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 <linux/bitops.h>
  39 #include <asm/uaccess.h>
  40
  41 #include "internal.h"
  42 #include "mount.h"
  43
  44 /* [Feb-1997 T. Schoebel-Theuer]
  45  * Fundamental changes in the pathname lookup mechanisms (namei)
  46  * were necessary because of omirr.  The reason is that omirr needs
  47  * to know the _real_ pathname, not the user-supplied one, in case
  48  * of symlinks (and also when transname replacements occur).
  49  *
  50  * The new code replaces the old recursive symlink resolution with
  51  * an iterative one (in case of non-nested symlink chains).  It does
  52  * this with calls to <fs>_follow_link().
  53  * As a side effect, dir_namei(), _namei() and follow_link() are now
  54  * replaced with a single function lookup_dentry() that can handle all
  55  * the special cases of the former code.
  56  *
  57  * With the new dcache, the pathname is stored at each inode, at least as
  58  * long as the refcount of the inode is positive.  As a side effect, the
  59  * size of the dcache depends on the inode cache and thus is dynamic.
  60  *
  61  * [29-Apr-1998 C. Scott Ananian] Updated above description of symlink
  62  * resolution to correspond with current state of the code.
  63  *
  64  * Note that the symlink resolution is not *completely* iterative.
  65  * There is still a significant amount of tail- and mid- recursion in
  66  * the algorithm.  Also, note that <fs>_readlink() is not used in
  67  * lookup_dentry(): lookup_dentry() on the result of <fs>_readlink()
  68  * may return different results than <fs>_follow_link().  Many virtual
  69  * filesystems (including /proc) exhibit this behavior.
  70  */
  71
  72 /* [24-Feb-97 T. Schoebel-Theuer] Side effects caused by new implementation:
  73  * New symlink semantics: when open() is called with flags O_CREAT | O_EXCL
  74  * and the name already exists in form of a symlink, try to create the new
  75  * name indicated by the symlink. The old code always complained that the
  76  * name already exists, due to not following the symlink even if its target
  77  * is nonexistent.  The new semantics affects also mknod() and link() when
  78  * the name is a symlink pointing to a non-existent name.
  79  *
  80  * I don't know which semantics is the right one, since I have no access
  81  * to standards. But I found by trial that HP-UX 9.0 has the full "new"
  82  * semantics implemented, while SunOS 4.1.1 and Solaris (SunOS 5.4) have the
  83  * "old" one. Personally, I think the new semantics is much more logical.
  84  * Note that "ln old new" where "new" is a symlink pointing to a non-existing
  85  * file does succeed in both HP-UX and SunOs, but not in Solaris
  86  * and in the old Linux semantics.
  87  */
  88
  89 /* [16-Dec-97 Kevin Buhr] For security reasons, we change some symlink
  90  * semantics.  See the comments in "open_namei" and "do_link" below.
  91  *
  92  * [10-Sep-98 Alan Modra] Another symlink change.
  93  */
  94
  95 /* [Feb-Apr 2000 AV] Complete rewrite. Rules for symlinks:
  96  *      inside the path - always follow.
  97  *      in the last component in creation/removal/renaming - never follow.
  98  *      if LOOKUP_FOLLOW passed - follow.
  99  *      if the pathname has trailing slashes - follow.
 100  *      otherwise - don't follow.
 101  * (applied in that order).
 102  *
 103  * [Jun 2000 AV] Inconsistent behaviour of open() in case if flags==O_CREAT
 104  * restored for 2.4. This is the last surviving part of old 4.2BSD bug.
 105  * During the 2.4 we need to fix the userland stuff depending on it -
 106  * hopefully we will be able to get rid of that wart in 2.5. So far only
 107  * XEmacs seems to be relying on it...
 108  */
 109 /*
 110  * [Sep 2001 AV] Single-semaphore locking scheme (kudos to David Holland)
 111  * implemented.  Let's see if raised priority of ->s_vfs_rename_mutex gives
 112  * any extra contention...
 113  */
 114
 115 /* In order to reduce some races, while at the same time doing additional
 116  * checking and hopefully speeding things up, we copy filenames to the
 117  * kernel data space before using them..
 118  *
 119  * POSIX.1 2.4: an empty pathname is invalid (ENOENT).
 120  * PATH_MAX includes the nul terminator --RR.
 121  */
 122
 123 #define EMBEDDED_NAME_MAX       (PATH_MAX - offsetof(struct filename, iname))
 124
 125 struct filename *
 126 getname_flags(const char __user *filename, int flags, int *empty)
 127 {
 128         struct filename *result;
 129         char *kname;
 130         int len;
 131
 132         result = audit_reusename(filename);
 133         if (result)
 134                 return result;
 135
 136         result = __getname();
 137         if (unlikely(!result))
 138                 return ERR_PTR(-ENOMEM);
 139
 140         /*
 141          * First, try to embed the struct filename inside the names_cache
 142          * allocation
 143          */
 144         kname = (char *)result->iname;
 145         result->name = kname;
 146
 147         len = strncpy_from_user(kname, filename, EMBEDDED_NAME_MAX);
 148         if (unlikely(len < 0)) {
 149                 __putname(result);
 150                 return ERR_PTR(len);
 151         }
 152
 153         /*
 154          * Uh-oh. We have a name that's approaching PATH_MAX. Allocate a
 155          * separate struct filename so we can dedicate the entire
 156          * names_cache allocation for the pathname, and re-do the copy from
 157          * userland.
 158          */
 159         if (unlikely(len == EMBEDDED_NAME_MAX)) {
 160                 const size_t size = offsetof(struct filename, iname[1]);
 161                 kname = (char *)result;
 162
 163                 /*
 164                  * size is chosen that way we to guarantee that
 165                  * result->iname[0] is within the same object and that
 166                  * kname can't be equal to result->iname, no matter what.
 167                  */
 168                 result = kzalloc(size, GFP_KERNEL);
 169                 if (unlikely(!result)) {
 170                         __putname(kname);
 171                         return ERR_PTR(-ENOMEM);
 172                 }
 173                 result->name = kname;
 174                 len = strncpy_from_user(kname, filename, PATH_MAX);
 175                 if (unlikely(len < 0)) {
 176                         __putname(kname);
 177                         kfree(result);
 178                         return ERR_PTR(len);
 179                 }
 180                 if (unlikely(len == PATH_MAX)) {
 181                         __putname(kname);
 182                         kfree(result);
 183                         return ERR_PTR(-ENAMETOOLONG);
 184                 }
 185         }
 186
 187         result->refcnt = 1;
 188         /* The empty path is special. */
 189         if (unlikely(!len)) {
 190                 if (empty)
 191                         *empty = 1;
 192                 if (!(flags & LOOKUP_EMPTY)) {
 193                         putname(result);
 194                         return ERR_PTR(-ENOENT);
 195                 }
 196         }
 197
 198         result->uptr = filename;
 199         result->aname = NULL;
 200         audit_getname(result);
 201         return result;
 202 }
 203
 204 struct filename *
 205 getname(const char __user * filename)
 206 {
 207         return getname_flags(filename, 0, NULL);
 208 }
 209
 210 struct filename *
 211 getname_kernel(const char * filename)
 212 {
 213         struct filename *result;
 214         int len = strlen(filename) + 1;
 215
 216         result = __getname();
 217         if (unlikely(!result))
 218                 return ERR_PTR(-ENOMEM);
 219
 220         if (len <= EMBEDDED_NAME_MAX) {
 221                 result->name = (char *)result->iname;
 222         } else if (len <= PATH_MAX) {
 223                 struct filename *tmp;
 224
 225                 tmp = kmalloc(sizeof(*tmp), GFP_KERNEL);
 226                 if (unlikely(!tmp)) {
 227                         __putname(result);
 228                         return ERR_PTR(-ENOMEM);
 229                 }
 230                 tmp->name = (char *)result;
 231                 result = tmp;
 232         } else {
 233                 __putname(result);
 234                 return ERR_PTR(-ENAMETOOLONG);
 235         }
 236         memcpy((char *)result->name, filename, len);
 237         result->uptr = NULL;
 238         result->aname = NULL;
 239         result->refcnt = 1;
 240         audit_getname(result);
 241
 242         return result;
 243 }
 244
 245 void putname(struct filename *name)
 246 {
 247         BUG_ON(name->refcnt <= 0);
 248
 249         if (--name->refcnt > 0)
 250                 return;
 251
 252         if (name->name != name->iname) {
 253                 __putname(name->name);
 254                 kfree(name);
 255         } else
 256                 __putname(name);
 257 }
 258
 259 static int check_acl(struct inode *inode, int mask)
 260 {
 261 #ifdef CONFIG_FS_POSIX_ACL
 262         struct posix_acl *acl;
 263
 264         if (mask & MAY_NOT_BLOCK) {
 265                 acl = get_cached_acl_rcu(inode, ACL_TYPE_ACCESS);
 266                 if (!acl)
 267                         return -EAGAIN;
 268                 /* no ->get_acl() calls in RCU mode... */
 269                 if (is_uncached_acl(acl))
 270                         return -ECHILD;
 271                 return posix_acl_permission(inode, acl, mask & ~MAY_NOT_BLOCK);
 272         }
 273
 274         acl = get_acl(inode, ACL_TYPE_ACCESS);
 275         if (IS_ERR(acl))
 276                 return PTR_ERR(acl);
 277         if (acl) {
 278                 int error = posix_acl_permission(inode, acl, mask);
 279                 posix_acl_release(acl);
 280                 return error;
 281         }
 282 #endif
 283
 284         return -EAGAIN;
 285 }
 286
 287 /*
 288  * This does the basic permission checking
 289  */
 290 static int acl_permission_check(struct inode *inode, int mask)
 291 {
 292         unsigned int mode = inode->i_mode;
 293
 294         if (likely(uid_eq(current_fsuid(), inode->i_uid)))
 295                 mode >>= 6;
 296         else {
 297                 if (IS_POSIXACL(inode) && (mode & S_IRWXG)) {
 298                         int error = check_acl(inode, mask);
 299                         if (error != -EAGAIN)
 300                                 return error;
 301                 }
 302
 303                 if (in_group_p(inode->i_gid))
 304                         mode >>= 3;
 305         }
 306
 307         /*
 308          * If the DACs are ok we don't need any capability check.
 309          */
 310         if ((mask & ~mode & (MAY_READ | MAY_WRITE | MAY_EXEC)) == 0)
 311                 return 0;
 312         return -EACCES;
 313 }
 314
 315 /**
 316  * generic_permission -  check for access rights on a Posix-like filesystem
 317  * @inode:      inode to check access rights for
 318  * @mask:       right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC, ...)
 319  *
 320  * Used to check for read/write/execute permissions on a file.
 321  * We use "fsuid" for this, letting us set arbitrary permissions
 322  * for filesystem access without changing the "normal" uids which
 323  * are used for other things.
 324  *
 325  * generic_permission is rcu-walk aware. It returns -ECHILD in case an rcu-walk
 326  * request cannot be satisfied (eg. requires blocking or too much complexity).
 327  * It would then be called again in ref-walk mode.
 328  */
 329 int generic_permission(struct inode *inode, int mask)
 330 {
 331         int ret;
 332
 333         /*
 334          * Do the basic permission checks.
 335          */
 336         ret = acl_permission_check(inode, mask);
 337         if (ret != -EACCES)
 338                 return ret;
 339
 340         if (S_ISDIR(inode->i_mode)) {
 341                 /* DACs are overridable for directories */
 342                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
 343                         return 0;
 344                 if (!(mask & MAY_WRITE))
 345                         if (capable_wrt_inode_uidgid(inode,
 346                                                      CAP_DAC_READ_SEARCH))
 347                                 return 0;
 348                 return -EACCES;
 349         }
 350         /*
 351          * Read/write DACs are always overridable.
 352          * Executable DACs are overridable when there is
 353          * at least one exec bit set.
 354          */
 355         if (!(mask & MAY_EXEC) || (inode->i_mode & S_IXUGO))
 356                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_OVERRIDE))
 357                         return 0;
 358
 359         /*
 360          * Searching includes executable on directories, else just read.
 361          */
 362         mask &= MAY_READ | MAY_WRITE | MAY_EXEC;
 363         if (mask == MAY_READ)
 364                 if (capable_wrt_inode_uidgid(inode, CAP_DAC_READ_SEARCH))
 365                         return 0;
 366
 367         return -EACCES;
 368 }
 369 EXPORT_SYMBOL(generic_permission);
 370
 371 /*
 372  * We _really_ want to just do "generic_permission()" without
 373  * even looking at the inode->i_op values. So we keep a cache
 374  * flag in inode->i_opflags, that says "this has not special
 375  * permission function, use the fast case".
 376  */
 377 static inline int do_inode_permission(struct inode *inode, int mask)
 378 {
 379         if (unlikely(!(inode->i_opflags & IOP_FASTPERM))) {
 380                 if (likely(inode->i_op->permission))
 381                         return inode->i_op->permission(inode, mask);
 382
 383                 /* This gets set once for the inode lifetime */
 384                 spin_lock(&inode->i_lock);
 385                 inode->i_opflags |= IOP_FASTPERM;
 386                 spin_unlock(&inode->i_lock);
 387         }
 388         return generic_permission(inode, mask);
 389 }
 390
 391 /**
 392  * __inode_permission - Check for access rights to a given inode
 393  * @inode: Inode to check permission on
 394  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 395  *
 396  * Check for read/write/execute permissions on an inode.
 397  *
 398  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
 399  *
 400  * This does not check for a read-only file system.  You probably want
 401  * inode_permission().
 402  */
 403 int __inode_permission(struct inode *inode, int mask)
 404 {
 405         int retval;
 406
 407         if (unlikely(mask & MAY_WRITE)) {
 408                 /*
 409                  * Nobody gets write access to an immutable file.
 410                  */
 411                 if (IS_IMMUTABLE(inode))
 412                         return -EACCES;
 413         }
 414
 415         retval = do_inode_permission(inode, mask);
 416         if (retval)
 417                 return retval;
 418
 419         retval = devcgroup_inode_permission(inode, mask);
 420         if (retval)
 421                 return retval;
 422
 423         return security_inode_permission(inode, mask);
 424 }
 425 EXPORT_SYMBOL(__inode_permission);
 426
 427 /**
 428  * sb_permission - Check superblock-level permissions
 429  * @sb: Superblock of inode to check permission on
 430  * @inode: Inode to check permission on
 431  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 432  *
 433  * Separate out file-system wide checks from inode-specific permission checks.
 434  */
 435 static int sb_permission(struct super_block *sb, struct inode *inode, int mask)
 436 {
 437         if (unlikely(mask & MAY_WRITE)) {
 438                 umode_t mode = inode->i_mode;
 439
 440                 /* Nobody gets write access to a read-only fs. */
 441                 if ((sb->s_flags & MS_RDONLY) &&
 442                     (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
 443                         return -EROFS;
 444         }
 445         return 0;
 446 }
 447
 448 /**
 449  * inode_permission - Check for access rights to a given inode
 450  * @inode: Inode to check permission on
 451  * @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
 452  *
 453  * Check for read/write/execute permissions on an inode.  We use fs[ug]id for
 454  * this, letting us set arbitrary permissions for filesystem access without
 455  * changing the "normal" UIDs which are used for other things.
 456  *
 457  * When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
 458  */
 459 int inode_permission(struct inode *inode, int mask)
 460 {
 461         int retval;
 462
 463         retval = sb_permission(inode->i_sb, inode, mask);
 464         if (retval)
 465                 return retval;
 466         return __inode_permission(inode, mask);
 467 }
 468 EXPORT_SYMBOL(inode_permission);
 469
 470 /**
 471  * path_get - get a reference to a path
 472  * @path: path to get the reference to
 473  *
 474  * Given a path increment the reference count to the dentry and the vfsmount.
 475  */
 476 void path_get(const struct path *path)
 477 {
 478         mntget(path->mnt);
 479         dget(path->dentry);
 480 }
 481 EXPORT_SYMBOL(path_get);
 482
 483 /**
 484  * path_put - put a reference to a path
 485  * @path: path to put the reference to
 486  *
 487  * Given a path decrement the reference count to the dentry and the vfsmount.
 488  */
 489 void path_put(const struct path *path)
 490 {
 491         dput(path->dentry);
 492         mntput(path->mnt);
 493 }
 494 EXPORT_SYMBOL(path_put);
 495
 496 #define EMBEDDED_LEVELS 2
 497 struct nameidata {
 498         struct path     path;
 499         struct qstr     last;
 500         struct path     root;
 501         struct inode    *inode; /* path.dentry.d_inode */
 502         unsigned int    flags;
 503         unsigned        seq, m_seq;
 504         int             last_type;
 505         unsigned        depth;
 506         int             total_link_count;
 507         struct saved {
 508                 struct path link;
 509                 struct delayed_call done;
 510                 const char *name;
 511                 unsigned seq;
 512         } *stack, internal[EMBEDDED_LEVELS];
 513         struct filename *name;
 514         struct nameidata *saved;
 515         struct inode    *link_inode;
 516         unsigned        root_seq;
 517         int             dfd;
 518 };
 519
 520 static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
 521 {
 522         struct nameidata *old = current->nameidata;
 523         p->stack = p->internal;
 524         p->dfd = dfd;
 525         p->name = name;
 526         p->total_link_count = old ? old->total_link_count : 0;
 527         p->saved = old;
 528         current->nameidata = p;
 529 }
 530
 531 static void restore_nameidata(void)
 532 {
 533         struct nameidata *now = current->nameidata, *old = now->saved;
 534
 535         current->nameidata = old;
 536         if (old)
 537                 old->total_link_count = now->total_link_count;
 538         if (now->stack != now->internal)
 539                 kfree(now->stack);
 540 }
 541
 542 static int __nd_alloc_stack(struct nameidata *nd)
 543 {
 544         struct saved *p;
 545
 546         if (nd->flags & LOOKUP_RCU) {
 547                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
 548                                   GFP_ATOMIC);
 549                 if (unlikely(!p))
 550                         return -ECHILD;
 551         } else {
 552                 p= kmalloc(MAXSYMLINKS * sizeof(struct saved),
 553                                   GFP_KERNEL);
 554                 if (unlikely(!p))
 555                         return -ENOMEM;
 556         }
 557         memcpy(p, nd->internal, sizeof(nd->internal));
 558         nd->stack = p;
 559         return 0;
 560 }
 561
 562 /**
 563  * path_connected - Verify that a path->dentry is below path->mnt.mnt_root
 564  * @path: nameidate to verify
 565  *
 566  * Rename can sometimes move a file or directory outside of a bind
 567  * mount, path_connected allows those cases to be detected.
 568  */
 569 static bool path_connected(const struct path *path)
 570 {
 571         struct vfsmount *mnt = path->mnt;
 572
 573         /* Only bind mounts can have disconnected paths */
 574         if (mnt->mnt_root == mnt->mnt_sb->s_root)
 575                 return true;
 576
 577         return is_subdir(path->dentry, mnt->mnt_root);
 578 }
 579
 580 static inline int nd_alloc_stack(struct nameidata *nd)
 581 {
 582         if (likely(nd->depth != EMBEDDED_LEVELS))
 583                 return 0;
 584         if (likely(nd->stack != nd->internal))
 585                 return 0;
 586         return __nd_alloc_stack(nd);
 587 }
 588
 589 static void drop_links(struct nameidata *nd)
 590 {
 591         int i = nd->depth;
 592         while (i--) {
 593                 struct saved *last = nd->stack + i;
 594                 do_delayed_call(&last->done);
 595                 clear_delayed_call(&last->done);
 596         }
 597 }
 598
 599 static void terminate_walk(struct nameidata *nd)
 600 {
 601         drop_links(nd);
 602         if (!(nd->flags & LOOKUP_RCU)) {
 603                 int i;
 604                 path_put(&nd->path);
 605                 for (i = 0; i < nd->depth; i++)
 606                         path_put(&nd->stack[i].link);
 607                 if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 608                         path_put(&nd->root);
 609                         nd->root.mnt = NULL;
 610                 }
 611         } else {
 612                 nd->flags &= ~LOOKUP_RCU;
 613                 if (!(nd->flags & LOOKUP_ROOT))
 614                         nd->root.mnt = NULL;
 615                 rcu_read_unlock();
 616         }
 617         nd->depth = 0;
 618 }
 619
 620 /* path_put is needed afterwards regardless of success or failure */
 621 static bool legitimize_path(struct nameidata *nd,
 622                             struct path *path, unsigned seq)
 623 {
 624         int res = __legitimize_mnt(path->mnt, nd->m_seq);
 625         if (unlikely(res)) {
 626                 if (res > 0)
 627                         path->mnt = NULL;
 628                 path->dentry = NULL;
 629                 return false;
 630         }
 631         if (unlikely(!lockref_get_not_dead(&path->dentry->d_lockref))) {
 632                 path->dentry = NULL;
 633                 return false;
 634         }
 635         return !read_seqcount_retry(&path->dentry->d_seq, seq);
 636 }
 637
 638 static bool legitimize_links(struct nameidata *nd)
 639 {
 640         int i;
 641         for (i = 0; i < nd->depth; i++) {
 642                 struct saved *last = nd->stack + i;
 643                 if (unlikely(!legitimize_path(nd, &last->link, last->seq))) {
 644                         drop_links(nd);
 645                         nd->depth = i + 1;
 646                         return false;
 647                 }
 648         }
 649         return true;
 650 }
 651
 652 /*
 653  * Path walking has 2 modes, rcu-walk and ref-walk (see
 654  * Documentation/filesystems/path-lookup.txt).  In situations when we can't
 655  * continue in RCU mode, we attempt to drop out of rcu-walk mode and grab
 656  * normal reference counts on dentries and vfsmounts to transition to ref-walk
 657  * mode.  Refcounts are grabbed at the last known good point before rcu-walk
 658  * got stuck, so ref-walk may continue from there. If this is not successful
 659  * (eg. a seqcount has changed), then failure is returned and it's up to caller
 660  * to restart the path walk from the beginning in ref-walk mode.
 661  */
 662
 663 /**
 664  * unlazy_walk - try to switch to ref-walk mode.
 665  * @nd: nameidata pathwalk data
 666  * @dentry: child of nd->path.dentry or NULL
 667  * @seq: seq number to check dentry against
 668  * Returns: 0 on success, -ECHILD on failure
 669  *
 670  * unlazy_walk attempts to legitimize the current nd->path, nd->root and dentry
 671  * for ref-walk mode.  @dentry must be a path found by a do_lookup call on
 672  * @nd or NULL.  Must be called from rcu-walk context.
 673  * Nothing should touch nameidata between unlazy_walk() failure and
 674  * terminate_walk().
 675  */
 676 static int unlazy_walk(struct nameidata *nd, struct dentry *dentry, unsigned seq)
 677 {
 678         struct dentry *parent = nd->path.dentry;
 679
 680         BUG_ON(!(nd->flags & LOOKUP_RCU));
 681
 682         nd->flags &= ~LOOKUP_RCU;
 683         if (unlikely(!legitimize_links(nd)))
 684                 goto out2;
 685         if (unlikely(!legitimize_mnt(nd->path.mnt, nd->m_seq)))
 686                 goto out2;
 687         if (unlikely(!lockref_get_not_dead(&parent->d_lockref)))
 688                 goto out1;
 689
 690         /*
 691          * For a negative lookup, the lookup sequence point is the parents
 692          * sequence point, and it only needs to revalidate the parent dentry.
 693          *
 694          * For a positive lookup, we need to move both the parent and the
 695          * dentry from the RCU domain to be properly refcounted. And the
 696          * sequence number in the dentry validates *both* dentry counters,
 697          * since we checked the sequence number of the parent after we got
 698          * the child sequence number. So we know the parent must still
 699          * be valid if the child sequence number is still valid.
 700          */
 701         if (!dentry) {
 702                 if (read_seqcount_retry(&parent->d_seq, nd->seq))
 703                         goto out;
 704                 BUG_ON(nd->inode != parent->d_inode);
 705         } else {
 706                 if (!lockref_get_not_dead(&dentry->d_lockref))
 707                         goto out;
 708                 if (read_seqcount_retry(&dentry->d_seq, seq))
 709                         goto drop_dentry;
 710         }
 711
 712         /*
 713          * Sequence counts matched. Now make sure that the root is
 714          * still valid and get it if required.
 715          */
 716         if (nd->root.mnt && !(nd->flags & LOOKUP_ROOT)) {
 717                 if (unlikely(!legitimize_path(nd, &nd->root, nd->root_seq))) {
 718                         rcu_read_unlock();
 719                         dput(dentry);
 720                         return -ECHILD;
 721                 }
 722         }
 723
 724         rcu_read_unlock();
 725         return 0;
 726
 727 drop_dentry:
 728         rcu_read_unlock();
 729         dput(dentry);
 730         goto drop_root_mnt;
 731 out2:
 732         nd->path.mnt = NULL;
 733 out1:
 734         nd->path.dentry = NULL;
 735 out:
 736         rcu_read_unlock();
 737 drop_root_mnt:
 738         if (!(nd->flags & LOOKUP_ROOT))
 739                 nd->root.mnt = NULL;
 740         return -ECHILD;
 741 }
 742
 743 static int unlazy_link(struct nameidata *nd, struct path *link, unsigned seq)
 744 {
 745         if (unlikely(!legitimize_path(nd, link, seq))) {
 746                 drop_links(nd);
 747                 nd->depth = 0;
 748                 nd->flags &= ~LOOKUP_RCU;
 749                 nd->path.mnt = NULL;
 750                 nd->path.dentry = NULL;
 751                 if (!(nd->flags & LOOKUP_ROOT))
 752                         nd->root.mnt = NULL;
 753                 rcu_read_unlock();
 754         } else if (likely(unlazy_walk(nd, NULL, 0)) == 0) {
 755                 return 0;
 756         }
 757         path_put(link);
 758         return -ECHILD;
 759 }
 760
 761 static inline int d_revalidate(struct dentry *dentry, unsigned int flags)
 762 {
 763         return dentry->d_op->d_revalidate(dentry, flags);
 764 }
 765
 766 /**
 767  * complete_walk - successful completion of path walk
 768  * @nd:  pointer nameidata
 769  *
 770  * If we had been in RCU mode, drop out of it and legitimize nd->path.
 771  * Revalidate the final result, unless we'd already done that during
 772  * the path walk or the filesystem doesn't ask for it.  Return 0 on
 773  * success, -error on failure.  In case of failure caller does not
 774  * need to drop nd->path.
 775  */
 776 static int complete_walk(struct nameidata *nd)
 777 {
 778         struct dentry *dentry = nd->path.dentry;
 779         int status;
 780
 781         if (nd->flags & LOOKUP_RCU) {
 782                 if (!(nd->flags & LOOKUP_ROOT))
 783                         nd->root.mnt = NULL;
 784                 if (unlikely(unlazy_walk(nd, NULL, 0)))
 785                         return -ECHILD;
 786         }
 787
 788         if (likely(!(nd->flags & LOOKUP_JUMPED)))
 789                 return 0;
 790
 791         if (likely(!(dentry->d_flags & DCACHE_OP_WEAK_REVALIDATE)))
 792                 return 0;
 793
 794         status = dentry->d_op->d_weak_revalidate(dentry, nd->flags);
 795         if (status > 0)
 796                 return 0;
 797
 798         if (!status)
 799                 status = -ESTALE;
 800
 801         return status;
 802 }
 803
 804 static void set_root(struct nameidata *nd)
 805 {
 806         struct fs_struct *fs = current->fs;
 807
 808         if (nd->flags & LOOKUP_RCU) {
 809                 unsigned seq;
 810
 811                 do {
 812                         seq = read_seqcount_begin(&fs->seq);
 813                         nd->root = fs->root;
 814                         nd->root_seq = __read_seqcount_begin(&nd->root.dentry->d_seq);
 815                 } while (read_seqcount_retry(&fs->seq, seq));
 816         } else {
 817                 get_fs_root(fs, &nd->root);
 818         }
 819 }
 820
 821 static void path_put_conditional(struct path *path, struct nameidata *nd)
 822 {
 823         dput(path->dentry);
 824         if (path->mnt != nd->path.mnt)
 825                 mntput(path->mnt);
 826 }
 827
 828 static inline void path_to_nameidata(const struct path *path,
 829                                         struct nameidata *nd)
 830 {
 831         if (!(nd->flags & LOOKUP_RCU)) {
 832                 dput(nd->path.dentry);
 833                 if (nd->path.mnt != path->mnt)
 834                         mntput(nd->path.mnt);
 835         }
 836         nd->path.mnt = path->mnt;
 837         nd->path.dentry = path->dentry;
 838 }
 839
 840 static int nd_jump_root(struct nameidata *nd)
 841 {
 842         if (nd->flags & LOOKUP_RCU) {
 843                 struct dentry *d;
 844                 nd->path = nd->root;
 845                 d = nd->path.dentry;
 846                 nd->inode = d->d_inode;
 847                 nd->seq = nd->root_seq;
 848                 if (unlikely(read_seqcount_retry(&d->d_seq, nd->seq)))
 849                         return -ECHILD;
 850         } else {
 851                 path_put(&nd->path);
 852                 nd->path = nd->root;
 853                 path_get(&nd->path);
 854                 nd->inode = nd->path.dentry->d_inode;
 855         }
 856         nd->flags |= LOOKUP_JUMPED;
 857         return 0;
 858 }
 859
 860 /*
 861  * Helper to directly jump to a known parsed path from ->get_link,
 862  * caller must have taken a reference to path beforehand.
 863  */
 864 void nd_jump_link(struct path *path)
 865 {
 866         struct nameidata *nd = current->nameidata;
 867         path_put(&nd->path);
 868
 869         nd->path = *path;
 870         nd->inode = nd->path.dentry->d_inode;
 871         nd->flags |= LOOKUP_JUMPED;
 872 }
 873
 874 static inline void put_link(struct nameidata *nd)
 875 {
 876         struct saved *last = nd->stack + --nd->depth;
 877         do_delayed_call(&last->done);
 878         if (!(nd->flags & LOOKUP_RCU))
 879                 path_put(&last->link);
 880 }
 881
 882 int sysctl_protected_symlinks __read_mostly = 0;
 883 int sysctl_protected_hardlinks __read_mostly = 0;
 884
 885 /**
 886  * may_follow_link - Check symlink following for unsafe situations
 887  * @nd: nameidata pathwalk data
 888  *
 889  * In the case of the sysctl_protected_symlinks sysctl being enabled,
 890  * CAP_DAC_OVERRIDE needs to be specifically ignored if the symlink is
 891  * in a sticky world-writable directory. This is to protect privileged
 892  * processes from failing races against path names that may change out
 893  * from under them by way of other users creating malicious symlinks.
 894  * It will permit symlinks to be followed only when outside a sticky
 895  * world-writable directory, or when the uid of the symlink and follower
 896  * match, or when the directory owner matches the symlink's owner.
 897  *
 898  * Returns 0 if following the symlink is allowed, -ve on error.
 899  */
 900 static inline int may_follow_link(struct nameidata *nd)
 901 {
 902         const struct inode *inode;
 903         const struct inode *parent;
 904
 905         if (!sysctl_protected_symlinks)
 906                 return 0;
 907
 908         /* Allowed if owner and follower match. */
 909         inode = nd->link_inode;
 910         if (uid_eq(current_cred()->fsuid, inode->i_uid))
 911                 return 0;
 912
 913         /* Allowed if parent directory not sticky and world-writable. */
 914         parent = nd->inode;
 915         if ((parent->i_mode & (S_ISVTX|S_IWOTH)) != (S_ISVTX|S_IWOTH))
 916                 return 0;
 917
 918         /* Allowed if parent directory and link owner match. */
 919         if (uid_eq(parent->i_uid, inode->i_uid))
 920                 return 0;
 921
 922         if (nd->flags & LOOKUP_RCU)
 923                 return -ECHILD;
 924
 925         audit_log_link_denied("follow_link", &nd->stack[0].link);
 926         return -EACCES;
 927 }
 928
 929 /**
 930  * safe_hardlink_source - Check for safe hardlink conditions
 931  * @inode: the source inode to hardlink from
 932  *
 933  * Return false if at least one of the following conditions:
 934  *    - inode is not a regular file
 935  *    - inode is setuid
 936  *    - inode is setgid and group-exec
 937  *    - access failure for read and write
 938  *
 939  * Otherwise returns true.
 940  */
 941 static bool safe_hardlink_source(struct inode *inode)
 942 {
 943         umode_t mode = inode->i_mode;
 944
 945         /* Special files should not get pinned to the filesystem. */
 946         if (!S_ISREG(mode))
 947                 return false;
 948
 949         /* Setuid files should not get pinned to the filesystem. */
 950         if (mode & S_ISUID)
 951                 return false;
 952
 953         /* Executable setgid files should not get pinned to the filesystem. */
 954         if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP))
 955                 return false;
 956
 957         /* Hardlinking to unreadable or unwritable sources is dangerous. */
 958         if (inode_permission(inode, MAY_READ | MAY_WRITE))
 959                 return false;
 960
 961         return true;
 962 }
 963
 964 /**
 965  * may_linkat - Check permissions for creating a hardlink
 966  * @link: the source to hardlink from
 967  *
 968  * Block hardlink when all of:
 969  *  - sysctl_protected_hardlinks enabled
 970  *  - fsuid does not match inode
 971  *  - hardlink source is unsafe (see safe_hardlink_source() above)
 972  *  - not CAP_FOWNER in a namespace with the inode owner uid mapped
 973  *
 974  * Returns 0 if successful, -ve on error.
 975  */
 976 static int may_linkat(struct path *link)
 977 {
 978         struct inode *inode;
 979
 980         if (!sysctl_protected_hardlinks)
 981                 return 0;
 982
 983         inode = link->dentry->d_inode;
 984
 985         /* Source inode owner (or CAP_FOWNER) can hardlink all they like,
 986          * otherwise, it must be a safe source.
 987          */
 988         if (inode_owner_or_capable(inode) || safe_hardlink_source(inode))
 989                 return 0;
 990
 991         audit_log_link_denied("linkat", link);
 992         return -EPERM;
 993 }
 994
 995 static __always_inline
 996 const char *get_link(struct nameidata *nd)
 997 {
 998         struct saved *last = nd->stack + nd->depth - 1;
 999         struct dentry *dentry = last->link.dentry;
1000         struct inode *inode = nd->link_inode;
1001         int error;
1002         const char *res;
1003
1004         if (!(nd->flags & LOOKUP_RCU)) {
1005                 touch_atime(&last->link);
1006                 cond_resched();
1007         } else if (atime_needs_update(&last->link, inode)) {
1008                 if (unlikely(unlazy_walk(nd, NULL, 0)))
1009                         return ERR_PTR(-ECHILD);
1010                 touch_atime(&last->link);
1011         }
1012
1013         error = security_inode_follow_link(dentry, inode,
1014                                            nd->flags & LOOKUP_RCU);
1015         if (unlikely(error))
1016                 return ERR_PTR(error);
1017
1018         nd->last_type = LAST_BIND;
1019         res = inode->i_link;
1020         if (!res) {
1021                 const char * (*get)(struct dentry *, struct inode *,
1022                                 struct delayed_call *);
1023                 get = inode->i_op->get_link;
1024                 if (nd->flags & LOOKUP_RCU) {
1025                         res = get(NULL, inode, &last->done);
1026                         if (res == ERR_PTR(-ECHILD)) {
1027                                 if (unlikely(unlazy_walk(nd, NULL, 0)))
1028                                         return ERR_PTR(-ECHILD);
1029                                 res = get(dentry, inode, &last->done);
1030                         }
1031                 } else {
1032                         res = get(dentry, inode, &last->done);
1033                 }
1034                 if (IS_ERR_OR_NULL(res))
1035                         return res;
1036         }
1037         if (*res == '/') {
1038                 if (!nd->root.mnt)
1039                         set_root(nd);
1040                 if (unlikely(nd_jump_root(nd)))
1041                         return ERR_PTR(-ECHILD);
1042                 while (unlikely(*++res == '/'))
1043                         ;
1044         }
1045         if (!*res)
1046                 res = NULL;
1047         return res;
1048 }
1049
1050 /*
1051  * follow_up - Find the mountpoint of path's vfsmount
1052  *
1053  * Given a path, find the mountpoint of its source file system.
1054  * Replace @path with the path of the mountpoint in the parent mount.
1055  * Up is towards /.
1056  *
1057  * Return 1 if we went up a level and 0 if we were already at the
1058  * root.
1059  */
1060 int follow_up(struct path *path)
1061 {
1062         struct mount *mnt = real_mount(path->mnt);
1063         struct mount *parent;
1064         struct dentry *mountpoint;
1065
1066         read_seqlock_excl(&mount_lock);
1067         parent = mnt->mnt_parent;
1068         if (parent == mnt) {
1069                 read_sequnlock_excl(&mount_lock);
1070                 return 0;
1071         }
1072         mntget(&parent->mnt);
1073         mountpoint = dget(mnt->mnt_mountpoint);
1074         read_sequnlock_excl(&mount_lock);
1075         dput(path->dentry);
1076         path->dentry = mountpoint;
1077         mntput(path->mnt);
1078         path->mnt = &parent->mnt;
1079         return 1;
1080 }
1081 EXPORT_SYMBOL(follow_up);
1082
1083 /*
1084  * Perform an automount
1085  * - return -EISDIR to tell follow_managed() to stop and return the path we
1086  *   were called with.
1087  */
1088 static int follow_automount(struct path *path, struct nameidata *nd,
1089                             bool *need_mntput)
1090 {
1091         struct vfsmount *mnt;
1092         int err;
1093
1094         if (!path->dentry->d_op || !path->dentry->d_op->d_automount)
1095                 return -EREMOTE;
1096
1097         /* We don't want to mount if someone's just doing a stat -
1098          * unless they're stat'ing a directory and appended a '/' to
1099          * the name.
1100          *
1101          * We do, however, want to mount if someone wants to open or
1102          * create a file of any type under the mountpoint, wants to
1103          * traverse through the mountpoint or wants to open the
1104          * mounted directory.  Also, autofs may mark negative dentries
1105          * as being automount points.  These will need the attentions
1106          * of the daemon to instantiate them before they can be used.
1107          */
1108         if (!(nd->flags & (LOOKUP_PARENT | LOOKUP_DIRECTORY |
1109                            LOOKUP_OPEN | LOOKUP_CREATE | LOOKUP_AUTOMOUNT)) &&
1110             path->dentry->d_inode)
1111                 return -EISDIR;
1112
1113         nd->total_link_count++;
1114         if (nd->total_link_count >= 40)
1115                 return -ELOOP;
1116
1117         mnt = path->dentry->d_op->d_automount(path);
1118         if (IS_ERR(mnt)) {
1119                 /*
1120                  * The filesystem is allowed to return -EISDIR here to indicate
1121                  * it doesn't want to automount.  For instance, autofs would do
1122                  * this so that its userspace daemon can mount on this dentry.
1123                  *
1124                  * However, we can only permit this if it's a terminal point in
1125                  * the path being looked up; if it wasn't then the remainder of
1126                  * the path is inaccessible and we should say so.
1127                  */
1128                 if (PTR_ERR(mnt) == -EISDIR && (nd->flags & LOOKUP_PARENT))
1129                         return -EREMOTE;
1130                 return PTR_ERR(mnt);
1131         }
1132
1133         if (!mnt) /* mount collision */
1134                 return 0;
1135
1136         if (!*need_mntput) {
1137                 /* lock_mount() may release path->mnt on error */
1138                 mntget(path->mnt);
1139                 *need_mntput = true;
1140         }
1141         err = finish_automount(mnt, path);
1142
1143         switch (err) {
1144         case -EBUSY:
1145                 /* Someone else made a mount here whilst we were busy */
1146                 return 0;
1147         case 0:
1148                 path_put(path);
1149                 path->mnt = mnt;
1150                 path->dentry = dget(mnt->mnt_root);
1151                 return 0;
1152         default:
1153                 return err;
1154         }
1155
1156 }
1157
1158 /*
1159  * Handle a dentry that is managed in some way.
1160  * - Flagged for transit management (autofs)
1161  * - Flagged as mountpoint
1162  * - Flagged as automount point
1163  *
1164  * This may only be called in refwalk mode.
1165  *
1166  * Serialization is taken care of in namespace.c
1167  */
1168 static int follow_managed(struct path *path, struct nameidata *nd)
1169 {
1170         struct vfsmount *mnt = path->mnt; /* held by caller, must be left alone */
1171         unsigned managed;
1172         bool need_mntput = false;
1173         int ret = 0;
1174
1175         /* Given that we're not holding a lock here, we retain the value in a
1176          * local variable for each dentry as we look at it so that we don't see
1177          * the components of that value change under us */
1178         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1179                managed &= DCACHE_MANAGED_DENTRY,
1180                unlikely(managed != 0)) {
1181                 /* Allow the filesystem to manage the transit without i_mutex
1182                  * being held. */
1183                 if (managed & DCACHE_MANAGE_TRANSIT) {
1184                         BUG_ON(!path->dentry->d_op);
1185                         BUG_ON(!path->dentry->d_op->d_manage);
1186                         ret = path->dentry->d_op->d_manage(path->dentry, false);
1187                         if (ret < 0)
1188                                 break;
1189                 }
1190
1191                 /* Transit to a mounted filesystem. */
1192                 if (managed & DCACHE_MOUNTED) {
1193                         struct vfsmount *mounted = lookup_mnt(path);
1194                         if (mounted) {
1195                                 dput(path->dentry);
1196                                 if (need_mntput)
1197                                         mntput(path->mnt);
1198                                 path->mnt = mounted;
1199                                 path->dentry = dget(mounted->mnt_root);
1200                                 need_mntput = true;
1201                                 continue;
1202                         }
1203
1204                         /* Something is mounted on this dentry in another
1205                          * namespace and/or whatever was mounted there in this
1206                          * namespace got unmounted before lookup_mnt() could
1207                          * get it */
1208                 }
1209
1210                 /* Handle an automount point */
1211                 if (managed & DCACHE_NEED_AUTOMOUNT) {
1212                         ret = follow_automount(path, nd, &need_mntput);
1213                         if (ret < 0)
1214                                 break;
1215                         continue;
1216                 }
1217
1218                 /* We didn't change the current path point */
1219                 break;
1220         }
1221
1222         if (need_mntput && path->mnt == mnt)
1223                 mntput(path->mnt);
1224         if (ret == -EISDIR || !ret)
1225                 ret = 1;
1226         if (need_mntput)
1227                 nd->flags |= LOOKUP_JUMPED;
1228         if (unlikely(ret < 0))
1229                 path_put_conditional(path, nd);
1230         return ret;
1231 }
1232
1233 int follow_down_one(struct path *path)
1234 {
1235         struct vfsmount *mounted;
1236
1237         mounted = lookup_mnt(path);
1238         if (mounted) {
1239                 dput(path->dentry);
1240                 mntput(path->mnt);
1241                 path->mnt = mounted;
1242                 path->dentry = dget(mounted->mnt_root);
1243                 return 1;
1244         }
1245         return 0;
1246 }
1247 EXPORT_SYMBOL(follow_down_one);
1248
1249 static inline int managed_dentry_rcu(struct dentry *dentry)
1250 {
1251         return (dentry->d_flags & DCACHE_MANAGE_TRANSIT) ?
1252                 dentry->d_op->d_manage(dentry, true) : 0;
1253 }
1254
1255 /*
1256  * Try to skip to top of mountpoint pile in rcuwalk mode.  Fail if
1257  * we meet a managed dentry that would need blocking.
1258  */
1259 static bool __follow_mount_rcu(struct nameidata *nd, struct path *path,
1260                                struct inode **inode, unsigned *seqp)
1261 {
1262         for (;;) {
1263                 struct mount *mounted;
1264                 /*
1265                  * Don't forget we might have a non-mountpoint managed dentry
1266                  * that wants to block transit.
1267                  */
1268                 switch (managed_dentry_rcu(path->dentry)) {
1269                 case -ECHILD:
1270                 default:
1271                         return false;
1272                 case -EISDIR:
1273                         return true;
1274                 case 0:
1275                         break;
1276                 }
1277
1278                 if (!d_mountpoint(path->dentry))
1279                         return !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1280
1281                 mounted = __lookup_mnt(path->mnt, path->dentry);
1282                 if (!mounted)
1283                         break;
1284                 path->mnt = &mounted->mnt;
1285                 path->dentry = mounted->mnt.mnt_root;
1286                 nd->flags |= LOOKUP_JUMPED;
1287                 *seqp = read_seqcount_begin(&path->dentry->d_seq);
1288                 /*
1289                  * Update the inode too. We don't need to re-check the
1290                  * dentry sequence number here after this d_inode read,
1291                  * because a mount-point is always pinned.
1292                  */
1293                 *inode = path->dentry->d_inode;
1294         }
1295         return !read_seqretry(&mount_lock, nd->m_seq) &&
1296                 !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT);
1297 }
1298
1299 static int follow_dotdot_rcu(struct nameidata *nd)
1300 {
1301         struct inode *inode = nd->inode;
1302
1303         while (1) {
1304                 if (path_equal(&nd->path, &nd->root))
1305                         break;
1306                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1307                         struct dentry *old = nd->path.dentry;
1308                         struct dentry *parent = old->d_parent;
1309                         unsigned seq;
1310
1311                         inode = parent->d_inode;
1312                         seq = read_seqcount_begin(&parent->d_seq);
1313                         if (unlikely(read_seqcount_retry(&old->d_seq, nd->seq)))
1314                                 return -ECHILD;
1315                         nd->path.dentry = parent;
1316                         nd->seq = seq;
1317                         if (unlikely(!path_connected(&nd->path)))
1318                                 return -ENOENT;
1319                         break;
1320                 } else {
1321                         struct mount *mnt = real_mount(nd->path.mnt);
1322                         struct mount *mparent = mnt->mnt_parent;
1323                         struct dentry *mountpoint = mnt->mnt_mountpoint;
1324                         struct inode *inode2 = mountpoint->d_inode;
1325                         unsigned seq = read_seqcount_begin(&mountpoint->d_seq);
1326                         if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1327                                 return -ECHILD;
1328                         if (&mparent->mnt == nd->path.mnt)
1329                                 break;
1330                         /* we know that mountpoint was pinned */
1331                         nd->path.dentry = mountpoint;
1332                         nd->path.mnt = &mparent->mnt;
1333                         inode = inode2;
1334                         nd->seq = seq;
1335                 }
1336         }
1337         while (unlikely(d_mountpoint(nd->path.dentry))) {
1338                 struct mount *mounted;
1339                 mounted = __lookup_mnt(nd->path.mnt, nd->path.dentry);
1340                 if (unlikely(read_seqretry(&mount_lock, nd->m_seq)))
1341                         return -ECHILD;
1342                 if (!mounted)
1343                         break;
1344                 nd->path.mnt = &mounted->mnt;
1345                 nd->path.dentry = mounted->mnt.mnt_root;
1346                 inode = nd->path.dentry->d_inode;
1347                 nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
1348         }
1349         nd->inode = inode;
1350         return 0;
1351 }
1352
1353 /*
1354  * Follow down to the covering mount currently visible to userspace.  At each
1355  * point, the filesystem owning that dentry may be queried as to whether the
1356  * caller is permitted to proceed or not.
1357  */
1358 int follow_down(struct path *path)
1359 {
1360         unsigned managed;
1361         int ret;
1362
1363         while (managed = ACCESS_ONCE(path->dentry->d_flags),
1364                unlikely(managed & DCACHE_MANAGED_DENTRY)) {
1365                 /* Allow the filesystem to manage the transit without i_mutex
1366                  * being held.
1367                  *
1368                  * We indicate to the filesystem if someone is trying to mount
1369                  * something here.  This gives autofs the chance to deny anyone
1370                  * other than its daemon the right to mount on its
1371                  * superstructure.
1372                  *
1373                  * The filesystem may sleep at this point.
1374                  */
1375                 if (managed & DCACHE_MANAGE_TRANSIT) {
1376                         BUG_ON(!path->dentry->d_op);
1377                         BUG_ON(!path->dentry->d_op->d_manage);
1378                         ret = path->dentry->d_op->d_manage(
1379                                 path->dentry, false);
1380                         if (ret < 0)
1381                                 return ret == -EISDIR ? 0 : ret;
1382                 }
1383
1384                 /* Transit to a mounted filesystem. */
1385                 if (managed & DCACHE_MOUNTED) {
1386                         struct vfsmount *mounted = lookup_mnt(path);
1387                         if (!mounted)
1388                                 break;
1389                         dput(path->dentry);
1390                         mntput(path->mnt);
1391                         path->mnt = mounted;
1392                         path->dentry = dget(mounted->mnt_root);
1393                         continue;
1394                 }
1395
1396                 /* Don't handle automount points here */
1397                 break;
1398         }
1399         return 0;
1400 }
1401 EXPORT_SYMBOL(follow_down);
1402
1403 /*
1404  * Skip to top of mountpoint pile in refwalk mode for follow_dotdot()
1405  */
1406 static void follow_mount(struct path *path)
1407 {
1408         while (d_mountpoint(path->dentry)) {
1409                 struct vfsmount *mounted = lookup_mnt(path);
1410                 if (!mounted)
1411                         break;
1412                 dput(path->dentry);
1413                 mntput(path->mnt);
1414                 path->mnt = mounted;
1415                 path->dentry = dget(mounted->mnt_root);
1416         }
1417 }
1418
1419 static int path_parent_directory(struct path *path)
1420 {
1421         struct dentry *old = path->dentry;
1422         /* rare case of legitimate dget_parent()... */
1423         path->dentry = dget_parent(path->dentry);
1424         dput(old);
1425         if (unlikely(!path_connected(path)))
1426                 return -ENOENT;
1427         return 0;
1428 }
1429
1430 static int follow_dotdot(struct nameidata *nd)
1431 {
1432         while(1) {
1433                 if (nd->path.dentry == nd->root.dentry &&
1434                     nd->path.mnt == nd->root.mnt) {
1435                         break;
1436                 }
1437                 if (nd->path.dentry != nd->path.mnt->mnt_root) {
1438                         int ret = path_parent_directory(&nd->path);
1439                         if (ret)
1440                                 return ret;
1441                         break;
1442                 }
1443                 if (!follow_up(&nd->path))
1444                         break;
1445         }
1446         follow_mount(&nd->path);
1447         nd->inode = nd->path.dentry->d_inode;
1448         return 0;
1449 }
1450
1451 /*
1452  * This looks up the name in dcache and possibly revalidates the found dentry.
1453  * NULL is returned if the dentry does not exist in the cache.
1454  */
1455 static struct dentry *lookup_dcache(const struct qstr *name,
1456                                     struct dentry *dir,
1457                                     unsigned int flags)
1458 {
1459         struct dentry *dentry;
1460         int error;
1461
1462         dentry = d_lookup(dir, name);
1463         if (dentry) {
1464                 if (dentry->d_flags & DCACHE_OP_REVALIDATE) {
1465                         error = d_revalidate(dentry, flags);
1466                         if (unlikely(error <= 0)) {
1467                                 if (!error)
1468                                         d_invalidate(dentry);
1469                                 dput(dentry);
1470                                 return ERR_PTR(error);
1471                         }
1472                 }
1473         }
1474         return dentry;
1475 }
1476
1477 /*
1478  * Call i_op->lookup on the dentry.  The dentry must be negative and
1479  * unhashed.
1480  *
1481  * dir->d_inode->i_mutex must be held
1482  */
1483 static struct dentry *lookup_real(struct inode *dir, struct dentry *dentry,
1484                                   unsigned int flags)
1485 {
1486         struct dentry *old;
1487
1488         /* Don't create child dentry for a dead directory. */
1489         if (unlikely(IS_DEADDIR(dir))) {
1490                 dput(dentry);
1491                 return ERR_PTR(-ENOENT);
1492         }
1493
1494         old = dir->i_op->lookup(dir, dentry, flags);
1495         if (unlikely(old)) {
1496                 dput(dentry);
1497                 dentry = old;
1498         }
1499         return dentry;
1500 }
1501
1502 static struct dentry *__lookup_hash(const struct qstr *name,
1503                 struct dentry *base, unsigned int flags)
1504 {
1505         struct dentry *dentry = lookup_dcache(name, base, flags);
1506
1507         if (dentry)
1508                 return dentry;
1509
1510         dentry = d_alloc(base, name);
1511         if (unlikely(!dentry))
1512                 return ERR_PTR(-ENOMEM);
1513
1514         return lookup_real(base->d_inode, dentry, flags);
1515 }
1516
1517 static int lookup_fast(struct nameidata *nd,
1518                        struct path *path, struct inode **inode,
1519                        unsigned *seqp)
1520 {
1521         struct vfsmount *mnt = nd->path.mnt;
1522         struct dentry *dentry, *parent = nd->path.dentry;
1523         int status = 1;
1524         int err;
1525
1526         /*
1527          * Rename seqlock is not required here because in the off chance
1528          * of a false negative due to a concurrent rename, the caller is
1529          * going to fall back to non-racy lookup.
1530          */
1531         if (nd->flags & LOOKUP_RCU) {
1532                 unsigned seq;
1533                 bool negative;
1534                 dentry = __d_lookup_rcu(parent, &nd->last, &seq);
1535                 if (unlikely(!dentry)) {
1536                         if (unlazy_walk(nd, NULL, 0))
1537                                 return -ECHILD;
1538                         return 0;
1539                 }
1540
1541                 /*
1542                  * This sequence count validates that the inode matches
1543                  * the dentry name information from lookup.
1544                  */
1545                 *inode = d_backing_inode(dentry);
1546                 negative = d_is_negative(dentry);
1547                 if (unlikely(read_seqcount_retry(&dentry->d_seq, seq)))
1548                         return -ECHILD;
1549
1550                 /*
1551                  * This sequence count validates that the parent had no
1552                  * changes while we did the lookup of the dentry above.
1553                  *
1554                  * The memory barrier in read_seqcount_begin of child is
1555                  *  enough, we can use __read_seqcount_retry here.
1556                  */
1557                 if (unlikely(__read_seqcount_retry(&parent->d_seq, nd->seq)))
1558                         return -ECHILD;
1559
1560                 *seqp = seq;
1561                 if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
1562                         status = d_revalidate(dentry, nd->flags);
1563                 if (unlikely(status <= 0)) {
1564                         if (unlazy_walk(nd, dentry, seq))
1565                                 return -ECHILD;
1566                         if (status == -ECHILD)
1567                                 status = d_revalidate(dentry, nd->flags);
1568                 } else {
1569                         /*
1570                          * Note: do negative dentry check after revalidation in
1571                          * case that drops it.
1572                          */
1573                         if (unlikely(negative))
1574                                 return -ENOENT;
1575                         path->mnt = mnt;
1576                         path->dentry = dentry;
1577                         if (likely(__follow_mount_rcu(nd, path, inode, seqp)))
1578                                 return 1;
1579                         if (unlazy_walk(nd, dentry, seq))
1580                                 return -ECHILD;
1581                 }
1582         } else {
1583                 dentry = __d_lookup(parent, &nd->last);
1584                 if (unlikely(!dentry))
1585                         return 0;
1586                 if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
1587                         status = d_revalidate(dentry, nd->flags);
1588         }
1589         if (unlikely(status <= 0)) {
1590                 if (!status)
1591                         d_invalidate(dentry);
1592                 dput(dentry);
1593                 return status;
1594         }
1595         if (unlikely(d_is_negative(dentry))) {
1596                 dput(dentry);
1597                 return -ENOENT;
1598         }
1599
1600         path->mnt = mnt;
1601         path->dentry = dentry;
1602         err = follow_managed(path, nd);
1603         if (likely(err > 0))
1604                 *inode = d_backing_inode(path->dentry);
1605         return err;
1606 }
1607
1608 /* Fast lookup failed, do it the slow way */
1609 static struct dentry *lookup_slow(const struct qstr *name,
1610                                   struct dentry *dir,
1611                                   unsigned int flags)
1612 {
1613         struct dentry *dentry = ERR_PTR(-ENOENT), *old;
1614         struct inode *inode = dir->d_inode;
1615         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
1616
1617         inode_lock_shared(inode);
1618         /* Don't go there if it's already dead */
1619         if (unlikely(IS_DEADDIR(inode)))
1620                 goto out;
1621 again:
1622         dentry = d_alloc_parallel(dir, name, &wq);
1623         if (IS_ERR(dentry))
1624                 goto out;
1625         if (unlikely(!d_in_lookup(dentry))) {
1626                 if ((dentry->d_flags & DCACHE_OP_REVALIDATE) &&
1627                     !(flags & LOOKUP_NO_REVAL)) {
1628                         int error = d_revalidate(dentry, flags);
1629                         if (unlikely(error <= 0)) {
1630                                 if (!error) {
1631                                         d_invalidate(dentry);
1632                                         dput(dentry);
1633                                         goto again;
1634                                 }
1635                                 dput(dentry);
1636                                 dentry = ERR_PTR(error);
1637                         }
1638                 }
1639         } else {
1640                 old = inode->i_op->lookup(inode, dentry, flags);
1641                 d_lookup_done(dentry);
1642                 if (unlikely(old)) {
1643                         dput(dentry);
1644                         dentry = old;
1645                 }
1646         }
1647 out:
1648         inode_unlock_shared(inode);
1649         return dentry;
1650 }
1651
1652 static inline int may_lookup(struct nameidata *nd)
1653 {
1654         if (nd->flags & LOOKUP_RCU) {
1655                 int err = inode_permission(nd->inode, MAY_EXEC|MAY_NOT_BLOCK);
1656                 if (err != -ECHILD)
1657                         return err;
1658                 if (unlazy_walk(nd, NULL, 0))
1659                         return -ECHILD;
1660         }
1661         return inode_permission(nd->inode, MAY_EXEC);
1662 }
1663
1664 static inline int handle_dots(struct nameidata *nd, int type)
1665 {
1666         if (type == LAST_DOTDOT) {
1667                 if (!nd->root.mnt)
1668                         set_root(nd);
1669                 if (nd->flags & LOOKUP_RCU) {
1670                         return follow_dotdot_rcu(nd);
1671                 } else
1672                         return follow_dotdot(nd);
1673         }
1674         return 0;
1675 }
1676
1677 static int pick_link(struct nameidata *nd, struct path *link,
1678                      struct inode *inode, unsigned seq)
1679 {
1680         int error;
1681         struct saved *last;
1682         if (unlikely(nd->total_link_count++ >= MAXSYMLINKS)) {
1683                 path_to_nameidata(link, nd);
1684                 return -ELOOP;
1685         }
1686         if (!(nd->flags & LOOKUP_RCU)) {
1687                 if (link->mnt == nd->path.mnt)
1688                         mntget(link->mnt);
1689         }
1690         error = nd_alloc_stack(nd);
1691         if (unlikely(error)) {
1692                 if (error == -ECHILD) {
1693                         if (unlikely(unlazy_link(nd, link, seq)))
1694                                 return -ECHILD;
1695                         error = nd_alloc_stack(nd);
1696                 }
1697                 if (error) {
1698                         path_put(link);
1699                         return error;
1700                 }
1701         }
1702
1703         last = nd->stack + nd->depth++;
1704         last->link = *link;
1705         clear_delayed_call(&last->done);
1706         nd->link_inode = inode;
1707         last->seq = seq;
1708         return 1;
1709 }
1710
1711 /*
1712  * Do we need to follow links? We _really_ want to be able
1713  * to do this check without having to look at inode->i_op,
1714  * so we keep a cache of "no, this doesn't need follow_link"
1715  * for the common case.
1716  */
1717 static inline int should_follow_link(struct nameidata *nd, struct path *link,
1718                                      int follow,
1719                                      struct inode *inode, unsigned seq)
1720 {
1721         if (likely(!d_is_symlink(link->dentry)))
1722                 return 0;
1723         if (!follow)
1724                 return 0;
1725         /* make sure that d_is_symlink above matches inode */
1726         if (nd->flags & LOOKUP_RCU) {
1727                 if (read_seqcount_retry(&link->dentry->d_seq, seq))
1728                         return -ECHILD;
1729         }
1730         return pick_link(nd, link, inode, seq);
1731 }
1732
1733 enum {WALK_GET = 1, WALK_PUT = 2};
1734
1735 static int walk_component(struct nameidata *nd, int flags)
1736 {
1737         struct path path;
1738         struct inode *inode;
1739         unsigned seq;
1740         int err;
1741         /*
1742          * "." and ".." are special - ".." especially so because it has
1743          * to be able to know about the current root directory and
1744          * parent relationships.
1745          */
1746         if (unlikely(nd->last_type != LAST_NORM)) {
1747                 err = handle_dots(nd, nd->last_type);
1748                 if (flags & WALK_PUT)
1749                         put_link(nd);
1750                 return err;
1751         }
1752         err = lookup_fast(nd, &path, &inode, &seq);
1753         if (unlikely(err <= 0)) {
1754                 if (err < 0)
1755                         return err;
1756                 path.dentry = lookup_slow(&nd->last, nd->path.dentry,
1757                                           nd->flags);
1758                 if (IS_ERR(path.dentry))
1759                         return PTR_ERR(path.dentry);
1760
1761                 path.mnt = nd->path.mnt;
1762                 err = follow_managed(&path, nd);
1763                 if (unlikely(err < 0))
1764                         return err;
1765
1766                 if (unlikely(d_is_negative(path.dentry))) {
1767                         path_to_nameidata(&path, nd);
1768                         return -ENOENT;
1769                 }
1770
1771                 seq = 0;        /* we are already out of RCU mode */
1772                 inode = d_backing_inode(path.dentry);
1773         }
1774
1775         if (flags & WALK_PUT)
1776                 put_link(nd);
1777         err = should_follow_link(nd, &path, flags & WALK_GET, inode, seq);
1778         if (unlikely(err))
1779                 return err;
1780         path_to_nameidata(&path, nd);
1781         nd->inode = inode;
1782         nd->seq = seq;
1783         return 0;
1784 }
1785
1786 /*
1787  * We can do the critical dentry name comparison and hashing
1788  * operations one word at a time, but we are limited to:
1789  *
1790  * - Architectures with fast unaligned word accesses. We could
1791  *   do a "get_unaligned()" if this helps and is sufficiently
1792  *   fast.
1793  *
1794  * - non-CONFIG_DEBUG_PAGEALLOC configurations (so that we
1795  *   do not trap on the (extremely unlikely) case of a page
1796  *   crossing operation.
1797  *
1798  * - Furthermore, we need an efficient 64-bit compile for the
1799  *   64-bit case in order to generate the "number of bytes in
1800  *   the final mask". Again, that could be replaced with a
1801  *   efficient population count instruction or similar.
1802  */
1803 #ifdef CONFIG_DCACHE_WORD_ACCESS
1804
1805 #include <asm/word-at-a-time.h>
1806
1807 #ifdef HASH_MIX
1808
1809 /* Architecture provides HASH_MIX and fold_hash() in <asm/hash.h> */
1810
1811 #elif defined(CONFIG_64BIT)
1812 /*
1813  * Register pressure in the mixing function is an issue, particularly
1814  * on 32-bit x86, but almost any function requires one state value and
1815  * one temporary.  Instead, use a function designed for two state values
1816  * and no temporaries.
1817  *
1818  * This function cannot create a collision in only two iterations, so
1819  * we have two iterations to achieve avalanche.  In those two iterations,
1820  * we have six layers of mixing, which is enough to spread one bit's
1821  * influence out to 2^6 = 64 state bits.
1822  *
1823  * Rotate constants are scored by considering either 64 one-bit input
1824  * deltas or 64*63/2 = 2016 two-bit input deltas, and finding the
1825  * probability of that delta causing a change to each of the 128 output
1826  * bits, using a sample of random initial states.
1827  *
1828  * The Shannon entropy of the computed probabilities is then summed
1829  * to produce a score.  Ideally, any input change has a 50% chance of
1830  * toggling any given output bit.
1831  *
1832  * Mixing scores (in bits) for (12,45):
1833  * Input delta: 1-bit      2-bit
1834  * 1 round:     713.3    42542.6
1835  * 2 rounds:   2753.7   140389.8
1836  * 3 rounds:   5954.1   233458.2
1837  * 4 rounds:   7862.6   256672.2
1838  * Perfect:    8192     258048
1839  *            (64*128) (64*63/2 * 128)
1840  */
1841 #define HASH_MIX(x, y, a)       \
1842         (       x ^= (a),       \
1843         y ^= x, x = rol64(x,12),\
1844         x += y, y = rol64(y,45),\
1845         y *= 9                  )
1846
1847 /*
1848  * Fold two longs into one 32-bit hash value.  This must be fast, but
1849  * latency isn't quite as critical, as there is a fair bit of additional
1850  * work done before the hash value is used.
1851  */
1852 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1853 {
1854         y ^= x * GOLDEN_RATIO_64;
1855         y *= GOLDEN_RATIO_64;
1856         return y >> 32;
1857 }
1858
1859 #else   /* 32-bit case */
1860
1861 /*
1862  * Mixing scores (in bits) for (7,20):
1863  * Input delta: 1-bit      2-bit
1864  * 1 round:     330.3     9201.6
1865  * 2 rounds:   1246.4    25475.4
1866  * 3 rounds:   1907.1    31295.1
1867  * 4 rounds:   2042.3    31718.6
1868  * Perfect:    2048      31744
1869  *            (32*64)   (32*31/2 * 64)
1870  */
1871 #define HASH_MIX(x, y, a)       \
1872         (       x ^= (a),       \
1873         y ^= x, x = rol32(x, 7),\
1874         x += y, y = rol32(y,20),\
1875         y *= 9                  )
1876
1877 static inline unsigned int fold_hash(unsigned long x, unsigned long y)
1878 {
1879         /* Use arch-optimized multiply if one exists */
1880         return __hash_32(y ^ __hash_32(x));
1881 }
1882
1883 #endif
1884
1885 /*
1886  * Return the hash of a string of known length.  This is carfully
1887  * designed to match hash_name(), which is the more critical function.
1888  * In particular, we must end by hashing a final word containing 0..7
1889  * payload bytes, to match the way that hash_name() iterates until it
1890  * finds the delimiter after the name.
1891  */
1892 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1893 {
1894         unsigned long a, x = 0, y = (unsigned long)salt;
1895
1896         for (;;) {
1897                 if (!len)
1898                         goto done;
1899                 a = load_unaligned_zeropad(name);
1900                 if (len < sizeof(unsigned long))
1901                         break;
1902                 HASH_MIX(x, y, a);
1903                 name += sizeof(unsigned long);
1904                 len -= sizeof(unsigned long);
1905         }
1906         x ^= a & bytemask_from_count(len);
1907 done:
1908         return fold_hash(x, y);
1909 }
1910 EXPORT_SYMBOL(full_name_hash);
1911
1912 /* Return the "hash_len" (hash and length) of a null-terminated string */
1913 u64 hashlen_string(const void *salt, const char *name)
1914 {
1915         unsigned long a = 0, x = 0, y = (unsigned long)salt;
1916         unsigned long adata, mask, len;
1917         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1918
1919         len = 0;
1920         goto inside;
1921
1922         do {
1923                 HASH_MIX(x, y, a);
1924                 len += sizeof(unsigned long);
1925 inside:
1926                 a = load_unaligned_zeropad(name+len);
1927         } while (!has_zero(a, &adata, &constants));
1928
1929         adata = prep_zero_mask(a, adata, &constants);
1930         mask = create_zero_mask(adata);
1931         x ^= a & zero_bytemask(mask);
1932
1933         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1934 }
1935 EXPORT_SYMBOL(hashlen_string);
1936
1937 /*
1938  * Calculate the length and hash of the path component, and
1939  * return the "hash_len" as the result.
1940  */
1941 static inline u64 hash_name(const void *salt, const char *name)
1942 {
1943         unsigned long a = 0, b, x = 0, y = (unsigned long)salt;
1944         unsigned long adata, bdata, mask, len;
1945         const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
1946
1947         len = 0;
1948         goto inside;
1949
1950         do {
1951                 HASH_MIX(x, y, a);
1952                 len += sizeof(unsigned long);
1953 inside:
1954                 a = load_unaligned_zeropad(name+len);
1955                 b = a ^ REPEAT_BYTE('/');
1956         } while (!(has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)));
1957
1958         adata = prep_zero_mask(a, adata, &constants);
1959         bdata = prep_zero_mask(b, bdata, &constants);
1960         mask = create_zero_mask(adata | bdata);
1961         x ^= a & zero_bytemask(mask);
1962
1963         return hashlen_create(fold_hash(x, y), len + find_zero(mask));
1964 }
1965
1966 #else   /* !CONFIG_DCACHE_WORD_ACCESS: Slow, byte-at-a-time version */
1967
1968 /* Return the hash of a string of known length */
1969 unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
1970 {
1971         unsigned long hash = init_name_hash(salt);
1972         while (len--)
1973                 hash = partial_name_hash((unsigned char)*name++, hash);
1974         return end_name_hash(hash);
1975 }
1976 EXPORT_SYMBOL(full_name_hash);
1977
1978 /* Return the "hash_len" (hash and length) of a null-terminated string */
1979 u64 hashlen_string(const void *salt, const char *name)
1980 {
1981         unsigned long hash = init_name_hash(salt);
1982         unsigned long len = 0, c;
1983
1984         c = (unsigned char)*name;
1985         while (c) {
1986                 len++;
1987                 hash = partial_name_hash(c, hash);
1988                 c = (unsigned char)name[len];
1989         }
1990         return hashlen_create(end_name_hash(hash), len);
1991 }
1992 EXPORT_SYMBOL(hashlen_string);
1993
1994 /*
1995  * We know there's a real path component here of at least
1996  * one character.
1997  */
1998 static inline u64 hash_name(const void *salt, const char *name)
1999 {
2000         unsigned long hash = init_name_hash(salt);
2001         unsigned long len = 0, c;
2002
2003         c = (unsigned char)*name;
2004         do {
2005                 len++;
2006                 hash = partial_name_hash(c, hash);
2007                 c = (unsigned char)name[len];
2008         } while (c && c != '/');
2009         return hashlen_create(end_name_hash(hash), len);
2010 }
2011
2012 #endif
2013
2014 /*
2015  * Name resolution.
2016  * This is the basic name resolution function, turning a pathname into
2017  * the final dentry. We expect 'base' to be positive and a directory.
2018  *
2019  * Returns 0 and nd will have valid dentry and mnt on success.
2020  * Returns error and drops reference to input namei data on failure.
2021  */
2022 static int link_path_walk(const char *name, struct nameidata *nd)
2023 {
2024         int err;
2025
2026         while (*name=='/')
2027                 name++;
2028         if (!*name)
2029                 return 0;
2030
2031         /* At this point we know we have a real path component. */
2032         for(;;) {
2033                 u64 hash_len;
2034                 int type;
2035
2036                 err = may_lookup(nd);
2037                 if (err)
2038                         return err;
2039
2040                 hash_len = hash_name(nd->path.dentry, name);
2041
2042                 type = LAST_NORM;
2043                 if (name[0] == '.') switch (hashlen_len(hash_len)) {
2044                         case 2:
2045                                 if (name[1] == '.') {
2046                                         type = LAST_DOTDOT;
2047                                         nd->flags |= LOOKUP_JUMPED;
2048                                 }
2049                                 break;
2050                         case 1:
2051                                 type = LAST_DOT;
2052                 }
2053                 if (likely(type == LAST_NORM)) {
2054                         struct dentry *parent = nd->path.dentry;
2055                         nd->flags &= ~LOOKUP_JUMPED;
2056                         if (unlikely(parent->d_flags & DCACHE_OP_HASH)) {
2057                                 struct qstr this = { { .hash_len = hash_len }, .name = name };
2058                                 err = parent->d_op->d_hash(parent, &this);
2059                                 if (err < 0)
2060                                         return err;
2061                                 hash_len = this.hash_len;
2062                                 name = this.name;
2063                         }
2064                 }
2065
2066                 nd->last.hash_len = hash_len;
2067                 nd->last.name = name;
2068                 nd->last_type = type;
2069
2070                 name += hashlen_len(hash_len);
2071                 if (!*name)
2072                         goto OK;
2073                 /*
2074                  * If it wasn't NUL, we know it was '/'. Skip that
2075                  * slash, and continue until no more slashes.
2076                  */
2077                 do {
2078                         name++;
2079                 } while (unlikely(*name == '/'));
2080                 if (unlikely(!*name)) {
2081 OK:
2082                         /* pathname body, done */
2083                         if (!nd->depth)
2084                                 return 0;
2085                         name = nd->stack[nd->depth - 1].name;
2086                         /* trailing symlink, done */
2087                         if (!name)
2088                                 return 0;
2089                         /* last component of nested symlink */
2090                         err = walk_component(nd, WALK_GET | WALK_PUT);
2091                 } else {
2092                         err = walk_component(nd, WALK_GET);
2093                 }
2094                 if (err < 0)
2095                         return err;
2096
2097                 if (err) {
2098                         const char *s = get_link(nd);
2099
2100                         if (IS_ERR(s))
2101                                 return PTR_ERR(s);
2102                         err = 0;
2103                         if (unlikely(!s)) {
2104                                 /* jumped */
2105                                 put_link(nd);
2106                         } else {
2107                                 nd->stack[nd->depth - 1].name = name;
2108                                 name = s;
2109                                 continue;
2110                         }
2111                 }
2112                 if (unlikely(!d_can_lookup(nd->path.dentry))) {
2113                         if (nd->flags & LOOKUP_RCU) {
2114                                 if (unlazy_walk(nd, NULL, 0))
2115                                         return -ECHILD;
2116                         }
2117                         return -ENOTDIR;
2118                 }
2119         }
2120 }
2121
2122 static const char *path_init(struct nameidata *nd, unsigned flags)
2123 {
2124         int retval = 0;
2125         const char *s = nd->name->name;
2126
2127         nd->last_type = LAST_ROOT; /* if there are only slashes... */
2128         nd->flags = flags | LOOKUP_JUMPED | LOOKUP_PARENT;
2129         nd->depth = 0;
2130         if (flags & LOOKUP_ROOT) {
2131                 struct dentry *root = nd->root.dentry;
2132                 struct inode *inode = root->d_inode;
2133                 if (*s) {
2134                         if (!d_can_lookup(root))
2135                                 return ERR_PTR(-ENOTDIR);
2136                         retval = inode_permission(inode, MAY_EXEC);
2137                         if (retval)
2138                                 return ERR_PTR(retval);
2139                 }
2140                 nd->path = nd->root;
2141                 nd->inode = inode;
2142                 if (flags & LOOKUP_RCU) {
2143                         rcu_read_lock();
2144                         nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2145                         nd->root_seq = nd->seq;
2146                         nd->m_seq = read_seqbegin(&mount_lock);
2147                 } else {
2148                         path_get(&nd->path);
2149                 }
2150                 return s;
2151         }
2152
2153         nd->root.mnt = NULL;
2154         nd->path.mnt = NULL;
2155         nd->path.dentry = NULL;
2156
2157         nd->m_seq = read_seqbegin(&mount_lock);
2158         if (*s == '/') {
2159                 if (flags & LOOKUP_RCU)
2160                         rcu_read_lock();
2161                 set_root(nd);
2162                 if (likely(!nd_jump_root(nd)))
2163                         return s;
2164                 nd->root.mnt = NULL;
2165                 rcu_read_unlock();
2166                 return ERR_PTR(-ECHILD);
2167         } else if (nd->dfd == AT_FDCWD) {
2168                 if (flags & LOOKUP_RCU) {
2169                         struct fs_struct *fs = current->fs;
2170                         unsigned seq;
2171
2172                         rcu_read_lock();
2173
2174                         do {
2175                                 seq = read_seqcount_begin(&fs->seq);
2176                                 nd->path = fs->pwd;
2177                                 nd->inode = nd->path.dentry->d_inode;
2178                                 nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
2179                         } while (read_seqcount_retry(&fs->seq, seq));
2180                 } else {
2181                         get_fs_pwd(current->fs, &nd->path);
2182                         nd->inode = nd->path.dentry->d_inode;
2183                 }
2184                 return s;
2185         } else {
2186                 /* Caller must check execute permissions on the starting path component */
2187                 struct fd f = fdget_raw(nd->dfd);
2188                 struct dentry *dentry;
2189
2190                 if (!f.file)
2191                         return ERR_PTR(-EBADF);
2192
2193                 dentry = f.file->f_path.dentry;
2194
2195                 if (*s) {
2196                         if (!d_can_lookup(dentry)) {
2197                                 fdput(f);
2198                                 return ERR_PTR(-ENOTDIR);
2199                         }
2200                 }
2201
2202                 nd->path = f.file->f_path;
2203                 if (flags & LOOKUP_RCU) {
2204                         rcu_read_lock();
2205                         nd->inode = nd->path.dentry->d_inode;
2206                         nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
2207                 } else {
2208                         path_get(&nd->path);
2209                         nd->inode = nd->path.dentry->d_inode;
2210                 }
2211                 fdput(f);
2212                 return s;
2213         }
2214 }
2215
2216 static const char *trailing_symlink(struct nameidata *nd)
2217 {
2218         const char *s;
2219         int error = may_follow_link(nd);
2220         if (unlikely(error))
2221                 return ERR_PTR(error);
2222         nd->flags |= LOOKUP_PARENT;
2223         nd->stack[0].name = NULL;
2224         s = get_link(nd);
2225         return s ? s : "";
2226 }
2227
2228 static inline int lookup_last(struct nameidata *nd)
2229 {
2230         if (nd->last_type == LAST_NORM && nd->last.name[nd->last.len])
2231                 nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
2232
2233         nd->flags &= ~LOOKUP_PARENT;
2234         return walk_component(nd,
2235                         nd->flags & LOOKUP_FOLLOW
2236                                 ? nd->depth
2237                                         ? WALK_PUT | WALK_GET
2238                                         : WALK_GET
2239                                 : 0);
2240 }
2241
2242 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2243 static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
2244 {
2245         const char *s = path_init(nd, flags);
2246         int err;
2247
2248         if (IS_ERR(s))
2249                 return PTR_ERR(s);
2250         while (!(err = link_path_walk(s, nd))
2251                 && ((err = lookup_last(nd)) > 0)) {
2252                 s = trailing_symlink(nd);
2253                 if (IS_ERR(s)) {
2254                         err = PTR_ERR(s);
2255                         break;
2256                 }
2257         }
2258         if (!err)
2259                 err = complete_walk(nd);
2260
2261         if (!err && nd->flags & LOOKUP_DIRECTORY)
2262                 if (!d_can_lookup(nd->path.dentry))
2263                         err = -ENOTDIR;
2264         if (!err) {
2265                 *path = nd->path;
2266                 nd->path.mnt = NULL;
2267                 nd->path.dentry = NULL;
2268         }
2269         terminate_walk(nd);
2270         return err;
2271 }
2272
2273 static int filename_lookup(int dfd, struct filename *name, unsigned flags,
2274                            struct path *path, struct path *root)
2275 {
2276         int retval;
2277         struct nameidata nd;
2278         if (IS_ERR(name))
2279                 return PTR_ERR(name);
2280         if (unlikely(root)) {
2281                 nd.root = *root;
2282                 flags |= LOOKUP_ROOT;
2283         }
2284         set_nameidata(&nd, dfd, name);
2285         retval = path_lookupat(&nd, flags | LOOKUP_RCU, path);
2286         if (unlikely(retval == -ECHILD))
2287                 retval = path_lookupat(&nd, flags, path);
2288         if (unlikely(retval == -ESTALE))
2289                 retval = path_lookupat(&nd, flags | LOOKUP_REVAL, path);
2290
2291         if (likely(!retval))
2292                 audit_inode(name, path->dentry, flags & LOOKUP_PARENT);
2293         restore_nameidata();
2294         putname(name);
2295         return retval;
2296 }
2297
2298 /* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
2299 static int path_parentat(struct nameidata *nd, unsigned flags,
2300                                 struct path *parent)
2301 {
2302         const char *s = path_init(nd, flags);
2303         int err;
2304         if (IS_ERR(s))
2305                 return PTR_ERR(s);
2306         err = link_path_walk(s, nd);
2307         if (!err)
2308                 err = complete_walk(nd);
2309         if (!err) {
2310                 *parent = nd->path;
2311                 nd->path.mnt = NULL;
2312                 nd->path.dentry = NULL;
2313         }
2314         terminate_walk(nd);
2315         return err;
2316 }
2317
2318 static struct filename *filename_parentat(int dfd, struct filename *name,
2319                                 unsigned int flags, struct path *parent,
2320                                 struct qstr *last, int *type)
2321 {
2322         int retval;
2323         struct nameidata nd;
2324
2325         if (IS_ERR(name))
2326                 return name;
2327         set_nameidata(&nd, dfd, name);
2328         retval = path_parentat(&nd, flags | LOOKUP_RCU, parent);
2329         if (unlikely(retval == -ECHILD))
2330                 retval = path_parentat(&nd, flags, parent);
2331         if (unlikely(retval == -ESTALE))
2332                 retval = path_parentat(&nd, flags | LOOKUP_REVAL, parent);
2333         if (likely(!retval)) {
2334                 *last = nd.last;
2335                 *type = nd.last_type;
2336                 audit_inode(name, parent->dentry, LOOKUP_PARENT);
2337         } else {
2338                 putname(name);
2339                 name = ERR_PTR(retval);
2340         }
2341         restore_nameidata();
2342         return name;
2343 }
2344
2345 /* does lookup, returns the object with parent locked */
2346 struct dentry *kern_path_locked(const char *name, struct path *path)
2347 {
2348         struct filename *filename;
2349         struct dentry *d;
2350         struct qstr last;
2351         int type;
2352
2353         filename = filename_parentat(AT_FDCWD, getname_kernel(name), 0, path,
2354                                     &last, &type);
2355         if (IS_ERR(filename))
2356                 return ERR_CAST(filename);
2357         if (unlikely(type != LAST_NORM)) {
2358                 path_put(path);
2359                 putname(filename);
2360                 return ERR_PTR(-EINVAL);
2361         }
2362         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
2363         d = __lookup_hash(&last, path->dentry, 0);
2364         if (IS_ERR(d)) {
2365                 inode_unlock(path->dentry->d_inode);
2366                 path_put(path);
2367         }
2368         putname(filename);
2369         return d;
2370 }
2371
2372 int kern_path(const char *name, unsigned int flags, struct path *path)
2373 {
2374         return filename_lookup(AT_FDCWD, getname_kernel(name),
2375                                flags, path, NULL);
2376 }
2377 EXPORT_SYMBOL(kern_path);
2378
2379 /**
2380  * vfs_path_lookup - lookup a file path relative to a dentry-vfsmount pair
2381  * @dentry:  pointer to dentry of the base directory
2382  * @mnt: pointer to vfs mount of the base directory
2383  * @name: pointer to file name
2384  * @flags: lookup flags
2385  * @path: pointer to struct path to fill
2386  */
2387 int vfs_path_lookup(struct dentry *dentry, struct vfsmount *mnt,
2388                     const char *name, unsigned int flags,
2389                     struct path *path)
2390 {
2391         struct path root = {.mnt = mnt, .dentry = dentry};
2392         /* the first argument of filename_lookup() is ignored with root */
2393         return filename_lookup(AT_FDCWD, getname_kernel(name),
2394                                flags , path, &root);
2395 }
2396 EXPORT_SYMBOL(vfs_path_lookup);
2397
2398 /**
2399  * lookup_one_len - filesystem helper to lookup single pathname component
2400  * @name:       pathname component to lookup
2401  * @base:       base directory to lookup from
2402  * @len:        maximum length @len should be interpreted to
2403  *
2404  * Note that this routine is purely a helper for filesystem usage and should
2405  * not be called by generic code.
2406  *
2407  * The caller must hold base->i_mutex.
2408  */
2409 struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
2410 {
2411         struct qstr this;
2412         unsigned int c;
2413         int err;
2414
2415         WARN_ON_ONCE(!inode_is_locked(base->d_inode));
2416
2417         this.name = name;
2418         this.len = len;
2419         this.hash = full_name_hash(base, name, len);
2420         if (!len)
2421                 return ERR_PTR(-EACCES);
2422
2423         if (unlikely(name[0] == '.')) {
2424                 if (len < 2 || (len == 2 && name[1] == '.'))
2425                         return ERR_PTR(-EACCES);
2426         }
2427
2428         while (len--) {
2429                 c = *(const unsigned char *)name++;
2430                 if (c == '/' || c == '\0')
2431                         return ERR_PTR(-EACCES);
2432         }
2433         /*
2434          * See if the low-level filesystem might want
2435          * to use its own hash..
2436          */
2437         if (base->d_flags & DCACHE_OP_HASH) {
2438                 int err = base->d_op->d_hash(base, &this);
2439                 if (err < 0)
2440                         return ERR_PTR(err);
2441         }
2442
2443         err = inode_permission(base->d_inode, MAY_EXEC);
2444         if (err)
2445                 return ERR_PTR(err);
2446
2447         return __lookup_hash(&this, base, 0);
2448 }
2449 EXPORT_SYMBOL(lookup_one_len);
2450
2451 /**
2452  * lookup_one_len_unlocked - filesystem helper to lookup single pathname component
2453  * @name:       pathname component to lookup
2454  * @base:       base directory to lookup from
2455  * @len:        maximum length @len should be interpreted to
2456  *
2457  * Note that this routine is purely a helper for filesystem usage and should
2458  * not be called by generic code.
2459  *
2460  * Unlike lookup_one_len, it should be called without the parent
2461  * i_mutex held, and will take the i_mutex itself if necessary.
2462  */
2463 struct dentry *lookup_one_len_unlocked(const char *name,
2464                                        struct dentry *base, int len)
2465 {
2466         struct qstr this;
2467         unsigned int c;
2468         int err;
2469         struct dentry *ret;
2470
2471         this.name = name;
2472         this.len = len;
2473         this.hash = full_name_hash(base, name, len);
2474         if (!len)
2475                 return ERR_PTR(-EACCES);
2476
2477         if (unlikely(name[0] == '.')) {
2478                 if (len < 2 || (len == 2 && name[1] == '.'))
2479                         return ERR_PTR(-EACCES);
2480         }
2481
2482         while (len--) {
2483                 c = *(const unsigned char *)name++;
2484                 if (c == '/' || c == '\0')
2485                         return ERR_PTR(-EACCES);
2486         }
2487         /*
2488          * See if the low-level filesystem might want
2489          * to use its own hash..
2490          */
2491         if (base->d_flags & DCACHE_OP_HASH) {
2492                 int err = base->d_op->d_hash(base, &this);
2493                 if (err < 0)
2494                         return ERR_PTR(err);
2495         }
2496
2497         err = inode_permission(base->d_inode, MAY_EXEC);
2498         if (err)
2499                 return ERR_PTR(err);
2500
2501         ret = lookup_dcache(&this, base, 0);
2502         if (!ret)
2503                 ret = lookup_slow(&this, base, 0);
2504         return ret;
2505 }
2506 EXPORT_SYMBOL(lookup_one_len_unlocked);
2507
2508 #ifdef CONFIG_UNIX98_PTYS
2509 int path_pts(struct path *path)
2510 {
2511         /* Find something mounted on "pts" in the same directory as
2512          * the input path.
2513          */
2514         struct dentry *child, *parent;
2515         struct qstr this;
2516         int ret;
2517
2518         ret = path_parent_directory(path);
2519         if (ret)
2520                 return ret;
2521
2522         parent = path->dentry;
2523         this.name = "pts";
2524         this.len = 3;
2525         child = d_hash_and_lookup(parent, &this);
2526         if (!child)
2527                 return -ENOENT;
2528
2529         path->dentry = child;
2530         dput(parent);
2531         follow_mount(path);
2532         return 0;
2533 }
2534 #endif
2535
2536 int user_path_at_empty(int dfd, const char __user *name, unsigned flags,
2537                  struct path *path, int *empty)
2538 {
2539         return filename_lookup(dfd, getname_flags(name, flags, empty),
2540                                flags, path, NULL);
2541 }
2542 EXPORT_SYMBOL(user_path_at_empty);
2543
2544 /*
2545  * NB: most callers don't do anything directly with the reference to the
2546  *     to struct filename, but the nd->last pointer points into the name string
2547  *     allocated by getname. So we must hold the reference to it until all
2548  *     path-walking is complete.
2549  */
2550 static inline struct filename *
2551 user_path_parent(int dfd, const char __user *path,
2552                  struct path *parent,
2553                  struct qstr *last,
2554                  int *type,
2555                  unsigned int flags)
2556 {
2557         /* only LOOKUP_REVAL is allowed in extra flags */
2558         return filename_parentat(dfd, getname(path), flags & LOOKUP_REVAL,
2559                                  parent, last, type);
2560 }
2561
2562 /**
2563  * mountpoint_last - look up last component for umount
2564  * @nd:   pathwalk nameidata - currently pointing at parent directory of "last"
2565  * @path: pointer to container for result
2566  *
2567  * This is a special lookup_last function just for umount. In this case, we
2568  * need to resolve the path without doing any revalidation.
2569  *
2570  * The nameidata should be the result of doing a LOOKUP_PARENT pathwalk. Since
2571  * mountpoints are always pinned in the dcache, their ancestors are too. Thus,
2572  * in almost all cases, this lookup will be served out of the dcache. The only
2573  * cases where it won't are if nd->last refers to a symlink or the path is
2574  * bogus and it doesn't exist.
2575  *
2576  * Returns:
2577  * -error: if there was an error during lookup. This includes -ENOENT if the
2578  *         lookup found a negative dentry. The nd->path reference will also be
2579  *         put in this case.
2580  *
2581  * 0:      if we successfully resolved nd->path and found it to not to be a
2582  *         symlink that needs to be followed. "path" will also be populated.
2583  *         The nd->path reference will also be put.
2584  *
2585  * 1:      if we successfully resolved nd->last and found it to be a symlink
2586  *         that needs to be followed. "path" will be populated with the path
2587  *         to the link, and nd->path will *not* be put.
2588  */
2589 static int
2590 mountpoint_last(struct nameidata *nd, struct path *path)
2591 {
2592         int error = 0;
2593         struct dentry *dentry;
2594         struct dentry *dir = nd->path.dentry;
2595
2596         /* If we're in rcuwalk, drop out of it to handle last component */
2597         if (nd->flags & LOOKUP_RCU) {
2598                 if (unlazy_walk(nd, NULL, 0))
2599                         return -ECHILD;
2600         }
2601
2602         nd->flags &= ~LOOKUP_PARENT;
2603
2604         if (unlikely(nd->last_type != LAST_NORM)) {
2605                 error = handle_dots(nd, nd->last_type);
2606                 if (error)
2607                         return error;
2608                 dentry = dget(nd->path.dentry);
2609         } else {
2610                 dentry = d_lookup(dir, &nd->last);
2611                 if (!dentry) {
2612                         /*
2613                          * No cached dentry. Mounted dentries are pinned in the
2614                          * cache, so that means that this dentry is probably
2615                          * a symlink or the path doesn't actually point
2616                          * to a mounted dentry.
2617                          */
2618                         dentry = lookup_slow(&nd->last, dir,
2619                                              nd->flags | LOOKUP_NO_REVAL);
2620                         if (IS_ERR(dentry))
2621                                 return PTR_ERR(dentry);
2622                 }
2623         }
2624         if (d_is_negative(dentry)) {
2625                 dput(dentry);
2626                 return -ENOENT;
2627         }
2628         if (nd->depth)
2629                 put_link(nd);
2630         path->dentry = dentry;
2631         path->mnt = nd->path.mnt;
2632         error = should_follow_link(nd, path, nd->flags & LOOKUP_FOLLOW,
2633                                    d_backing_inode(dentry), 0);
2634         if (unlikely(error))
2635                 return error;
2636         mntget(path->mnt);
2637         follow_mount(path);
2638         return 0;
2639 }
2640
2641 /**
2642  * path_mountpoint - look up a path to be umounted
2643  * @nd:         lookup context
2644  * @flags:      lookup flags
2645  * @path:       pointer to container for result
2646  *
2647  * Look up the given name, but don't attempt to revalidate the last component.
2648  * Returns 0 and "path" will be valid on success; Returns error otherwise.
2649  */
2650 static int
2651 path_mountpoint(struct nameidata *nd, unsigned flags, struct path *path)
2652 {
2653         const char *s = path_init(nd, flags);
2654         int err;
2655         if (IS_ERR(s))
2656                 return PTR_ERR(s);
2657         while (!(err = link_path_walk(s, nd)) &&
2658                 (err = mountpoint_last(nd, path)) > 0) {
2659                 s = trailing_symlink(nd);
2660                 if (IS_ERR(s)) {
2661                         err = PTR_ERR(s);
2662                         break;
2663                 }
2664         }
2665         terminate_walk(nd);
2666         return err;
2667 }
2668
2669 static int
2670 filename_mountpoint(int dfd, struct filename *name, struct path *path,
2671                         unsigned int flags)
2672 {
2673         struct nameidata nd;
2674         int error;
2675         if (IS_ERR(name))
2676                 return PTR_ERR(name);
2677         set_nameidata(&nd, dfd, name);
2678         error = path_mountpoint(&nd, flags | LOOKUP_RCU, path);
2679         if (unlikely(error == -ECHILD))
2680                 error = path_mountpoint(&nd, flags, path);
2681         if (unlikely(error == -ESTALE))
2682                 error = path_mountpoint(&nd, flags | LOOKUP_REVAL, path);
2683         if (likely(!error))
2684                 audit_inode(name, path->dentry, 0);
2685         restore_nameidata();
2686         putname(name);
2687         return error;
2688 }
2689
2690 /**
2691  * user_path_mountpoint_at - lookup a path from userland in order to umount it
2692  * @dfd:        directory file descriptor
2693  * @name:       pathname from userland
2694  * @flags:      lookup flags
2695  * @path:       pointer to container to hold result
2696  *
2697  * A umount is a special case for path walking. We're not actually interested
2698  * in the inode in this situation, and ESTALE errors can be a problem. We
2699  * simply want track down the dentry and vfsmount attached at the mountpoint
2700  * and avoid revalidating the last component.
2701  *
2702  * Returns 0 and populates "path" on success.
2703  */
2704 int
2705 user_path_mountpoint_at(int dfd, const char __user *name, unsigned int flags,
2706                         struct path *path)
2707 {
2708         return filename_mountpoint(dfd, getname(name), path, flags);
2709 }
2710
2711 int
2712 kern_path_mountpoint(int dfd, const char *name, struct path *path,
2713                         unsigned int flags)
2714 {
2715         return filename_mountpoint(dfd, getname_kernel(name), path, flags);
2716 }
2717 EXPORT_SYMBOL(kern_path_mountpoint);
2718
2719 int __check_sticky(struct inode *dir, struct inode *inode)
2720 {
2721         kuid_t fsuid = current_fsuid();
2722
2723         if (uid_eq(inode->i_uid, fsuid))
2724                 return 0;
2725         if (uid_eq(dir->i_uid, fsuid))
2726                 return 0;
2727         return !capable_wrt_inode_uidgid(inode, CAP_FOWNER);
2728 }
2729 EXPORT_SYMBOL(__check_sticky);
2730
2731 /*
2732  *      Check whether we can remove a link victim from directory dir, check
2733  *  whether the type of victim is right.
2734  *  1. We can't do it if dir is read-only (done in permission())
2735  *  2. We should have write and exec permissions on dir
2736  *  3. We can't remove anything from append-only dir
2737  *  4. We can't do anything with immutable dir (done in permission())
2738  *  5. If the sticky bit on dir is set we should either
2739  *      a. be owner of dir, or
2740  *      b. be owner of victim, or
2741  *      c. have CAP_FOWNER capability
2742  *  6. If the victim is append-only or immutable we can't do antyhing with
2743  *     links pointing to it.
2744  *  7. If we were asked to remove a directory and victim isn't one - ENOTDIR.
2745  *  8. If we were asked to remove a non-directory and victim isn't one - EISDIR.
2746  *  9. We can't remove a root or mountpoint.
2747  * 10. We don't allow removal of NFS sillyrenamed files; it's handled by
2748  *     nfs_async_unlink().
2749  */
2750 static int may_delete(struct inode *dir, struct dentry *victim, bool isdir)
2751 {
2752         struct inode *inode = d_backing_inode(victim);
2753         int error;
2754
2755         if (d_is_negative(victim))
2756                 return -ENOENT;
2757         BUG_ON(!inode);
2758
2759         BUG_ON(victim->d_parent->d_inode != dir);
2760         audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);
2761
2762         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
2763         if (error)
2764                 return error;
2765         if (IS_APPEND(dir))
2766                 return -EPERM;
2767
2768         if (check_sticky(dir, inode) || IS_APPEND(inode) ||
2769             IS_IMMUTABLE(inode) || IS_SWAPFILE(inode))
2770                 return -EPERM;
2771         if (isdir) {
2772                 if (!d_is_dir(victim))
2773                         return -ENOTDIR;
2774                 if (IS_ROOT(victim))
2775                         return -EBUSY;
2776         } else if (d_is_dir(victim))
2777                 return -EISDIR;
2778         if (IS_DEADDIR(dir))
2779                 return -ENOENT;
2780         if (victim->d_flags & DCACHE_NFSFS_RENAMED)
2781                 return -EBUSY;
2782         return 0;
2783 }
2784
2785 /*      Check whether we can create an object with dentry child in directory
2786  *  dir.
2787  *  1. We can't do it if child already exists (open has special treatment for
2788  *     this case, but since we are inlined it's OK)
2789  *  2. We can't do it if dir is read-only (done in permission())
2790  *  3. We should have write and exec permissions on dir
2791  *  4. We can't do it if dir is immutable (done in permission())
2792  */
2793 static inline int may_create(struct inode *dir, struct dentry *child)
2794 {
2795         audit_inode_child(dir, child, AUDIT_TYPE_CHILD_CREATE);
2796         if (child->d_inode)
2797                 return -EEXIST;
2798         if (IS_DEADDIR(dir))
2799                 return -ENOENT;
2800         return inode_permission(dir, MAY_WRITE | MAY_EXEC);
2801 }
2802
2803 /*
2804  * p1 and p2 should be directories on the same fs.
2805  */
2806 struct dentry *lock_rename(struct dentry *p1, struct dentry *p2)
2807 {
2808         struct dentry *p;
2809
2810         if (p1 == p2) {
2811                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2812                 return NULL;
2813         }
2814
2815         mutex_lock(&p1->d_sb->s_vfs_rename_mutex);
2816
2817         p = d_ancestor(p2, p1);
2818         if (p) {
2819                 inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
2820                 inode_lock_nested(p1->d_inode, I_MUTEX_CHILD);
2821                 return p;
2822         }
2823
2824         p = d_ancestor(p1, p2);
2825         if (p) {
2826                 inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2827                 inode_lock_nested(p2->d_inode, I_MUTEX_CHILD);
2828                 return p;
2829         }
2830
2831         inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
2832         inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
2833         return NULL;
2834 }
2835 EXPORT_SYMBOL(lock_rename);
2836
2837 void unlock_rename(struct dentry *p1, struct dentry *p2)
2838 {
2839         inode_unlock(p1->d_inode);
2840         if (p1 != p2) {
2841                 inode_unlock(p2->d_inode);
2842                 mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
2843         }
2844 }
2845 EXPORT_SYMBOL(unlock_rename);
2846
2847 int vfs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
2848                 bool want_excl)
2849 {
2850         int error = may_create(dir, dentry);
2851         if (error)
2852                 return error;
2853
2854         if (!dir->i_op->create)
2855                 return -EACCES; /* shouldn't it be ENOSYS? */
2856         mode &= S_IALLUGO;
2857         mode |= S_IFREG;
2858         error = security_inode_create(dir, dentry, mode);
2859         if (error)
2860                 return error;
2861         error = dir->i_op->create(dir, dentry, mode, want_excl);
2862         if (!error)
2863                 fsnotify_create(dir, dentry);
2864         return error;
2865 }
2866 EXPORT_SYMBOL(vfs_create);
2867
2868 static int may_open(struct path *path, int acc_mode, int flag)
2869 {
2870         struct dentry *dentry = path->dentry;
2871         struct inode *inode = dentry->d_inode;
2872         int error;
2873
2874         if (!inode)
2875                 return -ENOENT;
2876
2877         switch (inode->i_mode & S_IFMT) {
2878         case S_IFLNK:
2879                 return -ELOOP;
2880         case S_IFDIR:
2881                 if (acc_mode & MAY_WRITE)
2882                         return -EISDIR;
2883                 break;
2884         case S_IFBLK:
2885         case S_IFCHR:
2886                 if (path->mnt->mnt_flags & MNT_NODEV)
2887                         return -EACCES;
2888                 /*FALLTHRU*/
2889         case S_IFIFO:
2890         case S_IFSOCK:
2891                 flag &= ~O_TRUNC;
2892                 break;
2893         }
2894
2895         error = inode_permission(inode, MAY_OPEN | acc_mode);
2896         if (error)
2897                 return error;
2898
2899         /*
2900          * An append-only file must be opened in append mode for writing.
2901          */
2902         if (IS_APPEND(inode)) {
2903                 if  ((flag & O_ACCMODE) != O_RDONLY && !(flag & O_APPEND))
2904                         return -EPERM;
2905                 if (flag & O_TRUNC)
2906                         return -EPERM;
2907         }
2908
2909         /* O_NOATIME can only be set by the owner or superuser */
2910         if (flag & O_NOATIME && !inode_owner_or_capable(inode))
2911                 return -EPERM;
2912
2913         return 0;
2914 }
2915
2916 static int handle_truncate(struct file *filp)
2917 {
2918         struct path *path = &filp->f_path;
2919         struct inode *inode = path->dentry->d_inode;
2920         int error = get_write_access(inode);
2921         if (error)
2922                 return error;
2923         /*
2924          * Refuse to truncate files with mandatory locks held on them.
2925          */
2926         error = locks_verify_locked(filp);
2927         if (!error)
2928                 error = security_path_truncate(path);
2929         if (!error) {
2930                 error = do_truncate(path->dentry, 0,
2931                                     ATTR_MTIME|ATTR_CTIME|ATTR_OPEN,
2932                                     filp);
2933         }
2934         put_write_access(inode);
2935         return error;
2936 }
2937
2938 static inline int open_to_namei_flags(int flag)
2939 {
2940         if ((flag & O_ACCMODE) == 3)
2941                 flag--;
2942         return flag;
2943 }
2944
2945 static int may_o_create(const struct path *dir, struct dentry *dentry, umode_t mode)
2946 {
2947         int error = security_path_mknod(dir, dentry, mode, 0);
2948         if (error)
2949                 return error;
2950
2951         error = inode_permission(dir->dentry->d_inode, MAY_WRITE | MAY_EXEC);
2952         if (error)
2953                 return error;
2954
2955         return security_inode_create(dir->dentry->d_inode, dentry, mode);
2956 }
2957
2958 /*
2959  * Attempt to atomically look up, create and open a file from a negative
2960  * dentry.
2961  *
2962  * Returns 0 if successful.  The file will have been created and attached to
2963  * @file by the filesystem calling finish_open().
2964  *
2965  * Returns 1 if the file was looked up only or didn't need creating.  The
2966  * caller will need to perform the open themselves.  @path will have been
2967  * updated to point to the new dentry.  This may be negative.
2968  *
2969  * Returns an error code otherwise.
2970  */
2971 static int atomic_open(struct nameidata *nd, struct dentry *dentry,
2972                         struct path *path, struct file *file,
2973                         const struct open_flags *op,
2974                         int open_flag, umode_t mode,
2975                         int *opened)
2976 {
2977         struct dentry *const DENTRY_NOT_SET = (void *) -1UL;
2978         struct inode *dir =  nd->path.dentry->d_inode;
2979         int error;
2980
2981         if (!(~open_flag & (O_EXCL | O_CREAT))) /* both O_EXCL and O_CREAT */
2982                 open_flag &= ~O_TRUNC;
2983
2984         if (nd->flags & LOOKUP_DIRECTORY)
2985                 open_flag |= O_DIRECTORY;
2986
2987         file->f_path.dentry = DENTRY_NOT_SET;
2988         file->f_path.mnt = nd->path.mnt;
2989         error = dir->i_op->atomic_open(dir, dentry, file,
2990                                        open_to_namei_flags(open_flag),
2991                                        mode, opened);
2992         d_lookup_done(dentry);
2993         if (!error) {
2994                 /*
2995                  * We didn't have the inode before the open, so check open
2996                  * permission here.
2997                  */
2998                 int acc_mode = op->acc_mode;
2999                 if (*opened & FILE_CREATED) {
3000                         WARN_ON(!(open_flag & O_CREAT));
3001                         fsnotify_create(dir, dentry);
3002                         acc_mode = 0;
3003                 }
3004                 error = may_open(&file->f_path, acc_mode, open_flag);
3005                 if (WARN_ON(error > 0))
3006                         error = -EINVAL;
3007         } else if (error > 0) {
3008                 if (WARN_ON(file->f_path.dentry == DENTRY_NOT_SET)) {
3009                         error = -EIO;
3010                 } else {
3011                         if (file->f_path.dentry) {
3012                                 dput(dentry);
3013                                 dentry = file->f_path.dentry;
3014                         }
3015                         if (*opened & FILE_CREATED)
3016                                 fsnotify_create(dir, dentry);
3017                         if (unlikely(d_is_negative(dentry))) {
3018                                 error = -ENOENT;
3019                         } else {
3020                                 path->dentry = dentry;
3021                                 path->mnt = nd->path.mnt;
3022                                 return 1;
3023                         }
3024                 }
3025         }
3026         dput(dentry);
3027         return error;
3028 }
3029
3030 /*
3031  * Look up and maybe create and open the last component.
3032  *
3033  * Must be called with i_mutex held on parent.
3034  *
3035  * Returns 0 if the file was successfully atomically created (if necessary) and
3036  * opened.  In this case the file will be returned attached to @file.
3037  *
3038  * Returns 1 if the file was not completely opened at this time, though lookups
3039  * and creations will have been performed and the dentry returned in @path will
3040  * be positive upon return if O_CREAT was specified.  If O_CREAT wasn't
3041  * specified then a negative dentry may be returned.
3042  *
3043  * An error code is returned otherwise.
3044  *
3045  * FILE_CREATE will be set in @*opened if the dentry was created and will be
3046  * cleared otherwise prior to returning.
3047  */
3048 static int lookup_open(struct nameidata *nd, struct path *path,
3049                         struct file *file,
3050                         const struct open_flags *op,
3051                         bool got_write, int *opened)
3052 {
3053         struct dentry *dir = nd->path.dentry;
3054         struct inode *dir_inode = dir->d_inode;
3055         int open_flag = op->open_flag;
3056         struct dentry *dentry;
3057         int error, create_error = 0;
3058         umode_t mode = op->mode;
3059         DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
3060
3061         if (unlikely(IS_DEADDIR(dir_inode)))
3062                 return -ENOENT;
3063
3064         *opened &= ~FILE_CREATED;
3065         dentry = d_lookup(dir, &nd->last);
3066         for (;;) {
3067                 if (!dentry) {
3068                         dentry = d_alloc_parallel(dir, &nd->last, &wq);
3069                         if (IS_ERR(dentry))
3070                                 return PTR_ERR(dentry);
3071                 }
3072                 if (d_in_lookup(dentry))
3073                         break;
3074
3075                 if (!(dentry->d_flags & DCACHE_OP_REVALIDATE))
3076                         break;
3077
3078                 error = d_revalidate(dentry, nd->flags);
3079                 if (likely(error > 0))
3080                         break;
3081                 if (error)
3082                         goto out_dput;
3083                 d_invalidate(dentry);
3084                 dput(dentry);
3085                 dentry = NULL;
3086         }
3087         if (dentry->d_inode) {
3088                 /* Cached positive dentry: will open in f_op->open */
3089                 goto out_no_open;
3090         }
3091
3092         /*
3093          * Checking write permission is tricky, bacuse we don't know if we are
3094          * going to actually need it: O_CREAT opens should work as long as the
3095          * file exists.  But checking existence breaks atomicity.  The trick is
3096          * to check access and if not granted clear O_CREAT from the flags.
3097          *
3098          * Another problem is returing the "right" error value (e.g. for an
3099          * O_EXCL open we want to return EEXIST not EROFS).
3100          */
3101         if (open_flag & O_CREAT) {
3102                 if (!IS_POSIXACL(dir->d_inode))
3103                         mode &= ~current_umask();
3104                 if (unlikely(!got_write)) {
3105                         create_error = -EROFS;
3106                         open_flag &= ~O_CREAT;
3107                         if (open_flag & (O_EXCL | O_TRUNC))
3108                                 goto no_open;
3109                         /* No side effects, safe to clear O_CREAT */
3110                 } else {
3111                         create_error = may_o_create(&nd->path, dentry, mode);
3112                         if (create_error) {
3113                                 open_flag &= ~O_CREAT;
3114                                 if (open_flag & O_EXCL)
3115                                         goto no_open;
3116                         }
3117                 }
3118         } else if ((open_flag & (O_TRUNC|O_WRONLY|O_RDWR)) &&
3119                    unlikely(!got_write)) {
3120                 /*
3121                  * No O_CREATE -> atomicity not a requirement -> fall
3122                  * back to lookup + open
3123                  */
3124                 goto no_open;
3125         }
3126
3127         if (dir_inode->i_op->atomic_open) {
3128                 error = atomic_open(nd, dentry, path, file, op, open_flag,
3129                                     mode, opened);
3130                 if (unlikely(error == -ENOENT) && create_error)
3131                         error = create_error;
3132                 return error;
3133         }
3134
3135 no_open:
3136         if (d_in_lookup(dentry)) {
3137                 struct dentry *res = dir_inode->i_op->lookup(dir_inode, dentry,
3138                                                              nd->flags);
3139                 d_lookup_done(dentry);
3140                 if (unlikely(res)) {
3141                         if (IS_ERR(res)) {
3142                                 error = PTR_ERR(res);
3143                                 goto out_dput;
3144                         }
3145                         dput(dentry);
3146                         dentry = res;
3147                 }
3148         }
3149
3150         /* Negative dentry, just create the file */
3151         if (!dentry->d_inode && (open_flag & O_CREAT)) {
3152                 *opened |= FILE_CREATED;
3153                 audit_inode_child(dir_inode, dentry, AUDIT_TYPE_CHILD_CREATE);
3154                 if (!dir_inode->i_op->create) {
3155                         error = -EACCES;
3156                         goto out_dput;
3157                 }
3158                 error = dir_inode->i_op->create(dir_inode, dentry, mode,
3159                                                 open_flag & O_EXCL);
3160                 if (error)
3161                         goto out_dput;
3162                 fsnotify_create(dir_inode, dentry);
3163         }
3164         if (unlikely(create_error) && !dentry->d_inode) {
3165                 error = create_error;
3166                 goto out_dput;
3167         }
3168 out_no_open:
3169         path->dentry = dentry;
3170         path->mnt = nd->path.mnt;
3171         return 1;
3172
3173 out_dput:
3174         dput(dentry);
3175         return error;
3176 }
3177
3178 /*
3179  * Handle the last step of open()
3180  */
3181 static int do_last(struct nameidata *nd,
3182                    struct file *file, const struct open_flags *op,
3183                    int *opened)
3184 {
3185         struct dentry *dir = nd->path.dentry;
3186         int open_flag = op->open_flag;
3187         bool will_truncate = (open_flag & O_TRUNC) != 0;
3188         bool got_write = false;
3189         int acc_mode = op->acc_mode;
3190         unsigned seq;
3191         struct inode *inode;
3192         struct path path;
3193         int error;
3194
3195         nd->flags &= ~LOOKUP_PARENT;
3196         nd->flags |= op->intent;
3197
3198         if (nd->last_type != LAST_NORM) {
3199                 error = handle_dots(nd, nd->last_type);
3200                 if (unlikely(error))
3201                         return error;
3202                 goto finish_open;
3203         }
3204
3205         if (!(open_flag & O_CREAT)) {
3206                 if (nd->last.name[nd->last.len])
3207                         nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
3208                 /* we _can_ be in RCU mode here */
3209                 error = lookup_fast(nd, &path, &inode, &seq);
3210                 if (likely(error > 0))
3211                         goto finish_lookup;
3212
3213                 if (error < 0)
3214                         return error;
3215
3216                 BUG_ON(nd->inode != dir->d_inode);
3217                 BUG_ON(nd->flags & LOOKUP_RCU);
3218         } else {
3219                 /* create side of things */
3220                 /*
3221                  * This will *only* deal with leaving RCU mode - LOOKUP_JUMPED
3222                  * has been cleared when we got to the last component we are
3223                  * about to look up
3224                  */
3225                 error = complete_walk(nd);
3226                 if (error)
3227                         return error;
3228
3229                 audit_inode(nd->name, dir, LOOKUP_PARENT);
3230                 /* trailing slashes? */
3231                 if (unlikely(nd->last.name[nd->last.len]))
3232                         return -EISDIR;
3233         }
3234
3235         if (open_flag & (O_CREAT | O_TRUNC | O_WRONLY | O_RDWR)) {
3236                 error = mnt_want_write(nd->path.mnt);
3237                 if (!error)
3238                         got_write = true;
3239                 /*
3240                  * do _not_ fail yet - we might not need that or fail with
3241                  * a different error; let lookup_open() decide; we'll be
3242                  * dropping this one anyway.
3243                  */
3244         }
3245         if (open_flag & O_CREAT)
3246                 inode_lock(dir->d_inode);
3247         else
3248                 inode_lock_shared(dir->d_inode);
3249         error = lookup_open(nd, &path, file, op, got_write, opened);
3250         if (open_flag & O_CREAT)
3251                 inode_unlock(dir->d_inode);
3252         else
3253                 inode_unlock_shared(dir->d_inode);
3254
3255         if (error <= 0) {
3256                 if (error)
3257                         goto out;
3258
3259                 if ((*opened & FILE_CREATED) ||
3260                     !S_ISREG(file_inode(file)->i_mode))
3261                         will_truncate = false;
3262
3263                 audit_inode(nd->name, file->f_path.dentry, 0);
3264                 goto opened;
3265         }
3266
3267         if (*opened & FILE_CREATED) {
3268                 /* Don't check for write permission, don't truncate */
3269                 open_flag &= ~O_TRUNC;
3270                 will_truncate = false;
3271                 acc_mode = 0;
3272                 path_to_nameidata(&path, nd);
3273                 goto finish_open_created;
3274         }
3275
3276         /*
3277          * If atomic_open() acquired write access it is dropped now due to
3278          * possible mount and symlink following (this might be optimized away if
3279          * necessary...)
3280          */
3281         if (got_write) {
3282                 mnt_drop_write(nd->path.mnt);
3283                 got_write = false;
3284         }
3285
3286         error = follow_managed(&path, nd);
3287         if (unlikely(error < 0))
3288                 return error;
3289
3290         if (unlikely(d_is_negative(path.dentry))) {
3291                 path_to_nameidata(&path, nd);
3292                 return -ENOENT;
3293         }
3294
3295         /*
3296          * create/update audit record if it already exists.
3297          */
3298         audit_inode(nd->name, path.dentry, 0);
3299
3300         if (unlikely((open_flag & (O_EXCL | O_CREAT)) == (O_EXCL | O_CREAT))) {
3301                 path_to_nameidata(&path, nd);
3302                 return -EEXIST;
3303         }
3304
3305         seq = 0;        /* out of RCU mode, so the value doesn't matter */
3306         inode = d_backing_inode(path.dentry);
3307 finish_lookup:
3308         if (nd->depth)
3309                 put_link(nd);
3310         error = should_follow_link(nd, &path, nd->flags & LOOKUP_FOLLOW,
3311                                    inode, seq);
3312         if (unlikely(error))
3313                 return error;
3314
3315         path_to_nameidata(&path, nd);
3316         nd->inode = inode;
3317         nd->seq = seq;
3318         /* Why this, you ask?  _Now_ we might have grown LOOKUP_JUMPED... */
3319 finish_open:
3320         error = complete_walk(nd);
3321         if (error)
3322                 return error;
3323         audit_inode(nd->name, nd->path.dentry, 0);
3324         error = -EISDIR;
3325         if ((open_flag & O_CREAT) && d_is_dir(nd->path.dentry))
3326                 goto out;
3327         error = -ENOTDIR;
3328         if ((nd->flags & LOOKUP_DIRECTORY) && !d_can_lookup(nd->path.dentry))
3329                 goto out;
3330         if (!d_is_reg(nd->path.dentry))
3331                 will_truncate = false;
3332
3333         if (will_truncate) {
3334                 error = mnt_want_write(nd->path.mnt);
3335                 if (error)
3336                         goto out;
3337                 got_write = true;
3338         }
3339 finish_open_created:
3340         error = may_open(&nd->path, acc_mode, open_flag);
3341         if (error)
3342                 goto out;
3343         BUG_ON(*opened & FILE_OPENED); /* once it's opened, it's opened */
3344         error = vfs_open(&nd->path, file, current_cred());
3345         if (error)
3346                 goto out;
3347         *opened |= FILE_OPENED;
3348 opened:
3349         error = open_check_o_direct(file);
3350         if (!error)
3351                 error = ima_file_check(file, op->acc_mode, *opened);
3352         if (!error && will_truncate)
3353                 error = handle_truncate(file);
3354 out:
3355         if (unlikely(error) && (*opened & FILE_OPENED))
3356                 fput(file);
3357         if (unlikely(error > 0)) {
3358                 WARN_ON(1);
3359                 error = -EINVAL;
3360         }
3361         if (got_write)
3362                 mnt_drop_write(nd->path.mnt);
3363         return error;
3364 }
3365
3366 static int do_tmpfile(struct nameidata *nd, unsigned flags,
3367                 const struct open_flags *op,
3368                 struct file *file, int *opened)
3369 {
3370         static const struct qstr name = QSTR_INIT("/", 1);
3371         struct dentry *child;
3372         struct inode *dir;
3373         struct path path;
3374         int error = path_lookupat(nd, flags | LOOKUP_DIRECTORY, &path);
3375         if (unlikely(error))
3376                 return error;
3377         error = mnt_want_write(path.mnt);
3378         if (unlikely(error))
3379                 goto out;
3380         dir = path.dentry->d_inode;
3381         /* we want directory to be writable */
3382         error = inode_permission(dir, MAY_WRITE | MAY_EXEC);
3383         if (error)
3384                 goto out2;
3385         if (!dir->i_op->tmpfile) {
3386                 error = -EOPNOTSUPP;
3387                 goto out2;
3388         }
3389         child = d_alloc(path.dentry, &name);
3390         if (unlikely(!child)) {
3391                 error = -ENOMEM;
3392                 goto out2;
3393         }
3394         dput(path.dentry);
3395         path.dentry = child;
3396         error = dir->i_op->tmpfile(dir, child, op->mode);
3397         if (error)
3398                 goto out2;
3399         audit_inode(nd->name, child, 0);
3400         /* Don't check for other permissions, the inode was just created */
3401         error = may_open(&path, 0, op->open_flag);
3402         if (error)
3403                 goto out2;
3404         file->f_path.mnt = path.mnt;
3405         error = finish_open(file, child, NULL, opened);
3406         if (error)
3407                 goto out2;
3408         error = open_check_o_direct(file);
3409         if (error) {
3410                 fput(file);
3411         } else if (!(op->open_flag & O_EXCL)) {
3412                 struct inode *inode = file_inode(file);
3413                 spin_lock(&inode->i_lock);
3414                 inode->i_state |= I_LINKABLE;
3415                 spin_unlock(&inode->i_lock);
3416         }
3417 out2:
3418         mnt_drop_write(path.mnt);
3419 out:
3420         path_put(&path);
3421         return error;
3422 }
3423
3424 static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
3425 {
3426         struct path path;
3427         int error = path_lookupat(nd, flags, &path);
3428         if (!error) {
3429                 audit_inode(nd->name, path.dentry, 0);
3430                 error = vfs_open(&path, file, current_cred());
3431                 path_put(&path);
3432         }
3433         return error;
3434 }
3435
3436 static struct file *path_openat(struct nameidata *nd,
3437                         const struct open_flags *op, unsigned flags)
3438 {
3439         const char *s;
3440         struct file *file;
3441         int opened = 0;
3442         int error;
3443
3444         file = get_empty_filp();
3445         if (IS_ERR(file))
3446                 return file;
3447
3448         file->f_flags = op->open_flag;
3449
3450         if (unlikely(file->f_flags & __O_TMPFILE)) {
3451                 error = do_tmpfile(nd, flags, op, file, &opened);
3452                 goto out2;
3453         }
3454
3455         if (unlikely(file->f_flags & O_PATH)) {
3456                 error = do_o_path(nd, flags, file);
3457                 if (!error)
3458                         opened |= FILE_OPENED;
3459                 goto out2;
3460         }
3461
3462         s = path_init(nd, flags);
3463         if (IS_ERR(s)) {
3464                 put_filp(file);
3465                 return ERR_CAST(s);
3466         }
3467         while (!(error = link_path_walk(s, nd)) &&
3468                 (error = do_last(nd, file, op, &opened)) > 0) {
3469                 nd->flags &= ~(LOOKUP_OPEN|LOOKUP_CREATE|LOOKUP_EXCL);
3470                 s = trailing_symlink(nd);
3471                 if (IS_ERR(s)) {
3472                         error = PTR_ERR(s);
3473                         break;
3474                 }
3475         }
3476         terminate_walk(nd);
3477 out2:
3478         if (!(opened & FILE_OPENED)) {
3479                 BUG_ON(!error);
3480                 put_filp(file);
3481         }
3482         if (unlikely(error)) {
3483                 if (error == -EOPENSTALE) {
3484                         if (flags & LOOKUP_RCU)
3485                                 error = -ECHILD;
3486                         else
3487                                 error = -ESTALE;
3488                 }
3489                 file = ERR_PTR(error);
3490         }
3491         return file;
3492 }
3493
3494 struct file *do_filp_open(int dfd, struct filename *pathname,
3495                 const struct open_flags *op)
3496 {
3497         struct nameidata nd;
3498         int flags = op->lookup_flags;
3499         struct file *filp;
3500
3501         set_nameidata(&nd, dfd, pathname);
3502         filp = path_openat(&nd, op, flags | LOOKUP_RCU);
3503         if (unlikely(filp == ERR_PTR(-ECHILD)))
3504                 filp = path_openat(&nd, op, flags);
3505         if (unlikely(filp == ERR_PTR(-ESTALE)))
3506                 filp = path_openat(&nd, op, flags | LOOKUP_REVAL);
3507         restore_nameidata();
3508         return filp;
3509 }
3510
3511 struct file *do_file_open_root(struct dentry *dentry, struct vfsmount *mnt,
3512                 const char *name, const struct open_flags *op)
3513 {
3514         struct nameidata nd;
3515         struct file *file;
3516         struct filename *filename;
3517         int flags = op->lookup_flags | LOOKUP_ROOT;
3518
3519         nd.root.mnt = mnt;
3520         nd.root.dentry = dentry;
3521
3522         if (d_is_symlink(dentry) && op->intent & LOOKUP_OPEN)
3523                 return ERR_PTR(-ELOOP);
3524
3525         filename = getname_kernel(name);
3526         if (IS_ERR(filename))
3527                 return ERR_CAST(filename);
3528
3529         set_nameidata(&nd, -1, filename);
3530         file = path_openat(&nd, op, flags | LOOKUP_RCU);
3531         if (unlikely(file == ERR_PTR(-ECHILD)))
3532                 file = path_openat(&nd, op, flags);
3533         if (unlikely(file == ERR_PTR(-ESTALE)))
3534                 file = path_openat(&nd, op, flags | LOOKUP_REVAL);
3535         restore_nameidata();
3536         putname(filename);
3537         return file;
3538 }
3539
3540 static struct dentry *filename_create(int dfd, struct filename *name,
3541                                 struct path *path, unsigned int lookup_flags)
3542 {
3543         struct dentry *dentry = ERR_PTR(-EEXIST);
3544         struct qstr last;
3545         int type;
3546         int err2;
3547         int error;
3548         bool is_dir = (lookup_flags & LOOKUP_DIRECTORY);
3549
3550         /*
3551          * Note that only LOOKUP_REVAL and LOOKUP_DIRECTORY matter here. Any
3552          * other flags passed in are ignored!
3553          */
3554         lookup_flags &= LOOKUP_REVAL;
3555
3556         name = filename_parentat(dfd, name, lookup_flags, path, &last, &type);
3557         if (IS_ERR(name))
3558                 return ERR_CAST(name);
3559
3560         /*
3561          * Yucky last component or no last component at all?
3562          * (foo/., foo/.., /////)
3563          */
3564         if (unlikely(type != LAST_NORM))
3565                 goto out;
3566
3567         /* don't fail immediately if it's r/o, at least try to report other errors */
3568         err2 = mnt_want_write(path->mnt);
3569         /*
3570          * Do the final lookup.
3571          */
3572         lookup_flags |= LOOKUP_CREATE | LOOKUP_EXCL;
3573         inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
3574         dentry = __lookup_hash(&last, path->dentry, lookup_flags);
3575         if (IS_ERR(dentry))
3576                 goto unlock;
3577
3578         error = -EEXIST;
3579         if (d_is_positive(dentry))
3580                 goto fail;
3581
3582         /*
3583          * Special case - lookup gave negative, but... we had foo/bar/
3584          * From the vfs_mknod() POV we just have a negative dentry -
3585          * all is fine. Let's be bastards - you had / on the end, you've
3586          * been asking for (non-existent) directory. -ENOENT for you.
3587          */
3588         if (unlikely(!is_dir && last.name[last.len])) {
3589                 error = -ENOENT;
3590                 goto fail;
3591         }
3592         if (unlikely(err2)) {
3593                 error = err2;
3594                 goto fail;
3595         }
3596         putname(name);
3597         return dentry;
3598 fail:
3599         dput(dentry);
3600         dentry = ERR_PTR(error);
3601 unlock:
3602         inode_unlock(path->dentry->d_inode);
3603         if (!err2)
3604                 mnt_drop_write(path->mnt);
3605 out:
3606         path_put(path);
3607         putname(name);
3608         return dentry;
3609 }
3610
3611 struct dentry *kern_path_create(int dfd, const char *pathname,
3612                                 struct path *path, unsigned int lookup_flags)
3613 {
3614         return filename_create(dfd, getname_kernel(pathname),
3615                                 path, lookup_flags);
3616 }
3617 EXPORT_SYMBOL(kern_path_create);
3618
3619 void done_path_create(struct path *path, struct dentry *dentry)
3620 {
3621         dput(dentry);
3622         inode_unlock(path->dentry->d_inode);
3623         mnt_drop_write(path->mnt);
3624         path_put(path);
3625 }
3626 EXPORT_SYMBOL(done_path_create);
3627
3628 inline struct dentry *user_path_create(int dfd, const char __user *pathname,
3629                                 struct path *path, unsigned int lookup_flags)
3630 {
3631         return filename_create(dfd, getname(pathname), path, lookup_flags);
3632 }
3633 EXPORT_SYMBOL(user_path_create);
3634
3635 int vfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
3636 {
3637         int error = may_create(dir, dentry);
3638
3639         if (error)
3640                 return error;
3641
3642         if ((S_ISCHR(mode) || S_ISBLK(mode)) && !capable(CAP_MKNOD))
3643                 return -EPERM;
3644
3645         if (!dir->i_op->mknod)
3646                 return -EPERM;
3647
3648         error = devcgroup_inode_mknod(mode, dev);
3649         if (error)
3650                 return error;
3651
3652         error = security_inode_mknod(dir, dentry, mode, dev);
3653         if (error)
3654                 return error;
3655
3656         error = dir->i_op->mknod(dir, dentry, mode, dev);
3657         if (!error)
3658                 fsnotify_create(dir, dentry);
3659         return error;
3660 }
3661 EXPORT_SYMBOL(vfs_mknod);
3662
3663 static int may_mknod(umode_t mode)
3664 {
3665         switch (mode & S_IFMT) {
3666         case S_IFREG:
3667         case S_IFCHR:
3668         case S_IFBLK:
3669         case S_IFIFO:
3670         case S_IFSOCK:
3671         case 0: /* zero mode translates to S_IFREG */
3672                 return 0;
3673         case S_IFDIR:
3674                 return -EPERM;
3675         default:
3676                 return -EINVAL;
3677         }
3678 }
3679
3680 SYSCALL_DEFINE4(mknodat, int, dfd, const char __user *, filename, umode_t, mode,
3681                 unsigned, dev)
3682 {
3683         struct dentry *dentry;
3684         struct path path;
3685         int error;
3686         unsigned int lookup_flags = 0;
3687
3688         error = may_mknod(mode);
3689         if (error)
3690                 return error;
3691 retry:
3692         dentry = user_path_create(dfd, filename, &path, lookup_flags);
3693         if (IS_ERR(dentry))
3694                 return PTR_ERR(dentry);
3695
3696         if (!IS_POSIXACL(path.dentry->d_inode))
3697                 mode &= ~current_umask();
3698         error = security_path_mknod(&path, dentry, mode, dev);
3699         if (error)
3700                 goto out;
3701         switch (mode & S_IFMT) {
3702                 case 0: case S_IFREG:
3703                         error = vfs_create(path.dentry->d_inode,dentry,mode,true);
3704                         if (!error)
3705                                 ima_post_path_mknod(dentry);
3706                         break;
3707                 case S_IFCHR: case S_IFBLK:
3708                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,
3709                                         new_decode_dev(dev));
3710                         break;
3711                 case S_IFIFO: case S_IFSOCK:
3712                         error = vfs_mknod(path.dentry->d_inode,dentry,mode,0);
3713                         break;
3714         }
3715 out:
3716         done_path_create(&path, dentry);
3717         if (retry_estale(error, lookup_flags)) {
3718                 lookup_flags |= LOOKUP_REVAL;
3719                 goto retry;
3720         }
3721         return error;
3722 }
3723
3724 SYSCALL_DEFINE3(mknod, const char __user *, filename, umode_t, mode, unsigned, dev)
3725 {
3726         return sys_mknodat(AT_FDCWD, filename, mode, dev);
3727 }
3728
3729 int vfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3730 {
3731         int error = may_create(dir, dentry);
3732         unsigned max_links = dir->i_sb->s_max_links;
3733
3734         if (error)
3735                 return error;
3736
3737         if (!dir->i_op->mkdir)
3738                 return -EPERM;
3739
3740         mode &= (S_IRWXUGO|S_ISVTX);
3741         error = security_inode_mkdir(dir, dentry, mode);
3742         if (error)
3743                 return error;
3744
3745         if (max_links && dir->i_nlink >= max_links)
3746                 return -EMLINK;
3747
3748         error = dir->i_op->mkdir(dir, dentry, mode);
3749         if (!error)
3750                 fsnotify_mkdir(dir, dentry);
3751         return error;
3752 }
3753 EXPORT_SYMBOL(vfs_mkdir);
3754
3755 SYSCALL_DEFINE3(mkdirat, int, dfd, const char __user *, pathname, umode_t, mode)
3756 {
3757         struct dentry *dentry;
3758         struct path path;
3759         int error;
3760         unsigned int lookup_flags = LOOKUP_DIRECTORY;
3761
3762 retry:
3763         dentry = user_path_create(dfd, pathname, &path, lookup_flags);
3764         if (IS_ERR(dentry))
3765                 return PTR_ERR(dentry);
3766
3767         if (!IS_POSIXACL(path.dentry->d_inode))
3768                 mode &= ~current_umask();
3769         error = security_path_mkdir(&path, dentry, mode);
3770         if (!error)
3771                 error = vfs_mkdir(path.dentry->d_inode, dentry, mode);
3772         done_path_create(&path, dentry);
3773         if (retry_estale(error, lookup_flags)) {
3774                 lookup_flags |= LOOKUP_REVAL;
3775                 goto retry;
3776         }
3777         return error;
3778 }
3779
3780 SYSCALL_DEFINE2(mkdir, const char __user *, pathname, umode_t, mode)
3781 {
3782         return sys_mkdirat(AT_FDCWD, pathname, mode);
3783 }
3784
3785 int vfs_rmdir(struct inode *dir, struct dentry *dentry)
3786 {
3787         int error = may_delete(dir, dentry, 1);
3788
3789         if (error)
3790                 return error;
3791
3792         if (!dir->i_op->rmdir)
3793                 return -EPERM;
3794
3795         dget(dentry);
3796         inode_lock(dentry->d_inode);
3797
3798         error = -EBUSY;
3799         if (is_local_mountpoint(dentry))
3800                 goto out;
3801
3802         error = security_inode_rmdir(dir, dentry);
3803         if (error)
3804                 goto out;
3805
3806         shrink_dcache_parent(dentry);
3807         error = dir->i_op->rmdir(dir, dentry);
3808         if (error)
3809                 goto out;
3810
3811         dentry->d_inode->i_flags |= S_DEAD;
3812         dont_mount(dentry);
3813         detach_mounts(dentry);
3814
3815 out:
3816         inode_unlock(dentry->d_inode);
3817         dput(dentry);
3818         if (!error)
3819                 d_delete(dentry);
3820         return error;
3821 }
3822 EXPORT_SYMBOL(vfs_rmdir);
3823
3824 static long do_rmdir(int dfd, const char __user *pathname)
3825 {
3826         int error = 0;
3827         struct filename *name;
3828         struct dentry *dentry;
3829         struct path path;
3830         struct qstr last;
3831         int type;
3832         unsigned int lookup_flags = 0;
3833 retry:
3834         name = user_path_parent(dfd, pathname,
3835                                 &path, &last, &type, lookup_flags);
3836         if (IS_ERR(name))
3837                 return PTR_ERR(name);
3838
3839         switch (type) {
3840         case LAST_DOTDOT:
3841                 error = -ENOTEMPTY;
3842                 goto exit1;
3843         case LAST_DOT:
3844                 error = -EINVAL;
3845                 goto exit1;
3846         case LAST_ROOT:
3847                 error = -EBUSY;
3848                 goto exit1;
3849         }
3850
3851         error = mnt_want_write(path.mnt);
3852         if (error)
3853                 goto exit1;
3854
3855         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3856         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3857         error = PTR_ERR(dentry);
3858         if (IS_ERR(dentry))
3859                 goto exit2;
3860         if (!dentry->d_inode) {
3861                 error = -ENOENT;
3862                 goto exit3;
3863         }
3864         error = security_path_rmdir(&path, dentry);
3865         if (error)
3866                 goto exit3;
3867         error = vfs_rmdir(path.dentry->d_inode, dentry);
3868 exit3:
3869         dput(dentry);
3870 exit2:
3871         inode_unlock(path.dentry->d_inode);
3872         mnt_drop_write(path.mnt);
3873 exit1:
3874         path_put(&path);
3875         putname(name);
3876         if (retry_estale(error, lookup_flags)) {
3877                 lookup_flags |= LOOKUP_REVAL;
3878                 goto retry;
3879         }
3880         return error;
3881 }
3882
3883 SYSCALL_DEFINE1(rmdir, const char __user *, pathname)
3884 {
3885         return do_rmdir(AT_FDCWD, pathname);
3886 }
3887
3888 /**
3889  * vfs_unlink - unlink a filesystem object
3890  * @dir:        parent directory
3891  * @dentry:     victim
3892  * @delegated_inode: returns victim inode, if the inode is delegated.
3893  *
3894  * The caller must hold dir->i_mutex.
3895  *
3896  * If vfs_unlink discovers a delegation, it will return -EWOULDBLOCK and
3897  * return a reference to the inode in delegated_inode.  The caller
3898  * should then break the delegation on that inode and retry.  Because
3899  * breaking a delegation may take a long time, the caller should drop
3900  * dir->i_mutex before doing so.
3901  *
3902  * Alternatively, a caller may pass NULL for delegated_inode.  This may
3903  * be appropriate for callers that expect the underlying filesystem not
3904  * to be NFS exported.
3905  */
3906 int vfs_unlink(struct inode *dir, struct dentry *dentry, struct inode **delegated_inode)
3907 {
3908         struct inode *target = dentry->d_inode;
3909         int error = may_delete(dir, dentry, 0);
3910
3911         if (error)
3912                 return error;
3913
3914         if (!dir->i_op->unlink)
3915                 return -EPERM;
3916
3917         inode_lock(target);
3918         if (is_local_mountpoint(dentry))
3919                 error = -EBUSY;
3920         else {
3921                 error = security_inode_unlink(dir, dentry);
3922                 if (!error) {
3923                         error = try_break_deleg(target, delegated_inode);
3924                         if (error)
3925                                 goto out;
3926                         error = dir->i_op->unlink(dir, dentry);
3927                         if (!error) {
3928                                 dont_mount(dentry);
3929                                 detach_mounts(dentry);
3930                         }
3931                 }
3932         }
3933 out:
3934         inode_unlock(target);
3935
3936         /* We don't d_delete() NFS sillyrenamed files--they still exist. */
3937         if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
3938                 fsnotify_link_count(target);
3939                 d_delete(dentry);
3940         }
3941
3942         return error;
3943 }
3944 EXPORT_SYMBOL(vfs_unlink);
3945
3946 /*
3947  * Make sure that the actual truncation of the file will occur outside its
3948  * directory's i_mutex.  Truncate can take a long time if there is a lot of
3949  * writeout happening, and we don't want to prevent access to the directory
3950  * while waiting on the I/O.
3951  */
3952 static long do_unlinkat(int dfd, const char __user *pathname)
3953 {
3954         int error;
3955         struct filename *name;
3956         struct dentry *dentry;
3957         struct path path;
3958         struct qstr last;
3959         int type;
3960         struct inode *inode = NULL;
3961         struct inode *delegated_inode = NULL;
3962         unsigned int lookup_flags = 0;
3963 retry:
3964         name = user_path_parent(dfd, pathname,
3965                                 &path, &last, &type, lookup_flags);
3966         if (IS_ERR(name))
3967                 return PTR_ERR(name);
3968
3969         error = -EISDIR;
3970         if (type != LAST_NORM)
3971                 goto exit1;
3972
3973         error = mnt_want_write(path.mnt);
3974         if (error)
3975                 goto exit1;
3976 retry_deleg:
3977         inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
3978         dentry = __lookup_hash(&last, path.dentry, lookup_flags);
3979         error = PTR_ERR(dentry);
3980         if (!IS_ERR(dentry)) {
3981                 /* Why not before? Because we want correct error value */
3982                 if (last.name[last.len])
3983                         goto slashes;
3984                 inode = dentry->d_inode;
3985                 if (d_is_negative(dentry))
3986                         goto slashes;
3987                 ihold(inode);
3988                 error = security_path_unlink(&path, dentry);
3989                 if (error)
3990                         goto exit2;
3991                 error = vfs_unlink(path.dentry->d_inode, dentry, &delegated_inode);
3992 exit2:
3993                 dput(dentry);
3994         }
3995         inode_unlock(path.dentry->d_inode);
3996         if (inode)
3997                 iput(inode);    /* truncate the inode here */
3998         inode = NULL;
3999         if (delegated_inode) {
4000                 error = break_deleg_wait(&delegated_inode);
4001                 if (!error)
4002                         goto retry_deleg;
4003         }
4004         mnt_drop_write(path.mnt);
4005 exit1:
4006         path_put(&path);
4007         putname(name);
4008         if (retry_estale(error, lookup_flags)) {
4009                 lookup_flags |= LOOKUP_REVAL;
4010                 inode = NULL;
4011                 goto retry;
4012         }
4013         return error;
4014
4015 slashes:
4016         if (d_is_negative(dentry))
4017                 error = -ENOENT;
4018         else if (d_is_dir(dentry))
4019                 error = -EISDIR;
4020         else
4021                 error = -ENOTDIR;
4022         goto exit2;
4023 }
4024
4025 SYSCALL_DEFINE3(unlinkat, int, dfd, const char __user *, pathname, int, flag)
4026 {
4027         if ((flag & ~AT_REMOVEDIR) != 0)
4028                 return -EINVAL;
4029
4030         if (flag & AT_REMOVEDIR)
4031                 return do_rmdir(dfd, pathname);
4032
4033         return do_unlinkat(dfd, pathname);
4034 }
4035
4036 SYSCALL_DEFINE1(unlink, const char __user *, pathname)
4037 {
4038         return do_unlinkat(AT_FDCWD, pathname);
4039 }
4040
4041 int vfs_symlink(struct inode *dir, struct dentry *dentry, const char *oldname)
4042 {
4043         int error = may_create(dir, dentry);
4044
4045         if (error)
4046                 return error;
4047
4048         if (!dir->i_op->symlink)
4049                 return -EPERM;
4050
4051         error = security_inode_symlink(dir, dentry, oldname);
4052         if (error)
4053                 return error;
4054
4055         error = dir->i_op->symlink(dir, dentry, oldname);
4056         if (!error)
4057                 fsnotify_create(dir, dentry);
4058         return error;
4059 }
4060 EXPORT_SYMBOL(vfs_symlink);
4061
4062 SYSCALL_DEFINE3(symlinkat, const char __user *, oldname,
4063                 int, newdfd, const char __user *, newname)
4064 {
4065         int error;
4066         struct filename *from;
4067         struct dentry *dentry;
4068         struct path path;
4069         unsigned int lookup_flags = 0;
4070
4071         from = getname(oldname);
4072         if (IS_ERR(from))
4073                 return PTR_ERR(from);
4074 retry:
4075         dentry = user_path_create(newdfd, newname, &path, lookup_flags);
4076         error = PTR_ERR(dentry);
4077         if (IS_ERR(dentry))
4078                 goto out_putname;
4079
4080         error = security_path_symlink(&path, dentry, from->name);
4081         if (!error)
4082                 error = vfs_symlink(path.dentry->d_inode, dentry, from->name);
4083         done_path_create(&path, dentry);
4084         if (retry_estale(error, lookup_flags)) {
4085                 lookup_flags |= LOOKUP_REVAL;
4086                 goto retry;
4087         }
4088 out_putname:
4089         putname(from);
4090         return error;
4091 }
4092
4093 SYSCALL_DEFINE2(symlink, const char __user *, oldname, const char __user *, newname)
4094 {
4095         return sys_symlinkat(oldname, AT_FDCWD, newname);
4096 }
4097
4098 /**
4099  * vfs_link - create a new link
4100  * @old_dentry: object to be linked
4101  * @dir:        new parent
4102  * @new_dentry: where to create the new link
4103  * @delegated_inode: returns inode needing a delegation break
4104  *
4105  * The caller must hold dir->i_mutex
4106  *
4107  * If vfs_link discovers a delegation on the to-be-linked file in need
4108  * of breaking, it will return -EWOULDBLOCK and return a reference to the
4109  * inode in delegated_inode.  The caller should then break the delegation
4110  * and retry.  Because breaking a delegation may take a long time, the
4111  * caller should drop the i_mutex before doing so.
4112  *
4113  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4114  * be appropriate for callers that expect the underlying filesystem not
4115  * to be NFS exported.
4116  */
4117 int vfs_link(struct dentry *old_dentry, struct inode *dir, struct dentry *new_dentry, struct inode **delegated_inode)
4118 {
4119         struct inode *inode = old_dentry->d_inode;
4120         unsigned max_links = dir->i_sb->s_max_links;
4121         int error;
4122
4123         if (!inode)
4124                 return -ENOENT;
4125
4126         error = may_create(dir, new_dentry);
4127         if (error)
4128                 return error;
4129
4130         if (dir->i_sb != inode->i_sb)
4131                 return -EXDEV;
4132
4133         /*
4134          * A link to an append-only or immutable file cannot be created.
4135          */
4136         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4137                 return -EPERM;
4138         if (!dir->i_op->link)
4139                 return -EPERM;
4140         if (S_ISDIR(inode->i_mode))
4141                 return -EPERM;
4142
4143         error = security_inode_link(old_dentry, dir, new_dentry);
4144         if (error)
4145                 return error;
4146
4147         inode_lock(inode);
4148         /* Make sure we don't allow creating hardlink to an unlinked file */
4149         if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))
4150                 error =  -ENOENT;
4151         else if (max_links && inode->i_nlink >= max_links)
4152                 error = -EMLINK;
4153         else {
4154                 error = try_break_deleg(inode, delegated_inode);
4155                 if (!error)
4156                         error = dir->i_op->link(old_dentry, dir, new_dentry);
4157         }
4158
4159         if (!error && (inode->i_state & I_LINKABLE)) {
4160                 spin_lock(&inode->i_lock);
4161                 inode->i_state &= ~I_LINKABLE;
4162                 spin_unlock(&inode->i_lock);
4163         }
4164         inode_unlock(inode);
4165         if (!error)
4166                 fsnotify_link(dir, inode, new_dentry);
4167         return error;
4168 }
4169 EXPORT_SYMBOL(vfs_link);
4170
4171 /*
4172  * Hardlinks are often used in delicate situations.  We avoid
4173  * security-related surprises by not following symlinks on the
4174  * newname.  --KAB
4175  *
4176  * We don't follow them on the oldname either to be compatible
4177  * with linux 2.0, and to avoid hard-linking to directories
4178  * and other special files.  --ADM
4179  */
4180 SYSCALL_DEFINE5(linkat, int, olddfd, const char __user *, oldname,
4181                 int, newdfd, const char __user *, newname, int, flags)
4182 {
4183         struct dentry *new_dentry;
4184         struct path old_path, new_path;
4185         struct inode *delegated_inode = NULL;
4186         int how = 0;
4187         int error;
4188
4189         if ((flags & ~(AT_SYMLINK_FOLLOW | AT_EMPTY_PATH)) != 0)
4190                 return -EINVAL;
4191         /*
4192          * To use null names we require CAP_DAC_READ_SEARCH
4193          * This ensures that not everyone will be able to create
4194          * handlink using the passed filedescriptor.
4195          */
4196         if (flags & AT_EMPTY_PATH) {
4197                 if (!capable(CAP_DAC_READ_SEARCH))
4198                         return -ENOENT;
4199                 how = LOOKUP_EMPTY;
4200         }
4201
4202         if (flags & AT_SYMLINK_FOLLOW)
4203                 how |= LOOKUP_FOLLOW;
4204 retry:
4205         error = user_path_at(olddfd, oldname, how, &old_path);
4206         if (error)
4207                 return error;
4208
4209         new_dentry = user_path_create(newdfd, newname, &new_path,
4210                                         (how & LOOKUP_REVAL));
4211         error = PTR_ERR(new_dentry);
4212         if (IS_ERR(new_dentry))
4213                 goto out;
4214
4215         error = -EXDEV;
4216         if (old_path.mnt != new_path.mnt)
4217                 goto out_dput;
4218         error = may_linkat(&old_path);
4219         if (unlikely(error))
4220                 goto out_dput;
4221         error = security_path_link(old_path.dentry, &new_path, new_dentry);
4222         if (error)
4223                 goto out_dput;
4224         error = vfs_link(old_path.dentry, new_path.dentry->d_inode, new_dentry, &delegated_inode);
4225 out_dput:
4226         done_path_create(&new_path, new_dentry);
4227         if (delegated_inode) {
4228                 error = break_deleg_wait(&delegated_inode);
4229                 if (!error) {
4230                         path_put(&old_path);
4231                         goto retry;
4232                 }
4233         }
4234         if (retry_estale(error, how)) {
4235                 path_put(&old_path);
4236                 how |= LOOKUP_REVAL;
4237                 goto retry;
4238         }
4239 out:
4240         path_put(&old_path);
4241
4242         return error;
4243 }
4244
4245 SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname)
4246 {
4247         return sys_linkat(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4248 }
4249
4250 /**
4251  * vfs_rename - rename a filesystem object
4252  * @old_dir:    parent of source
4253  * @old_dentry: source
4254  * @new_dir:    parent of destination
4255  * @new_dentry: destination
4256  * @delegated_inode: returns an inode needing a delegation break
4257  * @flags:      rename flags
4258  *
4259  * The caller must hold multiple mutexes--see lock_rename()).
4260  *
4261  * If vfs_rename discovers a delegation in need of breaking at either
4262  * the source or destination, it will return -EWOULDBLOCK and return a
4263  * reference to the inode in delegated_inode.  The caller should then
4264  * break the delegation and retry.  Because breaking a delegation may
4265  * take a long time, the caller should drop all locks before doing
4266  * so.
4267  *
4268  * Alternatively, a caller may pass NULL for delegated_inode.  This may
4269  * be appropriate for callers that expect the underlying filesystem not
4270  * to be NFS exported.
4271  *
4272  * The worst of all namespace operations - renaming directory. "Perverted"
4273  * doesn't even start to describe it. Somebody in UCB had a heck of a trip...
4274  * Problems:
4275  *      a) we can get into loop creation.
4276  *      b) race potential - two innocent renames can create a loop together.
4277  *         That's where 4.4 screws up. Current fix: serialization on
4278  *         sb->s_vfs_rename_mutex. We might be more accurate, but that's another
4279  *         story.
4280  *      c) we have to lock _four_ objects - parents and victim (if it exists),
4281  *         and source (if it is not a directory).
4282  *         And that - after we got ->i_mutex on parents (until then we don't know
4283  *         whether the target exists).  Solution: try to be smart with locking
4284  *         order for inodes.  We rely on the fact that tree topology may change
4285  *         only under ->s_vfs_rename_mutex _and_ that parent of the object we
4286  *         move will be locked.  Thus we can rank directories by the tree
4287  *         (ancestors first) and rank all non-directories after them.
4288  *         That works since everybody except rename does "lock parent, lookup,
4289  *         lock child" and rename is under ->s_vfs_rename_mutex.
4290  *         HOWEVER, it relies on the assumption that any object with ->lookup()
4291  *         has no more than 1 dentry.  If "hybrid" objects will ever appear,
4292  *         we'd better make sure that there's no link(2) for them.
4293  *      d) conversion from fhandle to dentry may come in the wrong moment - when
4294  *         we are removing the target. Solution: we will have to grab ->i_mutex
4295  *         in the fhandle_to_dentry code. [FIXME - current nfsfh.c relies on
4296  *         ->i_mutex on parents, which works but leads to some truly excessive
4297  *         locking].
4298  */
4299 int vfs_rename(struct inode *old_dir, struct dentry *old_dentry,
4300                struct inode *new_dir, struct dentry *new_dentry,
4301                struct inode **delegated_inode, unsigned int flags)
4302 {
4303         int error;
4304         bool is_dir = d_is_dir(old_dentry);
4305         const unsigned char *old_name;
4306         struct inode *source = old_dentry->d_inode;
4307         struct inode *target = new_dentry->d_inode;
4308         bool new_is_dir = false;
4309         unsigned max_links = new_dir->i_sb->s_max_links;
4310
4311         /*
4312          * Check source == target.
4313          * On overlayfs need to look at underlying inodes.
4314          */
4315         if (d_real_inode(old_dentry) == d_real_inode(new_dentry))
4316                 return 0;
4317
4318         error = may_delete(old_dir, old_dentry, is_dir);
4319         if (error)
4320                 return error;
4321
4322         if (!target) {
4323                 error = may_create(new_dir, new_dentry);
4324         } else {
4325                 new_is_dir = d_is_dir(new_dentry);
4326
4327                 if (!(flags & RENAME_EXCHANGE))
4328                         error = may_delete(new_dir, new_dentry, is_dir);
4329                 else
4330                         error = may_delete(new_dir, new_dentry, new_is_dir);
4331         }
4332         if (error)
4333                 return error;
4334
4335         if (!old_dir->i_op->rename && !old_dir->i_op->rename2)
4336                 return -EPERM;
4337
4338         if (flags && !old_dir->i_op->rename2)
4339                 return -EINVAL;
4340
4341         /*
4342          * If we are going to change the parent - check write permissions,
4343          * we'll need to flip '..'.
4344          */
4345         if (new_dir != old_dir) {
4346                 if (is_dir) {
4347                         error = inode_permission(source, MAY_WRITE);
4348                         if (error)
4349                                 return error;
4350                 }
4351                 if ((flags & RENAME_EXCHANGE) && new_is_dir) {
4352                         error = inode_permission(target, MAY_WRITE);
4353                         if (error)
4354                                 return error;
4355                 }
4356         }
4357
4358         error = security_inode_rename(old_dir, old_dentry, new_dir, new_dentry,
4359                                       flags);
4360         if (error)
4361                 return error;
4362
4363         old_name = fsnotify_oldname_init(old_dentry->d_name.name);
4364         dget(new_dentry);
4365         if (!is_dir || (flags & RENAME_EXCHANGE))
4366                 lock_two_nondirectories(source, target);
4367         else if (target)
4368                 inode_lock(target);
4369
4370         error = -EBUSY;
4371         if (is_local_mountpoint(old_dentry) || is_local_mountpoint(new_dentry))
4372                 goto out;
4373
4374         if (max_links && new_dir != old_dir) {
4375                 error = -EMLINK;
4376                 if (is_dir && !new_is_dir && new_dir->i_nlink >= max_links)
4377                         goto out;
4378                 if ((flags & RENAME_EXCHANGE) && !is_dir && new_is_dir &&
4379                     old_dir->i_nlink >= max_links)
4380                         goto out;
4381         }
4382         if (is_dir && !(flags & RENAME_EXCHANGE) && target)
4383                 shrink_dcache_parent(new_dentry);
4384         if (!is_dir) {
4385                 error = try_break_deleg(source, delegated_inode);
4386                 if (error)
4387                         goto out;
4388         }
4389         if (target && !new_is_dir) {
4390                 error = try_break_deleg(target, delegated_inode);
4391                 if (error)
4392                         goto out;
4393         }
4394         if (!old_dir->i_op->rename2) {
4395                 error = old_dir->i_op->rename(old_dir, old_dentry,
4396                                               new_dir, new_dentry);
4397         } else {
4398                 WARN_ON(old_dir->i_op->rename != NULL);
4399                 error = old_dir->i_op->rename2(old_dir, old_dentry,
4400                                                new_dir, new_dentry, flags);
4401         }
4402         if (error)
4403                 goto out;
4404
4405         if (!(flags & RENAME_EXCHANGE) && target) {
4406                 if (is_dir)
4407                         target->i_flags |= S_DEAD;
4408                 dont_mount(new_dentry);
4409                 detach_mounts(new_dentry);
4410         }
4411         if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE)) {
4412                 if (!(flags & RENAME_EXCHANGE))
4413                         d_move(old_dentry, new_dentry);
4414                 else
4415                         d_exchange(old_dentry, new_dentry);
4416         }
4417 out:
4418         if (!is_dir || (flags & RENAME_EXCHANGE))
4419                 unlock_two_nondirectories(source, target);
4420         else if (target)
4421                 inode_unlock(target);
4422         dput(new_dentry);
4423         if (!error) {
4424                 fsnotify_move(old_dir, new_dir, old_name, is_dir,
4425                               !(flags & RENAME_EXCHANGE) ? target : NULL, old_dentry);
4426                 if (flags & RENAME_EXCHANGE) {
4427                         fsnotify_move(new_dir, old_dir, old_dentry->d_name.name,
4428                                       new_is_dir, NULL, new_dentry);
4429                 }
4430         }
4431         fsnotify_oldname_free(old_name);
4432
4433         return error;
4434 }
4435 EXPORT_SYMBOL(vfs_rename);
4436
4437 SYSCALL_DEFINE5(renameat2, int, olddfd, const char __user *, oldname,
4438                 int, newdfd, const char __user *, newname, unsigned int, flags)
4439 {
4440         struct dentry *old_dentry, *new_dentry;
4441         struct dentry *trap;
4442         struct path old_path, new_path;
4443         struct qstr old_last, new_last;
4444         int old_type, new_type;
4445         struct inode *delegated_inode = NULL;
4446         struct filename *from;
4447         struct filename *to;
4448         unsigned int lookup_flags = 0, target_flags = LOOKUP_RENAME_TARGET;
4449         bool should_retry = false;
4450         int error;
4451
4452         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
4453                 return -EINVAL;
4454
4455         if ((flags & (RENAME_NOREPLACE | RENAME_WHITEOUT)) &&
4456             (flags & RENAME_EXCHANGE))
4457                 return -EINVAL;
4458
4459         if ((flags & RENAME_WHITEOUT) && !capable(CAP_MKNOD))
4460                 return -EPERM;
4461
4462         if (flags & RENAME_EXCHANGE)
4463                 target_flags = 0;
4464
4465 retry:
4466         from = user_path_parent(olddfd, oldname,
4467                                 &old_path, &old_last, &old_type, lookup_flags);
4468         if (IS_ERR(from)) {
4469                 error = PTR_ERR(from);
4470                 goto exit;
4471         }
4472
4473         to = user_path_parent(newdfd, newname,
4474                                 &new_path, &new_last, &new_type, lookup_flags);
4475         if (IS_ERR(to)) {
4476                 error = PTR_ERR(to);
4477                 goto exit1;
4478         }
4479
4480         error = -EXDEV;
4481         if (old_path.mnt != new_path.mnt)
4482                 goto exit2;
4483
4484         error = -EBUSY;
4485         if (old_type != LAST_NORM)
4486                 goto exit2;
4487
4488         if (flags & RENAME_NOREPLACE)
4489                 error = -EEXIST;
4490         if (new_type != LAST_NORM)
4491                 goto exit2;
4492
4493         error = mnt_want_write(old_path.mnt);
4494         if (error)
4495                 goto exit2;
4496
4497 retry_deleg:
4498         trap = lock_rename(new_path.dentry, old_path.dentry);
4499
4500         old_dentry = __lookup_hash(&old_last, old_path.dentry, lookup_flags);
4501         error = PTR_ERR(old_dentry);
4502         if (IS_ERR(old_dentry))
4503                 goto exit3;
4504         /* source must exist */
4505         error = -ENOENT;
4506         if (d_is_negative(old_dentry))
4507                 goto exit4;
4508         new_dentry = __lookup_hash(&new_last, new_path.dentry, lookup_flags | target_flags);
4509         error = PTR_ERR(new_dentry);
4510         if (IS_ERR(new_dentry))
4511                 goto exit4;
4512         error = -EEXIST;
4513         if ((flags & RENAME_NOREPLACE) && d_is_positive(new_dentry))
4514                 goto exit5;
4515         if (flags & RENAME_EXCHANGE) {
4516                 error = -ENOENT;
4517                 if (d_is_negative(new_dentry))
4518                         goto exit5;
4519
4520                 if (!d_is_dir(new_dentry)) {
4521                         error = -ENOTDIR;
4522                         if (new_last.name[new_last.len])
4523                                 goto exit5;
4524                 }
4525         }
4526         /* unless the source is a directory trailing slashes give -ENOTDIR */
4527         if (!d_is_dir(old_dentry)) {
4528                 error = -ENOTDIR;
4529                 if (old_last.name[old_last.len])
4530                         goto exit5;
4531                 if (!(flags & RENAME_EXCHANGE) && new_last.name[new_last.len])
4532                         goto exit5;
4533         }
4534         /* source should not be ancestor of target */
4535         error = -EINVAL;
4536         if (old_dentry == trap)
4537                 goto exit5;
4538         /* target should not be an ancestor of source */
4539         if (!(flags & RENAME_EXCHANGE))
4540                 error = -ENOTEMPTY;
4541         if (new_dentry == trap)
4542                 goto exit5;
4543
4544         error = security_path_rename(&old_path, old_dentry,
4545                                      &new_path, new_dentry, flags);
4546         if (error)
4547                 goto exit5;
4548         error = vfs_rename(old_path.dentry->d_inode, old_dentry,
4549                            new_path.dentry->d_inode, new_dentry,
4550                            &delegated_inode, flags);
4551 exit5:
4552         dput(new_dentry);
4553 exit4:
4554         dput(old_dentry);
4555 exit3:
4556         unlock_rename(new_path.dentry, old_path.dentry);
4557         if (delegated_inode) {
4558                 error = break_deleg_wait(&delegated_inode);
4559                 if (!error)
4560                         goto retry_deleg;
4561         }
4562         mnt_drop_write(old_path.mnt);
4563 exit2:
4564         if (retry_estale(error, lookup_flags))
4565                 should_retry = true;
4566         path_put(&new_path);
4567         putname(to);
4568 exit1:
4569         path_put(&old_path);
4570         putname(from);
4571         if (should_retry) {
4572                 should_retry = false;
4573                 lookup_flags |= LOOKUP_REVAL;
4574                 goto retry;
4575         }
4576 exit:
4577         return error;
4578 }
4579
4580 SYSCALL_DEFINE4(renameat, int, olddfd, const char __user *, oldname,
4581                 int, newdfd, const char __user *, newname)
4582 {
4583         return sys_renameat2(olddfd, oldname, newdfd, newname, 0);
4584 }
4585
4586 SYSCALL_DEFINE2(rename, const char __user *, oldname, const char __user *, newname)
4587 {
4588         return sys_renameat2(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
4589 }
4590
4591 int vfs_whiteout(struct inode *dir, struct dentry *dentry)
4592 {
4593         int error = may_create(dir, dentry);
4594         if (error)
4595                 return error;
4596
4597         if (!dir->i_op->mknod)
4598                 return -EPERM;
4599
4600         return dir->i_op->mknod(dir, dentry,
4601                                 S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
4602 }
4603 EXPORT_SYMBOL(vfs_whiteout);
4604
4605 int readlink_copy(char __user *buffer, int buflen, const char *link)
4606 {
4607         int len = PTR_ERR(link);
4608         if (IS_ERR(link))
4609                 goto out;
4610
4611         len = strlen(link);
4612         if (len > (unsigned) buflen)
4613                 len = buflen;
4614         if (copy_to_user(buffer, link, len))
4615                 len = -EFAULT;
4616 out:
4617         return len;
4618 }
4619
4620 /*
4621  * A helper for ->readlink().  This should be used *ONLY* for symlinks that
4622  * have ->get_link() not calling nd_jump_link().  Using (or not using) it
4623  * for any given inode is up to filesystem.
4624  */
4625 int generic_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4626 {
4627         DEFINE_DELAYED_CALL(done);
4628         struct inode *inode = d_inode(dentry);
4629         const char *link = inode->i_link;
4630         int res;
4631
4632         if (!link) {
4633                 link = inode->i_op->get_link(dentry, inode, &done);
4634                 if (IS_ERR(link))
4635                         return PTR_ERR(link);
4636         }
4637         res = readlink_copy(buffer, buflen, link);
4638         do_delayed_call(&done);
4639         return res;
4640 }
4641 EXPORT_SYMBOL(generic_readlink);
4642
4643 /* get the link contents into pagecache */
4644 const char *page_get_link(struct dentry *dentry, struct inode *inode,
4645                           struct delayed_call *callback)
4646 {
4647         char *kaddr;
4648         struct page *page;
4649         struct address_space *mapping = inode->i_mapping;
4650
4651         if (!dentry) {
4652                 page = find_get_page(mapping, 0);
4653                 if (!page)
4654                         return ERR_PTR(-ECHILD);
4655                 if (!PageUptodate(page)) {
4656                         put_page(page);
4657                         return ERR_PTR(-ECHILD);
4658                 }
4659         } else {
4660                 page = read_mapping_page(mapping, 0, NULL);
4661                 if (IS_ERR(page))
4662                         return (char*)page;
4663         }
4664         set_delayed_call(callback, page_put_link, page);
4665         BUG_ON(mapping_gfp_mask(mapping) & __GFP_HIGHMEM);
4666         kaddr = page_address(page);
4667         nd_terminate_link(kaddr, inode->i_size, PAGE_SIZE - 1);
4668         return kaddr;
4669 }
4670
4671 EXPORT_SYMBOL(page_get_link);
4672
4673 void page_put_link(void *arg)
4674 {
4675         put_page(arg);
4676 }
4677 EXPORT_SYMBOL(page_put_link);
4678
4679 int page_readlink(struct dentry *dentry, char __user *buffer, int buflen)
4680 {
4681         DEFINE_DELAYED_CALL(done);
4682         int res = readlink_copy(buffer, buflen,
4683                                 page_get_link(dentry, d_inode(dentry),
4684                                               &done));
4685         do_delayed_call(&done);
4686         return res;
4687 }
4688 EXPORT_SYMBOL(page_readlink);
4689
4690 /*
4691  * The nofs argument instructs pagecache_write_begin to pass AOP_FLAG_NOFS
4692  */
4693 int __page_symlink(struct inode *inode, const char *symname, int len, int nofs)
4694 {
4695         struct address_space *mapping = inode->i_mapping;
4696         struct page *page;
4697         void *fsdata;
4698         int err;
4699         unsigned int flags = AOP_FLAG_UNINTERRUPTIBLE;
4700         if (nofs)
4701                 flags |= AOP_FLAG_NOFS;
4702
4703 retry:
4704         err = pagecache_write_begin(NULL, mapping, 0, len-1,
4705                                 flags, &page, &fsdata);
4706         if (err)
4707                 goto fail;
4708
4709         memcpy(page_address(page), symname, len-1);
4710
4711         err = pagecache_write_end(NULL, mapping, 0, len-1, len-1,
4712                                                         page, fsdata);
4713         if (err < 0)
4714                 goto fail;
4715         if (err < len-1)
4716                 goto retry;
4717
4718         mark_inode_dirty(inode);
4719         return 0;
4720 fail:
4721         return err;
4722 }
4723 EXPORT_SYMBOL(__page_symlink);
4724
4725 int page_symlink(struct inode *inode, const char *symname, int len)
4726 {
4727         return __page_symlink(inode, symname, len,
4728                         !mapping_gfp_constraint(inode->i_mapping, __GFP_FS));
4729 }
4730 EXPORT_SYMBOL(page_symlink);
4731
4732 const struct inode_operations page_symlink_inode_operations = {
4733         .readlink       = generic_readlink,
4734         .get_link       = page_get_link,
4735 };
4736 EXPORT_SYMBOL(page_symlink_inode_operations);