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 <linux/overflow.h>
  17#include "overlayfs.h"
  18
  19struct ovl_cache_entry {
  20	unsigned int len;
  21	unsigned int type;
  22	u64 real_ino;
  23	u64 ino;
  24	struct list_head l_node;
  25	struct rb_node node;
  26	struct ovl_cache_entry *next_maybe_whiteout;
  27	bool is_upper;
  28	bool is_whiteout;
  29	bool check_xwhiteout;
  30	const char *c_name;
  31	int c_len;
  32	char name[];
  33};
  34
  35struct ovl_dir_cache {
  36	long refcount;
  37	u64 version;
  38	struct list_head entries;
  39	struct rb_root root;
  40};
  41
  42struct ovl_readdir_data {
  43	struct dir_context ctx;
  44	struct dentry *dentry;
  45	bool is_lowest;
  46	struct rb_root *root;
  47	struct list_head *list;
  48	struct list_head middle;
  49	struct ovl_cache_entry *first_maybe_whiteout;
  50	struct unicode_map *map;
  51	int count;
  52	int err;
  53	bool is_upper;
  54	bool d_type_supported;
  55	bool in_xwhiteouts_dir;
  56};
  57
  58struct ovl_dir_file {
  59	bool is_real;
  60	bool is_upper;
  61	struct ovl_dir_cache *cache;
  62	struct list_head *cursor;
  63	struct file *realfile;
  64	struct file *upperfile;
  65};
  66
  67static struct ovl_cache_entry *ovl_cache_entry_from_node(struct rb_node *n)
  68{
  69	return rb_entry(n, struct ovl_cache_entry, node);
  70}
  71
  72static int ovl_casefold(struct ovl_readdir_data *rdd, const char *str, int len,
  73			char **dst)
  74{
  75	const struct qstr qstr = { .name = str, .len = len };
  76	char *cf_name;
  77	int cf_len;
  78
  79	if (!IS_ENABLED(CONFIG_UNICODE) || !rdd->map || is_dot_dotdot(str, len))
  80		return 0;
  81
  82	cf_name = kmalloc(NAME_MAX, GFP_KERNEL);
  83	if (!cf_name) {
  84		rdd->err = -ENOMEM;
  85		return -ENOMEM;
  86	}
  87
  88	cf_len = utf8_casefold(rdd->map, &qstr, cf_name, NAME_MAX);
  89	if (cf_len > 0)
  90		*dst = cf_name;
  91	else
  92		kfree(cf_name);
  93
  94	return cf_len;
  95}
  96
  97static bool ovl_cache_entry_find_link(const char *name, int len,
  98				      struct rb_node ***link,
  99				      struct rb_node **parent)
 100{
 101	bool found = false;
 102	struct rb_node **newp = *link;
 103
 104	while (!found && *newp) {
 105		int cmp;
 106		struct ovl_cache_entry *tmp;
 107
 108		*parent = *newp;
 109		tmp = ovl_cache_entry_from_node(*newp);
 110		cmp = strncmp(name, tmp->c_name, len);
 111		if (cmp > 0)
 112			newp = &tmp->node.rb_right;
 113		else if (cmp < 0 || len < tmp->c_len)
 114			newp = &tmp->node.rb_left;
 115		else
 116			found = true;
 117	}
 118	*link = newp;
 119
 120	return found;
 121}
 122
 123static struct ovl_cache_entry *ovl_cache_entry_find(struct rb_root *root,
 124						    const char *name, int len)
 125{
 126	struct rb_node *node = root->rb_node;
 127	int cmp;
 128
 129	while (node) {
 130		struct ovl_cache_entry *p = ovl_cache_entry_from_node(node);
 131
 132		cmp = strncmp(name, p->c_name, len);
 133		if (cmp > 0)
 134			node = p->node.rb_right;
 135		else if (cmp < 0 || len < p->c_len)
 136			node = p->node.rb_left;
 137		else
 138			return p;
 139	}
 140
 141	return NULL;
 142}
 143
 144static bool ovl_calc_d_ino(struct ovl_readdir_data *rdd,
 145			   struct ovl_cache_entry *p)
 146{
 147	/* Don't care if not doing ovl_iter() */
 148	if (!rdd->dentry)
 149		return false;
 150
 151	/* Always recalc d_ino when remapping lower inode numbers */
 152	if (ovl_xino_bits(OVL_FS(rdd->dentry->d_sb)))
 153		return true;
 154
 155	/* Always recalc d_ino for parent */
 156	if (strcmp(p->name, "..") == 0)
 157		return true;
 158
 159	/* If this is lower, then native d_ino will do */
 160	if (!rdd->is_upper)
 161		return false;
 162
 163	/*
 164	 * Recalc d_ino for '.' and for all entries if dir is impure (contains
 165	 * copied up entries)
 166	 */
 167	if ((p->name[0] == '.' && p->len == 1) ||
 168	    ovl_test_flag(OVL_IMPURE, d_inode(rdd->dentry)))
 169		return true;
 170
 171	return false;
 172}
 173
 174static struct ovl_cache_entry *ovl_cache_entry_new(struct ovl_readdir_data *rdd,
 175						   const char *name, int len,
 176						   const char *c_name, int c_len,
 177						   u64 ino, unsigned int d_type)
 178{
 179	struct ovl_cache_entry *p;
 180
 181	p = kmalloc(struct_size(p, name, len + 1), GFP_KERNEL);
 182	if (!p)
 183		return NULL;
 184
 185	memcpy(p->name, name, len);
 186	p->name[len] = '\0';
 187	p->len = len;
 188	p->type = d_type;
 189	p->real_ino = ino;
 190	p->ino = ino;
 191	/* Defer setting d_ino for upper entry to ovl_iterate() */
 192	if (ovl_calc_d_ino(rdd, p))
 193		p->ino = 0;
 194	p->is_upper = rdd->is_upper;
 195	p->is_whiteout = false;
 196	/* Defer check for overlay.whiteout to ovl_iterate() */
 197	p->check_xwhiteout = rdd->in_xwhiteouts_dir && d_type == DT_REG;
 198
 199	if (c_name && c_name != name) {
 200		p->c_name = c_name;
 201		p->c_len = c_len;
 202	} else {
 203		p->c_name = p->name;
 204		p->c_len = len;
 205	}
 206
 207	if (d_type == DT_CHR) {
 208		p->next_maybe_whiteout = rdd->first_maybe_whiteout;
 209		rdd->first_maybe_whiteout = p;
 210	}
 211	return p;
 212}
 213
 214/* Return 0 for found, 1 for added, <0 for error */
 215static int ovl_cache_entry_add_rb(struct ovl_readdir_data *rdd,
 216				  const char *name, int len,
 217				  const char *c_name, int c_len,
 218				  u64 ino,
 219				  unsigned int d_type)
 220{
 221	struct rb_node **newp = &rdd->root->rb_node;
 222	struct rb_node *parent = NULL;
 223	struct ovl_cache_entry *p;
 224
 225	if (ovl_cache_entry_find_link(c_name, c_len, &newp, &parent))
 226		return 0;
 227
 228	p = ovl_cache_entry_new(rdd, name, len, c_name, c_len, ino, d_type);
 229	if (p == NULL) {
 230		rdd->err = -ENOMEM;
 231		return -ENOMEM;
 232	}
 233
 234	list_add_tail(&p->l_node, rdd->list);
 235	rb_link_node(&p->node, parent, newp);
 236	rb_insert_color(&p->node, rdd->root);
 237
 238	return 1;
 239}
 240
 241/* Return 0 for found, 1 for added, <0 for error */
 242static int ovl_fill_lowest(struct ovl_readdir_data *rdd,
 243			   const char *name, int namelen,
 244			   const char *c_name, int c_len,
 245			   loff_t offset, u64 ino, unsigned int d_type)
 246{
 247	struct ovl_cache_entry *p;
 248
 249	p = ovl_cache_entry_find(rdd->root, c_name, c_len);
 250	if (p) {
 251		list_move_tail(&p->l_node, &rdd->middle);
 252		return 0;
 253	} else {
 254		p = ovl_cache_entry_new(rdd, name, namelen, c_name, c_len,
 255					ino, d_type);
 256		if (p == NULL)
 257			rdd->err = -ENOMEM;
 258		else
 259			list_add_tail(&p->l_node, &rdd->middle);
 260	}
 261
 262	return rdd->err ?: 1;
 263}
 264
 265static void ovl_cache_entry_free(struct ovl_cache_entry *p)
 266{
 267	if (p->c_name != p->name)
 268		kfree(p->c_name);
 269	kfree(p);
 270}
 271
 272void ovl_cache_free(struct list_head *list)
 273{
 274	struct ovl_cache_entry *p;
 275	struct ovl_cache_entry *n;
 276
 277	list_for_each_entry_safe(p, n, list, l_node)
 278		ovl_cache_entry_free(p);
 279
 280	INIT_LIST_HEAD(list);
 281}
 282
 283void ovl_dir_cache_free(struct inode *inode)
 284{
 285	struct ovl_dir_cache *cache = ovl_dir_cache(inode);
 286
 287	if (cache) {
 288		ovl_cache_free(&cache->entries);
 289		kfree(cache);
 290	}
 291}
 292
 293static void ovl_cache_put(struct ovl_dir_file *od, struct inode *inode)
 294{
 295	struct ovl_dir_cache *cache = od->cache;
 296
 297	WARN_ON(cache->refcount <= 0);
 298	cache->refcount--;
 299	if (!cache->refcount) {
 300		if (ovl_dir_cache(inode) == cache)
 301			ovl_set_dir_cache(inode, NULL);
 302
 303		ovl_cache_free(&cache->entries);
 304		kfree(cache);
 305	}
 306}
 307
 308static bool ovl_fill_merge(struct dir_context *ctx, const char *name,
 309			  int namelen, loff_t offset, u64 ino,
 310			  unsigned int d_type)
 311{
 312	struct ovl_readdir_data *rdd =
 313		container_of(ctx, struct ovl_readdir_data, ctx);
 314	struct ovl_fs *ofs = OVL_FS(rdd->dentry->d_sb);
 315	const char *c_name = NULL;
 316	char *cf_name = NULL;
 317	int c_len = 0, ret;
 318
 319	if (ofs->casefold)
 320		c_len = ovl_casefold(rdd, name, namelen, &cf_name);
 321
 322	if (rdd->err)
 323		return false;
 324
 325	if (c_len <= 0) {
 326		c_name = name;
 327		c_len = namelen;
 328	} else {
 329		c_name = cf_name;
 330	}
 331
 332	rdd->count++;
 333	if (!rdd->is_lowest)
 334		ret = ovl_cache_entry_add_rb(rdd, name, namelen, c_name, c_len, ino, d_type);
 335	else
 336		ret = ovl_fill_lowest(rdd, name, namelen, c_name, c_len, offset, ino, d_type);
 337
 338	/*
 339	 * If ret == 1, that means that c_name is being used as part of struct
 340	 * ovl_cache_entry and will be freed at ovl_cache_free(). Otherwise,
 341	 * c_name was found in the rb-tree so we can free it here.
 342	 */
 343	if (ret != 1 && c_name != name)
 344		kfree(c_name);
 345
 346	return ret >= 0;
 347}
 348
 349static int ovl_check_whiteouts(const struct path *path, struct ovl_readdir_data *rdd)
 350{
 351	struct dentry *dentry, *dir = path->dentry;
 352
 353	while (rdd->first_maybe_whiteout) {
 354		struct ovl_cache_entry *p =
 355			rdd->first_maybe_whiteout;
 356		rdd->first_maybe_whiteout = p->next_maybe_whiteout;
 357		dentry = lookup_one_positive_killable(mnt_idmap(path->mnt),
 358						      &QSTR_LEN(p->name, p->len),
 359						      dir);
 360		if (!IS_ERR(dentry)) {
 361			p->is_whiteout = ovl_is_whiteout(dentry);
 362			dput(dentry);
 363		} else if (PTR_ERR(dentry) == -EINTR) {
 364			return -EINTR;
 365		}
 366	}
 367
 368	return 0;
 369}
 370
 371static inline int ovl_dir_read(const struct path *realpath,
 372			       struct ovl_readdir_data *rdd)
 373{
 374	struct file *realfile;
 375	int err;
 376
 377	realfile = ovl_path_open(realpath, O_RDONLY | O_LARGEFILE);
 378	if (IS_ERR(realfile))
 379		return PTR_ERR(realfile);
 380
 381	rdd->first_maybe_whiteout = NULL;
 382	rdd->ctx.pos = 0;
 383	do {
 384		rdd->count = 0;
 385		rdd->err = 0;
 386		err = iterate_dir(realfile, &rdd->ctx);
 387		if (err >= 0)
 388			err = rdd->err;
 389	} while (!err && rdd->count);
 390
 391	if (!err && rdd->first_maybe_whiteout && rdd->dentry)
 392		err = ovl_check_whiteouts(realpath, rdd);
 393
 394	fput(realfile);
 395
 396	return err;
 397}
 398
 399static void ovl_dir_reset(struct file *file)
 400{
 401	struct ovl_dir_file *od = file->private_data;
 402	struct ovl_dir_cache *cache = od->cache;
 403	struct inode *inode = file_inode(file);
 404	bool is_real;
 405
 406	if (cache && ovl_inode_version_get(inode) != cache->version) {
 407		ovl_cache_put(od, inode);
 408		od->cache = NULL;
 409		od->cursor = NULL;
 410	}
 411	is_real = ovl_dir_is_real(inode);
 412	if (od->is_real != is_real) {
 413		/* is_real can only become false when dir is copied up */
 414		if (WARN_ON(is_real))
 415			return;
 416		od->is_real = false;
 417	}
 418}
 419
 420static int ovl_dir_read_merged(struct dentry *dentry, struct list_head *list,
 421	struct rb_root *root)
 422{
 423	int err;
 424	struct path realpath;
 425	struct ovl_readdir_data rdd = {
 426		.ctx.actor = ovl_fill_merge,
 427		.ctx.count = INT_MAX,
 428		.dentry = dentry,
 429		.list = list,
 430		.root = root,
 431		.is_lowest = false,
 432		.map = NULL,
 433	};
 434	int idx, next;
 435	const struct ovl_layer *layer;
 436	struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 437
 438	for (idx = 0; idx != -1; idx = next) {
 439		next = ovl_path_next(idx, dentry, &realpath, &layer);
 440
 441		if (ofs->casefold)
 442			rdd.map = sb_encoding(realpath.dentry->d_sb);
 443
 444		rdd.is_upper = ovl_dentry_upper(dentry) == realpath.dentry;
 445		rdd.in_xwhiteouts_dir = layer->has_xwhiteouts &&
 446					ovl_dentry_has_xwhiteouts(dentry);
 447
 448		if (next != -1) {
 449			err = ovl_dir_read(&realpath, &rdd);
 450			if (err)
 451				break;
 452		} else {
 453			/*
 454			 * Insert lowest layer entries before upper ones, this
 455			 * allows offsets to be reasonably constant
 456			 */
 457			list_add(&rdd.middle, rdd.list);
 458			rdd.is_lowest = true;
 459			err = ovl_dir_read(&realpath, &rdd);
 460			list_del(&rdd.middle);
 461		}
 462	}
 463	return err;
 464}
 465
 466static void ovl_seek_cursor(struct ovl_dir_file *od, loff_t pos)
 467{
 468	struct list_head *p;
 469	loff_t off = 0;
 470
 471	list_for_each(p, &od->cache->entries) {
 472		if (off >= pos)
 473			break;
 474		off++;
 475	}
 476	/* Cursor is safe since the cache is stable */
 477	od->cursor = p;
 478}
 479
 480static struct ovl_dir_cache *ovl_cache_get(struct dentry *dentry)
 481{
 482	int res;
 483	struct ovl_dir_cache *cache;
 484	struct inode *inode = d_inode(dentry);
 485
 486	cache = ovl_dir_cache(inode);
 487	if (cache && ovl_inode_version_get(inode) == cache->version) {
 488		WARN_ON(!cache->refcount);
 489		cache->refcount++;
 490		return cache;
 491	}
 492	ovl_set_dir_cache(d_inode(dentry), NULL);
 493
 494	cache = kzalloc(sizeof(struct ovl_dir_cache), GFP_KERNEL);
 495	if (!cache)
 496		return ERR_PTR(-ENOMEM);
 497
 498	cache->refcount = 1;
 499	INIT_LIST_HEAD(&cache->entries);
 500	cache->root = RB_ROOT;
 501
 502	res = ovl_dir_read_merged(dentry, &cache->entries, &cache->root);
 503	if (res) {
 504		ovl_cache_free(&cache->entries);
 505		kfree(cache);
 506		return ERR_PTR(res);
 507	}
 508
 509	cache->version = ovl_inode_version_get(inode);
 510	ovl_set_dir_cache(inode, cache);
 511
 512	return cache;
 513}
 514
 515/* Map inode number to lower fs unique range */
 516static u64 ovl_remap_lower_ino(u64 ino, int xinobits, int fsid,
 517			       const char *name, int namelen, bool warn)
 518{
 519	unsigned int xinoshift = 64 - xinobits;
 520
 521	if (unlikely(ino >> xinoshift)) {
 522		if (warn) {
 523			pr_warn_ratelimited("d_ino too big (%.*s, ino=%llu, xinobits=%d)\n",
 524					    namelen, name, ino, xinobits);
 525		}
 526		return ino;
 527	}
 528
 529	/*
 530	 * The lowest xinobit is reserved for mapping the non-peresistent inode
 531	 * numbers range, but this range is only exposed via st_ino, not here.
 532	 */
 533	return ino | ((u64)fsid) << (xinoshift + 1);
 534}
 535
 536/*
 537 * Set d_ino for upper entries if needed. Non-upper entries should always report
 538 * the uppermost real inode ino and should not call this function.
 539 *
 540 * When not all layer are on same fs, report real ino also for upper.
 541 *
 542 * When all layers are on the same fs, and upper has a reference to
 543 * copy up origin, call vfs_getattr() on the overlay entry to make
 544 * sure that d_ino will be consistent with st_ino from stat(2).
 545 *
 546 * Also checks the overlay.whiteout xattr by doing a full lookup which will return
 547 * negative in this case.
 548 */
 549static int ovl_cache_update(const struct path *path, struct ovl_cache_entry *p, bool update_ino)
 550
 551{
 552	struct dentry *dir = path->dentry;
 553	struct ovl_fs *ofs = OVL_FS(dir->d_sb);
 554	struct dentry *this = NULL;
 555	enum ovl_path_type type;
 556	u64 ino = p->real_ino;
 557	int xinobits = ovl_xino_bits(ofs);
 558	int err = 0;
 559
 560	if (!ovl_same_dev(ofs) && !p->check_xwhiteout)
 561		goto out;
 562
 563	if (p->name[0] == '.') {
 564		if (p->len == 1) {
 565			this = dget(dir);
 566			goto get;
 567		}
 568		if (p->len == 2 && p->name[1] == '.') {
 569			/* we shall not be moved */
 570			this = dget(dir->d_parent);
 571			goto get;
 572		}
 573	}
 574	/* This checks also for xwhiteouts */
 575	this = lookup_one(mnt_idmap(path->mnt), &QSTR_LEN(p->name, p->len), dir);
 576	if (IS_ERR_OR_NULL(this) || !this->d_inode) {
 577		/* Mark a stale entry */
 578		p->is_whiteout = true;
 579		if (IS_ERR(this)) {
 580			err = PTR_ERR(this);
 581			this = NULL;
 582			goto fail;
 583		}
 584		goto out;
 585	}
 586
 587get:
 588	if (!ovl_same_dev(ofs) || !update_ino)
 589		goto out;
 590
 591	type = ovl_path_type(this);
 592	if (OVL_TYPE_ORIGIN(type)) {
 593		struct kstat stat;
 594		struct path statpath = *path;
 595
 596		statpath.dentry = this;
 597		err = vfs_getattr(&statpath, &stat, STATX_INO, 0);
 598		if (err)
 599			goto fail;
 600
 601		/*
 602		 * Directory inode is always on overlay st_dev.
 603		 * Non-dir with ovl_same_dev() could be on pseudo st_dev in case
 604		 * of xino bits overflow.
 605		 */
 606		WARN_ON_ONCE(S_ISDIR(stat.mode) &&
 607			     dir->d_sb->s_dev != stat.dev);
 608		ino = stat.ino;
 609	} else if (xinobits && !OVL_TYPE_UPPER(type)) {
 610		ino = ovl_remap_lower_ino(ino, xinobits,
 611					  ovl_layer_lower(this)->fsid,
 612					  p->name, p->len,
 613					  ovl_xino_warn(ofs));
 614	}
 615
 616out:
 617	p->ino = ino;
 618	dput(this);
 619	return err;
 620
 621fail:
 622	pr_warn_ratelimited("failed to look up (%s) for ino (%i)\n",
 623			    p->name, err);
 624	goto out;
 625}
 626
 627static bool ovl_fill_plain(struct dir_context *ctx, const char *name,
 628			  int namelen, loff_t offset, u64 ino,
 629			  unsigned int d_type)
 630{
 631	struct ovl_cache_entry *p;
 632	struct ovl_readdir_data *rdd =
 633		container_of(ctx, struct ovl_readdir_data, ctx);
 634
 635	rdd->count++;
 636	p = ovl_cache_entry_new(rdd, name, namelen, NULL, 0, ino, d_type);
 637	if (p == NULL) {
 638		rdd->err = -ENOMEM;
 639		return false;
 640	}
 641	list_add_tail(&p->l_node, rdd->list);
 642
 643	return true;
 644}
 645
 646static int ovl_dir_read_impure(const struct path *path,  struct list_head *list,
 647			       struct rb_root *root)
 648{
 649	int err;
 650	struct path realpath;
 651	struct ovl_cache_entry *p, *n;
 652	struct ovl_readdir_data rdd = {
 653		.ctx.actor = ovl_fill_plain,
 654		.ctx.count = INT_MAX,
 655		.list = list,
 656		.root = root,
 657	};
 658
 659	INIT_LIST_HEAD(list);
 660	*root = RB_ROOT;
 661	ovl_path_upper(path->dentry, &realpath);
 662
 663	err = ovl_dir_read(&realpath, &rdd);
 664	if (err)
 665		return err;
 666
 667	list_for_each_entry_safe(p, n, list, l_node) {
 668		if (strcmp(p->name, ".") != 0 &&
 669		    strcmp(p->name, "..") != 0) {
 670			err = ovl_cache_update(path, p, true);
 671			if (err)
 672				return err;
 673		}
 674		if (p->ino == p->real_ino) {
 675			list_del(&p->l_node);
 676			ovl_cache_entry_free(p);
 677		} else {
 678			struct rb_node **newp = &root->rb_node;
 679			struct rb_node *parent = NULL;
 680
 681			if (WARN_ON(ovl_cache_entry_find_link(p->name, p->len,
 682							      &newp, &parent)))
 683				return -EIO;
 684
 685			rb_link_node(&p->node, parent, newp);
 686			rb_insert_color(&p->node, root);
 687		}
 688	}
 689	return 0;
 690}
 691
 692static struct ovl_dir_cache *ovl_cache_get_impure(const struct path *path)
 693{
 694	int res;
 695	struct dentry *dentry = path->dentry;
 696	struct inode *inode = d_inode(dentry);
 697	struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 698	struct ovl_dir_cache *cache;
 699
 700	cache = ovl_dir_cache(inode);
 701	if (cache && ovl_inode_version_get(inode) == cache->version)
 702		return cache;
 703
 704	/* Impure cache is not refcounted, free it here */
 705	ovl_dir_cache_free(inode);
 706	ovl_set_dir_cache(inode, NULL);
 707
 708	cache = kzalloc(sizeof(struct ovl_dir_cache), GFP_KERNEL);
 709	if (!cache)
 710		return ERR_PTR(-ENOMEM);
 711
 712	res = ovl_dir_read_impure(path, &cache->entries, &cache->root);
 713	if (res) {
 714		ovl_cache_free(&cache->entries);
 715		kfree(cache);
 716		return ERR_PTR(res);
 717	}
 718	if (list_empty(&cache->entries)) {
 719		/*
 720		 * A good opportunity to get rid of an unneeded "impure" flag.
 721		 * Removing the "impure" xattr is best effort.
 722		 */
 723		if (!ovl_want_write(dentry)) {
 724			ovl_removexattr(ofs, ovl_dentry_upper(dentry),
 725					OVL_XATTR_IMPURE);
 726			ovl_drop_write(dentry);
 727		}
 728		ovl_clear_flag(OVL_IMPURE, inode);
 729		kfree(cache);
 730		return NULL;
 731	}
 732
 733	cache->version = ovl_inode_version_get(inode);
 734	ovl_set_dir_cache(inode, cache);
 735
 736	return cache;
 737}
 738
 739struct ovl_readdir_translate {
 740	struct dir_context *orig_ctx;
 741	struct ovl_dir_cache *cache;
 742	struct dir_context ctx;
 743	u64 parent_ino;
 744	int fsid;
 745	int xinobits;
 746	bool xinowarn;
 747};
 748
 749static bool ovl_fill_real(struct dir_context *ctx, const char *name,
 750			   int namelen, loff_t offset, u64 ino,
 751			   unsigned int d_type)
 752{
 753	struct ovl_readdir_translate *rdt =
 754		container_of(ctx, struct ovl_readdir_translate, ctx);
 755	struct dir_context *orig_ctx = rdt->orig_ctx;
 756	bool res;
 757
 758	if (rdt->parent_ino && namelen == 2 && !strncmp(name, "..", 2)) {
 759		ino = rdt->parent_ino;
 760	} else if (rdt->cache) {
 761		struct ovl_cache_entry *p;
 762
 763		p = ovl_cache_entry_find(&rdt->cache->root, name, namelen);
 764		if (p)
 765			ino = p->ino;
 766	} else if (rdt->xinobits) {
 767		ino = ovl_remap_lower_ino(ino, rdt->xinobits, rdt->fsid,
 768					  name, namelen, rdt->xinowarn);
 769	}
 770
 771	res = orig_ctx->actor(orig_ctx, name, namelen, offset, ino, d_type);
 772	ctx->count = orig_ctx->count;
 773
 774	return res;
 775}
 776
 777static bool ovl_is_impure_dir(struct file *file)
 778{
 779	struct ovl_dir_file *od = file->private_data;
 780	struct inode *dir = file_inode(file);
 781
 782	/*
 783	 * Only upper dir can be impure, but if we are in the middle of
 784	 * iterating a lower real dir, dir could be copied up and marked
 785	 * impure. We only want the impure cache if we started iterating
 786	 * a real upper dir to begin with.
 787	 */
 788	return od->is_upper && ovl_test_flag(OVL_IMPURE, dir);
 789
 790}
 791
 792static int ovl_iterate_real(struct file *file, struct dir_context *ctx)
 793{
 794	int err;
 795	struct ovl_dir_file *od = file->private_data;
 796	struct dentry *dir = file->f_path.dentry;
 797	struct ovl_fs *ofs = OVL_FS(dir->d_sb);
 798	const struct ovl_layer *lower_layer = ovl_layer_lower(dir);
 799	struct ovl_readdir_translate rdt = {
 800		.ctx.actor = ovl_fill_real,
 801		.ctx.count = ctx->count,
 802		.orig_ctx = ctx,
 803		.xinobits = ovl_xino_bits(ofs),
 804		.xinowarn = ovl_xino_warn(ofs),
 805	};
 806
 807	if (rdt.xinobits && lower_layer)
 808		rdt.fsid = lower_layer->fsid;
 809
 810	if (OVL_TYPE_MERGE(ovl_path_type(dir->d_parent))) {
 811		struct kstat stat;
 812		struct path statpath = file->f_path;
 813
 814		statpath.dentry = dir->d_parent;
 815		err = vfs_getattr(&statpath, &stat, STATX_INO, 0);
 816		if (err)
 817			return err;
 818
 819		WARN_ON_ONCE(dir->d_sb->s_dev != stat.dev);
 820		rdt.parent_ino = stat.ino;
 821	}
 822
 823	if (ovl_is_impure_dir(file)) {
 824		rdt.cache = ovl_cache_get_impure(&file->f_path);
 825		if (IS_ERR(rdt.cache))
 826			return PTR_ERR(rdt.cache);
 827	}
 828
 829	err = iterate_dir(od->realfile, &rdt.ctx);
 830	ctx->pos = rdt.ctx.pos;
 831
 832	return err;
 833}
 834
 835static int ovl_iterate_merged(struct file *file, struct dir_context *ctx)
 836{
 837	struct ovl_dir_file *od = file->private_data;
 838	struct dentry *dentry = file->f_path.dentry;
 839	struct ovl_cache_entry *p;
 840	int err = 0;
 841
 842	if (!od->cache) {
 843		struct ovl_dir_cache *cache;
 844
 845		cache = ovl_cache_get(dentry);
 846		err = PTR_ERR(cache);
 847		if (IS_ERR(cache))
 848			return err;
 849
 850		od->cache = cache;
 851		ovl_seek_cursor(od, ctx->pos);
 852	}
 853
 854	while (od->cursor != &od->cache->entries) {
 855		p = list_entry(od->cursor, struct ovl_cache_entry, l_node);
 856		if (!p->is_whiteout) {
 857			if (!p->ino || p->check_xwhiteout) {
 858				err = ovl_cache_update(&file->f_path, p, !p->ino);
 859				if (err)
 860					return err;
 861			}
 862		}
 863		/* ovl_cache_update() sets is_whiteout on stale entry */
 864		if (!p->is_whiteout) {
 865			if (!dir_emit(ctx, p->name, p->len, p->ino, p->type))
 866				break;
 867		}
 868		od->cursor = p->l_node.next;
 869		ctx->pos++;
 870	}
 871	return err;
 872}
 873
 874static bool ovl_need_adjust_d_ino(struct file *file)
 875{
 876	struct dentry *dentry = file->f_path.dentry;
 877	struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
 878
 879	/* If parent is merge, then need to adjust d_ino for '..' */
 880	if (ovl_xino_bits(ofs))
 881		return true;
 882
 883	/* Can't do consistent inode numbering */
 884	if (!ovl_same_fs(ofs))
 885		return false;
 886
 887	/* If dir is impure then need to adjust d_ino for copied up entries */
 888	if (ovl_is_impure_dir(file) ||
 889	    OVL_TYPE_MERGE(ovl_path_type(dentry->d_parent)))
 890		return true;
 891
 892	/* Pure: no need to adjust d_ino */
 893	return false;
 894}
 895
 896
 897static int ovl_iterate(struct file *file, struct dir_context *ctx)
 898{
 899	struct ovl_dir_file *od = file->private_data;
 900
 901	if (!ctx->pos)
 902		ovl_dir_reset(file);
 903
 904	with_ovl_creds(file_dentry(file)->d_sb) {
 905		if (!od->is_real)
 906			return ovl_iterate_merged(file, ctx);
 907
 908		if (ovl_need_adjust_d_ino(file))
 909			return ovl_iterate_real(file, ctx);
 910
 911		return iterate_dir(od->realfile, ctx);
 912	}
 913}
 914
 915static loff_t ovl_dir_llseek(struct file *file, loff_t offset, int origin)
 916{
 917	loff_t res;
 918	struct ovl_dir_file *od = file->private_data;
 919
 920	inode_lock(file_inode(file));
 921	if (!file->f_pos)
 922		ovl_dir_reset(file);
 923
 924	if (od->is_real) {
 925		res = vfs_llseek(od->realfile, offset, origin);
 926		file->f_pos = od->realfile->f_pos;
 927	} else {
 928		res = -EINVAL;
 929
 930		switch (origin) {
 931		case SEEK_CUR:
 932			offset += file->f_pos;
 933			break;
 934		case SEEK_SET:
 935			break;
 936		default:
 937			goto out_unlock;
 938		}
 939		if (offset < 0)
 940			goto out_unlock;
 941
 942		if (offset != file->f_pos) {
 943			file->f_pos = offset;
 944			if (od->cache)
 945				ovl_seek_cursor(od, offset);
 946		}
 947		res = offset;
 948	}
 949out_unlock:
 950	inode_unlock(file_inode(file));
 951
 952	return res;
 953}
 954
 955static struct file *ovl_dir_open_realfile(const struct file *file,
 956					  const struct path *realpath)
 957{
 958	with_ovl_creds(file_inode(file)->i_sb)
 959		return ovl_path_open(realpath, O_RDONLY | (file->f_flags & O_LARGEFILE));
 960}
 961
 962/*
 963 * Like ovl_real_fdget(), returns upperfile if dir was copied up since open.
 964 * Unlike ovl_real_fdget(), this caches upperfile in file->private_data.
 965 *
 966 * TODO: use same abstract type for file->private_data of dir and file so
 967 * upperfile could also be cached for files as well.
 968 */
 969struct file *ovl_dir_real_file(const struct file *file, bool want_upper)
 970{
 971
 972	struct ovl_dir_file *od = file->private_data;
 973	struct dentry *dentry = file->f_path.dentry;
 974	struct file *old, *realfile = od->realfile;
 975
 976	if (!OVL_TYPE_UPPER(ovl_path_type(dentry)))
 977		return want_upper ? NULL : realfile;
 978
 979	/*
 980	 * Need to check if we started out being a lower dir, but got copied up
 981	 */
 982	if (!od->is_upper) {
 983		realfile = READ_ONCE(od->upperfile);
 984		if (!realfile) {
 985			struct path upperpath;
 986
 987			ovl_path_upper(dentry, &upperpath);
 988			realfile = ovl_dir_open_realfile(file, &upperpath);
 989			if (IS_ERR(realfile))
 990				return realfile;
 991
 992			old = cmpxchg_release(&od->upperfile, NULL, realfile);
 993			if (old) {
 994				fput(realfile);
 995				realfile = old;
 996			}
 997		}
 998	}
 999
1000	return realfile;
1001}
1002
1003static int ovl_dir_fsync(struct file *file, loff_t start, loff_t end,
1004			 int datasync)
1005{
1006	struct file *realfile;
1007	int err;
1008
1009	err = ovl_sync_status(OVL_FS(file_inode(file)->i_sb));
1010	if (err <= 0)
1011		return err;
1012
1013	realfile = ovl_dir_real_file(file, true);
1014	err = PTR_ERR_OR_ZERO(realfile);
1015
1016	/* Nothing to sync for lower */
1017	if (!realfile || err)
1018		return err;
1019
1020	return vfs_fsync_range(realfile, start, end, datasync);
1021}
1022
1023static int ovl_dir_release(struct inode *inode, struct file *file)
1024{
1025	struct ovl_dir_file *od = file->private_data;
1026
1027	if (od->cache) {
1028		inode_lock(inode);
1029		ovl_cache_put(od, inode);
1030		inode_unlock(inode);
1031	}
1032	fput(od->realfile);
1033	if (od->upperfile)
1034		fput(od->upperfile);
1035	kfree(od);
1036
1037	return 0;
1038}
1039
1040static int ovl_dir_open(struct inode *inode, struct file *file)
1041{
1042	struct path realpath;
1043	struct file *realfile;
1044	struct ovl_dir_file *od;
1045	enum ovl_path_type type;
1046
1047	od = kzalloc(sizeof(struct ovl_dir_file), GFP_KERNEL);
1048	if (!od)
1049		return -ENOMEM;
1050
1051	type = ovl_path_real(file->f_path.dentry, &realpath);
1052	realfile = ovl_dir_open_realfile(file, &realpath);
1053	if (IS_ERR(realfile)) {
1054		kfree(od);
1055		return PTR_ERR(realfile);
1056	}
1057	od->realfile = realfile;
1058	od->is_real = ovl_dir_is_real(inode);
1059	od->is_upper = OVL_TYPE_UPPER(type);
1060	file->private_data = od;
1061
1062	return 0;
1063}
1064
1065WRAP_DIR_ITER(ovl_iterate) // FIXME!
1066const struct file_operations ovl_dir_operations = {
1067	.read		= generic_read_dir,
1068	.open		= ovl_dir_open,
1069	.iterate_shared	= shared_ovl_iterate,
1070	.llseek		= ovl_dir_llseek,
1071	.fsync		= ovl_dir_fsync,
1072	.release	= ovl_dir_release,
1073};
1074
1075int ovl_check_empty_dir(struct dentry *dentry, struct list_head *list)
1076{
1077	int err;
1078	struct ovl_cache_entry *p, *n;
1079	struct rb_root root = RB_ROOT;
1080
1081	with_ovl_creds(dentry->d_sb)
1082		err = ovl_dir_read_merged(dentry, list, &root);
1083	if (err)
1084		return err;
1085
1086	err = 0;
1087
1088	list_for_each_entry_safe(p, n, list, l_node) {
1089		/*
1090		 * Select whiteouts in upperdir, they should
1091		 * be cleared when deleting this directory.
1092		 */
1093		if (p->is_whiteout) {
1094			if (p->is_upper)
1095				continue;
1096			goto del_entry;
1097		}
1098
1099		if (p->name[0] == '.') {
1100			if (p->len == 1)
1101				goto del_entry;
1102			if (p->len == 2 && p->name[1] == '.')
1103				goto del_entry;
1104		}
1105		err = -ENOTEMPTY;
1106		break;
1107
1108del_entry:
1109		list_del(&p->l_node);
1110		ovl_cache_entry_free(p);
1111	}
1112
1113	return err;
1114}
1115
1116void ovl_cleanup_whiteouts(struct ovl_fs *ofs, struct dentry *upper,
1117			   struct list_head *list)
1118{
1119	struct ovl_cache_entry *p;
1120
1121	list_for_each_entry(p, list, l_node) {
1122		struct dentry *dentry;
1123
1124		if (WARN_ON(!p->is_whiteout || !p->is_upper))
1125			continue;
1126
1127		dentry = ovl_lookup_upper_unlocked(ofs, p->name, upper, p->len);
1128		if (IS_ERR(dentry)) {
1129			pr_err("lookup '%s/%.*s' failed (%i)\n",
1130			       upper->d_name.name, p->len, p->name,
1131			       (int) PTR_ERR(dentry));
1132			continue;
1133		}
1134		if (dentry->d_inode)
1135			ovl_cleanup(ofs, upper, dentry);
1136		dput(dentry);
1137	}
1138}
1139
1140static bool ovl_check_d_type(struct dir_context *ctx, const char *name,
1141			  int namelen, loff_t offset, u64 ino,
1142			  unsigned int d_type)
1143{
1144	struct ovl_readdir_data *rdd =
1145		container_of(ctx, struct ovl_readdir_data, ctx);
1146
1147	/* Even if d_type is not supported, DT_DIR is returned for . and .. */
1148	if (!strncmp(name, ".", namelen) || !strncmp(name, "..", namelen))
1149		return true;
1150
1151	if (d_type != DT_UNKNOWN)
1152		rdd->d_type_supported = true;
1153
1154	return true;
1155}
1156
1157/*
1158 * Returns 1 if d_type is supported, 0 not supported/unknown. Negative values
1159 * if error is encountered.
1160 */
1161int ovl_check_d_type_supported(const struct path *realpath)
1162{
1163	int err;
1164	struct ovl_readdir_data rdd = {
1165		.ctx.actor = ovl_check_d_type,
1166		.ctx.count = INT_MAX,
1167		.d_type_supported = false,
1168	};
1169
1170	err = ovl_dir_read(realpath, &rdd);
1171	if (err)
1172		return err;
1173
1174	return rdd.d_type_supported;
1175}
1176
1177#define OVL_INCOMPATDIR_NAME "incompat"
1178
1179static int ovl_workdir_cleanup_recurse(struct ovl_fs *ofs, const struct path *path,
1180				       int level)
1181{
1182	int err;
1183	LIST_HEAD(list);
1184	struct ovl_cache_entry *p;
1185	struct ovl_readdir_data rdd = {
1186		.ctx.actor = ovl_fill_plain,
1187		.ctx.count = INT_MAX,
1188		.list = &list,
1189	};
1190	bool incompat = false;
1191
1192	/*
1193	 * The "work/incompat" directory is treated specially - if it is not
1194	 * empty, instead of printing a generic error and mounting read-only,
1195	 * we will error about incompat features and fail the mount.
1196	 *
1197	 * When called from ovl_indexdir_cleanup(), path->dentry->d_name.name
1198	 * starts with '#'.
1199	 */
1200	if (level == 2 &&
1201	    !strcmp(path->dentry->d_name.name, OVL_INCOMPATDIR_NAME))
1202		incompat = true;
1203
1204	err = ovl_dir_read(path, &rdd);
1205	if (err)
1206		goto out;
1207
1208	list_for_each_entry(p, &list, l_node) {
1209		struct dentry *dentry;
1210
1211		if (p->name[0] == '.') {
1212			if (p->len == 1)
1213				continue;
1214			if (p->len == 2 && p->name[1] == '.')
1215				continue;
1216		} else if (incompat) {
1217			pr_err("overlay with incompat feature '%s' cannot be mounted\n",
1218				p->name);
1219			err = -EINVAL;
1220			break;
1221		}
1222		dentry = ovl_lookup_upper_unlocked(ofs, p->name, path->dentry, p->len);
1223		if (IS_ERR(dentry))
1224			continue;
1225		if (dentry->d_inode)
1226			err = ovl_workdir_cleanup(ofs, path->dentry, path->mnt,
1227						  dentry, level);
1228		dput(dentry);
1229		if (err)
1230			break;
1231	}
1232out:
1233	ovl_cache_free(&list);
1234	return err;
1235}
1236
1237int ovl_workdir_cleanup(struct ovl_fs *ofs, struct dentry *parent,
1238			struct vfsmount *mnt, struct dentry *dentry, int level)
1239{
1240	int err;
1241
1242	if (!d_is_dir(dentry) || level > 1)
1243		return ovl_cleanup(ofs, parent, dentry);
1244
1245	dentry = start_removing_dentry(parent, dentry);
1246	if (IS_ERR(dentry))
1247		return PTR_ERR(dentry);
1248	err = ovl_do_rmdir(ofs, parent->d_inode, dentry);
1249	end_removing(dentry);
1250	if (err) {
1251		struct path path = { .mnt = mnt, .dentry = dentry };
1252
1253		err = ovl_workdir_cleanup_recurse(ofs, &path, level + 1);
1254		if (!err)
1255			err = ovl_cleanup(ofs, parent, dentry);
1256	}
1257
1258	return err;
1259}
1260
1261int ovl_indexdir_cleanup(struct ovl_fs *ofs)
1262{
1263	int err;
1264	struct dentry *indexdir = ofs->workdir;
1265	struct dentry *index = NULL;
1266	struct path path = { .mnt = ovl_upper_mnt(ofs), .dentry = indexdir };
1267	LIST_HEAD(list);
1268	struct ovl_cache_entry *p;
1269	struct ovl_readdir_data rdd = {
1270		.ctx.actor = ovl_fill_plain,
1271		.ctx.count = INT_MAX,
1272		.list = &list,
1273	};
1274
1275	err = ovl_dir_read(&path, &rdd);
1276	if (err)
1277		goto out;
1278
1279	list_for_each_entry(p, &list, l_node) {
1280		if (p->name[0] == '.') {
1281			if (p->len == 1)
1282				continue;
1283			if (p->len == 2 && p->name[1] == '.')
1284				continue;
1285		}
1286		index = ovl_lookup_upper_unlocked(ofs, p->name, indexdir, p->len);
1287		if (IS_ERR(index)) {
1288			err = PTR_ERR(index);
1289			index = NULL;
1290			break;
1291		}
1292		/* Cleanup leftover from index create/cleanup attempt */
1293		if (index->d_name.name[0] == '#') {
1294			err = ovl_workdir_cleanup(ofs, indexdir, path.mnt, index, 1);
1295			if (err)
1296				break;
1297			goto next;
1298		}
1299		err = ovl_verify_index(ofs, index);
1300		if (!err) {
1301			goto next;
1302		} else if (err == -ESTALE) {
1303			/* Cleanup stale index entries */
1304			err = ovl_cleanup(ofs, indexdir, index);
1305		} else if (err != -ENOENT) {
1306			/*
1307			 * Abort mount to avoid corrupting the index if
1308			 * an incompatible index entry was found or on out
1309			 * of memory.
1310			 */
1311			break;
1312		} else if (ofs->config.nfs_export) {
1313			/*
1314			 * Whiteout orphan index to block future open by
1315			 * handle after overlay nlink dropped to zero.
1316			 */
1317			err = ovl_cleanup_and_whiteout(ofs, indexdir, index);
1318		} else {
1319			/* Cleanup orphan index entries */
1320			err = ovl_cleanup(ofs, indexdir, index);
1321		}
1322
1323		if (err)
1324			break;
1325
1326next:
1327		dput(index);
1328		index = NULL;
1329	}
1330	dput(index);
1331out:
1332	ovl_cache_free(&list);
1333	if (err)
1334		pr_err("failed index dir cleanup (%i)\n", err);
1335	return err;
1336}