fs/overlayfs/dir.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *
   4  * Copyright (C) 2011 Novell Inc.
   5  */
   6
   7 #include <linux/fs.h>
   8 #include <linux/namei.h>
   9 #include <linux/xattr.h>
  10 #include <linux/security.h>
  11 #include <linux/cred.h>
  12 #include <linux/module.h>
  13 #include <linux/posix_acl.h>
  14 #include <linux/posix_acl_xattr.h>
  15 #include <linux/atomic.h>
  16 #include <linux/ratelimit.h>
  17 #include "overlayfs.h"
  18
  19 static unsigned short ovl_redirect_max = 256;
  20 module_param_named(redirect_max, ovl_redirect_max, ushort, 0644);
  21 MODULE_PARM_DESC(redirect_max,
  22                  "Maximum length of absolute redirect xattr value");
  23
  24 static int ovl_set_redirect(struct dentry *dentry, bool samedir);
  25
  26 int ovl_cleanup(struct inode *wdir, struct dentry *wdentry)
  27 {
  28         int err;
  29
  30         dget(wdentry);
  31         if (d_is_dir(wdentry))
  32                 err = ovl_do_rmdir(wdir, wdentry);
  33         else
  34                 err = ovl_do_unlink(wdir, wdentry);
  35         dput(wdentry);
  36
  37         if (err) {
  38                 pr_err("cleanup of '%pd2' failed (%i)\n",
  39                        wdentry, err);
  40         }
  41
  42         return err;
  43 }
  44
  45 struct dentry *ovl_lookup_temp(struct dentry *workdir)
  46 {
  47         struct dentry *temp;
  48         char name[20];
  49         static atomic_t temp_id = ATOMIC_INIT(0);
  50
  51         /* counter is allowed to wrap, since temp dentries are ephemeral */
  52         snprintf(name, sizeof(name), "#%x", atomic_inc_return(&temp_id));
  53
  54         temp = lookup_one_len(name, workdir, strlen(name));
  55         if (!IS_ERR(temp) && temp->d_inode) {
  56                 pr_err("workdir/%s already exists\n", name);
  57                 dput(temp);
  58                 temp = ERR_PTR(-EIO);
  59         }
  60
  61         return temp;
  62 }
  63
  64 /* caller holds i_mutex on workdir */
  65 static struct dentry *ovl_whiteout(struct ovl_fs *ofs)
  66 {
  67         int err;
  68         struct dentry *whiteout;
  69         struct dentry *workdir = ofs->workdir;
  70         struct inode *wdir = workdir->d_inode;
  71
  72         if (!ofs->whiteout) {
  73                 whiteout = ovl_lookup_temp(workdir);
  74                 if (IS_ERR(whiteout))
  75                         goto out;
  76
  77                 err = ovl_do_whiteout(wdir, whiteout);
  78                 if (err) {
  79                         dput(whiteout);
  80                         whiteout = ERR_PTR(err);
  81                         goto out;
  82                 }
  83                 ofs->whiteout = whiteout;
  84         }
  85
  86         if (ofs->share_whiteout) {
  87                 whiteout = ovl_lookup_temp(workdir);
  88                 if (IS_ERR(whiteout))
  89                         goto out;
  90
  91                 err = ovl_do_link(ofs->whiteout, wdir, whiteout);
  92                 if (!err)
  93                         goto out;
  94
  95                 if (err != -EMLINK) {
  96                         pr_warn("Failed to link whiteout - disabling whiteout inode sharing(nlink=%u, err=%i)\n",
  97                                 ofs->whiteout->d_inode->i_nlink, err);
  98                         ofs->share_whiteout = false;
  99                 }
 100                 dput(whiteout);
 101         }
 102         whiteout = ofs->whiteout;
 103         ofs->whiteout = NULL;
 104 out:
 105         return whiteout;
 106 }
 107
 108 /* Caller must hold i_mutex on both workdir and dir */
 109 int ovl_cleanup_and_whiteout(struct ovl_fs *ofs, struct inode *dir,
 110                              struct dentry *dentry)
 111 {
 112         struct inode *wdir = ofs->workdir->d_inode;
 113         struct dentry *whiteout;
 114         int err;
 115         int flags = 0;
 116
 117         whiteout = ovl_whiteout(ofs);
 118         err = PTR_ERR(whiteout);
 119         if (IS_ERR(whiteout))
 120                 return err;
 121
 122         if (d_is_dir(dentry))
 123                 flags = RENAME_EXCHANGE;
 124
 125         err = ovl_do_rename(wdir, whiteout, dir, dentry, flags);
 126         if (err)
 127                 goto kill_whiteout;
 128         if (flags)
 129                 ovl_cleanup(wdir, dentry);
 130
 131 out:
 132         dput(whiteout);
 133         return err;
 134
 135 kill_whiteout:
 136         ovl_cleanup(wdir, whiteout);
 137         goto out;
 138 }
 139
 140 static int ovl_mkdir_real(struct inode *dir, struct dentry **newdentry,
 141                           umode_t mode)
 142 {
 143         int err;
 144         struct dentry *d, *dentry = *newdentry;
 145
 146         err = ovl_do_mkdir(dir, dentry, mode);
 147         if (err)
 148                 return err;
 149
 150         if (likely(!d_unhashed(dentry)))
 151                 return 0;
 152
 153         /*
 154          * vfs_mkdir() may succeed and leave the dentry passed
 155          * to it unhashed and negative. If that happens, try to
 156          * lookup a new hashed and positive dentry.
 157          */
 158         d = lookup_one_len(dentry->d_name.name, dentry->d_parent,
 159                            dentry->d_name.len);
 160         if (IS_ERR(d)) {
 161                 pr_warn("failed lookup after mkdir (%pd2, err=%i).\n",
 162                         dentry, err);
 163                 return PTR_ERR(d);
 164         }
 165         dput(dentry);
 166         *newdentry = d;
 167
 168         return 0;
 169 }
 170
 171 struct dentry *ovl_create_real(struct inode *dir, struct dentry *newdentry,
 172                                struct ovl_cattr *attr)
 173 {
 174         int err;
 175
 176         if (IS_ERR(newdentry))
 177                 return newdentry;
 178
 179         err = -ESTALE;
 180         if (newdentry->d_inode)
 181                 goto out;
 182
 183         if (attr->hardlink) {
 184                 err = ovl_do_link(attr->hardlink, dir, newdentry);
 185         } else {
 186                 switch (attr->mode & S_IFMT) {
 187                 case S_IFREG:
 188                         err = ovl_do_create(dir, newdentry, attr->mode);
 189                         break;
 190
 191                 case S_IFDIR:
 192                         /* mkdir is special... */
 193                         err =  ovl_mkdir_real(dir, &newdentry, attr->mode);
 194                         break;
 195
 196                 case S_IFCHR:
 197                 case S_IFBLK:
 198                 case S_IFIFO:
 199                 case S_IFSOCK:
 200                         err = ovl_do_mknod(dir, newdentry, attr->mode,
 201                                            attr->rdev);
 202                         break;
 203
 204                 case S_IFLNK:
 205                         err = ovl_do_symlink(dir, newdentry, attr->link);
 206                         break;
 207
 208                 default:
 209                         err = -EPERM;
 210                 }
 211         }
 212         if (!err && WARN_ON(!newdentry->d_inode)) {
 213                 /*
 214                  * Not quite sure if non-instantiated dentry is legal or not.
 215                  * VFS doesn't seem to care so check and warn here.
 216                  */
 217                 err = -EIO;
 218         }
 219 out:
 220         if (err) {
 221                 dput(newdentry);
 222                 return ERR_PTR(err);
 223         }
 224         return newdentry;
 225 }
 226
 227 struct dentry *ovl_create_temp(struct dentry *workdir, struct ovl_cattr *attr)
 228 {
 229         return ovl_create_real(d_inode(workdir), ovl_lookup_temp(workdir),
 230                                attr);
 231 }
 232
 233 static int ovl_set_opaque_xerr(struct dentry *dentry, struct dentry *upper,
 234                                int xerr)
 235 {
 236         int err;
 237
 238         err = ovl_check_setxattr(dentry, upper, OVL_XATTR_OPAQUE, "y", 1, xerr);
 239         if (!err)
 240                 ovl_dentry_set_opaque(dentry);
 241
 242         return err;
 243 }
 244
 245 static int ovl_set_opaque(struct dentry *dentry, struct dentry *upperdentry)
 246 {
 247         /*
 248          * Fail with -EIO when trying to create opaque dir and upper doesn't
 249          * support xattrs. ovl_rename() calls ovl_set_opaque_xerr(-EXDEV) to
 250          * return a specific error for noxattr case.
 251          */
 252         return ovl_set_opaque_xerr(dentry, upperdentry, -EIO);
 253 }
 254
 255 /*
 256  * Common operations required to be done after creation of file on upper.
 257  * If @hardlink is false, then @inode is a pre-allocated inode, we may or
 258  * may not use to instantiate the new dentry.
 259  */
 260 static int ovl_instantiate(struct dentry *dentry, struct inode *inode,
 261                            struct dentry *newdentry, bool hardlink)
 262 {
 263         struct ovl_inode_params oip = {
 264                 .upperdentry = newdentry,
 265                 .newinode = inode,
 266         };
 267
 268         ovl_dir_modified(dentry->d_parent, false);
 269         ovl_dentry_set_upper_alias(dentry);
 270         ovl_dentry_update_reval(dentry, newdentry,
 271                         DCACHE_OP_REVALIDATE | DCACHE_OP_WEAK_REVALIDATE);
 272
 273         if (!hardlink) {
 274                 /*
 275                  * ovl_obtain_alias() can be called after ovl_create_real()
 276                  * and before we get here, so we may get an inode from cache
 277                  * with the same real upperdentry that is not the inode we
 278                  * pre-allocated.  In this case we will use the cached inode
 279                  * to instantiate the new dentry.
 280                  *
 281                  * XXX: if we ever use ovl_obtain_alias() to decode directory
 282                  * file handles, need to use ovl_get_inode_locked() and
 283                  * d_instantiate_new() here to prevent from creating two
 284                  * hashed directory inode aliases.
 285                  */
 286                 inode = ovl_get_inode(dentry->d_sb, &oip);
 287                 if (IS_ERR(inode))
 288                         return PTR_ERR(inode);
 289                 if (inode == oip.newinode)
 290                         ovl_set_flag(OVL_UPPERDATA, inode);
 291         } else {
 292                 WARN_ON(ovl_inode_real(inode) != d_inode(newdentry));
 293                 dput(newdentry);
 294                 inc_nlink(inode);
 295         }
 296
 297         d_instantiate(dentry, inode);
 298         if (inode != oip.newinode) {
 299                 pr_warn_ratelimited("newly created inode found in cache (%pd2)\n",
 300                                     dentry);
 301         }
 302
 303         /* Force lookup of new upper hardlink to find its lower */
 304         if (hardlink)
 305                 d_drop(dentry);
 306
 307         return 0;
 308 }
 309
 310 static bool ovl_type_merge(struct dentry *dentry)
 311 {
 312         return OVL_TYPE_MERGE(ovl_path_type(dentry));
 313 }
 314
 315 static bool ovl_type_origin(struct dentry *dentry)
 316 {
 317         return OVL_TYPE_ORIGIN(ovl_path_type(dentry));
 318 }
 319
 320 static int ovl_create_upper(struct dentry *dentry, struct inode *inode,
 321                             struct ovl_cattr *attr)
 322 {
 323         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 324         struct inode *udir = upperdir->d_inode;
 325         struct dentry *newdentry;
 326         int err;
 327
 328         if (!attr->hardlink && !IS_POSIXACL(udir))
 329                 attr->mode &= ~current_umask();
 330
 331         inode_lock_nested(udir, I_MUTEX_PARENT);
 332         newdentry = ovl_create_real(udir,
 333                                     lookup_one_len(dentry->d_name.name,
 334                                                    upperdir,
 335                                                    dentry->d_name.len),
 336                                     attr);
 337         err = PTR_ERR(newdentry);
 338         if (IS_ERR(newdentry))
 339                 goto out_unlock;
 340
 341         if (ovl_type_merge(dentry->d_parent) && d_is_dir(newdentry)) {
 342                 /* Setting opaque here is just an optimization, allow to fail */
 343                 ovl_set_opaque(dentry, newdentry);
 344         }
 345
 346         err = ovl_instantiate(dentry, inode, newdentry, !!attr->hardlink);
 347         if (err)
 348                 goto out_cleanup;
 349 out_unlock:
 350         inode_unlock(udir);
 351         return err;
 352
 353 out_cleanup:
 354         ovl_cleanup(udir, newdentry);
 355         dput(newdentry);
 356         goto out_unlock;
 357 }
 358
 359 static struct dentry *ovl_clear_empty(struct dentry *dentry,
 360                                       struct list_head *list)
 361 {
 362         struct dentry *workdir = ovl_workdir(dentry);
 363         struct inode *wdir = workdir->d_inode;
 364         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 365         struct inode *udir = upperdir->d_inode;
 366         struct path upperpath;
 367         struct dentry *upper;
 368         struct dentry *opaquedir;
 369         struct kstat stat;
 370         int err;
 371
 372         if (WARN_ON(!workdir))
 373                 return ERR_PTR(-EROFS);
 374
 375         err = ovl_lock_rename_workdir(workdir, upperdir);
 376         if (err)
 377                 goto out;
 378
 379         ovl_path_upper(dentry, &upperpath);
 380         err = vfs_getattr(&upperpath, &stat,
 381                           STATX_BASIC_STATS, AT_STATX_SYNC_AS_STAT);
 382         if (err)
 383                 goto out_unlock;
 384
 385         err = -ESTALE;
 386         if (!S_ISDIR(stat.mode))
 387                 goto out_unlock;
 388         upper = upperpath.dentry;
 389         if (upper->d_parent->d_inode != udir)
 390                 goto out_unlock;
 391
 392         opaquedir = ovl_create_temp(workdir, OVL_CATTR(stat.mode));
 393         err = PTR_ERR(opaquedir);
 394         if (IS_ERR(opaquedir))
 395                 goto out_unlock;
 396
 397         err = ovl_copy_xattr(dentry->d_sb, upper, opaquedir);
 398         if (err)
 399                 goto out_cleanup;
 400
 401         err = ovl_set_opaque(dentry, opaquedir);
 402         if (err)
 403                 goto out_cleanup;
 404
 405         inode_lock(opaquedir->d_inode);
 406         err = ovl_set_attr(opaquedir, &stat);
 407         inode_unlock(opaquedir->d_inode);
 408         if (err)
 409                 goto out_cleanup;
 410
 411         err = ovl_do_rename(wdir, opaquedir, udir, upper, RENAME_EXCHANGE);
 412         if (err)
 413                 goto out_cleanup;
 414
 415         ovl_cleanup_whiteouts(upper, list);
 416         ovl_cleanup(wdir, upper);
 417         unlock_rename(workdir, upperdir);
 418
 419         /* dentry's upper doesn't match now, get rid of it */
 420         d_drop(dentry);
 421
 422         return opaquedir;
 423
 424 out_cleanup:
 425         ovl_cleanup(wdir, opaquedir);
 426         dput(opaquedir);
 427 out_unlock:
 428         unlock_rename(workdir, upperdir);
 429 out:
 430         return ERR_PTR(err);
 431 }
 432
 433 static int ovl_set_upper_acl(struct dentry *upperdentry, const char *name,
 434                              const struct posix_acl *acl)
 435 {
 436         void *buffer;
 437         size_t size;
 438         int err;
 439
 440         if (!IS_ENABLED(CONFIG_FS_POSIX_ACL) || !acl)
 441                 return 0;
 442
 443         size = posix_acl_xattr_size(acl->a_count);
 444         buffer = kmalloc(size, GFP_KERNEL);
 445         if (!buffer)
 446                 return -ENOMEM;
 447
 448         err = posix_acl_to_xattr(&init_user_ns, acl, buffer, size);
 449         if (err < 0)
 450                 goto out_free;
 451
 452         err = vfs_setxattr(upperdentry, name, buffer, size, XATTR_CREATE);
 453 out_free:
 454         kfree(buffer);
 455         return err;
 456 }
 457
 458 static int ovl_create_over_whiteout(struct dentry *dentry, struct inode *inode,
 459                                     struct ovl_cattr *cattr)
 460 {
 461         struct dentry *workdir = ovl_workdir(dentry);
 462         struct inode *wdir = workdir->d_inode;
 463         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 464         struct inode *udir = upperdir->d_inode;
 465         struct dentry *upper;
 466         struct dentry *newdentry;
 467         int err;
 468         struct posix_acl *acl, *default_acl;
 469         bool hardlink = !!cattr->hardlink;
 470
 471         if (WARN_ON(!workdir))
 472                 return -EROFS;
 473
 474         if (!hardlink) {
 475                 err = posix_acl_create(dentry->d_parent->d_inode,
 476                                        &cattr->mode, &default_acl, &acl);
 477                 if (err)
 478                         return err;
 479         }
 480
 481         err = ovl_lock_rename_workdir(workdir, upperdir);
 482         if (err)
 483                 goto out;
 484
 485         upper = lookup_one_len(dentry->d_name.name, upperdir,
 486                                dentry->d_name.len);
 487         err = PTR_ERR(upper);
 488         if (IS_ERR(upper))
 489                 goto out_unlock;
 490
 491         err = -ESTALE;
 492         if (d_is_negative(upper) || !IS_WHITEOUT(d_inode(upper)))
 493                 goto out_dput;
 494
 495         newdentry = ovl_create_temp(workdir, cattr);
 496         err = PTR_ERR(newdentry);
 497         if (IS_ERR(newdentry))
 498                 goto out_dput;
 499
 500         /*
 501          * mode could have been mutilated due to umask (e.g. sgid directory)
 502          */
 503         if (!hardlink &&
 504             !S_ISLNK(cattr->mode) &&
 505             newdentry->d_inode->i_mode != cattr->mode) {
 506                 struct iattr attr = {
 507                         .ia_valid = ATTR_MODE,
 508                         .ia_mode = cattr->mode,
 509                 };
 510                 inode_lock(newdentry->d_inode);
 511                 err = notify_change(newdentry, &attr, NULL);
 512                 inode_unlock(newdentry->d_inode);
 513                 if (err)
 514                         goto out_cleanup;
 515         }
 516         if (!hardlink) {
 517                 err = ovl_set_upper_acl(newdentry, XATTR_NAME_POSIX_ACL_ACCESS,
 518                                         acl);
 519                 if (err)
 520                         goto out_cleanup;
 521
 522                 err = ovl_set_upper_acl(newdentry, XATTR_NAME_POSIX_ACL_DEFAULT,
 523                                         default_acl);
 524                 if (err)
 525                         goto out_cleanup;
 526         }
 527
 528         if (!hardlink && S_ISDIR(cattr->mode)) {
 529                 err = ovl_set_opaque(dentry, newdentry);
 530                 if (err)
 531                         goto out_cleanup;
 532
 533                 err = ovl_do_rename(wdir, newdentry, udir, upper,
 534                                     RENAME_EXCHANGE);
 535                 if (err)
 536                         goto out_cleanup;
 537
 538                 ovl_cleanup(wdir, upper);
 539         } else {
 540                 err = ovl_do_rename(wdir, newdentry, udir, upper, 0);
 541                 if (err)
 542                         goto out_cleanup;
 543         }
 544         err = ovl_instantiate(dentry, inode, newdentry, hardlink);
 545         if (err)
 546                 goto out_cleanup;
 547 out_dput:
 548         dput(upper);
 549 out_unlock:
 550         unlock_rename(workdir, upperdir);
 551 out:
 552         if (!hardlink) {
 553                 posix_acl_release(acl);
 554                 posix_acl_release(default_acl);
 555         }
 556         return err;
 557
 558 out_cleanup:
 559         ovl_cleanup(wdir, newdentry);
 560         dput(newdentry);
 561         goto out_dput;
 562 }
 563
 564 static int ovl_create_or_link(struct dentry *dentry, struct inode *inode,
 565                               struct ovl_cattr *attr, bool origin)
 566 {
 567         int err;
 568         const struct cred *old_cred;
 569         struct cred *override_cred;
 570         struct dentry *parent = dentry->d_parent;
 571
 572         err = ovl_copy_up(parent);
 573         if (err)
 574                 return err;
 575
 576         old_cred = ovl_override_creds(dentry->d_sb);
 577
 578         /*
 579          * When linking a file with copy up origin into a new parent, mark the
 580          * new parent dir "impure".
 581          */
 582         if (origin) {
 583                 err = ovl_set_impure(parent, ovl_dentry_upper(parent));
 584                 if (err)
 585                         goto out_revert_creds;
 586         }
 587
 588         err = -ENOMEM;
 589         override_cred = prepare_creds();
 590         if (override_cred) {
 591                 override_cred->fsuid = inode->i_uid;
 592                 override_cred->fsgid = inode->i_gid;
 593                 if (!attr->hardlink) {
 594                         err = security_dentry_create_files_as(dentry,
 595                                         attr->mode, &dentry->d_name, old_cred,
 596                                         override_cred);
 597                         if (err) {
 598                                 put_cred(override_cred);
 599                                 goto out_revert_creds;
 600                         }
 601                 }
 602                 put_cred(override_creds(override_cred));
 603                 put_cred(override_cred);
 604
 605                 if (!ovl_dentry_is_whiteout(dentry))
 606                         err = ovl_create_upper(dentry, inode, attr);
 607                 else
 608                         err = ovl_create_over_whiteout(dentry, inode, attr);
 609         }
 610 out_revert_creds:
 611         revert_creds(old_cred);
 612         return err;
 613 }
 614
 615 static int ovl_create_object(struct dentry *dentry, int mode, dev_t rdev,
 616                              const char *link)
 617 {
 618         int err;
 619         struct inode *inode;
 620         struct ovl_cattr attr = {
 621                 .rdev = rdev,
 622                 .link = link,
 623         };
 624
 625         err = ovl_want_write(dentry);
 626         if (err)
 627                 goto out;
 628
 629         /* Preallocate inode to be used by ovl_get_inode() */
 630         err = -ENOMEM;
 631         inode = ovl_new_inode(dentry->d_sb, mode, rdev);
 632         if (!inode)
 633                 goto out_drop_write;
 634
 635         spin_lock(&inode->i_lock);
 636         inode->i_state |= I_CREATING;
 637         spin_unlock(&inode->i_lock);
 638
 639         inode_init_owner(inode, dentry->d_parent->d_inode, mode);
 640         attr.mode = inode->i_mode;
 641
 642         err = ovl_create_or_link(dentry, inode, &attr, false);
 643         /* Did we end up using the preallocated inode? */
 644         if (inode != d_inode(dentry))
 645                 iput(inode);
 646
 647 out_drop_write:
 648         ovl_drop_write(dentry);
 649 out:
 650         return err;
 651 }
 652
 653 static int ovl_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 654                       bool excl)
 655 {
 656         return ovl_create_object(dentry, (mode & 07777) | S_IFREG, 0, NULL);
 657 }
 658
 659 static int ovl_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 660 {
 661         return ovl_create_object(dentry, (mode & 07777) | S_IFDIR, 0, NULL);
 662 }
 663
 664 static int ovl_mknod(struct inode *dir, struct dentry *dentry, umode_t mode,
 665                      dev_t rdev)
 666 {
 667         /* Don't allow creation of "whiteout" on overlay */
 668         if (S_ISCHR(mode) && rdev == WHITEOUT_DEV)
 669                 return -EPERM;
 670
 671         return ovl_create_object(dentry, mode, rdev, NULL);
 672 }
 673
 674 static int ovl_symlink(struct inode *dir, struct dentry *dentry,
 675                        const char *link)
 676 {
 677         return ovl_create_object(dentry, S_IFLNK, 0, link);
 678 }
 679
 680 static int ovl_set_link_redirect(struct dentry *dentry)
 681 {
 682         const struct cred *old_cred;
 683         int err;
 684
 685         old_cred = ovl_override_creds(dentry->d_sb);
 686         err = ovl_set_redirect(dentry, false);
 687         revert_creds(old_cred);
 688
 689         return err;
 690 }
 691
 692 static int ovl_link(struct dentry *old, struct inode *newdir,
 693                     struct dentry *new)
 694 {
 695         int err;
 696         struct inode *inode;
 697
 698         err = ovl_want_write(old);
 699         if (err)
 700                 goto out;
 701
 702         err = ovl_copy_up(old);
 703         if (err)
 704                 goto out_drop_write;
 705
 706         err = ovl_copy_up(new->d_parent);
 707         if (err)
 708                 goto out_drop_write;
 709
 710         if (ovl_is_metacopy_dentry(old)) {
 711                 err = ovl_set_link_redirect(old);
 712                 if (err)
 713                         goto out_drop_write;
 714         }
 715
 716         err = ovl_nlink_start(old);
 717         if (err)
 718                 goto out_drop_write;
 719
 720         inode = d_inode(old);
 721         ihold(inode);
 722
 723         err = ovl_create_or_link(new, inode,
 724                         &(struct ovl_cattr) {.hardlink = ovl_dentry_upper(old)},
 725                         ovl_type_origin(old));
 726         if (err)
 727                 iput(inode);
 728
 729         ovl_nlink_end(old);
 730 out_drop_write:
 731         ovl_drop_write(old);
 732 out:
 733         return err;
 734 }
 735
 736 static bool ovl_matches_upper(struct dentry *dentry, struct dentry *upper)
 737 {
 738         return d_inode(ovl_dentry_upper(dentry)) == d_inode(upper);
 739 }
 740
 741 static int ovl_remove_and_whiteout(struct dentry *dentry,
 742                                    struct list_head *list)
 743 {
 744         struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 745         struct dentry *workdir = ovl_workdir(dentry);
 746         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 747         struct dentry *upper;
 748         struct dentry *opaquedir = NULL;
 749         int err;
 750
 751         if (WARN_ON(!workdir))
 752                 return -EROFS;
 753
 754         if (!list_empty(list)) {
 755                 opaquedir = ovl_clear_empty(dentry, list);
 756                 err = PTR_ERR(opaquedir);
 757                 if (IS_ERR(opaquedir))
 758                         goto out;
 759         }
 760
 761         err = ovl_lock_rename_workdir(workdir, upperdir);
 762         if (err)
 763                 goto out_dput;
 764
 765         upper = lookup_one_len(dentry->d_name.name, upperdir,
 766                                dentry->d_name.len);
 767         err = PTR_ERR(upper);
 768         if (IS_ERR(upper))
 769                 goto out_unlock;
 770
 771         err = -ESTALE;
 772         if ((opaquedir && upper != opaquedir) ||
 773             (!opaquedir && ovl_dentry_upper(dentry) &&
 774              !ovl_matches_upper(dentry, upper))) {
 775                 goto out_dput_upper;
 776         }
 777
 778         err = ovl_cleanup_and_whiteout(ofs, d_inode(upperdir), upper);
 779         if (err)
 780                 goto out_d_drop;
 781
 782         ovl_dir_modified(dentry->d_parent, true);
 783 out_d_drop:
 784         d_drop(dentry);
 785 out_dput_upper:
 786         dput(upper);
 787 out_unlock:
 788         unlock_rename(workdir, upperdir);
 789 out_dput:
 790         dput(opaquedir);
 791 out:
 792         return err;
 793 }
 794
 795 static int ovl_remove_upper(struct dentry *dentry, bool is_dir,
 796                             struct list_head *list)
 797 {
 798         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 799         struct inode *dir = upperdir->d_inode;
 800         struct dentry *upper;
 801         struct dentry *opaquedir = NULL;
 802         int err;
 803
 804         if (!list_empty(list)) {
 805                 opaquedir = ovl_clear_empty(dentry, list);
 806                 err = PTR_ERR(opaquedir);
 807                 if (IS_ERR(opaquedir))
 808                         goto out;
 809         }
 810
 811         inode_lock_nested(dir, I_MUTEX_PARENT);
 812         upper = lookup_one_len(dentry->d_name.name, upperdir,
 813                                dentry->d_name.len);
 814         err = PTR_ERR(upper);
 815         if (IS_ERR(upper))
 816                 goto out_unlock;
 817
 818         err = -ESTALE;
 819         if ((opaquedir && upper != opaquedir) ||
 820             (!opaquedir && !ovl_matches_upper(dentry, upper)))
 821                 goto out_dput_upper;
 822
 823         if (is_dir)
 824                 err = vfs_rmdir(dir, upper);
 825         else
 826                 err = vfs_unlink(dir, upper, NULL);
 827         ovl_dir_modified(dentry->d_parent, ovl_type_origin(dentry));
 828
 829         /*
 830          * Keeping this dentry hashed would mean having to release
 831          * upperpath/lowerpath, which could only be done if we are the
 832          * sole user of this dentry.  Too tricky...  Just unhash for
 833          * now.
 834          */
 835         if (!err)
 836                 d_drop(dentry);
 837 out_dput_upper:
 838         dput(upper);
 839 out_unlock:
 840         inode_unlock(dir);
 841         dput(opaquedir);
 842 out:
 843         return err;
 844 }
 845
 846 static bool ovl_pure_upper(struct dentry *dentry)
 847 {
 848         return !ovl_dentry_lower(dentry) &&
 849                !ovl_test_flag(OVL_WHITEOUTS, d_inode(dentry));
 850 }
 851
 852 static void ovl_drop_nlink(struct dentry *dentry)
 853 {
 854         struct inode *inode = d_inode(dentry);
 855         struct dentry *alias;
 856
 857         /* Try to find another, hashed alias */
 858         spin_lock(&inode->i_lock);
 859         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
 860                 if (alias != dentry && !d_unhashed(alias))
 861                         break;
 862         }
 863         spin_unlock(&inode->i_lock);
 864
 865         /*
 866          * Changes to underlying layers may cause i_nlink to lose sync with
 867          * reality.  In this case prevent the link count from going to zero
 868          * prematurely.
 869          */
 870         if (inode->i_nlink > !!alias)
 871                 drop_nlink(inode);
 872 }
 873
 874 static int ovl_do_remove(struct dentry *dentry, bool is_dir)
 875 {
 876         int err;
 877         const struct cred *old_cred;
 878         struct dentry *upperdentry;
 879         bool lower_positive = ovl_lower_positive(dentry);
 880         LIST_HEAD(list);
 881
 882         /* No need to clean pure upper removed by vfs_rmdir() */
 883         if (is_dir && (lower_positive || !ovl_pure_upper(dentry))) {
 884                 err = ovl_check_empty_dir(dentry, &list);
 885                 if (err)
 886                         goto out;
 887         }
 888
 889         err = ovl_want_write(dentry);
 890         if (err)
 891                 goto out;
 892
 893         err = ovl_copy_up(dentry->d_parent);
 894         if (err)
 895                 goto out_drop_write;
 896
 897         err = ovl_nlink_start(dentry);
 898         if (err)
 899                 goto out_drop_write;
 900
 901         old_cred = ovl_override_creds(dentry->d_sb);
 902         if (!lower_positive)
 903                 err = ovl_remove_upper(dentry, is_dir, &list);
 904         else
 905                 err = ovl_remove_and_whiteout(dentry, &list);
 906         revert_creds(old_cred);
 907         if (!err) {
 908                 if (is_dir)
 909                         clear_nlink(dentry->d_inode);
 910                 else
 911                         ovl_drop_nlink(dentry);
 912         }
 913         ovl_nlink_end(dentry);
 914
 915         /*
 916          * Copy ctime
 917          *
 918          * Note: we fail to update ctime if there was no copy-up, only a
 919          * whiteout
 920          */
 921         upperdentry = ovl_dentry_upper(dentry);
 922         if (upperdentry)
 923                 ovl_copyattr(d_inode(upperdentry), d_inode(dentry));
 924
 925 out_drop_write:
 926         ovl_drop_write(dentry);
 927 out:
 928         ovl_cache_free(&list);
 929         return err;
 930 }
 931
 932 static int ovl_unlink(struct inode *dir, struct dentry *dentry)
 933 {
 934         return ovl_do_remove(dentry, false);
 935 }
 936
 937 static int ovl_rmdir(struct inode *dir, struct dentry *dentry)
 938 {
 939         return ovl_do_remove(dentry, true);
 940 }
 941
 942 static bool ovl_type_merge_or_lower(struct dentry *dentry)
 943 {
 944         enum ovl_path_type type = ovl_path_type(dentry);
 945
 946         return OVL_TYPE_MERGE(type) || !OVL_TYPE_UPPER(type);
 947 }
 948
 949 static bool ovl_can_move(struct dentry *dentry)
 950 {
 951         return ovl_redirect_dir(dentry->d_sb) ||
 952                 !d_is_dir(dentry) || !ovl_type_merge_or_lower(dentry);
 953 }
 954
 955 static char *ovl_get_redirect(struct dentry *dentry, bool abs_redirect)
 956 {
 957         char *buf, *ret;
 958         struct dentry *d, *tmp;
 959         int buflen = ovl_redirect_max + 1;
 960
 961         if (!abs_redirect) {
 962                 ret = kstrndup(dentry->d_name.name, dentry->d_name.len,
 963                                GFP_KERNEL);
 964                 goto out;
 965         }
 966
 967         buf = ret = kmalloc(buflen, GFP_KERNEL);
 968         if (!buf)
 969                 goto out;
 970
 971         buflen--;
 972         buf[buflen] = '\0';
 973         for (d = dget(dentry); !IS_ROOT(d);) {
 974                 const char *name;
 975                 int thislen;
 976
 977                 spin_lock(&d->d_lock);
 978                 name = ovl_dentry_get_redirect(d);
 979                 if (name) {
 980                         thislen = strlen(name);
 981                 } else {
 982                         name = d->d_name.name;
 983                         thislen = d->d_name.len;
 984                 }
 985
 986                 /* If path is too long, fall back to userspace move */
 987                 if (thislen + (name[0] != '/') > buflen) {
 988                         ret = ERR_PTR(-EXDEV);
 989                         spin_unlock(&d->d_lock);
 990                         goto out_put;
 991                 }
 992
 993                 buflen -= thislen;
 994                 memcpy(&buf[buflen], name, thislen);
 995                 tmp = dget_dlock(d->d_parent);
 996                 spin_unlock(&d->d_lock);
 997
 998                 dput(d);
 999                 d = tmp;
1000
1001                 /* Absolute redirect: finished */
1002                 if (buf[buflen] == '/')
1003                         break;
1004                 buflen--;
1005                 buf[buflen] = '/';
1006         }
1007         ret = kstrdup(&buf[buflen], GFP_KERNEL);
1008 out_put:
1009         dput(d);
1010         kfree(buf);
1011 out:
1012         return ret ? ret : ERR_PTR(-ENOMEM);
1013 }
1014
1015 static bool ovl_need_absolute_redirect(struct dentry *dentry, bool samedir)
1016 {
1017         struct dentry *lowerdentry;
1018
1019         if (!samedir)
1020                 return true;
1021
1022         if (d_is_dir(dentry))
1023                 return false;
1024
1025         /*
1026          * For non-dir hardlinked files, we need absolute redirects
1027          * in general as two upper hardlinks could be in different
1028          * dirs. We could put a relative redirect now and convert
1029          * it to absolute redirect later. But when nlink > 1 and
1030          * indexing is on, that means relative redirect needs to be
1031          * converted to absolute during copy up of another lower
1032          * hardllink as well.
1033          *
1034          * So without optimizing too much, just check if lower is
1035          * a hard link or not. If lower is hard link, put absolute
1036          * redirect.
1037          */
1038         lowerdentry = ovl_dentry_lower(dentry);
1039         return (d_inode(lowerdentry)->i_nlink > 1);
1040 }
1041
1042 static int ovl_set_redirect(struct dentry *dentry, bool samedir)
1043 {
1044         int err;
1045         const char *redirect = ovl_dentry_get_redirect(dentry);
1046         bool absolute_redirect = ovl_need_absolute_redirect(dentry, samedir);
1047
1048         if (redirect && (!absolute_redirect || redirect[0] == '/'))
1049                 return 0;
1050
1051         redirect = ovl_get_redirect(dentry, absolute_redirect);
1052         if (IS_ERR(redirect))
1053                 return PTR_ERR(redirect);
1054
1055         err = ovl_check_setxattr(dentry, ovl_dentry_upper(dentry),
1056                                  OVL_XATTR_REDIRECT,
1057                                  redirect, strlen(redirect), -EXDEV);
1058         if (!err) {
1059                 spin_lock(&dentry->d_lock);
1060                 ovl_dentry_set_redirect(dentry, redirect);
1061                 spin_unlock(&dentry->d_lock);
1062         } else {
1063                 kfree(redirect);
1064                 pr_warn_ratelimited("failed to set redirect (%i)\n",
1065                                     err);
1066                 /* Fall back to userspace copy-up */
1067                 err = -EXDEV;
1068         }
1069         return err;
1070 }
1071
1072 static int ovl_rename(struct inode *olddir, struct dentry *old,
1073                       struct inode *newdir, struct dentry *new,
1074                       unsigned int flags)
1075 {
1076         int err;
1077         struct dentry *old_upperdir;
1078         struct dentry *new_upperdir;
1079         struct dentry *olddentry;
1080         struct dentry *newdentry;
1081         struct dentry *trap;
1082         bool old_opaque;
1083         bool new_opaque;
1084         bool cleanup_whiteout = false;
1085         bool update_nlink = false;
1086         bool overwrite = !(flags & RENAME_EXCHANGE);
1087         bool is_dir = d_is_dir(old);
1088         bool new_is_dir = d_is_dir(new);
1089         bool samedir = olddir == newdir;
1090         struct dentry *opaquedir = NULL;
1091         const struct cred *old_cred = NULL;
1092         LIST_HEAD(list);
1093
1094         err = -EINVAL;
1095         if (flags & ~(RENAME_EXCHANGE | RENAME_NOREPLACE))
1096                 goto out;
1097
1098         flags &= ~RENAME_NOREPLACE;
1099
1100         /* Don't copy up directory trees */
1101         err = -EXDEV;
1102         if (!ovl_can_move(old))
1103                 goto out;
1104         if (!overwrite && !ovl_can_move(new))
1105                 goto out;
1106
1107         if (overwrite && new_is_dir && !ovl_pure_upper(new)) {
1108                 err = ovl_check_empty_dir(new, &list);
1109                 if (err)
1110                         goto out;
1111         }
1112
1113         if (overwrite) {
1114                 if (ovl_lower_positive(old)) {
1115                         if (!ovl_dentry_is_whiteout(new)) {
1116                                 /* Whiteout source */
1117                                 flags |= RENAME_WHITEOUT;
1118                         } else {
1119                                 /* Switch whiteouts */
1120                                 flags |= RENAME_EXCHANGE;
1121                         }
1122                 } else if (is_dir && ovl_dentry_is_whiteout(new)) {
1123                         flags |= RENAME_EXCHANGE;
1124                         cleanup_whiteout = true;
1125                 }
1126         }
1127
1128         err = ovl_want_write(old);
1129         if (err)
1130                 goto out;
1131
1132         err = ovl_copy_up(old);
1133         if (err)
1134                 goto out_drop_write;
1135
1136         err = ovl_copy_up(new->d_parent);
1137         if (err)
1138                 goto out_drop_write;
1139         if (!overwrite) {
1140                 err = ovl_copy_up(new);
1141                 if (err)
1142                         goto out_drop_write;
1143         } else if (d_inode(new)) {
1144                 err = ovl_nlink_start(new);
1145                 if (err)
1146                         goto out_drop_write;
1147
1148                 update_nlink = true;
1149         }
1150
1151         old_cred = ovl_override_creds(old->d_sb);
1152
1153         if (!list_empty(&list)) {
1154                 opaquedir = ovl_clear_empty(new, &list);
1155                 err = PTR_ERR(opaquedir);
1156                 if (IS_ERR(opaquedir)) {
1157                         opaquedir = NULL;
1158                         goto out_revert_creds;
1159                 }
1160         }
1161
1162         old_upperdir = ovl_dentry_upper(old->d_parent);
1163         new_upperdir = ovl_dentry_upper(new->d_parent);
1164
1165         if (!samedir) {
1166                 /*
1167                  * When moving a merge dir or non-dir with copy up origin into
1168                  * a new parent, we are marking the new parent dir "impure".
1169                  * When ovl_iterate() iterates an "impure" upper dir, it will
1170                  * lookup the origin inodes of the entries to fill d_ino.
1171                  */
1172                 if (ovl_type_origin(old)) {
1173                         err = ovl_set_impure(new->d_parent, new_upperdir);
1174                         if (err)
1175                                 goto out_revert_creds;
1176                 }
1177                 if (!overwrite && ovl_type_origin(new)) {
1178                         err = ovl_set_impure(old->d_parent, old_upperdir);
1179                         if (err)
1180                                 goto out_revert_creds;
1181                 }
1182         }
1183
1184         trap = lock_rename(new_upperdir, old_upperdir);
1185
1186         olddentry = lookup_one_len(old->d_name.name, old_upperdir,
1187                                    old->d_name.len);
1188         err = PTR_ERR(olddentry);
1189         if (IS_ERR(olddentry))
1190                 goto out_unlock;
1191
1192         err = -ESTALE;
1193         if (!ovl_matches_upper(old, olddentry))
1194                 goto out_dput_old;
1195
1196         newdentry = lookup_one_len(new->d_name.name, new_upperdir,
1197                                    new->d_name.len);
1198         err = PTR_ERR(newdentry);
1199         if (IS_ERR(newdentry))
1200                 goto out_dput_old;
1201
1202         old_opaque = ovl_dentry_is_opaque(old);
1203         new_opaque = ovl_dentry_is_opaque(new);
1204
1205         err = -ESTALE;
1206         if (d_inode(new) && ovl_dentry_upper(new)) {
1207                 if (opaquedir) {
1208                         if (newdentry != opaquedir)
1209                                 goto out_dput;
1210                 } else {
1211                         if (!ovl_matches_upper(new, newdentry))
1212                                 goto out_dput;
1213                 }
1214         } else {
1215                 if (!d_is_negative(newdentry) &&
1216                     (!new_opaque || !ovl_is_whiteout(newdentry)))
1217                         goto out_dput;
1218         }
1219
1220         if (olddentry == trap)
1221                 goto out_dput;
1222         if (newdentry == trap)
1223                 goto out_dput;
1224
1225         if (olddentry->d_inode == newdentry->d_inode)
1226                 goto out_dput;
1227
1228         err = 0;
1229         if (ovl_type_merge_or_lower(old))
1230                 err = ovl_set_redirect(old, samedir);
1231         else if (is_dir && !old_opaque && ovl_type_merge(new->d_parent))
1232                 err = ovl_set_opaque_xerr(old, olddentry, -EXDEV);
1233         if (err)
1234                 goto out_dput;
1235
1236         if (!overwrite && ovl_type_merge_or_lower(new))
1237                 err = ovl_set_redirect(new, samedir);
1238         else if (!overwrite && new_is_dir && !new_opaque &&
1239                  ovl_type_merge(old->d_parent))
1240                 err = ovl_set_opaque_xerr(new, newdentry, -EXDEV);
1241         if (err)
1242                 goto out_dput;
1243
1244         err = ovl_do_rename(old_upperdir->d_inode, olddentry,
1245                             new_upperdir->d_inode, newdentry, flags);
1246         if (err)
1247                 goto out_dput;
1248
1249         if (cleanup_whiteout)
1250                 ovl_cleanup(old_upperdir->d_inode, newdentry);
1251
1252         if (overwrite && d_inode(new)) {
1253                 if (new_is_dir)
1254                         clear_nlink(d_inode(new));
1255                 else
1256                         ovl_drop_nlink(new);
1257         }
1258
1259         ovl_dir_modified(old->d_parent, ovl_type_origin(old) ||
1260                          (!overwrite && ovl_type_origin(new)));
1261         ovl_dir_modified(new->d_parent, ovl_type_origin(old) ||
1262                          (d_inode(new) && ovl_type_origin(new)));
1263
1264         /* copy ctime: */
1265         ovl_copyattr(d_inode(olddentry), d_inode(old));
1266         if (d_inode(new) && ovl_dentry_upper(new))
1267                 ovl_copyattr(d_inode(newdentry), d_inode(new));
1268
1269 out_dput:
1270         dput(newdentry);
1271 out_dput_old:
1272         dput(olddentry);
1273 out_unlock:
1274         unlock_rename(new_upperdir, old_upperdir);
1275 out_revert_creds:
1276         revert_creds(old_cred);
1277         if (update_nlink)
1278                 ovl_nlink_end(new);
1279 out_drop_write:
1280         ovl_drop_write(old);
1281 out:
1282         dput(opaquedir);
1283         ovl_cache_free(&list);
1284         return err;
1285 }
1286
1287 const struct inode_operations ovl_dir_inode_operations = {
1288         .lookup         = ovl_lookup,
1289         .mkdir          = ovl_mkdir,
1290         .symlink        = ovl_symlink,
1291         .unlink         = ovl_unlink,
1292         .rmdir          = ovl_rmdir,
1293         .rename         = ovl_rename,
1294         .link           = ovl_link,
1295         .setattr        = ovl_setattr,
1296         .create         = ovl_create,
1297         .mknod          = ovl_mknod,
1298         .permission     = ovl_permission,
1299         .getattr        = ovl_getattr,
1300         .listxattr      = ovl_listxattr,
1301         .get_acl        = ovl_get_acl,
1302         .update_time    = ovl_update_time,
1303 };