fs/namei.c

   1 /*
   2  *  linux/fs/namei.c
   3  *
   4  *  Copyright (C) 1991, 1992  Linus Torvalds
   5  */
   6
   7 /*
   8  * Some corrections by tytso.
   9  */
  10
  11 /* [Feb 1997 T. Schoebel-Theuer] Complete rewrite of the pathname
  12  * lookup logic.
  13  */
  14 /* [Feb-Apr 2000, AV] Rewrite to the new namespace architecture.
  15  */
  16
  17 #include <linux/init.h>
  18 #include <linux/export.h>
  19 #include <linux/kernel.h>
  20 #include <linux/slab.h>
  21 #include <linux/fs.h>
  22 #include <linux/namei.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/fsnotify.h>
  25 #include <linux/personality.h>
  26 #include <linux/security.h>
  27 #include <linux/ima.h>
  28 #include <linux/syscalls.h>
  29 #include <linux/mount.h>
  30 #include <linux/audit.h>
  31 #include <linux/capability.h>
  32 #include <linux/file.h>
  33 #include <linux/fcntl.h>
  34 #include <linux/device_cgroup.h>
  35 #include <linux/fs_struct.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/hash.h>
  38 #include <asm/uaccess.h>
  39
  40 #include "internal.h"
  41 #include "mount.h"
  42
  43 /* [Feb-1997 T. Schoebel-Theuer]
  44  * Fundamental changes in the pathname lookup mechanisms (namei)
  45  * were necessary because of omirr.  The reason is that omirr needs
  46  * to know the _real_ pathname, not the user-supplied one, in case
  47  * of symlinks (and also when transname replacements occur).
  48  *
  49  * The new code replaces the old recursive symlink resolution with
  50  * an iterative one (in case of non-nested symlink chains).  It does
  51  * this with calls to <fs>_follow_link().
  52  * As a side effect, dir_namei(), _namei() and follow_link() are now
  53  * replaced with a single function lookup_dentry() that can handle all
  54  * the special cases of the former code.
  55  *
  56  * With the new dcache, the pathname is stored at each inode, at least as
  57  * long as the refcount of the inode is positive.  As a side effect, the
  58  * size of the dcache depends on the inode cache and thus is dynamic.
  59  *
  60  * [29-Apr-1998 C. Scott Ananian] Updated above description of symlink
  61  * resolution to correspond with current state of the code.
  62  *
  63  * Note that the symlink resolution is not *completely* iterative.
  64  * There is still a significant amount of tail- and mid- recursion in
  65  * the algorithm.  Also, note that <fs>_readlink() is not used in
  66  * lookup_dentry(): lookup_dentry() on the result of <fs>_readlink()
  67  * may return different results than <fs>_follow_link().  Many virtual
  68  * filesystems (including /proc) exhibit this behavior.
  69  */
  70
  71 /* [24-Feb-97 T. Schoebel-Theuer] Side effects caused by new implementation:
  72  * New symlink semantics: when open() is called with flags O_CREAT | O_EXCL
  73  * and the name already exists in form of a symlink, try to create the new
  74  * name indicated by the symlink. The old code always complained that the
  75  * name already exists, due to not following the symlink even if its target
  76  * is nonexistent.  The new semantics affects also mknod() and link() when
  77  * the name is a symlink pointing to a non-existent name.
  78  *
  79  * I don't know which semantics is the right one, since I have no access
  80  * to standards. But I found by trial that HP-UX 9.0 has the full "new"
  81  * semantics implemented, while SunOS 4.1.1 and Solaris (SunOS 5.4) have the
  82  * "old" one. Personally, I think the new semantics is much more logical.
  83  * Note that "ln old new" where "new" is a symlink pointing to a non-existing
  84  * file does succeed in both HP-UX and SunOs, but not in Solaris
  85  * and in the old Linux semantics.
  86  */
  87
  88 /* [16-Dec-97 Kevin Buhr] For security reasons, we change some symlink
  89  * semantics.  See the comments in "open_namei" and "do_link" below.
  90  *
  91  * [10-Sep-98 Alan Modra] Another symlink change.
  92  */
  93
  94 /* [Feb-Apr 2000 AV] Complete rewrite. Rules for symlinks:
  95  *      inside the path - always follow.
  96  *      in the last component in creation/removal/renaming - never follow.
  97  *      if LOOKUP_FOLLOW passed - follow.
  98  *      if the pathname has trailing slashes - follow.
  99  *      otherwise - don't follow.
 100  * (applied in that order).
 101  *
 102  * [Jun 2000 AV] Inconsistent behaviour of open() in case if flags==O_CREAT
 103  * restored for 2.4. This is the last surviving part of old 4.2BSD bug.
 104  * During the 2.4 we need to fix the userland stuff depending on it -
 105  * hopefully we will be able to get rid of that wart in 2.5. So far only
 106  * XEmacs seems to be relying on it...
 107  */
 108 /*
 109  * [Sep 2001 AV] Single-semaphore locking scheme (kudos to David Holland)
 110  * implemented.  Let's see if raised priority of ->s_vfs_rename_mutex gives
 111  * any extra contention...
 112  */
 113
 114 /* In order to reduce some races, while at the same time doing additional
 115  * checking and hopefully speeding things up, we copy filenames to the
 116  * kernel data space before using them..
 117  *
 118  * POSIX.1 2.4: an empty pathname is invalid (ENOENT).
 119  * PATH_MAX includes the nul terminator --RR.
 120  */
 121
 122 #define EMBEDDED_NAME_MAX       (PATH_MAX - offsetof(struct filename, iname))
 123
 124 struct filename *
 125 getname_flags(const char __user *filename, int flags, int *empty)
 126 {
 127         struct filename *result;
 128         char *kname;
 129         int len;
 130
 131         result = audit_reusename(filename);
 132         if (result)
 133                 return result;
 134
 135         result = __getname();
 136         if (unlikely(!result))
 137                 return ERR_PTR(-ENOMEM);
 138
 139         /*
 140          * First, try to embed the struct filename inside the names_cache
 141          * allocation
 142          */
 143         kname = (char *)result->iname;
 144         result->name = kname;
 145
 146         len = strncpy_from_user(kname, filename, EMBEDDED_NAME_MAX);
 147         if (unlikely(len < 0)) {
 148                 __putname(result);
 149                 return ERR_PTR(len);
 150         }
 151
 152         /*
 153          * Uh-oh. We have a name that's approaching PATH_MAX. Allocate a
 154          * separate struct filename so we can dedicate the entire
 155          * names_cache allocation for the pathname, and re-do the copy from
 156          * userland.
 157          */
 158         if (unlikely(len == EMBEDDED_NAME_MAX)) {
 159                 const size_t size = offsetof(struct filename, iname[1]);
 160                 kname = (char *)result;
 161
 162                 /*
 163                  * size is chosen that way we to guarantee that
 164                  * result->iname[0] is within the same object and that
 165                  * kname can't be equal to result->iname, no matter what.
 166                  */
 167                 result = kzalloc(size, GFP_KERNEL);
 168                 if (unlikely(!result)) {
 169                         __putname(kname);
 170                         return ERR_PTR(-ENOMEM);
 171                 }
 172                 result->name = kname;
 173                 len = strncpy_from_user(kname, filename, PATH_MAX);
 174                 if (unlikely(len < 0)) {
 175                         __putname(kname);
 176                         kfree(result);
 177                         return ERR_PTR(len);
 178                 }
 179                 if (unlikely(len == PATH_MAX)) {
 180                         __putname(kname);
 181                         kfree(result);
 182                         return ERR_PTR(-ENAMETOOLONG);
 183                 }
 184         }
 185
 186         result->refcnt = 1;
 187         /* The empty path is special. */
 188         if (unlikely(!len)) {
 189                 if (empty)
 190                         *empty = 1;
 191                 if (!(flags & LOOKUP_EMPTY)) {
 192                         putname(result);
 193                         return ERR_PTR(-ENOENT);
 194                 }
 195         }
 196
 197         result->uptr = filename;
 198         result->aname = NULL;
 199         audit_getname(result);
 200         return result;
 201 }
 202
 203 struct filename *
 204 getname(const char __user * filename)
 205 {
 206         return getname_flags(filename, 0, NULL);
 207 }
 208
 209 struct filename *
 210 getname_kernel(const char * filename)
 211 {
 212         struct filename *result;
 213         int len = strlen(filename) + 1;
 214
 215         result = __getname();
 216         if (unlikely(!result))
 217                 return ERR_PTR(-ENOMEM);
 218
 219         if (len <= EMBEDDED_NAME_MAX) {
 220                 result->name = (char *)result->iname;
 221         } else if (len <= PATH_MAX) {
 222                 struct filename *tmp;
 223
 224                 tmp = kmalloc(sizeof(*tmp), GFP_KERNEL);
 225                 if (unlikely(!tmp)) {
 226                         __putname(result);
 227                         return ERR_PTR(-ENOMEM);
 228                 }
 229                 tmp->name = (char *)result;
 230                 result = tmp;
 231         } else {
 232                 __putname(result);
 233                 return ERR_PTR(-ENAMETOOLONG);
 234         }
 235         memcpy((char *)result->name, filename, len);
 236         result->uptr = NULL;
 237         result->aname = NULL;
 238         result->refcnt = 1;
 239         audit_getname(result);
 240
 241         return result;
 242 }
 243
 244 void putname(struct filename *name)
 245 {
 246         BUG_ON(name->refcnt <= 0);
 247
 248         if (--name->refcnt > 0)
 249                 return;
 250
 251         if (name->name != name->iname) {
 252                 __putname(name->name);
 253                 kfree(name);
 254         } else
 255                 __putname(name);
 256 }
 257
 258 static int check_acl(struct inode *inode, int mask)
 259 {
 260 #ifdef CONFIG_FS_POSIX_ACL
 261         struct posix_acl *acl;
 262
 263         if (mask & MAY_NOT_BLOCK) {
 264                 acl = get_cached_acl_rcu(inode, ACL_TYPE_ACCESS);
 265                 if (!acl)
 266                         return -EAGAIN;
 267                 /* no ->get_acl() calls in RCU mode... */
 268                 if (is_uncached_acl(acl))
 269                         return -ECHILD;
 270                 return posix_acl_permission(inode, acl, mask & ~MAY_NOT_BLOCK);
 271         }
 272
 273         acl = get_acl(inode, ACL_TYPE_ACCESS);
 274         if (IS_ERR(acl))
 275                 return PTR_ERR(acl);
 276         if (acl) {
 277                 int error = posix_acl_permission(inode, acl, mask);
 278                 posix_acl_release(acl);
 279                 return error;
 280         }
 281 #endif
 282
 283         return -EAGAIN;
 284 }
 285
 286 /*
 287  * This does the basic permission checking
 288  */
 289 static int acl_permission_check(struct inode *inode, int mask)
 290 {
 291         unsigned int mode = inode->i_mode;
 292
 293         if (likely(uid_eq(current_fsuid(), inode->i_uid)))
 294                 mode >>= 6;
 295         else {
 296                 if (IS_POSIXACL(inode) && (mode & S_IRWXG)) {
 297                         int error = check_acl(inode, mask);
 298                         if (error != -EAGAIN)
 299                                 return error;
 300                 }
 301
 302                 if (in_group_p(inode->i_gid))
 303                         mode >>= 3;
 304         }
 305
 306         /*
 307          * If the DACs are ok we don't need any capability check.
 308          */
 309         if ((mask & ~mode & (MAY_READ | MAY_WRITE | MAY_EXEC)) == 0)
 310                 return 0;
 311         return -EACCES;
 312 }
 313
 314 /**
 315  * generic_permission -  check for access rights on a Posix-like filesystem
 316  * @inode:      inode to check access rights for
 317  * @mask:       right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC, ...)
 318  *
 319  * Used to check for read/write/execute permissions on a file.
 320  * We use "fsuid" for this, letting us set arbitrary permissions
 321  * for filesystem access without changing the "normal" uids which
 322  * are used for other things.
 323  *
 324  * generic_permission is rcu-walk aware. It returns -ECHILD in case an rcu-walk
 325  * request cannot be satisfied (eg. requires blocking or too much complexity).
 326  * It would then be called again in ref-walk mode.
 327  */
 328 int generic_permission(struct inode *inode, int mask)
 329 {
 330         int ret;
 331
 332         /*
 333          * Do the basic permission checks.
 334          */
 335         ret = acl_permission_check(inode, mask);
 336         if (ret != -EACCES)
 337                 return ret;
 338
 339         if (S_ISDIR(inode->i_mode)) {
 340                 /* DACs are overridable for directories */
 341                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
 342                         return 0;
 343                 if (!(mask & MAY_WRITE))
 344                         if (capable_wrt_inode_uidgid(inode,
 345                                                      CAP_DAC_READ_SEARCH))
 346                                 return 0;
 347                 return -EACCES;
 348         }
 349         /*
 350          * Read/write DACs are always overridable.
 351          * Executable DACs are overridable when there is
 352          * at least one exec bit set.
 353          */
 354         if (!(mask & MAY_EXEC) || (inode->i_mode & S_IXUGO))
 355                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
 356                         return 0;
 357
 358         /*
 359          * Searching includes executable on directories, else just read.
 360          */
 361         mask &= MAY_READ | MAY_WRITE | MAY_EXEC;
 362         if (mask == MAY_READ)
 363                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_READ_SEARCH))
 364                         return 0;
 365
 366         return -EACCES;
 367 }
 368 EXPORT_SYMBOL(generic_permission);
 369
 370 /*
 371  * We _really_ want to just do "generic_permission()" without
 372  * even looking at the inode->i_op values. So we keep a cache
 373  * flag in inode->i_opflags, that says "this has not special
 374  * permission function, use the fast case".
 375  */
 376 static inline int do_inode_permission(struct inode *inode, int mask)
 377 {
 378         if (unlikely(!(inode->i_opflags & IOP_FASTPERM))) {
 379                 if (likely(inode->i_op->permission))
 380                         return inode->i_op->permission(inode, mask);
 381
 382                 /* This gets set once for the inode lifetime */
 383                 spin_lock(&inode->i_lock);
 384                 inode->i_opflags |= IOP_FASTPERM;
 385                 spin_unlock(&inode->i_lock);
 386         }
 387         return generic_permission(inode, mask);
 388 }
 389
 390 /**
 391  * __inode_permission - Check for access rights to a given inode
 392  * @inode: Inode to check permission on
 393  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 394  *
 395  * Check for read/write/execute permissions on an inode.
 396  *
 397  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
 398  *
 399  * This does not check for a read-only file system.  You probably want
 400  * inode_permission().
 401  */
 402 int __inode_permission(struct inode *inode, int mask)
 403 {
 404         int retval;
 405
 406         if (unlikely(mask & MAY_WRITE)) {
 407                 /*
 408                  * Nobody gets write access to an immutable file.
 409                  */
 410                 if (IS_IMMUTABLE(inode))
 411                         return -EACCES;
 412         }
 413
 414         retval = do_inode_permission(inode, mask);
 415         if (retval)
 416                 return retval;
 417
 418         retval = devcgroup_inode_permission(inode, mask);
 419         if (retval)
 420                 return retval;
 421
 422         return security_inode_permission(inode, mask);
 423 }
 424 EXPORT_SYMBOL(__inode_permission);
 425
 426 /**
 427  * sb_permission - Check superblock-level permissions
 428  * @sb: Superblock of inode to check permission on
 429  * @inode: Inode to check permission on
 430  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 431  *
 432  * Separate out file-system wide checks from inode-specific permission checks.
 433  */
 434 static int sb_permission(struct super_block *sb, struct inode *inode, int mask)
 435 {
 436         if (unlikely(mask & MAY_WRITE)) {
 437                 umode_t mode = inode->i_mode;
 438
 439                 /* Nobody gets write access to a read-only fs. */
 440                 if ((sb->s_flags & MS_RDONLY) &&
 441                     (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
 442                         return -EROFS;
 443         }
 444         return 0;
 445 }
 446
 447 /**
 448  * inode_permission - Check for access rights to a given inode
 449  * @inode: Inode to check permission on
 450  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 451  *
 452  * Check for read/write/execute permissions on an inode.  We use fs[ug]id for
 453  * this, letting us set arbitrary permissions for filesystem access without
 454  * changing the "normal" UIDs which are used for other things.
 455  *
 456  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
 457  */
 458 int inode_permission(struct inode *inode, int mask)
 459 {
 460         int retval;
 461
 462         retval = sb_permission(inode->i_sb, inode, mask);
 463         if (retval)
 464                 return retval;
 465         return __inode_permission(inode, mask);
 466 }
 467 EXPORT_SYMBOL(inode_permission);
 468
 469 /**
 470  * path_get - get a reference to a path
 471  * @path: path to get the reference to
 472  *
 473  * Given a path increment the reference count to the dentry and the vfsmount.
 474  */
 475 void path_get(const struct path *path)
 476 {
 477         mntget(path->mnt);
 478         dget(path->dentry);
 479 }
 480 EXPORT_SYMBOL(path_get);
 481
 482 /**
 483  * path_put - put a reference to a path
 484  * @path: path to put the reference to
 485  *
 486  * Given a path decrement the reference count to the dentry and the vfsmount.
 487  */
 488 void path_put(const struct path *path)
 489 {
 490         dput(path->dentry);
 491         mntput(path->mnt);
 492 }
 493 EXPORT_SYMBOL(path_put);
 494
 495 #define EMBEDDED_LEVELS 2
 496 struct nameidata {
 497         struct path     path;
 498         struct qstr     last;
 499         struct path     root;
 500         struct inode    *inode; /* path.dentry.d_inode */
 501         unsigned int    flags;
 502         unsigned        seq, m_seq;
 503         int             last_type;
 504         unsigned        depth;
 505         int             total_link_count;
 506         struct saved {
 507                 struct path link;
 508                 struct delayed_call done;
 509                 const char *name;
 510                 unsigned seq;
 511         } *stack, internal[EMBEDDED_LEVELS];
 512         struct filename *name;
 513         struct nameidata *saved;
 514         struct inode    *link_inode;
 515         unsigned        root_seq;
 516         int             dfd;
 517 };
 518
 519 static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
 520 {
 521         struct nameidata *old = current->nameidata;
 522         p->stack = p->internal;
 523         p->dfd = dfd;
 524         p->name = name;
 525         p->total_link_count = old ? old->total_link_count : 0;
 526         p->saved = old;
 527         current->nameidata = p;
 528 }
 529
 530 static void restore_nameidata(void)
 531 {
 532         struct nameidata *now = current->nameidata, *old = now->saved;
 533
 534         current->nameidata = old;
 535         if (old)
 536                 old->total_link_count = now->total_link_count;
 537         if (now->stack != now->internal)
 538                 kfree(now->stack);
 539 }
 540
 541 static int __nd_alloc_stack(struct nameidata *nd)
 542 {
 543         struct saved *p;
 544
 545         if (nd->flags & LOOKUP_RCU) {
 546                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
 547                                   GFP_ATOMIC);
 548                 if (unlikely(!p))
 549                         return -ECHILD;
 550         } else {
 551                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
 552                                   GFP_KERNEL);
 553                 if (unlikely(!p))
 554                         return -ENOMEM;
 555         }
 556         memcpy(p, nd->internal, sizeof(nd->internal));
 557         nd->stack = p;
 558         return 0;
 559 }
 560
 561 /**
 562  * path_connected - Verify that a path->dentry is below path->mnt.mnt_root
 563  * @path: nameidate to verify
 564  *
 565  * Rename can sometimes move a file or directory outside of a bind
 566  * mount, path_connected allows those cases to be detected.
 567  */
 568 static bool path_connected(const struct path *path)
 569 {
 570         struct vfsmount *mnt = path->mnt;
 571
 572         /* Only bind mounts can have disconnected paths */
 573         if (mnt->mnt_root == mnt->mnt_sb->s_root)
 574                 return true;
 575
 576         return is_subdir(path->dentry, mnt->mnt_root);
 577 }
 578
 579 static inline int nd_alloc_stack(struct nameidata *nd)
 580 {
 581         if (likely(nd->depth != EMBEDDED_LEVELS))
 582                 return 0;
 583         if (likely(nd->stack != nd->internal))
 584                 return 0;
 585         return __nd_alloc_stack(nd);
 586 }
 587
 588 static void drop_links(struct nameidata *nd)
 589 {
 590         int i = nd->depth;
 591         while (i--) {
 592                 struct saved *last = nd->stack + i;
 593                 do_delayed_call(&last->done);
 594                 clear_delayed_call(&last->done);
 595         }
 596 }
 597
 598 static void terminate_walk(struct nameidata *nd)
 599 {
 600         drop_links(nd);
 601         if (!(nd->flags & LOOKUP_RCU)) {
 602                 int i;
 603                 path_put(&nd->path);
 604                 for (i = 0; i < nd->depth; i++)
 605                         path_put(&nd->stack[i].link);
 606                 if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 607                         path_put(&nd->root);
 608                         nd->root.mnt = NULL;
 609                 }
 610         } else {
 611                 nd->flags &= ~LOOKUP_RCU;
 612                 if (!(nd->flags & LOOKUP_ROOT))
 613                         nd->root.mnt = NULL;
 614                 rcu_read_unlock();
 615         }
 616         nd->depth = 0;
 617 }
 618
 619 /* path_put is needed afterwards regardless of success or failure */
 620 static bool legitimize_path(struct nameidata *nd,
 621                             struct path *path, unsigned seq)
 622 {
 623         int res = __legitimize_mnt(path->mnt, nd->m_seq);
 624         if (unlikely(res)) {
 625                 if (res > 0)
 626                         path->mnt = NULL;
 627                 path->dentry = NULL;
 628                 return false;
 629         }
 630         if (unlikely(!lockref_get_not_dead(&path->dentry->d_lockref))) {
 631                 path->dentry = NULL;
 632                 return false;
 633         }
 634         return !read_seqcount_retry(&path->dentry->d_seq, seq);
 635 }
 636
 637 static bool legitimize_links(struct nameidata *nd)
 638 {
 639         int i;
 640         for (i = 0; i < nd->depth; i++) {
 641                 struct saved *last = nd->stack + i;
 642                 if (unlikely(!legitimize_path(nd, &last->link, last->seq))) {
 643                         drop_links(nd);
 644                         nd->depth = i + 1;
 645                         return false;
 646                 }
 647         }
 648         return true;
 649 }
 650
 651 /*
 652  * Path walking has 2 modes, rcu-walk and ref-walk (see
 653  * Documentation/filesystems/path-lookup.txt).  In situations when we can't
 654  * continue in RCU mode, we attempt to drop out of rcu-walk mode and grab
 655  * normal reference counts on dentries and vfsmounts to transition to ref-walk
 656  * mode.  Refcounts are grabbed at the last known good point before rcu-walk
 657  * got stuck, so ref-walk may continue from there. If this is not successful
 658  * (eg. a seqcount has changed), then failure is returned and it's up to caller
 659  * to restart the path walk from the beginning in ref-walk mode.
 660  */
 661
 662 /**
 663  * unlazy_walk - try to switch to ref-walk mode.
 664  * @nd: nameidata pathwalk data
 665  * @dentry: child of nd->path.dentry or NULL
 666  * @seq: seq number to check dentry against
 667  * Returns: 0 on success, -ECHILD on failure
 668  *
 669  * unlazy_walk attempts to legitimize the current nd->path, nd->root and dentry
 670  * for ref-walk mode.  @dentry must be a path found by a do_lookup call on
 671  * @nd or NULL.  Must be called from rcu-walk context.
 672  * Nothing should touch nameidata between unlazy_walk() failure and
 673  * terminate_walk().
 674  */
 675 static int unlazy_walk(struct nameidata *nd, struct dentry *dentry, unsigned seq)
 676 {
 677         struct dentry *parent = nd->path.dentry;
 678
 679         BUG_ON(!(nd->flags & LOOKUP_RCU));
 680
 681         nd->flags &= ~LOOKUP_RCU;
 682         if (unlikely(!legitimize_links(nd)))
 683                 goto out2;
 684         if (unlikely(!legitimize_mnt(nd->path.mnt, nd->m_seq)))
 685                 goto out2;
 686         if (unlikely(!lockref_get_not_dead(&parent->d_lockref)))
 687                 goto out1;
 688
 689         /*
 690          * For a negative lookup, the lookup sequence point is the parents
 691          * sequence point, and it only needs to revalidate the parent dentry.
 692          *
 693          * For a positive lookup, we need to move both the parent and the
 694          * dentry from the RCU domain to be properly refcounted. And the
 695          * sequence number in the dentry validates *both* dentry counters,
 696          * since we checked the sequence number of the parent after we got
 697          * the child sequence number. So we know the parent must still
 698          * be valid if the child sequence number is still valid.
 699          */
 700         if (!dentry) {
 701                 if (read_seqcount_retry(&parent->d_seq, nd->seq))
 702                         goto out;
 703                 BUG_ON(nd->inode != parent->d_inode);
 704         } else {
 705                 if (!lockref_get_not_dead(&dentry->d_lockref))
 706                         goto out;
 707                 if (read_seqcount_retry(&dentry->d_seq, seq))
 708                         goto drop_dentry;
 709         }
 710
 711         /*
 712          * Sequence counts matched. Now make sure that the root is
 713          * still valid and get it if required.
 714          */
 715         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 716                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq))) {
 717                         rcu_read_unlock();
 718                         dput(dentry);
 719                         return -ECHILD;
 720                 }
 721         }
 722
 723         rcu_read_unlock();
 724         return 0;
 725
 726 drop_dentry:
 727         rcu_read_unlock();
 728         dput(dentry);
 729         goto drop_root_mnt;
 730 out2:
 731         nd->path.mnt = NULL;
 732 out1:
 733         nd->path.dentry = NULL;
 734 out:
 735         rcu_read_unlock();
 736 drop_root_mnt:
 737         if (!(nd->flags & LOOKUP_ROOT))
 738                 nd->root.mnt = NULL;
 739         return -ECHILD;
 740 }
 741
 742 static int unlazy_link(struct nameidata *nd, struct path *link, unsigned seq)
 743 {
 744         if (unlikely(!legitimize_path(nd, link, seq))) {
 745                 drop_links(nd);
 746                 nd->depth = 0;
 747                 nd->flags &= ~LOOKUP_RCU;
 748                 nd->path.mnt = NULL;
 749                 nd->path.dentry = NULL;
 750                 if (!(nd->flags & LOOKUP_ROOT))
 751                         nd->root.mnt = NULL;
 752                 rcu_read_unlock();
 753         } else if (likely(unlazy_walk(nd, NULL, 0)) == 0) {
 754                 return 0;
 755         }
 756         path_put(link);
 757         return -ECHILD;
 758 }
 759
 760 static inline int d_revalidate(struct dentry *dentry, unsigned int flags)
 761 {
 762         return dentry->d_op->d_revalidate(dentry, flags);
 763 }
 764
 765 /**
 766  * complete_walk - successful completion of path walk
 767  * @nd:  pointer nameidata
 768  *
 769  * If we had been in RCU mode, drop out of it and legitimize nd->path.
 770  * Revalidate the final result, unless we'd already done that during
 771  * the path walk or the filesystem doesn't ask for it.  Return 0 on
 772  * success, -error on failure.  In case of failure caller does not
 773  * need to drop nd->path.
 774  */
 775 static int complete_walk(struct nameidata *nd)
 776 {
 777         struct dentry *dentry = nd->path.dentry;
 778         int status;
 779
 780         if (nd->flags & LOOKUP_RCU) {
 781                 if (!(nd->flags & LOOKUP_ROOT))
 782                         nd->root.mnt = NULL;
 783                 if (unlikely(unlazy_walk(nd, NULL, 0)))
 784                         return -ECHILD;
 785         }
 786
 787         if (likely(!(nd->flags & LOOKUP_JUMPED)))
 788                 return 0;
 789
 790         if (likely(!(dentry->d_flags & DCACHE_OP_WEAK_REVALIDATE)))
 791                 return 0;
 792
 793         status = dentry->d_op->d_weak_revalidate(dentry, nd->flags);
 794         if (status > 0)
 795                 return 0;
 796
 797         if (!status)
 798                 status = -ESTALE;
 799
 800         return status;
 801 }
 802
 803 static void set_root(struct nameidata *nd)
 804 {
 805         struct fs_struct *fs = current->fs;
 806
 807         if (nd->flags & LOOKUP_RCU) {
 808                 unsigned seq;
 809
 810                 do {
 811                         seq = read_seqcount_begin(&fs->seq);
 812                         nd->root = fs->root;
 813                         nd->root_seq = __read_seqcount_begin(&nd->root.dentry->d_seq);
 814                 } while (read_seqcount_retry(&fs->seq, seq));
 815         } else {
 816                 get_fs_root(fs, &nd->root);
 817         }
 818 }
 819
 820 static void path_put_conditional(struct path *path, struct nameidata *nd)
 821 {
 822         dput(path->dentry);
 823         if (path->mnt != nd->path.mnt)
 824                 mntput(path->mnt);
 825 }
 826
 827 static inline void path_to_nameidata(const struct path *path,
 828                                         struct nameidata *nd)
 829 {
 830         if (!(nd->flags & LOOKUP_RCU)) {
 831                 dput(nd->path.dentry);
 832                 if (nd->path.mnt != path->mnt)
 833                         mntput(nd->path.mnt);
 834         }
 835         nd->path.mnt = path->mnt;
 836         nd->path.dentry = path->dentry;
 837 }
 838
 839 static int nd_jump_root(struct nameidata *nd)
 840 {
 841         if (nd->flags & LOOKUP_RCU) {
 842                 struct dentry *d;
 843                 nd->path = nd->root;
 844                 d = nd->path.dentry;
 845                 nd->inode = d->d_inode;
 846                 nd->seq = nd->root_seq;
 847                 if (unlikely(read_seqcount_retry(&d->d_seq, nd->seq)))
 848                         return -ECHILD;
 849         } else {
 850                 path_put(&nd->path);
 851                 nd->path = nd->root;
 852                 path_get(&nd->path);
 853                 nd->inode = nd->path.dentry->d_inode;
 854         }
 855         nd->flags |= LOOKUP_JUMPED;
 856         return 0;
 857 }
 858
 859 /*
 860  * Helper to directly jump to a known parsed path from ->get_link,
 861  * caller must have taken a reference to path beforehand.
 862  */
 863 void nd_jump_link(struct path *path)
 864 {
 865         struct nameidata *nd = current->nameidata;
 866         path_put(&nd->path);
 867
 868         nd->path = *path;
 869         nd->inode = nd->path.dentry->d_inode;
 870         nd->flags |= LOOKUP_JUMPED;
 871 }
 872
 873 static inline void put_link(struct nameidata *nd)
 874 {
 875         struct saved *last = nd->stack + --nd->depth;
 876         do_delayed_call(&last->done);
 877         if (!(nd->flags & LOOKUP_RCU))
 878                 path_put(&last->link);
 879 }
 880
 881 int sysctl_protected_symlinks __read_mostly = 0;
 882 int sysctl_protected_hardlinks __read_mostly = 0;
 883
 884 /**
 885  * may_follow_link - Check symlink following for unsafe situations
 886  * @nd: nameidata pathwalk data
 887  *
 888  * In the case of the sysctl_protected_symlinks sysctl being enabled,
 889  * CAP_DAC_OVERRIDE needs to be specifically ignored if the symlink is
 890  * in a sticky world-writable directory. This is to protect privileged
 891  * processes from failing races against path names that may change out
 892  * from under them by way of other users creating malicious symlinks.
 893  * It will permit symlinks to be followed only when outside a sticky
 894  * world-writable directory, or when the uid of the symlink and follower
 895  * match, or when the directory owner matches the symlink's owner.
 896  *
 897  * Returns 0 if following the symlink is allowed, -ve on error.
 898  */
 899 static inline int may_follow_link(struct nameidata *nd)
 900 {
 901         const struct inode *inode;
 902         const struct inode *parent;
 903
 904         if (!sysctl_protected_symlinks)
 905                 return 0;
 906
 907         /* Allowed if owner and follower match. */
 908         inode = nd->link_inode;
 909         if (uid_eq(current_cred()->fsuid, inode->i_uid))
 910                 return 0;
 911
 912         /* Allowed if parent directory not sticky and world-writable. */
 913         parent = nd->inode;
 914         if ((parent->i_mode & (S_ISVTX|S_IWOTH)) != (S_ISVTX|S_IWOTH))
 915                 return 0;
 916
 917         /* Allowed if parent directory and link owner match. */
 918         if (uid_eq(parent->i_uid, inode->i_uid))
 919                 return 0;
 920
 921         if (nd->flags & LOOKUP_RCU)
 922                 return -ECHILD;
 923
 924         audit_log_link_denied("follow_link", &nd->stack[0].link);
 925         return -EACCES;
 926 }
 927
 928 /**
 929  * safe_hardlink_source - Check for safe hardlink conditions
 930  * @inode: the source inode to hardlink from
 931  *
 932  * Return false if at least one of the following conditions:
 933  *    - inode is not a regular file
 934  *    - inode is setuid
 935  *    - inode is setgid and group-exec
 936  *    - access failure for read and write
 937  *
 938  * Otherwise returns true.
 939  */
 940 static bool safe_hardlink_source(struct inode *inode)
 941 {
 942         umode_t mode = inode->i_mode;
 943
 944         /* Special files should not get pinned to the filesystem. */
 945         if (!S_ISREG(mode))
 946                 return false;
 947
 948         /* Setuid files should not get pinned to the filesystem. */
 949         if (mode & S_ISUID)
 950                 return false;
 951
 952         /* Executable setgid files should not get pinned to the filesystem. */
 953         if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP))
 954                 return false;
 955
 956         /* Hardlinking to unreadable or unwritable sources is dangerous. */
 957         if (inode_permission(inode, MAY_READ | MAY_WRITE))
 958                 return false;
 959
 960         return true;
 961 }
 962
 963 /**
 964  * may_linkat - Check permissions for creating a hardlink
 965  * @link: the source to hardlink from
 966  *
 967  * Block hardlink when all of:
 968  *  - sysctl_protected_hardlinks enabled
 969  *  - fsuid does not match inode
 970  *  - hardlink source is unsafe (see safe_hardlink_source() above)
 971  *  - not CAP_FOWNER in a namespace with the inode owner uid mapped
 972  *
 973  * Returns 0 if successful, -ve on error.
 974  */
 975 static int may_linkat(struct path *link)
 976 {
 977         struct inode *inode;
 978
 979         if (!sysctl_protected_hardlinks)
 980                 return 0;
 981
 982         inode = link->dentry->d_inode;
 983
 984         /* Source inode owner (or CAP_FOWNER) can hardlink all they like,
 985          * otherwise, it must be a safe source.
 986          */
 987         if (inode_owner_or_capable(inode) || safe_hardlink_source(inode))
 988                 return 0;
 989
 990         audit_log_link_denied("linkat", link);
 991         return -EPERM;
 992 }
 993
 994 static __always_inline
 995 const char *get_link(struct nameidata *nd)
 996 {
 997         struct saved *last = nd->stack + nd->depth - 1;
 998         struct dentry *dentry = last->link.dentry;
 999         struct inode *inode = nd->link_inode;
1000         int error;
1001         const char *res;
1002
1003         if (!(nd->flags & LOOKUP_RCU)) {
1004                 touch_atime(&last->link);
1005                 cond_resched();
1006         } else if (atime_needs_update(&last->link, inode)) {
1007                 if (unlikely(unlazy_walk(nd, NULL, 0)))
1008                         return ERR_PTR(-ECHILD);
1009                 touch_atime(&last->link);
1010         }
1011
1012         error = security_inode_follow_link(dentry, inode,
1013                                            nd->flags & LOOKUP_RCU);
1014         if (unlikely(error))
1015                 return ERR_PTR(error);
1016
1017         nd->last_type = LAST_BIND;
1018         res = inode->i_link;
1019         if (!res) {
1020                 const char * (*get)(struct dentry *, struct inode *,
1021                                 struct delayed_call *);
1022                 get = inode->i_op->get_link;
1023                 if (nd->flags & LOOKUP_RCU) {
1024                         res = get(NULL, inode, &last->done);
1025                         if (res == ERR_PTR(-ECHILD)) {
1026                                 if (unlikely(unlazy_walk(nd, NULL, 0)))
1027                                         return ERR_PTR(-ECHILD);
1028                                 res = get(dentry, inode, &last->done);
1029                         }
1030                 } else {
1031                         res = get(dentry, inode, &last->done);
1032                 }
1033                 if (IS_ERR_OR_NULL(res))
1034                         return res;
1035         }
1036         if (*res == '/') {
1037                 if (!nd->root.mnt)
1038                         set_root(nd);
1039                 if (unlikely(nd_jump_root(nd)))
1040                         return ERR_PTR(-ECHILD);
1041                 while (unlikely(*++res == '/'))
1042                         ;
1043         }
1044         if (!*res)
1045                 res = NULL;
1046         return res;
1047 }
1048
1049 /*
1050  * follow_up - Find the mountpoint of path's vfsmount
1051  *
1052  * Given a path, find the mountpoint of its source file system.
1053  * Replace @path with the path of the mountpoint in the parent mount.
1054  * Up is towards /.
1055  *
1056  * Return 1 if we went up a level and 0 if we were already at the
1057  * root.
1058  */
1059 int follow_up(struct path *path)
1060 {
1061         struct mount *mnt = real_mount(path->mnt);
1062         struct mount *parent;
1063         struct dentry *mountpoint;
1064
1065         read_seqlock_excl(&mount_lock);
1066         parent = mnt->mnt_parent;
1067         if (parent == mnt) {
1068                 read_sequnlock_excl(&mount_lock);
1069                 return 0;
1070         }
1071         mntget(&parent->mnt);
1072         mountpoint = dget(mnt->mnt_mountpoint);
1073         read_sequnlock_excl(&mount_lock);
1074         dput(path->dentry);
1075         path->dentry = mountpoint;
1076         mntput(path->mnt);
1077         path->mnt = &parent->mnt;
1078         return 1;
1079 }
1080 EXPORT_SYMBOL(follow_up);
1081
1082 /*
1083  * Perform an automount
1084  * - return -EISDIR to tell follow_managed() to stop and return the path we
1085  *   were called with.
1086  */
1087 static int follow_automount(struct path *path, struct nameidata *nd,
1088                             bool *need_mntput)
1089 {
1090         struct vfsmount *mnt;
1091         int err;
1092
1093         if (!path->dentry->d_op || !path->dentry->d_op->d_automount)
1094                 return -EREMOTE;
1095
1096         /* We don't want to mount if someone's just doing a stat -
1097          * unless they're stat'ing a directory and appended a '/' to
1098          * the name.
1099          *
1100          * We do, however, want to mount if someone wants to open or
1101          * create a file of any type under the mountpoint, wants to
1102          * traverse through the mountpoint or wants to open the
1103          * mounted directory.  Also, autofs may mark negative dentries
1104          * as being automount points.  These will need the attentions
1105          * of the daemon to instantiate them before they can be used.
1106          */
1107         if (!(nd->flags & (LOOKUP_PARENT | LOOKUP_DIRECTORY |
1108                            LOOKUP_OPEN | LOOKUP_CREATE | LOOKUP_AUTOMOUNT)) &&
1109             path->dentry->d_inode)
1110                 return -EISDIR;
1111
1112         nd->total_link_count++;
1113         if (nd->total_link_count >= 40)
1114                 return -ELOOP;
1115
1116         mnt = path->dentry->d_op->d_automount(path);
1117         if (IS_ERR(mnt)) {
1118                 /*
1119                  * The filesystem is allowed to return -EISDIR here to indicate
1120                  * it doesn't want to automount.  For instance, autofs would do
1121                  * this so that its userspace daemon can mount on this dentry.
1122                  *
1123                  * However, we can only permit this if it's a terminal point in
1124                  * the path being looked up; if it wasn't then the remainder of
1125                  * the path is inaccessible and we should say so.
1126                  */
1127                 if (PTR_ERR(mnt) == -EISDIR && (nd->flags & LOOKUP_PARENT))
1128                         return -EREMOTE;
1129                 return PTR_ERR(mnt);
1130         }
1131
1132         if (!mnt) /* mount collision */
1133                 return 0;
1134
1135         if (!*need_mntput) {
1136                 /* lock_mount() may release path->mnt on error */
1137                 mntget(path->mnt);
1138                 *need_mntput = true;
1139         }
1140         err = finish_automount(mnt, path);
1141
1142         switch (err) {
1143         case -EBUSY:
1144                 /* Someone else made a mount here whilst we were busy */
1145                 return 0;
1146         case 0:
1147                 path_put(path);
1148                 path->mnt = mnt;
1149                 path->dentry = dget(mnt->mnt_root);
1150                 return 0;
1151         default:
1152                 return err;
1153         }
1154
1155 }
1156
1157 /*
1158  * Handle a dentry that is managed in some way.
1159  * - Flagged for transit management (autofs)
1160  * - Flagged as mountpoint
1161  * - Flagged as automount point
1162  *
1163  * This may only be called in refwalk mode.
1164  *
1165  * Serialization is taken care of in namespace.c
1166  */
1167 static int follow_managed(struct path *path, struct nameidata *nd)
1168 {
1169         struct vfsmount *mnt = path->mnt; /* held by caller, must be left alone */
1170         unsigned managed;
1171         bool need_mntput = false;
1172         int ret = 0;
1173
1174         /* Given that we're not holding a lock here, we retain the value in a
1175          * local variable for each dentry as we look at it so that we don't see
1176          * the components of that value change under us */
1177         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1178                managed &= DCACHE_MANAGED_DENTRY,
1179                unlikely(managed != 0)) {
1180                 /* Allow the filesystem to manage the transit without i_mutex
1181                  * being held. */
1182                 if (managed & DCACHE_MANAGE_TRANSIT) {
1183                         BUG_ON(!path->dentry->d_op);
1184                         BUG_ON(!path->dentry->d_op->d_manage);
1185                         ret = path->dentry->d_op->d_manage(path->dentry, false);
1186                         if (ret < 0)
1187                                 break;
1188                 }
1189
1190                 /* Transit to a mounted filesystem. */
1191                 if (managed & DCACHE_MOUNTED) {
1192                         struct vfsmount *mounted = lookup_mnt(path);
1193                         if (mounted) {
1194                                 dput(path->dentry);
1195                                 if (need_mntput)
1196                                         mntput(path->mnt);
1197                                 path->mnt = mounted;
1198                                 path->dentry = dget(mounted->mnt_root);
1199                                 need_mntput = true;
1200                                 continue;
1201                         }
1202
1203                         /* Something is mounted on this dentry in another
1204                          * namespace and/or whatever was mounted there in this
1205                          * namespace got unmounted before lookup_mnt() could
1206                          * get it */
1207                 }
1208
1209                 /* Handle an automount point */
1210                 if (managed & DCACHE_NEED_AUTOMOUNT) {
1211                         ret = follow_automount(path, nd, &need_mntput);
1212                         if (ret < 0)
1213                                 break;
1214                         continue;
1215                 }
1216
1217                 /* We didn't change the current path point */
1218                 break;
1219         }
1220
1221         if (need_mntput && path->mnt == mnt)
1222                 mntput(path->mnt);
1223         if (ret == -EISDIR || !ret)
1224                 ret = 1;
1225         if (need_mntput)
1226                 nd->flags |= LOOKUP_JUMPED;
1227         if (unlikely(ret < 0))
1228                 path_put_conditional(path, nd);
1229         return ret;
1230 }
1231
1232 int follow_down_one(struct path *path)
1233 {
1234         struct vfsmount *mounted;
1235
1236         mounted = lookup_mnt(path);
1237         if (mounted) {
1238                 dput(path->dentry);
1239                 mntput(path->mnt);
1240                 path->mnt = mounted;
1241                 path->dentry = dget(mounted->mnt_root);
1242                 return 1;
1243         }
1244         return 0;
1245 }
1246 EXPORT_SYMBOL(follow_down_one);
1247
1248 static inline int managed_dentry_rcu(struct dentry *dentry)
1249 {
1250         return (dentry->d_flags & DCACHE_MANAGE_TRANSIT) ?
1251                 dentry->d_op->d_manage(dentry, true) : 0;
1252 }
1253
1254 /*
1255  * Try to skip to top of mountpoint pile in rcuwalk mode.  Fail if
1256  * we meet a managed dentry that would need blocking.
1257  */
1258 static bool __follow_mount_rcu(struct nameidata *nd, struct path *path,
1259                                struct inode **inode, unsigned *seqp)
1260 {
1261         for (;;) {
1262                 struct mount *mounted;
1263                 /*
1264                  * Don't forget we might have a non-mountpoint managed dentry
1265                  * that wants to block transit.
1266                  */
1267                 switch (managed_dentry_rcu(path->dentry)) {
1268                 case -ECHILD:
1269                 default:
1270                         return false;
1271                 case -EISDIR:
1272                         return true;
1273                 case 0:
1274                         break;
1275                 }
1276
1277                 if (!d_mountpoint(path->dentry))
1278                         return !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1279
1280                 mounted = __lookup_mnt(path->mnt, path->dentry);
1281                 if (!mounted)
1282                         break;
1283                 path->mnt = &mounted->mnt;
1284                 path->dentry = mounted->mnt.mnt_root;
1285                 nd->flags |= LOOKUP_JUMPED;
1286                 *seqp = read_seqcount_begin(&path->dentry->d_seq);
1287                 /*
1288                  * Update the inode too. We don't need to re-check the
1289                  * dentry sequence number here after this d_inode read,
1290                  * because a mount-point is always pinned.
1291                  */
1292                 *inode = path->dentry->d_inode;
1293         }
1294         return !read_seqretry(&mount_lock, nd->m_seq) &&
1295                 !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1296 }
1297
1298 static int follow_dotdot_rcu(struct nameidata *nd)
1299 {
1300         struct inode *inode = nd->inode;
1301
1302         while (1) {
1303                 if (path_equal(&nd->path, &nd->root))
1304                         break;
1305                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1306                         struct dentry *old = nd->path.dentry;
1307                         struct dentry *parent = old->d_parent;
1308                         unsigned seq;
1309
1310                         inode = parent->d_inode;
1311                         seq = read_seqcount_begin(&parent->d_seq);
1312                         if (unlikely(read_seqcount_retry(&old->d_seq, nd->seq)))
1313                                 return -ECHILD;
1314                         nd->path.dentry = parent;
1315                         nd->seq = seq;
1316                         if (unlikely(!path_connected(&nd->path)))
1317                                 return -ENOENT;
1318                         break;
1319                 } else {
1320                         struct mount *mnt = real_mount(nd->path.mnt);
1321                         struct mount *mparent = mnt->mnt_parent;
1322                         struct dentry *mountpoint = mnt->mnt_mountpoint;
1323                         struct inode *inode2 = mountpoint->d_inode;
1324                         unsigned seq = read_seqcount_begin(&mountpoint->d_seq);
1325                         if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1326                                 return -ECHILD;
1327                         if (&mparent->mnt == nd->path.mnt)
1328                                 break;
1329                         /* we know that mountpoint was pinned */
1330                         nd->path.dentry = mountpoint;
1331                         nd->path.mnt = &mparent->mnt;
1332                         inode = inode2;
1333                         nd->seq = seq;
1334                 }
1335         }
1336         while (unlikely(d_mountpoint(nd->path.dentry))) {
1337                 struct mount *mounted;
1338                 mounted = __lookup_mnt(nd->path.mnt, nd->path.dentry);
1339                 if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1340                         return -ECHILD;
1341                 if (!mounted)
1342                         break;
1343                 nd->path.mnt = &mounted->mnt;
1344                 nd->path.dentry = mounted->mnt.mnt_root;
1345                 inode = nd->path.dentry->d_inode;
1346                 nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
1347         }
1348         nd->inode = inode;
1349         return 0;
1350 }
1351
1352 /*
1353  * Follow down to the covering mount currently visible to userspace.  At each
1354  * point, the filesystem owning that dentry may be queried as to whether the
1355  * caller is permitted to proceed or not.
1356  */
1357 int follow_down(struct path *path)
1358 {
1359         unsigned managed;
1360         int ret;
1361
1362         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1363                unlikely(managed & DCACHE_MANAGED_DENTRY)) {
1364                 /* Allow the filesystem to manage the transit without i_mutex
1365                  * being held.
1366                  *
1367                  * We indicate to the filesystem if someone is trying to mount
1368                  * something here.  This gives autofs the chance to deny anyone
1369                  * other than its daemon the right to mount on its
1370                  * superstructure.
1371                  *
1372                  * The filesystem may sleep at this point.
1373                  */
1374                 if (managed & DCACHE_MANAGE_TRANSIT) {
1375                         BUG_ON(!path->dentry->d_op);
1376                         BUG_ON(!path->dentry->d_op->d_manage);
1377                         ret = path->dentry->d_op->d_manage(
1378                                 path->dentry, false);
1379                         if (ret < 0)
1380                                 return ret == -EISDIR ? 0 : ret;
1381                 }
1382
1383                 /* Transit to a mounted filesystem. */
1384                 if (managed & DCACHE_MOUNTED) {
1385                         struct vfsmount *mounted = lookup_mnt(path);
1386                         if (!mounted)
1387                                 break;
1388                         dput(path->dentry);
1389                         mntput(path->mnt);
1390                         path->mnt = mounted;
1391                         path->dentry = dget(mounted->mnt_root);
1392                         continue;
1393                 }
1394
1395                 /* Don't handle automount points here */
1396                 break;
1397         }
1398         return 0;
1399 }
1400 EXPORT_SYMBOL(follow_down);
1401
1402 /*
1403  * Skip to top of mountpoint pile in refwalk mode for follow_dotdot()
1404  */
1405 static void follow_mount(struct path *path)
1406 {
1407         while (d_mountpoint(path->dentry)) {
1408                 struct vfsmount *mounted = lookup_mnt(path);
1409                 if (!mounted)
1410                         break;
1411                 dput(path->dentry);
1412                 mntput(path->mnt);
1413                 path->mnt = mounted;
1414                 path->dentry = dget(mounted->mnt_root);
1415         }
1416 }
1417
1418 static int follow_dotdot(struct nameidata *nd)
1419 {
1420         while(1) {
1421                 struct dentry *old = nd->path.dentry;
1422
1423                 if (nd->path.dentry == nd->root.dentry &&
1424                     nd->path.mnt == nd->root.mnt) {
1425                         break;
1426                 }
1427                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1428                         /* rare case of legitimate dget_parent()... */
1429                         nd->path.dentry = dget_parent(nd->path.dentry);
1430                         dput(old);
1431                         if (unlikely(!path_connected(&nd->path)))
1432                                 return -ENOENT;
1433                         break;
1434                 }
1435                 if (!follow_up(&nd->path))
1436                         break;
1437         }
1438         follow_mount(&nd->path);
1439         nd->inode = nd->path.dentry->d_inode;
1440         return 0;
1441 }
1442
1443 /*
1444  * This looks up the name in dcache, possibly revalidates the old dentry and
1445  * allocates a new one if not found or not valid.  In the need_lookup argument
1446  * returns whether i_op->lookup is necessary.
1447  */
1448 static struct dentry *lookup_dcache(const struct qstr *name,
1449                                     struct dentry *dir,
1450                                     unsigned int flags)
1451 {
1452         struct dentry *dentry;
1453         int error;
1454
1455         dentry = d_lookup(dir, name);
1456         if (dentry) {
1457                 if (dentry->d_flags & DCACHE_OP_REVALIDATE) {
1458                         error = d_revalidate(dentry, flags);
1459                         if (unlikely(error <= 0)) {
1460                                 if (!error)
1461                                         d_invalidate(dentry);
1462                                 dput(dentry);
1463                                 return ERR_PTR(error);
1464                         }
1465                 }
1466         }
1467         return dentry;
1468 }
1469
1470 /*
1471  * Call i_op->lookup on the dentry.  The dentry must be negative and
1472  * unhashed.
1473  *
1474  * dir->d_inode->i_mutex must be held
1475  */
1476 static struct dentry *lookup_real(struct inode *dir, struct dentry *dentry,
1477                                   unsigned int flags)
1478 {
1479         struct dentry *old;
1480
1481         /* Don't create child dentry for a dead directory. */
1482         if (unlikely(IS_DEADDIR(dir))) {
1483                 dput(dentry);
1484                 return ERR_PTR(-ENOENT);
1485         }
1486
1487         old = dir->i_op->lookup(dir, dentry, flags);
1488         if (unlikely(old)) {
1489                 dput(dentry);
1490                 dentry = old;
1491         }
1492         return dentry;
1493 }
1494
1495 static struct dentry *__lookup_hash(const struct qstr *name,
1496                 struct dentry *base, unsigned int flags)
1497 {
1498         struct dentry *dentry = lookup_dcache(name, base, flags);
1499
1500         if (dentry)
1501                 return dentry;
1502
1503         dentry = d_alloc(base, name);
1504         if (unlikely(!dentry))
1505                 return ERR_PTR(-ENOMEM);
1506
1507         return lookup_real(base->d_inode, dentry, flags);
1508 }
1509
1510 static int lookup_fast(struct nameidata *nd,
1511                        struct path *path, struct inode **inode,
1512                        unsigned *seqp)
1513 {
1514         struct vfsmount *mnt = nd->path.mnt;
1515         struct dentry *dentry, *parent = nd->path.dentry;
1516         int status = 1;
1517         int err;
1518
1519         /*
1520          * Rename seqlock is not required here because in the off chance
1521          * of a false negative due to a concurrent rename, the caller is
1522          * going to fall back to non-racy lookup.
1523          */
1524         if (nd->flags & LOOKUP_RCU) {
1525                 unsigned seq;
1526                 bool negative;
1527                 dentry = __d_lookup_rcu(parent, &nd->last, &seq);
1528                 if (unlikely(!dentry)) {
1529                         if (unlazy_walk(nd, NULL, 0))
1530                                 return -ECHILD;
1531                         return 0;
1532                 }
1533
1534                 /*
1535                  * This sequence count validates that the inode matches
1536                  * the dentry name information from lookup.
1537                  */
1538                 *inode = d_backing_inode(dentry);
1539                 negative = d_is_negative(dentry);
1540                 if (unlikely(read_seqcount_retry(&dentry->d_seq, seq)))
1541                         return -ECHILD;
1542
1543                 /*
1544                  * This sequence count validates that the parent had no
1545                  * changes while we did the lookup of the dentry above.
1546                  *
1547                  * The memory barrier in read_seqcount_begin of child is
1548                  *  enough, we can use __read_seqcount_retry here.
1549                  */
1550                 if (unlikely(__read_seqcount_retry(&parent->d_seq, nd->seq)))
1551                         return -ECHILD;
1552
1553                 *seqp = seq;
1554                 if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
1555                         status = d_revalidate(dentry, nd->flags);
1556                 if (unlikely(status <= 0)) {
1557                         if (unlazy_walk(nd, dentry, seq))
1558                                 return -ECHILD;
1559                         if (status == -ECHILD)
1560                                 status = d_revalidate(dentry, nd->flags);
1561                 } else {
1562                         /*
1563                          * Note: do negative dentry check after revalidation in
1564                          * case that drops it.
1565                          */
1566                         if (unlikely(negative))
1567                                 return -ENOENT;
1568                         path->mnt = mnt;
1569                         path->dentry = dentry;
1570                         if (likely(__follow_mount_rcu(nd, path, inode, seqp)))
1571                                 return 1;
1572                         if (unlazy_walk(nd, dentry, seq))
1573                                 return -ECHILD;
1574                 }
1575         } else {
1576                 dentry = __d_lookup(parent, &nd->last);
1577                 if (unlikely(!dentry))
1578                         return 0;
1579                 if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
1580                         status = d_revalidate(dentry, nd->flags);
1581         }
1582         if (unlikely(status <= 0)) {
1583                 if (!status)
1584                         d_invalidate(dentry);
1585                 dput(dentry);
1586                 return status;
1587         }
1588         if (unlikely(d_is_negative(dentry))) {
1589                 dput(dentry);
1590                 return -ENOENT;
1591         }
1592
1593         path->mnt = mnt;
1594         path->dentry = dentry;
1595         err = follow_managed(path, nd);
1596         if (likely(err > 0))
1597                 *inode = d_backing_inode(path->dentry);
1598         return err;
1599 }
1600
1601 /* Fast lookup failed, do it the slow way */
1602 static struct dentry *lookup_slow(const struct qstr *name,
1603                                   struct dentry *dir,
1604                                   unsigned int flags)
1605 {
1606         struct dentry *dentry = ERR_PTR(-ENOENT), *old;
1607         struct inode *inode = dir->d_inode;
1608         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
1609
1610         inode_lock_shared(inode);
1611         /* Don't go there if it's already dead */
1612         if (unlikely(IS_DEADDIR(inode)))
1613                 goto out;
1614 again:
1615         dentry = d_alloc_parallel(dir, name, &wq);
1616         if (IS_ERR(dentry))
1617                 goto out;
1618         if (unlikely(!d_in_lookup(dentry))) {
1619                 if ((dentry->d_flags & DCACHE_OP_REVALIDATE) &&
1620                     !(flags & LOOKUP_NO_REVAL)) {
1621                         int error = d_revalidate(dentry, flags);
1622                         if (unlikely(error <= 0)) {
1623                                 if (!error) {
1624                                         d_invalidate(dentry);
1625                                         dput(dentry);
1626                                         goto again;
1627                                 }
1628                                 dput(dentry);
1629                                 dentry = ERR_PTR(error);
1630                         }
1631                 }
1632         } else {
1633                 old = inode->i_op->lookup(inode, dentry, flags);
1634                 d_lookup_done(dentry);
1635                 if (unlikely(old)) {
1636                         dput(dentry);
1637                         dentry = old;
1638                 }
1639         }
1640 out:
1641         inode_unlock_shared(inode);
1642         return dentry;
1643 }
1644
1645 static inline int may_lookup(struct nameidata *nd)
1646 {
1647         if (nd->flags & LOOKUP_RCU) {
1648                 int err = inode_permission(nd->inode, MAY_EXEC|MAY_NOT_BLOCK);
1649                 if (err != -ECHILD)
1650                         return err;
1651                 if (unlazy_walk(nd, NULL, 0))
1652                         return -ECHILD;
1653         }
1654         return inode_permission(nd->inode, MAY_EXEC);
1655 }
1656
1657 static inline int handle_dots(struct nameidata *nd, int type)
1658 {
1659         if (type == LAST_DOTDOT) {
1660                 if (!nd->root.mnt)
1661                         set_root(nd);
1662                 if (nd->flags & LOOKUP_RCU) {
1663                         return follow_dotdot_rcu(nd);
1664                 } else
1665                         return follow_dotdot(nd);
1666         }
1667         return 0;
1668 }
1669
1670 static int pick_link(struct nameidata *nd, struct path *link,
1671                      struct inode *inode, unsigned seq)
1672 {
1673         int error;
1674         struct saved *last;
1675         if (unlikely(nd->total_link_count++ >= MAXSYMLINKS)) {
1676                 path_to_nameidata(link, nd);
1677                 return -ELOOP;
1678         }
1679         if (!(nd->flags & LOOKUP_RCU)) {
1680                 if (link->mnt == nd->path.mnt)
1681                         mntget(link->mnt);
1682         }
1683         error = nd_alloc_stack(nd);
1684         if (unlikely(error)) {
1685                 if (error == -ECHILD) {
1686                         if (unlikely(unlazy_link(nd, link, seq)))
1687                                 return -ECHILD;
1688                         error = nd_alloc_stack(nd);
1689                 }
1690                 if (error) {
1691                         path_put(link);
1692                         return error;
1693                 }
1694         }
1695
1696         last = nd->stack + nd->depth++;
1697         last->link = *link;
1698         clear_delayed_call(&last->done);
1699         nd->link_inode = inode;
1700         last->seq = seq;
1701         return 1;
1702 }
1703
1704 /*
1705  * Do we need to follow links? We _really_ want to be able
1706  * to do this check without having to look at inode->i_op,
1707  * so we keep a cache of "no, this doesn't need follow_link"
1708  * for the common case.
1709  */
1710 static inline int should_follow_link(struct nameidata *nd, struct path *link,
1711                                      int follow,
1712                                      struct inode *inode, unsigned seq)
1713 {
1714         if (likely(!d_is_symlink(link->dentry)))
1715                 return 0;
1716         if (!follow)
1717                 return 0;
1718         /* make sure that d_is_symlink above matches inode */
1719         if (nd->flags & LOOKUP_RCU) {
1720                 if (read_seqcount_retry(&link->dentry->d_seq, seq))
1721                         return -ECHILD;
1722         }
1723         return pick_link(nd, link, inode, seq);
1724 }
1725
1726 enum {WALK_GET = 1, WALK_PUT = 2};
1727
1728 static int walk_component(struct nameidata *nd, int flags)
1729 {
1730         struct path path;
1731         struct inode *inode;
1732         unsigned seq;
1733         int err;
1734         /*
1735          * "." and ".." are special - ".." especially so because it has
1736          * to be able to know about the current root directory and
1737          * parent relationships.
1738          */
1739         if (unlikely(nd->last_type != LAST_NORM)) {
1740                 err = handle_dots(nd, nd->last_type);
1741                 if (flags & WALK_PUT)
1742                         put_link(nd);
1743                 return err;
1744         }
1745         err = lookup_fast(nd, &path, &inode, &seq);
1746         if (unlikely(err <= 0)) {
1747                 if (err < 0)
1748                         return err;
1749                 path.dentry = lookup_slow(&nd->last, nd->path.dentry,
1750                                           nd->flags);
1751                 if (IS_ERR(path.dentry))
1752                         return PTR_ERR(path.dentry);
1753
1754                 path.mnt = nd->path.mnt;
1755                 err = follow_managed(&path, nd);
1756                 if (unlikely(err < 0))
1757                         return err;
1758
1759                 if (unlikely(d_is_negative(path.dentry))) {
1760                         path_to_nameidata(&path, nd);
1761                         return -ENOENT;
1762                 }
1763
1764                 seq = 0;        /* we are already out of RCU mode */
1765                 inode = d_backing_inode(path.dentry);
1766         }
1767
1768         if (flags & WALK_PUT)
1769                 put_link(nd);
1770         err = should_follow_link(nd, &path, flags & WALK_GET, inode, seq);
1771         if (unlikely(err))
1772                 return err;
1773         path_to_nameidata(&path, nd);
1774         nd->inode = inode;
1775         nd->seq = seq;
1776         return 0;
1777 }
1778
1779 /*
1780  * We can do the critical dentry name comparison and hashing
1781  * operations one word at a time, but we are limited to:
1782  *
1783  * - Architectures with fast unaligned word accesses. We could
1784  *   do a "get_unaligned()" if this helps and is sufficiently
1785  *   fast.
1786  *
1787  * - non-CONFIG_DEBUG_PAGEALLOC configurations (so that we
1788  *   do not trap on the (extremely unlikely) case of a page
1789  *   crossing operation.
1790  *
1791  * - Furthermore, we need an efficient 64-bit compile for the
1792  *   64-bit case in order to generate the "number of bytes in
1793  *   the final mask". Again, that could be replaced with a
1794  *   efficient population count instruction or similar.
1795  */
1796 #ifdef CONFIG_DCACHE_WORD_ACCESS
1797
1798 #include <asm/word-at-a-time.h>
1799
1800 #ifdef CONFIG_64BIT
1801
1802 static inline unsigned int fold_hash(unsigned long hash)
1803 {
1804         return hash_64(hash, 32);
1805 }
1806
1807 /*
1808  * This is George Marsaglia's XORSHIFT generator.
1809  * It implements a maximum-period LFSR in only a few
1810  * instructions.  It also has the property (required
1811  * by hash_name()) that mix_hash(0) = 0.
1812  */
1813 static inline unsigned long mix_hash(unsigned long hash)
1814 {
1815         hash ^= hash << 13;
1816         hash ^= hash >> 7;
1817         hash ^= hash << 17;
1818         return hash;
1819 }
1820
1821 #else   /* 32-bit case */
1822
1823 #define fold_hash(x) (x)
1824
1825 static inline unsigned long mix_hash(unsigned long hash)
1826 {
1827         hash ^= hash << 13;
1828         hash ^= hash >> 17;
1829         hash ^= hash << 5;
1830         return hash;
1831 }
1832
1833 #endif
1834
1835 unsigned int full_name_hash(const unsigned char *name, unsigned int len)
1836 {
1837         unsigned long a, hash = 0;
1838
1839         for (;;) {
1840                 a = load_unaligned_zeropad(name);
1841                 if (len < sizeof(unsigned long))
1842                         break;
1843                 hash = mix_hash(hash + a);
1844                 name += sizeof(unsigned long);
1845                 len -= sizeof(unsigned long);
1846                 if (!len)
1847                         goto done;
1848         }
1849         hash += a & bytemask_from_count(len);
1850 done:
1851         return fold_hash(hash);
1852 }
1853 EXPORT_SYMBOL(full_name_hash);
1854
1855 /*
1856  * Calculate the length and hash of the path component, and
1857  * return the "hash_len" as the result.
1858  */
1859 static inline u64 hash_name(const char *name)
1860 {
1861         unsigned long a, b, adata, bdata, mask, hash, len;
1862         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1863
1864         hash = a = 0;
1865         len = -sizeof(unsigned long);
1866         do {
1867                 hash = mix_hash(hash + a);
1868                 len += sizeof(unsigned long);
1869                 a = load_unaligned_zeropad(name+len);
1870                 b = a ^ REPEAT_BYTE('/');
1871         } while (!(has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)));
1872
1873         adata = prep_zero_mask(a, adata, &constants);
1874         bdata = prep_zero_mask(b, bdata, &constants);
1875
1876         mask = create_zero_mask(adata | bdata);
1877
1878         hash += a & zero_bytemask(mask);
1879         len += find_zero(mask);
1880         return hashlen_create(fold_hash(hash), len);
1881 }
1882
1883 #else
1884
1885 unsigned int full_name_hash(const unsigned char *name, unsigned int len)
1886 {
1887         unsigned long hash = init_name_hash();
1888         while (len--)
1889                 hash = partial_name_hash(*name++, hash);
1890         return end_name_hash(hash);
1891 }
1892 EXPORT_SYMBOL(full_name_hash);
1893
1894 /*
1895  * We know there's a real path component here of at least
1896  * one character.
1897  */
1898 static inline u64 hash_name(const char *name)
1899 {
1900         unsigned long hash = init_name_hash();
1901         unsigned long len = 0, c;
1902
1903         c = (unsigned char)*name;
1904         do {
1905                 len++;
1906                 hash = partial_name_hash(c, hash);
1907                 c = (unsigned char)name[len];
1908         } while (c && c != '/');
1909         return hashlen_create(end_name_hash(hash), len);
1910 }
1911
1912 #endif
1913
1914 /*
1915  * Name resolution.
1916  * This is the basic name resolution function, turning a pathname into
1917  * the final dentry. We expect 'base' to be positive and a directory.
1918  *
1919  * Returns 0 and nd will have valid dentry and mnt on success.
1920  * Returns error and drops reference to input namei data on failure.
1921  */
1922 static int link_path_walk(const char *name, struct nameidata *nd)
1923 {
1924         int err;
1925
1926         while (*name=='/')
1927                 name++;
1928         if (!*name)
1929                 return 0;
1930
1931         /* At this point we know we have a real path component. */
1932         for(;;) {
1933                 u64 hash_len;
1934                 int type;
1935
1936                 err = may_lookup(nd);
1937                 if (err)
1938                         return err;
1939
1940                 hash_len = hash_name(name);
1941
1942                 type = LAST_NORM;
1943                 if (name[0] == '.') switch (hashlen_len(hash_len)) {
1944                         case 2:
1945                                 if (name[1] == '.') {
1946                                         type = LAST_DOTDOT;
1947                                         nd->flags |= LOOKUP_JUMPED;
1948                                 }
1949                                 break;
1950                         case 1:
1951                                 type = LAST_DOT;
1952                 }
1953                 if (likely(type == LAST_NORM)) {
1954                         struct dentry *parent = nd->path.dentry;
1955                         nd->flags &= ~LOOKUP_JUMPED;
1956                         if (unlikely(parent->d_flags & DCACHE_OP_HASH)) {
1957                                 struct qstr this = { { .hash_len = hash_len }, .name = name };
1958                                 err = parent->d_op->d_hash(parent, &this);
1959                                 if (err < 0)
1960                                         return err;
1961                                 hash_len = this.hash_len;
1962                                 name = this.name;
1963                         }
1964                 }
1965
1966                 nd->last.hash_len = hash_len;
1967                 nd->last.name = name;
1968                 nd->last_type = type;
1969
1970                 name += hashlen_len(hash_len);
1971                 if (!*name)
1972                         goto OK;
1973                 /*
1974                  * If it wasn't NUL, we know it was '/'. Skip that
1975                  * slash, and continue until no more slashes.
1976                  */
1977                 do {
1978                         name++;
1979                 } while (unlikely(*name == '/'));
1980                 if (unlikely(!*name)) {
1981 OK:
1982                         /* pathname body, done */
1983                         if (!nd->depth)
1984                                 return 0;
1985                         name = nd->stack[nd->depth - 1].name;
1986                         /* trailing symlink, done */
1987                         if (!name)
1988                                 return 0;
1989                         /* last component of nested symlink */
1990                         err = walk_component(nd, WALK_GET | WALK_PUT);
1991                 } else {
1992                         err = walk_component(nd, WALK_GET);
1993                 }
1994                 if (err < 0)
1995                         return err;
1996
1997                 if (err) {
1998                         const char *s = get_link(nd);
1999
2000                         if (IS_ERR(s))
2001                                 return PTR_ERR(s);
2002                         err = 0;
2003                         if (unlikely(!s)) {
2004                                 /* jumped */
2005                                 put_link(nd);
2006                         } else {
2007                                 nd->stack[nd->depth - 1].name = name;
2008                                 name = s;
2009                                 continue;
2010                         }
2011                 }
2012                 if (unlikely(!d_can_lookup(nd->path.dentry))) {
2013                         if (nd->flags & LOOKUP_RCU) {
2014                                 if (unlazy_walk(nd, NULL, 0))
2015                                         return -ECHILD;
2016                         }
2017                         return -ENOTDIR;
2018                 }
2019         }
2020 }
2021
2022 static const char *path_init(struct nameidata *nd, unsigned flags)
2023 {
2024         int retval = 0;
2025         const char *s = nd->name->name;
2026
2027         nd->last_type = LAST_ROOT; /* if there are only slashes... */
2028         nd->flags = flags | LOOKUP_JUMPED | LOOKUP_PARENT;
2029         nd->depth = 0;
2030         if (flags & LOOKUP_ROOT) {
2031                 struct dentry *root = nd->root.dentry;
2032                 struct inode *inode = root->d_inode;
2033                 if (*s) {
2034                         if (!d_can_lookup(root))
2035                                 return ERR_PTR(-ENOTDIR);
2036                         retval = inode_permission(inode, MAY_EXEC);
2037                         if (retval)
2038                                 return ERR_PTR(retval);
2039                 }
2040                 nd->path = nd->root;
2041                 nd->inode = inode;
2042                 if (flags & LOOKUP_RCU) {
2043                         rcu_read_lock();
2044                         nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2045                         nd->root_seq = nd->seq;
2046                         nd->m_seq = read_seqbegin(&mount_lock);
2047                 } else {
2048                         path_get(&nd->path);
2049                 }
2050                 return s;
2051         }
2052
2053         nd->root.mnt = NULL;
2054         nd->path.mnt = NULL;
2055         nd->path.dentry = NULL;
2056
2057         nd->m_seq = read_seqbegin(&mount_lock);
2058         if (*s == '/') {
2059                 if (flags & LOOKUP_RCU)
2060                         rcu_read_lock();
2061                 set_root(nd);
2062                 if (likely(!nd_jump_root(nd)))
2063                         return s;
2064                 nd->root.mnt = NULL;
2065                 rcu_read_unlock();
2066                 return ERR_PTR(-ECHILD);
2067         } else if (nd->dfd == AT_FDCWD) {
2068                 if (flags & LOOKUP_RCU) {
2069                         struct fs_struct *fs = current->fs;
2070                         unsigned seq;
2071
2072                         rcu_read_lock();
2073
2074                         do {
2075                                 seq = read_seqcount_begin(&fs->seq);
2076                                 nd->path = fs->pwd;
2077                                 nd->inode = nd->path.dentry->d_inode;
2078                                 nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2079                         } while (read_seqcount_retry(&fs->seq, seq));
2080                 } else {
2081                         get_fs_pwd(current->fs, &nd->path);
2082                         nd->inode = nd->path.dentry->d_inode;
2083                 }
2084                 return s;
2085         } else {
2086                 /* Caller must check execute permissions on the starting path component */
2087                 struct fd f = fdget_raw(nd->dfd);
2088                 struct dentry *dentry;
2089
2090                 if (!f.file)
2091                         return ERR_PTR(-EBADF);
2092
2093                 dentry = f.file->f_path.dentry;
2094
2095                 if (*s) {
2096                         if (!d_can_lookup(dentry)) {
2097                                 fdput(f);
2098                                 return ERR_PTR(-ENOTDIR);
2099                         }
2100                 }
2101
2102                 nd->path = f.file->f_path;
2103                 if (flags & LOOKUP_RCU) {
2104                         rcu_read_lock();
2105                         nd->inode = nd->path.dentry->d_inode;
2106                         nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
2107                 } else {
2108                         path_get(&nd->path);
2109                         nd->inode = nd->path.dentry->d_inode;
2110                 }
2111                 fdput(f);
2112                 return s;
2113         }
2114 }
2115
2116 static const char *trailing_symlink(struct nameidata *nd)
2117 {
2118         const char *s;
2119         int error = may_follow_link(nd);
2120         if (unlikely(error))
2121                 return ERR_PTR(error);
2122         nd->flags |= LOOKUP_PARENT;
2123         nd->stack[0].name = NULL;
2124         s = get_link(nd);
2125         return s ? s : "";
2126 }
2127
2128 static inline int lookup_last(struct nameidata *nd)
2129 {
2130         if (nd->last_type == LAST_NORM && nd->last.name[nd->last.len])
2131                 nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
2132
2133         nd->flags &= ~LOOKUP_PARENT;
2134         return walk_component(nd,
2135                         nd->flags & LOOKUP_FOLLOW
2136                                 ? nd->depth
2137                                         ? WALK_PUT | WALK_GET
2138                                         : WALK_GET
2139                                 : 0);
2140 }
2141
2142 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2143 static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
2144 {
2145         const char *s = path_init(nd, flags);
2146         int err;
2147
2148         if (IS_ERR(s))
2149                 return PTR_ERR(s);
2150         while (!(err = link_path_walk(s, nd))
2151                 && ((err = lookup_last(nd)) > 0)) {
2152                 s = trailing_symlink(nd);
2153                 if (IS_ERR(s)) {
2154                         err = PTR_ERR(s);
2155                         break;
2156                 }
2157         }
2158         if (!err)
2159                 err = complete_walk(nd);
2160
2161         if (!err && nd->flags & LOOKUP_DIRECTORY)
2162                 if (!d_can_lookup(nd->path.dentry))
2163                         err = -ENOTDIR;
2164         if (!err) {
2165                 *path = nd->path;
2166                 nd->path.mnt = NULL;
2167                 nd->path.dentry = NULL;
2168         }
2169         terminate_walk(nd);
2170         return err;
2171 }
2172
2173 static int filename_lookup(int dfd, struct filename *name, unsigned flags,
2174                            struct path *path, struct path *root)
2175 {
2176         int retval;
2177         struct nameidata nd;
2178         if (IS_ERR(name))
2179                 return PTR_ERR(name);
2180         if (unlikely(root)) {
2181                 nd.root = *root;
2182                 flags |= LOOKUP_ROOT;
2183         }
2184         set_nameidata(&nd, dfd, name);
2185         retval = path_lookupat(&nd, flags | LOOKUP_RCU, path);
2186         if (unlikely(retval == -ECHILD))
2187                 retval = path_lookupat(&nd, flags, path);
2188         if (unlikely(retval == -ESTALE))
2189                 retval = path_lookupat(&nd, flags | LOOKUP_REVAL, path);
2190
2191         if (likely(!retval))
2192                 audit_inode(name, path->dentry, flags & LOOKUP_PARENT);
2193         restore_nameidata();
2194         putname(name);
2195         return retval;
2196 }
2197
2198 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2199 static int path_parentat(struct nameidata *nd, unsigned flags,
2200                                 struct path *parent)
2201 {
2202         const char *s = path_init(nd, flags);
2203         int err;
2204         if (IS_ERR(s))
2205                 return PTR_ERR(s);
2206         err = link_path_walk(s, nd);
2207         if (!err)
2208                 err = complete_walk(nd);
2209         if (!err) {
2210                 *parent = nd->path;
2211                 nd->path.mnt = NULL;
2212                 nd->path.dentry = NULL;
2213         }
2214         terminate_walk(nd);
2215         return err;
2216 }
2217
2218 static struct filename *filename_parentat(int dfd, struct filename *name,
2219                                 unsigned int flags, struct path *parent,
2220                                 struct qstr *last, int *type)
2221 {
2222         int retval;
2223         struct nameidata nd;
2224
2225         if (IS_ERR(name))
2226                 return name;
2227         set_nameidata(&nd, dfd, name);
2228         retval = path_parentat(&nd, flags | LOOKUP_RCU, parent);
2229         if (unlikely(retval == -ECHILD))
2230                 retval = path_parentat(&nd, flags, parent);
2231         if (unlikely(retval == -ESTALE))
2232                 retval = path_parentat(&nd, flags | LOOKUP_REVAL, parent);
2233         if (likely(!retval)) {
2234                 *last = nd.last;
2235                 *type = nd.last_type;
2236                 audit_inode(name, parent->dentry, LOOKUP_PARENT);
2237         } else {
2238                 putname(name);
2239                 name = ERR_PTR(retval);
2240         }
2241         restore_nameidata();
2242         return name;
2243 }
2244
2245 /* does lookup, returns the object with parent locked */
2246 struct dentry *kern_path_locked(const char *name, struct path *path)
2247 {
2248         struct filename *filename;
2249         struct dentry *d;
2250         struct qstr last;
2251         int type;
2252
2253         filename = filename_parentat(AT_FDCWD, getname_kernel(name), 0, path,
2254                                     &last, &type);
2255         if (IS_ERR(filename))
2256                 return ERR_CAST(filename);
2257         if (unlikely(type != LAST_NORM)) {
2258                 path_put(path);
2259                 putname(filename);
2260                 return ERR_PTR(-EINVAL);
2261         }
2262         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
2263         d = __lookup_hash(&last, path->dentry, 0);
2264         if (IS_ERR(d)) {
2265                 inode_unlock(path->dentry->d_inode);
2266                 path_put(path);
2267         }
2268         putname(filename);
2269         return d;
2270 }
2271
2272 int kern_path(const char *name, unsigned int flags, struct path *path)
2273 {
2274         return filename_lookup(AT_FDCWD, getname_kernel(name),
2275                                flags, path, NULL);
2276 }
2277 EXPORT_SYMBOL(kern_path);
2278
2279 /**
2280  * vfs_path_lookup - lookup a file path relative to a dentry-vfsmount pair
2281  * @dentry:  pointer to dentry of the base directory
2282  * @mnt: pointer to vfs mount of the base directory
2283  * @name: pointer to file name
2284  * @flags: lookup flags
2285  * @path: pointer to struct path to fill
2286  */
2287 int vfs_path_lookup(struct dentry *dentry, struct vfsmount *mnt,
2288                     const char *name, unsigned int flags,
2289                     struct path *path)
2290 {
2291         struct path root = {.mnt = mnt, .dentry = dentry};
2292         /* the first argument of filename_lookup() is ignored with root */
2293         return filename_lookup(AT_FDCWD, getname_kernel(name),
2294                                flags , path, &root);
2295 }
2296 EXPORT_SYMBOL(vfs_path_lookup);
2297
2298 /**
2299  * lookup_hash - lookup single pathname component on already hashed name
2300  * @name:       name and hash to lookup
2301  * @base:       base directory to lookup from
2302  *
2303  * The name must have been verified and hashed (see lookup_one_len()).  Using
2304  * this after just full_name_hash() is unsafe.
2305  *
2306  * This function also doesn't check for search permission on base directory.
2307  *
2308  * Use lookup_one_len_unlocked() instead, unless you really know what you are
2309  * doing.
2310  *
2311  * Do not hold i_mutex; this helper takes i_mutex if necessary.
2312  */
2313 struct dentry *lookup_hash(const struct qstr *name, struct dentry *base)
2314 {
2315         struct dentry *ret;
2316
2317         ret = lookup_dcache(name, base, 0);
2318         if (!ret)
2319                 ret = lookup_slow(name, base, 0);
2320
2321         return ret;
2322 }
2323 EXPORT_SYMBOL(lookup_hash);
2324
2325 /**
2326  * lookup_one_len - filesystem helper to lookup single pathname component
2327  * @name:       pathname component to lookup
2328  * @base:       base directory to lookup from
2329  * @len:        maximum length @len should be interpreted to
2330  *
2331  * Note that this routine is purely a helper for filesystem usage and should
2332  * not be called by generic code.
2333  *
2334  * The caller must hold base->i_mutex.
2335  */
2336 struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
2337 {
2338         struct qstr this;
2339         unsigned int c;
2340         int err;
2341
2342         WARN_ON_ONCE(!inode_is_locked(base->d_inode));
2343
2344         this.name = name;
2345         this.len = len;
2346         this.hash = full_name_hash(name, len);
2347         if (!len)
2348                 return ERR_PTR(-EACCES);
2349
2350         if (unlikely(name[0] == '.')) {
2351                 if (len < 2 || (len == 2 && name[1] == '.'))
2352                         return ERR_PTR(-EACCES);
2353         }
2354
2355         while (len--) {
2356                 c = *(const unsigned char *)name++;
2357                 if (c == '/' || c == '\0')
2358                         return ERR_PTR(-EACCES);
2359         }
2360         /*
2361          * See if the low-level filesystem might want
2362          * to use its own hash..
2363          */
2364         if (base->d_flags & DCACHE_OP_HASH) {
2365                 int err = base->d_op->d_hash(base, &this);
2366                 if (err < 0)
2367                         return ERR_PTR(err);
2368         }
2369
2370         err = inode_permission(base->d_inode, MAY_EXEC);
2371         if (err)
2372                 return ERR_PTR(err);
2373
2374         return __lookup_hash(&this, base, 0);
2375 }
2376 EXPORT_SYMBOL(lookup_one_len);
2377
2378 /**
2379  * lookup_one_len_unlocked - filesystem helper to lookup single pathname component
2380  * @name:       pathname component to lookup
2381  * @base:       base directory to lookup from
2382  * @len:        maximum length @len should be interpreted to
2383  *
2384  * Note that this routine is purely a helper for filesystem usage and should
2385  * not be called by generic code.
2386  *
2387  * Unlike lookup_one_len, it should be called without the parent
2388  * i_mutex held, and will take the i_mutex itself if necessary.
2389  */
2390 struct dentry *lookup_one_len_unlocked(const char *name,
2391                                        struct dentry *base, int len)
2392 {
2393         struct qstr this;
2394         unsigned int c;
2395         int err;
2396
2397         this.name = name;
2398         this.len = len;
2399         this.hash = full_name_hash(name, len);
2400         if (!len)
2401                 return ERR_PTR(-EACCES);
2402
2403         if (unlikely(name[0] == '.')) {
2404                 if (len < 2 || (len == 2 && name[1] == '.'))
2405                         return ERR_PTR(-EACCES);
2406         }
2407
2408         while (len--) {
2409                 c = *(const unsigned char *)name++;
2410                 if (c == '/' || c == '\0')
2411                         return ERR_PTR(-EACCES);
2412         }
2413         /*
2414          * See if the low-level filesystem might want
2415          * to use its own hash..
2416          */
2417         if (base->d_flags & DCACHE_OP_HASH) {
2418                 int err = base->d_op->d_hash(base, &this);
2419                 if (err < 0)
2420                         return ERR_PTR(err);
2421         }
2422
2423         err = inode_permission(base->d_inode, MAY_EXEC);
2424         if (err)
2425                 return ERR_PTR(err);
2426
2427         return lookup_hash(&this, base);
2428 }
2429 EXPORT_SYMBOL(lookup_one_len_unlocked);
2430
2431 int user_path_at_empty(int dfd, const char __user *name, unsigned flags,
2432                  struct path *path, int *empty)
2433 {
2434         return filename_lookup(dfd, getname_flags(name, flags, empty),
2435                                flags, path, NULL);
2436 }
2437 EXPORT_SYMBOL(user_path_at_empty);
2438
2439 /*
2440  * NB: most callers don't do anything directly with the reference to the
2441  *     to struct filename, but the nd->last pointer points into the name string
2442  *     allocated by getname. So we must hold the reference to it until all
2443  *     path-walking is complete.
2444  */
2445 static inline struct filename *
2446 user_path_parent(int dfd, const char __user *path,
2447                  struct path *parent,
2448                  struct qstr *last,
2449                  int *type,
2450                  unsigned int flags)
2451 {
2452         /* only LOOKUP_REVAL is allowed in extra flags */
2453         return filename_parentat(dfd, getname(path), flags & LOOKUP_REVAL,
2454                                  parent, last, type);
2455 }
2456
2457 /**
2458  * mountpoint_last - look up last component for umount
2459  * @nd:   pathwalk nameidata - currently pointing at parent directory of "last"
2460  * @path: pointer to container for result
2461  *
2462  * This is a special lookup_last function just for umount. In this case, we
2463  * need to resolve the path without doing any revalidation.
2464  *
2465  * The nameidata should be the result of doing a LOOKUP_PARENT pathwalk. Since
2466  * mountpoints are always pinned in the dcache, their ancestors are too. Thus,
2467  * in almost all cases, this lookup will be served out of the dcache. The only
2468  * cases where it won't are if nd->last refers to a symlink or the path is
2469  * bogus and it doesn't exist.
2470  *
2471  * Returns:
2472  * -error: if there was an error during lookup. This includes -ENOENT if the
2473  *         lookup found a negative dentry. The nd->path reference will also be
2474  *         put in this case.
2475  *
2476  * 0:      if we successfully resolved nd->path and found it to not to be a
2477  *         symlink that needs to be followed. "path" will also be populated.
2478  *         The nd->path reference will also be put.
2479  *
2480  * 1:      if we successfully resolved nd->last and found it to be a symlink
2481  *         that needs to be followed. "path" will be populated with the path
2482  *         to the link, and nd->path will *not* be put.
2483  */
2484 static int
2485 mountpoint_last(struct nameidata *nd, struct path *path)
2486 {
2487         int error = 0;
2488         struct dentry *dentry;
2489         struct dentry *dir = nd->path.dentry;
2490
2491         /* If we're in rcuwalk, drop out of it to handle last component */
2492         if (nd->flags & LOOKUP_RCU) {
2493                 if (unlazy_walk(nd, NULL, 0))
2494                         return -ECHILD;
2495         }
2496
2497         nd->flags &= ~LOOKUP_PARENT;
2498
2499         if (unlikely(nd->last_type != LAST_NORM)) {
2500                 error = handle_dots(nd, nd->last_type);
2501                 if (error)
2502                         return error;
2503                 dentry = dget(nd->path.dentry);
2504         } else {
2505                 dentry = d_lookup(dir, &nd->last);
2506                 if (!dentry) {
2507                         /*
2508                          * No cached dentry. Mounted dentries are pinned in the
2509                          * cache, so that means that this dentry is probably
2510                          * a symlink or the path doesn't actually point
2511                          * to a mounted dentry.
2512                          */
2513                         dentry = lookup_slow(&nd->last, dir,
2514                                              nd->flags | LOOKUP_NO_REVAL);
2515                         if (IS_ERR(dentry))
2516                                 return PTR_ERR(dentry);
2517                 }
2518         }
2519         if (d_is_negative(dentry)) {
2520                 dput(dentry);
2521                 return -ENOENT;
2522         }
2523         if (nd->depth)
2524                 put_link(nd);
2525         path->dentry = dentry;
2526         path->mnt = nd->path.mnt;
2527         error = should_follow_link(nd, path, nd->flags & LOOKUP_FOLLOW,
2528                                    d_backing_inode(dentry), 0);
2529         if (unlikely(error))
2530                 return error;
2531         mntget(path->mnt);
2532         follow_mount(path);
2533         return 0;
2534 }
2535
2536 /**
2537  * path_mountpoint - look up a path to be umounted
2538  * @nd:         lookup context
2539  * @flags:      lookup flags
2540  * @path:       pointer to container for result
2541  *
2542  * Look up the given name, but don't attempt to revalidate the last component.
2543  * Returns 0 and "path" will be valid on success; Returns error otherwise.
2544  */
2545 static int
2546 path_mountpoint(struct nameidata *nd, unsigned flags, struct path *path)
2547 {
2548         const char *s = path_init(nd, flags);
2549         int err;
2550         if (IS_ERR(s))
2551                 return PTR_ERR(s);
2552         while (!(err = link_path_walk(s, nd)) &&
2553                 (err = mountpoint_last(nd, path)) > 0) {
2554                 s = trailing_symlink(nd);
2555                 if (IS_ERR(s)) {
2556                         err = PTR_ERR(s);
2557                         break;
2558                 }
2559         }
2560         terminate_walk(nd);
2561         return err;
2562 }
2563
2564 static int
2565 filename_mountpoint(int dfd, struct filename *name, struct path *path,
2566                         unsigned int flags)
2567 {
2568         struct nameidata nd;
2569         int error;
2570         if (IS_ERR(name))
2571                 return PTR_ERR(name);
2572         set_nameidata(&nd, dfd, name);
2573         error = path_mountpoint(&nd, flags | LOOKUP_RCU, path);
2574         if (unlikely(error == -ECHILD))
2575                 error = path_mountpoint(&nd, flags, path);
2576         if (unlikely(error == -ESTALE))
2577                 error = path_mountpoint(&nd, flags | LOOKUP_REVAL, path);
2578         if (likely(!error))
2579                 audit_inode(name, path->dentry, 0);
2580         restore_nameidata();
2581         putname(name);
2582         return error;
2583 }
2584
2585 /**
2586  * user_path_mountpoint_at - lookup a path from userland in order to umount it
2587  * @dfd:        directory file descriptor
2588  * @name:       pathname from userland
2589  * @flags:      lookup flags
2590  * @path:       pointer to container to hold result
2591  *
2592  * A umount is a special case for path walking. We're not actually interested
2593  * in the inode in this situation, and ESTALE errors can be a problem. We
2594  * simply want track down the dentry and vfsmount attached at the mountpoint
2595  * and avoid revalidating the last component.
2596  *
2597  * Returns 0 and populates "path" on success.
2598  */
2599 int
2600 user_path_mountpoint_at(int dfd, const char __user *name, unsigned int flags,
2601                         struct path *path)
2602 {
2603         return filename_mountpoint(dfd, getname(name), path, flags);
2604 }
2605
2606 int
2607 kern_path_mountpoint(int dfd, const char *name, struct path *path,
2608                         unsigned int flags)
2609 {
2610         return filename_mountpoint(dfd, getname_kernel(name), path, flags);
2611 }
2612 EXPORT_SYMBOL(kern_path_mountpoint);
2613
2614 int __check_sticky(struct inode *dir, struct inode *inode)
2615 {
2616         kuid_t fsuid = current_fsuid();
2617
2618         if (uid_eq(inode->i_uid, fsuid))
2619                 return 0;
2620         if (uid_eq(dir->i_uid, fsuid))
2621                 return 0;
2622         return !capable_wrt_inode_uidgid(inode, CAP_FOWNER);
2623 }
2624 EXPORT_SYMBOL(__check_sticky);
2625
2626 /*
2627  *      Check whether we can remove a link victim from directory dir, check
2628  *  whether the type of victim is right.
2629  *  1. We can't do it if dir is read-only (done in permission())
2630  *  2. We should have write and exec permissions on dir
2631  *  3. We can't remove anything from append-only dir
2632  *  4. We can't do anything with immutable dir (done in permission())
2633  *  5. If the sticky bit on dir is set we should either
2634  *      a. be owner of dir, or
2635  *      b. be owner of victim, or
2636  *      c. have CAP_FOWNER capability
2637  *  6. If the victim is append-only or immutable we can't do antyhing with
2638  *     links pointing to it.
2639  *  7. If we were asked to remove a directory and victim isn't one - ENOTDIR.
2640  *  8. If we were asked to remove a non-directory and victim isn't one - EISDIR.
2641  *  9. We can't remove a root or mountpoint.
2642  * 10. We don't allow removal of NFS sillyrenamed files; it's handled by
2643  *     nfs_async_unlink().
2644  */
2645 static int may_delete(struct inode *dir, struct dentry *victim, bool isdir)
2646 {
2647         struct inode *inode = d_backing_inode(victim);
2648         int error;
2649
2650         if (d_is_negative(victim))
2651                 return -ENOENT;
2652         BUG_ON(!inode);
2653
2654         BUG_ON(victim->d_parent->d_inode != dir);
2655         audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);
2656
2657         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
2658         if (error)
2659                 return error;
2660         if (IS_APPEND(dir))
2661                 return -EPERM;
2662
2663         if (check_sticky(dir, inode) || IS_APPEND(inode) ||
2664             IS_IMMUTABLE(inode) || IS_SWAPFILE(inode))
2665                 return -EPERM;
2666         if (isdir) {
2667                 if (!d_is_dir(victim))
2668                         return -ENOTDIR;
2669                 if (IS_ROOT(victim))
2670                         return -EBUSY;
2671         } else if (d_is_dir(victim))
2672                 return -EISDIR;
2673         if (IS_DEADDIR(dir))
2674                 return -ENOENT;
2675         if (victim->d_flags & DCACHE_NFSFS_RENAMED)
2676                 return -EBUSY;
2677         return 0;
2678 }
2679
2680 /*      Check whether we can create an object with dentry child in directory
2681  *  dir.
2682  *  1. We can't do it if child already exists (open has special treatment for
2683  *     this case, but since we are inlined it's OK)
2684  *  2. We can't do it if dir is read-only (done in permission())
2685  *  3. We should have write and exec permissions on dir
2686  *  4. We can't do it if dir is immutable (done in permission())
2687  */
2688 static inline int may_create(struct inode *dir, struct dentry *child)
2689 {
2690         audit_inode_child(dir, child, AUDIT_TYPE_CHILD_CREATE);
2691         if (child->d_inode)
2692                 return -EEXIST;
2693         if (IS_DEADDIR(dir))
2694                 return -ENOENT;
2695         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
2696 }
2697
2698 /*
2699  * p1 and p2 should be directories on the same fs.
2700  */
2701 struct dentry *lock_rename(struct dentry *p1, struct dentry *p2)
2702 {
2703         struct dentry *p;
2704
2705         if (p1 == p2) {
2706                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2707                 return NULL;
2708         }
2709
2710         mutex_lock(&p1->d_sb->s_vfs_rename_mutex);
2711
2712         p = d_ancestor(p2, p1);
2713         if (p) {
2714                 inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
2715                 inode_lock_nested(p1->d_inode, I_MUTEX_CHILD);
2716                 return p;
2717         }
2718
2719         p = d_ancestor(p1, p2);
2720         if (p) {
2721                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2722                 inode_lock_nested(p2->d_inode, I_MUTEX_CHILD);
2723                 return p;
2724         }
2725
2726         inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2727         inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
2728         return NULL;
2729 }
2730 EXPORT_SYMBOL(lock_rename);
2731
2732 void unlock_rename(struct dentry *p1, struct dentry *p2)
2733 {
2734         inode_unlock(p1->d_inode);
2735         if (p1 != p2) {
2736                 inode_unlock(p2->d_inode);
2737                 mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
2738         }
2739 }
2740 EXPORT_SYMBOL(unlock_rename);
2741
2742 int vfs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2743                 bool want_excl)
2744 {
2745         int error = may_create(dir, dentry);
2746         if (error)
2747                 return error;
2748
2749         if (!dir->i_op->create)
2750                 return -EACCES; /* shouldn't it be ENOSYS? */
2751         mode &= S_IALLUGO;
2752         mode |= S_IFREG;
2753         error = security_inode_create(dir, dentry, mode);
2754         if (error)
2755                 return error;
2756         error = dir->i_op->create(dir, dentry, mode, want_excl);
2757         if (!error)
2758                 fsnotify_create(dir, dentry);
2759         return error;
2760 }
2761 EXPORT_SYMBOL(vfs_create);
2762
2763 static int may_open(struct path *path, int acc_mode, int flag)
2764 {
2765         struct dentry *dentry = path->dentry;
2766         struct inode *inode = dentry->d_inode;
2767         int error;
2768
2769         if (!inode)
2770                 return -ENOENT;
2771
2772         switch (inode->i_mode & S_IFMT) {
2773         case S_IFLNK:
2774                 return -ELOOP;
2775         case S_IFDIR:
2776                 if (acc_mode & MAY_WRITE)
2777                         return -EISDIR;
2778                 break;
2779         case S_IFBLK:
2780         case S_IFCHR:
2781                 if (path->mnt->mnt_flags & MNT_NODEV)
2782                         return -EACCES;
2783                 /*FALLTHRU*/
2784         case S_IFIFO:
2785         case S_IFSOCK:
2786                 flag &= ~O_TRUNC;
2787                 break;
2788         }
2789
2790         error = inode_permission(inode, MAY_OPEN | acc_mode);
2791         if (error)
2792                 return error;
2793
2794         /*
2795          * An append-only file must be opened in append mode for writing.
2796          */
2797         if (IS_APPEND(inode)) {
2798                 if  ((flag & O_ACCMODE) != O_RDONLY && !(flag & O_APPEND))
2799                         return -EPERM;
2800                 if (flag & O_TRUNC)
2801                         return -EPERM;
2802         }
2803
2804         /* O_NOATIME can only be set by the owner or superuser */
2805         if (flag & O_NOATIME && !inode_owner_or_capable(inode))
2806                 return -EPERM;
2807
2808         return 0;
2809 }
2810
2811 static int handle_truncate(struct file *filp)
2812 {
2813         struct path *path = &filp->f_path;
2814         struct inode *inode = path->dentry->d_inode;
2815         int error = get_write_access(inode);
2816         if (error)
2817                 return error;
2818         /*
2819          * Refuse to truncate files with mandatory locks held on them.
2820          */
2821         error = locks_verify_locked(filp);
2822         if (!error)
2823                 error = security_path_truncate(path);
2824         if (!error) {
2825                 error = do_truncate(path->dentry, 0,
2826                                     ATTR_MTIME|ATTR_CTIME|ATTR_OPEN,
2827                                     filp);
2828         }
2829         put_write_access(inode);
2830         return error;
2831 }
2832
2833 static inline int open_to_namei_flags(int flag)
2834 {
2835         if ((flag & O_ACCMODE) == 3)
2836                 flag--;
2837         return flag;
2838 }
2839
2840 static int may_o_create(const struct path *dir, struct dentry *dentry, umode_t mode)
2841 {
2842         int error = security_path_mknod(dir, dentry, mode, 0);
2843         if (error)
2844                 return error;
2845
2846         error = inode_permission(dir->dentry->d_inode, MAY_WRITE | MAY_EXEC);
2847         if (error)
2848                 return error;
2849
2850         return security_inode_create(dir->dentry->d_inode, dentry, mode);
2851 }
2852
2853 /*
2854  * Attempt to atomically look up, create and open a file from a negative
2855  * dentry.
2856  *
2857  * Returns 0 if successful.  The file will have been created and attached to
2858  * @file by the filesystem calling finish_open().
2859  *
2860  * Returns 1 if the file was looked up only or didn't need creating.  The
2861  * caller will need to perform the open themselves.  @path will have been
2862  * updated to point to the new dentry.  This may be negative.
2863  *
2864  * Returns an error code otherwise.
2865  */
2866 static int atomic_open(struct nameidata *nd, struct dentry *dentry,
2867                         struct path *path, struct file *file,
2868                         const struct open_flags *op,
2869                         int open_flag, umode_t mode,
2870                         int *opened)
2871 {
2872         struct dentry *const DENTRY_NOT_SET = (void *) -1UL;
2873         struct inode *dir =  nd->path.dentry->d_inode;
2874         int error;
2875
2876         if (!(~open_flag & (O_EXCL | O_CREAT))) /* both O_EXCL and O_CREAT */
2877                 open_flag &= ~O_TRUNC;
2878
2879         if (nd->flags & LOOKUP_DIRECTORY)
2880                 open_flag |= O_DIRECTORY;
2881
2882         file->f_path.dentry = DENTRY_NOT_SET;
2883         file->f_path.mnt = nd->path.mnt;
2884         error = dir->i_op->atomic_open(dir, dentry, file,
2885                                        open_to_namei_flags(open_flag),
2886                                        mode, opened);
2887         d_lookup_done(dentry);
2888         if (!error) {
2889                 /*
2890                  * We didn't have the inode before the open, so check open
2891                  * permission here.
2892                  */
2893                 int acc_mode = op->acc_mode;
2894                 if (*opened & FILE_CREATED) {
2895                         WARN_ON(!(open_flag & O_CREAT));
2896                         fsnotify_create(dir, dentry);
2897                         acc_mode = 0;
2898                 }
2899                 error = may_open(&file->f_path, acc_mode, open_flag);
2900                 if (WARN_ON(error > 0))
2901                         error = -EINVAL;
2902         } else if (error > 0) {
2903                 if (WARN_ON(file->f_path.dentry == DENTRY_NOT_SET)) {
2904                         error = -EIO;
2905                 } else {
2906                         if (file->f_path.dentry) {
2907                                 dput(dentry);
2908                                 dentry = file->f_path.dentry;
2909                         }
2910                         if (*opened & FILE_CREATED)
2911                                 fsnotify_create(dir, dentry);
2912                         path->dentry = dentry;
2913                         path->mnt = nd->path.mnt;
2914                         return 1;
2915                 }
2916         }
2917         dput(dentry);
2918         return error;
2919 }
2920
2921 /*
2922  * Look up and maybe create and open the last component.
2923  *
2924  * Must be called with i_mutex held on parent.
2925  *
2926  * Returns 0 if the file was successfully atomically created (if necessary) and
2927  * opened.  In this case the file will be returned attached to @file.
2928  *
2929  * Returns 1 if the file was not completely opened at this time, though lookups
2930  * and creations will have been performed and the dentry returned in @path will
2931  * be positive upon return if O_CREAT was specified.  If O_CREAT wasn't
2932  * specified then a negative dentry may be returned.
2933  *
2934  * An error code is returned otherwise.
2935  *
2936  * FILE_CREATE will be set in @*opened if the dentry was created and will be
2937  * cleared otherwise prior to returning.
2938  */
2939 static int lookup_open(struct nameidata *nd, struct path *path,
2940                         struct file *file,
2941                         const struct open_flags *op,
2942                         bool got_write, int *opened)
2943 {
2944         struct dentry *dir = nd->path.dentry;
2945         struct inode *dir_inode = dir->d_inode;
2946         int open_flag = op->open_flag;
2947         struct dentry *dentry;
2948         int error, create_error = 0;
2949         umode_t mode = op->mode;
2950         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
2951
2952         if (unlikely(IS_DEADDIR(dir_inode)))
2953                 return -ENOENT;
2954
2955         *opened &= ~FILE_CREATED;
2956         dentry = d_lookup(dir, &nd->last);
2957         for (;;) {
2958                 if (!dentry) {
2959                         dentry = d_alloc_parallel(dir, &nd->last, &wq);
2960                         if (IS_ERR(dentry))
2961                                 return PTR_ERR(dentry);
2962                 }
2963                 if (d_in_lookup(dentry))
2964                         break;
2965
2966                 if (!(dentry->d_flags & DCACHE_OP_REVALIDATE))
2967                         break;
2968
2969                 error = d_revalidate(dentry, nd->flags);
2970                 if (likely(error > 0))
2971                         break;
2972                 if (error)
2973                         goto out_dput;
2974                 d_invalidate(dentry);
2975                 dput(dentry);
2976                 dentry = NULL;
2977         }
2978         if (dentry->d_inode) {
2979                 /* Cached positive dentry: will open in f_op->open */
2980                 goto out_no_open;
2981         }
2982
2983         /*
2984          * Checking write permission is tricky, bacuse we don't know if we are
2985          * going to actually need it: O_CREAT opens should work as long as the
2986          * file exists.  But checking existence breaks atomicity.  The trick is
2987          * to check access and if not granted clear O_CREAT from the flags.
2988          *
2989          * Another problem is returing the "right" error value (e.g. for an
2990          * O_EXCL open we want to return EEXIST not EROFS).
2991          */
2992         if (open_flag & O_CREAT) {
2993                 if (!IS_POSIXACL(dir->d_inode))
2994                         mode &= ~current_umask();
2995                 if (unlikely(!got_write)) {
2996                         create_error = -EROFS;
2997                         open_flag &= ~O_CREAT;
2998                         if (open_flag & (O_EXCL | O_TRUNC))
2999                                 goto no_open;
3000                         /* No side effects, safe to clear O_CREAT */
3001                 } else {
3002                         create_error = may_o_create(&nd->path, dentry, mode);
3003                         if (create_error) {
3004                                 open_flag &= ~O_CREAT;
3005                                 if (open_flag & O_EXCL)
3006                                         goto no_open;
3007                         }
3008                 }
3009         } else if ((open_flag & (O_TRUNC|O_WRONLY|O_RDWR)) &&
3010                    unlikely(!got_write)) {
3011                 /*
3012                  * No O_CREATE -> atomicity not a requirement -> fall
3013                  * back to lookup + open
3014                  */
3015                 goto no_open;
3016         }
3017
3018         if (dir_inode->i_op->atomic_open) {
3019                 error = atomic_open(nd, dentry, path, file, op, open_flag,
3020                                     mode, opened);
3021                 if (unlikely(error == -ENOENT) && create_error)
3022                         error = create_error;
3023                 return error;
3024         }
3025
3026 no_open:
3027         if (d_in_lookup(dentry)) {
3028                 struct dentry *res = dir_inode->i_op->lookup(dir_inode, dentry,
3029                                                              nd->flags);
3030                 d_lookup_done(dentry);
3031                 if (unlikely(res)) {
3032                         if (IS_ERR(res)) {
3033                                 error = PTR_ERR(res);
3034                                 goto out_dput;
3035                         }
3036                         dput(dentry);
3037                         dentry = res;
3038                 }
3039         }
3040
3041         /* Negative dentry, just create the file */
3042         if (!dentry->d_inode && (open_flag & O_CREAT)) {
3043                 *opened |= FILE_CREATED;
3044                 audit_inode_child(dir_inode, dentry, AUDIT_TYPE_CHILD_CREATE);
3045                 if (!dir_inode->i_op->create) {
3046                         error = -EACCES;
3047                         goto out_dput;
3048                 }
3049                 error = dir_inode->i_op->create(dir_inode, dentry, mode,
3050                                                 open_flag & O_EXCL);
3051                 if (error)
3052                         goto out_dput;
3053                 fsnotify_create(dir_inode, dentry);
3054         }
3055         if (unlikely(create_error) && !dentry->d_inode) {
3056                 error = create_error;
3057                 goto out_dput;
3058         }
3059 out_no_open:
3060         path->dentry = dentry;
3061         path->mnt = nd->path.mnt;
3062         return 1;
3063
3064 out_dput:
3065         dput(dentry);
3066         return error;
3067 }
3068
3069 /*
3070  * Handle the last step of open()
3071  */
3072 static int do_last(struct nameidata *nd,
3073                    struct file *file, const struct open_flags *op,
3074                    int *opened)
3075 {
3076         struct dentry *dir = nd->path.dentry;
3077         int open_flag = op->open_flag;
3078         bool will_truncate = (open_flag & O_TRUNC) != 0;
3079         bool got_write = false;
3080         int acc_mode = op->acc_mode;
3081         unsigned seq;
3082         struct inode *inode;
3083         struct path save_parent = { .dentry = NULL, .mnt = NULL };
3084         struct path path;
3085         bool retried = false;
3086         int error;
3087
3088         nd->flags &= ~LOOKUP_PARENT;
3089         nd->flags |= op->intent;
3090
3091         if (nd->last_type != LAST_NORM) {
3092                 error = handle_dots(nd, nd->last_type);
3093                 if (unlikely(error))
3094                         return error;
3095                 goto finish_open;
3096         }
3097
3098         if (!(open_flag & O_CREAT)) {
3099                 if (nd->last.name[nd->last.len])
3100                         nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
3101                 /* we _can_ be in RCU mode here */
3102                 error = lookup_fast(nd, &path, &inode, &seq);
3103                 if (likely(error > 0))
3104                         goto finish_lookup;
3105
3106                 if (error < 0)
3107                         return error;
3108
3109                 BUG_ON(nd->inode != dir->d_inode);
3110                 BUG_ON(nd->flags & LOOKUP_RCU);
3111         } else {
3112                 /* create side of things */
3113                 /*
3114                  * This will *only* deal with leaving RCU mode - LOOKUP_JUMPED
3115                  * has been cleared when we got to the last component we are
3116                  * about to look up
3117                  */
3118                 error = complete_walk(nd);
3119                 if (error)
3120                         return error;
3121
3122                 audit_inode(nd->name, dir, LOOKUP_PARENT);
3123                 /* trailing slashes? */
3124                 if (unlikely(nd->last.name[nd->last.len]))
3125                         return -EISDIR;
3126         }
3127
3128 retry_lookup:
3129         if (open_flag & (O_CREAT | O_TRUNC | O_WRONLY | O_RDWR)) {
3130                 error = mnt_want_write(nd->path.mnt);
3131                 if (!error)
3132                         got_write = true;
3133                 /*
3134                  * do _not_ fail yet - we might not need that or fail with
3135                  * a different error; let lookup_open() decide; we'll be
3136                  * dropping this one anyway.
3137                  */
3138         }
3139         if (open_flag & O_CREAT)
3140                 inode_lock(dir->d_inode);
3141         else
3142                 inode_lock_shared(dir->d_inode);
3143         error = lookup_open(nd, &path, file, op, got_write, opened);
3144         if (open_flag & O_CREAT)
3145                 inode_unlock(dir->d_inode);
3146         else
3147                 inode_unlock_shared(dir->d_inode);
3148
3149         if (error <= 0) {
3150                 if (error)
3151                         goto out;
3152
3153                 if ((*opened & FILE_CREATED) ||
3154                     !S_ISREG(file_inode(file)->i_mode))
3155                         will_truncate = false;
3156
3157                 audit_inode(nd->name, file->f_path.dentry, 0);
3158                 goto opened;
3159         }
3160
3161         if (*opened & FILE_CREATED) {
3162                 /* Don't check for write permission, don't truncate */
3163                 open_flag &= ~O_TRUNC;
3164                 will_truncate = false;
3165                 acc_mode = 0;
3166                 path_to_nameidata(&path, nd);
3167                 goto finish_open_created;
3168         }
3169
3170         /*
3171          * If atomic_open() acquired write access it is dropped now due to
3172          * possible mount and symlink following (this might be optimized away if
3173          * necessary...)
3174          */
3175         if (got_write) {
3176                 mnt_drop_write(nd->path.mnt);
3177                 got_write = false;
3178         }
3179
3180         if (unlikely(d_is_negative(path.dentry))) {
3181                 path_to_nameidata(&path, nd);
3182                 return -ENOENT;
3183         }
3184
3185         /*
3186          * create/update audit record if it already exists.
3187          */
3188         audit_inode(nd->name, path.dentry, 0);
3189
3190         if (unlikely((open_flag & (O_EXCL | O_CREAT)) == (O_EXCL | O_CREAT))) {
3191                 path_to_nameidata(&path, nd);
3192                 return -EEXIST;
3193         }
3194
3195         error = follow_managed(&path, nd);
3196         if (unlikely(error < 0))
3197                 return error;
3198
3199         seq = 0;        /* out of RCU mode, so the value doesn't matter */
3200         inode = d_backing_inode(path.dentry);
3201 finish_lookup:
3202         if (nd->depth)
3203                 put_link(nd);
3204         error = should_follow_link(nd, &path, nd->flags & LOOKUP_FOLLOW,
3205                                    inode, seq);
3206         if (unlikely(error))
3207                 return error;
3208
3209         if ((nd->flags & LOOKUP_RCU) || nd->path.mnt != path.mnt) {
3210                 path_to_nameidata(&path, nd);
3211         } else {
3212                 save_parent.dentry = nd->path.dentry;
3213                 save_parent.mnt = mntget(path.mnt);
3214                 nd->path.dentry = path.dentry;
3215
3216         }
3217         nd->inode = inode;
3218         nd->seq = seq;
3219         /* Why this, you ask?  _Now_ we might have grown LOOKUP_JUMPED... */
3220 finish_open:
3221         error = complete_walk(nd);
3222         if (error) {
3223                 path_put(&save_parent);
3224                 return error;
3225         }
3226         audit_inode(nd->name, nd->path.dentry, 0);
3227         error = -EISDIR;
3228         if ((open_flag & O_CREAT) && d_is_dir(nd->path.dentry))
3229                 goto out;
3230         error = -ENOTDIR;
3231         if ((nd->flags & LOOKUP_DIRECTORY) && !d_can_lookup(nd->path.dentry))
3232                 goto out;
3233         if (!d_is_reg(nd->path.dentry))
3234                 will_truncate = false;
3235
3236         if (will_truncate) {
3237                 error = mnt_want_write(nd->path.mnt);
3238                 if (error)
3239                         goto out;
3240                 got_write = true;
3241         }
3242 finish_open_created:
3243         error = may_open(&nd->path, acc_mode, open_flag);
3244         if (error)
3245                 goto out;
3246         BUG_ON(*opened & FILE_OPENED); /* once it's opened, it's opened */
3247         error = vfs_open(&nd->path, file, current_cred());
3248         if (!error) {
3249                 *opened |= FILE_OPENED;
3250         } else {
3251                 if (error == -EOPENSTALE)
3252                         goto stale_open;
3253                 goto out;
3254         }
3255 opened:
3256         error = open_check_o_direct(file);
3257         if (!error)
3258                 error = ima_file_check(file, op->acc_mode, *opened);
3259         if (!error && will_truncate)
3260                 error = handle_truncate(file);
3261 out:
3262         if (unlikely(error) && (*opened & FILE_OPENED))
3263                 fput(file);
3264         if (unlikely(error > 0)) {
3265                 WARN_ON(1);
3266                 error = -EINVAL;
3267         }
3268         if (got_write)
3269                 mnt_drop_write(nd->path.mnt);
3270         path_put(&save_parent);
3271         return error;
3272
3273 stale_open:
3274         /* If no saved parent or already retried then can't retry */
3275         if (!save_parent.dentry || retried)
3276                 goto out;
3277
3278         BUG_ON(save_parent.dentry != dir);
3279         path_put(&nd->path);
3280         nd->path = save_parent;
3281         nd->inode = dir->d_inode;
3282         save_parent.mnt = NULL;
3283         save_parent.dentry = NULL;
3284         if (got_write) {
3285                 mnt_drop_write(nd->path.mnt);
3286                 got_write = false;
3287         }
3288         retried = true;
3289         goto retry_lookup;
3290 }
3291
3292 static int do_tmpfile(struct nameidata *nd, unsigned flags,
3293                 const struct open_flags *op,
3294                 struct file *file, int *opened)
3295 {
3296         static const struct qstr name = QSTR_INIT("/", 1);
3297         struct dentry *child;
3298         struct inode *dir;
3299         struct path path;
3300         int error = path_lookupat(nd, flags | LOOKUP_DIRECTORY, &path);
3301         if (unlikely(error))
3302                 return error;
3303         error = mnt_want_write(path.mnt);
3304         if (unlikely(error))
3305                 goto out;
3306         dir = path.dentry->d_inode;
3307         /* we want directory to be writable */
3308         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
3309         if (error)
3310                 goto out2;
3311         if (!dir->i_op->tmpfile) {
3312                 error = -EOPNOTSUPP;
3313                 goto out2;
3314         }
3315         child = d_alloc(path.dentry, &name);
3316         if (unlikely(!child)) {
3317                 error = -ENOMEM;
3318                 goto out2;
3319         }
3320         dput(path.dentry);
3321         path.dentry = child;
3322         error = dir->i_op->tmpfile(dir, child, op->mode);
3323         if (error)
3324                 goto out2;
3325         audit_inode(nd->name, child, 0);
3326         /* Don't check for other permissions, the inode was just created */
3327         error = may_open(&path, 0, op->open_flag);
3328         if (error)
3329                 goto out2;
3330         file->f_path.mnt = path.mnt;
3331         error = finish_open(file, child, NULL, opened);
3332         if (error)
3333                 goto out2;
3334         error = open_check_o_direct(file);
3335         if (error) {
3336                 fput(file);
3337         } else if (!(op->open_flag & O_EXCL)) {
3338                 struct inode *inode = file_inode(file);
3339                 spin_lock(&inode->i_lock);
3340                 inode->i_state |= I_LINKABLE;
3341                 spin_unlock(&inode->i_lock);
3342         }
3343 out2:
3344         mnt_drop_write(path.mnt);
3345 out:
3346         path_put(&path);
3347         return error;
3348 }
3349
3350 static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
3351 {
3352         struct path path;
3353         int error = path_lookupat(nd, flags, &path);
3354         if (!error) {
3355                 audit_inode(nd->name, path.dentry, 0);
3356                 error = vfs_open(&path, file, current_cred());
3357                 path_put(&path);
3358         }
3359         return error;
3360 }
3361
3362 static struct file *path_openat(struct nameidata *nd,
3363                         const struct open_flags *op, unsigned flags)
3364 {
3365         const char *s;
3366         struct file *file;
3367         int opened = 0;
3368         int error;
3369
3370         file = get_empty_filp();
3371         if (IS_ERR(file))
3372                 return file;
3373
3374         file->f_flags = op->open_flag;
3375
3376         if (unlikely(file->f_flags & __O_TMPFILE)) {
3377                 error = do_tmpfile(nd, flags, op, file, &opened);
3378                 goto out2;
3379         }
3380
3381         if (unlikely(file->f_flags & O_PATH)) {
3382                 error = do_o_path(nd, flags, file);
3383                 if (!error)
3384                         opened |= FILE_OPENED;
3385                 goto out2;
3386         }
3387
3388         s = path_init(nd, flags);
3389         if (IS_ERR(s)) {
3390                 put_filp(file);
3391                 return ERR_CAST(s);
3392         }
3393         while (!(error = link_path_walk(s, nd)) &&
3394                 (error = do_last(nd, file, op, &opened)) > 0) {
3395                 nd->flags &= ~(LOOKUP_OPEN|LOOKUP_CREATE|LOOKUP_EXCL);
3396                 s = trailing_symlink(nd);
3397                 if (IS_ERR(s)) {
3398                         error = PTR_ERR(s);
3399                         break;
3400                 }
3401         }
3402         terminate_walk(nd);
3403 out2:
3404         if (!(opened & FILE_OPENED)) {
3405                 BUG_ON(!error);
3406                 put_filp(file);
3407         }
3408         if (unlikely(error)) {
3409                 if (error == -EOPENSTALE) {
3410                         if (flags & LOOKUP_RCU)
3411                                 error = -ECHILD;
3412                         else
3413                                 error = -ESTALE;
3414                 }
3415                 file = ERR_PTR(error);
3416         }
3417         return file;
3418 }
3419
3420 struct file *do_filp_open(int dfd, struct filename *pathname,
3421                 const struct open_flags *op)
3422 {
3423         struct nameidata nd;
3424         int flags = op->lookup_flags;
3425         struct file *filp;
3426
3427         set_nameidata(&nd, dfd, pathname);
3428         filp = path_openat(&nd, op, flags | LOOKUP_RCU);
3429         if (unlikely(filp == ERR_PTR(-ECHILD)))
3430                 filp = path_openat(&nd, op, flags);
3431         if (unlikely(filp == ERR_PTR(-ESTALE)))
3432                 filp = path_openat(&nd, op, flags | LOOKUP_REVAL);
3433         restore_nameidata();
3434         return filp;
3435 }
3436
3437 struct file *do_file_open_root(struct dentry *dentry, struct vfsmount *mnt,
3438                 const char *name, const struct open_flags *op)
3439 {
3440         struct nameidata nd;
3441         struct file *file;
3442         struct filename *filename;
3443         int flags = op->lookup_flags | LOOKUP_ROOT;
3444
3445         nd.root.mnt = mnt;
3446         nd.root.dentry = dentry;
3447
3448         if (d_is_symlink(dentry) && op->intent & LOOKUP_OPEN)
3449                 return ERR_PTR(-ELOOP);
3450
3451         filename = getname_kernel(name);
3452         if (IS_ERR(filename))
3453                 return ERR_CAST(filename);
3454
3455         set_nameidata(&nd, -1, filename);
3456         file = path_openat(&nd, op, flags | LOOKUP_RCU);
3457         if (unlikely(file == ERR_PTR(-ECHILD)))
3458                 file = path_openat(&nd, op, flags);
3459         if (unlikely(file == ERR_PTR(-ESTALE)))
3460                 file = path_openat(&nd, op, flags | LOOKUP_REVAL);
3461         restore_nameidata();
3462         putname(filename);
3463         return file;
3464 }
3465
3466 static struct dentry *filename_create(int dfd, struct filename *name,
3467                                 struct path *path, unsigned int lookup_flags)
3468 {
3469         struct dentry *dentry = ERR_PTR(-EEXIST);
3470         struct qstr last;
3471         int type;
3472         int err2;
3473         int error;
3474         bool is_dir = (lookup_flags & LOOKUP_DIRECTORY);
3475
3476         /*
3477          * Note that only LOOKUP_REVAL and LOOKUP_DIRECTORY matter here. Any
3478          * other flags passed in are ignored!
3479          */
3480         lookup_flags &= LOOKUP_REVAL;
3481
3482         name = filename_parentat(dfd, name, lookup_flags, path, &last, &type);
3483         if (IS_ERR(name))
3484                 return ERR_CAST(name);
3485
3486         /*
3487          * Yucky last component or no last component at all?
3488          * (foo/., foo/.., /////)
3489          */
3490         if (unlikely(type != LAST_NORM))
3491                 goto out;
3492
3493         /* don't fail immediately if it's r/o, at least try to report other errors */
3494         err2 = mnt_want_write(path->mnt);
3495         /*
3496          * Do the final lookup.
3497          */
3498         lookup_flags |= LOOKUP_CREATE | LOOKUP_EXCL;
3499         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
3500         dentry = __lookup_hash(&last, path->dentry, lookup_flags);
3501         if (IS_ERR(dentry))
3502                 goto unlock;
3503
3504         error = -EEXIST;
3505         if (d_is_positive(dentry))
3506                 goto fail;
3507
3508         /*
3509          * Special case - lookup gave negative, but... we had foo/bar/
3510          * From the vfs_mknod() POV we just have a negative dentry -
3511          * all is fine. Let's be bastards - you had / on the end, you've
3512          * been asking for (non-existent) directory. -ENOENT for you.
3513          */
3514         if (unlikely(!is_dir && last.name[last.len])) {
3515                 error = -ENOENT;
3516                 goto fail;
3517         }
3518         if (unlikely(err2)) {
3519                 error = err2;
3520                 goto fail;
3521         }
3522         putname(name);
3523         return dentry;
3524 fail:
3525         dput(dentry);
3526         dentry = ERR_PTR(error);
3527 unlock:
3528         inode_unlock(path->dentry->d_inode);
3529         if (!err2)
3530                 mnt_drop_write(path->mnt);
3531 out:
3532         path_put(path);
3533         putname(name);
3534         return dentry;
3535 }
3536
3537 struct dentry *kern_path_create(int dfd, const char *pathname,
3538                                 struct path *path, unsigned int lookup_flags)
3539 {
3540         return filename_create(dfd, getname_kernel(pathname),
3541                                 path, lookup_flags);
3542 }
3543 EXPORT_SYMBOL(kern_path_create);
3544
3545 void done_path_create(struct path *path, struct dentry *dentry)
3546 {
3547         dput(dentry);
3548         inode_unlock(path->dentry->d_inode);
3549         mnt_drop_write(path->mnt);
3550         path_put(path);
3551 }
3552 EXPORT_SYMBOL(done_path_create);
3553
3554 inline struct dentry *user_path_create(int dfd, const char __user *pathname,
3555                                 struct path *path, unsigned int lookup_flags)
3556 {
3557         return filename_create(dfd, getname(pathname), path, lookup_flags);
3558 }
3559 EXPORT_SYMBOL(user_path_create);
3560
3561 int vfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3562 {
3563         int error = may_create(dir, dentry);
3564
3565         if (error)
3566                 return error;
3567
3568         if ((S_ISCHR(mode) || S_ISBLK(mode)) && !capable(CAP_MKNOD))
3569                 return -EPERM;
3570
3571         if (!dir->i_op->mknod)
3572                 return -EPERM;
3573
3574         error = devcgroup_inode_mknod(mode, dev);
3575         if (error)
3576                 return error;
3577
3578         error = security_inode_mknod(dir, dentry, mode, dev);
3579         if (error)
3580                 return error;
3581
3582         error = dir->i_op->mknod(dir, dentry, mode, dev);
3583         if (!error)
3584                 fsnotify_create(dir, dentry);
3585         return error;
3586 }
3587 EXPORT_SYMBOL(vfs_mknod);
3588
3589 static int may_mknod(umode_t mode)
3590 {
3591         switch (mode & S_IFMT) {
3592         case S_IFREG:
3593         case S_IFCHR:
3594         case S_IFBLK:
3595         case S_IFIFO:
3596         case S_IFSOCK:
3597         case 0: /* zero mode translates to S_IFREG */
3598                 return 0;
3599         case S_IFDIR:
3600                 return -EPERM;
3601         default:
3602                 return -EINVAL;
3603         }
3604 }
3605
3606 SYSCALL_DEFINE4(mknodat, int, dfd, const char __user *, filename, umode_t, mode,
3607                 unsigned, dev)
3608 {
3609         struct dentry *dentry;
3610         struct path path;
3611         int error;
3612         unsigned int lookup_flags = 0;
3613
3614         error = may_mknod(mode);
3615         if (error)
3616                 return error;
3617 retry:
3618         dentry = user_path_create(dfd, filename, &path, lookup_flags);
3619         if (IS_ERR(dentry))
3620                 return PTR_ERR(dentry);
3621
3622         if (!IS_POSIXACL(path.dentry->d_inode))
3623                 mode &= ~current_umask();
3624         error = security_path_mknod(&path, dentry, mode, dev);
3625         if (error)
3626                 goto out;
3627         switch (mode & S_IFMT) {
3628                 case 0: case S_IFREG:
3629                         error = vfs_create(path.dentry->d_inode,dentry,mode,true);
3630                         if (!error)
3631                                 ima_post_path_mknod(dentry);
3632                         break;
3633                 case S_IFCHR: case S_IFBLK:
3634                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,
3635                                         new_decode_dev(dev));
3636                         break;
3637                 case S_IFIFO: case S_IFSOCK:
3638                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,0);
3639                         break;
3640         }
3641 out:
3642         done_path_create(&path, dentry);
3643         if (retry_estale(error, lookup_flags)) {
3644                 lookup_flags |= LOOKUP_REVAL;
3645                 goto retry;
3646         }
3647         return error;
3648 }
3649
3650 SYSCALL_DEFINE3(mknod, const char __user *, filename, umode_t, mode, unsigned, dev)
3651 {
3652         return sys_mknodat(AT_FDCWD, filename, mode, dev);
3653 }
3654
3655 int vfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3656 {
3657         int error = may_create(dir, dentry);
3658         unsigned max_links = dir->i_sb->s_max_links;
3659
3660         if (error)
3661                 return error;
3662
3663         if (!dir->i_op->mkdir)
3664                 return -EPERM;
3665
3666         mode &= (S_IRWXUGO|S_ISVTX);
3667         error = security_inode_mkdir(dir, dentry, mode);
3668         if (error)
3669                 return error;
3670
3671         if (max_links && dir->i_nlink >= max_links)
3672                 return -EMLINK;
3673
3674         error = dir->i_op->mkdir(dir, dentry, mode);
3675         if (!error)
3676                 fsnotify_mkdir(dir, dentry);
3677         return error;
3678 }
3679 EXPORT_SYMBOL(vfs_mkdir);
3680
3681 SYSCALL_DEFINE3(mkdirat, int, dfd, const char __user *, pathname, umode_t, mode)
3682 {
3683         struct dentry *dentry;
3684         struct path path;
3685         int error;
3686         unsigned int lookup_flags = LOOKUP_DIRECTORY;
3687
3688 retry:
3689         dentry = user_path_create(dfd, pathname, &path, lookup_flags);
3690         if (IS_ERR(dentry))
3691                 return PTR_ERR(dentry);
3692
3693         if (!IS_POSIXACL(path.dentry->d_inode))
3694                 mode &= ~current_umask();
3695         error = security_path_mkdir(&path, dentry, mode);
3696         if (!error)
3697                 error = vfs_mkdir(path.dentry->d_inode, dentry, mode);
3698         done_path_create(&path, dentry);
3699         if (retry_estale(error, lookup_flags)) {
3700                 lookup_flags |= LOOKUP_REVAL;
3701                 goto retry;
3702         }
3703         return error;
3704 }
3705
3706 SYSCALL_DEFINE2(mkdir, const char __user *, pathname, umode_t, mode)
3707 {
3708         return sys_mkdirat(AT_FDCWD, pathname, mode);
3709 }
3710
3711 int vfs_rmdir(struct inode *dir, struct dentry *dentry)
3712 {
3713         int error = may_delete(dir, dentry, 1);
3714
3715         if (error)
3716                 return error;
3717
3718         if (!dir->i_op->rmdir)
3719                 return -EPERM;
3720
3721         dget(dentry);
3722         inode_lock(dentry->d_inode);
3723
3724         error = -EBUSY;
3725         if (is_local_mountpoint(dentry))
3726                 goto out;
3727
3728         error = security_inode_rmdir(dir, dentry);
3729         if (error)
3730                 goto out;
3731
3732         shrink_dcache_parent(dentry);
3733         error = dir->i_op->rmdir(dir, dentry);
3734         if (error)
3735                 goto out;
3736
3737         dentry->d_inode->i_flags |= S_DEAD;
3738         dont_mount(dentry);
3739         detach_mounts(dentry);
3740
3741 out:
3742         inode_unlock(dentry->d_inode);
3743         dput(dentry);
3744         if (!error)
3745                 d_delete(dentry);
3746         return error;
3747 }
3748 EXPORT_SYMBOL(vfs_rmdir);
3749
3750 static long do_rmdir(int dfd, const char __user *pathname)
3751 {
3752         int error = 0;
3753         struct filename *name;
3754         struct dentry *dentry;
3755         struct path path;
3756         struct qstr last;
3757         int type;
3758         unsigned int lookup_flags = 0;
3759 retry:
3760         name = user_path_parent(dfd, pathname,
3761                                 &path, &last, &type, lookup_flags);
3762         if (IS_ERR(name))
3763                 return PTR_ERR(name);
3764
3765         switch (type) {
3766         case LAST_DOTDOT:
3767                 error = -ENOTEMPTY;
3768                 goto exit1;
3769         case LAST_DOT:
3770                 error = -EINVAL;
3771                 goto exit1;
3772         case LAST_ROOT:
3773                 error = -EBUSY;
3774                 goto exit1;
3775         }
3776
3777         error = mnt_want_write(path.mnt);
3778         if (error)
3779                 goto exit1;
3780
3781         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3782         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3783         error = PTR_ERR(dentry);
3784         if (IS_ERR(dentry))
3785                 goto exit2;
3786         if (!dentry->d_inode) {
3787                 error = -ENOENT;
3788                 goto exit3;
3789         }
3790         error = security_path_rmdir(&path, dentry);
3791         if (error)
3792                 goto exit3;
3793         error = vfs_rmdir(path.dentry->d_inode, dentry);
3794 exit3:
3795         dput(dentry);
3796 exit2:
3797         inode_unlock(path.dentry->d_inode);
3798         mnt_drop_write(path.mnt);
3799 exit1:
3800         path_put(&path);
3801         putname(name);
3802         if (retry_estale(error, lookup_flags)) {
3803                 lookup_flags |= LOOKUP_REVAL;
3804                 goto retry;
3805         }
3806         return error;
3807 }
3808
3809 SYSCALL_DEFINE1(rmdir, const char __user *, pathname)
3810 {
3811         return do_rmdir(AT_FDCWD, pathname);
3812 }
3813
3814 /**
3815  * vfs_unlink - unlink a filesystem object
3816  * @dir:        parent directory
3817  * @dentry:     victim
3818  * @delegated_inode: returns victim inode, if the inode is delegated.
3819  *
3820  * The caller must hold dir->i_mutex.
3821  *
3822  * If vfs_unlink discovers a delegation, it will return -EWOULDBLOCK and
3823  * return a reference to the inode in delegated_inode.  The caller
3824  * should then break the delegation on that inode and retry.  Because
3825  * breaking a delegation may take a long time, the caller should drop
3826  * dir->i_mutex before doing so.
3827  *
3828  * Alternatively, a caller may pass NULL for delegated_inode.  This may
3829  * be appropriate for callers that expect the underlying filesystem not
3830  * to be NFS exported.
3831  */
3832 int vfs_unlink(struct inode *dir, struct dentry *dentry, struct inode **delegated_inode)
3833 {
3834         struct inode *target = dentry->d_inode;
3835         int error = may_delete(dir, dentry, 0);
3836
3837         if (error)
3838                 return error;
3839
3840         if (!dir->i_op->unlink)
3841                 return -EPERM;
3842
3843         inode_lock(target);
3844         if (is_local_mountpoint(dentry))
3845                 error = -EBUSY;
3846         else {
3847                 error = security_inode_unlink(dir, dentry);
3848                 if (!error) {
3849                         error = try_break_deleg(target, delegated_inode);
3850                         if (error)
3851                                 goto out;
3852                         error = dir->i_op->unlink(dir, dentry);
3853                         if (!error) {
3854                                 dont_mount(dentry);
3855                                 detach_mounts(dentry);
3856                         }
3857                 }
3858         }
3859 out:
3860         inode_unlock(target);
3861
3862         /* We don't d_delete() NFS sillyrenamed files--they still exist. */
3863         if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
3864                 fsnotify_link_count(target);
3865                 d_delete(dentry);
3866         }
3867
3868         return error;
3869 }
3870 EXPORT_SYMBOL(vfs_unlink);
3871
3872 /*
3873  * Make sure that the actual truncation of the file will occur outside its
3874  * directory's i_mutex.  Truncate can take a long time if there is a lot of
3875  * writeout happening, and we don't want to prevent access to the directory
3876  * while waiting on the I/O.
3877  */
3878 static long do_unlinkat(int dfd, const char __user *pathname)
3879 {
3880         int error;
3881         struct filename *name;
3882         struct dentry *dentry;
3883         struct path path;
3884         struct qstr last;
3885         int type;
3886         struct inode *inode = NULL;
3887         struct inode *delegated_inode = NULL;
3888         unsigned int lookup_flags = 0;
3889 retry:
3890         name = user_path_parent(dfd, pathname,
3891                                 &path, &last, &type, lookup_flags);
3892         if (IS_ERR(name))
3893                 return PTR_ERR(name);
3894
3895         error = -EISDIR;
3896         if (type != LAST_NORM)
3897                 goto exit1;
3898
3899         error = mnt_want_write(path.mnt);
3900         if (error)
3901                 goto exit1;
3902 retry_deleg:
3903         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3904         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3905         error = PTR_ERR(dentry);
3906         if (!IS_ERR(dentry)) {
3907                 /* Why not before? Because we want correct error value */
3908                 if (last.name[last.len])
3909                         goto slashes;
3910                 inode = dentry->d_inode;
3911                 if (d_is_negative(dentry))
3912                         goto slashes;
3913                 ihold(inode);
3914                 error = security_path_unlink(&path, dentry);
3915                 if (error)
3916                         goto exit2;
3917                 error = vfs_unlink(path.dentry->d_inode, dentry, &delegated_inode);
3918 exit2:
3919                 dput(dentry);
3920         }
3921         inode_unlock(path.dentry->d_inode);
3922         if (inode)
3923                 iput(inode);    /* truncate the inode here */
3924         inode = NULL;
3925         if (delegated_inode) {
3926                 error = break_deleg_wait(&delegated_inode);
3927                 if (!error)
3928                         goto retry_deleg;
3929         }
3930         mnt_drop_write(path.mnt);
3931 exit1:
3932         path_put(&path);
3933         putname(name);
3934         if (retry_estale(error, lookup_flags)) {
3935                 lookup_flags |= LOOKUP_REVAL;
3936                 inode = NULL;
3937                 goto retry;
3938         }
3939         return error;
3940
3941 slashes:
3942         if (d_is_negative(dentry))
3943                 error = -ENOENT;
3944         else if (d_is_dir(dentry))
3945                 error = -EISDIR;
3946         else
3947                 error = -ENOTDIR;
3948         goto exit2;
3949 }
3950
3951 SYSCALL_DEFINE3(unlinkat, int, dfd, const char __user *, pathname, int, flag)
3952 {
3953         if ((flag & ~AT_REMOVEDIR) != 0)
3954                 return -EINVAL;
3955
3956         if (flag & AT_REMOVEDIR)
3957                 return do_rmdir(dfd, pathname);
3958
3959         return do_unlinkat(dfd, pathname);
3960 }
3961
3962 SYSCALL_DEFINE1(unlink, const char __user *, pathname)
3963 {
3964         return do_unlinkat(AT_FDCWD, pathname);
3965 }
3966
3967 int vfs_symlink(struct inode *dir, struct dentry *dentry, const char *oldname)
3968 {
3969         int error = may_create(dir, dentry);
3970
3971         if (error)
3972                 return error;
3973
3974         if (!dir->i_op->symlink)
3975                 return -EPERM;
3976
3977         error = security_inode_symlink(dir, dentry, oldname);
3978         if (error)
3979                 return error;
3980
3981         error = dir->i_op->symlink(dir, dentry, oldname);
3982         if (!error)
3983                 fsnotify_create(dir, dentry);
3984         return error;
3985 }
3986 EXPORT_SYMBOL(vfs_symlink);
3987
3988 SYSCALL_DEFINE3(symlinkat, const char __user *, oldname,
3989                 int, newdfd, const char __user *, newname)
3990 {
3991         int error;
3992         struct filename *from;
3993         struct dentry *dentry;
3994         struct path path;
3995         unsigned int lookup_flags = 0;
3996
3997         from = getname(oldname);
3998         if (IS_ERR(from))
3999                 return PTR_ERR(from);
4000 retry:
4001         dentry = user_path_create(newdfd, newname, &path, lookup_flags);
4002         error = PTR_ERR(dentry);
4003         if (IS_ERR(dentry))
4004                 goto out_putname;
4005
4006         error = security_path_symlink(&path, dentry, from->name);
4007         if (!error)
4008                 error = vfs_symlink(path.dentry->d_inode, dentry, from->name);
4009         done_path_create(&path, dentry);
4010         if (retry_estale(error, lookup_flags)) {
4011                 lookup_flags |= LOOKUP_REVAL;
4012                 goto retry;
4013         }
4014 out_putname:
4015         putname(from);
4016         return error;
4017 }
4018
4019 SYSCALL_DEFINE2(symlink, const char __user *, oldname, const char __user *, newname)
4020 {
4021         return sys_symlinkat(oldname, AT_FDCWD, newname);
4022 }
4023
4024 /**
4025  * vfs_link - create a new link
4026  * @old_dentry: object to be linked
4027  * @dir:        new parent
4028  * @new_dentry: where to create the new link
4029  * @delegated_inode: returns inode needing a delegation break
4030  *
4031  * The caller must hold dir->i_mutex
4032  *
4033  * If vfs_link discovers a delegation on the to-be-linked file in need
4034  * of breaking, it will return -EWOULDBLOCK and return a reference to the
4035  * inode in delegated_inode.  The caller should then break the delegation
4036  * and retry.  Because breaking a delegation may take a long time, the
4037  * caller should drop the i_mutex before doing so.
4038  *
4039  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4040  * be appropriate for callers that expect the underlying filesystem not
4041  * to be NFS exported.
4042  */
4043 int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_dentry, struct inode **delegated_inode)
4044 {
4045         struct inode *inode = old_dentry->d_inode;
4046         unsigned max_links = dir->i_sb->s_max_links;
4047         int error;
4048
4049         if (!inode)
4050                 return -ENOENT;
4051
4052         error = may_create(dir, new_dentry);
4053         if (error)
4054                 return error;
4055
4056         if (dir->i_sb != inode->i_sb)
4057                 return -EXDEV;
4058
4059         /*
4060          * A link to an append-only or immutable file cannot be created.
4061          */
4062         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4063                 return -EPERM;
4064         if (!dir->i_op->link)
4065                 return -EPERM;
4066         if (S_ISDIR(inode->i_mode))
4067                 return -EPERM;
4068
4069         error = security_inode_link(old_dentry, dir, new_dentry);
4070         if (error)
4071                 return error;
4072
4073         inode_lock(inode);
4074         /* Make sure we don't allow creating hardlink to an unlinked file */
4075         if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))
4076                 error =  -ENOENT;
4077         else if (max_links && inode->i_nlink >= max_links)
4078                 error = -EMLINK;
4079         else {
4080                 error = try_break_deleg(inode, delegated_inode);
4081                 if (!error)
4082                         error = dir->i_op->link(old_dentry, dir, new_dentry);
4083         }
4084
4085         if (!error && (inode->i_state & I_LINKABLE)) {
4086                 spin_lock(&inode->i_lock);
4087                 inode->i_state &= ~I_LINKABLE;
4088                 spin_unlock(&inode->i_lock);
4089         }
4090         inode_unlock(inode);
4091         if (!error)
4092                 fsnotify_link(dir, inode, new_dentry);
4093         return error;
4094 }
4095 EXPORT_SYMBOL(vfs_link);
4096
4097 /*
4098  * Hardlinks are often used in delicate situations.  We avoid
4099  * security-related surprises by not following symlinks on the
4100  * newname.  --KAB
4101  *
4102  * We don't follow them on the oldname either to be compatible
4103  * with linux 2.0, and to avoid hard-linking to directories
4104  * and other special files.  --ADM
4105  */
4106 SYSCALL_DEFINE5(linkat, int, olddfd, const char __user *, oldname,
4107                 int, newdfd, const char __user *, newname, int, flags)
4108 {
4109         struct dentry *new_dentry;
4110         struct path old_path, new_path;
4111         struct inode *delegated_inode = NULL;
4112         int how = 0;
4113         int error;
4114
4115         if ((flags & ~(AT_SYMLINK_FOLLOW | AT_EMPTY_PATH)) != 0)
4116                 return -EINVAL;
4117         /*
4118          * To use null names we require CAP_DAC_READ_SEARCH
4119          * This ensures that not everyone will be able to create
4120          * handlink using the passed filedescriptor.
4121          */
4122         if (flags & AT_EMPTY_PATH) {
4123                 if (!capable(CAP_DAC_READ_SEARCH))
4124                         return -ENOENT;
4125                 how = LOOKUP_EMPTY;
4126         }
4127
4128         if (flags & AT_SYMLINK_FOLLOW)
4129                 how |= LOOKUP_FOLLOW;
4130 retry:
4131         error = user_path_at(olddfd, oldname, how, &old_path);
4132         if (error)
4133                 return error;
4134
4135         new_dentry = user_path_create(newdfd, newname, &new_path,
4136                                         (how & LOOKUP_REVAL));
4137         error = PTR_ERR(new_dentry);
4138         if (IS_ERR(new_dentry))
4139                 goto out;
4140
4141         error = -EXDEV;
4142         if (old_path.mnt != new_path.mnt)
4143                 goto out_dput;
4144         error = may_linkat(&old_path);
4145         if (unlikely(error))
4146                 goto out_dput;
4147         error = security_path_link(old_path.dentry, &new_path, new_dentry);
4148         if (error)
4149                 goto out_dput;
4150         error = vfs_link(old_path.dentry, new_path.dentry->d_inode, new_dentry, &delegated_inode);
4151 out_dput:
4152         done_path_create(&new_path, new_dentry);
4153         if (delegated_inode) {
4154                 error = break_deleg_wait(&delegated_inode);
4155                 if (!error) {
4156                         path_put(&old_path);
4157                         goto retry;
4158                 }
4159         }
4160         if (retry_estale(error, how)) {
4161                 path_put(&old_path);
4162                 how |= LOOKUP_REVAL;
4163                 goto retry;
4164         }
4165 out:
4166         path_put(&old_path);
4167
4168         return error;
4169 }
4170
4171 SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname)
4172 {
4173         return sys_linkat(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4174 }
4175
4176 /**
4177  * vfs_rename - rename a filesystem object
4178  * @old_dir:    parent of source
4179  * @old_dentry: source
4180  * @new_dir:    parent of destination
4181  * @new_dentry: destination
4182  * @delegated_inode: returns an inode needing a delegation break
4183  * @flags:      rename flags
4184  *
4185  * The caller must hold multiple mutexes--see lock_rename()).
4186  *
4187  * If vfs_rename discovers a delegation in need of breaking at either
4188  * the source or destination, it will return -EWOULDBLOCK and return a
4189  * reference to the inode in delegated_inode.  The caller should then
4190  * break the delegation and retry.  Because breaking a delegation may
4191  * take a long time, the caller should drop all locks before doing
4192  * so.
4193  *
4194  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4195  * be appropriate for callers that expect the underlying filesystem not
4196  * to be NFS exported.
4197  *
4198  * The worst of all namespace operations - renaming directory. "Perverted"
4199  * doesn't even start to describe it. Somebody in UCB had a heck of a trip...
4200  * Problems:
4201  *      a) we can get into loop creation.
4202  *      b) race potential - two innocent renames can create a loop together.
4203  *         That's where 4.4 screws up. Current fix: serialization on
4204  *         sb->s_vfs_rename_mutex. We might be more accurate, but that's another
4205  *         story.
4206  *      c) we have to lock _four_ objects - parents and victim (if it exists),
4207  *         and source (if it is not a directory).
4208  *         And that - after we got ->i_mutex on parents (until then we don't know
4209  *         whether the target exists).  Solution: try to be smart with locking
4210  *         order for inodes.  We rely on the fact that tree topology may change
4211  *         only under ->s_vfs_rename_mutex _and_ that parent of the object we
4212  *         move will be locked.  Thus we can rank directories by the tree
4213  *         (ancestors first) and rank all non-directories after them.
4214  *         That works since everybody except rename does "lock parent, lookup,
4215  *         lock child" and rename is under ->s_vfs_rename_mutex.
4216  *         HOWEVER, it relies on the assumption that any object with ->lookup()
4217  *         has no more than 1 dentry.  If "hybrid" objects will ever appear,
4218  *         we'd better make sure that there's no link(2) for them.
4219  *      d) conversion from fhandle to dentry may come in the wrong moment - when
4220  *         we are removing the target. Solution: we will have to grab ->i_mutex
4221  *         in the fhandle_to_dentry code. [FIXME - current nfsfh.c relies on
4222  *         ->i_mutex on parents, which works but leads to some truly excessive
4223  *         locking].
4224  */
4225 int vfs_rename(struct inode *old_dir, struct dentry *old_dentry,
4226                struct inode *new_dir, struct dentry *new_dentry,
4227                struct inode **delegated_inode, unsigned int flags)
4228 {
4229         int error;
4230         bool is_dir = d_is_dir(old_dentry);
4231         const unsigned char *old_name;
4232         struct inode *source = old_dentry->d_inode;
4233         struct inode *target = new_dentry->d_inode;
4234         bool new_is_dir = false;
4235         unsigned max_links = new_dir->i_sb->s_max_links;
4236
4237         /*
4238          * Check source == target.
4239          * On overlayfs need to look at underlying inodes.
4240          */
4241         if (vfs_select_inode(old_dentry, 0) == vfs_select_inode(new_dentry, 0))
4242                 return 0;
4243
4244         error = may_delete(old_dir, old_dentry, is_dir);
4245         if (error)
4246                 return error;
4247
4248         if (!target) {
4249                 error = may_create(new_dir, new_dentry);
4250         } else {
4251                 new_is_dir = d_is_dir(new_dentry);
4252
4253                 if (!(flags & RENAME_EXCHANGE))
4254                         error = may_delete(new_dir, new_dentry, is_dir);
4255                 else
4256                         error = may_delete(new_dir, new_dentry, new_is_dir);
4257         }
4258         if (error)
4259                 return error;
4260
4261         if (!old_dir->i_op->rename && !old_dir->i_op->rename2)
4262                 return -EPERM;
4263
4264         if (flags && !old_dir->i_op->rename2)
4265                 return -EINVAL;
4266
4267         /*
4268          * If we are going to change the parent - check write permissions,
4269          * we'll need to flip '..'.
4270          */
4271         if (new_dir != old_dir) {
4272                 if (is_dir) {
4273                         error = inode_permission(source, MAY_WRITE);
4274                         if (error)
4275                                 return error;
4276                 }
4277                 if ((flags & RENAME_EXCHANGE) && new_is_dir) {
4278                         error = inode_permission(target, MAY_WRITE);
4279                         if (error)
4280                                 return error;
4281                 }
4282         }
4283
4284         error = security_inode_rename(old_dir, old_dentry, new_dir, new_dentry,
4285                                       flags);
4286         if (error)
4287                 return error;
4288
4289         old_name = fsnotify_oldname_init(old_dentry->d_name.name);
4290         dget(new_dentry);
4291         if (!is_dir || (flags & RENAME_EXCHANGE))
4292                 lock_two_nondirectories(source, target);
4293         else if (target)
4294                 inode_lock(target);
4295
4296         error = -EBUSY;
4297         if (is_local_mountpoint(old_dentry) || is_local_mountpoint(new_dentry))
4298                 goto out;
4299
4300         if (max_links && new_dir != old_dir) {
4301                 error = -EMLINK;
4302                 if (is_dir && !new_is_dir && new_dir->i_nlink >= max_links)
4303                         goto out;
4304                 if ((flags & RENAME_EXCHANGE) && !is_dir && new_is_dir &&
4305                     old_dir->i_nlink >= max_links)
4306                         goto out;
4307         }
4308         if (is_dir && !(flags & RENAME_EXCHANGE) && target)
4309                 shrink_dcache_parent(new_dentry);
4310         if (!is_dir) {
4311                 error = try_break_deleg(source, delegated_inode);
4312                 if (error)
4313                         goto out;
4314         }
4315         if (target && !new_is_dir) {
4316                 error = try_break_deleg(target, delegated_inode);
4317                 if (error)
4318                         goto out;
4319         }
4320         if (!old_dir->i_op->rename2) {
4321                 error = old_dir->i_op->rename(old_dir, old_dentry,
4322                                               new_dir, new_dentry);
4323         } else {
4324                 WARN_ON(old_dir->i_op->rename != NULL);
4325                 error = old_dir->i_op->rename2(old_dir, old_dentry,
4326                                                new_dir, new_dentry, flags);
4327         }
4328         if (error)
4329                 goto out;
4330
4331         if (!(flags & RENAME_EXCHANGE) && target) {
4332                 if (is_dir)
4333                         target->i_flags |= S_DEAD;
4334                 dont_mount(new_dentry);
4335                 detach_mounts(new_dentry);
4336         }
4337         if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE)) {
4338                 if (!(flags & RENAME_EXCHANGE))
4339                         d_move(old_dentry, new_dentry);
4340                 else
4341                         d_exchange(old_dentry, new_dentry);
4342         }
4343 out:
4344         if (!is_dir || (flags & RENAME_EXCHANGE))
4345                 unlock_two_nondirectories(source, target);
4346         else if (target)
4347                 inode_unlock(target);
4348         dput(new_dentry);
4349         if (!error) {
4350                 fsnotify_move(old_dir, new_dir, old_name, is_dir,
4351                               !(flags & RENAME_EXCHANGE) ? target : NULL, old_dentry);
4352                 if (flags & RENAME_EXCHANGE) {
4353                         fsnotify_move(new_dir, old_dir, old_dentry->d_name.name,
4354                                       new_is_dir, NULL, new_dentry);
4355                 }
4356         }
4357         fsnotify_oldname_free(old_name);
4358
4359         return error;
4360 }
4361 EXPORT_SYMBOL(vfs_rename);
4362
4363 SYSCALL_DEFINE5(renameat2, int, olddfd, const char __user *, oldname,
4364                 int, newdfd, const char __user *, newname, unsigned int, flags)
4365 {
4366         struct dentry *old_dentry, *new_dentry;
4367         struct dentry *trap;
4368         struct path old_path, new_path;
4369         struct qstr old_last, new_last;
4370         int old_type, new_type;
4371         struct inode *delegated_inode = NULL;
4372         struct filename *from;
4373         struct filename *to;
4374         unsigned int lookup_flags = 0, target_flags = LOOKUP_RENAME_TARGET;
4375         bool should_retry = false;
4376         int error;
4377
4378         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
4379                 return -EINVAL;
4380
4381         if ((flags & (RENAME_NOREPLACE | RENAME_WHITEOUT)) &&
4382             (flags & RENAME_EXCHANGE))
4383                 return -EINVAL;
4384
4385         if ((flags & RENAME_WHITEOUT) && !capable(CAP_MKNOD))
4386                 return -EPERM;
4387
4388         if (flags & RENAME_EXCHANGE)
4389                 target_flags = 0;
4390
4391 retry:
4392         from = user_path_parent(olddfd, oldname,
4393                                 &old_path, &old_last, &old_type, lookup_flags);
4394         if (IS_ERR(from)) {
4395                 error = PTR_ERR(from);
4396                 goto exit;
4397         }
4398
4399         to = user_path_parent(newdfd, newname,
4400                                 &new_path, &new_last, &new_type, lookup_flags);
4401         if (IS_ERR(to)) {
4402                 error = PTR_ERR(to);
4403                 goto exit1;
4404         }
4405
4406         error = -EXDEV;
4407         if (old_path.mnt != new_path.mnt)
4408                 goto exit2;
4409
4410         error = -EBUSY;
4411         if (old_type != LAST_NORM)
4412                 goto exit2;
4413
4414         if (flags & RENAME_NOREPLACE)
4415                 error = -EEXIST;
4416         if (new_type != LAST_NORM)
4417                 goto exit2;
4418
4419         error = mnt_want_write(old_path.mnt);
4420         if (error)
4421                 goto exit2;
4422
4423 retry_deleg:
4424         trap = lock_rename(new_path.dentry, old_path.dentry);
4425
4426         old_dentry = __lookup_hash(&old_last, old_path.dentry, lookup_flags);
4427         error = PTR_ERR(old_dentry);
4428         if (IS_ERR(old_dentry))
4429                 goto exit3;
4430         /* source must exist */
4431         error = -ENOENT;
4432         if (d_is_negative(old_dentry))
4433                 goto exit4;
4434         new_dentry = __lookup_hash(&new_last, new_path.dentry, lookup_flags | target_flags);
4435         error = PTR_ERR(new_dentry);
4436         if (IS_ERR(new_dentry))
4437                 goto exit4;
4438         error = -EEXIST;
4439         if ((flags & RENAME_NOREPLACE) && d_is_positive(new_dentry))
4440                 goto exit5;
4441         if (flags & RENAME_EXCHANGE) {
4442                 error = -ENOENT;
4443                 if (d_is_negative(new_dentry))
4444                         goto exit5;
4445
4446                 if (!d_is_dir(new_dentry)) {
4447                         error = -ENOTDIR;
4448                         if (new_last.name[new_last.len])
4449                                 goto exit5;
4450                 }
4451         }
4452         /* unless the source is a directory trailing slashes give -ENOTDIR */
4453         if (!d_is_dir(old_dentry)) {
4454                 error = -ENOTDIR;
4455                 if (old_last.name[old_last.len])
4456                         goto exit5;
4457                 if (!(flags & RENAME_EXCHANGE) && new_last.name[new_last.len])
4458                         goto exit5;
4459         }
4460         /* source should not be ancestor of target */
4461         error = -EINVAL;
4462         if (old_dentry == trap)
4463                 goto exit5;
4464         /* target should not be an ancestor of source */
4465         if (!(flags & RENAME_EXCHANGE))
4466                 error = -ENOTEMPTY;
4467         if (new_dentry == trap)
4468                 goto exit5;
4469
4470         error = security_path_rename(&old_path, old_dentry,
4471                                      &new_path, new_dentry, flags);
4472         if (error)
4473                 goto exit5;
4474         error = vfs_rename(old_path.dentry->d_inode, old_dentry,
4475                            new_path.dentry->d_inode, new_dentry,
4476                            &delegated_inode, flags);
4477 exit5:
4478         dput(new_dentry);
4479 exit4:
4480         dput(old_dentry);
4481 exit3:
4482         unlock_rename(new_path.dentry, old_path.dentry);
4483         if (delegated_inode) {
4484                 error = break_deleg_wait(&delegated_inode);
4485                 if (!error)
4486                         goto retry_deleg;
4487         }
4488         mnt_drop_write(old_path.mnt);
4489 exit2:
4490         if (retry_estale(error, lookup_flags))
4491                 should_retry = true;
4492         path_put(&new_path);
4493         putname(to);
4494 exit1:
4495         path_put(&old_path);
4496         putname(from);
4497         if (should_retry) {
4498                 should_retry = false;
4499                 lookup_flags |= LOOKUP_REVAL;
4500                 goto retry;
4501         }
4502 exit:
4503         return error;
4504 }
4505
4506 SYSCALL_DEFINE4(renameat, int, olddfd, const char __user *, oldname,
4507                 int, newdfd, const char __user *, newname)
4508 {
4509         return sys_renameat2(olddfd, oldname, newdfd, newname, 0);
4510 }
4511
4512 SYSCALL_DEFINE2(rename, const char __user *, oldname, const char __user *, newname)
4513 {
4514         return sys_renameat2(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4515 }
4516
4517 int vfs_whiteout(struct inode *dir, struct dentry *dentry)
4518 {
4519         int error = may_create(dir, dentry);
4520         if (error)
4521                 return error;
4522
4523         if (!dir->i_op->mknod)
4524                 return -EPERM;
4525
4526         return dir->i_op->mknod(dir, dentry,
4527                                 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
4528 }
4529 EXPORT_SYMBOL(vfs_whiteout);
4530
4531 int readlink_copy(char __user *buffer, int buflen, const char *link)
4532 {
4533         int len = PTR_ERR(link);
4534         if (IS_ERR(link))
4535                 goto out;
4536
4537         len = strlen(link);
4538         if (len > (unsigned) buflen)
4539                 len = buflen;
4540         if (copy_to_user(buffer, link, len))
4541                 len = -EFAULT;
4542 out:
4543         return len;
4544 }
4545 EXPORT_SYMBOL(readlink_copy);
4546
4547 /*
4548  * A helper for ->readlink().  This should be used *ONLY* for symlinks that
4549  * have ->get_link() not calling nd_jump_link().  Using (or not using) it
4550  * for any given inode is up to filesystem.
4551  */
4552 int generic_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4553 {
4554         DEFINE_DELAYED_CALL(done);
4555         struct inode *inode = d_inode(dentry);
4556         const char *link = inode->i_link;
4557         int res;
4558
4559         if (!link) {
4560                 link = inode->i_op->get_link(dentry, inode, &done);
4561                 if (IS_ERR(link))
4562                         return PTR_ERR(link);
4563         }
4564         res = readlink_copy(buffer, buflen, link);
4565         do_delayed_call(&done);
4566         return res;
4567 }
4568 EXPORT_SYMBOL(generic_readlink);
4569
4570 /* get the link contents into pagecache */
4571 const char *page_get_link(struct dentry *dentry, struct inode *inode,
4572                           struct delayed_call *callback)
4573 {
4574         char *kaddr;
4575         struct page *page;
4576         struct address_space *mapping = inode->i_mapping;
4577
4578         if (!dentry) {
4579                 page = find_get_page(mapping, 0);
4580                 if (!page)
4581                         return ERR_PTR(-ECHILD);
4582                 if (!PageUptodate(page)) {
4583                         put_page(page);
4584                         return ERR_PTR(-ECHILD);
4585                 }
4586         } else {
4587                 page = read_mapping_page(mapping, 0, NULL);
4588                 if (IS_ERR(page))
4589                         return (char*)page;
4590         }
4591         set_delayed_call(callback, page_put_link, page);
4592         BUG_ON(mapping_gfp_mask(mapping) & __GFP_HIGHMEM);
4593         kaddr = page_address(page);
4594         nd_terminate_link(kaddr, inode->i_size, PAGE_SIZE - 1);
4595         return kaddr;
4596 }
4597
4598 EXPORT_SYMBOL(page_get_link);
4599
4600 void page_put_link(void *arg)
4601 {
4602         put_page(arg);
4603 }
4604 EXPORT_SYMBOL(page_put_link);
4605
4606 int page_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4607 {
4608         DEFINE_DELAYED_CALL(done);
4609         int res = readlink_copy(buffer, buflen,
4610                                 page_get_link(dentry, d_inode(dentry),
4611                                               &done));
4612         do_delayed_call(&done);
4613         return res;
4614 }
4615 EXPORT_SYMBOL(page_readlink);
4616
4617 /*
4618  * The nofs argument instructs pagecache_write_begin to pass AOP_FLAG_NOFS
4619  */
4620 int __page_symlink(struct inode *inode, const char *symname, int len, int nofs)
4621 {
4622         struct address_space *mapping = inode->i_mapping;
4623         struct page *page;
4624         void *fsdata;
4625         int err;
4626         unsigned int flags = AOP_FLAG_UNINTERRUPTIBLE;
4627         if (nofs)
4628                 flags |= AOP_FLAG_NOFS;
4629
4630 retry:
4631         err = pagecache_write_begin(NULL, mapping, 0, len-1,
4632                                 flags, &page, &fsdata);
4633         if (err)
4634                 goto fail;
4635
4636         memcpy(page_address(page), symname, len-1);
4637
4638         err = pagecache_write_end(NULL, mapping, 0, len-1, len-1,
4639                                                         page, fsdata);
4640         if (err < 0)
4641                 goto fail;
4642         if (err < len-1)
4643                 goto retry;
4644
4645         mark_inode_dirty(inode);
4646         return 0;
4647 fail:
4648         return err;
4649 }
4650 EXPORT_SYMBOL(__page_symlink);
4651
4652 int page_symlink(struct inode *inode, const char *symname, int len)
4653 {
4654         return __page_symlink(inode, symname, len,
4655                         !mapping_gfp_constraint(inode->i_mapping, __GFP_FS));
4656 }
4657 EXPORT_SYMBOL(page_symlink);
4658
4659 const struct inode_operations page_symlink_inode_operations = {
4660         .readlink       = generic_readlink,
4661         .get_link       = page_get_link,
4662 };
4663 EXPORT_SYMBOL(page_symlink_inode_operations);