FILE_OPENED))
3375                 fput(file);
3376         if (unlikely(error > 0)) {
3377                 WARN_ON(1);
3378                 error = -EINVAL;
3379         }
3380         if (got_write)
3381                 mnt_drop_write(nd->path.mnt);
3382         return error;
3383 }
3384
3385 struct dentry *vfs_tmpfile(struct dentry *dentry, umode_t mode, int open_flag)
3386 {
3387         struct dentry *child = NULL;
3388         struct inode *dir = dentry->d_inode;
3389         struct inode *inode;
3390         int error;
3391
3392         /* we want directory to be writable */
3393         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
3394         if (error)
3395                 goto out_err;
3396         error = -EOPNOTSUPP;
3397         if (!dir->i_op->tmpfile)
3398                 goto out_err;
3399         error = -ENOMEM;
3400         child = d_alloc(dentry, &slash_name);
3401         if (unlikely(!child))
3402                 goto out_err;
3403         error = dir->i_op->tmpfile(dir, child, mode);
3404         if (error)
3405                 goto out_err;
3406         error = -ENOENT;
3407         inode = child->d_inode;
3408         if (unlikely(!inode))
3409                 goto out_err;
3410         if (!(open_flag & O_EXCL)) {
3411                 spin_lock(&inode->i_lock);
3412                 inode->i_state |= I_LINKABLE;
3413                 spin_unlock(&inode->i_lock);
3414         }
3415         return child;
3416
3417 out_err:
3418         dput(child);
3419         return ERR_PTR(error);
3420 }
3421 EXPORT_SYMBOL(vfs_tmpfile);
3422
3423 static int do_tmpfile(struct nameidata *nd, unsigned flags,
3424                 const struct open_flags *op,
3425                 struct file *file, int *opened)
3426 {
3427         struct dentry *child;
3428         struct path path;
3429         int error = path_lookupat(nd, flags | LOOKUP_DIRECTORY, &path);
3430         if (unlikely(error))
3431                 return error;
3432         error = mnt_want_write(path.mnt);
3433         if (unlikely(error))
3434                 goto out;
3435         child = vfs_tmpfile(path.dentry, op->mode, op->open_flag);
3436         error = PTR_ERR(child);
3437         if (IS_ERR(child))
3438                 goto out2;
3439         dput(path.dentry);
3440         path.dentry = child;
3441         audit_inode(nd->name, child, 0);
3442         /* Don't check for other permissions, the inode was just created */
3443         error = may_open(&path, 0, op->open_flag);
3444         if (error)
3445                 goto out2;
3446         file->f_path.mnt = path.mnt;
3447         error = finish_open(file, child, NULL, opened);
3448         if (error)
3449                 goto out2;
3450 out2:
3451         mnt_drop_write(path.mnt);
3452 out:
3453         path_put(&path);
3454         return error;
3455 }
3456
3457 static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
3458 {
3459         struct path path;
3460         int error = path_lookupat(nd, flags, &path);
3461         if (!error) {
3462                 audit_inode(nd->name, path.dentry, 0);
3463                 error = vfs_open(&path, file, current_cred());
3464                 path_put(&path);
3465         }
3466         return error;
3467 }
3468
3469 static struct file *path_openat(struct nameidata *nd,
3470                         const struct open_flags *op, unsigned flags)
3471 {
3472         const char *s;
3473         struct file *file;
3474         int opened = 0;
3475         int error;
3476
3477         file = get_empty_filp();
3478         if (IS_ERR(file))
3479                 return file;
3480
3481         file->f_flags = op->open_flag;
3482
3483         if (unlikely(file->f_flags & __O_TMPFILE)) {
3484                 error = do_tmpfile(nd, flags, op, file, &opened);
3485                 goto out2;
3486         }
3487
3488         if (unlikely(file->f_flags & O_PATH)) {
3489                 error = do_o_path(nd, flags, file);
3490                 if (!error)
3491                         opened |= FILE_OPENED;
3492                 goto out2;
3493         }
3494
3495         s = path_init(nd, flags);
3496         if (IS_ERR(s)) {
3497                 put_filp(file);
3498                 return ERR_CAST(s);
3499         }
3500         while (!(error = link_path_walk(s, nd)) &&
3501                 (error = do_last(nd, file, op, &opened)) > 0) {
3502                 nd->flags &= ~(LOOKUP_OPEN|LOOKUP_CREATE|LOOKUP_EXCL);
3503                 s = trailing_symlink(nd);
3504                 if (IS_ERR(s)) {
3505                         error = PTR_ERR(s);
3506                         break;
3507                 }
3508         }
3509         terminate_walk(nd);
3510 out2:
3511         if (!(opened & FILE_OPENED)) {
3512                 BUG_ON(!error);
3513                 put_filp(file);
3514         }
3515         if (unlikely(error)) {
3516                 if (error == -EOPENSTALE) {
3517                         if (flags & LOOKUP_RCU)
3518                                 error = -ECHILD;
3519                         else
3520                                 error = -ESTALE;
3521                 }
3522                 file = ERR_PTR(error);
3523         }
3524         return file;
3525 }
3526
3527 struct file *do_filp_open(int dfd, struct filename *pathname,
3528                 const struct open_flags *op)
3529 {
3530         struct nameidata nd;
3531         int flags = op->lookup_flags;
3532         struct file *filp;
3533
3534         set_nameidata(&nd, dfd, pathname);
3535         filp = path_openat(&nd, op, flags | LOOKUP_RCU);
3536         if (unlikely(filp == ERR_PTR(-ECHILD)))
3537                 filp = path_openat(&nd, op, flags);
3538         if (unlikely(filp == ERR_PTR(-ESTALE)))
3539                 filp = path_openat(&nd, op, flags | LOOKUP_REVAL);
3540         restore_nameidata();
3541         return filp;
3542 }
3543
3544 struct file *do_file_open_root(struct dentry *dentry, struct vfsmount *mnt,
3545                 const char *name, const struct open_flags *op)
3546 {
3547         struct nameidata nd;
3548         struct file *file;
3549         struct filename *filename;
3550         int flags = op->lookup_flags | LOOKUP_ROOT;
3551
3552         nd.root.mnt = mnt;
3553         nd.root.dentry = dentry;
3554
3555         if (d_is_symlink(dentry) && op->intent & LOOKUP_OPEN)
3556                 return ERR_PTR(-ELOOP);
3557
3558         filename = getname_kernel(name);
3559         if (IS_ERR(filename))
3560                 return ERR_CAST(filename);
3561
3562         set_nameidata(&nd, -1, filename);
3563         file = path_openat(&nd, op, flags | LOOKUP_RCU);
3564         if (unlikely(file == ERR_PTR(-ECHILD)))
3565                 file = path_openat(&nd, op, flags);
3566         if (unlikely(file == ERR_PTR(-ESTALE)))
3567                 file = path_openat(&nd, op, flags | LOOKUP_REVAL);
3568         restore_nameidata();
3569         putname(filename);
3570         return file;
3571 }
3572
3573 static struct dentry *filename_create(int dfd, struct filename *name,
3574                                 struct path *path, unsigned int lookup_flags)
3575 {
3576         struct dentry *dentry = ERR_PTR(-EEXIST);
3577         struct qstr last;
3578         int type;
3579         int err2;
3580         int error;
3581         bool is_dir = (lookup_flags & LOOKUP_DIRECTORY);
3582
3583         /*
3584          * Note that only LOOKUP_REVAL and LOOKUP_DIRECTORY matter here. Any
3585          * other flags passed in are ignored!
3586          */
3587         lookup_flags &= LOOKUP_REVAL;
3588
3589         name = filename_parentat(dfd, name, lookup_flags, path, &last, &type);
3590         if (IS_ERR(name))
3591                 return ERR_CAST(name);
3592
3593         /*
3594          * Yucky last component or no last component at all?
3595          * (foo/., foo/.., /////)
3596          */
3597         if (unlikely(type != LAST_NORM))
3598                 goto out;
3599
3600         /* don't fail immediately if it's r/o, at least try to report other errors */
3601         err2 = mnt_want_write(path->mnt);
3602         /*
3603          * Do the final lookup.
3604          */
3605         lookup_flags |= LOOKUP_CREATE | LOOKUP_EXCL;
3606         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
3607         dentry = __lookup_hash(&last, path->dentry, lookup_flags);
3608         if (IS_ERR(dentry))
3609                 goto unlock;
3610
3611         error = -EEXIST;
3612         if (d_is_positive(dentry))
3613                 goto fail;
3614
3615         /*
3616          * Special case - lookup gave negative, but... we had foo/bar/
3617          * From the vfs_mknod() POV we just have a negative dentry -
3618          * all is fine. Let's be bastards - you had / on the end, you've
3619          * been asking for (non-existent) directory. -ENOENT for you.
3620          */
3621         if (unlikely(!is_dir && last.name[last.len])) {
3622                 error = -ENOENT;
3623                 goto fail;
3624         }
3625         if (unlikely(err2)) {
3626                 error = err2;
3627                 goto fail;
3628         }
3629         putname(name);
3630         return dentry;
3631 fail:
3632         dput(dentry);
3633         dentry = ERR_PTR(error);
3634 unlock:
3635         inode_unlock(path->dentry->d_inode);
3636         if (!err2)
3637                 mnt_drop_write(path->mnt);
3638 out:
3639         path_put(path);
3640         putname(name);
3641         return dentry;
3642 }
3643
3644 struct dentry *kern_path_create(int dfd, const char *pathname,
3645                                 struct path *path, unsigned int lookup_flags)
3646 {
3647         return filename_create(dfd, getname_kernel(pathname),
3648                                 path, lookup_flags);
3649 }
3650 EXPORT_SYMBOL(kern_path_create);
3651
3652 void done_path_create(struct path *path, struct dentry *dentry)
3653 {
3654         dput(dentry);
3655         inode_unlock(path->dentry->d_inode);
3656         mnt_drop_write(path->mnt);
3657         path_put(path);
3658 }
3659 EXPORT_SYMBOL(done_path_create);
3660
3661 inline struct dentry *user_path_create(int dfd, const char __user *pathname,
3662                                 struct path *path, unsigned int lookup_flags)
3663 {
3664         return filename_create(dfd, getname(pathname), path, lookup_flags);
3665 }
3666 EXPORT_SYMBOL(user_path_create);
3667
3668 int vfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3669 {
3670         int error = may_create(dir, dentry);
3671
3672         if (error)
3673                 return error;
3674
3675         if ((S_ISCHR(mode) || S_ISBLK(mode)) && !capable(CAP_MKNOD))
3676                 return -EPERM;
3677
3678         if (!dir->i_op->mknod)
3679                 return -EPERM;
3680
3681         error = devcgroup_inode_mknod(mode, dev);
3682         if (error)
3683                 return error;
3684
3685         error = security_inode_mknod(dir, dentry, mode, dev);
3686         if (error)
3687                 return error;
3688
3689         error = dir->i_op->mknod(dir, dentry, mode, dev);
3690         if (!error)
3691                 fsnotify_create(dir, dentry);
3692         return error;
3693 }
3694 EXPORT_SYMBOL(vfs_mknod);
3695
3696 static int may_mknod(umode_t mode)
3697 {
3698         switch (mode & S_IFMT) {
3699         case S_IFREG:
3700         case S_IFCHR:
3701         case S_IFBLK:
3702         case S_IFIFO:
3703         case S_IFSOCK:
3704         case 0: /* zero mode translates to S_IFREG */
3705                 return 0;
3706         case S_IFDIR:
3707                 return -EPERM;
3708         default:
3709                 return -EINVAL;
3710         }
3711 }
3712
3713 long do_mknodat(int dfd, const char __user *filename, umode_t mode,
3714                 unsigned int dev)
3715 {
3716         struct dentry *dentry;
3717         struct path path;
3718         int error;
3719         unsigned int lookup_flags = 0;
3720
3721         error = may_mknod(mode);
3722         if (error)
3723                 return error;
3724 retry:
3725         dentry = user_path_create(dfd, filename, &path, lookup_flags);
3726         if (IS_ERR(dentry))
3727                 return PTR_ERR(dentry);
3728
3729         if (!IS_POSIXACL(path.dentry->d_inode))
3730                 mode &= ~current_umask();
3731         error = security_path_mknod(&path, dentry, mode, dev);
3732         if (error)
3733                 goto out;
3734         switch (mode & S_IFMT) {
3735                 case 0: case S_IFREG:
3736                         error = vfs_create(path.dentry->d_inode,dentry,mode,true);
3737                         if (!error)
3738                                 ima_post_path_mknod(dentry);
3739                         break;
3740                 case S_IFCHR: case S_IFBLK:
3741                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,
3742                                         new_decode_dev(dev));
3743                         break;
3744                 case S_IFIFO: case S_IFSOCK:
3745                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,0);
3746                         break;
3747         }
3748 out:
3749         done_path_create(&path, dentry);
3750         if (retry_estale(error, lookup_flags)) {
3751                 lookup_flags |= LOOKUP_REVAL;
3752                 goto retry;
3753         }
3754         return error;
3755 }
3756
3757 SYSCALL_DEFINE4(mknodat, int, dfd, const char __user *, filename, umode_t, mode,
3758                 unsigned int, dev)
3759 {
3760         return do_mknodat(dfd, filename, mode, dev);
3761 }
3762
3763 SYSCALL_DEFINE3(mknod, const char __user *, filename, umode_t, mode, unsigned, dev)
3764 {
3765         return do_mknodat(AT_FDCWD, filename, mode, dev);
3766 }
3767
3768 int vfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3769 {
3770         int error = may_create(dir, dentry);
3771         unsigned max_links = dir->i_sb->s_max_links;
3772
3773         if (error)
3774                 return error;
3775
3776         if (!dir->i_op->mkdir)
3777                 return -EPERM;
3778
3779         mode &= (S_IRWXUGO|S_ISVTX);
3780         error = security_inode_mkdir(dir, dentry, mode);
3781         if (error)
3782                 return error;
3783
3784         if (max_links && dir->i_nlink >= max_links)
3785                 return -EMLINK;
3786
3787         error = dir->i_op->mkdir(dir, dentry, mode);
3788         if (!error)
3789                 fsnotify_mkdir(dir, dentry);
3790         return error;
3791 }
3792 EXPORT_SYMBOL(vfs_mkdir);
3793
3794 long do_mkdirat(int dfd, const char __user *pathname, umode_t mode)
3795 {
3796         struct dentry *dentry;
3797         struct path path;
3798         int error;
3799         unsigned int lookup_flags = LOOKUP_DIRECTORY;
3800
3801 retry:
3802         dentry = user_path_create(dfd, pathname, &path, lookup_flags);
3803         if (IS_ERR(dentry))
3804                 return PTR_ERR(dentry);
3805
3806         if (!IS_POSIXACL(path.dentry->d_inode))
3807                 mode &= ~current_umask();
3808         error = security_path_mkdir(&path, dentry, mode);
3809         if (!error)
3810                 error = vfs_mkdir(path.dentry->d_inode, dentry, mode);
3811         done_path_create(&path, dentry);
3812         if (retry_estale(error, lookup_flags)) {
3813                 lookup_flags |= LOOKUP_REVAL;
3814                 goto retry;
3815         }
3816         return error;
3817 }
3818
3819 SYSCALL_DEFINE3(mkdirat, int, dfd, const char __user *, pathname, umode_t, mode)
3820 {
3821         return do_mkdirat(dfd, pathname, mode);
3822 }
3823
3824 SYSCALL_DEFINE2(mkdir, const char __user *, pathname, umode_t, mode)
3825 {
3826         return do_mkdirat(AT_FDCWD, pathname, mode);
3827 }
3828
3829 int vfs_rmdir(struct inode *dir, struct dentry *dentry)
3830 {
3831         int error = may_delete(dir, dentry, 1);
3832
3833         if (error)
3834                 return error;
3835
3836         if (!dir->i_op->rmdir)
3837                 return -EPERM;
3838
3839         dget(dentry);
3840         inode_lock(dentry->d_inode);
3841
3842         error = -EBUSY;
3843         if (is_local_mountpoint(dentry))
3844                 goto out;
3845
3846         error = security_inode_rmdir(dir, dentry);
3847         if (error)
3848                 goto out;
3849
3850         shrink_dcache_parent(dentry);
3851         error = dir->i_op->rmdir(dir, dentry);
3852         if (error)
3853                 goto out;
3854
3855         dentry->d_inode->i_flags |= S_DEAD;
3856         dont_mount(dentry);
3857         detach_mounts(dentry);
3858
3859 out:
3860         inode_unlock(dentry->d_inode);
3861         dput(dentry);
3862         if (!error)
3863                 d_delete(dentry);
3864         return error;
3865 }
3866 EXPORT_SYMBOL(vfs_rmdir);
3867
3868 long do_rmdir(int dfd, const char __user *pathname)
3869 {
3870         int error = 0;
3871         struct filename *name;
3872         struct dentry *dentry;
3873         struct path path;
3874         struct qstr last;
3875         int type;
3876         unsigned int lookup_flags = 0;
3877 retry:
3878         name = filename_parentat(dfd, getname(pathname), lookup_flags,
3879                                 &path, &last, &type);
3880         if (IS_ERR(name))
3881                 return PTR_ERR(name);
3882
3883         switch (type) {
3884         case LAST_DOTDOT:
3885                 error = -ENOTEMPTY;
3886                 goto exit1;
3887         case LAST_DOT:
3888                 error = -EINVAL;
3889                 goto exit1;
3890         case LAST_ROOT:
3891                 error = -EBUSY;
3892                 goto exit1;
3893         }
3894
3895         error = mnt_want_write(path.mnt);
3896         if (error)
3897                 goto exit1;
3898
3899         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3900         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3901         error = PTR_ERR(dentry);
3902         if (IS_ERR(dentry))
3903                 goto exit2;
3904         if (!dentry->d_inode) {
3905                 error = -ENOENT;
3906                 goto exit3;
3907         }
3908         error = security_path_rmdir(&path, dentry);
3909         if (error)
3910                 goto exit3;
3911         error = vfs_rmdir(path.dentry->d_inode, dentry);
3912 exit3:
3913         dput(dentry);
3914 exit2:
3915         inode_unlock(path.dentry->d_inode);
3916         mnt_drop_write(path.mnt);
3917 exit1:
3918         path_put(&path);
3919         putname(name);
3920         if (retry_estale(error, lookup_flags)) {
3921                 lookup_flags |= LOOKUP_REVAL;
3922                 goto retry;
3923         }
3924         return error;
3925 }
3926
3927 SYSCALL_DEFINE1(rmdir, const char __user *, pathname)
3928 {
3929         return do_rmdir(AT_FDCWD, pathname);
3930 }
3931
3932 /**
3933  * vfs_unlink - unlink a filesystem object
3934  * @dir:        parent directory
3935  * @dentry:     victim
3936  * @delegated_inode: returns victim inode, if the inode is delegated.
3937  *
3938  * The caller must hold dir->i_mutex.
3939  *
3940  * If vfs_unlink discovers a delegation, it will return -EWOULDBLOCK and
3941  * return a reference to the inode in delegated_inode.  The caller
3942  * should then break the delegation on that inode and retry.  Because
3943  * breaking a delegation may take a long time, the caller should drop
3944  * dir->i_mutex before doing so.
3945  *
3946  * Alternatively, a caller may pass NULL for delegated_inode.  This may
3947  * be appropriate for callers that expect the underlying filesystem not
3948  * to be NFS exported.
3949  */
3950 int vfs_unlink(struct inode *dir, struct dentry *dentry, struct inode **delegated_inode)
3951 {
3952         struct inode *target = dentry->d_inode;
3953         int error = may_delete(dir, dentry, 0);
3954
3955         if (error)
3956                 return error;
3957
3958         if (!dir->i_op->unlink)
3959                 return -EPERM;
3960
3961         inode_lock(target);
3962         if (is_local_mountpoint(dentry))
3963                 error = -EBUSY;
3964         else {
3965                 error = security_inode_unlink(dir, dentry);
3966                 if (!error) {
3967                         error = try_break_deleg(target, delegated_inode);
3968                         if (error)
3969                                 goto out;
3970                         error = dir->i_op->unlink(dir, dentry);
3971                         if (!error) {
3972                                 dont_mount(dentry);
3973                                 detach_mounts(dentry);
3974                         }
3975                 }
3976         }
3977 out:
3978         inode_unlock(target);
3979
3980         /* We don't d_delete() NFS sillyrenamed files--they still exist. */
3981         if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
3982                 fsnotify_link_count(target);
3983                 d_delete(dentry);
3984         }
3985
3986         return error;
3987 }
3988 EXPORT_SYMBOL(vfs_unlink);
3989
3990 /*
3991  * Make sure that the actual truncation of the file will occur outside its
3992  * directory's i_mutex.  Truncate can take a long time if there is a lot of
3993  * writeout happening, and we don't want to prevent access to the directory
3994  * while waiting on the I/O.
3995  */
3996 long do_unlinkat(int dfd, struct filename *name)
3997 {
3998         int error;
3999         struct dentry *dentry;
4000         struct path path;
4001         struct qstr last;
4002         int type;
4003         struct inode *inode = NULL;
4004         struct inode *delegated_inode = NULL;
4005         unsigned int lookup_flags = 0;
4006 retry:
4007         name = filename_parentat(dfd, name, lookup_flags, &path, &last, &type);
4008         if (IS_ERR(name))
4009                 return PTR_ERR(name);
4010
4011         error = -EISDIR;
4012         if (type != LAST_NORM)
4013                 goto exit1;
4014
4015         error = mnt_want_write(path.mnt);
4016         if (error)
4017                 goto exit1;
4018 retry_deleg:
4019         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
4020         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
4021         error = PTR_ERR(dentry);
4022         if (!IS_ERR(dentry)) {
4023                 /* Why not before? Because we want correct error value */
4024                 if (last.name[last.len])
4025                         goto slashes;
4026                 inode = dentry->d_inode;
4027                 if (d_is_negative(dentry))
4028                         goto slashes;
4029                 ihold(inode);
4030                 error = security_path_unlink(&path, dentry);
4031                 if (error)
4032                         goto exit2;
4033                 error = vfs_unlink(path.dentry->d_inode, dentry, &delegated_inode);
4034 exit2:
4035                 dput(dentry);
4036         }
4037         inode_unlock(path.dentry->d_inode);
4038         if (inode)
4039                 iput(inode);    /* truncate the inode here */
4040         inode = NULL;
4041         if (delegated_inode) {
4042                 error = break_deleg_wait(&delegated_inode);
4043                 if (!error)
4044                         goto retry_deleg;
4045         }
4046         mnt_drop_write(path.mnt);
4047 exit1:
4048         path_put(&path);
4049         if (retry_estale(error, lookup_flags)) {
4050                 lookup_flags |= LOOKUP_REVAL;
4051                 inode = NULL;
4052                 goto retry;
4053         }
4054         putname(name);
4055         return error;
4056
4057 slashes:
4058         if (d_is_negative(dentry))
4059                 error = -ENOENT;
4060         else if (d_is_dir(dentry))
4061                 error = -EISDIR;
4062         else
4063                 error = -ENOTDIR;
4064         goto exit2;
4065 }
4066
4067 SYSCALL_DEFINE3(unlinkat, int, dfd, const char __user *, pathname, int, flag)
4068 {
4069         if ((flag & ~AT_REMOVEDIR) != 0)
4070                 return -EINVAL;
4071
4072         if (flag & AT_REMOVEDIR)
4073                 return do_rmdir(dfd, pathname);
4074
4075         return do_unlinkat(dfd, getname(pathname));
4076 }
4077
4078 SYSCALL_DEFINE1(unlink, const char __user *, pathname)
4079 {
4080         return do_unlinkat(AT_FDCWD, getname(pathname));
4081 }
4082
4083 int vfs_symlink(struct inode *dir, struct dentry *dentry, const char *oldname)
4084 {
4085         int error = may_create(dir, dentry);
4086
4087         if (error)
4088                 return error;
4089
4090         if (!dir->i_op->symlink)
4091                 return -EPERM;
4092
4093         error = security_inode_symlink(dir, dentry, oldname);
4094         if (error)
4095                 return error;
4096
4097         error = dir->i_op->symlink(dir, dentry, oldname);
4098         if (!error)
4099                 fsnotify_create(dir, dentry);
4100         return error;
4101 }
4102 EXPORT_SYMBOL(vfs_symlink);
4103
4104 long do_symlinkat(const char __user *oldname, int newdfd,
4105                   const char __user *newname)
4106 {
4107         int error;
4108         struct filename *from;
4109         struct dentry *dentry;
4110         struct path path;
4111         unsigned int lookup_flags = 0;
4112
4113         from = getname(oldname);
4114         if (IS_ERR(from))
4115                 return PTR_ERR(from);
4116 retry:
4117         dentry = user_path_create(newdfd, newname, &path, lookup_flags);
4118         error = PTR_ERR(dentry);
4119         if (IS_ERR(dentry))
4120                 goto out_putname;
4121
4122         error = security_path_symlink(&path, dentry, from->name);
4123         if (!error)
4124                 error = vfs_symlink(path.dentry->d_inode, dentry, from->name);
4125         done_path_create(&path, dentry);
4126         if (retry_estale(error, lookup_flags)) {
4127                 lookup_flags |= LOOKUP_REVAL;
4128                 goto retry;
4129         }
4130 out_putname:
4131         putname(from);
4132         return error;
4133 }
4134
4135 SYSCALL_DEFINE3(symlinkat, const char __user *, oldname,
4136                 int, newdfd, const char __user *, newname)
4137 {
4138         return do_symlinkat(oldname, newdfd, newname);
4139 }
4140
4141 SYSCALL_DEFINE2(symlink, const char __user *, oldname, const char __user *, newname)
4142 {
4143         return do_symlinkat(oldname, AT_FDCWD, newname);
4144 }
4145
4146 /**
4147  * vfs_link - create a new link
4148  * @old_dentry: object to be linked
4149  * @dir:        new parent
4150  * @new_dentry: where to create the new link
4151  * @delegated_inode: returns inode needing a delegation break
4152  *
4153  * The caller must hold dir->i_mutex
4154  *
4155  * If vfs_link discovers a delegation on the to-be-linked file in need
4156  * of breaking, it will return -EWOULDBLOCK and return a reference to the
4157  * inode in delegated_inode.  The caller should then break the delegation
4158  * and retry.  Because breaking a delegation may take a long time, the
4159  * caller should drop the i_mutex before doing so.
4160  *
4161  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4162  * be appropriate for callers that expect the underlying filesystem not
4163  * to be NFS exported.
4164  */
4165 int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_dentry, struct inode **delegated_inode)
4166 {
4167         struct inode *inode = old_dentry->d_inode;
4168         unsigned max_links = dir->i_sb->s_max_links;
4169         int error;
4170
4171         if (!inode)
4172                 return -ENOENT;
4173
4174         error = may_create(dir, new_dentry);
4175         if (error)
4176                 return error;
4177
4178         if (dir->i_sb != inode->i_sb)
4179                 return -EXDEV;
4180
4181         /*
4182          * A link to an append-only or immutable file cannot be created.
4183          */
4184         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4185                 return -EPERM;
4186         /*
4187          * Updating the link count will likely cause i_uid and i_gid to
4188          * be writen back improperly if their true value is unknown to
4189          * the vfs.
4190          */
4191         if (HAS_UNMAPPED_ID(inode))
4192                 return -EPERM;
4193         if (!dir->i_op->link)
4194                 return -EPERM;
4195         if (S_ISDIR(inode->i_mode))
4196                 return -EPERM;
4197
4198         error = security_inode_link(old_dentry, dir, new_dentry);
4199         if (error)
4200                 return error;
4201
4202         inode_lock(inode);
4203         /* Make sure we don't allow creating hardlink to an unlinked file */
4204         if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))
4205                 error =  -ENOENT;
4206         else if (max_links && inode->i_nlink >= max_links)
4207                 error = -EMLINK;
4208         else {
4209                 error = try_break_deleg(inode, delegated_inode);
4210                 if (!error)
4211                         error = dir->i_op->link(old_dentry, dir, new_dentry);
4212         }
4213
4214         if (!error && (inode->i_state & I_LINKABLE)) {
4215                 spin_lock(&inode->i_lock);
4216                 inode->i_state &= ~I_LINKABLE;
4217                 spin_unlock(&inode->i_lock);
4218         }
4219         inode_unlock(inode);
4220         if (!error)
4221                 fsnotify_link(dir, inode, new_dentry);
4222         return error;
4223 }
4224 EXPORT_SYMBOL(vfs_link);
4225
4226 /*
4227  * Hardlinks are often used in delicate situations.  We avoid
4228  * security-related surprises by not following symlinks on the
4229  * newname.  --KAB
4230  *
4231  * We don't follow them on the oldname either to be compatible
4232  * with linux 2.0, and to avoid hard-linking to directories
4233  * and other special files.  --ADM
4234  */
4235 int do_linkat(int olddfd, const char __user *oldname, int newdfd,
4236               const char __user *newname, int flags)
4237 {
4238         struct dentry *new_dentry;
4239         struct path old_path, new_path;
4240         struct inode *delegated_inode = NULL;
4241         int how = 0;
4242         int error;
4243
4244         if ((flags & ~(AT_SYMLINK_FOLLOW | AT_EMPTY_PATH)) != 0)
4245                 return -EINVAL;
4246         /*
4247          * To use null names we require CAP_DAC_READ_SEARCH
4248          * This ensures that not everyone will be able to create
4249          * handlink using the passed filedescriptor.
4250          */
4251         if (flags & AT_EMPTY_PATH) {
4252                 if (!capable(CAP_DAC_READ_SEARCH))
4253                         return -ENOENT;
4254                 how = LOOKUP_EMPTY;
4255         }
4256
4257         if (flags & AT_SYMLINK_FOLLOW)
4258                 how |= LOOKUP_FOLLOW;
4259 retry:
4260         error = user_path_at(olddfd, oldname, how, &old_path);
4261         if (error)
4262                 return error;
4263
4264         new_dentry = user_path_create(newdfd, newname, &new_path,
4265                                         (how & LOOKUP_REVAL));
4266         error = PTR_ERR(new_dentry);
4267         if (IS_ERR(new_dentry))
4268                 goto out;
4269
4270         error = -EXDEV;
4271         if (old_path.mnt != new_path.mnt)
4272                 goto out_dput;
4273         error = may_linkat(&old_path);
4274         if (unlikely(error))
4275                 goto out_dput;
4276         error = security_path_link(old_path.dentry, &new_path, new_dentry);
4277         if (error)
4278                 goto out_dput;
4279         error = vfs_link(old_path.dentry, new_path.dentry->d_inode, new_dentry, &delegated_inode);
4280 out_dput:
4281         done_path_create(&new_path, new_dentry);
4282         if (delegated_inode) {
4283                 error = break_deleg_wait(&delegated_inode);
4284                 if (!error) {
4285                         path_put(&old_path);
4286                         goto retry;
4287                 }
4288         }
4289         if (retry_estale(error, how)) {
4290                 path_put(&old_path);
4291                 how |= LOOKUP_REVAL;
4292                 goto retry;
4293         }
4294 out:
4295         path_put(&old_path);
4296
4297         return error;
4298 }
4299
4300 SYSCALL_DEFINE5(linkat, int, olddfd, const char __user *, oldname,
4301                 int, newdfd, const char __user *, newname, int, flags)
4302 {
4303         return do_linkat(olddfd, oldname, newdfd, newname, flags);
4304 }
4305
4306 SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname)
4307 {
4308         return do_linkat(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4309 }
4310
4311 /**
4312  * vfs_rename - rename a filesystem object
4313  * @old_dir:    parent of source
4314  * @old_dentry: source
4315  * @new_dir:    parent of destination
4316  * @new_dentry: destination
4317  * @delegated_inode: returns an inode needing a delegation break
4318  * @flags:      rename flags
4319  *
4320  * The caller must hold multiple mutexes--see lock_rename()).
4321  *
4322  * If vfs_rename discovers a delegation in need of breaking at either
4323  * the source or destination, it will return -EWOULDBLOCK and return a
4324  * reference to the inode in delegated_inode.  The caller should then
4325  * break the delegation and retry.  Because breaking a delegation may
4326  * take a long time, the caller should drop all locks before doing
4327  * so.
4328  *
4329  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4330  * be appropriate for callers that expect the underlying filesystem not
4331  * to be NFS exported.
4332  *
4333  * The worst of all namespace operations - renaming directory. "Perverted"
4334  * doesn't even start to describe it. Somebody in UCB had a heck of a trip...
4335  * Problems:
4336  *
4337  *      a) we can get into loop creation.
4338  *      b) race potential - two innocent renames can create a loop together.
4339  *         That's where 4.4 screws up. Current fix: serialization on
4340  *         sb->s_vfs_rename_mutex. We might be more accurate, but that's another
4341  *         story.
4342  *      c) we have to lock _four_ objects - parents and victim (if it exists),
4343  *         and source (if it is not a directory).
4344  *         And that - after we got ->i_mutex on parents (until then we don't know
4345  *         whether the target exists).  Solution: try to be smart with locking
4346  *         order for inodes.  We rely on the fact that tree topology may change
4347  *         only under ->s_vfs_rename_mutex _and_ that parent of the object we
4348  *         move will be locked.  Thus we can rank directories by the tree
4349  *         (ancestors first) and rank all non-directories after them.
4350  *         That works since everybody except rename does "lock parent, lookup,
4351  *         lock child" and rename is under ->s_vfs_rename_mutex.
4352  *         HOWEVER, it relies on the assumption that any object with ->lookup()
4353  *         has no more than 1 dentry.  If "hybrid" objects will ever appear,
4354  *         we'd better make sure that there's no link(2) for them.
4355  *      d) conversion from fhandle to dentry may come in the wrong moment - when
4356  *         we are removing the target. Solution: we will have to grab ->i_mutex
4357  *         in the fhandle_to_dentry code. [FIXME - current nfsfh.c relies on
4358  *         ->i_mutex on parents, which works but leads to some truly excessive
4359  *         locking].
4360  */
4361 int vfs_rename(struct inode *old_dir, struct dentry *old_dentry,
4362                struct inode *new_dir, struct dentry *new_dentry,
4363                struct inode **delegated_inode, unsigned int flags)
4364 {
4365         int error;
4366         bool is_dir = d_is_dir(old_dentry);
4367         struct inode *source = old_dentry->d_inode;
4368         struct inode *target = new_dentry->d_inode;
4369         bool new_is_dir = false;
4370         unsigned max_links = new_dir->i_sb->s_max_links;
4371         struct name_snapshot old_name;
4372
4373         if (source == target)
4374                 return 0;
4375
4376         error = may_delete(old_dir, old_dentry, is_dir);
4377         if (error)
4378                 return error;
4379
4380         if (!target) {
4381                 error = may_create(new_dir, new_dentry);
4382         } else {
4383                 new_is_dir = d_is_dir(new_dentry);
4384
4385                 if (!(flags & RENAME_EXCHANGE))
4386                         error = may_delete(new_dir, new_dentry, is_dir);
4387                 else
4388                         error = may_delete(new_dir, new_dentry, new_is_dir);
4389         }
4390         if (error)
4391                 return error;
4392
4393         if (!old_dir->i_op->rename)
4394                 return -EPERM;
4395
4396         /*
4397          * If we are going to change the parent - check write permissions,
4398          * we'll need to flip '..'.
4399          */
4400         if (new_dir != old_dir) {
4401                 if (is_dir) {
4402                         error = inode_permission(source, MAY_WRITE);
4403                         if (error)
4404                                 return error;
4405                 }
4406                 if ((flags & RENAME_EXCHANGE) && new_is_dir) {
4407                         error = inode_permission(target, MAY_WRITE);
4408                         if (error)
4409                                 return error;
4410                 }
4411         }
4412
4413         error = security_inode_rename(old_dir, old_dentry, new_dir, new_dentry,
4414                                       flags);
4415         if (error)
4416                 return error;
4417
4418         take_dentry_name_snapshot(&old_name, old_dentry);
4419         dget(new_dentry);
4420         if (!is_dir || (flags & RENAME_EXCHANGE))
4421                 lock_two_nondirectories(source, target);
4422         else if (target)
4423                 inode_lock(target);
4424
4425         error = -EBUSY;
4426         if (is_local_mountpoint(old_dentry) || is_local_mountpoint(new_dentry))
4427                 goto out;
4428
4429         if (max_links && new_dir != old_dir) {
4430                 error = -EMLINK;
4431                 if (is_dir && !new_is_dir && new_dir->i_nlink >= max_links)
4432                         goto out;
4433                 if ((flags & RENAME_EXCHANGE) && !is_dir && new_is_dir &&
4434                     old_dir->i_nlink >= max_links)
4435                         goto out;
4436         }
4437         if (is_dir && !(flags & RENAME_EXCHANGE) && target)
4438                 shrink_dcache_parent(new_dentry);
4439         if (!is_dir) {
4440                 error = try_break_deleg(source, delegated_inode);
4441                 if (error)
4442                         goto out;
4443         }
4444         if (target && !new_is_dir) {
4445                 error = try_break_deleg(target, delegated_inode);
4446                 if (error)
4447                         goto out;
4448         }
4449         error = old_dir->i_op->rename(old_dir, old_dentry,
4450                                        new_dir, new_dentry, flags);
4451         if (error)
4452                 goto out;
4453
4454         if (!(flags & RENAME_EXCHANGE) && target) {
4455                 if (is_dir)
4456                         target->i_flags |= S_DEAD;
4457                 dont_mount(new_dentry);
4458                 detach_mounts(new_dentry);
4459         }
4460         if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE)) {
4461                 if (!(flags & RENAME_EXCHANGE))
4462                         d_move(old_dentry, new_dentry);
4463                 else
4464                         d_exchange(old_dentry, new_dentry);
4465         }
4466 out:
4467         if (!is_dir || (flags & RENAME_EXCHANGE))
4468                 unlock_two_nondirectories(source, target);
4469         else if (target)
4470                 inode_unlock(target);
4471         dput(new_dentry);
4472         if (!error) {
4473                 fsnotify_move(old_dir, new_dir, old_name.name, is_dir,
4474                               !(flags & RENAME_EXCHANGE) ? target : NULL, old_dentry);
4475                 if (flags & RENAME_EXCHANGE) {
4476                         fsnotify_move(new_dir, old_dir, old_dentry->d_name.name,
4477                                       new_is_dir, NULL, new_dentry);
4478                 }
4479         }
4480         release_dentry_name_snapshot(&old_name);
4481
4482         return error;
4483 }
4484 EXPORT_SYMBOL(vfs_rename);
4485
4486 static int do_renameat2(int olddfd, const char __user *oldname, int newdfd,
4487                         const char __user *newname, unsigned int flags)
4488 {
4489         struct dentry *old_dentry, *new_dentry;
4490         struct dentry *trap;
4491         struct path old_path, new_path;
4492         struct qstr old_last, new_last;
4493         int old_type, new_type;
4494         struct inode *delegated_inode = NULL;
4495         struct filename *from;
4496         struct filename *to;
4497         unsigned int lookup_flags = 0, target_flags = LOOKUP_RENAME_TARGET;
4498         bool should_retry = false;
4499         int error;
4500
4501         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
4502                 return -EINVAL;
4503
4504         if ((flags & (RENAME_NOREPLACE | RENAME_WHITEOUT)) &&
4505             (flags & RENAME_EXCHANGE))
4506                 return -EINVAL;
4507
4508         if ((flags & RENAME_WHITEOUT) && !capable(CAP_MKNOD))
4509                 return -EPERM;
4510
4511         if (flags & RENAME_EXCHANGE)
4512                 target_flags = 0;
4513
4514 retry:
4515         from = filename_parentat(olddfd, getname(oldname), lookup_flags,
4516                                 &old_path, &old_last, &old_type);
4517         if (IS_ERR(from)) {
4518                 error = PTR_ERR(from);
4519                 goto exit;
4520         }
4521
4522         to = filename_parentat(newdfd, getname(newname), lookup_flags,
4523                                 &new_path, &new_last, &new_type);
4524         if (IS_ERR(to)) {
4525                 error = PTR_ERR(to);
4526                 goto exit1;
4527         }
4528
4529         error = -EXDEV;
4530         if (old_path.mnt != new_path.mnt)
4531                 goto exit2;
4532
4533         error = -EBUSY;
4534         if (old_type != LAST_NORM)
4535                 goto exit2;
4536
4537         if (flags & RENAME_NOREPLACE)
4538                 error = -EEXIST;
4539         if (new_type != LAST_NORM)
4540                 goto exit2;
4541
4542         error = mnt_want_write(old_path.mnt);
4543         if (error)
4544                 goto exit2;
4545
4546 retry_deleg:
4547         trap = lock_rename(new_path.dentry, old_path.dentry);
4548
4549         old_dentry = __lookup_hash(&old_last, old_path.dentry, lookup_flags);
4550         error = PTR_ERR(old_dentry);
4551         if (IS_ERR(old_dentry))
4552                 goto exit3;
4553         /* source must exist */
4554         error = -ENOENT;
4555         if (d_is_negative(old_dentry))
4556                 goto exit4;
4557         new_dentry = __lookup_hash(&new_last, new_path.dentry, lookup_flags | target_flags);
4558         error = PTR_ERR(new_dentry);
4559         if (IS_ERR(new_dentry))
4560                 goto exit4;
4561         error = -EEXIST;
4562         if ((flags & RENAME_NOREPLACE) && d_is_positive(new_dentry))
4563                 goto exit5;
4564         if (flags & RENAME_EXCHANGE) {
4565                 error = -ENOENT;
4566                 if (d_is_negative(new_dentry))
4567                         goto exit5;
4568
4569                 if (!d_is_dir(new_dentry)) {
4570                         error = -ENOTDIR;
4571                         if (new_last.name[new_last.len])
4572                                 goto exit5;
4573                 }
4574         }
4575         /* unless the source is a directory trailing slashes give -ENOTDIR */
4576         if (!d_is_dir(old_dentry)) {
4577                 error = -ENOTDIR;
4578                 if (old_last.name[old_last.len])
4579                         goto exit5;
4580                 if (!(flags & RENAME_EXCHANGE) && new_last.name[new_last.len])
4581                         goto exit5;
4582         }
4583         /* source should not be ancestor of target */
4584         error = -EINVAL;
4585         if (old_dentry == trap)
4586                 goto exit5;
4587         /* target should not be an ancestor of source */
4588         if (!(flags & RENAME_EXCHANGE))
4589                 error = -ENOTEMPTY;
4590         if (new_dentry == trap)
4591                 goto exit5;
4592
4593         error = security_path_rename(&old_path, old_dentry,
4594                                      &new_path, new_dentry, flags);
4595         if (error)
4596                 goto exit5;
4597         error = vfs_rename(old_path.dentry->d_inode, old_dentry,
4598                            new_path.dentry->d_inode, new_dentry,
4599                            &delegated_inode, flags);
4600 exit5:
4601         dput(new_dentry);
4602 exit4:
4603         dput(old_dentry);
4604 exit3:
4605         unlock_rename(new_path.dentry, old_path.dentry);
4606         if (delegated_inode) {
4607                 error = break_deleg_wait(&delegated_inode);
4608                 if (!error)
4609                         goto retry_deleg;
4610         }
4611         mnt_drop_write(old_path.mnt);
4612 exit2:
4613         if (retry_estale(error, lookup_flags))
4614                 should_retry = true;
4615         path_put(&new_path);
4616         putname(to);
4617 exit1:
4618         path_put(&old_path);
4619         putname(from);
4620         if (should_retry) {
4621                 should_retry = false;
4622                 lookup_flags |= LOOKUP_REVAL;
4623                 goto retry;
4624         }
4625 exit:
4626         return error;
4627 }
4628
4629 SYSCALL_DEFINE5(renameat2, int, olddfd, const char __user *, oldname,
4630                 int, newdfd, const char __user *, newname, unsigned int, flags)
4631 {
4632         return do_renameat2(olddfd, oldname, newdfd, newname, flags);
4633 }
4634
4635 SYSCALL_DEFINE4(renameat, int, olddfd, const char __user *, oldname,
4636                 int, newdfd, const char __user *, newname)
4637 {
4638         return do_renameat2(olddfd, oldname, newdfd, newname, 0);
4639 }
4640
4641 SYSCALL_DEFINE2(rename, const char __user *, oldname, const char __user *, newname)
4642 {
4643         return do_renameat2(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4644 }
4645
4646 int vfs_whiteout(struct inode *dir, struct dentry *dentry)
4647 {
4648         int error = may_create(dir, dentry);
4649         if (error)
4650                 return error;
4651
4652         if (!dir->i_op->mknod)
4653                 return -EPERM;
4654
4655         return dir->i_op->mknod(dir, dentry,
4656                                 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
4657 }
4658 EXPORT_SYMBOL(vfs_whiteout);
4659
4660 int readlink_copy(char __user *buffer, int buflen, const char *link)
4661 {
4662         int len = PTR_ERR(link);
4663         if (IS_ERR(link))
4664                 goto out;
4665
4666         len = strlen(link);
4667         if (len > (unsigned) buflen)
4668                 len = buflen;
4669         if (copy_to_user(buffer, link, len))
4670                 len = -EFAULT;
4671 out:
4672         return len;
4673 }
4674
4675 /*
4676  * A helper for ->readlink().  This should be used *ONLY* for symlinks that
4677  * have ->get_link() not calling nd_jump_link().  Using (or not using) it
4678  * for any given inode is up to filesystem.
4679  */
4680 static int generic_readlink(struct dentry *dentry, char __user *buffer,
4681                             int buflen)
4682 {
4683         DEFINE_DELAYED_CALL(done);
4684         struct inode *inode = d_inode(dentry);
4685         const char *link = inode->i_link;
4686         int res;
4687
4688         if (!link) {
4689                 link = inode->i_op->get_link(dentry, inode, &done);
4690                 if (IS_ERR(link))
4691                         return PTR_ERR(link);
4692         }
4693         res = readlink_copy(buffer, buflen, link);
4694         do_delayed_call(&done);
4695         return res;
4696 }
4697
4698 /**
4699  * vfs_readlink - copy symlink body into userspace buffer
4700  * @dentry: dentry on which to get symbolic link
4701  * @buffer: user memory pointer
4702  * @buflen: size of buffer
4703  *
4704  * Does not touch atime.  That's up to the caller if necessary
4705  *
4706  * Does not call security hook.
4707  */
4708 int vfs_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4709 {
4710         struct inode *inode = d_inode(dentry);
4711
4712         if (unlikely(!(inode->i_opflags & IOP_DEFAULT_READLINK))) {
4713                 if (unlikely(inode->i_op->readlink))
4714                         return inode->i_op->readlink(dentry, buffer, buflen);
4715
4716                 if (!d_is_symlink(dentry))
4717                         return -EINVAL;
4718
4719                 spin_lock(&inode->i_lock);
4720                 inode->i_opflags |= IOP_DEFAULT_READLINK;
4721                 spin_unlock(&inode->i_lock);
4722         }
4723
4724         return generic_readlink(dentry, buffer, buflen);
4725 }
4726 EXPORT_SYMBOL(vfs_readlink);
4727
4728 /**
4729  * vfs_get_link - get symlink body
4730  * @dentry: dentry on which to get symbolic link
4731  * @done: caller needs to free returned data with this
4732  *
4733  * Calls security hook and i_op->get_link() on the supplied inode.
4734  *
4735  * It does not touch atime.  That's up to the caller if necessary.
4736  *
4737  * Does not work on "special" symlinks like /proc/$$/fd/N
4738  */
4739 const char *vfs_get_link(struct dentry *dentry, struct delayed_call *done)
4740 {
4741         const char *res = ERR_PTR(-EINVAL);
4742         struct inode *inode = d_inode(dentry);
4743
4744         if (d_is_symlink(dentry)) {
4745                 res = ERR_PTR(security_inode_readlink(dentry));
4746                 if (!res)
4747                         res = inode->i_op->get_link(dentry, inode, done);
4748         }
4749         return res;
4750 }
4751 EXPORT_SYMBOL(vfs_get_link);
4752
4753 /* get the link contents into pagecache */
4754 const char *page_get_link(struct dentry *dentry, struct inode *inode,
4755                           struct delayed_call *callback)
4756 {
4757         char *kaddr;
4758         struct page *page;
4759         struct address_space *mapping = inode->i_mapping;
4760
4761         if (!dentry) {
4762                 page = find_get_page(mapping, 0);
4763                 if (!page)
4764                         return ERR_PTR(-ECHILD);
4765                 if (!PageUptodate(page)) {
4766                         put_page(page);
4767                         return ERR_PTR(-ECHILD);
4768                 }
4769         } else {
4770                 page = read_mapping_page(mapping, 0, NULL);
4771                 if (IS_ERR(page))
4772                         return (char*)page;
4773         }
4774         set_delayed_call(callback, page_put_link, page);
4775         BUG_ON(mapping_gfp_mask(mapping) & __GFP_HIGHMEM);
4776         kaddr = page_address(page);
4777         nd_terminate_link(kaddr, inode->i_size, PAGE_SIZE - 1);
4778         return kaddr;
4779 }
4780
4781 EXPORT_SYMBOL(page_get_link);
4782
4783 void page_put_link(void *arg)
4784 {
4785         put_page(arg);
4786 }
4787 EXPORT_SYMBOL(page_put_link);
4788
4789 int page_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4790 {
4791         DEFINE_DELAYED_CALL(done);
4792         int res = readlink_copy(buffer, buflen,
4793                                 page_get_link(dentry, d_inode(dentry),
4794                                               &done));
4795         do_delayed_call(&done);
4796         return res;
4797 }
4798 EXPORT_SYMBOL(page_readlink);
4799
4800 /*
4801  * The nofs argument instructs pagecache_write_begin to pass AOP_FLAG_NOFS
4802  */
4803 int __page_symlink(struct inode *inode, const char *symname, int len, int nofs)
4804 {
4805         struct address_space *mapping = inode->i_mapping;
4806         struct page *page;
4807         void *fsdata;
4808         int err;
4809         unsigned int flags = 0;
4810         if (nofs)
4811                 flags |= AOP_FLAG_NOFS;
4812
4813 retry:
4814         err = pagecache_write_begin(NULL, mapping, 0, len-1,
4815                                 flags, &page, &fsdata);
4816         if (err)
4817                 goto fail;
4818
4819         memcpy(page_address(page), symname, len-1);
4820
4821         err = pagecache_write_end(NULL, mapping, 0, len-1, len-1,
4822                                                         page, fsdata);
4823         if (err < 0)
4824                 goto fail;
4825         if (err < len-1)
4826                 goto retry;
4827
4828         mark_inode_dirty(inode);
4829         return 0;
4830 fail:
4831         return err;
4832 }
4833 EXPORT_SYMBOL(__page_symlink);
4834
4835 int page_symlink(struct inode *inode, const char *symname, int len)
4836 {
4837         return __page_symlink(inode, symname, len,
4838                         !mapping_gfp_constraint(inode->i_mapping, __GFP_FS));
4839 }
4840 EXPORT_SYMBOL(page_symlink);
4841
4842 const struct inode_operations page_symlink_inode_operations = {
4843         .get_link       = page_get_link,
4844 };
4845 EXPORT_SYMBOL(page_symlink_inode_operations);