kernel/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 "overlayfs.h"
  16
  17 void ovl_cleanup(struct inode *wdir, struct dentry *wdentry)
  18 {
  19         int err;
  20
  21         dget(wdentry);
  22         if (d_is_dir(wdentry))
  23                 err = ovl_do_rmdir(wdir, wdentry);
  24         else
  25                 err = ovl_do_unlink(wdir, wdentry);
  26         dput(wdentry);
  27
  28         if (err) {
  29                 pr_err("overlayfs: cleanup of '%pd2' failed (%i)\n",
  30                        wdentry, err);
  31         }
  32 }
  33
  34 struct dentry *ovl_lookup_temp(struct dentry *workdir, struct dentry *dentry)
  35 {
  36         struct dentry *temp;
  37         char name[20];
  38
  39         snprintf(name, sizeof(name), "#%lx", (unsigned long) dentry);
  40
  41         temp = lookup_one_len(name, workdir, strlen(name));
  42         if (!IS_ERR(temp) && temp->d_inode) {
  43                 pr_err("overlayfs: workdir/%s already exists\n", name);
  44                 dput(temp);
  45                 temp = ERR_PTR(-EIO);
  46         }
  47
  48         return temp;
  49 }
  50
  51 /* caller holds i_mutex on workdir */
  52 static struct dentry *ovl_whiteout(struct dentry *workdir,
  53                                    struct dentry *dentry)
  54 {
  55         int err;
  56         struct dentry *whiteout;
  57         struct inode *wdir = workdir->d_inode;
  58
  59         whiteout = ovl_lookup_temp(workdir, dentry);
  60         if (IS_ERR(whiteout))
  61                 return whiteout;
  62
  63         err = ovl_do_whiteout(wdir, whiteout);
  64         if (err) {
  65                 dput(whiteout);
  66                 whiteout = ERR_PTR(err);
  67         }
  68
  69         return whiteout;
  70 }
  71
  72 int ovl_create_real(struct inode *dir, struct dentry *newdentry,
  73                     struct kstat *stat, const char *link,
  74                     struct dentry *hardlink, bool debug)
  75 {
  76         int err;
  77
  78         if (newdentry->d_inode)
  79                 return -ESTALE;
  80
  81         if (hardlink) {
  82                 err = ovl_do_link(hardlink, dir, newdentry, debug);
  83         } else {
  84                 switch (stat->mode & S_IFMT) {
  85                 case S_IFREG:
  86                         err = ovl_do_create(dir, newdentry, stat->mode, debug);
  87                         break;
  88
  89                 case S_IFDIR:
  90                         err = ovl_do_mkdir(dir, newdentry, stat->mode, debug);
  91                         break;
  92
  93                 case S_IFCHR:
  94                 case S_IFBLK:
  95                 case S_IFIFO:
  96                 case S_IFSOCK:
  97                         err = ovl_do_mknod(dir, newdentry,
  98                                            stat->mode, stat->rdev, debug);
  99                         break;
 100
 101                 case S_IFLNK:
 102                         err = ovl_do_symlink(dir, newdentry, link, debug);
 103                         break;
 104
 105                 default:
 106                         err = -EPERM;
 107                 }
 108         }
 109         if (!err && WARN_ON(!newdentry->d_inode)) {
 110                 /*
 111                  * Not quite sure if non-instantiated dentry is legal or not.
 112                  * VFS doesn't seem to care so check and warn here.
 113                  */
 114                 err = -ENOENT;
 115         }
 116         return err;
 117 }
 118
 119 static int ovl_set_opaque(struct dentry *upperdentry)
 120 {
 121         return ovl_do_setxattr(upperdentry, OVL_XATTR_OPAQUE, "y", 1, 0);
 122 }
 123
 124 static void ovl_remove_opaque(struct dentry *upperdentry)
 125 {
 126         int err;
 127
 128         err = ovl_do_removexattr(upperdentry, OVL_XATTR_OPAQUE);
 129         if (err) {
 130                 pr_warn("overlayfs: failed to remove opaque from '%s' (%i)\n",
 131                         upperdentry->d_name.name, err);
 132         }
 133 }
 134
 135 static int ovl_dir_getattr(struct vfsmount *mnt, struct dentry *dentry,
 136                          struct kstat *stat)
 137 {
 138         int err;
 139         enum ovl_path_type type;
 140         struct path realpath;
 141
 142         type = ovl_path_real(dentry, &realpath);
 143         err = vfs_getattr(&realpath, stat);
 144         if (err)
 145                 return err;
 146
 147         stat->dev = dentry->d_sb->s_dev;
 148         stat->ino = dentry->d_inode->i_ino;
 149
 150         /*
 151          * It's probably not worth it to count subdirs to get the
 152          * correct link count.  nlink=1 seems to pacify 'find' and
 153          * other utilities.
 154          */
 155         if (OVL_TYPE_MERGE(type))
 156                 stat->nlink = 1;
 157
 158         return 0;
 159 }
 160
 161 static int ovl_create_upper(struct dentry *dentry, struct inode *inode,
 162                             struct kstat *stat, const char *link,
 163                             struct dentry *hardlink)
 164 {
 165         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 166         struct inode *udir = upperdir->d_inode;
 167         struct dentry *newdentry;
 168         int err;
 169
 170         mutex_lock_nested(&udir->i_mutex, I_MUTEX_PARENT);
 171         newdentry = lookup_one_len(dentry->d_name.name, upperdir,
 172                                    dentry->d_name.len);
 173         err = PTR_ERR(newdentry);
 174         if (IS_ERR(newdentry))
 175                 goto out_unlock;
 176         err = ovl_create_real(udir, newdentry, stat, link, hardlink, false);
 177         if (err)
 178                 goto out_dput;
 179
 180         ovl_dentry_version_inc(dentry->d_parent);
 181         ovl_dentry_update(dentry, newdentry);
 182         ovl_copyattr(newdentry->d_inode, inode);
 183         d_instantiate(dentry, inode);
 184         newdentry = NULL;
 185 out_dput:
 186         dput(newdentry);
 187 out_unlock:
 188         mutex_unlock(&udir->i_mutex);
 189         return err;
 190 }
 191
 192 static int ovl_lock_rename_workdir(struct dentry *workdir,
 193                                    struct dentry *upperdir)
 194 {
 195         /* Workdir should not be the same as upperdir */
 196         if (workdir == upperdir)
 197                 goto err;
 198
 199         /* Workdir should not be subdir of upperdir and vice versa */
 200         if (lock_rename(workdir, upperdir) != NULL)
 201                 goto err_unlock;
 202
 203         return 0;
 204
 205 err_unlock:
 206         unlock_rename(workdir, upperdir);
 207 err:
 208         pr_err("overlayfs: failed to lock workdir+upperdir\n");
 209         return -EIO;
 210 }
 211
 212 static struct dentry *ovl_clear_empty(struct dentry *dentry,
 213                                       struct list_head *list)
 214 {
 215         struct dentry *workdir = ovl_workdir(dentry);
 216         struct inode *wdir = workdir->d_inode;
 217         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 218         struct inode *udir = upperdir->d_inode;
 219         struct path upperpath;
 220         struct dentry *upper;
 221         struct dentry *opaquedir;
 222         struct kstat stat;
 223         int err;
 224
 225         if (WARN_ON(!workdir))
 226                 return ERR_PTR(-EROFS);
 227
 228         err = ovl_lock_rename_workdir(workdir, upperdir);
 229         if (err)
 230                 goto out;
 231
 232         ovl_path_upper(dentry, &upperpath);
 233         err = vfs_getattr(&upperpath, &stat);
 234         if (err)
 235                 goto out_unlock;
 236
 237         err = -ESTALE;
 238         if (!S_ISDIR(stat.mode))
 239                 goto out_unlock;
 240         upper = upperpath.dentry;
 241         if (upper->d_parent->d_inode != udir)
 242                 goto out_unlock;
 243
 244         opaquedir = ovl_lookup_temp(workdir, dentry);
 245         err = PTR_ERR(opaquedir);
 246         if (IS_ERR(opaquedir))
 247                 goto out_unlock;
 248
 249         err = ovl_create_real(wdir, opaquedir, &stat, NULL, NULL, true);
 250         if (err)
 251                 goto out_dput;
 252
 253         err = ovl_copy_xattr(upper, opaquedir);
 254         if (err)
 255                 goto out_cleanup;
 256
 257         err = ovl_set_opaque(opaquedir);
 258         if (err)
 259                 goto out_cleanup;
 260
 261         mutex_lock(&opaquedir->d_inode->i_mutex);
 262         err = ovl_set_attr(opaquedir, &stat);
 263         mutex_unlock(&opaquedir->d_inode->i_mutex);
 264         if (err)
 265                 goto out_cleanup;
 266
 267         err = ovl_do_rename(wdir, opaquedir, udir, upper, RENAME_EXCHANGE);
 268         if (err)
 269                 goto out_cleanup;
 270
 271         ovl_cleanup_whiteouts(upper, list);
 272         ovl_cleanup(wdir, upper);
 273         unlock_rename(workdir, upperdir);
 274
 275         /* dentry's upper doesn't match now, get rid of it */
 276         d_drop(dentry);
 277
 278         return opaquedir;
 279
 280 out_cleanup:
 281         ovl_cleanup(wdir, opaquedir);
 282 out_dput:
 283         dput(opaquedir);
 284 out_unlock:
 285         unlock_rename(workdir, upperdir);
 286 out:
 287         return ERR_PTR(err);
 288 }
 289
 290 static struct dentry *ovl_check_empty_and_clear(struct dentry *dentry)
 291 {
 292         int err;
 293         struct dentry *ret = NULL;
 294         LIST_HEAD(list);
 295
 296         err = ovl_check_empty_dir(dentry, &list);
 297         if (err)
 298                 ret = ERR_PTR(err);
 299         else {
 300                 /*
 301                  * If no upperdentry then skip clearing whiteouts.
 302                  *
 303                  * Can race with copy-up, since we don't hold the upperdir
 304                  * mutex.  Doesn't matter, since copy-up can't create a
 305                  * non-empty directory from an empty one.
 306                  */
 307                 if (ovl_dentry_upper(dentry))
 308                         ret = ovl_clear_empty(dentry, &list);
 309         }
 310
 311         ovl_cache_free(&list);
 312
 313         return ret;
 314 }
 315
 316 static int ovl_create_over_whiteout(struct dentry *dentry, struct inode *inode,
 317                                     struct kstat *stat, const char *link,
 318                                     struct dentry *hardlink)
 319 {
 320         struct dentry *workdir = ovl_workdir(dentry);
 321         struct inode *wdir = workdir->d_inode;
 322         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 323         struct inode *udir = upperdir->d_inode;
 324         struct dentry *upper;
 325         struct dentry *newdentry;
 326         int err;
 327
 328         if (WARN_ON(!workdir))
 329                 return -EROFS;
 330
 331         err = ovl_lock_rename_workdir(workdir, upperdir);
 332         if (err)
 333                 goto out;
 334
 335         newdentry = ovl_lookup_temp(workdir, dentry);
 336         err = PTR_ERR(newdentry);
 337         if (IS_ERR(newdentry))
 338                 goto out_unlock;
 339
 340         upper = lookup_one_len(dentry->d_name.name, upperdir,
 341                                dentry->d_name.len);
 342         err = PTR_ERR(upper);
 343         if (IS_ERR(upper))
 344                 goto out_dput;
 345
 346         err = ovl_create_real(wdir, newdentry, stat, link, hardlink, true);
 347         if (err)
 348                 goto out_dput2;
 349
 350         if (S_ISDIR(stat->mode)) {
 351                 err = ovl_set_opaque(newdentry);
 352                 if (err)
 353                         goto out_cleanup;
 354
 355                 err = ovl_do_rename(wdir, newdentry, udir, upper,
 356                                     RENAME_EXCHANGE);
 357                 if (err)
 358                         goto out_cleanup;
 359
 360                 ovl_cleanup(wdir, upper);
 361         } else {
 362                 err = ovl_do_rename(wdir, newdentry, udir, upper, 0);
 363                 if (err)
 364                         goto out_cleanup;
 365         }
 366         ovl_dentry_version_inc(dentry->d_parent);
 367         ovl_dentry_update(dentry, newdentry);
 368         ovl_copyattr(newdentry->d_inode, inode);
 369         d_instantiate(dentry, inode);
 370         newdentry = NULL;
 371 out_dput2:
 372         dput(upper);
 373 out_dput:
 374         dput(newdentry);
 375 out_unlock:
 376         unlock_rename(workdir, upperdir);
 377 out:
 378         return err;
 379
 380 out_cleanup:
 381         ovl_cleanup(wdir, newdentry);
 382         goto out_dput2;
 383 }
 384
 385 static int ovl_create_or_link(struct dentry *dentry, int mode, dev_t rdev,
 386                               const char *link, struct dentry *hardlink)
 387 {
 388         int err;
 389         struct inode *inode;
 390         struct kstat stat = {
 391                 .mode = mode,
 392                 .rdev = rdev,
 393         };
 394
 395         err = -ENOMEM;
 396         inode = ovl_new_inode(dentry->d_sb, mode, dentry->d_fsdata);
 397         if (!inode)
 398                 goto out;
 399
 400         err = ovl_copy_up(dentry->d_parent);
 401         if (err)
 402                 goto out_iput;
 403
 404         if (!ovl_dentry_is_opaque(dentry)) {
 405                 err = ovl_create_upper(dentry, inode, &stat, link, hardlink);
 406         } else {
 407                 const struct cred *old_cred;
 408                 struct cred *override_cred;
 409
 410                 err = -ENOMEM;
 411                 override_cred = prepare_creds();
 412                 if (!override_cred)
 413                         goto out_iput;
 414
 415                 /*
 416                  * CAP_SYS_ADMIN for setting opaque xattr
 417                  * CAP_DAC_OVERRIDE for create in workdir, rename
 418                  * CAP_FOWNER for removing whiteout from sticky dir
 419                  */
 420                 cap_raise(override_cred->cap_effective, CAP_SYS_ADMIN);
 421                 cap_raise(override_cred->cap_effective, CAP_DAC_OVERRIDE);
 422                 cap_raise(override_cred->cap_effective, CAP_FOWNER);
 423                 old_cred = override_creds(override_cred);
 424
 425                 err = ovl_create_over_whiteout(dentry, inode, &stat, link,
 426                                                hardlink);
 427
 428                 revert_creds(old_cred);
 429                 put_cred(override_cred);
 430         }
 431
 432         if (!err)
 433                 inode = NULL;
 434 out_iput:
 435         iput(inode);
 436 out:
 437         return err;
 438 }
 439
 440 static int ovl_create_object(struct dentry *dentry, int mode, dev_t rdev,
 441                              const char *link)
 442 {
 443         int err;
 444
 445         err = ovl_want_write(dentry);
 446         if (!err) {
 447                 err = ovl_create_or_link(dentry, mode, rdev, link, NULL);
 448                 ovl_drop_write(dentry);
 449         }
 450
 451         return err;
 452 }
 453
 454 static int ovl_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 455                       bool excl)
 456 {
 457         return ovl_create_object(dentry, (mode & 07777) | S_IFREG, 0, NULL);
 458 }
 459
 460 static int ovl_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 461 {
 462         return ovl_create_object(dentry, (mode & 07777) | S_IFDIR, 0, NULL);
 463 }
 464
 465 static int ovl_mknod(struct inode *dir, struct dentry *dentry, umode_t mode,
 466                      dev_t rdev)
 467 {
 468         /* Don't allow creation of "whiteout" on overlay */
 469         if (S_ISCHR(mode) && rdev == WHITEOUT_DEV)
 470                 return -EPERM;
 471
 472         return ovl_create_object(dentry, mode, rdev, NULL);
 473 }
 474
 475 static int ovl_symlink(struct inode *dir, struct dentry *dentry,
 476                        const char *link)
 477 {
 478         return ovl_create_object(dentry, S_IFLNK, 0, link);
 479 }
 480
 481 static int ovl_link(struct dentry *old, struct inode *newdir,
 482                     struct dentry *new)
 483 {
 484         int err;
 485         struct dentry *upper;
 486
 487         err = ovl_want_write(old);
 488         if (err)
 489                 goto out;
 490
 491         err = ovl_copy_up(old);
 492         if (err)
 493                 goto out_drop_write;
 494
 495         upper = ovl_dentry_upper(old);
 496         err = ovl_create_or_link(new, upper->d_inode->i_mode, 0, NULL, upper);
 497
 498 out_drop_write:
 499         ovl_drop_write(old);
 500 out:
 501         return err;
 502 }
 503
 504 static int ovl_remove_and_whiteout(struct dentry *dentry, bool is_dir)
 505 {
 506         struct dentry *workdir = ovl_workdir(dentry);
 507         struct inode *wdir = workdir->d_inode;
 508         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 509         struct inode *udir = upperdir->d_inode;
 510         struct dentry *whiteout;
 511         struct dentry *upper;
 512         struct dentry *opaquedir = NULL;
 513         int err;
 514
 515         if (WARN_ON(!workdir))
 516                 return -EROFS;
 517
 518         if (is_dir) {
 519                 if (OVL_TYPE_MERGE_OR_LOWER(ovl_path_type(dentry))) {
 520                         opaquedir = ovl_check_empty_and_clear(dentry);
 521                         err = PTR_ERR(opaquedir);
 522                         if (IS_ERR(opaquedir))
 523                                 goto out;
 524                 } else {
 525                         LIST_HEAD(list);
 526
 527                         /*
 528                          * When removing an empty opaque directory, then it
 529                          * makes no sense to replace it with an exact replica of
 530                          * itself.  But emptiness still needs to be checked.
 531                          */
 532                         err = ovl_check_empty_dir(dentry, &list);
 533                         ovl_cache_free(&list);
 534                         if (err)
 535                                 goto out;
 536                 }
 537         }
 538
 539         err = ovl_lock_rename_workdir(workdir, upperdir);
 540         if (err)
 541                 goto out_dput;
 542
 543         whiteout = ovl_whiteout(workdir, dentry);
 544         err = PTR_ERR(whiteout);
 545         if (IS_ERR(whiteout))
 546                 goto out_unlock;
 547
 548         upper = ovl_dentry_upper(dentry);
 549         if (!upper) {
 550                 upper = lookup_one_len(dentry->d_name.name, upperdir,
 551                                        dentry->d_name.len);
 552                 err = PTR_ERR(upper);
 553                 if (IS_ERR(upper))
 554                         goto kill_whiteout;
 555
 556                 err = ovl_do_rename(wdir, whiteout, udir, upper, 0);
 557                 dput(upper);
 558                 if (err)
 559                         goto kill_whiteout;
 560         } else {
 561                 int flags = 0;
 562
 563                 if (opaquedir)
 564                         upper = opaquedir;
 565                 err = -ESTALE;
 566                 if (upper->d_parent != upperdir)
 567                         goto kill_whiteout;
 568
 569                 if (is_dir)
 570                         flags |= RENAME_EXCHANGE;
 571
 572                 err = ovl_do_rename(wdir, whiteout, udir, upper, flags);
 573                 if (err)
 574                         goto kill_whiteout;
 575
 576                 if (is_dir)
 577                         ovl_cleanup(wdir, upper);
 578         }
 579         ovl_dentry_version_inc(dentry->d_parent);
 580 out_d_drop:
 581         d_drop(dentry);
 582         dput(whiteout);
 583 out_unlock:
 584         unlock_rename(workdir, upperdir);
 585 out_dput:
 586         dput(opaquedir);
 587 out:
 588         return err;
 589
 590 kill_whiteout:
 591         ovl_cleanup(wdir, whiteout);
 592         goto out_d_drop;
 593 }
 594
 595 static int ovl_remove_upper(struct dentry *dentry, bool is_dir)
 596 {
 597         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 598         struct inode *dir = upperdir->d_inode;
 599         struct dentry *upper = ovl_dentry_upper(dentry);
 600         int err;
 601
 602         mutex_lock_nested(&dir->i_mutex, I_MUTEX_PARENT);
 603         err = -ESTALE;
 604         if (upper->d_parent == upperdir) {
 605                 /* Don't let d_delete() think it can reset d_inode */
 606                 dget(upper);
 607                 if (is_dir)
 608                         err = vfs_rmdir(dir, upper);
 609                 else
 610                         err = vfs_unlink(dir, upper, NULL);
 611                 dput(upper);
 612                 ovl_dentry_version_inc(dentry->d_parent);
 613         }
 614
 615         /*
 616          * Keeping this dentry hashed would mean having to release
 617          * upperpath/lowerpath, which could only be done if we are the
 618          * sole user of this dentry.  Too tricky...  Just unhash for
 619          * now.
 620          */
 621         if (!err)
 622                 d_drop(dentry);
 623         mutex_unlock(&dir->i_mutex);
 624
 625         return err;
 626 }
 627
 628 static inline int ovl_check_sticky(struct dentry *dentry)
 629 {
 630         struct inode *dir = ovl_dentry_real(dentry->d_parent)->d_inode;
 631         struct inode *inode = ovl_dentry_real(dentry)->d_inode;
 632
 633         if (check_sticky(dir, inode))
 634                 return -EPERM;
 635
 636         return 0;
 637 }
 638
 639 static int ovl_do_remove(struct dentry *dentry, bool is_dir)
 640 {
 641         enum ovl_path_type type;
 642         int err;
 643
 644         err = ovl_check_sticky(dentry);
 645         if (err)
 646                 goto out;
 647
 648         err = ovl_want_write(dentry);
 649         if (err)
 650                 goto out;
 651
 652         err = ovl_copy_up(dentry->d_parent);
 653         if (err)
 654                 goto out_drop_write;
 655
 656         type = ovl_path_type(dentry);
 657         if (OVL_TYPE_PURE_UPPER(type)) {
 658                 err = ovl_remove_upper(dentry, is_dir);
 659         } else {
 660                 const struct cred *old_cred;
 661                 struct cred *override_cred;
 662
 663                 err = -ENOMEM;
 664                 override_cred = prepare_creds();
 665                 if (!override_cred)
 666                         goto out_drop_write;
 667
 668                 /*
 669                  * CAP_SYS_ADMIN for setting xattr on whiteout, opaque dir
 670                  * CAP_DAC_OVERRIDE for create in workdir, rename
 671                  * CAP_FOWNER for removing whiteout from sticky dir
 672                  * CAP_FSETID for chmod of opaque dir
 673                  * CAP_CHOWN for chown of opaque dir
 674                  */
 675                 cap_raise(override_cred->cap_effective, CAP_SYS_ADMIN);
 676                 cap_raise(override_cred->cap_effective, CAP_DAC_OVERRIDE);
 677                 cap_raise(override_cred->cap_effective, CAP_FOWNER);
 678                 cap_raise(override_cred->cap_effective, CAP_FSETID);
 679                 cap_raise(override_cred->cap_effective, CAP_CHOWN);
 680                 old_cred = override_creds(override_cred);
 681
 682                 err = ovl_remove_and_whiteout(dentry, is_dir);
 683
 684                 revert_creds(old_cred);
 685                 put_cred(override_cred);
 686         }
 687 out_drop_write:
 688         ovl_drop_write(dentry);
 689 out:
 690         return err;
 691 }
 692
 693 static int ovl_unlink(struct inode *dir, struct dentry *dentry)
 694 {
 695         return ovl_do_remove(dentry, false);
 696 }
 697
 698 static int ovl_rmdir(struct inode *dir, struct dentry *dentry)
 699 {
 700         return ovl_do_remove(dentry, true);
 701 }
 702
 703 static int ovl_rename2(struct inode *olddir, struct dentry *old,
 704                        struct inode *newdir, struct dentry *new,
 705                        unsigned int flags)
 706 {
 707         int err;
 708         enum ovl_path_type old_type;
 709         enum ovl_path_type new_type;
 710         struct dentry *old_upperdir;
 711         struct dentry *new_upperdir;
 712         struct dentry *olddentry;
 713         struct dentry *newdentry;
 714         struct dentry *trap;
 715         bool old_opaque;
 716         bool new_opaque;
 717         bool new_create = false;
 718         bool cleanup_whiteout = false;
 719         bool overwrite = !(flags & RENAME_EXCHANGE);
 720         bool is_dir = d_is_dir(old);
 721         bool new_is_dir = false;
 722         struct dentry *opaquedir = NULL;
 723         const struct cred *old_cred = NULL;
 724         struct cred *override_cred = NULL;
 725
 726         err = -EINVAL;
 727         if (flags & ~(RENAME_EXCHANGE | RENAME_NOREPLACE))
 728                 goto out;
 729
 730         flags &= ~RENAME_NOREPLACE;
 731
 732         err = ovl_check_sticky(old);
 733         if (err)
 734                 goto out;
 735
 736         /* Don't copy up directory trees */
 737         old_type = ovl_path_type(old);
 738         err = -EXDEV;
 739         if (OVL_TYPE_MERGE_OR_LOWER(old_type) && is_dir)
 740                 goto out;
 741
 742         if (new->d_inode) {
 743                 err = ovl_check_sticky(new);
 744                 if (err)
 745                         goto out;
 746
 747                 if (d_is_dir(new))
 748                         new_is_dir = true;
 749
 750                 new_type = ovl_path_type(new);
 751                 err = -EXDEV;
 752                 if (!overwrite && OVL_TYPE_MERGE_OR_LOWER(new_type) && new_is_dir)
 753                         goto out;
 754
 755                 err = 0;
 756                 if (!OVL_TYPE_UPPER(new_type) && !OVL_TYPE_UPPER(old_type)) {
 757                         if (ovl_dentry_lower(old)->d_inode ==
 758                             ovl_dentry_lower(new)->d_inode)
 759                                 goto out;
 760                 }
 761                 if (OVL_TYPE_UPPER(new_type) && OVL_TYPE_UPPER(old_type)) {
 762                         if (ovl_dentry_upper(old)->d_inode ==
 763                             ovl_dentry_upper(new)->d_inode)
 764                                 goto out;
 765                 }
 766         } else {
 767                 if (ovl_dentry_is_opaque(new))
 768                         new_type = __OVL_PATH_UPPER;
 769                 else
 770                         new_type = __OVL_PATH_UPPER | __OVL_PATH_PURE;
 771         }
 772
 773         err = ovl_want_write(old);
 774         if (err)
 775                 goto out;
 776
 777         err = ovl_copy_up(old);
 778         if (err)
 779                 goto out_drop_write;
 780
 781         err = ovl_copy_up(new->d_parent);
 782         if (err)
 783                 goto out_drop_write;
 784         if (!overwrite) {
 785                 err = ovl_copy_up(new);
 786                 if (err)
 787                         goto out_drop_write;
 788         }
 789
 790         old_opaque = !OVL_TYPE_PURE_UPPER(old_type);
 791         new_opaque = !OVL_TYPE_PURE_UPPER(new_type);
 792
 793         if (old_opaque || new_opaque) {
 794                 err = -ENOMEM;
 795                 override_cred = prepare_creds();
 796                 if (!override_cred)
 797                         goto out_drop_write;
 798
 799                 /*
 800                  * CAP_SYS_ADMIN for setting xattr on whiteout, opaque dir
 801                  * CAP_DAC_OVERRIDE for create in workdir
 802                  * CAP_FOWNER for removing whiteout from sticky dir
 803                  * CAP_FSETID for chmod of opaque dir
 804                  * CAP_CHOWN for chown of opaque dir
 805                  */
 806                 cap_raise(override_cred->cap_effective, CAP_SYS_ADMIN);
 807                 cap_raise(override_cred->cap_effective, CAP_DAC_OVERRIDE);
 808                 cap_raise(override_cred->cap_effective, CAP_FOWNER);
 809                 cap_raise(override_cred->cap_effective, CAP_FSETID);
 810                 cap_raise(override_cred->cap_effective, CAP_CHOWN);
 811                 old_cred = override_creds(override_cred);
 812         }
 813
 814         if (overwrite && OVL_TYPE_MERGE_OR_LOWER(new_type) && new_is_dir) {
 815                 opaquedir = ovl_check_empty_and_clear(new);
 816                 err = PTR_ERR(opaquedir);
 817                 if (IS_ERR(opaquedir)) {
 818                         opaquedir = NULL;
 819                         goto out_revert_creds;
 820                 }
 821         }
 822
 823         if (overwrite) {
 824                 if (old_opaque) {
 825                         if (new->d_inode || !new_opaque) {
 826                                 /* Whiteout source */
 827                                 flags |= RENAME_WHITEOUT;
 828                         } else {
 829                                 /* Switch whiteouts */
 830                                 flags |= RENAME_EXCHANGE;
 831                         }
 832                 } else if (is_dir && !new->d_inode && new_opaque) {
 833                         flags |= RENAME_EXCHANGE;
 834                         cleanup_whiteout = true;
 835                 }
 836         }
 837
 838         old_upperdir = ovl_dentry_upper(old->d_parent);
 839         new_upperdir = ovl_dentry_upper(new->d_parent);
 840
 841         trap = lock_rename(new_upperdir, old_upperdir);
 842
 843         olddentry = ovl_dentry_upper(old);
 844         newdentry = ovl_dentry_upper(new);
 845         if (newdentry) {
 846                 if (opaquedir) {
 847                         newdentry = opaquedir;
 848                         opaquedir = NULL;
 849                 } else {
 850                         dget(newdentry);
 851                 }
 852         } else {
 853                 new_create = true;
 854                 newdentry = lookup_one_len(new->d_name.name, new_upperdir,
 855                                            new->d_name.len);
 856                 err = PTR_ERR(newdentry);
 857                 if (IS_ERR(newdentry))
 858                         goto out_unlock;
 859         }
 860
 861         err = -ESTALE;
 862         if (olddentry->d_parent != old_upperdir)
 863                 goto out_dput;
 864         if (newdentry->d_parent != new_upperdir)
 865                 goto out_dput;
 866         if (olddentry == trap)
 867                 goto out_dput;
 868         if (newdentry == trap)
 869                 goto out_dput;
 870
 871         if (is_dir && !old_opaque && new_opaque) {
 872                 err = ovl_set_opaque(olddentry);
 873                 if (err)
 874                         goto out_dput;
 875         }
 876         if (!overwrite && new_is_dir && old_opaque && !new_opaque) {
 877                 err = ovl_set_opaque(newdentry);
 878                 if (err)
 879                         goto out_dput;
 880         }
 881
 882         if (old_opaque || new_opaque) {
 883                 err = ovl_do_rename(old_upperdir->d_inode, olddentry,
 884                                     new_upperdir->d_inode, newdentry,
 885                                     flags);
 886         } else {
 887                 /* No debug for the plain case */
 888                 BUG_ON(flags & ~RENAME_EXCHANGE);
 889                 err = vfs_rename(old_upperdir->d_inode, olddentry,
 890                                  new_upperdir->d_inode, newdentry,
 891                                  NULL, flags);
 892         }
 893
 894         if (err) {
 895                 if (is_dir && !old_opaque && new_opaque)
 896                         ovl_remove_opaque(olddentry);
 897                 if (!overwrite && new_is_dir && old_opaque && !new_opaque)
 898                         ovl_remove_opaque(newdentry);
 899                 goto out_dput;
 900         }
 901
 902         if (is_dir && old_opaque && !new_opaque)
 903                 ovl_remove_opaque(olddentry);
 904         if (!overwrite && new_is_dir && !old_opaque && new_opaque)
 905                 ovl_remove_opaque(newdentry);
 906
 907         /*
 908          * Old dentry now lives in different location. Dentries in
 909          * lowerstack are stale. We cannot drop them here because
 910          * access to them is lockless. This could be only pure upper
 911          * or opaque directory - numlower is zero. Or upper non-dir
 912          * entry - its pureness is tracked by flag opaque.
 913          */
 914         if (old_opaque != new_opaque) {
 915                 ovl_dentry_set_opaque(old, new_opaque);
 916                 if (!overwrite)
 917                         ovl_dentry_set_opaque(new, old_opaque);
 918         }
 919
 920         if (cleanup_whiteout)
 921                 ovl_cleanup(old_upperdir->d_inode, newdentry);
 922
 923         ovl_dentry_version_inc(old->d_parent);
 924         ovl_dentry_version_inc(new->d_parent);
 925
 926 out_dput:
 927         dput(newdentry);
 928 out_unlock:
 929         unlock_rename(new_upperdir, old_upperdir);
 930 out_revert_creds:
 931         if (old_opaque || new_opaque) {
 932                 revert_creds(old_cred);
 933                 put_cred(override_cred);
 934         }
 935 out_drop_write:
 936         ovl_drop_write(old);
 937 out:
 938         dput(opaquedir);
 939         return err;
 940 }
 941
 942 const struct inode_operations ovl_dir_inode_operations = {
 943         .lookup         = ovl_lookup,
 944         .mkdir          = ovl_mkdir,
 945         .symlink        = ovl_symlink,
 946         .unlink         = ovl_unlink,
 947         .rmdir          = ovl_rmdir,
 948         .rename2        = ovl_rename2,
 949         .link           = ovl_link,
 950         .setattr        = ovl_setattr,
 951         .create         = ovl_create,
 952         .mknod          = ovl_mknod,
 953         .permission     = ovl_permission,
 954         .getattr        = ovl_dir_getattr,
 955         .setxattr       = ovl_setxattr,
 956         .getxattr       = ovl_getxattr,
 957         .listxattr      = ovl_listxattr,
 958         .removexattr    = ovl_removexattr,
 959 };