fs/overlayfs/readdir.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/slab.h>
   9 #include <linux/namei.h>
  10 #include <linux/file.h>
  11 #include <linux/xattr.h>
  12 #include <linux/rbtree.h>
  13 #include <linux/security.h>
  14 #include <linux/cred.h>
  15 #include <linux/ratelimit.h>
  16 #include "overlayfs.h"
  17
  18 struct ovl_cache_entry {
  19         unsigned int len;
  20         unsigned int type;
  21         u64 real_ino;
  22         u64 ino;
  23         struct list_head l_node;
  24         struct rb_node node;
  25         struct ovl_cache_entry *next_maybe_whiteout;
  26         bool is_upper;
  27         bool is_whiteout;
  28         char name[];
  29 };
  30
  31 struct ovl_dir_cache {
  32         long refcount;
  33         u64 version;
  34         struct list_head entries;
  35         struct rb_root root;
  36 };
  37
  38 struct ovl_readdir_data {
  39         struct dir_context ctx;
  40         struct dentry *dentry;
  41         bool is_lowest;
  42         struct rb_root *root;
  43         struct list_head *list;
  44         struct list_head middle;
  45         struct ovl_cache_entry *first_maybe_whiteout;
  46         int count;
  47         int err;
  48         bool is_upper;
  49         bool d_type_supported;
  50 };
  51
  52 struct ovl_dir_file {
  53         bool is_real;
  54         bool is_upper;
  55         struct ovl_dir_cache *cache;
  56         struct list_head *cursor;
  57         struct file *realfile;
  58         struct file *upperfile;
  59 };
  60
  61 static struct ovl_cache_entry *ovl_cache_entry_from_node(struct rb_node *n)
  62 {
  63         return rb_entry(n, struct ovl_cache_entry, node);
  64 }
  65
  66 static bool ovl_cache_entry_find_link(const char *name, int len,
  67                                       struct rb_node ***link,
  68                                       struct rb_node **parent)
  69 {
  70         bool found = false;
  71         struct rb_node **newp = *link;
  72
  73         while (!found && *newp) {
  74                 int cmp;
  75                 struct ovl_cache_entry *tmp;
  76
  77                 *parent = *newp;
  78                 tmp = ovl_cache_entry_from_node(*newp);
  79                 cmp = strncmp(name, tmp->name, len);
  80                 if (cmp > 0)
  81                         newp = &tmp->node.rb_right;
  82                 else if (cmp < 0 || len < tmp->len)
  83                         newp = &tmp->node.rb_left;
  84                 else
  85                         found = true;
  86         }
  87         *link = newp;
  88
  89         return found;
  90 }
  91
  92 static struct ovl_cache_entry *ovl_cache_entry_find(struct rb_root *root,
  93                                                     const char *name, int len)
  94 {
  95         struct rb_node *node = root->rb_node;
  96         int cmp;
  97
  98         while (node) {
  99                 struct ovl_cache_entry *p = ovl_cache_entry_from_node(node);
 100
 101                 cmp = strncmp(name, p->name, len);
 102                 if (cmp > 0)
 103                         node = p->node.rb_right;
 104                 else if (cmp < 0 || len < p->len)
 105                         node = p->node.rb_left;
 106                 else
 107                         return p;
 108         }
 109
 110         return NULL;
 111 }
 112
 113 static bool ovl_calc_d_ino(struct ovl_readdir_data *rdd,
 114                            struct ovl_cache_entry *p)
 115 {
 116         /* Don't care if not doing ovl_iter() */
 117         if (!rdd->dentry)
 118                 return false;
 119
 120         /* Always recalc d_ino when remapping lower inode numbers */
 121         if (ovl_xino_bits(rdd->dentry->d_sb))
 122                 return true;
 123
 124         /* Always recalc d_ino for parent */
 125         if (strcmp(p->name, "..") == 0)
 126                 return true;
 127
 128         /* If this is lower, then native d_ino will do */
 129         if (!rdd->is_upper)
 130                 return false;
 131
 132         /*
 133          * Recalc d_ino for '.' and for all entries if dir is impure (contains
 134          * copied up entries)
 135          */
 136         if ((p->name[0] == '.' && p->len == 1) ||
 137             ovl_test_flag(OVL_IMPURE, d_inode(rdd->dentry)))
 138                 return true;
 139
 140         return false;
 141 }
 142
 143 static struct ovl_cache_entry *ovl_cache_entry_new(struct ovl_readdir_data *rdd,
 144                                                    const char *name, int len,
 145                                                    u64 ino, unsigned int d_type)
 146 {
 147         struct ovl_cache_entry *p;
 148         size_t size = offsetof(struct ovl_cache_entry, name[len + 1]);
 149
 150         p = kmalloc(size, GFP_KERNEL);
 151         if (!p)
 152                 return NULL;
 153
 154         memcpy(p->name, name, len);
 155         p->name[len] = '\0';
 156         p->len = len;
 157         p->type = d_type;
 158         p->real_ino = ino;
 159         p->ino = ino;
 160         /* Defer setting d_ino for upper entry to ovl_iterate() */
 161         if (ovl_calc_d_ino(rdd, p))
 162                 p->ino = 0;
 163         p->is_upper = rdd->is_upper;
 164         p->is_whiteout = false;
 165
 166         if (d_type == DT_CHR) {
 167                 p->next_maybe_whiteout = rdd->first_maybe_whiteout;
 168                 rdd->first_maybe_whiteout = p;
 169         }
 170         return p;
 171 }
 172
 173 static int ovl_cache_entry_add_rb(struct ovl_readdir_data *rdd,
 174                                   const char *name, int len, u64 ino,
 175                                   unsigned int d_type)
 176 {
 177         struct rb_node **newp = &rdd->root->rb_node;
 178         struct rb_node *parent = NULL;
 179         struct ovl_cache_entry *p;
 180
 181         if (ovl_cache_entry_find_link(name, len, &newp, &parent))
 182                 return 0;
 183
 184         p = ovl_cache_entry_new(rdd, name, len, ino, d_type);
 185         if (p == NULL) {
 186                 rdd->err = -ENOMEM;
 187                 return -ENOMEM;
 188         }
 189
 190         list_add_tail(&p->l_node, rdd->list);
 191         rb_link_node(&p->node, parent, newp);
 192         rb_insert_color(&p->node, rdd->root);
 193
 194         return 0;
 195 }
 196
 197 static int ovl_fill_lowest(struct ovl_readdir_data *rdd,
 198                            const char *name, int namelen,
 199                            loff_t offset, u64 ino, unsigned int d_type)
 200 {
 201         struct ovl_cache_entry *p;
 202
 203         p = ovl_cache_entry_find(rdd->root, name, namelen);
 204         if (p) {
 205                 list_move_tail(&p->l_node, &rdd->middle);
 206         } else {
 207                 p = ovl_cache_entry_new(rdd, name, namelen, ino, d_type);
 208                 if (p == NULL)
 209                         rdd->err = -ENOMEM;
 210                 else
 211                         list_add_tail(&p->l_node, &rdd->middle);
 212         }
 213
 214         return rdd->err;
 215 }
 216
 217 void ovl_cache_free(struct list_head *list)
 218 {
 219         struct ovl_cache_entry *p;
 220         struct ovl_cache_entry *n;
 221
 222         list_for_each_entry_safe(p, n, list, l_node)
 223                 kfree(p);
 224
 225         INIT_LIST_HEAD(list);
 226 }
 227
 228 void ovl_dir_cache_free(struct inode *inode)
 229 {
 230         struct ovl_dir_cache *cache = ovl_dir_cache(inode);
 231
 232         if (cache) {
 233                 ovl_cache_free(&cache->entries);
 234                 kfree(cache);
 235         }
 236 }
 237
 238 static void ovl_cache_put(struct ovl_dir_file *od, struct dentry *dentry)
 239 {
 240         struct ovl_dir_cache *cache = od->cache;
 241
 242         WARN_ON(cache->refcount <= 0);
 243         cache->refcount--;
 244         if (!cache->refcount) {
 245                 if (ovl_dir_cache(d_inode(dentry)) == cache)
 246                         ovl_set_dir_cache(d_inode(dentry), NULL);
 247
 248                 ovl_cache_free(&cache->entries);
 249                 kfree(cache);
 250         }
 251 }
 252
 253 static int ovl_fill_merge(struct dir_context *ctx, const char *name,
 254                           int namelen, loff_t offset, u64 ino,
 255                           unsigned int d_type)
 256 {
 257         struct ovl_readdir_data *rdd =
 258                 container_of(ctx, struct ovl_readdir_data, ctx);
 259
 260         rdd->count++;
 261         if (!rdd->is_lowest)
 262                 return ovl_cache_entry_add_rb(rdd, name, namelen, ino, d_type);
 263         else
 264                 return ovl_fill_lowest(rdd, name, namelen, offset, ino, d_type);
 265 }
 266
 267 static int ovl_check_whiteouts(struct dentry *dir, struct ovl_readdir_data *rdd)
 268 {
 269         int err;
 270         struct ovl_cache_entry *p;
 271         struct dentry *dentry;
 272         const struct cred *old_cred;
 273
 274         old_cred = ovl_override_creds(rdd->dentry->d_sb);
 275
 276         err = down_write_killable(&dir->d_inode->i_rwsem);
 277         if (!err) {
 278                 while (rdd->first_maybe_whiteout) {
 279                         p = rdd->first_maybe_whiteout;
 280                         rdd->first_maybe_whiteout = p->next_maybe_whiteout;
 281                         dentry = lookup_one_len(p->name, dir, p->len);
 282                         if (!IS_ERR(dentry)) {
 283                                 p->is_whiteout = ovl_is_whiteout(dentry);
 284                                 dput(dentry);
 285                         }
 286                 }
 287                 inode_unlock(dir->d_inode);
 288         }
 289         revert_creds(old_cred);
 290
 291         return err;
 292 }
 293
 294 static inline int ovl_dir_read(struct path *realpath,
 295                                struct ovl_readdir_data *rdd)
 296 {
 297         struct file *realfile;
 298         int err;
 299
 300         realfile = ovl_path_open(realpath, O_RDONLY | O_LARGEFILE);
 301         if (IS_ERR(realfile))
 302                 return PTR_ERR(realfile);
 303
 304         rdd->first_maybe_whiteout = NULL;
 305         rdd->ctx.pos = 0;
 306         do {
 307                 rdd->count = 0;
 308                 rdd->err = 0;
 309                 err = iterate_dir(realfile, &rdd->ctx);
 310                 if (err >= 0)
 311                         err = rdd->err;
 312         } while (!err && rdd->count);
 313
 314         if (!err && rdd->first_maybe_whiteout && rdd->dentry)
 315                 err = ovl_check_whiteouts(realpath->dentry, rdd);
 316
 317         fput(realfile);
 318
 319         return err;
 320 }
 321
 322 static void ovl_dir_reset(struct file *file)
 323 {
 324         struct ovl_dir_file *od = file->private_data;
 325         struct ovl_dir_cache *cache = od->cache;
 326         struct dentry *dentry = file->f_path.dentry;
 327         bool is_real;
 328
 329         if (cache && ovl_dentry_version_get(dentry) != cache->version) {
 330                 ovl_cache_put(od, dentry);
 331                 od->cache = NULL;
 332                 od->cursor = NULL;
 333         }
 334         is_real = ovl_dir_is_real(dentry);
 335         if (od->is_real != is_real) {
 336                 /* is_real can only become false when dir is copied up */
 337                 if (WARN_ON(is_real))
 338                         return;
 339                 od->is_real = false;
 340         }
 341 }
 342
 343 static int ovl_dir_read_merged(struct dentry *dentry, struct list_head *list,
 344         struct rb_root *root)
 345 {
 346         int err;
 347         struct path realpath;
 348         struct ovl_readdir_data rdd = {
 349                 .ctx.actor = ovl_fill_merge,
 350                 .dentry = dentry,
 351                 .list = list,
 352                 .root = root,
 353                 .is_lowest = false,
 354         };
 355         int idx, next;
 356
 357         for (idx = 0; idx != -1; idx = next) {
 358                 next = ovl_path_next(idx, dentry, &realpath);
 359                 rdd.is_upper = ovl_dentry_upper(dentry) == realpath.dentry;
 360
 361                 if (next != -1) {
 362                         err = ovl_dir_read(&realpath, &rdd);
 363                         if (err)
 364                                 break;
 365                 } else {
 366                         /*
 367                          * Insert lowest layer entries before upper ones, this
 368                          * allows offsets to be reasonably constant
 369                          */
 370                         list_add(&rdd.middle, rdd.list);
 371                         rdd.is_lowest = true;
 372                         err = ovl_dir_read(&realpath, &rdd);
 373                         list_del(&rdd.middle);
 374                 }
 375         }
 376         return err;
 377 }
 378
 379 static void ovl_seek_cursor(struct ovl_dir_file *od, loff_t pos)
 380 {
 381         struct list_head *p;
 382         loff_t off = 0;
 383
 384         list_for_each(p, &od->cache->entries) {
 385                 if (off >= pos)
 386                         break;
 387                 off++;
 388         }
 389         /* Cursor is safe since the cache is stable */
 390         od->cursor = p;
 391 }
 392
 393 static struct ovl_dir_cache *ovl_cache_get(struct dentry *dentry)
 394 {
 395         int res;
 396         struct ovl_dir_cache *cache;
 397
 398         cache = ovl_dir_cache(d_inode(dentry));
 399         if (cache && ovl_dentry_version_get(dentry) == cache->version) {
 400                 WARN_ON(!cache->refcount);
 401                 cache->refcount++;
 402                 return cache;
 403         }
 404         ovl_set_dir_cache(d_inode(dentry), NULL);
 405
 406         cache = kzalloc(sizeof(struct ovl_dir_cache), GFP_KERNEL);
 407         if (!cache)
 408                 return ERR_PTR(-ENOMEM);
 409
 410         cache->refcount = 1;
 411         INIT_LIST_HEAD(&cache->entries);
 412         cache->root = RB_ROOT;
 413
 414         res = ovl_dir_read_merged(dentry, &cache->entries, &cache->root);
 415         if (res) {
 416                 ovl_cache_free(&cache->entries);
 417                 kfree(cache);
 418                 return ERR_PTR(res);
 419         }
 420
 421         cache->version = ovl_dentry_version_get(dentry);
 422         ovl_set_dir_cache(d_inode(dentry), cache);
 423
 424         return cache;
 425 }
 426
 427 /* Map inode number to lower fs unique range */
 428 static u64 ovl_remap_lower_ino(u64 ino, int xinobits, int fsid,
 429                                const char *name, int namelen, bool warn)
 430 {
 431         unsigned int xinoshift = 64 - xinobits;
 432
 433         if (unlikely(ino >> xinoshift)) {
 434                 if (warn) {
 435                         pr_warn_ratelimited("d_ino too big (%.*s, ino=%llu, xinobits=%d)\n",
 436                                             namelen, name, ino, xinobits);
 437                 }
 438                 return ino;
 439         }
 440
 441         /*
 442          * The lowest xinobit is reserved for mapping the non-peresistent inode
 443          * numbers range, but this range is only exposed via st_ino, not here.
 444          */
 445         return ino | ((u64)fsid) << (xinoshift + 1);
 446 }
 447
 448 /*
 449  * Set d_ino for upper entries. Non-upper entries should always report
 450  * the uppermost real inode ino and should not call this function.
 451  *
 452  * When not all layer are on same fs, report real ino also for upper.
 453  *
 454  * When all layers are on the same fs, and upper has a reference to
 455  * copy up origin, call vfs_getattr() on the overlay entry to make
 456  * sure that d_ino will be consistent with st_ino from stat(2).
 457  */
 458 static int ovl_cache_update_ino(struct path *path, struct ovl_cache_entry *p)
 459
 460 {
 461         struct dentry *dir = path->dentry;
 462         struct dentry *this = NULL;
 463         enum ovl_path_type type;
 464         u64 ino = p->real_ino;
 465         int xinobits = ovl_xino_bits(dir->d_sb);
 466         int err = 0;
 467
 468         if (!ovl_same_dev(dir->d_sb))
 469                 goto out;
 470
 471         if (p->name[0] == '.') {
 472                 if (p->len == 1) {
 473                         this = dget(dir);
 474                         goto get;
 475                 }
 476                 if (p->len == 2 && p->name[1] == '.') {
 477                         /* we shall not be moved */
 478                         this = dget(dir->d_parent);
 479                         goto get;
 480                 }
 481         }
 482         this = lookup_one_len(p->name, dir, p->len);
 483         if (IS_ERR_OR_NULL(this) || !this->d_inode) {
 484                 /* Mark a stale entry */
 485                 p->is_whiteout = true;
 486                 if (IS_ERR(this)) {
 487                         err = PTR_ERR(this);
 488                         this = NULL;
 489                         goto fail;
 490                 }
 491                 goto out;
 492         }
 493
 494 get:
 495         type = ovl_path_type(this);
 496         if (OVL_TYPE_ORIGIN(type)) {
 497                 struct kstat stat;
 498                 struct path statpath = *path;
 499
 500                 statpath.dentry = this;
 501                 err = vfs_getattr(&statpath, &stat, STATX_INO, 0);
 502                 if (err)
 503                         goto fail;
 504
 505                 /*
 506                  * Directory inode is always on overlay st_dev.
 507                  * Non-dir with ovl_same_dev() could be on pseudo st_dev in case
 508                  * of xino bits overflow.
 509                  */
 510                 WARN_ON_ONCE(S_ISDIR(stat.mode) &&
 511                              dir->d_sb->s_dev != stat.dev);
 512                 ino = stat.ino;
 513         } else if (xinobits && !OVL_TYPE_UPPER(type)) {
 514                 ino = ovl_remap_lower_ino(ino, xinobits,
 515                                           ovl_layer_lower(this)->fsid,
 516                                           p->name, p->len,
 517                                           ovl_xino_warn(dir->d_sb));
 518         }
 519
 520 out:
 521         p->ino = ino;
 522         dput(this);
 523         return err;
 524
 525 fail:
 526         pr_warn_ratelimited("failed to look up (%s) for ino (%i)\n",
 527                             p->name, err);
 528         goto out;
 529 }
 530
 531 static int ovl_fill_plain(struct dir_context *ctx, const char *name,
 532                           int namelen, loff_t offset, u64 ino,
 533                           unsigned int d_type)
 534 {
 535         struct ovl_cache_entry *p;
 536         struct ovl_readdir_data *rdd =
 537                 container_of(ctx, struct ovl_readdir_data, ctx);
 538
 539         rdd->count++;
 540         p = ovl_cache_entry_new(rdd, name, namelen, ino, d_type);
 541         if (p == NULL) {
 542                 rdd->err = -ENOMEM;
 543                 return -ENOMEM;
 544         }
 545         list_add_tail(&p->l_node, rdd->list);
 546
 547         return 0;
 548 }
 549
 550 static int ovl_dir_read_impure(struct path *path,  struct list_head *list,
 551                                struct rb_root *root)
 552 {
 553         int err;
 554         struct path realpath;
 555         struct ovl_cache_entry *p, *n;
 556         struct ovl_readdir_data rdd = {
 557                 .ctx.actor = ovl_fill_plain,
 558                 .list = list,
 559                 .root = root,
 560         };
 561
 562         INIT_LIST_HEAD(list);
 563         *root = RB_ROOT;
 564         ovl_path_upper(path->dentry, &realpath);
 565
 566         err = ovl_dir_read(&realpath, &rdd);
 567         if (err)
 568                 return err;
 569
 570         list_for_each_entry_safe(p, n, list, l_node) {
 571                 if (strcmp(p->name, ".") != 0 &&
 572                     strcmp(p->name, "..") != 0) {
 573                         err = ovl_cache_update_ino(path, p);
 574                         if (err)
 575                                 return err;
 576                 }
 577                 if (p->ino == p->real_ino) {
 578                         list_del(&p->l_node);
 579                         kfree(p);
 580                 } else {
 581                         struct rb_node **newp = &root->rb_node;
 582                         struct rb_node *parent = NULL;
 583
 584                         if (WARN_ON(ovl_cache_entry_find_link(p->name, p->len,
 585                                                               &newp, &parent)))
 586                                 return -EIO;
 587
 588                         rb_link_node(&p->node, parent, newp);
 589                         rb_insert_color(&p->node, root);
 590                 }
 591         }
 592         return 0;
 593 }
 594
 595 static struct ovl_dir_cache *ovl_cache_get_impure(struct path *path)
 596 {
 597         int res;
 598         struct dentry *dentry = path->dentry;
 599         struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 600         struct ovl_dir_cache *cache;
 601
 602         cache = ovl_dir_cache(d_inode(dentry));
 603         if (cache && ovl_dentry_version_get(dentry) == cache->version)
 604                 return cache;
 605
 606         /* Impure cache is not refcounted, free it here */
 607         ovl_dir_cache_free(d_inode(dentry));
 608         ovl_set_dir_cache(d_inode(dentry), NULL);
 609
 610         cache = kzalloc(sizeof(struct ovl_dir_cache), GFP_KERNEL);
 611         if (!cache)
 612                 return ERR_PTR(-ENOMEM);
 613
 614         res = ovl_dir_read_impure(path, &cache->entries, &cache->root);
 615         if (res) {
 616                 ovl_cache_free(&cache->entries);
 617                 kfree(cache);
 618                 return ERR_PTR(res);
 619         }
 620         if (list_empty(&cache->entries)) {
 621                 /*
 622                  * A good opportunity to get rid of an unneeded "impure" flag.
 623                  * Removing the "impure" xattr is best effort.
 624                  */
 625                 if (!ovl_want_write(dentry)) {
 626                         ovl_do_removexattr(ofs, ovl_dentry_upper(dentry),
 627                                            OVL_XATTR_IMPURE);
 628                         ovl_drop_write(dentry);
 629                 }
 630                 ovl_clear_flag(OVL_IMPURE, d_inode(dentry));
 631                 kfree(cache);
 632                 return NULL;
 633         }
 634
 635         cache->version = ovl_dentry_version_get(dentry);
 636         ovl_set_dir_cache(d_inode(dentry), cache);
 637
 638         return cache;
 639 }
 640
 641 struct ovl_readdir_translate {
 642         struct dir_context *orig_ctx;
 643         struct ovl_dir_cache *cache;
 644         struct dir_context ctx;
 645         u64 parent_ino;
 646         int fsid;
 647         int xinobits;
 648         bool xinowarn;
 649 };
 650
 651 static int ovl_fill_real(struct dir_context *ctx, const char *name,
 652                            int namelen, loff_t offset, u64 ino,
 653                            unsigned int d_type)
 654 {
 655         struct ovl_readdir_translate *rdt =
 656                 container_of(ctx, struct ovl_readdir_translate, ctx);
 657         struct dir_context *orig_ctx = rdt->orig_ctx;
 658
 659         if (rdt->parent_ino && strcmp(name, "..") == 0) {
 660                 ino = rdt->parent_ino;
 661         } else if (rdt->cache) {
 662                 struct ovl_cache_entry *p;
 663
 664                 p = ovl_cache_entry_find(&rdt->cache->root, name, namelen);
 665                 if (p)
 666                         ino = p->ino;
 667         } else if (rdt->xinobits) {
 668                 ino = ovl_remap_lower_ino(ino, rdt->xinobits, rdt->fsid,
 669                                           name, namelen, rdt->xinowarn);
 670         }
 671
 672         return orig_ctx->actor(orig_ctx, name, namelen, offset, ino, d_type);
 673 }
 674
 675 static bool ovl_is_impure_dir(struct file *file)
 676 {
 677         struct ovl_dir_file *od = file->private_data;
 678         struct inode *dir = d_inode(file->f_path.dentry);
 679
 680         /*
 681          * Only upper dir can be impure, but if we are in the middle of
 682          * iterating a lower real dir, dir could be copied up and marked
 683          * impure. We only want the impure cache if we started iterating
 684          * a real upper dir to begin with.
 685          */
 686         return od->is_upper && ovl_test_flag(OVL_IMPURE, dir);
 687
 688 }
 689
 690 static int ovl_iterate_real(struct file *file, struct dir_context *ctx)
 691 {
 692         int err;
 693         struct ovl_dir_file *od = file->private_data;
 694         struct dentry *dir = file->f_path.dentry;
 695         const struct ovl_layer *lower_layer = ovl_layer_lower(dir);
 696         struct ovl_readdir_translate rdt = {
 697                 .ctx.actor = ovl_fill_real,
 698                 .orig_ctx = ctx,
 699                 .xinobits = ovl_xino_bits(dir->d_sb),
 700                 .xinowarn = ovl_xino_warn(dir->d_sb),
 701         };
 702
 703         if (rdt.xinobits && lower_layer)
 704                 rdt.fsid = lower_layer->fsid;
 705
 706         if (OVL_TYPE_MERGE(ovl_path_type(dir->d_parent))) {
 707                 struct kstat stat;
 708                 struct path statpath = file->f_path;
 709
 710                 statpath.dentry = dir->d_parent;
 711                 err = vfs_getattr(&statpath, &stat, STATX_INO, 0);
 712                 if (err)
 713                         return err;
 714
 715                 WARN_ON_ONCE(dir->d_sb->s_dev != stat.dev);
 716                 rdt.parent_ino = stat.ino;
 717         }
 718
 719         if (ovl_is_impure_dir(file)) {
 720                 rdt.cache = ovl_cache_get_impure(&file->f_path);
 721                 if (IS_ERR(rdt.cache))
 722                         return PTR_ERR(rdt.cache);
 723         }
 724
 725         err = iterate_dir(od->realfile, &rdt.ctx);
 726         ctx->pos = rdt.ctx.pos;
 727
 728         return err;
 729 }
 730
 731
 732 static int ovl_iterate(struct file *file, struct dir_context *ctx)
 733 {
 734         struct ovl_dir_file *od = file->private_data;
 735         struct dentry *dentry = file->f_path.dentry;
 736         struct ovl_cache_entry *p;
 737         const struct cred *old_cred;
 738         int err;
 739
 740         old_cred = ovl_override_creds(dentry->d_sb);
 741         if (!ctx->pos)
 742                 ovl_dir_reset(file);
 743
 744         if (od->is_real) {
 745                 /*
 746                  * If parent is merge, then need to adjust d_ino for '..', if
 747                  * dir is impure then need to adjust d_ino for copied up
 748                  * entries.
 749                  */
 750                 if (ovl_xino_bits(dentry->d_sb) ||
 751                     (ovl_same_fs(dentry->d_sb) &&
 752                      (ovl_is_impure_dir(file) ||
 753                       OVL_TYPE_MERGE(ovl_path_type(dentry->d_parent))))) {
 754                         err = ovl_iterate_real(file, ctx);
 755                 } else {
 756                         err = iterate_dir(od->realfile, ctx);
 757                 }
 758                 goto out;
 759         }
 760
 761         if (!od->cache) {
 762                 struct ovl_dir_cache *cache;
 763
 764                 cache = ovl_cache_get(dentry);
 765                 err = PTR_ERR(cache);
 766                 if (IS_ERR(cache))
 767                         goto out;
 768
 769                 od->cache = cache;
 770                 ovl_seek_cursor(od, ctx->pos);
 771         }
 772
 773         while (od->cursor != &od->cache->entries) {
 774                 p = list_entry(od->cursor, struct ovl_cache_entry, l_node);
 775                 if (!p->is_whiteout) {
 776                         if (!p->ino) {
 777                                 err = ovl_cache_update_ino(&file->f_path, p);
 778                                 if (err)
 779                                         goto out;
 780                         }
 781                 }
 782                 /* ovl_cache_update_ino() sets is_whiteout on stale entry */
 783                 if (!p->is_whiteout) {
 784                         if (!dir_emit(ctx, p->name, p->len, p->ino, p->type))
 785                                 break;
 786                 }
 787                 od->cursor = p->l_node.next;
 788                 ctx->pos++;
 789         }
 790         err = 0;
 791 out:
 792         revert_creds(old_cred);
 793         return err;
 794 }
 795
 796 static loff_t ovl_dir_llseek(struct file *file, loff_t offset, int origin)
 797 {
 798         loff_t res;
 799         struct ovl_dir_file *od = file->private_data;
 800
 801         inode_lock(file_inode(file));
 802         if (!file->f_pos)
 803                 ovl_dir_reset(file);
 804
 805         if (od->is_real) {
 806                 res = vfs_llseek(od->realfile, offset, origin);
 807                 file->f_pos = od->realfile->f_pos;
 808         } else {
 809                 res = -EINVAL;
 810
 811                 switch (origin) {
 812                 case SEEK_CUR:
 813                         offset += file->f_pos;
 814                         break;
 815                 case SEEK_SET:
 816                         break;
 817                 default:
 818                         goto out_unlock;
 819                 }
 820                 if (offset < 0)
 821                         goto out_unlock;
 822
 823                 if (offset != file->f_pos) {
 824                         file->f_pos = offset;
 825                         if (od->cache)
 826                                 ovl_seek_cursor(od, offset);
 827                 }
 828                 res = offset;
 829         }
 830 out_unlock:
 831         inode_unlock(file_inode(file));
 832
 833         return res;
 834 }
 835
 836 static struct file *ovl_dir_open_realfile(const struct file *file,
 837                                           struct path *realpath)
 838 {
 839         struct file *res;
 840         const struct cred *old_cred;
 841
 842         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 843         res = ovl_path_open(realpath, O_RDONLY | (file->f_flags & O_LARGEFILE));
 844         revert_creds(old_cred);
 845
 846         return res;
 847 }
 848
 849 /*
 850  * Like ovl_real_fdget(), returns upperfile if dir was copied up since open.
 851  * Unlike ovl_real_fdget(), this caches upperfile in file->private_data.
 852  *
 853  * TODO: use same abstract type for file->private_data of dir and file so
 854  * upperfile could also be cached for files as well.
 855  */
 856 struct file *ovl_dir_real_file(const struct file *file, bool want_upper)
 857 {
 858
 859         struct ovl_dir_file *od = file->private_data;
 860         struct dentry *dentry = file->f_path.dentry;
 861         struct file *old, *realfile = od->realfile;
 862
 863         if (!OVL_TYPE_UPPER(ovl_path_type(dentry)))
 864                 return want_upper ? NULL : realfile;
 865
 866         /*
 867          * Need to check if we started out being a lower dir, but got copied up
 868          */
 869         if (!od->is_upper) {
 870                 realfile = READ_ONCE(od->upperfile);
 871                 if (!realfile) {
 872                         struct path upperpath;
 873
 874                         ovl_path_upper(dentry, &upperpath);
 875                         realfile = ovl_dir_open_realfile(file, &upperpath);
 876                         if (IS_ERR(realfile))
 877                                 return realfile;
 878
 879                         old = cmpxchg_release(&od->upperfile, NULL, realfile);
 880                         if (old) {
 881                                 fput(realfile);
 882                                 realfile = old;
 883                         }
 884                 }
 885         }
 886
 887         return realfile;
 888 }
 889
 890 static int ovl_dir_fsync(struct file *file, loff_t start, loff_t end,
 891                          int datasync)
 892 {
 893         struct file *realfile;
 894         int err;
 895
 896         err = ovl_sync_status(OVL_FS(file->f_path.dentry->d_sb));
 897         if (err <= 0)
 898                 return err;
 899
 900         realfile = ovl_dir_real_file(file, true);
 901         err = PTR_ERR_OR_ZERO(realfile);
 902
 903         /* Nothing to sync for lower */
 904         if (!realfile || err)
 905                 return err;
 906
 907         return vfs_fsync_range(realfile, start, end, datasync);
 908 }
 909
 910 static int ovl_dir_release(struct inode *inode, struct file *file)
 911 {
 912         struct ovl_dir_file *od = file->private_data;
 913
 914         if (od->cache) {
 915                 inode_lock(inode);
 916                 ovl_cache_put(od, file->f_path.dentry);
 917                 inode_unlock(inode);
 918         }
 919         fput(od->realfile);
 920         if (od->upperfile)
 921                 fput(od->upperfile);
 922         kfree(od);
 923
 924         return 0;
 925 }
 926
 927 static int ovl_dir_open(struct inode *inode, struct file *file)
 928 {
 929         struct path realpath;
 930         struct file *realfile;
 931         struct ovl_dir_file *od;
 932         enum ovl_path_type type;
 933
 934         od = kzalloc(sizeof(struct ovl_dir_file), GFP_KERNEL);
 935         if (!od)
 936                 return -ENOMEM;
 937
 938         type = ovl_path_real(file->f_path.dentry, &realpath);
 939         realfile = ovl_dir_open_realfile(file, &realpath);
 940         if (IS_ERR(realfile)) {
 941                 kfree(od);
 942                 return PTR_ERR(realfile);
 943         }
 944         od->realfile = realfile;
 945         od->is_real = ovl_dir_is_real(file->f_path.dentry);
 946         od->is_upper = OVL_TYPE_UPPER(type);
 947         file->private_data = od;
 948
 949         return 0;
 950 }
 951
 952 const struct file_operations ovl_dir_operations = {
 953         .read           = generic_read_dir,
 954         .open           = ovl_dir_open,
 955         .iterate        = ovl_iterate,
 956         .llseek         = ovl_dir_llseek,
 957         .fsync          = ovl_dir_fsync,
 958         .release        = ovl_dir_release,
 959 };
 960
 961 int ovl_check_empty_dir(struct dentry *dentry, struct list_head *list)
 962 {
 963         int err;
 964         struct ovl_cache_entry *p, *n;
 965         struct rb_root root = RB_ROOT;
 966         const struct cred *old_cred;
 967
 968         old_cred = ovl_override_creds(dentry->d_sb);
 969         err = ovl_dir_read_merged(dentry, list, &root);
 970         revert_creds(old_cred);
 971         if (err)
 972                 return err;
 973
 974         err = 0;
 975
 976         list_for_each_entry_safe(p, n, list, l_node) {
 977                 /*
 978                  * Select whiteouts in upperdir, they should
 979                  * be cleared when deleting this directory.
 980                  */
 981                 if (p->is_whiteout) {
 982                         if (p->is_upper)
 983                                 continue;
 984                         goto del_entry;
 985                 }
 986
 987                 if (p->name[0] == '.') {
 988                         if (p->len == 1)
 989                                 goto del_entry;
 990                         if (p->len == 2 && p->name[1] == '.')
 991                                 goto del_entry;
 992                 }
 993                 err = -ENOTEMPTY;
 994                 break;
 995
 996 del_entry:
 997                 list_del(&p->l_node);
 998                 kfree(p);
 999         }
1000
1001         return err;
1002 }
1003
1004 void ovl_cleanup_whiteouts(struct dentry *upper, struct list_head *list)
1005 {
1006         struct ovl_cache_entry *p;
1007
1008         inode_lock_nested(upper->d_inode, I_MUTEX_CHILD);
1009         list_for_each_entry(p, list, l_node) {
1010                 struct dentry *dentry;
1011
1012                 if (WARN_ON(!p->is_whiteout || !p->is_upper))
1013                         continue;
1014
1015                 dentry = lookup_one_len(p->name, upper, p->len);
1016                 if (IS_ERR(dentry)) {
1017                         pr_err("lookup '%s/%.*s' failed (%i)\n",
1018                                upper->d_name.name, p->len, p->name,
1019                                (int) PTR_ERR(dentry));
1020                         continue;
1021                 }
1022                 if (dentry->d_inode)
1023                         ovl_cleanup(upper->d_inode, dentry);
1024                 dput(dentry);
1025         }
1026         inode_unlock(upper->d_inode);
1027 }
1028
1029 static int ovl_check_d_type(struct dir_context *ctx, const char *name,
1030                           int namelen, loff_t offset, u64 ino,
1031                           unsigned int d_type)
1032 {
1033         struct ovl_readdir_data *rdd =
1034                 container_of(ctx, struct ovl_readdir_data, ctx);
1035
1036         /* Even if d_type is not supported, DT_DIR is returned for . and .. */
1037         if (!strncmp(name, ".", namelen) || !strncmp(name, "..", namelen))
1038                 return 0;
1039
1040         if (d_type != DT_UNKNOWN)
1041                 rdd->d_type_supported = true;
1042
1043         return 0;
1044 }
1045
1046 /*
1047  * Returns 1 if d_type is supported, 0 not supported/unknown. Negative values
1048  * if error is encountered.
1049  */
1050 int ovl_check_d_type_supported(struct path *realpath)
1051 {
1052         int err;
1053         struct ovl_readdir_data rdd = {
1054                 .ctx.actor = ovl_check_d_type,
1055                 .d_type_supported = false,
1056         };
1057
1058         err = ovl_dir_read(realpath, &rdd);
1059         if (err)
1060                 return err;
1061
1062         return rdd.d_type_supported;
1063 }
1064
1065 #define OVL_INCOMPATDIR_NAME "incompat"
1066
1067 static int ovl_workdir_cleanup_recurse(struct path *path, int level)
1068 {
1069         int err;
1070         struct inode *dir = path->dentry->d_inode;
1071         LIST_HEAD(list);
1072         struct rb_root root = RB_ROOT;
1073         struct ovl_cache_entry *p;
1074         struct ovl_readdir_data rdd = {
1075                 .ctx.actor = ovl_fill_merge,
1076                 .dentry = NULL,
1077                 .list = &list,
1078                 .root = &root,
1079                 .is_lowest = false,
1080         };
1081         bool incompat = false;
1082
1083         /*
1084          * The "work/incompat" directory is treated specially - if it is not
1085          * empty, instead of printing a generic error and mounting read-only,
1086          * we will error about incompat features and fail the mount.
1087          *
1088          * When called from ovl_indexdir_cleanup(), path->dentry->d_name.name
1089          * starts with '#'.
1090          */
1091         if (level == 2 &&
1092             !strcmp(path->dentry->d_name.name, OVL_INCOMPATDIR_NAME))
1093                 incompat = true;
1094
1095         err = ovl_dir_read(path, &rdd);
1096         if (err)
1097                 goto out;
1098
1099         inode_lock_nested(dir, I_MUTEX_PARENT);
1100         list_for_each_entry(p, &list, l_node) {
1101                 struct dentry *dentry;
1102
1103                 if (p->name[0] == '.') {
1104                         if (p->len == 1)
1105                                 continue;
1106                         if (p->len == 2 && p->name[1] == '.')
1107                                 continue;
1108                 } else if (incompat) {
1109                         pr_err("overlay with incompat feature '%s' cannot be mounted\n",
1110                                 p->name);
1111                         err = -EINVAL;
1112                         break;
1113                 }
1114                 dentry = lookup_one_len(p->name, path->dentry, p->len);
1115                 if (IS_ERR(dentry))
1116                         continue;
1117                 if (dentry->d_inode)
1118                         err = ovl_workdir_cleanup(dir, path->mnt, dentry, level);
1119                 dput(dentry);
1120                 if (err)
1121                         break;
1122         }
1123         inode_unlock(dir);
1124 out:
1125         ovl_cache_free(&list);
1126         return err;
1127 }
1128
1129 int ovl_workdir_cleanup(struct inode *dir, struct vfsmount *mnt,
1130                          struct dentry *dentry, int level)
1131 {
1132         int err;
1133
1134         if (!d_is_dir(dentry) || level > 1) {
1135                 return ovl_cleanup(dir, dentry);
1136         }
1137
1138         err = ovl_do_rmdir(dir, dentry);
1139         if (err) {
1140                 struct path path = { .mnt = mnt, .dentry = dentry };
1141
1142                 inode_unlock(dir);
1143                 err = ovl_workdir_cleanup_recurse(&path, level + 1);
1144                 inode_lock_nested(dir, I_MUTEX_PARENT);
1145                 if (!err)
1146                         err = ovl_cleanup(dir, dentry);
1147         }
1148
1149         return err;
1150 }
1151
1152 int ovl_indexdir_cleanup(struct ovl_fs *ofs)
1153 {
1154         int err;
1155         struct dentry *indexdir = ofs->indexdir;
1156         struct dentry *index = NULL;
1157         struct inode *dir = indexdir->d_inode;
1158         struct path path = { .mnt = ovl_upper_mnt(ofs), .dentry = indexdir };
1159         LIST_HEAD(list);
1160         struct rb_root root = RB_ROOT;
1161         struct ovl_cache_entry *p;
1162         struct ovl_readdir_data rdd = {
1163                 .ctx.actor = ovl_fill_merge,
1164                 .dentry = NULL,
1165                 .list = &list,
1166                 .root = &root,
1167                 .is_lowest = false,
1168         };
1169
1170         err = ovl_dir_read(&path, &rdd);
1171         if (err)
1172                 goto out;
1173
1174         inode_lock_nested(dir, I_MUTEX_PARENT);
1175         list_for_each_entry(p, &list, l_node) {
1176                 if (p->name[0] == '.') {
1177                         if (p->len == 1)
1178                                 continue;
1179                         if (p->len == 2 && p->name[1] == '.')
1180                                 continue;
1181                 }
1182                 index = lookup_one_len(p->name, indexdir, p->len);
1183                 if (IS_ERR(index)) {
1184                         err = PTR_ERR(index);
1185                         index = NULL;
1186                         break;
1187                 }
1188                 /* Cleanup leftover from index create/cleanup attempt */
1189                 if (index->d_name.name[0] == '#') {
1190                         err = ovl_workdir_cleanup(dir, path.mnt, index, 1);
1191                         if (err)
1192                                 break;
1193                         goto next;
1194                 }
1195                 err = ovl_verify_index(ofs, index);
1196                 if (!err) {
1197                         goto next;
1198                 } else if (err == -ESTALE) {
1199                         /* Cleanup stale index entries */
1200                         err = ovl_cleanup(dir, index);
1201                 } else if (err != -ENOENT) {
1202                         /*
1203                          * Abort mount to avoid corrupting the index if
1204                          * an incompatible index entry was found or on out
1205                          * of memory.
1206                          */
1207                         break;
1208                 } else if (ofs->config.nfs_export) {
1209                         /*
1210                          * Whiteout orphan index to block future open by
1211                          * handle after overlay nlink dropped to zero.
1212                          */
1213                         err = ovl_cleanup_and_whiteout(ofs, dir, index);
1214                 } else {
1215                         /* Cleanup orphan index entries */
1216                         err = ovl_cleanup(dir, index);
1217                 }
1218
1219                 if (err)
1220                         break;
1221
1222 next:
1223                 dput(index);
1224                 index = NULL;
1225         }
1226         dput(index);
1227         inode_unlock(dir);
1228 out:
1229         ovl_cache_free(&list);
1230         if (err)
1231                 pr_err("failed index dir cleanup (%i)\n", err);
1232         return err;
1233 }