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         ovl_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->first_maybe_whiteout = NULL;
 309         rdd->ctx.pos = 0;
 310         do {
 311                 rdd->count = 0;
 312                 rdd->err = 0;
 313                 err = iterate_dir(realfile, &rdd->ctx);
 314                 if (err >= 0)
 315                         err = rdd->err;
 316         } while (!err && rdd->count);
 317
 318         if (!err && rdd->first_maybe_whiteout && rdd->dentry)
 319                 err = ovl_check_whiteouts(realpath, rdd);
 320
 321         fput(realfile);
 322
 323         return err;
 324 }
 325
 326 static void ovl_dir_reset(struct file *file)
 327 {
 328         struct ovl_dir_file *od = file->private_data;
 329         struct ovl_dir_cache *cache = od->cache;
 330         struct inode *inode = file_inode(file);
 331         bool is_real;
 332
 333         if (cache && ovl_inode_version_get(inode) != cache->version) {
 334                 ovl_cache_put(od, inode);
 335                 od->cache = NULL;
 336                 od->cursor = NULL;
 337         }
 338         is_real = ovl_dir_is_real(inode);
 339         if (od->is_real != is_real) {
 340                 /* is_real can only become false when dir is copied up */
 341                 if (WARN_ON(is_real))
 342                         return;
 343                 od->is_real = false;
 344         }
 345 }
 346
 347 static int ovl_dir_read_merged(struct dentry *dentry, struct list_head *list,
 348         struct rb_root *root)
 349 {
 350         int err;
 351         struct path realpath;
 352         struct ovl_readdir_data rdd = {
 353                 .ctx.actor = ovl_fill_merge,
 354                 .dentry = dentry,
 355                 .list = list,
 356                 .root = root,
 357                 .is_lowest = false,
 358         };
 359         int idx, next;
 360         const struct ovl_layer *layer;
 361
 362         for (idx = 0; idx != -1; idx = next) {
 363                 next = ovl_path_next(idx, dentry, &realpath, &layer);
 364                 rdd.is_upper = ovl_dentry_upper(dentry) == realpath.dentry;
 365                 rdd.in_xwhiteouts_dir = layer->has_xwhiteouts &&
 366                                         ovl_dentry_has_xwhiteouts(dentry);
 367
 368                 if (next != -1) {
 369                         err = ovl_dir_read(&realpath, &rdd);
 370                         if (err)
 371                                 break;
 372                 } else {
 373                         /*
 374                          * Insert lowest layer entries before upper ones, this
 375                          * allows offsets to be reasonably constant
 376                          */
 377                         list_add(&rdd.middle, rdd.list);
 378                         rdd.is_lowest = true;
 379                         err = ovl_dir_read(&realpath, &rdd);
 380                         list_del(&rdd.middle);
 381                 }
 382         }
 383         return err;
 384 }
 385
 386 static void ovl_seek_cursor(struct ovl_dir_file *od, loff_t pos)
 387 {
 388         struct list_head *p;
 389         loff_t off = 0;
 390
 391         list_for_each(p, &od->cache->entries) {
 392                 if (off >= pos)
 393                         break;
 394                 off++;
 395         }
 396         /* Cursor is safe since the cache is stable */
 397         od->cursor = p;
 398 }
 399
 400 static struct ovl_dir_cache *ovl_cache_get(struct dentry *dentry)
 401 {
 402         int res;
 403         struct ovl_dir_cache *cache;
 404         struct inode *inode = d_inode(dentry);
 405
 406         cache = ovl_dir_cache(inode);
 407         if (cache && ovl_inode_version_get(inode) == cache->version) {
 408                 WARN_ON(!cache->refcount);
 409                 cache->refcount++;
 410                 return cache;
 411         }
 412         ovl_set_dir_cache(d_inode(dentry), NULL);
 413
 414         cache = kzalloc(sizeof(struct ovl_dir_cache), GFP_KERNEL);
 415         if (!cache)
 416                 return ERR_PTR(-ENOMEM);
 417
 418         cache->refcount = 1;
 419         INIT_LIST_HEAD(&cache->entries);
 420         cache->root = RB_ROOT;
 421
 422         res = ovl_dir_read_merged(dentry, &cache->entries, &cache->root);
 423         if (res) {
 424                 ovl_cache_free(&cache->entries);
 425                 kfree(cache);
 426                 return ERR_PTR(res);
 427         }
 428
 429         cache->version = ovl_inode_version_get(inode);
 430         ovl_set_dir_cache(inode, cache);
 431
 432         return cache;
 433 }
 434
 435 /* Map inode number to lower fs unique range */
 436 static u64 ovl_remap_lower_ino(u64 ino, int xinobits, int fsid,
 437                                const char *name, int namelen, bool warn)
 438 {
 439         unsigned int xinoshift = 64 - xinobits;
 440
 441         if (unlikely(ino >> xinoshift)) {
 442                 if (warn) {
 443                         pr_warn_ratelimited("d_ino too big (%.*s, ino=%llu, xinobits=%d)\n",
 444                                             namelen, name, ino, xinobits);
 445                 }
 446                 return ino;
 447         }
 448
 449         /*
 450          * The lowest xinobit is reserved for mapping the non-peresistent inode
 451          * numbers range, but this range is only exposed via st_ino, not here.
 452          */
 453         return ino | ((u64)fsid) << (xinoshift + 1);
 454 }
 455
 456 /*
 457  * Set d_ino for upper entries if needed. Non-upper entries should always report
 458  * the uppermost real inode ino and should not call this function.
 459  *
 460  * When not all layer are on same fs, report real ino also for upper.
 461  *
 462  * When all layers are on the same fs, and upper has a reference to
 463  * copy up origin, call vfs_getattr() on the overlay entry to make
 464  * sure that d_ino will be consistent with st_ino from stat(2).
 465  *
 466  * Also checks the overlay.whiteout xattr by doing a full lookup which will return
 467  * negative in this case.
 468  */
 469 static int ovl_cache_update(const struct path *path, struct ovl_cache_entry *p, bool update_ino)
 470
 471 {
 472         struct dentry *dir = path->dentry;
 473         struct ovl_fs *ofs = OVL_FS(dir->d_sb);
 474         struct dentry *this = NULL;
 475         enum ovl_path_type type;
 476         u64 ino = p->real_ino;
 477         int xinobits = ovl_xino_bits(ofs);
 478         int err = 0;
 479
 480         if (!ovl_same_dev(ofs) && !p->check_xwhiteout)
 481                 goto out;
 482
 483         if (p->name[0] == '.') {
 484                 if (p->len == 1) {
 485                         this = dget(dir);
 486                         goto get;
 487                 }
 488                 if (p->len == 2 && p->name[1] == '.') {
 489                         /* we shall not be moved */
 490                         this = dget(dir->d_parent);
 491                         goto get;
 492                 }
 493         }
 494         /* This checks also for xwhiteouts */
 495         this = lookup_one(mnt_idmap(path->mnt), p->name, dir, p->len);
 496         if (IS_ERR_OR_NULL(this) || !this->d_inode) {
 497                 /* Mark a stale entry */
 498                 p->is_whiteout = true;
 499                 if (IS_ERR(this)) {
 500                         err = PTR_ERR(this);
 501                         this = NULL;
 502                         goto fail;
 503                 }
 504                 goto out;
 505         }
 506
 507 get:
 508         if (!ovl_same_dev(ofs) || !update_ino)
 509                 goto out;
 510
 511         type = ovl_path_type(this);
 512         if (OVL_TYPE_ORIGIN(type)) {
 513                 struct kstat stat;
 514                 struct path statpath = *path;
 515
 516                 statpath.dentry = this;
 517                 err = vfs_getattr(&statpath, &stat, STATX_INO, 0);
 518                 if (err)
 519                         goto fail;
 520
 521                 /*
 522                  * Directory inode is always on overlay st_dev.
 523                  * Non-dir with ovl_same_dev() could be on pseudo st_dev in case
 524                  * of xino bits overflow.
 525                  */
 526                 WARN_ON_ONCE(S_ISDIR(stat.mode) &&
 527                              dir->d_sb->s_dev != stat.dev);
 528                 ino = stat.ino;
 529         } else if (xinobits && !OVL_TYPE_UPPER(type)) {
 530                 ino = ovl_remap_lower_ino(ino, xinobits,
 531                                           ovl_layer_lower(this)->fsid,
 532                                           p->name, p->len,
 533                                           ovl_xino_warn(ofs));
 534         }
 535
 536 out:
 537         p->ino = ino;
 538         dput(this);
 539         return err;
 540
 541 fail:
 542         pr_warn_ratelimited("failed to look up (%s) for ino (%i)\n",
 543                             p->name, err);
 544         goto out;
 545 }
 546
 547 static bool ovl_fill_plain(struct dir_context *ctx, const char *name,
 548                           int namelen, loff_t offset, u64 ino,
 549                           unsigned int d_type)
 550 {
 551         struct ovl_cache_entry *p;
 552         struct ovl_readdir_data *rdd =
 553                 container_of(ctx, struct ovl_readdir_data, ctx);
 554
 555         rdd->count++;
 556         p = ovl_cache_entry_new(rdd, name, namelen, ino, d_type);
 557         if (p == NULL) {
 558                 rdd->err = -ENOMEM;
 559                 return false;
 560         }
 561         list_add_tail(&p->l_node, rdd->list);
 562
 563         return true;
 564 }
 565
 566 static int ovl_dir_read_impure(const struct path *path,  struct list_head *list,
 567                                struct rb_root *root)
 568 {
 569         int err;
 570         struct path realpath;
 571         struct ovl_cache_entry *p, *n;
 572         struct ovl_readdir_data rdd = {
 573                 .ctx.actor = ovl_fill_plain,
 574                 .list = list,
 575                 .root = root,
 576         };
 577
 578         INIT_LIST_HEAD(list);
 579         *root = RB_ROOT;
 580         ovl_path_upper(path->dentry, &realpath);
 581
 582         err = ovl_dir_read(&realpath, &rdd);
 583         if (err)
 584                 return err;
 585
 586         list_for_each_entry_safe(p, n, list, l_node) {
 587                 if (strcmp(p->name, ".") != 0 &&
 588                     strcmp(p->name, "..") != 0) {
 589                         err = ovl_cache_update(path, p, true);
 590                         if (err)
 591                                 return err;
 592                 }
 593                 if (p->ino == p->real_ino) {
 594                         list_del(&p->l_node);
 595                         kfree(p);
 596                 } else {
 597                         struct rb_node **newp = &root->rb_node;
 598                         struct rb_node *parent = NULL;
 599
 600                         if (WARN_ON(ovl_cache_entry_find_link(p->name, p->len,
 601                                                               &newp, &parent)))
 602                                 return -EIO;
 603
 604                         rb_link_node(&p->node, parent, newp);
 605                         rb_insert_color(&p->node, root);
 606                 }
 607         }
 608         return 0;
 609 }
 610
 611 static struct ovl_dir_cache *ovl_cache_get_impure(const struct path *path)
 612 {
 613         int res;
 614         struct dentry *dentry = path->dentry;
 615         struct inode *inode = d_inode(dentry);
 616         struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 617         struct ovl_dir_cache *cache;
 618
 619         cache = ovl_dir_cache(inode);
 620         if (cache && ovl_inode_version_get(inode) == cache->version)
 621                 return cache;
 622
 623         /* Impure cache is not refcounted, free it here */
 624         ovl_dir_cache_free(inode);
 625         ovl_set_dir_cache(inode, NULL);
 626
 627         cache = kzalloc(sizeof(struct ovl_dir_cache), GFP_KERNEL);
 628         if (!cache)
 629                 return ERR_PTR(-ENOMEM);
 630
 631         res = ovl_dir_read_impure(path, &cache->entries, &cache->root);
 632         if (res) {
 633                 ovl_cache_free(&cache->entries);
 634                 kfree(cache);
 635                 return ERR_PTR(res);
 636         }
 637         if (list_empty(&cache->entries)) {
 638                 /*
 639                  * A good opportunity to get rid of an unneeded "impure" flag.
 640                  * Removing the "impure" xattr is best effort.
 641                  */
 642                 if (!ovl_want_write(dentry)) {
 643                         ovl_removexattr(ofs, ovl_dentry_upper(dentry),
 644                                         OVL_XATTR_IMPURE);
 645                         ovl_drop_write(dentry);
 646                 }
 647                 ovl_clear_flag(OVL_IMPURE, inode);
 648                 kfree(cache);
 649                 return NULL;
 650         }
 651
 652         cache->version = ovl_inode_version_get(inode);
 653         ovl_set_dir_cache(inode, cache);
 654
 655         return cache;
 656 }
 657
 658 struct ovl_readdir_translate {
 659         struct dir_context *orig_ctx;
 660         struct ovl_dir_cache *cache;
 661         struct dir_context ctx;
 662         u64 parent_ino;
 663         int fsid;
 664         int xinobits;
 665         bool xinowarn;
 666 };
 667
 668 static bool ovl_fill_real(struct dir_context *ctx, const char *name,
 669                            int namelen, loff_t offset, u64 ino,
 670                            unsigned int d_type)
 671 {
 672         struct ovl_readdir_translate *rdt =
 673                 container_of(ctx, struct ovl_readdir_translate, ctx);
 674         struct dir_context *orig_ctx = rdt->orig_ctx;
 675
 676         if (rdt->parent_ino && strcmp(name, "..") == 0) {
 677                 ino = rdt->parent_ino;
 678         } else if (rdt->cache) {
 679                 struct ovl_cache_entry *p;
 680
 681                 p = ovl_cache_entry_find(&rdt->cache->root, name, namelen);
 682                 if (p)
 683                         ino = p->ino;
 684         } else if (rdt->xinobits) {
 685                 ino = ovl_remap_lower_ino(ino, rdt->xinobits, rdt->fsid,
 686                                           name, namelen, rdt->xinowarn);
 687         }
 688
 689         return orig_ctx->actor(orig_ctx, name, namelen, offset, ino, d_type);
 690 }
 691
 692 static bool ovl_is_impure_dir(struct file *file)
 693 {
 694         struct ovl_dir_file *od = file->private_data;
 695         struct inode *dir = file_inode(file);
 696
 697         /*
 698          * Only upper dir can be impure, but if we are in the middle of
 699          * iterating a lower real dir, dir could be copied up and marked
 700          * impure. We only want the impure cache if we started iterating
 701          * a real upper dir to begin with.
 702          */
 703         return od->is_upper && ovl_test_flag(OVL_IMPURE, dir);
 704
 705 }
 706
 707 static int ovl_iterate_real(struct file *file, struct dir_context *ctx)
 708 {
 709         int err;
 710         struct ovl_dir_file *od = file->private_data;
 711         struct dentry *dir = file->f_path.dentry;
 712         struct ovl_fs *ofs = OVL_FS(dir->d_sb);
 713         const struct ovl_layer *lower_layer = ovl_layer_lower(dir);
 714         struct ovl_readdir_translate rdt = {
 715                 .ctx.actor = ovl_fill_real,
 716                 .orig_ctx = ctx,
 717                 .xinobits = ovl_xino_bits(ofs),
 718                 .xinowarn = ovl_xino_warn(ofs),
 719         };
 720
 721         if (rdt.xinobits && lower_layer)
 722                 rdt.fsid = lower_layer->fsid;
 723
 724         if (OVL_TYPE_MERGE(ovl_path_type(dir->d_parent))) {
 725                 struct kstat stat;
 726                 struct path statpath = file->f_path;
 727
 728                 statpath.dentry = dir->d_parent;
 729                 err = vfs_getattr(&statpath, &stat, STATX_INO, 0);
 730                 if (err)
 731                         return err;
 732
 733                 WARN_ON_ONCE(dir->d_sb->s_dev != stat.dev);
 734                 rdt.parent_ino = stat.ino;
 735         }
 736
 737         if (ovl_is_impure_dir(file)) {
 738                 rdt.cache = ovl_cache_get_impure(&file->f_path);
 739                 if (IS_ERR(rdt.cache))
 740                         return PTR_ERR(rdt.cache);
 741         }
 742
 743         err = iterate_dir(od->realfile, &rdt.ctx);
 744         ctx->pos = rdt.ctx.pos;
 745
 746         return err;
 747 }
 748
 749
 750 static int ovl_iterate(struct file *file, struct dir_context *ctx)
 751 {
 752         struct ovl_dir_file *od = file->private_data;
 753         struct dentry *dentry = file->f_path.dentry;
 754         struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 755         struct ovl_cache_entry *p;
 756         const struct cred *old_cred;
 757         int err;
 758
 759         old_cred = ovl_override_creds(dentry->d_sb);
 760         if (!ctx->pos)
 761                 ovl_dir_reset(file);
 762
 763         if (od->is_real) {
 764                 /*
 765                  * If parent is merge, then need to adjust d_ino for '..', if
 766                  * dir is impure then need to adjust d_ino for copied up
 767                  * entries.
 768                  */
 769                 if (ovl_xino_bits(ofs) ||
 770                     (ovl_same_fs(ofs) &&
 771                      (ovl_is_impure_dir(file) ||
 772                       OVL_TYPE_MERGE(ovl_path_type(dentry->d_parent))))) {
 773                         err = ovl_iterate_real(file, ctx);
 774                 } else {
 775                         err = iterate_dir(od->realfile, ctx);
 776                 }
 777                 goto out;
 778         }
 779
 780         if (!od->cache) {
 781                 struct ovl_dir_cache *cache;
 782
 783                 cache = ovl_cache_get(dentry);
 784                 err = PTR_ERR(cache);
 785                 if (IS_ERR(cache))
 786                         goto out;
 787
 788                 od->cache = cache;
 789                 ovl_seek_cursor(od, ctx->pos);
 790         }
 791
 792         while (od->cursor != &od->cache->entries) {
 793                 p = list_entry(od->cursor, struct ovl_cache_entry, l_node);
 794                 if (!p->is_whiteout) {
 795                         if (!p->ino || p->check_xwhiteout) {
 796                                 err = ovl_cache_update(&file->f_path, p, !p->ino);
 797                                 if (err)
 798                                         goto out;
 799                         }
 800                 }
 801                 /* ovl_cache_update() sets is_whiteout on stale entry */
 802                 if (!p->is_whiteout) {
 803                         if (!dir_emit(ctx, p->name, p->len, p->ino, p->type))
 804                                 break;
 805                 }
 806                 od->cursor = p->l_node.next;
 807                 ctx->pos++;
 808         }
 809         err = 0;
 810 out:
 811         ovl_revert_creds(old_cred);
 812         return err;
 813 }
 814
 815 static loff_t ovl_dir_llseek(struct file *file, loff_t offset, int origin)
 816 {
 817         loff_t res;
 818         struct ovl_dir_file *od = file->private_data;
 819
 820         inode_lock(file_inode(file));
 821         if (!file->f_pos)
 822                 ovl_dir_reset(file);
 823
 824         if (od->is_real) {
 825                 res = vfs_llseek(od->realfile, offset, origin);
 826                 file->f_pos = od->realfile->f_pos;
 827         } else {
 828                 res = -EINVAL;
 829
 830                 switch (origin) {
 831                 case SEEK_CUR:
 832                         offset += file->f_pos;
 833                         break;
 834                 case SEEK_SET:
 835                         break;
 836                 default:
 837                         goto out_unlock;
 838                 }
 839                 if (offset < 0)
 840                         goto out_unlock;
 841
 842                 if (offset != file->f_pos) {
 843                         file->f_pos = offset;
 844                         if (od->cache)
 845                                 ovl_seek_cursor(od, offset);
 846                 }
 847                 res = offset;
 848         }
 849 out_unlock:
 850         inode_unlock(file_inode(file));
 851
 852         return res;
 853 }
 854
 855 static struct file *ovl_dir_open_realfile(const struct file *file,
 856                                           const struct path *realpath)
 857 {
 858         struct file *res;
 859         const struct cred *old_cred;
 860
 861         old_cred = ovl_override_creds(file_inode(file)->i_sb);
 862         res = ovl_path_open(realpath, O_RDONLY | (file->f_flags & O_LARGEFILE));
 863         ovl_revert_creds(old_cred);
 864
 865         return res;
 866 }
 867
 868 /*
 869  * Like ovl_real_fdget(), returns upperfile if dir was copied up since open.
 870  * Unlike ovl_real_fdget(), this caches upperfile in file->private_data.
 871  *
 872  * TODO: use same abstract type for file->private_data of dir and file so
 873  * upperfile could also be cached for files as well.
 874  */
 875 struct file *ovl_dir_real_file(const struct file *file, bool want_upper)
 876 {
 877
 878         struct ovl_dir_file *od = file->private_data;
 879         struct dentry *dentry = file->f_path.dentry;
 880         struct file *old, *realfile = od->realfile;
 881
 882         if (!OVL_TYPE_UPPER(ovl_path_type(dentry)))
 883                 return want_upper ? NULL : realfile;
 884
 885         /*
 886          * Need to check if we started out being a lower dir, but got copied up
 887          */
 888         if (!od->is_upper) {
 889                 realfile = READ_ONCE(od->upperfile);
 890                 if (!realfile) {
 891                         struct path upperpath;
 892
 893                         ovl_path_upper(dentry, &upperpath);
 894                         realfile = ovl_dir_open_realfile(file, &upperpath);
 895                         if (IS_ERR(realfile))
 896                                 return realfile;
 897
 898                         old = cmpxchg_release(&od->upperfile, NULL, realfile);
 899                         if (old) {
 900                                 fput(realfile);
 901                                 realfile = old;
 902                         }
 903                 }
 904         }
 905
 906         return realfile;
 907 }
 908
 909 static int ovl_dir_fsync(struct file *file, loff_t start, loff_t end,
 910                          int datasync)
 911 {
 912         struct file *realfile;
 913         int err;
 914
 915         err = ovl_sync_status(OVL_FS(file_inode(file)->i_sb));
 916         if (err <= 0)
 917                 return err;
 918
 919         realfile = ovl_dir_real_file(file, true);
 920         err = PTR_ERR_OR_ZERO(realfile);
 921
 922         /* Nothing to sync for lower */
 923         if (!realfile || err)
 924                 return err;
 925
 926         return vfs_fsync_range(realfile, start, end, datasync);
 927 }
 928
 929 static int ovl_dir_release(struct inode *inode, struct file *file)
 930 {
 931         struct ovl_dir_file *od = file->private_data;
 932
 933         if (od->cache) {
 934                 inode_lock(inode);
 935                 ovl_cache_put(od, inode);
 936                 inode_unlock(inode);
 937         }
 938         fput(od->realfile);
 939         if (od->upperfile)
 940                 fput(od->upperfile);
 941         kfree(od);
 942
 943         return 0;
 944 }
 945
 946 static int ovl_dir_open(struct inode *inode, struct file *file)
 947 {
 948         struct path realpath;
 949         struct file *realfile;
 950         struct ovl_dir_file *od;
 951         enum ovl_path_type type;
 952
 953         od = kzalloc(sizeof(struct ovl_dir_file), GFP_KERNEL);
 954         if (!od)
 955                 return -ENOMEM;
 956
 957         type = ovl_path_real(file->f_path.dentry, &realpath);
 958         realfile = ovl_dir_open_realfile(file, &realpath);
 959         if (IS_ERR(realfile)) {
 960                 kfree(od);
 961                 return PTR_ERR(realfile);
 962         }
 963         od->realfile = realfile;
 964         od->is_real = ovl_dir_is_real(inode);
 965         od->is_upper = OVL_TYPE_UPPER(type);
 966         file->private_data = od;
 967
 968         return 0;
 969 }
 970
 971 WRAP_DIR_ITER(ovl_iterate) // FIXME!
 972 const struct file_operations ovl_dir_operations = {
 973         .read           = generic_read_dir,
 974         .open           = ovl_dir_open,
 975         .iterate_shared = shared_ovl_iterate,
 976         .llseek         = ovl_dir_llseek,
 977         .fsync          = ovl_dir_fsync,
 978         .release        = ovl_dir_release,
 979 };
 980
 981 int ovl_check_empty_dir(struct dentry *dentry, struct list_head *list)
 982 {
 983         int err;
 984         struct ovl_cache_entry *p, *n;
 985         struct rb_root root = RB_ROOT;
 986         const struct cred *old_cred;
 987
 988         old_cred = ovl_override_creds(dentry->d_sb);
 989         err = ovl_dir_read_merged(dentry, list, &root);
 990         ovl_revert_creds(old_cred);
 991         if (err)
 992                 return err;
 993
 994         err = 0;
 995
 996         list_for_each_entry_safe(p, n, list, l_node) {
 997                 /*
 998                  * Select whiteouts in upperdir, they should
 999                  * be cleared when deleting this directory.
1000                  */
1001                 if (p->is_whiteout) {
1002                         if (p->is_upper)
1003                                 continue;
1004                         goto del_entry;
1005                 }
1006
1007                 if (p->name[0] == '.') {
1008                         if (p->len == 1)
1009                                 goto del_entry;
1010                         if (p->len == 2 && p->name[1] == '.')
1011                                 goto del_entry;
1012                 }
1013                 err = -ENOTEMPTY;
1014                 break;
1015
1016 del_entry:
1017                 list_del(&p->l_node);
1018                 kfree(p);
1019         }
1020
1021         return err;
1022 }
1023
1024 void ovl_cleanup_whiteouts(struct ovl_fs *ofs, struct dentry *upper,
1025                            struct list_head *list)
1026 {
1027         struct ovl_cache_entry *p;
1028
1029         inode_lock_nested(upper->d_inode, I_MUTEX_CHILD);
1030         list_for_each_entry(p, list, l_node) {
1031                 struct dentry *dentry;
1032
1033                 if (WARN_ON(!p->is_whiteout || !p->is_upper))
1034                         continue;
1035
1036                 dentry = ovl_lookup_upper(ofs, p->name, upper, p->len);
1037                 if (IS_ERR(dentry)) {
1038                         pr_err("lookup '%s/%.*s' failed (%i)\n",
1039                                upper->d_name.name, p->len, p->name,
1040                                (int) PTR_ERR(dentry));
1041                         continue;
1042                 }
1043                 if (dentry->d_inode)
1044                         ovl_cleanup(ofs, upper->d_inode, dentry);
1045                 dput(dentry);
1046         }
1047         inode_unlock(upper->d_inode);
1048 }
1049
1050 static bool ovl_check_d_type(struct dir_context *ctx, const char *name,
1051                           int namelen, loff_t offset, u64 ino,
1052                           unsigned int d_type)
1053 {
1054         struct ovl_readdir_data *rdd =
1055                 container_of(ctx, struct ovl_readdir_data, ctx);
1056
1057         /* Even if d_type is not supported, DT_DIR is returned for . and .. */
1058         if (!strncmp(name, ".", namelen) || !strncmp(name, "..", namelen))
1059                 return true;
1060
1061         if (d_type != DT_UNKNOWN)
1062                 rdd->d_type_supported = true;
1063
1064         return true;
1065 }
1066
1067 /*
1068  * Returns 1 if d_type is supported, 0 not supported/unknown. Negative values
1069  * if error is encountered.
1070  */
1071 int ovl_check_d_type_supported(const struct path *realpath)
1072 {
1073         int err;
1074         struct ovl_readdir_data rdd = {
1075                 .ctx.actor = ovl_check_d_type,
1076                 .d_type_supported = false,
1077         };
1078
1079         err = ovl_dir_read(realpath, &rdd);
1080         if (err)
1081                 return err;
1082
1083         return rdd.d_type_supported;
1084 }
1085
1086 #define OVL_INCOMPATDIR_NAME "incompat"
1087
1088 static int ovl_workdir_cleanup_recurse(struct ovl_fs *ofs, const struct path *path,
1089                                        int level)
1090 {
1091         int err;
1092         struct inode *dir = path->dentry->d_inode;
1093         LIST_HEAD(list);
1094         struct ovl_cache_entry *p;
1095         struct ovl_readdir_data rdd = {
1096                 .ctx.actor = ovl_fill_plain,
1097                 .list = &list,
1098         };
1099         bool incompat = false;
1100
1101         /*
1102          * The "work/incompat" directory is treated specially - if it is not
1103          * empty, instead of printing a generic error and mounting read-only,
1104          * we will error about incompat features and fail the mount.
1105          *
1106          * When called from ovl_indexdir_cleanup(), path->dentry->d_name.name
1107          * starts with '#'.
1108          */
1109         if (level == 2 &&
1110             !strcmp(path->dentry->d_name.name, OVL_INCOMPATDIR_NAME))
1111                 incompat = true;
1112
1113         err = ovl_dir_read(path, &rdd);
1114         if (err)
1115                 goto out;
1116
1117         inode_lock_nested(dir, I_MUTEX_PARENT);
1118         list_for_each_entry(p, &list, l_node) {
1119                 struct dentry *dentry;
1120
1121                 if (p->name[0] == '.') {
1122                         if (p->len == 1)
1123                                 continue;
1124                         if (p->len == 2 && p->name[1] == '.')
1125                                 continue;
1126                 } else if (incompat) {
1127                         pr_err("overlay with incompat feature '%s' cannot be mounted\n",
1128                                 p->name);
1129                         err = -EINVAL;
1130                         break;
1131                 }
1132                 dentry = ovl_lookup_upper(ofs, p->name, path->dentry, p->len);
1133                 if (IS_ERR(dentry))
1134                         continue;
1135                 if (dentry->d_inode)
1136                         err = ovl_workdir_cleanup(ofs, dir, path->mnt, dentry, level);
1137                 dput(dentry);
1138                 if (err)
1139                         break;
1140         }
1141         inode_unlock(dir);
1142 out:
1143         ovl_cache_free(&list);
1144         return err;
1145 }
1146
1147 int ovl_workdir_cleanup(struct ovl_fs *ofs, struct inode *dir,
1148                         struct vfsmount *mnt, struct dentry *dentry, int level)
1149 {
1150         int err;
1151
1152         if (!d_is_dir(dentry) || level > 1) {
1153                 return ovl_cleanup(ofs, dir, dentry);
1154         }
1155
1156         err = ovl_do_rmdir(ofs, dir, dentry);
1157         if (err) {
1158                 struct path path = { .mnt = mnt, .dentry = dentry };
1159
1160                 inode_unlock(dir);
1161                 err = ovl_workdir_cleanup_recurse(ofs, &path, level + 1);
1162                 inode_lock_nested(dir, I_MUTEX_PARENT);
1163                 if (!err)
1164                         err = ovl_cleanup(ofs, dir, dentry);
1165         }
1166
1167         return err;
1168 }
1169
1170 int ovl_indexdir_cleanup(struct ovl_fs *ofs)
1171 {
1172         int err;
1173         struct dentry *indexdir = ofs->workdir;
1174         struct dentry *index = NULL;
1175         struct inode *dir = indexdir->d_inode;
1176         struct path path = { .mnt = ovl_upper_mnt(ofs), .dentry = indexdir };
1177         LIST_HEAD(list);
1178         struct ovl_cache_entry *p;
1179         struct ovl_readdir_data rdd = {
1180                 .ctx.actor = ovl_fill_plain,
1181                 .list = &list,
1182         };
1183
1184         err = ovl_dir_read(&path, &rdd);
1185         if (err)
1186                 goto out;
1187
1188         inode_lock_nested(dir, I_MUTEX_PARENT);
1189         list_for_each_entry(p, &list, l_node) {
1190                 if (p->name[0] == '.') {
1191                         if (p->len == 1)
1192                                 continue;
1193                         if (p->len == 2 && p->name[1] == '.')
1194                                 continue;
1195                 }
1196                 index = ovl_lookup_upper(ofs, p->name, indexdir, p->len);
1197                 if (IS_ERR(index)) {
1198                         err = PTR_ERR(index);
1199                         index = NULL;
1200                         break;
1201                 }
1202                 /* Cleanup leftover from index create/cleanup attempt */
1203                 if (index->d_name.name[0] == '#') {
1204                         err = ovl_workdir_cleanup(ofs, dir, path.mnt, index, 1);
1205                         if (err)
1206                                 break;
1207                         goto next;
1208                 }
1209                 err = ovl_verify_index(ofs, index);
1210                 if (!err) {
1211                         goto next;
1212                 } else if (err == -ESTALE) {
1213                         /* Cleanup stale index entries */
1214                         err = ovl_cleanup(ofs, dir, index);
1215                 } else if (err != -ENOENT) {
1216                         /*
1217                          * Abort mount to avoid corrupting the index if
1218                          * an incompatible index entry was found or on out
1219                          * of memory.
1220                          */
1221                         break;
1222                 } else if (ofs->config.nfs_export) {
1223                         /*
1224                          * Whiteout orphan index to block future open by
1225                          * handle after overlay nlink dropped to zero.
1226                          */
1227                         err = ovl_cleanup_and_whiteout(ofs, dir, index);
1228                 } else {
1229                         /* Cleanup orphan index entries */
1230                         err = ovl_cleanup(ofs, dir, index);
1231                 }
1232
1233                 if (err)
1234                         break;
1235
1236 next:
1237                 dput(index);
1238                 index = NULL;
1239         }
1240         dput(index);
1241         inode_unlock(dir);
1242 out:
1243         ovl_cache_free(&list);
1244         if (err)
1245                 pr_err("failed index dir cleanup (%i)\n", err);
1246         return err;
1247 }