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         struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 237         int err;
 238
 239         err = ovl_check_setxattr(ofs, upper, OVL_XATTR_OPAQUE, "y", 1, xerr);
 240         if (!err)
 241                 ovl_dentry_set_opaque(dentry);
 242
 243         return err;
 244 }
 245
 246 static int ovl_set_opaque(struct dentry *dentry, struct dentry *upperdentry)
 247 {
 248         /*
 249          * Fail with -EIO when trying to create opaque dir and upper doesn't
 250          * support xattrs. ovl_rename() calls ovl_set_opaque_xerr(-EXDEV) to
 251          * return a specific error for noxattr case.
 252          */
 253         return ovl_set_opaque_xerr(dentry, upperdentry, -EIO);
 254 }
 255
 256 /*
 257  * Common operations required to be done after creation of file on upper.
 258  * If @hardlink is false, then @inode is a pre-allocated inode, we may or
 259  * may not use to instantiate the new dentry.
 260  */
 261 static int ovl_instantiate(struct dentry *dentry, struct inode *inode,
 262                            struct dentry *newdentry, bool hardlink)
 263 {
 264         struct ovl_inode_params oip = {
 265                 .upperdentry = newdentry,
 266                 .newinode = inode,
 267         };
 268
 269         ovl_dir_modified(dentry->d_parent, false);
 270         ovl_dentry_set_upper_alias(dentry);
 271         ovl_dentry_update_reval(dentry, newdentry,
 272                         DCACHE_OP_REVALIDATE | DCACHE_OP_WEAK_REVALIDATE);
 273
 274         if (!hardlink) {
 275                 /*
 276                  * ovl_obtain_alias() can be called after ovl_create_real()
 277                  * and before we get here, so we may get an inode from cache
 278                  * with the same real upperdentry that is not the inode we
 279                  * pre-allocated.  In this case we will use the cached inode
 280                  * to instantiate the new dentry.
 281                  *
 282                  * XXX: if we ever use ovl_obtain_alias() to decode directory
 283                  * file handles, need to use ovl_get_inode_locked() and
 284                  * d_instantiate_new() here to prevent from creating two
 285                  * hashed directory inode aliases.
 286                  */
 287                 inode = ovl_get_inode(dentry->d_sb, &oip);
 288                 if (IS_ERR(inode))
 289                         return PTR_ERR(inode);
 290                 if (inode == oip.newinode)
 291                         ovl_set_flag(OVL_UPPERDATA, inode);
 292         } else {
 293                 WARN_ON(ovl_inode_real(inode) != d_inode(newdentry));
 294                 dput(newdentry);
 295                 inc_nlink(inode);
 296         }
 297
 298         d_instantiate(dentry, inode);
 299         if (inode != oip.newinode) {
 300                 pr_warn_ratelimited("newly created inode found in cache (%pd2)\n",
 301                                     dentry);
 302         }
 303
 304         /* Force lookup of new upper hardlink to find its lower */
 305         if (hardlink)
 306                 d_drop(dentry);
 307
 308         return 0;
 309 }
 310
 311 static bool ovl_type_merge(struct dentry *dentry)
 312 {
 313         return OVL_TYPE_MERGE(ovl_path_type(dentry));
 314 }
 315
 316 static bool ovl_type_origin(struct dentry *dentry)
 317 {
 318         return OVL_TYPE_ORIGIN(ovl_path_type(dentry));
 319 }
 320
 321 static int ovl_create_upper(struct dentry *dentry, struct inode *inode,
 322                             struct ovl_cattr *attr)
 323 {
 324         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 325         struct inode *udir = upperdir->d_inode;
 326         struct dentry *newdentry;
 327         int err;
 328
 329         if (!attr->hardlink && !IS_POSIXACL(udir))
 330                 attr->mode &= ~current_umask();
 331
 332         inode_lock_nested(udir, I_MUTEX_PARENT);
 333         newdentry = ovl_create_real(udir,
 334                                     lookup_one_len(dentry->d_name.name,
 335                                                    upperdir,
 336                                                    dentry->d_name.len),
 337                                     attr);
 338         err = PTR_ERR(newdentry);
 339         if (IS_ERR(newdentry))
 340                 goto out_unlock;
 341
 342         if (ovl_type_merge(dentry->d_parent) && d_is_dir(newdentry)) {
 343                 /* Setting opaque here is just an optimization, allow to fail */
 344                 ovl_set_opaque(dentry, newdentry);
 345         }
 346
 347         err = ovl_instantiate(dentry, inode, newdentry, !!attr->hardlink);
 348         if (err)
 349                 goto out_cleanup;
 350 out_unlock:
 351         inode_unlock(udir);
 352         return err;
 353
 354 out_cleanup:
 355         ovl_cleanup(udir, newdentry);
 356         dput(newdentry);
 357         goto out_unlock;
 358 }
 359
 360 static struct dentry *ovl_clear_empty(struct dentry *dentry,
 361                                       struct list_head *list)
 362 {
 363         struct dentry *workdir = ovl_workdir(dentry);
 364         struct inode *wdir = workdir->d_inode;
 365         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 366         struct inode *udir = upperdir->d_inode;
 367         struct path upperpath;
 368         struct dentry *upper;
 369         struct dentry *opaquedir;
 370         struct kstat stat;
 371         int err;
 372
 373         if (WARN_ON(!workdir))
 374                 return ERR_PTR(-EROFS);
 375
 376         err = ovl_lock_rename_workdir(workdir, upperdir);
 377         if (err)
 378                 goto out;
 379
 380         ovl_path_upper(dentry, &upperpath);
 381         err = vfs_getattr(&upperpath, &stat,
 382                           STATX_BASIC_STATS, AT_STATX_SYNC_AS_STAT);
 383         if (err)
 384                 goto out_unlock;
 385
 386         err = -ESTALE;
 387         if (!S_ISDIR(stat.mode))
 388                 goto out_unlock;
 389         upper = upperpath.dentry;
 390         if (upper->d_parent->d_inode != udir)
 391                 goto out_unlock;
 392
 393         opaquedir = ovl_create_temp(workdir, OVL_CATTR(stat.mode));
 394         err = PTR_ERR(opaquedir);
 395         if (IS_ERR(opaquedir))
 396                 goto out_unlock;
 397
 398         err = ovl_copy_xattr(dentry->d_sb, upper, opaquedir);
 399         if (err)
 400                 goto out_cleanup;
 401
 402         err = ovl_set_opaque(dentry, opaquedir);
 403         if (err)
 404                 goto out_cleanup;
 405
 406         inode_lock(opaquedir->d_inode);
 407         err = ovl_set_attr(opaquedir, &stat);
 408         inode_unlock(opaquedir->d_inode);
 409         if (err)
 410                 goto out_cleanup;
 411
 412         err = ovl_do_rename(wdir, opaquedir, udir, upper, RENAME_EXCHANGE);
 413         if (err)
 414                 goto out_cleanup;
 415
 416         ovl_cleanup_whiteouts(upper, list);
 417         ovl_cleanup(wdir, upper);
 418         unlock_rename(workdir, upperdir);
 419
 420         /* dentry's upper doesn't match now, get rid of it */
 421         d_drop(dentry);
 422
 423         return opaquedir;
 424
 425 out_cleanup:
 426         ovl_cleanup(wdir, opaquedir);
 427         dput(opaquedir);
 428 out_unlock:
 429         unlock_rename(workdir, upperdir);
 430 out:
 431         return ERR_PTR(err);
 432 }
 433
 434 static int ovl_set_upper_acl(struct dentry *upperdentry, const char *name,
 435                              const struct posix_acl *acl)
 436 {
 437         void *buffer;
 438         size_t size;
 439         int err;
 440
 441         if (!IS_ENABLED(CONFIG_FS_POSIX_ACL) || !acl)
 442                 return 0;
 443
 444         size = posix_acl_xattr_size(acl->a_count);
 445         buffer = kmalloc(size, GFP_KERNEL);
 446         if (!buffer)
 447                 return -ENOMEM;
 448
 449         err = posix_acl_to_xattr(&init_user_ns, acl, buffer, size);
 450         if (err < 0)
 451                 goto out_free;
 452
 453         err = vfs_setxattr(&init_user_ns, upperdentry, name, buffer, size, XATTR_CREATE);
 454 out_free:
 455         kfree(buffer);
 456         return err;
 457 }
 458
 459 static int ovl_create_over_whiteout(struct dentry *dentry, struct inode *inode,
 460                                     struct ovl_cattr *cattr)
 461 {
 462         struct dentry *workdir = ovl_workdir(dentry);
 463         struct inode *wdir = workdir->d_inode;
 464         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 465         struct inode *udir = upperdir->d_inode;
 466         struct dentry *upper;
 467         struct dentry *newdentry;
 468         int err;
 469         struct posix_acl *acl, *default_acl;
 470         bool hardlink = !!cattr->hardlink;
 471
 472         if (WARN_ON(!workdir))
 473                 return -EROFS;
 474
 475         if (!hardlink) {
 476                 err = posix_acl_create(dentry->d_parent->d_inode,
 477                                        &cattr->mode, &default_acl, &acl);
 478                 if (err)
 479                         return err;
 480         }
 481
 482         err = ovl_lock_rename_workdir(workdir, upperdir);
 483         if (err)
 484                 goto out;
 485
 486         upper = lookup_one_len(dentry->d_name.name, upperdir,
 487                                dentry->d_name.len);
 488         err = PTR_ERR(upper);
 489         if (IS_ERR(upper))
 490                 goto out_unlock;
 491
 492         err = -ESTALE;
 493         if (d_is_negative(upper) || !IS_WHITEOUT(d_inode(upper)))
 494                 goto out_dput;
 495
 496         newdentry = ovl_create_temp(workdir, cattr);
 497         err = PTR_ERR(newdentry);
 498         if (IS_ERR(newdentry))
 499                 goto out_dput;
 500
 501         /*
 502          * mode could have been mutilated due to umask (e.g. sgid directory)
 503          */
 504         if (!hardlink &&
 505             !S_ISLNK(cattr->mode) &&
 506             newdentry->d_inode->i_mode != cattr->mode) {
 507                 struct iattr attr = {
 508                         .ia_valid = ATTR_MODE,
 509                         .ia_mode = cattr->mode,
 510                 };
 511                 inode_lock(newdentry->d_inode);
 512                 err = notify_change(&init_user_ns, newdentry, &attr, NULL);
 513                 inode_unlock(newdentry->d_inode);
 514                 if (err)
 515                         goto out_cleanup;
 516         }
 517         if (!hardlink) {
 518                 err = ovl_set_upper_acl(newdentry, XATTR_NAME_POSIX_ACL_ACCESS,
 519                                         acl);
 520                 if (err)
 521                         goto out_cleanup;
 522
 523                 err = ovl_set_upper_acl(newdentry, XATTR_NAME_POSIX_ACL_DEFAULT,
 524                                         default_acl);
 525                 if (err)
 526                         goto out_cleanup;
 527         }
 528
 529         if (!hardlink && S_ISDIR(cattr->mode)) {
 530                 err = ovl_set_opaque(dentry, newdentry);
 531                 if (err)
 532                         goto out_cleanup;
 533
 534                 err = ovl_do_rename(wdir, newdentry, udir, upper,
 535                                     RENAME_EXCHANGE);
 536                 if (err)
 537                         goto out_cleanup;
 538
 539                 ovl_cleanup(wdir, upper);
 540         } else {
 541                 err = ovl_do_rename(wdir, newdentry, udir, upper, 0);
 542                 if (err)
 543                         goto out_cleanup;
 544         }
 545         err = ovl_instantiate(dentry, inode, newdentry, hardlink);
 546         if (err)
 547                 goto out_cleanup;
 548 out_dput:
 549         dput(upper);
 550 out_unlock:
 551         unlock_rename(workdir, upperdir);
 552 out:
 553         if (!hardlink) {
 554                 posix_acl_release(acl);
 555                 posix_acl_release(default_acl);
 556         }
 557         return err;
 558
 559 out_cleanup:
 560         ovl_cleanup(wdir, newdentry);
 561         dput(newdentry);
 562         goto out_dput;
 563 }
 564
 565 static int ovl_create_or_link(struct dentry *dentry, struct inode *inode,
 566                               struct ovl_cattr *attr, bool origin)
 567 {
 568         int err;
 569         const struct cred *old_cred;
 570         struct cred *override_cred;
 571         struct dentry *parent = dentry->d_parent;
 572
 573         err = ovl_copy_up(parent);
 574         if (err)
 575                 return err;
 576
 577         old_cred = ovl_override_creds(dentry->d_sb);
 578
 579         /*
 580          * When linking a file with copy up origin into a new parent, mark the
 581          * new parent dir "impure".
 582          */
 583         if (origin) {
 584                 err = ovl_set_impure(parent, ovl_dentry_upper(parent));
 585                 if (err)
 586                         goto out_revert_creds;
 587         }
 588
 589         err = -ENOMEM;
 590         override_cred = prepare_creds();
 591         if (override_cred) {
 592                 override_cred->fsuid = inode->i_uid;
 593                 override_cred->fsgid = inode->i_gid;
 594                 if (!attr->hardlink) {
 595                         err = security_dentry_create_files_as(dentry,
 596                                         attr->mode, &dentry->d_name, old_cred,
 597                                         override_cred);
 598                         if (err) {
 599                                 put_cred(override_cred);
 600                                 goto out_revert_creds;
 601                         }
 602                 }
 603                 put_cred(override_creds(override_cred));
 604                 put_cred(override_cred);
 605
 606                 if (!ovl_dentry_is_whiteout(dentry))
 607                         err = ovl_create_upper(dentry, inode, attr);
 608                 else
 609                         err = ovl_create_over_whiteout(dentry, inode, attr);
 610         }
 611 out_revert_creds:
 612         revert_creds(old_cred);
 613         return err;
 614 }
 615
 616 static int ovl_create_object(struct dentry *dentry, int mode, dev_t rdev,
 617                              const char *link)
 618 {
 619         int err;
 620         struct inode *inode;
 621         struct ovl_cattr attr = {
 622                 .rdev = rdev,
 623                 .link = link,
 624         };
 625
 626         err = ovl_want_write(dentry);
 627         if (err)
 628                 goto out;
 629
 630         /* Preallocate inode to be used by ovl_get_inode() */
 631         err = -ENOMEM;
 632         inode = ovl_new_inode(dentry->d_sb, mode, rdev);
 633         if (!inode)
 634                 goto out_drop_write;
 635
 636         spin_lock(&inode->i_lock);
 637         inode->i_state |= I_CREATING;
 638         spin_unlock(&inode->i_lock);
 639
 640         inode_init_owner(&init_user_ns, inode, dentry->d_parent->d_inode, mode);
 641         attr.mode = inode->i_mode;
 642
 643         err = ovl_create_or_link(dentry, inode, &attr, false);
 644         /* Did we end up using the preallocated inode? */
 645         if (inode != d_inode(dentry))
 646                 iput(inode);
 647
 648 out_drop_write:
 649         ovl_drop_write(dentry);
 650 out:
 651         return err;
 652 }
 653
 654 static int ovl_create(struct user_namespace *mnt_userns, struct inode *dir,
 655                       struct dentry *dentry, umode_t mode, bool excl)
 656 {
 657         return ovl_create_object(dentry, (mode & 07777) | S_IFREG, 0, NULL);
 658 }
 659
 660 static int ovl_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
 661                      struct dentry *dentry, umode_t mode)
 662 {
 663         return ovl_create_object(dentry, (mode & 07777) | S_IFDIR, 0, NULL);
 664 }
 665
 666 static int ovl_mknod(struct user_namespace *mnt_userns, struct inode *dir,
 667                      struct dentry *dentry, umode_t mode, dev_t rdev)
 668 {
 669         /* Don't allow creation of "whiteout" on overlay */
 670         if (S_ISCHR(mode) && rdev == WHITEOUT_DEV)
 671                 return -EPERM;
 672
 673         return ovl_create_object(dentry, mode, rdev, NULL);
 674 }
 675
 676 static int ovl_symlink(struct user_namespace *mnt_userns, struct inode *dir,
 677                        struct dentry *dentry, const char *link)
 678 {
 679         return ovl_create_object(dentry, S_IFLNK, 0, link);
 680 }
 681
 682 static int ovl_set_link_redirect(struct dentry *dentry)
 683 {
 684         const struct cred *old_cred;
 685         int err;
 686
 687         old_cred = ovl_override_creds(dentry->d_sb);
 688         err = ovl_set_redirect(dentry, false);
 689         revert_creds(old_cred);
 690
 691         return err;
 692 }
 693
 694 static int ovl_link(struct dentry *old, struct inode *newdir,
 695                     struct dentry *new)
 696 {
 697         int err;
 698         struct inode *inode;
 699
 700         err = ovl_want_write(old);
 701         if (err)
 702                 goto out;
 703
 704         err = ovl_copy_up(old);
 705         if (err)
 706                 goto out_drop_write;
 707
 708         err = ovl_copy_up(new->d_parent);
 709         if (err)
 710                 goto out_drop_write;
 711
 712         if (ovl_is_metacopy_dentry(old)) {
 713                 err = ovl_set_link_redirect(old);
 714                 if (err)
 715                         goto out_drop_write;
 716         }
 717
 718         err = ovl_nlink_start(old);
 719         if (err)
 720                 goto out_drop_write;
 721
 722         inode = d_inode(old);
 723         ihold(inode);
 724
 725         err = ovl_create_or_link(new, inode,
 726                         &(struct ovl_cattr) {.hardlink = ovl_dentry_upper(old)},
 727                         ovl_type_origin(old));
 728         if (err)
 729                 iput(inode);
 730
 731         ovl_nlink_end(old);
 732 out_drop_write:
 733         ovl_drop_write(old);
 734 out:
 735         return err;
 736 }
 737
 738 static bool ovl_matches_upper(struct dentry *dentry, struct dentry *upper)
 739 {
 740         return d_inode(ovl_dentry_upper(dentry)) == d_inode(upper);
 741 }
 742
 743 static int ovl_remove_and_whiteout(struct dentry *dentry,
 744                                    struct list_head *list)
 745 {
 746         struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 747         struct dentry *workdir = ovl_workdir(dentry);
 748         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 749         struct dentry *upper;
 750         struct dentry *opaquedir = NULL;
 751         int err;
 752
 753         if (WARN_ON(!workdir))
 754                 return -EROFS;
 755
 756         if (!list_empty(list)) {
 757                 opaquedir = ovl_clear_empty(dentry, list);
 758                 err = PTR_ERR(opaquedir);
 759                 if (IS_ERR(opaquedir))
 760                         goto out;
 761         }
 762
 763         err = ovl_lock_rename_workdir(workdir, upperdir);
 764         if (err)
 765                 goto out_dput;
 766
 767         upper = lookup_one_len(dentry->d_name.name, upperdir,
 768                                dentry->d_name.len);
 769         err = PTR_ERR(upper);
 770         if (IS_ERR(upper))
 771                 goto out_unlock;
 772
 773         err = -ESTALE;
 774         if ((opaquedir && upper != opaquedir) ||
 775             (!opaquedir && ovl_dentry_upper(dentry) &&
 776              !ovl_matches_upper(dentry, upper))) {
 777                 goto out_dput_upper;
 778         }
 779
 780         err = ovl_cleanup_and_whiteout(ofs, d_inode(upperdir), upper);
 781         if (err)
 782                 goto out_d_drop;
 783
 784         ovl_dir_modified(dentry->d_parent, true);
 785 out_d_drop:
 786         d_drop(dentry);
 787 out_dput_upper:
 788         dput(upper);
 789 out_unlock:
 790         unlock_rename(workdir, upperdir);
 791 out_dput:
 792         dput(opaquedir);
 793 out:
 794         return err;
 795 }
 796
 797 static int ovl_remove_upper(struct dentry *dentry, bool is_dir,
 798                             struct list_head *list)
 799 {
 800         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 801         struct inode *dir = upperdir->d_inode;
 802         struct dentry *upper;
 803         struct dentry *opaquedir = NULL;
 804         int err;
 805
 806         if (!list_empty(list)) {
 807                 opaquedir = ovl_clear_empty(dentry, list);
 808                 err = PTR_ERR(opaquedir);
 809                 if (IS_ERR(opaquedir))
 810                         goto out;
 811         }
 812
 813         inode_lock_nested(dir, I_MUTEX_PARENT);
 814         upper = lookup_one_len(dentry->d_name.name, upperdir,
 815                                dentry->d_name.len);
 816         err = PTR_ERR(upper);
 817         if (IS_ERR(upper))
 818                 goto out_unlock;
 819
 820         err = -ESTALE;
 821         if ((opaquedir && upper != opaquedir) ||
 822             (!opaquedir && !ovl_matches_upper(dentry, upper)))
 823                 goto out_dput_upper;
 824
 825         if (is_dir)
 826                 err = vfs_rmdir(&init_user_ns, dir, upper);
 827         else
 828                 err = vfs_unlink(&init_user_ns, dir, upper, NULL);
 829         ovl_dir_modified(dentry->d_parent, ovl_type_origin(dentry));
 830
 831         /*
 832          * Keeping this dentry hashed would mean having to release
 833          * upperpath/lowerpath, which could only be done if we are the
 834          * sole user of this dentry.  Too tricky...  Just unhash for
 835          * now.
 836          */
 837         if (!err)
 838                 d_drop(dentry);
 839 out_dput_upper:
 840         dput(upper);
 841 out_unlock:
 842         inode_unlock(dir);
 843         dput(opaquedir);
 844 out:
 845         return err;
 846 }
 847
 848 static bool ovl_pure_upper(struct dentry *dentry)
 849 {
 850         return !ovl_dentry_lower(dentry) &&
 851                !ovl_test_flag(OVL_WHITEOUTS, d_inode(dentry));
 852 }
 853
 854 static void ovl_drop_nlink(struct dentry *dentry)
 855 {
 856         struct inode *inode = d_inode(dentry);
 857         struct dentry *alias;
 858
 859         /* Try to find another, hashed alias */
 860         spin_lock(&inode->i_lock);
 861         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
 862                 if (alias != dentry && !d_unhashed(alias))
 863                         break;
 864         }
 865         spin_unlock(&inode->i_lock);
 866
 867         /*
 868          * Changes to underlying layers may cause i_nlink to lose sync with
 869          * reality.  In this case prevent the link count from going to zero
 870          * prematurely.
 871          */
 872         if (inode->i_nlink > !!alias)
 873                 drop_nlink(inode);
 874 }
 875
 876 static int ovl_do_remove(struct dentry *dentry, bool is_dir)
 877 {
 878         int err;
 879         const struct cred *old_cred;
 880         struct dentry *upperdentry;
 881         bool lower_positive = ovl_lower_positive(dentry);
 882         LIST_HEAD(list);
 883
 884         /* No need to clean pure upper removed by vfs_rmdir() */
 885         if (is_dir && (lower_positive || !ovl_pure_upper(dentry))) {
 886                 err = ovl_check_empty_dir(dentry, &list);
 887                 if (err)
 888                         goto out;
 889         }
 890
 891         err = ovl_want_write(dentry);
 892         if (err)
 893                 goto out;
 894
 895         err = ovl_copy_up(dentry->d_parent);
 896         if (err)
 897                 goto out_drop_write;
 898
 899         err = ovl_nlink_start(dentry);
 900         if (err)
 901                 goto out_drop_write;
 902
 903         old_cred = ovl_override_creds(dentry->d_sb);
 904         if (!lower_positive)
 905                 err = ovl_remove_upper(dentry, is_dir, &list);
 906         else
 907                 err = ovl_remove_and_whiteout(dentry, &list);
 908         revert_creds(old_cred);
 909         if (!err) {
 910                 if (is_dir)
 911                         clear_nlink(dentry->d_inode);
 912                 else
 913                         ovl_drop_nlink(dentry);
 914         }
 915         ovl_nlink_end(dentry);
 916
 917         /*
 918          * Copy ctime
 919          *
 920          * Note: we fail to update ctime if there was no copy-up, only a
 921          * whiteout
 922          */
 923         upperdentry = ovl_dentry_upper(dentry);
 924         if (upperdentry)
 925                 ovl_copyattr(d_inode(upperdentry), d_inode(dentry));
 926
 927 out_drop_write:
 928         ovl_drop_write(dentry);
 929 out:
 930         ovl_cache_free(&list);
 931         return err;
 932 }
 933
 934 static int ovl_unlink(struct inode *dir, struct dentry *dentry)
 935 {
 936         return ovl_do_remove(dentry, false);
 937 }
 938
 939 static int ovl_rmdir(struct inode *dir, struct dentry *dentry)
 940 {
 941         return ovl_do_remove(dentry, true);
 942 }
 943
 944 static bool ovl_type_merge_or_lower(struct dentry *dentry)
 945 {
 946         enum ovl_path_type type = ovl_path_type(dentry);
 947
 948         return OVL_TYPE_MERGE(type) || !OVL_TYPE_UPPER(type);
 949 }
 950
 951 static bool ovl_can_move(struct dentry *dentry)
 952 {
 953         return ovl_redirect_dir(dentry->d_sb) ||
 954                 !d_is_dir(dentry) || !ovl_type_merge_or_lower(dentry);
 955 }
 956
 957 static char *ovl_get_redirect(struct dentry *dentry, bool abs_redirect)
 958 {
 959         char *buf, *ret;
 960         struct dentry *d, *tmp;
 961         int buflen = ovl_redirect_max + 1;
 962
 963         if (!abs_redirect) {
 964                 ret = kstrndup(dentry->d_name.name, dentry->d_name.len,
 965                                GFP_KERNEL);
 966                 goto out;
 967         }
 968
 969         buf = ret = kmalloc(buflen, GFP_KERNEL);
 970         if (!buf)
 971                 goto out;
 972
 973         buflen--;
 974         buf[buflen] = '\0';
 975         for (d = dget(dentry); !IS_ROOT(d);) {
 976                 const char *name;
 977                 int thislen;
 978
 979                 spin_lock(&d->d_lock);
 980                 name = ovl_dentry_get_redirect(d);
 981                 if (name) {
 982                         thislen = strlen(name);
 983                 } else {
 984                         name = d->d_name.name;
 985                         thislen = d->d_name.len;
 986                 }
 987
 988                 /* If path is too long, fall back to userspace move */
 989                 if (thislen + (name[0] != '/') > buflen) {
 990                         ret = ERR_PTR(-EXDEV);
 991                         spin_unlock(&d->d_lock);
 992                         goto out_put;
 993                 }
 994
 995                 buflen -= thislen;
 996                 memcpy(&buf[buflen], name, thislen);
 997                 spin_unlock(&d->d_lock);
 998                 tmp = dget_parent(d);
 999
1000                 dput(d);
1001                 d = tmp;
1002
1003                 /* Absolute redirect: finished */
1004                 if (buf[buflen] == '/')
1005                         break;
1006                 buflen--;
1007                 buf[buflen] = '/';
1008         }
1009         ret = kstrdup(&buf[buflen], GFP_KERNEL);
1010 out_put:
1011         dput(d);
1012         kfree(buf);
1013 out:
1014         return ret ? ret : ERR_PTR(-ENOMEM);
1015 }
1016
1017 static bool ovl_need_absolute_redirect(struct dentry *dentry, bool samedir)
1018 {
1019         struct dentry *lowerdentry;
1020
1021         if (!samedir)
1022                 return true;
1023
1024         if (d_is_dir(dentry))
1025                 return false;
1026
1027         /*
1028          * For non-dir hardlinked files, we need absolute redirects
1029          * in general as two upper hardlinks could be in different
1030          * dirs. We could put a relative redirect now and convert
1031          * it to absolute redirect later. But when nlink > 1 and
1032          * indexing is on, that means relative redirect needs to be
1033          * converted to absolute during copy up of another lower
1034          * hardllink as well.
1035          *
1036          * So without optimizing too much, just check if lower is
1037          * a hard link or not. If lower is hard link, put absolute
1038          * redirect.
1039          */
1040         lowerdentry = ovl_dentry_lower(dentry);
1041         return (d_inode(lowerdentry)->i_nlink > 1);
1042 }
1043
1044 static int ovl_set_redirect(struct dentry *dentry, bool samedir)
1045 {
1046         int err;
1047         struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
1048         const char *redirect = ovl_dentry_get_redirect(dentry);
1049         bool absolute_redirect = ovl_need_absolute_redirect(dentry, samedir);
1050
1051         if (redirect && (!absolute_redirect || redirect[0] == '/'))
1052                 return 0;
1053
1054         redirect = ovl_get_redirect(dentry, absolute_redirect);
1055         if (IS_ERR(redirect))
1056                 return PTR_ERR(redirect);
1057
1058         err = ovl_check_setxattr(ofs, ovl_dentry_upper(dentry),
1059                                  OVL_XATTR_REDIRECT,
1060                                  redirect, strlen(redirect), -EXDEV);
1061         if (!err) {
1062                 spin_lock(&dentry->d_lock);
1063                 ovl_dentry_set_redirect(dentry, redirect);
1064                 spin_unlock(&dentry->d_lock);
1065         } else {
1066                 kfree(redirect);
1067                 pr_warn_ratelimited("failed to set redirect (%i)\n",
1068                                     err);
1069                 /* Fall back to userspace copy-up */
1070                 err = -EXDEV;
1071         }
1072         return err;
1073 }
1074
1075 static int ovl_rename(struct user_namespace *mnt_userns, struct inode *olddir,
1076                       struct dentry *old, struct inode *newdir,
1077                       struct dentry *new, unsigned int flags)
1078 {
1079         int err;
1080         struct dentry *old_upperdir;
1081         struct dentry *new_upperdir;
1082         struct dentry *olddentry;
1083         struct dentry *newdentry;
1084         struct dentry *trap;
1085         bool old_opaque;
1086         bool new_opaque;
1087         bool cleanup_whiteout = false;
1088         bool update_nlink = false;
1089         bool overwrite = !(flags & RENAME_EXCHANGE);
1090         bool is_dir = d_is_dir(old);
1091         bool new_is_dir = d_is_dir(new);
1092         bool samedir = olddir == newdir;
1093         struct dentry *opaquedir = NULL;
1094         const struct cred *old_cred = NULL;
1095         LIST_HEAD(list);
1096
1097         err = -EINVAL;
1098         if (flags & ~(RENAME_EXCHANGE | RENAME_NOREPLACE))
1099                 goto out;
1100
1101         flags &= ~RENAME_NOREPLACE;
1102
1103         /* Don't copy up directory trees */
1104         err = -EXDEV;
1105         if (!ovl_can_move(old))
1106                 goto out;
1107         if (!overwrite && !ovl_can_move(new))
1108                 goto out;
1109
1110         if (overwrite && new_is_dir && !ovl_pure_upper(new)) {
1111                 err = ovl_check_empty_dir(new, &list);
1112                 if (err)
1113                         goto out;
1114         }
1115
1116         if (overwrite) {
1117                 if (ovl_lower_positive(old)) {
1118                         if (!ovl_dentry_is_whiteout(new)) {
1119                                 /* Whiteout source */
1120                                 flags |= RENAME_WHITEOUT;
1121                         } else {
1122                                 /* Switch whiteouts */
1123                                 flags |= RENAME_EXCHANGE;
1124                         }
1125                 } else if (is_dir && ovl_dentry_is_whiteout(new)) {
1126                         flags |= RENAME_EXCHANGE;
1127                         cleanup_whiteout = true;
1128                 }
1129         }
1130
1131         err = ovl_want_write(old);
1132         if (err)
1133                 goto out;
1134
1135         err = ovl_copy_up(old);
1136         if (err)
1137                 goto out_drop_write;
1138
1139         err = ovl_copy_up(new->d_parent);
1140         if (err)
1141                 goto out_drop_write;
1142         if (!overwrite) {
1143                 err = ovl_copy_up(new);
1144                 if (err)
1145                         goto out_drop_write;
1146         } else if (d_inode(new)) {
1147                 err = ovl_nlink_start(new);
1148                 if (err)
1149                         goto out_drop_write;
1150
1151                 update_nlink = true;
1152         }
1153
1154         old_cred = ovl_override_creds(old->d_sb);
1155
1156         if (!list_empty(&list)) {
1157                 opaquedir = ovl_clear_empty(new, &list);
1158                 err = PTR_ERR(opaquedir);
1159                 if (IS_ERR(opaquedir)) {
1160                         opaquedir = NULL;
1161                         goto out_revert_creds;
1162                 }
1163         }
1164
1165         old_upperdir = ovl_dentry_upper(old->d_parent);
1166         new_upperdir = ovl_dentry_upper(new->d_parent);
1167
1168         if (!samedir) {
1169                 /*
1170                  * When moving a merge dir or non-dir with copy up origin into
1171                  * a new parent, we are marking the new parent dir "impure".
1172                  * When ovl_iterate() iterates an "impure" upper dir, it will
1173                  * lookup the origin inodes of the entries to fill d_ino.
1174                  */
1175                 if (ovl_type_origin(old)) {
1176                         err = ovl_set_impure(new->d_parent, new_upperdir);
1177                         if (err)
1178                                 goto out_revert_creds;
1179                 }
1180                 if (!overwrite && ovl_type_origin(new)) {
1181                         err = ovl_set_impure(old->d_parent, old_upperdir);
1182                         if (err)
1183                                 goto out_revert_creds;
1184                 }
1185         }
1186
1187         trap = lock_rename(new_upperdir, old_upperdir);
1188
1189         olddentry = lookup_one_len(old->d_name.name, old_upperdir,
1190                                    old->d_name.len);
1191         err = PTR_ERR(olddentry);
1192         if (IS_ERR(olddentry))
1193                 goto out_unlock;
1194
1195         err = -ESTALE;
1196         if (!ovl_matches_upper(old, olddentry))
1197                 goto out_dput_old;
1198
1199         newdentry = lookup_one_len(new->d_name.name, new_upperdir,
1200                                    new->d_name.len);
1201         err = PTR_ERR(newdentry);
1202         if (IS_ERR(newdentry))
1203                 goto out_dput_old;
1204
1205         old_opaque = ovl_dentry_is_opaque(old);
1206         new_opaque = ovl_dentry_is_opaque(new);
1207
1208         err = -ESTALE;
1209         if (d_inode(new) && ovl_dentry_upper(new)) {
1210                 if (opaquedir) {
1211                         if (newdentry != opaquedir)
1212                                 goto out_dput;
1213                 } else {
1214                         if (!ovl_matches_upper(new, newdentry))
1215                                 goto out_dput;
1216                 }
1217         } else {
1218                 if (!d_is_negative(newdentry) &&
1219                     (!new_opaque || !ovl_is_whiteout(newdentry)))
1220                         goto out_dput;
1221         }
1222
1223         if (olddentry == trap)
1224                 goto out_dput;
1225         if (newdentry == trap)
1226                 goto out_dput;
1227
1228         if (olddentry->d_inode == newdentry->d_inode)
1229                 goto out_dput;
1230
1231         err = 0;
1232         if (ovl_type_merge_or_lower(old))
1233                 err = ovl_set_redirect(old, samedir);
1234         else if (is_dir && !old_opaque && ovl_type_merge(new->d_parent))
1235                 err = ovl_set_opaque_xerr(old, olddentry, -EXDEV);
1236         if (err)
1237                 goto out_dput;
1238
1239         if (!overwrite && ovl_type_merge_or_lower(new))
1240                 err = ovl_set_redirect(new, samedir);
1241         else if (!overwrite && new_is_dir && !new_opaque &&
1242                  ovl_type_merge(old->d_parent))
1243                 err = ovl_set_opaque_xerr(new, newdentry, -EXDEV);
1244         if (err)
1245                 goto out_dput;
1246
1247         err = ovl_do_rename(old_upperdir->d_inode, olddentry,
1248                             new_upperdir->d_inode, newdentry, flags);
1249         if (err)
1250                 goto out_dput;
1251
1252         if (cleanup_whiteout)
1253                 ovl_cleanup(old_upperdir->d_inode, newdentry);
1254
1255         if (overwrite && d_inode(new)) {
1256                 if (new_is_dir)
1257                         clear_nlink(d_inode(new));
1258                 else
1259                         ovl_drop_nlink(new);
1260         }
1261
1262         ovl_dir_modified(old->d_parent, ovl_type_origin(old) ||
1263                          (!overwrite && ovl_type_origin(new)));
1264         ovl_dir_modified(new->d_parent, ovl_type_origin(old) ||
1265                          (d_inode(new) && ovl_type_origin(new)));
1266
1267         /* copy ctime: */
1268         ovl_copyattr(d_inode(olddentry), d_inode(old));
1269         if (d_inode(new) && ovl_dentry_upper(new))
1270                 ovl_copyattr(d_inode(newdentry), d_inode(new));
1271
1272 out_dput:
1273         dput(newdentry);
1274 out_dput_old:
1275         dput(olddentry);
1276 out_unlock:
1277         unlock_rename(new_upperdir, old_upperdir);
1278 out_revert_creds:
1279         revert_creds(old_cred);
1280         if (update_nlink)
1281                 ovl_nlink_end(new);
1282 out_drop_write:
1283         ovl_drop_write(old);
1284 out:
1285         dput(opaquedir);
1286         ovl_cache_free(&list);
1287         return err;
1288 }
1289
1290 const struct inode_operations ovl_dir_inode_operations = {
1291         .lookup         = ovl_lookup,
1292         .mkdir          = ovl_mkdir,
1293         .symlink        = ovl_symlink,
1294         .unlink         = ovl_unlink,
1295         .rmdir          = ovl_rmdir,
1296         .rename         = ovl_rename,
1297         .link           = ovl_link,
1298         .setattr        = ovl_setattr,
1299         .create         = ovl_create,
1300         .mknod          = ovl_mknod,
1301         .permission     = ovl_permission,
1302         .getattr        = ovl_getattr,
1303         .listxattr      = ovl_listxattr,
1304         .get_acl        = ovl_get_acl,
1305         .update_time    = ovl_update_time,
1306         .fileattr_get   = ovl_fileattr_get,
1307         .fileattr_set   = ovl_fileattr_set,
1308 };