fs/overlayfs/dir.c

   1 /*
   2  *
   3  * Copyright (C) 2011 Novell Inc.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 as published by
   7  * the Free Software Foundation.
   8  */
   9
  10 #include <linux/fs.h>
  11 #include <linux/namei.h>
  12 #include <linux/xattr.h>
  13 #include <linux/security.h>
  14 #include <linux/cred.h>
  15 #include <linux/module.h>
  16 #include <linux/posix_acl.h>
  17 #include <linux/posix_acl_xattr.h>
  18 #include <linux/atomic.h>
  19 #include <linux/ratelimit.h>
  20 #include "overlayfs.h"
  21
  22 static unsigned short ovl_redirect_max = 256;
  23 module_param_named(redirect_max, ovl_redirect_max, ushort, 0644);
  24 MODULE_PARM_DESC(ovl_redirect_max,
  25                  "Maximum length of absolute redirect xattr value");
  26
  27 static int ovl_set_redirect(struct dentry *dentry, bool samedir);
  28
  29 int ovl_cleanup(struct inode *wdir, struct dentry *wdentry)
  30 {
  31         int err;
  32
  33         dget(wdentry);
  34         if (d_is_dir(wdentry))
  35                 err = ovl_do_rmdir(wdir, wdentry);
  36         else
  37                 err = ovl_do_unlink(wdir, wdentry);
  38         dput(wdentry);
  39
  40         if (err) {
  41                 pr_err("overlayfs: cleanup of '%pd2' failed (%i)\n",
  42                        wdentry, err);
  43         }
  44
  45         return err;
  46 }
  47
  48 static struct dentry *ovl_lookup_temp(struct dentry *workdir)
  49 {
  50         struct dentry *temp;
  51         char name[20];
  52         static atomic_t temp_id = ATOMIC_INIT(0);
  53
  54         /* counter is allowed to wrap, since temp dentries are ephemeral */
  55         snprintf(name, sizeof(name), "#%x", atomic_inc_return(&temp_id));
  56
  57         temp = lookup_one_len(name, workdir, strlen(name));
  58         if (!IS_ERR(temp) && temp->d_inode) {
  59                 pr_err("overlayfs: workdir/%s already exists\n", name);
  60                 dput(temp);
  61                 temp = ERR_PTR(-EIO);
  62         }
  63
  64         return temp;
  65 }
  66
  67 /* caller holds i_mutex on workdir */
  68 static struct dentry *ovl_whiteout(struct dentry *workdir)
  69 {
  70         int err;
  71         struct dentry *whiteout;
  72         struct inode *wdir = workdir->d_inode;
  73
  74         whiteout = ovl_lookup_temp(workdir);
  75         if (IS_ERR(whiteout))
  76                 return whiteout;
  77
  78         err = ovl_do_whiteout(wdir, whiteout);
  79         if (err) {
  80                 dput(whiteout);
  81                 whiteout = ERR_PTR(err);
  82         }
  83
  84         return whiteout;
  85 }
  86
  87 /* Caller must hold i_mutex on both workdir and dir */
  88 int ovl_cleanup_and_whiteout(struct dentry *workdir, struct inode *dir,
  89                              struct dentry *dentry)
  90 {
  91         struct inode *wdir = workdir->d_inode;
  92         struct dentry *whiteout;
  93         int err;
  94         int flags = 0;
  95
  96         whiteout = ovl_whiteout(workdir);
  97         err = PTR_ERR(whiteout);
  98         if (IS_ERR(whiteout))
  99                 return err;
 100
 101         if (d_is_dir(dentry))
 102                 flags = RENAME_EXCHANGE;
 103
 104         err = ovl_do_rename(wdir, whiteout, dir, dentry, flags);
 105         if (err)
 106                 goto kill_whiteout;
 107         if (flags)
 108                 ovl_cleanup(wdir, dentry);
 109
 110 out:
 111         dput(whiteout);
 112         return err;
 113
 114 kill_whiteout:
 115         ovl_cleanup(wdir, whiteout);
 116         goto out;
 117 }
 118
 119 int ovl_mkdir_real(struct inode *dir, struct dentry **newdentry, umode_t mode)
 120 {
 121         int err;
 122         struct dentry *d, *dentry = *newdentry;
 123
 124         err = ovl_do_mkdir(dir, dentry, mode);
 125         if (err)
 126                 return err;
 127
 128         if (likely(!d_unhashed(dentry)))
 129                 return 0;
 130
 131         /*
 132          * vfs_mkdir() may succeed and leave the dentry passed
 133          * to it unhashed and negative. If that happens, try to
 134          * lookup a new hashed and positive dentry.
 135          */
 136         d = lookup_one_len(dentry->d_name.name, dentry->d_parent,
 137                            dentry->d_name.len);
 138         if (IS_ERR(d)) {
 139                 pr_warn("overlayfs: failed lookup after mkdir (%pd2, err=%i).\n",
 140                         dentry, err);
 141                 return PTR_ERR(d);
 142         }
 143         dput(dentry);
 144         *newdentry = d;
 145
 146         return 0;
 147 }
 148
 149 struct dentry *ovl_create_real(struct inode *dir, struct dentry *newdentry,
 150                                struct ovl_cattr *attr)
 151 {
 152         int err;
 153
 154         if (IS_ERR(newdentry))
 155                 return newdentry;
 156
 157         err = -ESTALE;
 158         if (newdentry->d_inode)
 159                 goto out;
 160
 161         if (attr->hardlink) {
 162                 err = ovl_do_link(attr->hardlink, dir, newdentry);
 163         } else {
 164                 switch (attr->mode & S_IFMT) {
 165                 case S_IFREG:
 166                         err = ovl_do_create(dir, newdentry, attr->mode);
 167                         break;
 168
 169                 case S_IFDIR:
 170                         /* mkdir is special... */
 171                         err =  ovl_mkdir_real(dir, &newdentry, attr->mode);
 172                         break;
 173
 174                 case S_IFCHR:
 175                 case S_IFBLK:
 176                 case S_IFIFO:
 177                 case S_IFSOCK:
 178                         err = ovl_do_mknod(dir, newdentry, attr->mode,
 179                                            attr->rdev);
 180                         break;
 181
 182                 case S_IFLNK:
 183                         err = ovl_do_symlink(dir, newdentry, attr->link);
 184                         break;
 185
 186                 default:
 187                         err = -EPERM;
 188                 }
 189         }
 190         if (!err && WARN_ON(!newdentry->d_inode)) {
 191                 /*
 192                  * Not quite sure if non-instantiated dentry is legal or not.
 193                  * VFS doesn't seem to care so check and warn here.
 194                  */
 195                 err = -EIO;
 196         }
 197 out:
 198         if (err) {
 199                 dput(newdentry);
 200                 return ERR_PTR(err);
 201         }
 202         return newdentry;
 203 }
 204
 205 struct dentry *ovl_create_temp(struct dentry *workdir, struct ovl_cattr *attr)
 206 {
 207         return ovl_create_real(d_inode(workdir), ovl_lookup_temp(workdir),
 208                                attr);
 209 }
 210
 211 static int ovl_set_opaque_xerr(struct dentry *dentry, struct dentry *upper,
 212                                int xerr)
 213 {
 214         int err;
 215
 216         err = ovl_check_setxattr(dentry, upper, OVL_XATTR_OPAQUE, "y", 1, xerr);
 217         if (!err)
 218                 ovl_dentry_set_opaque(dentry);
 219
 220         return err;
 221 }
 222
 223 static int ovl_set_opaque(struct dentry *dentry, struct dentry *upperdentry)
 224 {
 225         /*
 226          * Fail with -EIO when trying to create opaque dir and upper doesn't
 227          * support xattrs. ovl_rename() calls ovl_set_opaque_xerr(-EXDEV) to
 228          * return a specific error for noxattr case.
 229          */
 230         return ovl_set_opaque_xerr(dentry, upperdentry, -EIO);
 231 }
 232
 233 /*
 234  * Common operations required to be done after creation of file on upper.
 235  * If @hardlink is false, then @inode is a pre-allocated inode, we may or
 236  * may not use to instantiate the new dentry.
 237  */
 238 static int ovl_instantiate(struct dentry *dentry, struct inode *inode,
 239                            struct dentry *newdentry, bool hardlink)
 240 {
 241         struct ovl_inode_params oip = {
 242                 .upperdentry = newdentry,
 243                 .newinode = inode,
 244         };
 245
 246         ovl_dir_modified(dentry->d_parent, false);
 247         ovl_dentry_set_upper_alias(dentry);
 248         if (!hardlink) {
 249                 /*
 250                  * ovl_obtain_alias() can be called after ovl_create_real()
 251                  * and before we get here, so we may get an inode from cache
 252                  * with the same real upperdentry that is not the inode we
 253                  * pre-allocated.  In this case we will use the cached inode
 254                  * to instantiate the new dentry.
 255                  *
 256                  * XXX: if we ever use ovl_obtain_alias() to decode directory
 257                  * file handles, need to use ovl_get_inode_locked() and
 258                  * d_instantiate_new() here to prevent from creating two
 259                  * hashed directory inode aliases.
 260                  */
 261                 inode = ovl_get_inode(dentry->d_sb, &oip);
 262                 if (IS_ERR(inode))
 263                         return PTR_ERR(inode);
 264         } else {
 265                 WARN_ON(ovl_inode_real(inode) != d_inode(newdentry));
 266                 dput(newdentry);
 267                 inc_nlink(inode);
 268         }
 269
 270         d_instantiate(dentry, inode);
 271         if (inode != oip.newinode) {
 272                 pr_warn_ratelimited("overlayfs: newly created inode found in cache (%pd2)\n",
 273                                     dentry);
 274         }
 275
 276         /* Force lookup of new upper hardlink to find its lower */
 277         if (hardlink)
 278                 d_drop(dentry);
 279
 280         return 0;
 281 }
 282
 283 static bool ovl_type_merge(struct dentry *dentry)
 284 {
 285         return OVL_TYPE_MERGE(ovl_path_type(dentry));
 286 }
 287
 288 static bool ovl_type_origin(struct dentry *dentry)
 289 {
 290         return OVL_TYPE_ORIGIN(ovl_path_type(dentry));
 291 }
 292
 293 static int ovl_create_upper(struct dentry *dentry, struct inode *inode,
 294                             struct ovl_cattr *attr)
 295 {
 296         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 297         struct inode *udir = upperdir->d_inode;
 298         struct dentry *newdentry;
 299         int err;
 300
 301         if (!attr->hardlink && !IS_POSIXACL(udir))
 302                 attr->mode &= ~current_umask();
 303
 304         inode_lock_nested(udir, I_MUTEX_PARENT);
 305         newdentry = ovl_create_real(udir,
 306                                     lookup_one_len(dentry->d_name.name,
 307                                                    upperdir,
 308                                                    dentry->d_name.len),
 309                                     attr);
 310         err = PTR_ERR(newdentry);
 311         if (IS_ERR(newdentry))
 312                 goto out_unlock;
 313
 314         if (ovl_type_merge(dentry->d_parent) && d_is_dir(newdentry)) {
 315                 /* Setting opaque here is just an optimization, allow to fail */
 316                 ovl_set_opaque(dentry, newdentry);
 317         }
 318
 319         err = ovl_instantiate(dentry, inode, newdentry, !!attr->hardlink);
 320         if (err)
 321                 goto out_cleanup;
 322 out_unlock:
 323         inode_unlock(udir);
 324         return err;
 325
 326 out_cleanup:
 327         ovl_cleanup(udir, newdentry);
 328         dput(newdentry);
 329         goto out_unlock;
 330 }
 331
 332 static struct dentry *ovl_clear_empty(struct dentry *dentry,
 333                                       struct list_head *list)
 334 {
 335         struct dentry *workdir = ovl_workdir(dentry);
 336         struct inode *wdir = workdir->d_inode;
 337         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 338         struct inode *udir = upperdir->d_inode;
 339         struct path upperpath;
 340         struct dentry *upper;
 341         struct dentry *opaquedir;
 342         struct kstat stat;
 343         int err;
 344
 345         if (WARN_ON(!workdir))
 346                 return ERR_PTR(-EROFS);
 347
 348         err = ovl_lock_rename_workdir(workdir, upperdir);
 349         if (err)
 350                 goto out;
 351
 352         ovl_path_upper(dentry, &upperpath);
 353         err = vfs_getattr(&upperpath, &stat,
 354                           STATX_BASIC_STATS, AT_STATX_SYNC_AS_STAT);
 355         if (err)
 356                 goto out_unlock;
 357
 358         err = -ESTALE;
 359         if (!S_ISDIR(stat.mode))
 360                 goto out_unlock;
 361         upper = upperpath.dentry;
 362         if (upper->d_parent->d_inode != udir)
 363                 goto out_unlock;
 364
 365         opaquedir = ovl_create_temp(workdir, OVL_CATTR(stat.mode));
 366         err = PTR_ERR(opaquedir);
 367         if (IS_ERR(opaquedir))
 368                 goto out_unlock;
 369
 370         err = ovl_copy_xattr(upper, opaquedir);
 371         if (err)
 372                 goto out_cleanup;
 373
 374         err = ovl_set_opaque(dentry, opaquedir);
 375         if (err)
 376                 goto out_cleanup;
 377
 378         inode_lock(opaquedir->d_inode);
 379         err = ovl_set_attr(opaquedir, &stat);
 380         inode_unlock(opaquedir->d_inode);
 381         if (err)
 382                 goto out_cleanup;
 383
 384         err = ovl_do_rename(wdir, opaquedir, udir, upper, RENAME_EXCHANGE);
 385         if (err)
 386                 goto out_cleanup;
 387
 388         ovl_cleanup_whiteouts(upper, list);
 389         ovl_cleanup(wdir, upper);
 390         unlock_rename(workdir, upperdir);
 391
 392         /* dentry's upper doesn't match now, get rid of it */
 393         d_drop(dentry);
 394
 395         return opaquedir;
 396
 397 out_cleanup:
 398         ovl_cleanup(wdir, opaquedir);
 399         dput(opaquedir);
 400 out_unlock:
 401         unlock_rename(workdir, upperdir);
 402 out:
 403         return ERR_PTR(err);
 404 }
 405
 406 static int ovl_set_upper_acl(struct dentry *upperdentry, const char *name,
 407                              const struct posix_acl *acl)
 408 {
 409         void *buffer;
 410         size_t size;
 411         int err;
 412
 413         if (!IS_ENABLED(CONFIG_FS_POSIX_ACL) || !acl)
 414                 return 0;
 415
 416         size = posix_acl_to_xattr(NULL, acl, NULL, 0);
 417         buffer = kmalloc(size, GFP_KERNEL);
 418         if (!buffer)
 419                 return -ENOMEM;
 420
 421         size = posix_acl_to_xattr(&init_user_ns, acl, buffer, size);
 422         err = size;
 423         if (err < 0)
 424                 goto out_free;
 425
 426         err = vfs_setxattr(upperdentry, name, buffer, size, XATTR_CREATE);
 427 out_free:
 428         kfree(buffer);
 429         return err;
 430 }
 431
 432 static int ovl_create_over_whiteout(struct dentry *dentry, struct inode *inode,
 433                                     struct ovl_cattr *cattr)
 434 {
 435         struct dentry *workdir = ovl_workdir(dentry);
 436         struct inode *wdir = workdir->d_inode;
 437         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 438         struct inode *udir = upperdir->d_inode;
 439         struct dentry *upper;
 440         struct dentry *newdentry;
 441         int err;
 442         struct posix_acl *acl, *default_acl;
 443         bool hardlink = !!cattr->hardlink;
 444
 445         if (WARN_ON(!workdir))
 446                 return -EROFS;
 447
 448         if (!hardlink) {
 449                 err = posix_acl_create(dentry->d_parent->d_inode,
 450                                        &cattr->mode, &default_acl, &acl);
 451                 if (err)
 452                         return err;
 453         }
 454
 455         err = ovl_lock_rename_workdir(workdir, upperdir);
 456         if (err)
 457                 goto out;
 458
 459         upper = lookup_one_len(dentry->d_name.name, upperdir,
 460                                dentry->d_name.len);
 461         err = PTR_ERR(upper);
 462         if (IS_ERR(upper))
 463                 goto out_unlock;
 464
 465         err = -ESTALE;
 466         if (d_is_negative(upper) || !IS_WHITEOUT(d_inode(upper)))
 467                 goto out_dput;
 468
 469         newdentry = ovl_create_temp(workdir, cattr);
 470         err = PTR_ERR(newdentry);
 471         if (IS_ERR(newdentry))
 472                 goto out_dput;
 473
 474         /*
 475          * mode could have been mutilated due to umask (e.g. sgid directory)
 476          */
 477         if (!hardlink &&
 478             !S_ISLNK(cattr->mode) &&
 479             newdentry->d_inode->i_mode != cattr->mode) {
 480                 struct iattr attr = {
 481                         .ia_valid = ATTR_MODE,
 482                         .ia_mode = cattr->mode,
 483                 };
 484                 inode_lock(newdentry->d_inode);
 485                 err = notify_change(newdentry, &attr, NULL);
 486                 inode_unlock(newdentry->d_inode);
 487                 if (err)
 488                         goto out_cleanup;
 489         }
 490         if (!hardlink) {
 491                 err = ovl_set_upper_acl(newdentry, XATTR_NAME_POSIX_ACL_ACCESS,
 492                                         acl);
 493                 if (err)
 494                         goto out_cleanup;
 495
 496                 err = ovl_set_upper_acl(newdentry, XATTR_NAME_POSIX_ACL_DEFAULT,
 497                                         default_acl);
 498                 if (err)
 499                         goto out_cleanup;
 500         }
 501
 502         if (!hardlink && S_ISDIR(cattr->mode)) {
 503                 err = ovl_set_opaque(dentry, newdentry);
 504                 if (err)
 505                         goto out_cleanup;
 506
 507                 err = ovl_do_rename(wdir, newdentry, udir, upper,
 508                                     RENAME_EXCHANGE);
 509                 if (err)
 510                         goto out_cleanup;
 511
 512                 ovl_cleanup(wdir, upper);
 513         } else {
 514                 err = ovl_do_rename(wdir, newdentry, udir, upper, 0);
 515                 if (err)
 516                         goto out_cleanup;
 517         }
 518         err = ovl_instantiate(dentry, inode, newdentry, hardlink);
 519         if (err) {
 520                 ovl_cleanup(udir, newdentry);
 521                 dput(newdentry);
 522         }
 523 out_dput:
 524         dput(upper);
 525 out_unlock:
 526         unlock_rename(workdir, upperdir);
 527 out:
 528         if (!hardlink) {
 529                 posix_acl_release(acl);
 530                 posix_acl_release(default_acl);
 531         }
 532         return err;
 533
 534 out_cleanup:
 535         ovl_cleanup(wdir, newdentry);
 536         dput(newdentry);
 537         goto out_dput;
 538 }
 539
 540 static int ovl_create_or_link(struct dentry *dentry, struct inode *inode,
 541                               struct ovl_cattr *attr, bool origin)
 542 {
 543         int err;
 544         const struct cred *old_cred;
 545         struct cred *override_cred;
 546         struct dentry *parent = dentry->d_parent;
 547
 548         err = ovl_copy_up(parent);
 549         if (err)
 550                 return err;
 551
 552         old_cred = ovl_override_creds(dentry->d_sb);
 553
 554         /*
 555          * When linking a file with copy up origin into a new parent, mark the
 556          * new parent dir "impure".
 557          */
 558         if (origin) {
 559                 err = ovl_set_impure(parent, ovl_dentry_upper(parent));
 560                 if (err)
 561                         goto out_revert_creds;
 562         }
 563
 564         if (!attr->hardlink) {
 565                 err = -ENOMEM;
 566                 override_cred = prepare_creds();
 567                 if (!override_cred)
 568                         goto out_revert_creds;
 569                 /*
 570                  * In the creation cases(create, mkdir, mknod, symlink),
 571                  * ovl should transfer current's fs{u,g}id to underlying
 572                  * fs. Because underlying fs want to initialize its new
 573                  * inode owner using current's fs{u,g}id. And in this
 574                  * case, the @inode is a new inode that is initialized
 575                  * in inode_init_owner() to current's fs{u,g}id. So use
 576                  * the inode's i_{u,g}id to override the cred's fs{u,g}id.
 577                  *
 578                  * But in the other hardlink case, ovl_link() does not
 579                  * create a new inode, so just use the ovl mounter's
 580                  * fs{u,g}id.
 581                  */
 582                 override_cred->fsuid = inode->i_uid;
 583                 override_cred->fsgid = inode->i_gid;
 584                 err = security_dentry_create_files_as(dentry,
 585                                 attr->mode, &dentry->d_name, old_cred,
 586                                 override_cred);
 587                 if (err) {
 588                         put_cred(override_cred);
 589                         goto out_revert_creds;
 590                 }
 591                 put_cred(override_creds(override_cred));
 592                 put_cred(override_cred);
 593         }
 594
 595         if (!ovl_dentry_is_whiteout(dentry))
 596                 err = ovl_create_upper(dentry, inode, attr);
 597         else
 598                 err = ovl_create_over_whiteout(dentry, inode, attr);
 599
 600 out_revert_creds:
 601         revert_creds(old_cred);
 602         return err;
 603 }
 604
 605 static int ovl_create_object(struct dentry *dentry, int mode, dev_t rdev,
 606                              const char *link)
 607 {
 608         int err;
 609         struct inode *inode;
 610         struct ovl_cattr attr = {
 611                 .rdev = rdev,
 612                 .link = link,
 613         };
 614
 615         err = ovl_want_write(dentry);
 616         if (err)
 617                 goto out;
 618
 619         /* Preallocate inode to be used by ovl_get_inode() */
 620         err = -ENOMEM;
 621         inode = ovl_new_inode(dentry->d_sb, mode, rdev);
 622         if (!inode)
 623                 goto out_drop_write;
 624
 625         spin_lock(&inode->i_lock);
 626         inode->i_state |= I_CREATING;
 627         spin_unlock(&inode->i_lock);
 628
 629         inode_init_owner(inode, dentry->d_parent->d_inode, mode);
 630         attr.mode = inode->i_mode;
 631
 632         err = ovl_create_or_link(dentry, inode, &attr, false);
 633         /* Did we end up using the preallocated inode? */
 634         if (inode != d_inode(dentry))
 635                 iput(inode);
 636
 637 out_drop_write:
 638         ovl_drop_write(dentry);
 639 out:
 640         return err;
 641 }
 642
 643 static int ovl_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 644                       bool excl)
 645 {
 646         return ovl_create_object(dentry, (mode & 07777) | S_IFREG, 0, NULL);
 647 }
 648
 649 static int ovl_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 650 {
 651         return ovl_create_object(dentry, (mode & 07777) | S_IFDIR, 0, NULL);
 652 }
 653
 654 static int ovl_mknod(struct inode *dir, struct dentry *dentry, umode_t mode,
 655                      dev_t rdev)
 656 {
 657         /* Don't allow creation of "whiteout" on overlay */
 658         if (S_ISCHR(mode) && rdev == WHITEOUT_DEV)
 659                 return -EPERM;
 660
 661         return ovl_create_object(dentry, mode, rdev, NULL);
 662 }
 663
 664 static int ovl_symlink(struct inode *dir, struct dentry *dentry,
 665                        const char *link)
 666 {
 667         return ovl_create_object(dentry, S_IFLNK, 0, link);
 668 }
 669
 670 static int ovl_set_link_redirect(struct dentry *dentry)
 671 {
 672         const struct cred *old_cred;
 673         int err;
 674
 675         old_cred = ovl_override_creds(dentry->d_sb);
 676         err = ovl_set_redirect(dentry, false);
 677         revert_creds(old_cred);
 678
 679         return err;
 680 }
 681
 682 static int ovl_link(struct dentry *old, struct inode *newdir,
 683                     struct dentry *new)
 684 {
 685         int err;
 686         bool locked = false;
 687         struct inode *inode;
 688
 689         err = ovl_want_write(old);
 690         if (err)
 691                 goto out;
 692
 693         err = ovl_copy_up(old);
 694         if (err)
 695                 goto out_drop_write;
 696
 697         err = ovl_copy_up(new->d_parent);
 698         if (err)
 699                 goto out_drop_write;
 700
 701         if (ovl_is_metacopy_dentry(old)) {
 702                 err = ovl_set_link_redirect(old);
 703                 if (err)
 704                         goto out_drop_write;
 705         }
 706
 707         err = ovl_nlink_start(old, &locked);
 708         if (err)
 709                 goto out_drop_write;
 710
 711         inode = d_inode(old);
 712         ihold(inode);
 713
 714         err = ovl_create_or_link(new, inode,
 715                         &(struct ovl_cattr) {.hardlink = ovl_dentry_upper(old)},
 716                         ovl_type_origin(old));
 717         if (err)
 718                 iput(inode);
 719
 720         ovl_nlink_end(old, locked);
 721 out_drop_write:
 722         ovl_drop_write(old);
 723 out:
 724         return err;
 725 }
 726
 727 static bool ovl_matches_upper(struct dentry *dentry, struct dentry *upper)
 728 {
 729         return d_inode(ovl_dentry_upper(dentry)) == d_inode(upper);
 730 }
 731
 732 static int ovl_remove_and_whiteout(struct dentry *dentry,
 733                                    struct list_head *list)
 734 {
 735         struct dentry *workdir = ovl_workdir(dentry);
 736         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 737         struct dentry *upper;
 738         struct dentry *opaquedir = NULL;
 739         int err;
 740
 741         if (WARN_ON(!workdir))
 742                 return -EROFS;
 743
 744         if (!list_empty(list)) {
 745                 opaquedir = ovl_clear_empty(dentry, list);
 746                 err = PTR_ERR(opaquedir);
 747                 if (IS_ERR(opaquedir))
 748                         goto out;
 749         }
 750
 751         err = ovl_lock_rename_workdir(workdir, upperdir);
 752         if (err)
 753                 goto out_dput;
 754
 755         upper = lookup_one_len(dentry->d_name.name, upperdir,
 756                                dentry->d_name.len);
 757         err = PTR_ERR(upper);
 758         if (IS_ERR(upper))
 759                 goto out_unlock;
 760
 761         err = -ESTALE;
 762         if ((opaquedir && upper != opaquedir) ||
 763             (!opaquedir && ovl_dentry_upper(dentry) &&
 764              !ovl_matches_upper(dentry, upper))) {
 765                 goto out_dput_upper;
 766         }
 767
 768         err = ovl_cleanup_and_whiteout(workdir, d_inode(upperdir), upper);
 769         if (err)
 770                 goto out_d_drop;
 771
 772         ovl_dir_modified(dentry->d_parent, true);
 773 out_d_drop:
 774         d_drop(dentry);
 775 out_dput_upper:
 776         dput(upper);
 777 out_unlock:
 778         unlock_rename(workdir, upperdir);
 779 out_dput:
 780         dput(opaquedir);
 781 out:
 782         return err;
 783 }
 784
 785 static int ovl_remove_upper(struct dentry *dentry, bool is_dir,
 786                             struct list_head *list)
 787 {
 788         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 789         struct inode *dir = upperdir->d_inode;
 790         struct dentry *upper;
 791         struct dentry *opaquedir = NULL;
 792         int err;
 793
 794         if (!list_empty(list)) {
 795                 opaquedir = ovl_clear_empty(dentry, list);
 796                 err = PTR_ERR(opaquedir);
 797                 if (IS_ERR(opaquedir))
 798                         goto out;
 799         }
 800
 801         inode_lock_nested(dir, I_MUTEX_PARENT);
 802         upper = lookup_one_len(dentry->d_name.name, upperdir,
 803                                dentry->d_name.len);
 804         err = PTR_ERR(upper);
 805         if (IS_ERR(upper))
 806                 goto out_unlock;
 807
 808         err = -ESTALE;
 809         if ((opaquedir && upper != opaquedir) ||
 810             (!opaquedir && !ovl_matches_upper(dentry, upper)))
 811                 goto out_dput_upper;
 812
 813         if (is_dir)
 814                 err = vfs_rmdir(dir, upper);
 815         else
 816                 err = vfs_unlink(dir, upper, NULL);
 817         ovl_dir_modified(dentry->d_parent, ovl_type_origin(dentry));
 818
 819         /*
 820          * Keeping this dentry hashed would mean having to release
 821          * upperpath/lowerpath, which could only be done if we are the
 822          * sole user of this dentry.  Too tricky...  Just unhash for
 823          * now.
 824          */
 825         if (!err)
 826                 d_drop(dentry);
 827 out_dput_upper:
 828         dput(upper);
 829 out_unlock:
 830         inode_unlock(dir);
 831         dput(opaquedir);
 832 out:
 833         return err;
 834 }
 835
 836 static bool ovl_pure_upper(struct dentry *dentry)
 837 {
 838         return !ovl_dentry_lower(dentry) &&
 839                !ovl_test_flag(OVL_WHITEOUTS, d_inode(dentry));
 840 }
 841
 842 static int ovl_do_remove(struct dentry *dentry, bool is_dir)
 843 {
 844         int err;
 845         bool locked = false;
 846         const struct cred *old_cred;
 847         struct dentry *upperdentry;
 848         bool lower_positive = ovl_lower_positive(dentry);
 849         LIST_HEAD(list);
 850
 851         /* No need to clean pure upper removed by vfs_rmdir() */
 852         if (is_dir && (lower_positive || !ovl_pure_upper(dentry))) {
 853                 err = ovl_check_empty_dir(dentry, &list);
 854                 if (err)
 855                         goto out;
 856         }
 857
 858         err = ovl_want_write(dentry);
 859         if (err)
 860                 goto out;
 861
 862         err = ovl_copy_up(dentry->d_parent);
 863         if (err)
 864                 goto out_drop_write;
 865
 866         err = ovl_nlink_start(dentry, &locked);
 867         if (err)
 868                 goto out_drop_write;
 869
 870         old_cred = ovl_override_creds(dentry->d_sb);
 871         if (!lower_positive)
 872                 err = ovl_remove_upper(dentry, is_dir, &list);
 873         else
 874                 err = ovl_remove_and_whiteout(dentry, &list);
 875         revert_creds(old_cred);
 876         if (!err) {
 877                 if (is_dir)
 878                         clear_nlink(dentry->d_inode);
 879                 else
 880                         drop_nlink(dentry->d_inode);
 881         }
 882         ovl_nlink_end(dentry, locked);
 883
 884         /*
 885          * Copy ctime
 886          *
 887          * Note: we fail to update ctime if there was no copy-up, only a
 888          * whiteout
 889          */
 890         upperdentry = ovl_dentry_upper(dentry);
 891         if (upperdentry)
 892                 ovl_copyattr(d_inode(upperdentry), d_inode(dentry));
 893
 894 out_drop_write:
 895         ovl_drop_write(dentry);
 896 out:
 897         ovl_cache_free(&list);
 898         return err;
 899 }
 900
 901 static int ovl_unlink(struct inode *dir, struct dentry *dentry)
 902 {
 903         return ovl_do_remove(dentry, false);
 904 }
 905
 906 static int ovl_rmdir(struct inode *dir, struct dentry *dentry)
 907 {
 908         return ovl_do_remove(dentry, true);
 909 }
 910
 911 static bool ovl_type_merge_or_lower(struct dentry *dentry)
 912 {
 913         enum ovl_path_type type = ovl_path_type(dentry);
 914
 915         return OVL_TYPE_MERGE(type) || !OVL_TYPE_UPPER(type);
 916 }
 917
 918 static bool ovl_can_move(struct dentry *dentry)
 919 {
 920         return ovl_redirect_dir(dentry->d_sb) ||
 921                 !d_is_dir(dentry) || !ovl_type_merge_or_lower(dentry);
 922 }
 923
 924 static char *ovl_get_redirect(struct dentry *dentry, bool abs_redirect)
 925 {
 926         char *buf, *ret;
 927         struct dentry *d, *tmp;
 928         int buflen = ovl_redirect_max + 1;
 929
 930         if (!abs_redirect) {
 931                 ret = kstrndup(dentry->d_name.name, dentry->d_name.len,
 932                                GFP_KERNEL);
 933                 goto out;
 934         }
 935
 936         buf = ret = kmalloc(buflen, GFP_KERNEL);
 937         if (!buf)
 938                 goto out;
 939
 940         buflen--;
 941         buf[buflen] = '\0';
 942         for (d = dget(dentry); !IS_ROOT(d);) {
 943                 const char *name;
 944                 int thislen;
 945
 946                 spin_lock(&d->d_lock);
 947                 name = ovl_dentry_get_redirect(d);
 948                 if (name) {
 949                         thislen = strlen(name);
 950                 } else {
 951                         name = d->d_name.name;
 952                         thislen = d->d_name.len;
 953                 }
 954
 955                 /* If path is too long, fall back to userspace move */
 956                 if (thislen + (name[0] != '/') > buflen) {
 957                         ret = ERR_PTR(-EXDEV);
 958                         spin_unlock(&d->d_lock);
 959                         goto out_put;
 960                 }
 961
 962                 buflen -= thislen;
 963                 memcpy(&buf[buflen], name, thislen);
 964                 spin_unlock(&d->d_lock);
 965                 tmp = dget_parent(d);
 966
 967                 dput(d);
 968                 d = tmp;
 969
 970                 /* Absolute redirect: finished */
 971                 if (buf[buflen] == '/')
 972                         break;
 973                 buflen--;
 974                 buf[buflen] = '/';
 975         }
 976         ret = kstrdup(&buf[buflen], GFP_KERNEL);
 977 out_put:
 978         dput(d);
 979         kfree(buf);
 980 out:
 981         return ret ? ret : ERR_PTR(-ENOMEM);
 982 }
 983
 984 static bool ovl_need_absolute_redirect(struct dentry *dentry, bool samedir)
 985 {
 986         struct dentry *lowerdentry;
 987
 988         if (!samedir)
 989                 return true;
 990
 991         if (d_is_dir(dentry))
 992                 return false;
 993
 994         /*
 995          * For non-dir hardlinked files, we need absolute redirects
 996          * in general as two upper hardlinks could be in different
 997          * dirs. We could put a relative redirect now and convert
 998          * it to absolute redirect later. But when nlink > 1 and
 999          * indexing is on, that means relative redirect needs to be
1000          * converted to absolute during copy up of another lower
1001          * hardllink as well.
1002          *
1003          * So without optimizing too much, just check if lower is
1004          * a hard link or not. If lower is hard link, put absolute
1005          * redirect.
1006          */
1007         lowerdentry = ovl_dentry_lower(dentry);
1008         return (d_inode(lowerdentry)->i_nlink > 1);
1009 }
1010
1011 static int ovl_set_redirect(struct dentry *dentry, bool samedir)
1012 {
1013         int err;
1014         const char *redirect = ovl_dentry_get_redirect(dentry);
1015         bool absolute_redirect = ovl_need_absolute_redirect(dentry, samedir);
1016
1017         if (redirect && (!absolute_redirect || redirect[0] == '/'))
1018                 return 0;
1019
1020         redirect = ovl_get_redirect(dentry, absolute_redirect);
1021         if (IS_ERR(redirect))
1022                 return PTR_ERR(redirect);
1023
1024         err = ovl_check_setxattr(dentry, ovl_dentry_upper(dentry),
1025                                  OVL_XATTR_REDIRECT,
1026                                  redirect, strlen(redirect), -EXDEV);
1027         if (!err) {
1028                 spin_lock(&dentry->d_lock);
1029                 ovl_dentry_set_redirect(dentry, redirect);
1030                 spin_unlock(&dentry->d_lock);
1031         } else {
1032                 kfree(redirect);
1033                 pr_warn_ratelimited("overlayfs: failed to set redirect (%i)\n",
1034                                     err);
1035                 /* Fall back to userspace copy-up */
1036                 err = -EXDEV;
1037         }
1038         return err;
1039 }
1040
1041 static int ovl_rename(struct inode *olddir, struct dentry *old,
1042                       struct inode *newdir, struct dentry *new,
1043                       unsigned int flags)
1044 {
1045         int err;
1046         bool locked = false;
1047         struct dentry *old_upperdir;
1048         struct dentry *new_upperdir;
1049         struct dentry *olddentry;
1050         struct dentry *newdentry;
1051         struct dentry *trap;
1052         bool old_opaque;
1053         bool new_opaque;
1054         bool cleanup_whiteout = false;
1055         bool overwrite = !(flags & RENAME_EXCHANGE);
1056         bool is_dir = d_is_dir(old);
1057         bool new_is_dir = d_is_dir(new);
1058         bool samedir = olddir == newdir;
1059         struct dentry *opaquedir = NULL;
1060         const struct cred *old_cred = NULL;
1061         LIST_HEAD(list);
1062
1063         err = -EINVAL;
1064         if (flags & ~(RENAME_EXCHANGE | RENAME_NOREPLACE))
1065                 goto out;
1066
1067         flags &= ~RENAME_NOREPLACE;
1068
1069         /* Don't copy up directory trees */
1070         err = -EXDEV;
1071         if (!ovl_can_move(old))
1072                 goto out;
1073         if (!overwrite && !ovl_can_move(new))
1074                 goto out;
1075
1076         if (overwrite && new_is_dir && !ovl_pure_upper(new)) {
1077                 err = ovl_check_empty_dir(new, &list);
1078                 if (err)
1079                         goto out;
1080         }
1081
1082         if (overwrite) {
1083                 if (ovl_lower_positive(old)) {
1084                         if (!ovl_dentry_is_whiteout(new)) {
1085                                 /* Whiteout source */
1086                                 flags |= RENAME_WHITEOUT;
1087                         } else {
1088                                 /* Switch whiteouts */
1089                                 flags |= RENAME_EXCHANGE;
1090                         }
1091                 } else if (is_dir && ovl_dentry_is_whiteout(new)) {
1092                         flags |= RENAME_EXCHANGE;
1093                         cleanup_whiteout = true;
1094                 }
1095         }
1096
1097         err = ovl_want_write(old);
1098         if (err)
1099                 goto out;
1100
1101         err = ovl_copy_up(old);
1102         if (err)
1103                 goto out_drop_write;
1104
1105         err = ovl_copy_up(new->d_parent);
1106         if (err)
1107                 goto out_drop_write;
1108         if (!overwrite) {
1109                 err = ovl_copy_up(new);
1110                 if (err)
1111                         goto out_drop_write;
1112         } else {
1113                 err = ovl_nlink_start(new, &locked);
1114                 if (err)
1115                         goto out_drop_write;
1116         }
1117
1118         old_cred = ovl_override_creds(old->d_sb);
1119
1120         if (!list_empty(&list)) {
1121                 opaquedir = ovl_clear_empty(new, &list);
1122                 err = PTR_ERR(opaquedir);
1123                 if (IS_ERR(opaquedir)) {
1124                         opaquedir = NULL;
1125                         goto out_revert_creds;
1126                 }
1127         }
1128
1129         old_upperdir = ovl_dentry_upper(old->d_parent);
1130         new_upperdir = ovl_dentry_upper(new->d_parent);
1131
1132         if (!samedir) {
1133                 /*
1134                  * When moving a merge dir or non-dir with copy up origin into
1135                  * a new parent, we are marking the new parent dir "impure".
1136                  * When ovl_iterate() iterates an "impure" upper dir, it will
1137                  * lookup the origin inodes of the entries to fill d_ino.
1138                  */
1139                 if (ovl_type_origin(old)) {
1140                         err = ovl_set_impure(new->d_parent, new_upperdir);
1141                         if (err)
1142                                 goto out_revert_creds;
1143                 }
1144                 if (!overwrite && ovl_type_origin(new)) {
1145                         err = ovl_set_impure(old->d_parent, old_upperdir);
1146                         if (err)
1147                                 goto out_revert_creds;
1148                 }
1149         }
1150
1151         trap = lock_rename(new_upperdir, old_upperdir);
1152
1153         olddentry = lookup_one_len(old->d_name.name, old_upperdir,
1154                                    old->d_name.len);
1155         err = PTR_ERR(olddentry);
1156         if (IS_ERR(olddentry))
1157                 goto out_unlock;
1158
1159         err = -ESTALE;
1160         if (!ovl_matches_upper(old, olddentry))
1161                 goto out_dput_old;
1162
1163         newdentry = lookup_one_len(new->d_name.name, new_upperdir,
1164                                    new->d_name.len);
1165         err = PTR_ERR(newdentry);
1166         if (IS_ERR(newdentry))
1167                 goto out_dput_old;
1168
1169         old_opaque = ovl_dentry_is_opaque(old);
1170         new_opaque = ovl_dentry_is_opaque(new);
1171
1172         err = -ESTALE;
1173         if (d_inode(new) && ovl_dentry_upper(new)) {
1174                 if (opaquedir) {
1175                         if (newdentry != opaquedir)
1176                                 goto out_dput;
1177                 } else {
1178                         if (!ovl_matches_upper(new, newdentry))
1179                                 goto out_dput;
1180                 }
1181         } else {
1182                 if (!d_is_negative(newdentry)) {
1183                         if (!new_opaque || !ovl_is_whiteout(newdentry))
1184                                 goto out_dput;
1185                 } else {
1186                         if (flags & RENAME_EXCHANGE)
1187                                 goto out_dput;
1188                 }
1189         }
1190
1191         if (olddentry == trap)
1192                 goto out_dput;
1193         if (newdentry == trap)
1194                 goto out_dput;
1195
1196         if (olddentry->d_inode == newdentry->d_inode)
1197                 goto out_dput;
1198
1199         err = 0;
1200         if (ovl_type_merge_or_lower(old))
1201                 err = ovl_set_redirect(old, samedir);
1202         else if (is_dir && !old_opaque && ovl_type_merge(new->d_parent))
1203                 err = ovl_set_opaque_xerr(old, olddentry, -EXDEV);
1204         if (err)
1205                 goto out_dput;
1206
1207         if (!overwrite && ovl_type_merge_or_lower(new))
1208                 err = ovl_set_redirect(new, samedir);
1209         else if (!overwrite && new_is_dir && !new_opaque &&
1210                  ovl_type_merge(old->d_parent))
1211                 err = ovl_set_opaque_xerr(new, newdentry, -EXDEV);
1212         if (err)
1213                 goto out_dput;
1214
1215         err = ovl_do_rename(old_upperdir->d_inode, olddentry,
1216                             new_upperdir->d_inode, newdentry, flags);
1217         if (err)
1218                 goto out_dput;
1219
1220         if (cleanup_whiteout)
1221                 ovl_cleanup(old_upperdir->d_inode, newdentry);
1222
1223         if (overwrite && d_inode(new)) {
1224                 if (new_is_dir)
1225                         clear_nlink(d_inode(new));
1226                 else
1227                         drop_nlink(d_inode(new));
1228         }
1229
1230         ovl_dir_modified(old->d_parent, ovl_type_origin(old) ||
1231                          (!overwrite && ovl_type_origin(new)));
1232         ovl_dir_modified(new->d_parent, ovl_type_origin(old) ||
1233                          (d_inode(new) && ovl_type_origin(new)));
1234
1235         /* copy ctime: */
1236         ovl_copyattr(d_inode(olddentry), d_inode(old));
1237         if (d_inode(new) && ovl_dentry_upper(new))
1238                 ovl_copyattr(d_inode(newdentry), d_inode(new));
1239
1240 out_dput:
1241         dput(newdentry);
1242 out_dput_old:
1243         dput(olddentry);
1244 out_unlock:
1245         unlock_rename(new_upperdir, old_upperdir);
1246 out_revert_creds:
1247         revert_creds(old_cred);
1248         ovl_nlink_end(new, locked);
1249 out_drop_write:
1250         ovl_drop_write(old);
1251 out:
1252         dput(opaquedir);
1253         ovl_cache_free(&list);
1254         return err;
1255 }
1256
1257 const struct inode_operations ovl_dir_inode_operations = {
1258         .lookup         = ovl_lookup,
1259         .mkdir          = ovl_mkdir,
1260         .symlink        = ovl_symlink,
1261         .unlink         = ovl_unlink,
1262         .rmdir          = ovl_rmdir,
1263         .rename         = ovl_rename,
1264         .link           = ovl_link,
1265         .setattr        = ovl_setattr,
1266         .create         = ovl_create,
1267         .mknod          = ovl_mknod,
1268         .permission     = ovl_permission,
1269         .getattr        = ovl_getattr,
1270         .listxattr      = ovl_listxattr,
1271         .get_acl        = ovl_get_acl,
1272         .update_time    = ovl_update_time,
1273 };