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         bool check_xwhiteout;
  29         char name[];
  30 };
  31
  32 struct ovl_dir_cache {
  33         long refcount;
  34         u64 version;
  35         struct list_head entries;
  36         struct rb_root root;
  37 };
  38
  39 struct ovl_readdir_data {
  40         struct dir_context ctx;
  41         struct dentry *dentry;
  42         bool is_lowest;
  43         struct rb_root *root;
  44         struct list_head *list;
  45         struct list_head middle;
  46         struct ovl_cache_entry *first_maybe_whiteout;
  47         int count;
  48         int err;
  49         bool is_upper;
  50         bool d_type_supported;
  51         bool in_xwhiteouts_dir;
  52 };
  53
  54 struct ovl_dir_file {
  55         bool is_real;
  56         bool is_upper;
  57         struct ovl_dir_cache *cache;
  58         struct list_head *cursor;
  59         struct file *realfile;
  60         struct file *upperfile;
  61 };
  62
  63 static struct ovl_cache_entry *ovl_cache_entry_from_node(struct rb_node *n)
  64 {
  65         return rb_entry(n, struct ovl_cache_entry, node);
  66 }
  67
  68 static bool ovl_cache_entry_find_link(const char *name, int len,
  69                                       struct rb_node ***link,
  70                                       struct rb_node **parent)
  71 {
  72         bool found = false;
  73         struct rb_node **newp = *link;
  74
  75         while (!found && *newp) {
  76                 int cmp;
  77                 struct ovl_cache_entry *tmp;
  78
  79                 *parent = *newp;
  80                 tmp = ovl_cache_entry_from_node(*newp);
  81                 cmp = strncmp(name, tmp->name, len);
  82                 if (cmp > 0)
  83                         newp = &tmp->node.rb_right;
  84                 else if (cmp < 0 || len < tmp->len)
  85                         newp = &tmp->node.rb_left;
  86                 else
  87                         found = true;
  88         }
  89         *link = newp;
  90
  91         return found;
  92 }
  93
  94 static struct ovl_cache_entry *ovl_cache_entry_find(struct rb_root *root,
  95                                                     const char *name, int len)
  96 {
  97         struct rb_node *node = root->rb_node;
  98         int cmp;
  99
 100         while (node) {
 101                 struct ovl_cache_entry *p = ovl_cache_entry_from_node(node);
 102
 103                 cmp = strncmp(name, p->name, len);
 104                 if (cmp > 0)
 105                         node = p->node.rb_right;
 106                 else if (cmp < 0 || len < p->len)
 107                         node = p->node.rb_left;
 108                 else
 109                         return p;
 110         }
 111
 112         return NULL;
 113 }
 114
 115 static bool ovl_calc_d_ino(struct ovl_readdir_data *rdd,
 116                            struct ovl_cache_entry *p)
 117 {
 118         /* Don't care if not doing ovl_iter() */
 119         if (!rdd->dentry)
 120                 return false;
 121
 122         /* Always recalc d_ino when remapping lower inode numbers */
 123         if (ovl_xino_bits(OVL_FS(rdd->dentry->d_sb)))
 124                 return true;
 125
 126         /* Always recalc d_ino for parent */
 127         if (strcmp(p->name, "..") == 0)
 128                 return true;
 129
 130         /* If this is lower, then native d_ino will do */
 131         if (!rdd->is_upper)
 132                 return false;
 133
 134         /*
 135          * Recalc d_ino for '.' and for all entries if dir is impure (contains
 136          * copied up entries)
 137          */
 138         if ((p->name[0] == '.' && p->len == 1) ||
 139             ovl_test_flag(OVL_IMPURE, d_inode(rdd->dentry)))
 140                 return true;
 141
 142         return false;
 143 }
 144
 145 static struct ovl_cache_entry *ovl_cache_entry_new(struct ovl_readdir_data *rdd,
 146                                                    const char *name, int len,
 147                                                    u64 ino, unsigned int d_type)
 148 {
 149         struct ovl_cache_entry *p;
 150         size_t size = offsetof(struct ovl_cache_entry, name[len + 1]);
 151
 152         p = kmalloc(size, GFP_KERNEL);
 153         if (!p)
 154                 return NULL;
 155
 156         memcpy(p->name, name, len);
 157         p->name[len] = '\0';
 158         p->len = len;
 159         p->type = d_type;
 160         p->real_ino = ino;
 161         p->ino = ino;
 162         /* Defer setting d_ino for upper entry to ovl_iterate() */
 163         if (ovl_calc_d_ino(rdd, p))
 164                 p->ino = 0;
 165         p->is_upper = rdd->is_upper;
 166         p->is_whiteout = false;
 167         /* Defer check for overlay.whiteout to ovl_iterate() */
 168         p->check_xwhiteout = rdd->in_xwhiteouts_dir && d_type == DT_REG;
 169
 170         if (d_type == DT_CHR) {
 171                 p->next_maybe_whiteout = rdd->first_maybe_whiteout;
 172                 rdd->first_maybe_whiteout = p;
 173         }
 174         return p;
 175 }
 176
 177 static bool ovl_cache_entry_add_rb(struct ovl_readdir_data *rdd,
 178                                   const char *name, int len, u64 ino,
 179                                   unsigned int d_type)
 180 {
 181         struct rb_node **newp = &rdd->root->rb_node;
 182         struct rb_node *parent = NULL;
 183         struct ovl_cache_entry *p;
 184
 185         if (ovl_cache_entry_find_link(name, len, &newp, &parent))
 186                 return true;
 187
 188         p = ovl_cache_entry_new(rdd, name, len, ino, d_type);
 189         if (p == NULL) {
 190                 rdd->err = -ENOMEM;
 191                 return false;
 192         }
 193
 194         list_add_tail(&p->l_node, rdd->list);
 195         rb_link_node(&p->node, parent, newp);
 196         rb_insert_color(&p->node, rdd->root);
 197
 198         return true;
 199 }
 200
 201 static bool ovl_fill_lowest(struct ovl_readdir_data *rdd,
 202                            const char *name, int namelen,
 203                            loff_t offset, u64 ino, unsigned int d_type)
 204 {
 205         struct ovl_cache_entry *p;
 206
 207         p = ovl_cache_entry_find(rdd->root, name, namelen);
 208         if (p) {
 209                 list_move_tail(&p->l_node, &rdd->middle);
 210         } else {
 211                 p = ovl_cache_entry_new(rdd, name, namelen, ino, d_type);
 212                 if (p == NULL)
 213                         rdd->err = -ENOMEM;
 214                 else
 215                         list_add_tail(&p->l_node, &rdd->middle);
 216         }
 217
 218         return rdd->err == 0;
 219 }
 220
 221 void ovl_cache_free(struct list_head *list)
 222 {
 223         struct ovl_cache_entry *p;
 224         struct ovl_cache_entry *n;
 225
 226         list_for_each_entry_safe(p, n, list, l_node)
 227                 kfree(p);
 228
 229         INIT_LIST_HEAD(list);
 230 }
 231
 232 void ovl_dir_cache_free(struct inode *inode)
 233 {
 234         struct ovl_dir_cache *cache = ovl_dir_cache(inode);
 235
 236         if (cache) {
 237                 ovl_cache_free(&cache->entries);
 238                 kfree(cache);
 239         }
 240 }
 241
 242 static void ovl_cache_put(struct ovl_dir_file *od, struct inode *inode)
 243 {
 244         struct ovl_dir_cache *cache = od->cache;
 245
 246         WARN_ON(cache->refcount <= 0);
 247         cache->refcount--;
 248         if (!cache->refcount) {
 249                 if (ovl_dir_cache(inode) == cache)
 250                         ovl_set_dir_cache(inode, NULL);
 251
 252                 ovl_cache_free(&cache->entries);
 253                 kfree(cache);
 254         }
 255 }
 256
 257 static bool ovl_fill_merge(struct dir_context *ctx, const char *name,
 258                           int namelen, loff_t offset, u64 ino,
 259                           unsigned int d_type)
 260 {
 261         struct ovl_readdir_data *rdd =
 262                 container_of(ctx, struct ovl_readdir_data, ctx);
 263
 264         rdd->count++;
 265         if (!rdd->is_lowest)
 266                 return ovl_cache_entry_add_rb(rdd, name, namelen, ino, d_type);
 267         else
 268                 return ovl_fill_lowest(rdd, name, namelen, offset, ino, d_type);
 269 }
 270
 271 static int ovl_check_whiteouts(const struct path *path, struct ovl_readdir_data *rdd)
 272 {
 273         int err;
 274         struct ovl_cache_entry *p;
 275         struct dentry *dentry, *dir = path->dentry;
 276         const struct cred *old_cred;
 277
 278         old_cred = ovl_override_creds(rdd->dentry->d_sb);
 279
 280         err = down_write_killable(&dir->d_inode->i_rwsem);
 281         if (!err) {
 282                 while (rdd->first_maybe_whiteout) {
 283                         p = rdd->first_maybe_whiteout;
 284                         rdd->first_maybe_whiteout = p->next_maybe_whiteout;
 285                         dentry = lookup_one(mnt_idmap(path->mnt), p->name, dir, p->len);
 286                         if (!IS_ERR(dentry)) {
 287                                 p->is_whiteout = ovl_is_whiteout(dentry);
 288                                 dput(dentry);
 289                         }
 290                 }
 291                 inode_unlock(dir->d_inode);
 292         }
 293         revert_creds(old_cred);
 294
 295         return err;
 296 }
 297
 298 static inline int ovl_dir_read(const struct path *realpath,
 299                                struct ovl_readdir_data *rdd)
 300 {
 301         struct file *realfile;
 302         int err;
 303
 304         realfile = ovl_path_open(realpath, O_RDONLY | O_LARGEFILE);
 305         if (IS_ERR(realfile))
 306                 return PTR_ERR(realfile);
 307
 308         rdd->in_xwhiteouts_dir = rdd->dentry &&
 309                 ovl_path_check_xwhiteouts_xattr(OVL_FS(rdd->dentry->d_sb), realpath);
 310         rdd->first_maybe_whiteout = NULL;
 311         rdd->ctx.pos = 0;
 312         do {
 313                 rdd->count = 0;
 314                 rdd->err = 0;
 315                 err = iterate_dir(realfile, &rdd->ctx);
 316                 if (err >= 0)
 317                         err = rdd->err;
 318         } while (!err && rdd->count);
 319
 320         if (!err && rdd->first_maybe_whiteout && rdd->dentry)
 321                 err = ovl_check_whiteouts(realpath, rdd);
 322
 323         fput(realfile);
 324
 325         return err;
 326 }
 327
 328 static void ovl_dir_reset(struct file *file)
 329 {
 330         struct ovl_dir_file *od = file->private_data;
 331         struct ovl_dir_cache *cache = od->cache;
 332         struct inode *inode = file_inode(file);
 333         bool is_real;
 334
 335         if (cache && ovl_inode_version_get(inode) != cache->version) {
 336                 ovl_cache_put(od, inode);
 337                 od->cache = NULL;
 338                 od->cursor = NULL;
 339         }
 340         is_real = ovl_dir_is_real(inode);
 341         if (od->is_real != is_real) {
 342                 /* is_real can only become false when dir is copied up */
 343                 if (WARN_ON(is_real))
 344                         return;
 345                 od->is_real = false;
 346         }
 347 }
 348
 349 static int ovl_dir_read_merged(struct dentry *dentry, struct list_head *list,
 350         struct rb_root *root)
 351 {
 352         int err;
 353         struct path realpath;
 354         struct ovl_readdir_data rdd = {
 355                 .ctx.actor = ovl_fill_merge,
 356                 .dentry = dentry,
 357                 .list = list,
 358                 .root = root,
 359                 .is_lowest = false,
 360         };
 361         int idx, next;
 362
 363         for (idx = 0; idx != -1; idx = next) {
 364                 next = ovl_path_next(idx, dentry, &realpath);
 365                 rdd.is_upper = ovl_dentry_upper(dentry) == realpath.dentry;
 366
 367                 if (next != -1) {
 368                         err = ovl_dir_read(&realpath, &rdd);
 369                         if (err)
 370                                 break;
 371                 } else {
 372                         /*
 373                          * Insert lowest layer entries before upper ones, this
 374                          * allows offsets to be reasonably constant
 375                          */
 376                         list_add(&rdd.middle, rdd.list);
 377                         rdd.is_lowest = true;
 378                         err = ovl_dir_read(&realpath, &rdd);
 379                         list_del(&rdd.middle);
 380                 }
 381         }
 382         return err;
 383 }
 384
 385 static void ovl_seek_cursor(struct ovl_dir_file *od, loff_t pos)
 386 {
 387         struct list_head *p;
 388         loff_t off = 0;
 389
 390         list_for_each(p, &od->cache->entries) {
 391                 if (off >= pos)
 392                         break;
 393                 off++;
 394         }
 395         /* Cursor is safe since the cache is stable */
 396         od->cursor = p;
 397 }
 398
 399 static struct ovl_dir_cache *ovl_cache_get(struct dentry *dentry)
 400 {
 401         int res;
 402         struct ovl_dir_cache *cache;
 403         struct inode *inode = d_inode(dentry);
 404
 405         cache = ovl_dir_cache(inode);
 406         if (cache && ovl_inode_version_get(inode) == cache->version) {
 407                 WARN_ON(!cache->refcount);
 408                 cache->refcount++;
 409                 return cache;
 410         }
 411         ovl_set_dir_cache(d_inode(dentry), NULL);
 412
 413         cache = kzalloc(sizeof(struct ovl_dir_cache), GFP_KERNEL);
 414         if (!cache)
 415                 return ERR_PTR(-ENOMEM);
 416
 417         cache->refcount = 1;
 418         INIT_LIST_HEAD(&cache->entries);
 419         cache->root = RB_ROOT;
 420
 421         res = ovl_dir_read_merged(dentry, &cache->entries, &cache->root);
 422         if (res) {
 423                 ovl_cache_free(&cache->entries);
 424                 kfree(cache);
 425                 return ERR_PTR(res);
 426         }
 427
 428         cache->version = ovl_inode_version_get(inode);
 429         ovl_set_dir_cache(inode, cache);
 430
 431         return cache;
 432 }
 433
 434 /* Map inode number to lower fs unique range */
 435 static u64 ovl_remap_lower_ino(u64 ino, int xinobits, int fsid,
 436                                const char *name, int namelen, bool warn)
 437 {
 438         unsigned int xinoshift = 64 - xinobits;
 439
 440         if (unlikely(ino >> xinoshift)) {
 441                 if (warn) {
 442                         pr_warn_ratelimited("d_ino too big (%.*s, ino=%llu, xinobits=%d)\n",
 443                                             namelen, name, ino, xinobits);
 444                 }
 445                 return ino;
 446         }
 447
 448         /*
 449          * The lowest xinobit is reserved for mapping the non-peresistent inode
 450          * numbers range, but this range is only exposed via st_ino, not here.
 451          */
 452         return ino | ((u64)fsid) << (xinoshift + 1);
 453 }
 454
 455 /*
 456  * Set d_ino for upper entries if needed. Non-upper entries should always report
 457  * the uppermost real inode ino and should not call this function.
 458  *
 459  * When not all layer are on same fs, report real ino also for upper.
 460  *
 461  * When all layers are on the same fs, and upper has a reference to
 462  * copy up origin, call vfs_getattr() on the overlay entry to make
 463  * sure that d_ino will be consistent with st_ino from stat(2).
 464  *
 465  * Also checks the overlay.whiteout xattr by doing a full lookup which will return
 466  * negative in this case.
 467  */
 468 static int ovl_cache_update(const struct path *path, struct ovl_cache_entry *p, bool update_ino)
 469
 470 {
 471         struct dentry *dir = path->dentry;
 472         struct ovl_fs *ofs = OVL_FS(dir->d_sb);
 473         struct dentry *this = NULL;
 474         enum ovl_path_type type;
 475         u64 ino = p->real_ino;
 476         int xinobits = ovl_xino_bits(ofs);
 477         int err = 0;
 478
 479         if (!ovl_same_dev(ofs) && !p->check_xwhiteout)
 480                 goto out;
 481
 482         if (p->name[0] == '.') {
 483                 if (p->len == 1) {
 484                         this = dget(dir);
 485                         goto get;
 486                 }
 487                 if (p->len == 2 && p->name[1] == '.') {
 488                         /* we shall not be moved */
 489                         this = dget(dir->d_parent);
 490                         goto get;
 491                 }
 492         }
 493         /* This checks also for xwhiteouts */
 494         this = lookup_one(mnt_idmap(path->mnt), p->name, dir, p->len);
 495         if (IS_ERR_OR_NULL(this) || !this->d_inode) {
 496                 /* Mark a stale entry */
 497                 p->is_whiteout = true;
 498                 if (IS_ERR(this)) {
 499                         err = PTR_ERR(this);
 500                         this = NULL;
 501                         goto fail;
 502                 }
 503                 goto out;
 504         }
 505
 506 get:
 507         if (!ovl_same_dev(ofs) || !update_ino)
 508                 goto out;
 509
 510         type = ovl_path_type(this);
 511         if (OVL_TYPE_ORIGIN(type)) {
 512                 struct kstat stat;
 513                 struct path statpath = *path;
 514
 515                 statpath.dentry = this;
 516                 err = vfs_getattr(&statpath, &stat, STATX_INO, 0);
 517                 if (err)
 518                         goto fail;
 519
 520                 /*
 521                  * Directory inode is always on overlay st_dev.
 522                  * Non-dir with ovl_same_dev() could be on pseudo st_dev in case
 523                  * of xino bits overflow.
 524                  */
 525                 WARN_ON_ONCE(S_ISDIR(stat.mode) &&
 526                              dir->d_sb->s_dev != stat.dev);
 527                 ino = stat.ino;
 528         } else if (xinobits && !OVL_TYPE_UPPER(type)) {
 529                 ino = ovl_remap_lower_ino(ino, xinobits,
 530                                           ovl_layer_lower(this)->fsid,
 531                                           p->name, p->len,
 532                                           ovl_xino_warn(ofs));
 533         }
 534
 535 out:
 536         p->ino = ino;
 537         dput(this);
 538         return err;
 539
 540 fail:
 541         pr_warn_ratelimited("failed to look up (%s) for ino (%i)\n",
 542                             p->name, err);
 543         goto out;
 544 }
 545
 546 static bool ovl_fill_plain(struct dir_context *ctx, const char *name,
 547                           int namelen, loff_t offset, u64 ino,
 548                           unsigned int d_type)
 549 {
 550         struct ovl_cache_entry *p;
 551         struct ovl_readdir_data *rdd =
 552                 container_of(ctx, struct ovl_readdir_data, ctx);
 553
 554         rdd->count++;
 555         p = ovl_cache_entry_new(rdd, name, namelen, ino, d_type);
 556         if (p == NULL) {
 557                 rdd->err = -ENOMEM;
 558                 return false;
 559         }
 560         list_add_tail(&p->l_node, rdd->list);
 561
 562         return true;
 563 }
 564
 565 static int ovl_dir_read_impure(const struct path *path,  struct list_head *list,
 566                                struct rb_root *root)
 567 {
 568         int err;
 569         struct path realpath;
 570         struct ovl_cache_entry *p, *n;
 571         struct ovl_readdir_data rdd = {
 572                 .ctx.actor = ovl_fill_plain,
 573                 .list = list,
 574                 .root = root,
 575         };
 576
 577         INIT_LIST_HEAD(list);
 578         *root = RB_ROOT;
 579         ovl_path_upper(path->dentry, &realpath);
 580
 581         err = ovl_dir_read(&realpath, &rdd);
 582         if (err)
 583                 return err;
 584
 585         list_for_each_entry_safe(p, n, list, l_node) {
 586                 if (strcmp(p->name, ".") != 0 &&
 587                     strcmp(p->name, "..") != 0) {
 588                         err = ovl_cache_update(path, p, true);
 589                         if (err)
 590                                 return err;
 591                 }
 592                 if (p->ino == p->real_ino) {
 593                         list_del(&p->l_node);
 594                         kfree(p);
 595                 } else {
 596                         struct rb_node **newp = &root->rb_node;
 597                         struct rb_node *parent = NULL;
 598
 599                         if (WARN_ON(ovl_cache_entry_find_link(p->name, p->len,
 600                                                               &newp, &parent)))
 601                                 return -EIO;
 602
 603                         rb_link_node(&p->node, parent, newp);
 604                         rb_insert_color(&p->node, root);
 605                 }
 606         }
 607         return 0;
 608 }
 609
 610 static struct ovl_dir_cache *ovl_cache_get_impure(const struct path *path)
 611 {
 612         int res;
 613         struct dentry *dentry = path->dentry;
 614         struct inode *inode = d_inode(dentry);
 615         struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 616         struct ovl_dir_cache *cache;
 617
 618         cache = ovl_dir_cache(inode);
 619         if (cache && ovl_inode_version_get(inode) == cache->version)
 620                 return cache;
 621
 622         /* Impure cache is not refcounted, free it here */
 623         ovl_dir_cache_free(inode);
 624         ovl_set_dir_cache(inode, NULL);
 625
 626         cache = kzalloc(sizeof(struct ovl_dir_cache), GFP_KERNEL);
 627         if (!cache)
 628                 return ERR_PTR(-ENOMEM);
 629
 630         res = ovl_dir_read_impure(path, &cache->entries, &cache->root);
 631         if (res) {
 632                 ovl_cache_free(&cache->entries);
 633                 kfree(cache);
 634                 return ERR_PTR(res);
 635         }
 636         if (list_empty(&cache->entries)) {
 637                 /*
 638                  * A good opportunity to get rid of an unneeded "impure" flag.
 639                  * Removing the "impure" xattr is best effort.
 640                  */
 641                 if (!ovl_want_write(dentry)) {
 642                         ovl_removexattr(ofs, ovl_dentry_upper(dentry),
 643                                         OVL_XATTR_IMPURE);
 644                         ovl_drop_write(dentry);
 645                 }
 646                 ovl_clear_flag(OVL_IMPURE, inode);
 647                 kfree(cache);
 648                 return NULL;
 649         }
 650
 651         cache->version = ovl_inode_version_get(inode);
 652         ovl_set_dir_cache(inode, cache);
 653
 654         return cache;
 655 }
 656
 657 struct ovl_readdir_translate {
 658         struct dir_context *orig_ctx;
 659         struct ovl_dir_cache *cache;
 660         struct dir_context ctx;
 661         u64 parent_ino;
 662         int fsid;
 663         int xinobits;
 664         bool xinowarn;
 665 };
 666
 667 static bool ovl_fill_real(struct dir_context *ctx, const char *name,
 668                            int namelen, loff_t offset, u64 ino,
 669                            unsigned int d_type)
 670 {
 671         struct ovl_readdir_translate *rdt =
 672                 container_of(ctx, struct ovl_readdir_translate, ctx);
 673         struct dir_context *orig_ctx = rdt->orig_ctx;
 674
 675         if (rdt->parent_ino && strcmp(name, "..") == 0) {
 676                 ino = rdt->parent_ino;
 677         } else if (rdt->cache) {
 678                 struct ovl_cache_entry *p;
 679
 680                 p = ovl_cache_entry_find(&rdt->cache->root, name, namelen);
 681                 if (p)
 682                         ino = p->ino;
 683         } else if (rdt->xinobits) {
 684                 ino = ovl_remap_lower_ino(ino, rdt->xinobits, rdt->fsid,
 685                                           name, namelen, rdt->xinowarn);
 686         }
 687
 688         return orig_ctx->actor(orig_ctx, name, namelen, offset, ino, d_type);
 689 }
 690
 691 static bool ovl_is_impure_dir(struct file *file)
 692 {
 693         struct ovl_dir_file *od = file->private_data;
 694         struct inode *dir = file_inode(file);
 695
 696         /*
 697          * Only upper dir can be impure, but if we are in the middle of
 698          * iterating a lower real dir, dir could be copied up and marked
 699          * impure. We only want the impure cache if we started iterating
 700          * a real upper dir to begin with.
 701          */
 702         return od->is_upper && ovl_test_flag(OVL_IMPURE, dir);
 703
 704 }
 705
 706 static int ovl_iterate_real(struct file *file, struct dir_context *ctx)
 707 {
 708         int err;
 709         struct ovl_dir_file *od = file->private_data;
 710         struct dentry *dir = file->f_path.dentry;
 711         struct ovl_fs *ofs = OVL_FS(dir->d_sb);
 712         const struct ovl_layer *lower_layer = ovl_layer_lower(dir);
 713         struct ovl_readdir_translate rdt = {
 714                 .ctx.actor = ovl_fill_real,
 715                 .orig_ctx = ctx,
 716                 .xinobits = ovl_xino_bits(ofs),
 717                 .xinowarn = ovl_xino_warn(ofs),
 718         };
 719
 720         if (rdt.xinobits && lower_layer)
 721                 rdt.fsid = lower_layer->fsid;
 722
 723         if (OVL_TYPE_MERGE(ovl_path_type(dir->d_parent))) {
 724                 struct kstat stat;
 725                 struct path statpath = file->f_path;
 726
 727                 statpath.dentry = dir->d_parent;
 728                 err = vfs_getattr(&statpath, &stat, STATX_INO, 0);
 729                 if (err)
 730                         return err;
 731
 732                 WARN_ON_ONCE(dir->d_sb->s_dev != stat.dev);
 733                 rdt.parent_ino = stat.ino;
 734         }
 735
 736         if (ovl_is_impure_dir(file)) {
 737                 rdt.cache = ovl_cache_get_impure(&file->f_path);
 738                 if (IS_ERR(rdt.cache))
 739                         return PTR_ERR(rdt.cache);
 740         }
 741
 742         err = iterate_dir(od->realfile, &rdt.ctx);
 743         ctx->pos = rdt.ctx.pos;
 744
 745         return err;
 746 }
 747
 748
 749 static int ovl_iterate(struct file *file, struct dir_context *ctx)
 750 {
 751         struct ovl_dir_file *od = file->private_data;
 752         struct dentry *dentry = file->f_path.dentry;
 753         struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 754         struct ovl_cache_entry *p;
 755         const struct cred *old_cred;
 756         int err;
 757
 758         old_cred = ovl_override_creds(dentry->d_sb);
 759         if (!ctx->pos)
 760                 ovl_dir_reset(file);
 761
 762         if (od->is_real) {
 763                 /*
 764                  * If parent is merge, then need to adjust d_ino for '..', if
 765                  * dir is impure then need to adjust d_ino for copied up
 766                  * entries.
 767                  */
 768                 if (ovl_xino_bits(ofs) ||
 769                     (ovl_same_fs(ofs) &&
 770                      (ovl_is_impure_dir(file) ||
 771                       OVL_TYPE_MERGE(ovl_path_type(dentry->d_parent))))) {
 772                         err = ovl_iterate_real(file, ctx);
 773                 } else {
 774                         err = iterate_dir(od->realfile, ctx);
 775                 }
 776                 goto out;
 777         }
 778
 779         if (!od->cache) {
 780                 struct ovl_dir_cache *cache;
 781
 782                 cache = ovl_cache_get(dentry);
 783                 err = PTR_ERR(cache);
 784                 if (IS_ERR(cache))
 785                         goto out;
 786
 787                 od->cache = cache;
 788                 ovl_seek_cursor(od, ctx->pos);
 789         }
 790
 791         while (od->cursor != &od->cache->entries) {
 792                 p = list_entry(od->cursor, struct ovl_cache_entry, l_node);
 793                 if (!p->is_whiteout) {
 794                         if (!p->ino || p->check_xwhiteout) {
 795                                 err = ovl_cache_update(&file->f_path, p, !p->ino);
 796                                 if (err)
 797                                         goto out;
 798                         }
 799                 }
 800                 /* ovl_cache_update() sets is_whiteout on stale entry */
 801                 if (!p->is_whiteout) {
 802                         if (!dir_emit(ctx, p->name, p->len, p->ino, p->type))
 803                                 break;
 804                 }
 805                 od->cursor = p->l_node.next;
 806                 ctx->pos++;
 807         }
 808         err = 0;
 809 out:
 810         revert_creds(old_cred);
 811         return err;
 812 }
 813
 814 static loff_t ovl_dir_llseek(struct file *file, loff_t offset, int origin)
 815 {
 816         loff_t res;
 817         struct ovl_dir_file *od = file->private_data;
 818
 819         inode_lock(file_inode(file));
 820         if (!file->f_pos)
 821                 ovl_dir_reset(file);
 822
 823         if (od->is_real) {
 824                 res = vfs_llseek(od->realfile, offset, origin);
 825                 file->f_pos = od->realfile->f_pos;
 826         } else {
 827                 res = -EINVAL;
 828
 829                 switch (origin) {
 830                 case SEEK_CUR:
 831                         offset += file->f_pos;
 832                         break;
 833                 case SEEK_SET:
 834                         break;
 835                 default:
 836                         goto out_unlock;
 837                 }
 838                 if (offset < 0)
 839                         goto out_unlock;
 840
 841                 if (offset != file->f_pos) {
 842                         file->f_pos = offset;
 843                         if (od->cache)
 844                                 ovl_seek_cursor(od, offset);
 845                 }
 846                 res = offset;
 847         }
 848 out_unlock:
 849         inode_unlock(file_inode(file));
 850
 851         return res;
 852 }
 853
 854 static struct file *ovl_dir_open_realfile(const struct file *file,
 855                                           const struct path *realpath)
 856 {
 857         struct file *res;
 858         const struct cred *old_cred;
 859
 860         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 861         res = ovl_path_open(realpath, O_RDONLY | (file->f_flags & O_LARGEFILE));
 862         revert_creds(old_cred);
 863
 864         return res;
 865 }
 866
 867 /*
 868  * Like ovl_real_fdget(), returns upperfile if dir was copied up since open.
 869  * Unlike ovl_real_fdget(), this caches upperfile in file->private_data.
 870  *
 871  * TODO: use same abstract type for file->private_data of dir and file so
 872  * upperfile could also be cached for files as well.
 873  */
 874 struct file *ovl_dir_real_file(const struct file *file, bool want_upper)
 875 {
 876
 877         struct ovl_dir_file *od = file->private_data;
 878         struct dentry *dentry = file->f_path.dentry;
 879         struct file *old, *realfile = od->realfile;
 880
 881         if (!OVL_TYPE_UPPER(ovl_path_type(dentry)))
 882                 return want_upper ? NULL : realfile;
 883
 884         /*
 885          * Need to check if we started out being a lower dir, but got copied up
 886          */
 887         if (!od->is_upper) {
 888                 realfile = READ_ONCE(od->upperfile);
 889                 if (!realfile) {
 890                         struct path upperpath;
 891
 892                         ovl_path_upper(dentry, &upperpath);
 893                         realfile = ovl_dir_open_realfile(file, &upperpath);
 894                         if (IS_ERR(realfile))
 895                                 return realfile;
 896
 897                         old = cmpxchg_release(&od->upperfile, NULL, realfile);
 898                         if (old) {
 899                                 fput(realfile);
 900                                 realfile = old;
 901                         }
 902                 }
 903         }
 904
 905         return realfile;
 906 }
 907
 908 static int ovl_dir_fsync(struct file *file, loff_t start, loff_t end,
 909                          int datasync)
 910 {
 911         struct file *realfile;
 912         int err;
 913
 914         err = ovl_sync_status(OVL_FS(file_inode(file)->i_sb));
 915         if (err <= 0)
 916                 return err;
 917
 918         realfile = ovl_dir_real_file(file, true);
 919         err = PTR_ERR_OR_ZERO(realfile);
 920
 921         /* Nothing to sync for lower */
 922         if (!realfile || err)
 923                 return err;
 924
 925         return vfs_fsync_range(realfile, start, end, datasync);
 926 }
 927
 928 static int ovl_dir_release(struct inode *inode, struct file *file)
 929 {
 930         struct ovl_dir_file *od = file->private_data;
 931
 932         if (od->cache) {
 933                 inode_lock(inode);
 934                 ovl_cache_put(od, inode);
 935                 inode_unlock(inode);
 936         }
 937         fput(od->realfile);
 938         if (od->upperfile)
 939                 fput(od->upperfile);
 940         kfree(od);
 941
 942         return 0;
 943 }
 944
 945 static int ovl_dir_open(struct inode *inode, struct file *file)
 946 {
 947         struct path realpath;
 948         struct file *realfile;
 949         struct ovl_dir_file *od;
 950         enum ovl_path_type type;
 951
 952         od = kzalloc(sizeof(struct ovl_dir_file), GFP_KERNEL);
 953         if (!od)
 954                 return -ENOMEM;
 955
 956         type = ovl_path_real(file->f_path.dentry, &realpath);
 957         realfile = ovl_dir_open_realfile(file, &realpath);
 958         if (IS_ERR(realfile)) {
 959                 kfree(od);
 960                 return PTR_ERR(realfile);
 961         }
 962         od->realfile = realfile;
 963         od->is_real = ovl_dir_is_real(inode);
 964         od->is_upper = OVL_TYPE_UPPER(type);
 965         file->private_data = od;
 966
 967         return 0;
 968 }
 969
 970 WRAP_DIR_ITER(ovl_iterate) // FIXME!
 971 const struct file_operations ovl_dir_operations = {
 972         .read           = generic_read_dir,
 973         .open           = ovl_dir_open,
 974         .iterate_shared = shared_ovl_iterate,
 975         .llseek         = ovl_dir_llseek,
 976         .fsync          = ovl_dir_fsync,
 977         .release        = ovl_dir_release,
 978 };
 979
 980 int ovl_check_empty_dir(struct dentry *dentry, struct list_head *list)
 981 {
 982         int err;
 983         struct ovl_cache_entry *p, *n;
 984         struct rb_root root = RB_ROOT;
 985         const struct cred *old_cred;
 986
 987         old_cred = ovl_override_creds(dentry->d_sb);
 988         err = ovl_dir_read_merged(dentry, list, &root);
 989         revert_creds(old_cred);
 990         if (err)
 991                 return err;
 992
 993         err = 0;
 994
 995         list_for_each_entry_safe(p, n, list, l_node) {
 996                 /*
 997                  * Select whiteouts in upperdir, they should
 998                  * be cleared when deleting this directory.
 999                  */
1000                 if (p->is_whiteout) {
1001                         if (p->is_upper)
1002                                 continue;
1003                         goto del_entry;
1004                 }
1005
1006                 if (p->name[0] == '.') {
1007                         if (p->len == 1)
1008                                 goto del_entry;
1009                         if (p->len == 2 && p->name[1] == '.')
1010                                 goto del_entry;
1011                 }
1012                 err = -ENOTEMPTY;
1013                 break;
1014
1015 del_entry:
1016                 list_del(&p->l_node);
1017                 kfree(p);
1018         }
1019
1020         return err;
1021 }
1022
1023 void ovl_cleanup_whiteouts(struct ovl_fs *ofs, struct dentry *upper,
1024                            struct list_head *list)
1025 {
1026         struct ovl_cache_entry *p;
1027
1028         inode_lock_nested(upper->d_inode, I_MUTEX_CHILD);
1029         list_for_each_entry(p, list, l_node) {
1030                 struct dentry *dentry;
1031
1032                 if (WARN_ON(!p->is_whiteout || !p->is_upper))
1033                         continue;
1034
1035                 dentry = ovl_lookup_upper(ofs, p->name, upper, p->len);
1036                 if (IS_ERR(dentry)) {
1037                         pr_err("lookup '%s/%.*s' failed (%i)\n",
1038                                upper->d_name.name, p->len, p->name,
1039                                (int) PTR_ERR(dentry));
1040                         continue;
1041                 }
1042                 if (dentry->d_inode)
1043                         ovl_cleanup(ofs, upper->d_inode, dentry);
1044                 dput(dentry);
1045         }
1046         inode_unlock(upper->d_inode);
1047 }
1048
1049 static bool ovl_check_d_type(struct dir_context *ctx, const char *name,
1050                           int namelen, loff_t offset, u64 ino,
1051                           unsigned int d_type)
1052 {
1053         struct ovl_readdir_data *rdd =
1054                 container_of(ctx, struct ovl_readdir_data, ctx);
1055
1056         /* Even if d_type is not supported, DT_DIR is returned for . and .. */
1057         if (!strncmp(name, ".", namelen) || !strncmp(name, "..", namelen))
1058                 return true;
1059
1060         if (d_type != DT_UNKNOWN)
1061                 rdd->d_type_supported = true;
1062
1063         return true;
1064 }
1065
1066 /*
1067  * Returns 1 if d_type is supported, 0 not supported/unknown. Negative values
1068  * if error is encountered.
1069  */
1070 int ovl_check_d_type_supported(const struct path *realpath)
1071 {
1072         int err;
1073         struct ovl_readdir_data rdd = {
1074                 .ctx.actor = ovl_check_d_type,
1075                 .d_type_supported = false,
1076         };
1077
1078         err = ovl_dir_read(realpath, &rdd);
1079         if (err)
1080                 return err;
1081
1082         return rdd.d_type_supported;
1083 }
1084
1085 #define OVL_INCOMPATDIR_NAME "incompat"
1086
1087 static int ovl_workdir_cleanup_recurse(struct ovl_fs *ofs, const struct path *path,
1088                                        int level)
1089 {
1090         int err;
1091         struct inode *dir = path->dentry->d_inode;
1092         LIST_HEAD(list);
1093         struct ovl_cache_entry *p;
1094         struct ovl_readdir_data rdd = {
1095                 .ctx.actor = ovl_fill_plain,
1096                 .list = &list,
1097         };
1098         bool incompat = false;
1099
1100         /*
1101          * The "work/incompat" directory is treated specially - if it is not
1102          * empty, instead of printing a generic error and mounting read-only,
1103          * we will error about incompat features and fail the mount.
1104          *
1105          * When called from ovl_indexdir_cleanup(), path->dentry->d_name.name
1106          * starts with '#'.
1107          */
1108         if (level == 2 &&
1109             !strcmp(path->dentry->d_name.name, OVL_INCOMPATDIR_NAME))
1110                 incompat = true;
1111
1112         err = ovl_dir_read(path, &rdd);
1113         if (err)
1114                 goto out;
1115
1116         inode_lock_nested(dir, I_MUTEX_PARENT);
1117         list_for_each_entry(p, &list, l_node) {
1118                 struct dentry *dentry;
1119
1120                 if (p->name[0] == '.') {
1121                         if (p->len == 1)
1122                                 continue;
1123                         if (p->len == 2 && p->name[1] == '.')
1124                                 continue;
1125                 } else if (incompat) {
1126                         pr_err("overlay with incompat feature '%s' cannot be mounted\n",
1127                                 p->name);
1128                         err = -EINVAL;
1129                         break;
1130                 }
1131                 dentry = ovl_lookup_upper(ofs, p->name, path->dentry, p->len);
1132                 if (IS_ERR(dentry))
1133                         continue;
1134                 if (dentry->d_inode)
1135                         err = ovl_workdir_cleanup(ofs, dir, path->mnt, dentry, level);
1136                 dput(dentry);
1137                 if (err)
1138                         break;
1139         }
1140         inode_unlock(dir);
1141 out:
1142         ovl_cache_free(&list);
1143         return err;
1144 }
1145
1146 int ovl_workdir_cleanup(struct ovl_fs *ofs, struct inode *dir,
1147                         struct vfsmount *mnt, struct dentry *dentry, int level)
1148 {
1149         int err;
1150
1151         if (!d_is_dir(dentry) || level > 1) {
1152                 return ovl_cleanup(ofs, dir, dentry);
1153         }
1154
1155         err = ovl_do_rmdir(ofs, dir, dentry);
1156         if (err) {
1157                 struct path path = { .mnt = mnt, .dentry = dentry };
1158
1159                 inode_unlock(dir);
1160                 err = ovl_workdir_cleanup_recurse(ofs, &path, level + 1);
1161                 inode_lock_nested(dir, I_MUTEX_PARENT);
1162                 if (!err)
1163                         err = ovl_cleanup(ofs, dir, dentry);
1164         }
1165
1166         return err;
1167 }
1168
1169 int ovl_indexdir_cleanup(struct ovl_fs *ofs)
1170 {
1171         int err;
1172         struct dentry *indexdir = ofs->indexdir;
1173         struct dentry *index = NULL;
1174         struct inode *dir = indexdir->d_inode;
1175         struct path path = { .mnt = ovl_upper_mnt(ofs), .dentry = indexdir };
1176         LIST_HEAD(list);
1177         struct ovl_cache_entry *p;
1178         struct ovl_readdir_data rdd = {
1179                 .ctx.actor = ovl_fill_plain,
1180                 .list = &list,
1181         };
1182
1183         err = ovl_dir_read(&path, &rdd);
1184         if (err)
1185                 goto out;
1186
1187         inode_lock_nested(dir, I_MUTEX_PARENT);
1188         list_for_each_entry(p, &list, l_node) {
1189                 if (p->name[0] == '.') {
1190                         if (p->len == 1)
1191                                 continue;
1192                         if (p->len == 2 && p->name[1] == '.')
1193                                 continue;
1194                 }
1195                 index = ovl_lookup_upper(ofs, p->name, indexdir, p->len);
1196                 if (IS_ERR(index)) {
1197                         err = PTR_ERR(index);
1198                         index = NULL;
1199                         break;
1200                 }
1201                 /* Cleanup leftover from index create/cleanup attempt */
1202                 if (index->d_name.name[0] == '#') {
1203                         err = ovl_workdir_cleanup(ofs, dir, path.mnt, index, 1);
1204                         if (err)
1205                                 break;
1206                         goto next;
1207                 }
1208                 err = ovl_verify_index(ofs, index);
1209                 if (!err) {
1210                         goto next;
1211                 } else if (err == -ESTALE) {
1212                         /* Cleanup stale index entries */
1213                         err = ovl_cleanup(ofs, dir, index);
1214                 } else if (err != -ENOENT) {
1215                         /*
1216                          * Abort mount to avoid corrupting the index if
1217                          * an incompatible index entry was found or on out
1218                          * of memory.
1219                          */
1220                         break;
1221                 } else if (ofs->config.nfs_export) {
1222                         /*
1223                          * Whiteout orphan index to block future open by
1224                          * handle after overlay nlink dropped to zero.
1225                          */
1226                         err = ovl_cleanup_and_whiteout(ofs, dir, index);
1227                 } else {
1228                         /* Cleanup orphan index entries */
1229                         err = ovl_cleanup(ofs, dir, index);
1230                 }
1231
1232                 if (err)
1233                         break;
1234
1235 next:
1236                 dput(index);
1237                 index = NULL;
1238         }
1239         dput(index);
1240         inode_unlock(dir);
1241 out:
1242         ovl_cache_free(&list);
1243         if (err)
1244                 pr_err("failed index dir cleanup (%i)\n", err);
1245         return err;
1246 }