fs/overlayfs/dir.c

   1 /*
   2  *
   3  * Copyright (C) 2011 Novell Inc.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 as published by
   7  * the Free Software Foundation.
   8  */
   9
  10 #include <linux/fs.h>
  11 #include <linux/namei.h>
  12 #include <linux/xattr.h>
  13 #include <linux/security.h>
  14 #include <linux/cred.h>
  15 #include <linux/module.h>
  16 #include <linux/posix_acl.h>
  17 #include <linux/posix_acl_xattr.h>
  18 #include <linux/atomic.h>
  19 #include <linux/ratelimit.h>
  20 #include "overlayfs.h"
  21
  22 static unsigned short ovl_redirect_max = 256;
  23 module_param_named(redirect_max, ovl_redirect_max, ushort, 0644);
  24 MODULE_PARM_DESC(ovl_redirect_max,
  25                  "Maximum length of absolute redirect xattr value");
  26
  27 int ovl_cleanup(struct inode *wdir, struct dentry *wdentry)
  28 {
  29         int err;
  30
  31         dget(wdentry);
  32         if (d_is_dir(wdentry))
  33                 err = ovl_do_rmdir(wdir, wdentry);
  34         else
  35                 err = ovl_do_unlink(wdir, wdentry);
  36         dput(wdentry);
  37
  38         if (err) {
  39                 pr_err("overlayfs: cleanup of '%pd2' failed (%i)\n",
  40                        wdentry, err);
  41         }
  42
  43         return err;
  44 }
  45
  46 struct dentry *ovl_lookup_temp(struct dentry *workdir)
  47 {
  48         struct dentry *temp;
  49         char name[20];
  50         static atomic_t temp_id = ATOMIC_INIT(0);
  51
  52         /* counter is allowed to wrap, since temp dentries are ephemeral */
  53         snprintf(name, sizeof(name), "#%x", atomic_inc_return(&temp_id));
  54
  55         temp = lookup_one_len(name, workdir, strlen(name));
  56         if (!IS_ERR(temp) && temp->d_inode) {
  57                 pr_err("overlayfs: workdir/%s already exists\n", name);
  58                 dput(temp);
  59                 temp = ERR_PTR(-EIO);
  60         }
  61
  62         return temp;
  63 }
  64
  65 /* caller holds i_mutex on workdir */
  66 static struct dentry *ovl_whiteout(struct dentry *workdir,
  67                                    struct dentry *dentry)
  68 {
  69         int err;
  70         struct dentry *whiteout;
  71         struct inode *wdir = workdir->d_inode;
  72
  73         whiteout = ovl_lookup_temp(workdir);
  74         if (IS_ERR(whiteout))
  75                 return whiteout;
  76
  77         err = ovl_do_whiteout(wdir, whiteout);
  78         if (err) {
  79                 dput(whiteout);
  80                 whiteout = ERR_PTR(err);
  81         }
  82
  83         return whiteout;
  84 }
  85
  86 int ovl_create_real(struct inode *dir, struct dentry *newdentry,
  87                     struct cattr *attr, struct dentry *hardlink, bool debug)
  88 {
  89         int err;
  90
  91         if (newdentry->d_inode)
  92                 return -ESTALE;
  93
  94         if (hardlink) {
  95                 err = ovl_do_link(hardlink, dir, newdentry, debug);
  96         } else {
  97                 switch (attr->mode & S_IFMT) {
  98                 case S_IFREG:
  99                         err = ovl_do_create(dir, newdentry, attr->mode, debug);
 100                         break;
 101
 102                 case S_IFDIR:
 103                         err = ovl_do_mkdir(dir, newdentry, attr->mode, debug);
 104                         break;
 105
 106                 case S_IFCHR:
 107                 case S_IFBLK:
 108                 case S_IFIFO:
 109                 case S_IFSOCK:
 110                         err = ovl_do_mknod(dir, newdentry,
 111                                            attr->mode, attr->rdev, debug);
 112                         break;
 113
 114                 case S_IFLNK:
 115                         err = ovl_do_symlink(dir, newdentry, attr->link, debug);
 116                         break;
 117
 118                 default:
 119                         err = -EPERM;
 120                 }
 121         }
 122         if (!err && WARN_ON(!newdentry->d_inode)) {
 123                 /*
 124                  * Not quite sure if non-instantiated dentry is legal or not.
 125                  * VFS doesn't seem to care so check and warn here.
 126                  */
 127                 err = -ENOENT;
 128         }
 129         return err;
 130 }
 131
 132 static int ovl_set_opaque_xerr(struct dentry *dentry, struct dentry *upper,
 133                                int xerr)
 134 {
 135         int err;
 136
 137         err = ovl_check_setxattr(dentry, upper, OVL_XATTR_OPAQUE, "y", 1, xerr);
 138         if (!err)
 139                 ovl_dentry_set_opaque(dentry);
 140
 141         return err;
 142 }
 143
 144 static int ovl_set_opaque(struct dentry *dentry, struct dentry *upperdentry)
 145 {
 146         /*
 147          * Fail with -EIO when trying to create opaque dir and upper doesn't
 148          * support xattrs. ovl_rename() calls ovl_set_opaque_xerr(-EXDEV) to
 149          * return a specific error for noxattr case.
 150          */
 151         return ovl_set_opaque_xerr(dentry, upperdentry, -EIO);
 152 }
 153
 154 /* Common operations required to be done after creation of file on upper */
 155 static void ovl_instantiate(struct dentry *dentry, struct inode *inode,
 156                             struct dentry *newdentry, bool hardlink)
 157 {
 158         ovl_dentry_version_inc(dentry->d_parent, false);
 159         ovl_dentry_set_upper_alias(dentry);
 160         if (!hardlink) {
 161                 ovl_inode_update(inode, newdentry);
 162                 ovl_copyattr(newdentry->d_inode, inode);
 163         } else {
 164                 WARN_ON(ovl_inode_real(inode) != d_inode(newdentry));
 165                 dput(newdentry);
 166                 inc_nlink(inode);
 167         }
 168         d_instantiate(dentry, inode);
 169         /* Force lookup of new upper hardlink to find its lower */
 170         if (hardlink)
 171                 d_drop(dentry);
 172 }
 173
 174 static bool ovl_type_merge(struct dentry *dentry)
 175 {
 176         return OVL_TYPE_MERGE(ovl_path_type(dentry));
 177 }
 178
 179 static bool ovl_type_origin(struct dentry *dentry)
 180 {
 181         return OVL_TYPE_ORIGIN(ovl_path_type(dentry));
 182 }
 183
 184 static int ovl_create_upper(struct dentry *dentry, struct inode *inode,
 185                             struct cattr *attr, struct dentry *hardlink)
 186 {
 187         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 188         struct inode *udir = upperdir->d_inode;
 189         struct dentry *newdentry;
 190         int err;
 191
 192         if (!hardlink && !IS_POSIXACL(udir))
 193                 attr->mode &= ~current_umask();
 194
 195         inode_lock_nested(udir, I_MUTEX_PARENT);
 196         newdentry = lookup_one_len(dentry->d_name.name, upperdir,
 197                                    dentry->d_name.len);
 198         err = PTR_ERR(newdentry);
 199         if (IS_ERR(newdentry))
 200                 goto out_unlock;
 201         err = ovl_create_real(udir, newdentry, attr, hardlink, false);
 202         if (err)
 203                 goto out_dput;
 204
 205         if (ovl_type_merge(dentry->d_parent) && d_is_dir(newdentry)) {
 206                 /* Setting opaque here is just an optimization, allow to fail */
 207                 ovl_set_opaque(dentry, newdentry);
 208         }
 209
 210         ovl_instantiate(dentry, inode, newdentry, !!hardlink);
 211         newdentry = NULL;
 212 out_dput:
 213         dput(newdentry);
 214 out_unlock:
 215         inode_unlock(udir);
 216         return err;
 217 }
 218
 219 static struct dentry *ovl_clear_empty(struct dentry *dentry,
 220                                       struct list_head *list)
 221 {
 222         struct dentry *workdir = ovl_workdir(dentry);
 223         struct inode *wdir = workdir->d_inode;
 224         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 225         struct inode *udir = upperdir->d_inode;
 226         struct path upperpath;
 227         struct dentry *upper;
 228         struct dentry *opaquedir;
 229         struct kstat stat;
 230         int err;
 231
 232         if (WARN_ON(!workdir))
 233                 return ERR_PTR(-EROFS);
 234
 235         err = ovl_lock_rename_workdir(workdir, upperdir);
 236         if (err)
 237                 goto out;
 238
 239         ovl_path_upper(dentry, &upperpath);
 240         err = vfs_getattr(&upperpath, &stat,
 241                           STATX_BASIC_STATS, AT_STATX_SYNC_AS_STAT);
 242         if (err)
 243                 goto out_unlock;
 244
 245         err = -ESTALE;
 246         if (!S_ISDIR(stat.mode))
 247                 goto out_unlock;
 248         upper = upperpath.dentry;
 249         if (upper->d_parent->d_inode != udir)
 250                 goto out_unlock;
 251
 252         opaquedir = ovl_lookup_temp(workdir);
 253         err = PTR_ERR(opaquedir);
 254         if (IS_ERR(opaquedir))
 255                 goto out_unlock;
 256
 257         err = ovl_create_real(wdir, opaquedir,
 258                               &(struct cattr){.mode = stat.mode}, NULL, true);
 259         if (err)
 260                 goto out_dput;
 261
 262         err = ovl_copy_xattr(upper, opaquedir);
 263         if (err)
 264                 goto out_cleanup;
 265
 266         err = ovl_set_opaque(dentry, opaquedir);
 267         if (err)
 268                 goto out_cleanup;
 269
 270         inode_lock(opaquedir->d_inode);
 271         err = ovl_set_attr(opaquedir, &stat);
 272         inode_unlock(opaquedir->d_inode);
 273         if (err)
 274                 goto out_cleanup;
 275
 276         err = ovl_do_rename(wdir, opaquedir, udir, upper, RENAME_EXCHANGE);
 277         if (err)
 278                 goto out_cleanup;
 279
 280         ovl_cleanup_whiteouts(upper, list);
 281         ovl_cleanup(wdir, upper);
 282         unlock_rename(workdir, upperdir);
 283
 284         /* dentry's upper doesn't match now, get rid of it */
 285         d_drop(dentry);
 286
 287         return opaquedir;
 288
 289 out_cleanup:
 290         ovl_cleanup(wdir, opaquedir);
 291 out_dput:
 292         dput(opaquedir);
 293 out_unlock:
 294         unlock_rename(workdir, upperdir);
 295 out:
 296         return ERR_PTR(err);
 297 }
 298
 299 static struct dentry *ovl_check_empty_and_clear(struct dentry *dentry)
 300 {
 301         int err;
 302         struct dentry *ret = NULL;
 303         enum ovl_path_type type = ovl_path_type(dentry);
 304         LIST_HEAD(list);
 305
 306         err = ovl_check_empty_dir(dentry, &list);
 307         if (err) {
 308                 ret = ERR_PTR(err);
 309                 goto out_free;
 310         }
 311
 312         /*
 313          * When removing an empty opaque directory, then it makes no sense to
 314          * replace it with an exact replica of itself.
 315          *
 316          * If no upperdentry then skip clearing whiteouts.
 317          *
 318          * Can race with copy-up, since we don't hold the upperdir mutex.
 319          * Doesn't matter, since copy-up can't create a non-empty directory
 320          * from an empty one.
 321          */
 322         if (OVL_TYPE_UPPER(type) && OVL_TYPE_MERGE(type))
 323                 ret = ovl_clear_empty(dentry, &list);
 324
 325 out_free:
 326         ovl_cache_free(&list);
 327
 328         return ret;
 329 }
 330
 331 static int ovl_set_upper_acl(struct dentry *upperdentry, const char *name,
 332                              const struct posix_acl *acl)
 333 {
 334         void *buffer;
 335         size_t size;
 336         int err;
 337
 338         if (!IS_ENABLED(CONFIG_FS_POSIX_ACL) || !acl)
 339                 return 0;
 340
 341         size = posix_acl_to_xattr(NULL, acl, NULL, 0);
 342         buffer = kmalloc(size, GFP_KERNEL);
 343         if (!buffer)
 344                 return -ENOMEM;
 345
 346         size = posix_acl_to_xattr(&init_user_ns, acl, buffer, size);
 347         err = size;
 348         if (err < 0)
 349                 goto out_free;
 350
 351         err = vfs_setxattr(upperdentry, name, buffer, size, XATTR_CREATE);
 352 out_free:
 353         kfree(buffer);
 354         return err;
 355 }
 356
 357 static int ovl_create_over_whiteout(struct dentry *dentry, struct inode *inode,
 358                                     struct cattr *cattr,
 359                                     struct dentry *hardlink)
 360 {
 361         struct dentry *workdir = ovl_workdir(dentry);
 362         struct inode *wdir = workdir->d_inode;
 363         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 364         struct inode *udir = upperdir->d_inode;
 365         struct dentry *upper;
 366         struct dentry *newdentry;
 367         int err;
 368         struct posix_acl *acl, *default_acl;
 369
 370         if (WARN_ON(!workdir))
 371                 return -EROFS;
 372
 373         if (!hardlink) {
 374                 err = posix_acl_create(dentry->d_parent->d_inode,
 375                                        &cattr->mode, &default_acl, &acl);
 376                 if (err)
 377                         return err;
 378         }
 379
 380         err = ovl_lock_rename_workdir(workdir, upperdir);
 381         if (err)
 382                 goto out;
 383
 384         newdentry = ovl_lookup_temp(workdir);
 385         err = PTR_ERR(newdentry);
 386         if (IS_ERR(newdentry))
 387                 goto out_unlock;
 388
 389         upper = lookup_one_len(dentry->d_name.name, upperdir,
 390                                dentry->d_name.len);
 391         err = PTR_ERR(upper);
 392         if (IS_ERR(upper))
 393                 goto out_dput;
 394
 395         err = -ESTALE;
 396         if (d_is_negative(upper) || !IS_WHITEOUT(d_inode(upper)))
 397                 goto out_dput2;
 398
 399         err = ovl_create_real(wdir, newdentry, cattr, hardlink, true);
 400         if (err)
 401                 goto out_dput2;
 402
 403         /*
 404          * mode could have been mutilated due to umask (e.g. sgid directory)
 405          */
 406         if (!hardlink &&
 407             !S_ISLNK(cattr->mode) &&
 408             newdentry->d_inode->i_mode != cattr->mode) {
 409                 struct iattr attr = {
 410                         .ia_valid = ATTR_MODE,
 411                         .ia_mode = cattr->mode,
 412                 };
 413                 inode_lock(newdentry->d_inode);
 414                 err = notify_change(newdentry, &attr, NULL);
 415                 inode_unlock(newdentry->d_inode);
 416                 if (err)
 417                         goto out_cleanup;
 418         }
 419         if (!hardlink) {
 420                 err = ovl_set_upper_acl(newdentry, XATTR_NAME_POSIX_ACL_ACCESS,
 421                                         acl);
 422                 if (err)
 423                         goto out_cleanup;
 424
 425                 err = ovl_set_upper_acl(newdentry, XATTR_NAME_POSIX_ACL_DEFAULT,
 426                                         default_acl);
 427                 if (err)
 428                         goto out_cleanup;
 429         }
 430
 431         if (!hardlink && S_ISDIR(cattr->mode)) {
 432                 err = ovl_set_opaque(dentry, newdentry);
 433                 if (err)
 434                         goto out_cleanup;
 435
 436                 err = ovl_do_rename(wdir, newdentry, udir, upper,
 437                                     RENAME_EXCHANGE);
 438                 if (err)
 439                         goto out_cleanup;
 440
 441                 ovl_cleanup(wdir, upper);
 442         } else {
 443                 err = ovl_do_rename(wdir, newdentry, udir, upper, 0);
 444                 if (err)
 445                         goto out_cleanup;
 446         }
 447         ovl_instantiate(dentry, inode, newdentry, !!hardlink);
 448         newdentry = NULL;
 449 out_dput2:
 450         dput(upper);
 451 out_dput:
 452         dput(newdentry);
 453 out_unlock:
 454         unlock_rename(workdir, upperdir);
 455 out:
 456         if (!hardlink) {
 457                 posix_acl_release(acl);
 458                 posix_acl_release(default_acl);
 459         }
 460         return err;
 461
 462 out_cleanup:
 463         ovl_cleanup(wdir, newdentry);
 464         goto out_dput2;
 465 }
 466
 467 static int ovl_create_or_link(struct dentry *dentry, struct inode *inode,
 468                               struct cattr *attr, struct dentry *hardlink,
 469                               bool origin)
 470 {
 471         int err;
 472         const struct cred *old_cred;
 473         struct cred *override_cred;
 474         struct dentry *parent = dentry->d_parent;
 475
 476         err = ovl_copy_up(parent);
 477         if (err)
 478                 return err;
 479
 480         old_cred = ovl_override_creds(dentry->d_sb);
 481
 482         /*
 483          * When linking a file with copy up origin into a new parent, mark the
 484          * new parent dir "impure".
 485          */
 486         if (origin) {
 487                 err = ovl_set_impure(parent, ovl_dentry_upper(parent));
 488                 if (err)
 489                         goto out_revert_creds;
 490         }
 491
 492         err = -ENOMEM;
 493         override_cred = prepare_creds();
 494         if (override_cred) {
 495                 override_cred->fsuid = inode->i_uid;
 496                 override_cred->fsgid = inode->i_gid;
 497                 if (!hardlink) {
 498                         err = security_dentry_create_files_as(dentry,
 499                                         attr->mode, &dentry->d_name, old_cred,
 500                                         override_cred);
 501                         if (err) {
 502                                 put_cred(override_cred);
 503                                 goto out_revert_creds;
 504                         }
 505                 }
 506                 put_cred(override_creds(override_cred));
 507                 put_cred(override_cred);
 508
 509                 if (!ovl_dentry_is_whiteout(dentry))
 510                         err = ovl_create_upper(dentry, inode, attr,
 511                                                 hardlink);
 512                 else
 513                         err = ovl_create_over_whiteout(dentry, inode, attr,
 514                                                         hardlink);
 515         }
 516 out_revert_creds:
 517         revert_creds(old_cred);
 518         if (!err) {
 519                 struct inode *realinode = d_inode(ovl_dentry_upper(dentry));
 520
 521                 WARN_ON(inode->i_mode != realinode->i_mode);
 522                 WARN_ON(!uid_eq(inode->i_uid, realinode->i_uid));
 523                 WARN_ON(!gid_eq(inode->i_gid, realinode->i_gid));
 524         }
 525         return err;
 526 }
 527
 528 static int ovl_create_object(struct dentry *dentry, int mode, dev_t rdev,
 529                              const char *link)
 530 {
 531         int err;
 532         struct inode *inode;
 533         struct cattr attr = {
 534                 .rdev = rdev,
 535                 .link = link,
 536         };
 537
 538         err = ovl_want_write(dentry);
 539         if (err)
 540                 goto out;
 541
 542         err = -ENOMEM;
 543         inode = ovl_new_inode(dentry->d_sb, mode, rdev);
 544         if (!inode)
 545                 goto out_drop_write;
 546
 547         inode_init_owner(inode, dentry->d_parent->d_inode, mode);
 548         attr.mode = inode->i_mode;
 549
 550         err = ovl_create_or_link(dentry, inode, &attr, NULL, false);
 551         if (err)
 552                 iput(inode);
 553
 554 out_drop_write:
 555         ovl_drop_write(dentry);
 556 out:
 557         return err;
 558 }
 559
 560 static int ovl_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 561                       bool excl)
 562 {
 563         return ovl_create_object(dentry, (mode & 07777) | S_IFREG, 0, NULL);
 564 }
 565
 566 static int ovl_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 567 {
 568         return ovl_create_object(dentry, (mode & 07777) | S_IFDIR, 0, NULL);
 569 }
 570
 571 static int ovl_mknod(struct inode *dir, struct dentry *dentry, umode_t mode,
 572                      dev_t rdev)
 573 {
 574         /* Don't allow creation of "whiteout" on overlay */
 575         if (S_ISCHR(mode) && rdev == WHITEOUT_DEV)
 576                 return -EPERM;
 577
 578         return ovl_create_object(dentry, mode, rdev, NULL);
 579 }
 580
 581 static int ovl_symlink(struct inode *dir, struct dentry *dentry,
 582                        const char *link)
 583 {
 584         return ovl_create_object(dentry, S_IFLNK, 0, link);
 585 }
 586
 587 static int ovl_link(struct dentry *old, struct inode *newdir,
 588                     struct dentry *new)
 589 {
 590         int err;
 591         bool locked = false;
 592         struct inode *inode;
 593
 594         err = ovl_want_write(old);
 595         if (err)
 596                 goto out;
 597
 598         err = ovl_copy_up(old);
 599         if (err)
 600                 goto out_drop_write;
 601
 602         err = ovl_copy_up(new->d_parent);
 603         if (err)
 604                 goto out_drop_write;
 605
 606
 607         err = ovl_nlink_start(old, &locked);
 608         if (err)
 609                 goto out_drop_write;
 610
 611         inode = d_inode(old);
 612         ihold(inode);
 613
 614         err = ovl_create_or_link(new, inode, NULL, ovl_dentry_upper(old),
 615                                  ovl_type_origin(old));
 616         if (err)
 617                 iput(inode);
 618
 619         ovl_nlink_end(old, locked);
 620 out_drop_write:
 621         ovl_drop_write(old);
 622 out:
 623         return err;
 624 }
 625
 626 static bool ovl_matches_upper(struct dentry *dentry, struct dentry *upper)
 627 {
 628         return d_inode(ovl_dentry_upper(dentry)) == d_inode(upper);
 629 }
 630
 631 static int ovl_remove_and_whiteout(struct dentry *dentry, bool is_dir)
 632 {
 633         struct dentry *workdir = ovl_workdir(dentry);
 634         struct inode *wdir = workdir->d_inode;
 635         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 636         struct inode *udir = upperdir->d_inode;
 637         struct dentry *whiteout;
 638         struct dentry *upper;
 639         struct dentry *opaquedir = NULL;
 640         int err;
 641         int flags = 0;
 642
 643         if (WARN_ON(!workdir))
 644                 return -EROFS;
 645
 646         if (is_dir) {
 647                 opaquedir = ovl_check_empty_and_clear(dentry);
 648                 err = PTR_ERR(opaquedir);
 649                 if (IS_ERR(opaquedir))
 650                         goto out;
 651         }
 652
 653         err = ovl_lock_rename_workdir(workdir, upperdir);
 654         if (err)
 655                 goto out_dput;
 656
 657         upper = lookup_one_len(dentry->d_name.name, upperdir,
 658                                dentry->d_name.len);
 659         err = PTR_ERR(upper);
 660         if (IS_ERR(upper))
 661                 goto out_unlock;
 662
 663         err = -ESTALE;
 664         if ((opaquedir && upper != opaquedir) ||
 665             (!opaquedir && ovl_dentry_upper(dentry) &&
 666              !ovl_matches_upper(dentry, upper))) {
 667                 goto out_dput_upper;
 668         }
 669
 670         whiteout = ovl_whiteout(workdir, dentry);
 671         err = PTR_ERR(whiteout);
 672         if (IS_ERR(whiteout))
 673                 goto out_dput_upper;
 674
 675         if (d_is_dir(upper))
 676                 flags = RENAME_EXCHANGE;
 677
 678         err = ovl_do_rename(wdir, whiteout, udir, upper, flags);
 679         if (err)
 680                 goto kill_whiteout;
 681         if (flags)
 682                 ovl_cleanup(wdir, upper);
 683
 684         ovl_dentry_version_inc(dentry->d_parent, true);
 685 out_d_drop:
 686         d_drop(dentry);
 687         dput(whiteout);
 688 out_dput_upper:
 689         dput(upper);
 690 out_unlock:
 691         unlock_rename(workdir, upperdir);
 692 out_dput:
 693         dput(opaquedir);
 694 out:
 695         return err;
 696
 697 kill_whiteout:
 698         ovl_cleanup(wdir, whiteout);
 699         goto out_d_drop;
 700 }
 701
 702 static int ovl_remove_upper(struct dentry *dentry, bool is_dir)
 703 {
 704         struct dentry *upperdir = ovl_dentry_upper(dentry->d_parent);
 705         struct inode *dir = upperdir->d_inode;
 706         struct dentry *upper;
 707         struct dentry *opaquedir = NULL;
 708         int err;
 709
 710         /* Redirect dir can be !ovl_lower_positive && OVL_TYPE_MERGE */
 711         if (is_dir && ovl_dentry_get_redirect(dentry)) {
 712                 opaquedir = ovl_check_empty_and_clear(dentry);
 713                 err = PTR_ERR(opaquedir);
 714                 if (IS_ERR(opaquedir))
 715                         goto out;
 716         }
 717
 718         inode_lock_nested(dir, I_MUTEX_PARENT);
 719         upper = lookup_one_len(dentry->d_name.name, upperdir,
 720                                dentry->d_name.len);
 721         err = PTR_ERR(upper);
 722         if (IS_ERR(upper))
 723                 goto out_unlock;
 724
 725         err = -ESTALE;
 726         if ((opaquedir && upper != opaquedir) ||
 727             (!opaquedir && !ovl_matches_upper(dentry, upper)))
 728                 goto out_dput_upper;
 729
 730         if (is_dir)
 731                 err = vfs_rmdir(dir, upper);
 732         else
 733                 err = vfs_unlink(dir, upper, NULL);
 734         ovl_dentry_version_inc(dentry->d_parent, ovl_type_origin(dentry));
 735
 736         /*
 737          * Keeping this dentry hashed would mean having to release
 738          * upperpath/lowerpath, which could only be done if we are the
 739          * sole user of this dentry.  Too tricky...  Just unhash for
 740          * now.
 741          */
 742         if (!err)
 743                 d_drop(dentry);
 744 out_dput_upper:
 745         dput(upper);
 746 out_unlock:
 747         inode_unlock(dir);
 748         dput(opaquedir);
 749 out:
 750         return err;
 751 }
 752
 753 static int ovl_do_remove(struct dentry *dentry, bool is_dir)
 754 {
 755         int err;
 756         bool locked = false;
 757         const struct cred *old_cred;
 758
 759         err = ovl_want_write(dentry);
 760         if (err)
 761                 goto out;
 762
 763         err = ovl_copy_up(dentry->d_parent);
 764         if (err)
 765                 goto out_drop_write;
 766
 767         err = ovl_nlink_start(dentry, &locked);
 768         if (err)
 769                 goto out_drop_write;
 770
 771         old_cred = ovl_override_creds(dentry->d_sb);
 772         if (!ovl_lower_positive(dentry))
 773                 err = ovl_remove_upper(dentry, is_dir);
 774         else
 775                 err = ovl_remove_and_whiteout(dentry, is_dir);
 776         revert_creds(old_cred);
 777         if (!err) {
 778                 if (is_dir)
 779                         clear_nlink(dentry->d_inode);
 780                 else
 781                         drop_nlink(dentry->d_inode);
 782         }
 783         ovl_nlink_end(dentry, locked);
 784 out_drop_write:
 785         ovl_drop_write(dentry);
 786 out:
 787         return err;
 788 }
 789
 790 static int ovl_unlink(struct inode *dir, struct dentry *dentry)
 791 {
 792         return ovl_do_remove(dentry, false);
 793 }
 794
 795 static int ovl_rmdir(struct inode *dir, struct dentry *dentry)
 796 {
 797         return ovl_do_remove(dentry, true);
 798 }
 799
 800 static bool ovl_type_merge_or_lower(struct dentry *dentry)
 801 {
 802         enum ovl_path_type type = ovl_path_type(dentry);
 803
 804         return OVL_TYPE_MERGE(type) || !OVL_TYPE_UPPER(type);
 805 }
 806
 807 static bool ovl_can_move(struct dentry *dentry)
 808 {
 809         return ovl_redirect_dir(dentry->d_sb) ||
 810                 !d_is_dir(dentry) || !ovl_type_merge_or_lower(dentry);
 811 }
 812
 813 static char *ovl_get_redirect(struct dentry *dentry, bool samedir)
 814 {
 815         char *buf, *ret;
 816         struct dentry *d, *tmp;
 817         int buflen = ovl_redirect_max + 1;
 818
 819         if (samedir) {
 820                 ret = kstrndup(dentry->d_name.name, dentry->d_name.len,
 821                                GFP_KERNEL);
 822                 goto out;
 823         }
 824
 825         buf = ret = kmalloc(buflen, GFP_KERNEL);
 826         if (!buf)
 827                 goto out;
 828
 829         buflen--;
 830         buf[buflen] = '\0';
 831         for (d = dget(dentry); !IS_ROOT(d);) {
 832                 const char *name;
 833                 int thislen;
 834
 835                 spin_lock(&d->d_lock);
 836                 name = ovl_dentry_get_redirect(d);
 837                 if (name) {
 838                         thislen = strlen(name);
 839                 } else {
 840                         name = d->d_name.name;
 841                         thislen = d->d_name.len;
 842                 }
 843
 844                 /* If path is too long, fall back to userspace move */
 845                 if (thislen + (name[0] != '/') > buflen) {
 846                         ret = ERR_PTR(-EXDEV);
 847                         spin_unlock(&d->d_lock);
 848                         goto out_put;
 849                 }
 850
 851                 buflen -= thislen;
 852                 memcpy(&buf[buflen], name, thislen);
 853                 spin_unlock(&d->d_lock);
 854                 tmp = dget_parent(d);
 855
 856                 dput(d);
 857                 d = tmp;
 858
 859                 /* Absolute redirect: finished */
 860                 if (buf[buflen] == '/')
 861                         break;
 862                 buflen--;
 863                 buf[buflen] = '/';
 864         }
 865         ret = kstrdup(&buf[buflen], GFP_KERNEL);
 866 out_put:
 867         dput(d);
 868         kfree(buf);
 869 out:
 870         return ret ? ret : ERR_PTR(-ENOMEM);
 871 }
 872
 873 static int ovl_set_redirect(struct dentry *dentry, bool samedir)
 874 {
 875         int err;
 876         const char *redirect = ovl_dentry_get_redirect(dentry);
 877
 878         if (redirect && (samedir || redirect[0] == '/'))
 879                 return 0;
 880
 881         redirect = ovl_get_redirect(dentry, samedir);
 882         if (IS_ERR(redirect))
 883                 return PTR_ERR(redirect);
 884
 885         err = ovl_check_setxattr(dentry, ovl_dentry_upper(dentry),
 886                                  OVL_XATTR_REDIRECT,
 887                                  redirect, strlen(redirect), -EXDEV);
 888         if (!err) {
 889                 spin_lock(&dentry->d_lock);
 890                 ovl_dentry_set_redirect(dentry, redirect);
 891                 spin_unlock(&dentry->d_lock);
 892         } else {
 893                 kfree(redirect);
 894                 pr_warn_ratelimited("overlay: failed to set redirect (%i)\n", err);
 895                 /* Fall back to userspace copy-up */
 896                 err = -EXDEV;
 897         }
 898         return err;
 899 }
 900
 901 static int ovl_rename(struct inode *olddir, struct dentry *old,
 902                       struct inode *newdir, struct dentry *new,
 903                       unsigned int flags)
 904 {
 905         int err;
 906         bool locked = false;
 907         struct dentry *old_upperdir;
 908         struct dentry *new_upperdir;
 909         struct dentry *olddentry;
 910         struct dentry *newdentry;
 911         struct dentry *trap;
 912         bool old_opaque;
 913         bool new_opaque;
 914         bool cleanup_whiteout = false;
 915         bool overwrite = !(flags & RENAME_EXCHANGE);
 916         bool is_dir = d_is_dir(old);
 917         bool new_is_dir = d_is_dir(new);
 918         bool samedir = olddir == newdir;
 919         struct dentry *opaquedir = NULL;
 920         const struct cred *old_cred = NULL;
 921
 922         err = -EINVAL;
 923         if (flags & ~(RENAME_EXCHANGE | RENAME_NOREPLACE))
 924                 goto out;
 925
 926         flags &= ~RENAME_NOREPLACE;
 927
 928         /* Don't copy up directory trees */
 929         err = -EXDEV;
 930         if (!ovl_can_move(old))
 931                 goto out;
 932         if (!overwrite && !ovl_can_move(new))
 933                 goto out;
 934
 935         err = ovl_want_write(old);
 936         if (err)
 937                 goto out;
 938
 939         err = ovl_copy_up(old);
 940         if (err)
 941                 goto out_drop_write;
 942
 943         err = ovl_copy_up(new->d_parent);
 944         if (err)
 945                 goto out_drop_write;
 946         if (!overwrite) {
 947                 err = ovl_copy_up(new);
 948                 if (err)
 949                         goto out_drop_write;
 950         } else {
 951                 err = ovl_nlink_start(new, &locked);
 952                 if (err)
 953                         goto out_drop_write;
 954         }
 955
 956         old_cred = ovl_override_creds(old->d_sb);
 957
 958         if (overwrite && new_is_dir && ovl_type_merge_or_lower(new)) {
 959                 opaquedir = ovl_check_empty_and_clear(new);
 960                 err = PTR_ERR(opaquedir);
 961                 if (IS_ERR(opaquedir)) {
 962                         opaquedir = NULL;
 963                         goto out_revert_creds;
 964                 }
 965         }
 966
 967         if (overwrite) {
 968                 if (ovl_lower_positive(old)) {
 969                         if (!ovl_dentry_is_whiteout(new)) {
 970                                 /* Whiteout source */
 971                                 flags |= RENAME_WHITEOUT;
 972                         } else {
 973                                 /* Switch whiteouts */
 974                                 flags |= RENAME_EXCHANGE;
 975                         }
 976                 } else if (is_dir && ovl_dentry_is_whiteout(new)) {
 977                         flags |= RENAME_EXCHANGE;
 978                         cleanup_whiteout = true;
 979                 }
 980         }
 981
 982         old_upperdir = ovl_dentry_upper(old->d_parent);
 983         new_upperdir = ovl_dentry_upper(new->d_parent);
 984
 985         if (!samedir) {
 986                 /*
 987                  * When moving a merge dir or non-dir with copy up origin into
 988                  * a new parent, we are marking the new parent dir "impure".
 989                  * When ovl_iterate() iterates an "impure" upper dir, it will
 990                  * lookup the origin inodes of the entries to fill d_ino.
 991                  */
 992                 if (ovl_type_origin(old)) {
 993                         err = ovl_set_impure(new->d_parent, new_upperdir);
 994                         if (err)
 995                                 goto out_revert_creds;
 996                 }
 997                 if (!overwrite && ovl_type_origin(new)) {
 998                         err = ovl_set_impure(old->d_parent, old_upperdir);
 999                         if (err)
1000                                 goto out_revert_creds;
1001                 }
1002         }
1003
1004         trap = lock_rename(new_upperdir, old_upperdir);
1005
1006         olddentry = lookup_one_len(old->d_name.name, old_upperdir,
1007                                    old->d_name.len);
1008         err = PTR_ERR(olddentry);
1009         if (IS_ERR(olddentry))
1010                 goto out_unlock;
1011
1012         err = -ESTALE;
1013         if (!ovl_matches_upper(old, olddentry))
1014                 goto out_dput_old;
1015
1016         newdentry = lookup_one_len(new->d_name.name, new_upperdir,
1017                                    new->d_name.len);
1018         err = PTR_ERR(newdentry);
1019         if (IS_ERR(newdentry))
1020                 goto out_dput_old;
1021
1022         old_opaque = ovl_dentry_is_opaque(old);
1023         new_opaque = ovl_dentry_is_opaque(new);
1024
1025         err = -ESTALE;
1026         if (d_inode(new) && ovl_dentry_upper(new)) {
1027                 if (opaquedir) {
1028                         if (newdentry != opaquedir)
1029                                 goto out_dput;
1030                 } else {
1031                         if (!ovl_matches_upper(new, newdentry))
1032                                 goto out_dput;
1033                 }
1034         } else {
1035                 if (!d_is_negative(newdentry)) {
1036                         if (!new_opaque || !ovl_is_whiteout(newdentry))
1037                                 goto out_dput;
1038                 } else {
1039                         if (flags & RENAME_EXCHANGE)
1040                                 goto out_dput;
1041                 }
1042         }
1043
1044         if (olddentry == trap)
1045                 goto out_dput;
1046         if (newdentry == trap)
1047                 goto out_dput;
1048
1049         if (olddentry->d_inode == newdentry->d_inode)
1050                 goto out_dput;
1051
1052         err = 0;
1053         if (is_dir) {
1054                 if (ovl_type_merge_or_lower(old))
1055                         err = ovl_set_redirect(old, samedir);
1056                 else if (!old_opaque && ovl_type_merge(new->d_parent))
1057                         err = ovl_set_opaque_xerr(old, olddentry, -EXDEV);
1058                 if (err)
1059                         goto out_dput;
1060         }
1061         if (!overwrite && new_is_dir) {
1062                 if (ovl_type_merge_or_lower(new))
1063                         err = ovl_set_redirect(new, samedir);
1064                 else if (!new_opaque && ovl_type_merge(old->d_parent))
1065                         err = ovl_set_opaque_xerr(new, newdentry, -EXDEV);
1066                 if (err)
1067                         goto out_dput;
1068         }
1069
1070         err = ovl_do_rename(old_upperdir->d_inode, olddentry,
1071                             new_upperdir->d_inode, newdentry, flags);
1072         if (err)
1073                 goto out_dput;
1074
1075         if (cleanup_whiteout)
1076                 ovl_cleanup(old_upperdir->d_inode, newdentry);
1077
1078         if (overwrite && d_inode(new)) {
1079                 if (new_is_dir)
1080                         clear_nlink(d_inode(new));
1081                 else
1082                         drop_nlink(d_inode(new));
1083         }
1084
1085         ovl_dentry_version_inc(old->d_parent,
1086                                !overwrite && ovl_type_origin(new));
1087         ovl_dentry_version_inc(new->d_parent, ovl_type_origin(old));
1088
1089 out_dput:
1090         dput(newdentry);
1091 out_dput_old:
1092         dput(olddentry);
1093 out_unlock:
1094         unlock_rename(new_upperdir, old_upperdir);
1095 out_revert_creds:
1096         revert_creds(old_cred);
1097         ovl_nlink_end(new, locked);
1098 out_drop_write:
1099         ovl_drop_write(old);
1100 out:
1101         dput(opaquedir);
1102         return err;
1103 }
1104
1105 const struct inode_operations ovl_dir_inode_operations = {
1106         .lookup         = ovl_lookup,
1107         .mkdir          = ovl_mkdir,
1108         .symlink        = ovl_symlink,
1109         .unlink         = ovl_unlink,
1110         .rmdir          = ovl_rmdir,
1111         .rename         = ovl_rename,
1112         .link           = ovl_link,
1113         .setattr        = ovl_setattr,
1114         .create         = ovl_create,
1115         .mknod          = ovl_mknod,
1116         .permission     = ovl_permission,
1117         .getattr        = ovl_getattr,
1118         .listxattr      = ovl_listxattr,
1119         .get_acl        = ovl_get_acl,
1120         .update_time    = ovl_update_time,
1121 };