1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * dir.c
5 *
6 * Creates, reads, walks and deletes directory-nodes
7 *
8 * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9 *
10 *  Portions of this code from linux/fs/ext3/dir.c
11 *
12 *  Copyright (C) 1992, 1993, 1994, 1995
13 *  Remy Card (card@masi.ibp.fr)
14 *  Laboratoire MASI - Institut Blaise pascal
15 *  Universite Pierre et Marie Curie (Paris VI)
16 *
17 *   from
18 *
19 *   linux/fs/minix/dir.c
20 *
21 *   Copyright (C) 1991, 1992 Linus Torvalds
22 *
23 * This program is free software; you can redistribute it and/or
24 * modify it under the terms of the GNU General Public
25 * License as published by the Free Software Foundation; either
26 * version 2 of the License, or (at your option) any later version.
27 *
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
31 * General Public License for more details.
32 *
33 * You should have received a copy of the GNU General Public
34 * License along with this program; if not, write to the
35 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
36 * Boston, MA 021110-1307, USA.
37 */
38
39#include <linux/fs.h>
40#include <linux/types.h>
41#include <linux/slab.h>
42#include <linux/highmem.h>
43#include <linux/quotaops.h>
44#include <linux/sort.h>
45
46#include <cluster/masklog.h>
47
48#include "ocfs2.h"
49
50#include "alloc.h"
51#include "blockcheck.h"
52#include "dir.h"
53#include "dlmglue.h"
54#include "extent_map.h"
55#include "file.h"
56#include "inode.h"
57#include "journal.h"
58#include "namei.h"
59#include "suballoc.h"
60#include "super.h"
61#include "sysfile.h"
62#include "uptodate.h"
63#include "ocfs2_trace.h"
64
65#include "buffer_head_io.h"
66
67#define NAMEI_RA_CHUNKS  2
68#define NAMEI_RA_BLOCKS  4
69#define NAMEI_RA_SIZE        (NAMEI_RA_CHUNKS * NAMEI_RA_BLOCKS)
70
71static unsigned char ocfs2_filetype_table[] = {
72	DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
73};
74
75static int ocfs2_do_extend_dir(struct super_block *sb,
76			       handle_t *handle,
77			       struct inode *dir,
78			       struct buffer_head *parent_fe_bh,
79			       struct ocfs2_alloc_context *data_ac,
80			       struct ocfs2_alloc_context *meta_ac,
81			       struct buffer_head **new_bh);
82static int ocfs2_dir_indexed(struct inode *inode);
83
84/*
85 * These are distinct checks because future versions of the file system will
86 * want to have a trailing dirent structure independent of indexing.
87 */
88static int ocfs2_supports_dir_trailer(struct inode *dir)
89{
90	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
91
92	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
93		return 0;
94
95	return ocfs2_meta_ecc(osb) || ocfs2_dir_indexed(dir);
96}
97
98/*
99 * "new' here refers to the point at which we're creating a new
100 * directory via "mkdir()", but also when we're expanding an inline
101 * directory. In either case, we don't yet have the indexing bit set
102 * on the directory, so the standard checks will fail in when metaecc
103 * is turned off. Only directory-initialization type functions should
104 * use this then. Everything else wants ocfs2_supports_dir_trailer()
105 */
106static int ocfs2_new_dir_wants_trailer(struct inode *dir)
107{
108	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
109
110	return ocfs2_meta_ecc(osb) ||
111		ocfs2_supports_indexed_dirs(osb);
112}
113
114static inline unsigned int ocfs2_dir_trailer_blk_off(struct super_block *sb)
115{
116	return sb->s_blocksize - sizeof(struct ocfs2_dir_block_trailer);
117}
118
119#define ocfs2_trailer_from_bh(_bh, _sb) ((struct ocfs2_dir_block_trailer *) ((_bh)->b_data + ocfs2_dir_trailer_blk_off((_sb))))
120
121/* XXX ocfs2_block_dqtrailer() is similar but not quite - can we make
122 * them more consistent? */
123struct ocfs2_dir_block_trailer *ocfs2_dir_trailer_from_size(int blocksize,
124							    void *data)
125{
126	char *p = data;
127
128	p += blocksize - sizeof(struct ocfs2_dir_block_trailer);
129	return (struct ocfs2_dir_block_trailer *)p;
130}
131
132/*
133 * XXX: This is executed once on every dirent. We should consider optimizing
134 * it.
135 */
136static int ocfs2_skip_dir_trailer(struct inode *dir,
137				  struct ocfs2_dir_entry *de,
138				  unsigned long offset,
139				  unsigned long blklen)
140{
141	unsigned long toff = blklen - sizeof(struct ocfs2_dir_block_trailer);
142
143	if (!ocfs2_supports_dir_trailer(dir))
144		return 0;
145
146	if (offset != toff)
147		return 0;
148
149	return 1;
150}
151
152static void ocfs2_init_dir_trailer(struct inode *inode,
153				   struct buffer_head *bh, u16 rec_len)
154{
155	struct ocfs2_dir_block_trailer *trailer;
156
157	trailer = ocfs2_trailer_from_bh(bh, inode->i_sb);
158	strcpy(trailer->db_signature, OCFS2_DIR_TRAILER_SIGNATURE);
159	trailer->db_compat_rec_len =
160			cpu_to_le16(sizeof(struct ocfs2_dir_block_trailer));
161	trailer->db_parent_dinode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
162	trailer->db_blkno = cpu_to_le64(bh->b_blocknr);
163	trailer->db_free_rec_len = cpu_to_le16(rec_len);
164}
165/*
166 * Link an unindexed block with a dir trailer structure into the index free
167 * list. This function will modify dirdata_bh, but assumes you've already
168 * passed it to the journal.
169 */
170static int ocfs2_dx_dir_link_trailer(struct inode *dir, handle_t *handle,
171				     struct buffer_head *dx_root_bh,
172				     struct buffer_head *dirdata_bh)
173{
174	int ret;
175	struct ocfs2_dx_root_block *dx_root;
176	struct ocfs2_dir_block_trailer *trailer;
177
178	ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
179				      OCFS2_JOURNAL_ACCESS_WRITE);
180	if (ret) {
181		mlog_errno(ret);
182		goto out;
183	}
184	trailer = ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
185	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
186
187	trailer->db_free_next = dx_root->dr_free_blk;
188	dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
189
190	ocfs2_journal_dirty(handle, dx_root_bh);
191
192out:
193	return ret;
194}
195
196static int ocfs2_free_list_at_root(struct ocfs2_dir_lookup_result *res)
197{
198	return res->dl_prev_leaf_bh == NULL;
199}
200
201void ocfs2_free_dir_lookup_result(struct ocfs2_dir_lookup_result *res)
202{
203	brelse(res->dl_dx_root_bh);
204	brelse(res->dl_leaf_bh);
205	brelse(res->dl_dx_leaf_bh);
206	brelse(res->dl_prev_leaf_bh);
207}
208
209static int ocfs2_dir_indexed(struct inode *inode)
210{
211	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INDEXED_DIR_FL)
212		return 1;
213	return 0;
214}
215
216static inline int ocfs2_dx_root_inline(struct ocfs2_dx_root_block *dx_root)
217{
218	return dx_root->dr_flags & OCFS2_DX_FLAG_INLINE;
219}
220
221/*
222 * Hashing code adapted from ext3
223 */
224#define DELTA 0x9E3779B9
225
226static void TEA_transform(__u32 buf[4], __u32 const in[])
227{
228	__u32	sum = 0;
229	__u32	b0 = buf[0], b1 = buf[1];
230	__u32	a = in[0], b = in[1], c = in[2], d = in[3];
231	int	n = 16;
232
233	do {
234		sum += DELTA;
235		b0 += ((b1 << 4)+a) ^ (b1+sum) ^ ((b1 >> 5)+b);
236		b1 += ((b0 << 4)+c) ^ (b0+sum) ^ ((b0 >> 5)+d);
237	} while (--n);
238
239	buf[0] += b0;
240	buf[1] += b1;
241}
242
243static void str2hashbuf(const char *msg, int len, __u32 *buf, int num)
244{
245	__u32	pad, val;
246	int	i;
247
248	pad = (__u32)len | ((__u32)len << 8);
249	pad |= pad << 16;
250
251	val = pad;
252	if (len > num*4)
253		len = num * 4;
254	for (i = 0; i < len; i++) {
255		if ((i % 4) == 0)
256			val = pad;
257		val = msg[i] + (val << 8);
258		if ((i % 4) == 3) {
259			*buf++ = val;
260			val = pad;
261			num--;
262		}
263	}
264	if (--num >= 0)
265		*buf++ = val;
266	while (--num >= 0)
267		*buf++ = pad;
268}
269
270static void ocfs2_dx_dir_name_hash(struct inode *dir, const char *name, int len,
271				   struct ocfs2_dx_hinfo *hinfo)
272{
273	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
274	const char	*p;
275	__u32		in[8], buf[4];
276
277	/*
278	 * XXX: Is this really necessary, if the index is never looked
279	 * at by readdir? Is a hash value of '0' a bad idea?
280	 */
281	if ((len == 1 && !strncmp(".", name, 1)) ||
282	    (len == 2 && !strncmp("..", name, 2))) {
283		buf[0] = buf[1] = 0;
284		goto out;
285	}
286
287#ifdef OCFS2_DEBUG_DX_DIRS
288	/*
289	 * This makes it very easy to debug indexing problems. We
290	 * should never allow this to be selected without hand editing
291	 * this file though.
292	 */
293	buf[0] = buf[1] = len;
294	goto out;
295#endif
296
297	memcpy(buf, osb->osb_dx_seed, sizeof(buf));
298
299	p = name;
300	while (len > 0) {
301		str2hashbuf(p, len, in, 4);
302		TEA_transform(buf, in);
303		len -= 16;
304		p += 16;
305	}
306
307out:
308	hinfo->major_hash = buf[0];
309	hinfo->minor_hash = buf[1];
310}
311
312/*
313 * bh passed here can be an inode block or a dir data block, depending
314 * on the inode inline data flag.
315 */
316static int ocfs2_check_dir_entry(struct inode * dir,
317				 struct ocfs2_dir_entry * de,
318				 struct buffer_head * bh,
319				 unsigned long offset)
320{
321	const char *error_msg = NULL;
322	const int rlen = le16_to_cpu(de->rec_len);
323
324	if (unlikely(rlen < OCFS2_DIR_REC_LEN(1)))
325		error_msg = "rec_len is smaller than minimal";
326	else if (unlikely(rlen % 4 != 0))
327		error_msg = "rec_len % 4 != 0";
328	else if (unlikely(rlen < OCFS2_DIR_REC_LEN(de->name_len)))
329		error_msg = "rec_len is too small for name_len";
330	else if (unlikely(
331		 ((char *) de - bh->b_data) + rlen > dir->i_sb->s_blocksize))
332		error_msg = "directory entry across blocks";
333
334	if (unlikely(error_msg != NULL))
335		mlog(ML_ERROR, "bad entry in directory #%llu: %s - "
336		     "offset=%lu, inode=%llu, rec_len=%d, name_len=%d\n",
337		     (unsigned long long)OCFS2_I(dir)->ip_blkno, error_msg,
338		     offset, (unsigned long long)le64_to_cpu(de->inode), rlen,
339		     de->name_len);
340
341	return error_msg == NULL ? 1 : 0;
342}
343
344static inline int ocfs2_match(int len,
345			      const char * const name,
346			      struct ocfs2_dir_entry *de)
347{
348	if (len != de->name_len)
349		return 0;
350	if (!de->inode)
351		return 0;
352	return !memcmp(name, de->name, len);
353}
354
355/*
356 * Returns 0 if not found, -1 on failure, and 1 on success
357 */
358static inline int ocfs2_search_dirblock(struct buffer_head *bh,
359					struct inode *dir,
360					const char *name, int namelen,
361					unsigned long offset,
362					char *first_de,
363					unsigned int bytes,
364					struct ocfs2_dir_entry **res_dir)
365{
366	struct ocfs2_dir_entry *de;
367	char *dlimit, *de_buf;
368	int de_len;
369	int ret = 0;
370
371	de_buf = first_de;
372	dlimit = de_buf + bytes;
373
374	while (de_buf < dlimit) {
375		/* this code is executed quadratically often */
376		/* do minimal checking `by hand' */
377
378		de = (struct ocfs2_dir_entry *) de_buf;
379
380		if (de_buf + namelen <= dlimit &&
381		    ocfs2_match(namelen, name, de)) {
382			/* found a match - just to be sure, do a full check */
383			if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
384				ret = -1;
385				goto bail;
386			}
387			*res_dir = de;
388			ret = 1;
389			goto bail;
390		}
391
392		/* prevent looping on a bad block */
393		de_len = le16_to_cpu(de->rec_len);
394		if (de_len <= 0) {
395			ret = -1;
396			goto bail;
397		}
398
399		de_buf += de_len;
400		offset += de_len;
401	}
402
403bail:
404	trace_ocfs2_search_dirblock(ret);
405	return ret;
406}
407
408static struct buffer_head *ocfs2_find_entry_id(const char *name,
409					       int namelen,
410					       struct inode *dir,
411					       struct ocfs2_dir_entry **res_dir)
412{
413	int ret, found;
414	struct buffer_head *di_bh = NULL;
415	struct ocfs2_dinode *di;
416	struct ocfs2_inline_data *data;
417
418	ret = ocfs2_read_inode_block(dir, &di_bh);
419	if (ret) {
420		mlog_errno(ret);
421		goto out;
422	}
423
424	di = (struct ocfs2_dinode *)di_bh->b_data;
425	data = &di->id2.i_data;
426
427	found = ocfs2_search_dirblock(di_bh, dir, name, namelen, 0,
428				      data->id_data, i_size_read(dir), res_dir);
429	if (found == 1)
430		return di_bh;
431
432	brelse(di_bh);
433out:
434	return NULL;
435}
436
437static int ocfs2_validate_dir_block(struct super_block *sb,
438				    struct buffer_head *bh)
439{
440	int rc;
441	struct ocfs2_dir_block_trailer *trailer =
442		ocfs2_trailer_from_bh(bh, sb);
443
444
445	/*
446	 * We don't validate dirents here, that's handled
447	 * in-place when the code walks them.
448	 */
449	trace_ocfs2_validate_dir_block((unsigned long long)bh->b_blocknr);
450
451	BUG_ON(!buffer_uptodate(bh));
452
453	/*
454	 * If the ecc fails, we return the error but otherwise
455	 * leave the filesystem running.  We know any error is
456	 * local to this block.
457	 *
458	 * Note that we are safe to call this even if the directory
459	 * doesn't have a trailer.  Filesystems without metaecc will do
460	 * nothing, and filesystems with it will have one.
461	 */
462	rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &trailer->db_check);
463	if (rc)
464		mlog(ML_ERROR, "Checksum failed for dinode %llu\n",
465		     (unsigned long long)bh->b_blocknr);
466
467	return rc;
468}
469
470/*
471 * Validate a directory trailer.
472 *
473 * We check the trailer here rather than in ocfs2_validate_dir_block()
474 * because that function doesn't have the inode to test.
475 */
476static int ocfs2_check_dir_trailer(struct inode *dir, struct buffer_head *bh)
477{
478	int rc = 0;
479	struct ocfs2_dir_block_trailer *trailer;
480
481	trailer = ocfs2_trailer_from_bh(bh, dir->i_sb);
482	if (!OCFS2_IS_VALID_DIR_TRAILER(trailer)) {
483		rc = -EINVAL;
484		ocfs2_error(dir->i_sb,
485			    "Invalid dirblock #%llu: "
486			    "signature = %.*s\n",
487			    (unsigned long long)bh->b_blocknr, 7,
488			    trailer->db_signature);
489		goto out;
490	}
491	if (le64_to_cpu(trailer->db_blkno) != bh->b_blocknr) {
492		rc = -EINVAL;
493		ocfs2_error(dir->i_sb,
494			    "Directory block #%llu has an invalid "
495			    "db_blkno of %llu",
496			    (unsigned long long)bh->b_blocknr,
497			    (unsigned long long)le64_to_cpu(trailer->db_blkno));
498		goto out;
499	}
500	if (le64_to_cpu(trailer->db_parent_dinode) !=
501	    OCFS2_I(dir)->ip_blkno) {
502		rc = -EINVAL;
503		ocfs2_error(dir->i_sb,
504			    "Directory block #%llu on dinode "
505			    "#%llu has an invalid parent_dinode "
506			    "of %llu",
507			    (unsigned long long)bh->b_blocknr,
508			    (unsigned long long)OCFS2_I(dir)->ip_blkno,
509			    (unsigned long long)le64_to_cpu(trailer->db_blkno));
510		goto out;
511	}
512out:
513	return rc;
514}
515
516/*
517 * This function forces all errors to -EIO for consistency with its
518 * predecessor, ocfs2_bread().  We haven't audited what returning the
519 * real error codes would do to callers.  We log the real codes with
520 * mlog_errno() before we squash them.
521 */
522static int ocfs2_read_dir_block(struct inode *inode, u64 v_block,
523				struct buffer_head **bh, int flags)
524{
525	int rc = 0;
526	struct buffer_head *tmp = *bh;
527
528	rc = ocfs2_read_virt_blocks(inode, v_block, 1, &tmp, flags,
529				    ocfs2_validate_dir_block);
530	if (rc) {
531		mlog_errno(rc);
532		goto out;
533	}
534
535	if (!(flags & OCFS2_BH_READAHEAD) &&
536	    ocfs2_supports_dir_trailer(inode)) {
537		rc = ocfs2_check_dir_trailer(inode, tmp);
538		if (rc) {
539			if (!*bh)
540				brelse(tmp);
541			mlog_errno(rc);
542			goto out;
543		}
544	}
545
546	/* If ocfs2_read_virt_blocks() got us a new bh, pass it up. */
547	if (!*bh)
548		*bh = tmp;
549
550out:
551	return rc ? -EIO : 0;
552}
553
554/*
555 * Read the block at 'phys' which belongs to this directory
556 * inode. This function does no virtual->physical block translation -
557 * what's passed in is assumed to be a valid directory block.
558 */
559static int ocfs2_read_dir_block_direct(struct inode *dir, u64 phys,
560				       struct buffer_head **bh)
561{
562	int ret;
563	struct buffer_head *tmp = *bh;
564
565	ret = ocfs2_read_block(INODE_CACHE(dir), phys, &tmp,
566			       ocfs2_validate_dir_block);
567	if (ret) {
568		mlog_errno(ret);
569		goto out;
570	}
571
572	if (ocfs2_supports_dir_trailer(dir)) {
573		ret = ocfs2_check_dir_trailer(dir, tmp);
574		if (ret) {
575			if (!*bh)
576				brelse(tmp);
577			mlog_errno(ret);
578			goto out;
579		}
580	}
581
582	if (!ret && !*bh)
583		*bh = tmp;
584out:
585	return ret;
586}
587
588static int ocfs2_validate_dx_root(struct super_block *sb,
589				  struct buffer_head *bh)
590{
591	int ret;
592	struct ocfs2_dx_root_block *dx_root;
593
594	BUG_ON(!buffer_uptodate(bh));
595
596	dx_root = (struct ocfs2_dx_root_block *) bh->b_data;
597
598	ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_root->dr_check);
599	if (ret) {
600		mlog(ML_ERROR,
601		     "Checksum failed for dir index root block %llu\n",
602		     (unsigned long long)bh->b_blocknr);
603		return ret;
604	}
605
606	if (!OCFS2_IS_VALID_DX_ROOT(dx_root)) {
607		ocfs2_error(sb,
608			    "Dir Index Root # %llu has bad signature %.*s",
609			    (unsigned long long)le64_to_cpu(dx_root->dr_blkno),
610			    7, dx_root->dr_signature);
611		return -EINVAL;
612	}
613
614	return 0;
615}
616
617static int ocfs2_read_dx_root(struct inode *dir, struct ocfs2_dinode *di,
618			      struct buffer_head **dx_root_bh)
619{
620	int ret;
621	u64 blkno = le64_to_cpu(di->i_dx_root);
622	struct buffer_head *tmp = *dx_root_bh;
623
624	ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
625			       ocfs2_validate_dx_root);
626
627	/* If ocfs2_read_block() got us a new bh, pass it up. */
628	if (!ret && !*dx_root_bh)
629		*dx_root_bh = tmp;
630
631	return ret;
632}
633
634static int ocfs2_validate_dx_leaf(struct super_block *sb,
635				  struct buffer_head *bh)
636{
637	int ret;
638	struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)bh->b_data;
639
640	BUG_ON(!buffer_uptodate(bh));
641
642	ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_leaf->dl_check);
643	if (ret) {
644		mlog(ML_ERROR,
645		     "Checksum failed for dir index leaf block %llu\n",
646		     (unsigned long long)bh->b_blocknr);
647		return ret;
648	}
649
650	if (!OCFS2_IS_VALID_DX_LEAF(dx_leaf)) {
651		ocfs2_error(sb, "Dir Index Leaf has bad signature %.*s",
652			    7, dx_leaf->dl_signature);
653		return -EROFS;
654	}
655
656	return 0;
657}
658
659static int ocfs2_read_dx_leaf(struct inode *dir, u64 blkno,
660			      struct buffer_head **dx_leaf_bh)
661{
662	int ret;
663	struct buffer_head *tmp = *dx_leaf_bh;
664
665	ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
666			       ocfs2_validate_dx_leaf);
667
668	/* If ocfs2_read_block() got us a new bh, pass it up. */
669	if (!ret && !*dx_leaf_bh)
670		*dx_leaf_bh = tmp;
671
672	return ret;
673}
674
675/*
676 * Read a series of dx_leaf blocks. This expects all buffer_head
677 * pointers to be NULL on function entry.
678 */
679static int ocfs2_read_dx_leaves(struct inode *dir, u64 start, int num,
680				struct buffer_head **dx_leaf_bhs)
681{
682	int ret;
683
684	ret = ocfs2_read_blocks(INODE_CACHE(dir), start, num, dx_leaf_bhs, 0,
685				ocfs2_validate_dx_leaf);
686	if (ret)
687		mlog_errno(ret);
688
689	return ret;
690}
691
692static struct buffer_head *ocfs2_find_entry_el(const char *name, int namelen,
693					       struct inode *dir,
694					       struct ocfs2_dir_entry **res_dir)
695{
696	struct super_block *sb;
697	struct buffer_head *bh_use[NAMEI_RA_SIZE];
698	struct buffer_head *bh, *ret = NULL;
699	unsigned long start, block, b;
700	int ra_max = 0;		/* Number of bh's in the readahead
701				   buffer, bh_use[] */
702	int ra_ptr = 0;		/* Current index into readahead
703				   buffer */
704	int num = 0;
705	int nblocks, i, err;
706
707	sb = dir->i_sb;
708
709	nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
710	start = OCFS2_I(dir)->ip_dir_start_lookup;
711	if (start >= nblocks)
712		start = 0;
713	block = start;
714
715restart:
716	do {
717		/*
718		 * We deal with the read-ahead logic here.
719		 */
720		if (ra_ptr >= ra_max) {
721			/* Refill the readahead buffer */
722			ra_ptr = 0;
723			b = block;
724			for (ra_max = 0; ra_max < NAMEI_RA_SIZE; ra_max++) {
725				/*
726				 * Terminate if we reach the end of the
727				 * directory and must wrap, or if our
728				 * search has finished at this block.
729				 */
730				if (b >= nblocks || (num && block == start)) {
731					bh_use[ra_max] = NULL;
732					break;
733				}
734				num++;
735
736				bh = NULL;
737				err = ocfs2_read_dir_block(dir, b++, &bh,
738							   OCFS2_BH_READAHEAD);
739				bh_use[ra_max] = bh;
740			}
741		}
742		if ((bh = bh_use[ra_ptr++]) == NULL)
743			goto next;
744		if (ocfs2_read_dir_block(dir, block, &bh, 0)) {
745			/* read error, skip block & hope for the best.
746			 * ocfs2_read_dir_block() has released the bh. */
747			mlog(ML_ERROR, "reading directory %llu, "
748				    "offset %lu\n",
749				    (unsigned long long)OCFS2_I(dir)->ip_blkno,
750				    block);
751			goto next;
752		}
753		i = ocfs2_search_dirblock(bh, dir, name, namelen,
754					  block << sb->s_blocksize_bits,
755					  bh->b_data, sb->s_blocksize,
756					  res_dir);
757		if (i == 1) {
758			OCFS2_I(dir)->ip_dir_start_lookup = block;
759			ret = bh;
760			goto cleanup_and_exit;
761		} else {
762			brelse(bh);
763			if (i < 0)
764				goto cleanup_and_exit;
765		}
766	next:
767		if (++block >= nblocks)
768			block = 0;
769	} while (block != start);
770
771	/*
772	 * If the directory has grown while we were searching, then
773	 * search the last part of the directory before giving up.
774	 */
775	block = nblocks;
776	nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
777	if (block < nblocks) {
778		start = 0;
779		goto restart;
780	}
781
782cleanup_and_exit:
783	/* Clean up the read-ahead blocks */
784	for (; ra_ptr < ra_max; ra_ptr++)
785		brelse(bh_use[ra_ptr]);
786
787	trace_ocfs2_find_entry_el(ret);
788	return ret;
789}
790
791static int ocfs2_dx_dir_lookup_rec(struct inode *inode,
792				   struct ocfs2_extent_list *el,
793				   u32 major_hash,
794				   u32 *ret_cpos,
795				   u64 *ret_phys_blkno,
796				   unsigned int *ret_clen)
797{
798	int ret = 0, i, found;
799	struct buffer_head *eb_bh = NULL;
800	struct ocfs2_extent_block *eb;
801	struct ocfs2_extent_rec *rec = NULL;
802
803	if (el->l_tree_depth) {
804		ret = ocfs2_find_leaf(INODE_CACHE(inode), el, major_hash,
805				      &eb_bh);
806		if (ret) {
807			mlog_errno(ret);
808			goto out;
809		}
810
811		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
812		el = &eb->h_list;
813
814		if (el->l_tree_depth) {
815			ocfs2_error(inode->i_sb,
816				    "Inode %lu has non zero tree depth in "
817				    "btree tree block %llu\n", inode->i_ino,
818				    (unsigned long long)eb_bh->b_blocknr);
819			ret = -EROFS;
820			goto out;
821		}
822	}
823
824	found = 0;
825	for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
826		rec = &el->l_recs[i];
827
828		if (le32_to_cpu(rec->e_cpos) <= major_hash) {
829			found = 1;
830			break;
831		}
832	}
833
834	if (!found) {
835		ocfs2_error(inode->i_sb, "Inode %lu has bad extent "
836			    "record (%u, %u, 0) in btree", inode->i_ino,
837			    le32_to_cpu(rec->e_cpos),
838			    ocfs2_rec_clusters(el, rec));
839		ret = -EROFS;
840		goto out;
841	}
842
843	if (ret_phys_blkno)
844		*ret_phys_blkno = le64_to_cpu(rec->e_blkno);
845	if (ret_cpos)
846		*ret_cpos = le32_to_cpu(rec->e_cpos);
847	if (ret_clen)
848		*ret_clen = le16_to_cpu(rec->e_leaf_clusters);
849
850out:
851	brelse(eb_bh);
852	return ret;
853}
854
855/*
856 * Returns the block index, from the start of the cluster which this
857 * hash belongs too.
858 */
859static inline unsigned int __ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
860						   u32 minor_hash)
861{
862	return minor_hash & osb->osb_dx_mask;
863}
864
865static inline unsigned int ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
866					  struct ocfs2_dx_hinfo *hinfo)
867{
868	return __ocfs2_dx_dir_hash_idx(osb, hinfo->minor_hash);
869}
870
871static int ocfs2_dx_dir_lookup(struct inode *inode,
872			       struct ocfs2_extent_list *el,
873			       struct ocfs2_dx_hinfo *hinfo,
874			       u32 *ret_cpos,
875			       u64 *ret_phys_blkno)
876{
877	int ret = 0;
878	unsigned int cend, uninitialized_var(clen);
879	u32 uninitialized_var(cpos);
880	u64 uninitialized_var(blkno);
881	u32 name_hash = hinfo->major_hash;
882
883	ret = ocfs2_dx_dir_lookup_rec(inode, el, name_hash, &cpos, &blkno,
884				      &clen);
885	if (ret) {
886		mlog_errno(ret);
887		goto out;
888	}
889
890	cend = cpos + clen;
891	if (name_hash >= cend) {
892		/* We want the last cluster */
893		blkno += ocfs2_clusters_to_blocks(inode->i_sb, clen - 1);
894		cpos += clen - 1;
895	} else {
896		blkno += ocfs2_clusters_to_blocks(inode->i_sb,
897						  name_hash - cpos);
898		cpos = name_hash;
899	}
900
901	/*
902	 * We now have the cluster which should hold our entry. To
903	 * find the exact block from the start of the cluster to
904	 * search, we take the lower bits of the hash.
905	 */
906	blkno += ocfs2_dx_dir_hash_idx(OCFS2_SB(inode->i_sb), hinfo);
907
908	if (ret_phys_blkno)
909		*ret_phys_blkno = blkno;
910	if (ret_cpos)
911		*ret_cpos = cpos;
912
913out:
914
915	return ret;
916}
917
918static int ocfs2_dx_dir_search(const char *name, int namelen,
919			       struct inode *dir,
920			       struct ocfs2_dx_root_block *dx_root,
921			       struct ocfs2_dir_lookup_result *res)
922{
923	int ret, i, found;
924	u64 uninitialized_var(phys);
925	struct buffer_head *dx_leaf_bh = NULL;
926	struct ocfs2_dx_leaf *dx_leaf;
927	struct ocfs2_dx_entry *dx_entry = NULL;
928	struct buffer_head *dir_ent_bh = NULL;
929	struct ocfs2_dir_entry *dir_ent = NULL;
930	struct ocfs2_dx_hinfo *hinfo = &res->dl_hinfo;
931	struct ocfs2_extent_list *dr_el;
932	struct ocfs2_dx_entry_list *entry_list;
933
934	ocfs2_dx_dir_name_hash(dir, name, namelen, &res->dl_hinfo);
935
936	if (ocfs2_dx_root_inline(dx_root)) {
937		entry_list = &dx_root->dr_entries;
938		goto search;
939	}
940
941	dr_el = &dx_root->dr_list;
942
943	ret = ocfs2_dx_dir_lookup(dir, dr_el, hinfo, NULL, &phys);
944	if (ret) {
945		mlog_errno(ret);
946		goto out;
947	}
948
949	trace_ocfs2_dx_dir_search((unsigned long long)OCFS2_I(dir)->ip_blkno,
950				  namelen, name, hinfo->major_hash,
951				  hinfo->minor_hash, (unsigned long long)phys);
952
953	ret = ocfs2_read_dx_leaf(dir, phys, &dx_leaf_bh);
954	if (ret) {
955		mlog_errno(ret);
956		goto out;
957	}
958
959	dx_leaf = (struct ocfs2_dx_leaf *) dx_leaf_bh->b_data;
960
961	trace_ocfs2_dx_dir_search_leaf_info(
962			le16_to_cpu(dx_leaf->dl_list.de_num_used),
963			le16_to_cpu(dx_leaf->dl_list.de_count));
964
965	entry_list = &dx_leaf->dl_list;
966
967search:
968	/*
969	 * Empty leaf is legal, so no need to check for that.
970	 */
971	found = 0;
972	for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
973		dx_entry = &entry_list->de_entries[i];
974
975		if (hinfo->major_hash != le32_to_cpu(dx_entry->dx_major_hash)
976		    || hinfo->minor_hash != le32_to_cpu(dx_entry->dx_minor_hash))
977			continue;
978
979		/*
980		 * Search unindexed leaf block now. We're not
981		 * guaranteed to find anything.
982		 */
983		ret = ocfs2_read_dir_block_direct(dir,
984					  le64_to_cpu(dx_entry->dx_dirent_blk),
985					  &dir_ent_bh);
986		if (ret) {
987			mlog_errno(ret);
988			goto out;
989		}
990
991		/*
992		 * XXX: We should check the unindexed block here,
993		 * before using it.
994		 */
995
996		found = ocfs2_search_dirblock(dir_ent_bh, dir, name, namelen,
997					      0, dir_ent_bh->b_data,
998					      dir->i_sb->s_blocksize, &dir_ent);
999		if (found == 1)
1000			break;
1001
1002		if (found == -1) {
1003			/* This means we found a bad directory entry. */
1004			ret = -EIO;
1005			mlog_errno(ret);
1006			goto out;
1007		}
1008
1009		brelse(dir_ent_bh);
1010		dir_ent_bh = NULL;
1011	}
1012
1013	if (found <= 0) {
1014		ret = -ENOENT;
1015		goto out;
1016	}
1017
1018	res->dl_leaf_bh = dir_ent_bh;
1019	res->dl_entry = dir_ent;
1020	res->dl_dx_leaf_bh = dx_leaf_bh;
1021	res->dl_dx_entry = dx_entry;
1022
1023	ret = 0;
1024out:
1025	if (ret) {
1026		brelse(dx_leaf_bh);
1027		brelse(dir_ent_bh);
1028	}
1029	return ret;
1030}
1031
1032static int ocfs2_find_entry_dx(const char *name, int namelen,
1033			       struct inode *dir,
1034			       struct ocfs2_dir_lookup_result *lookup)
1035{
1036	int ret;
1037	struct buffer_head *di_bh = NULL;
1038	struct ocfs2_dinode *di;
1039	struct buffer_head *dx_root_bh = NULL;
1040	struct ocfs2_dx_root_block *dx_root;
1041
1042	ret = ocfs2_read_inode_block(dir, &di_bh);
1043	if (ret) {
1044		mlog_errno(ret);
1045		goto out;
1046	}
1047
1048	di = (struct ocfs2_dinode *)di_bh->b_data;
1049
1050	ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
1051	if (ret) {
1052		mlog_errno(ret);
1053		goto out;
1054	}
1055	dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
1056
1057	ret = ocfs2_dx_dir_search(name, namelen, dir, dx_root, lookup);
1058	if (ret) {
1059		if (ret != -ENOENT)
1060			mlog_errno(ret);
1061		goto out;
1062	}
1063
1064	lookup->dl_dx_root_bh = dx_root_bh;
1065	dx_root_bh = NULL;
1066out:
1067	brelse(di_bh);
1068	brelse(dx_root_bh);
1069	return ret;
1070}
1071
1072/*
1073 * Try to find an entry of the provided name within 'dir'.
1074 *
1075 * If nothing was found, -ENOENT is returned. Otherwise, zero is
1076 * returned and the struct 'res' will contain information useful to
1077 * other directory manipulation functions.
1078 *
1079 * Caller can NOT assume anything about the contents of the
1080 * buffer_heads - they are passed back only so that it can be passed
1081 * into any one of the manipulation functions (add entry, delete
1082 * entry, etc). As an example, bh in the extent directory case is a
1083 * data block, in the inline-data case it actually points to an inode,
1084 * in the indexed directory case, multiple buffers are involved.
1085 */
1086int ocfs2_find_entry(const char *name, int namelen,
1087		     struct inode *dir, struct ocfs2_dir_lookup_result *lookup)
1088{
1089	struct buffer_head *bh;
1090	struct ocfs2_dir_entry *res_dir = NULL;
1091
1092	if (ocfs2_dir_indexed(dir))
1093		return ocfs2_find_entry_dx(name, namelen, dir, lookup);
1094
1095	/*
1096	 * The unindexed dir code only uses part of the lookup
1097	 * structure, so there's no reason to push it down further
1098	 * than this.
1099	 */
1100	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1101		bh = ocfs2_find_entry_id(name, namelen, dir, &res_dir);
1102	else
1103		bh = ocfs2_find_entry_el(name, namelen, dir, &res_dir);
1104
1105	if (bh == NULL)
1106		return -ENOENT;
1107
1108	lookup->dl_leaf_bh = bh;
1109	lookup->dl_entry = res_dir;
1110	return 0;
1111}
1112
1113/*
1114 * Update inode number and type of a previously found directory entry.
1115 */
1116int ocfs2_update_entry(struct inode *dir, handle_t *handle,
1117		       struct ocfs2_dir_lookup_result *res,
1118		       struct inode *new_entry_inode)
1119{
1120	int ret;
1121	ocfs2_journal_access_func access = ocfs2_journal_access_db;
1122	struct ocfs2_dir_entry *de = res->dl_entry;
1123	struct buffer_head *de_bh = res->dl_leaf_bh;
1124
1125	/*
1126	 * The same code works fine for both inline-data and extent
1127	 * based directories, so no need to split this up.  The only
1128	 * difference is the journal_access function.
1129	 */
1130
1131	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1132		access = ocfs2_journal_access_di;
1133
1134	ret = access(handle, INODE_CACHE(dir), de_bh,
1135		     OCFS2_JOURNAL_ACCESS_WRITE);
1136	if (ret) {
1137		mlog_errno(ret);
1138		goto out;
1139	}
1140
1141	de->inode = cpu_to_le64(OCFS2_I(new_entry_inode)->ip_blkno);
1142	ocfs2_set_de_type(de, new_entry_inode->i_mode);
1143
1144	ocfs2_journal_dirty(handle, de_bh);
1145
1146out:
1147	return ret;
1148}
1149
1150/*
1151 * __ocfs2_delete_entry deletes a directory entry by merging it with the
1152 * previous entry
1153 */
1154static int __ocfs2_delete_entry(handle_t *handle, struct inode *dir,
1155				struct ocfs2_dir_entry *de_del,
1156				struct buffer_head *bh, char *first_de,
1157				unsigned int bytes)
1158{
1159	struct ocfs2_dir_entry *de, *pde;
1160	int i, status = -ENOENT;
1161	ocfs2_journal_access_func access = ocfs2_journal_access_db;
1162
1163	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1164		access = ocfs2_journal_access_di;
1165
1166	i = 0;
1167	pde = NULL;
1168	de = (struct ocfs2_dir_entry *) first_de;
1169	while (i < bytes) {
1170		if (!ocfs2_check_dir_entry(dir, de, bh, i)) {
1171			status = -EIO;
1172			mlog_errno(status);
1173			goto bail;
1174		}
1175		if (de == de_del)  {
1176			status = access(handle, INODE_CACHE(dir), bh,
1177					OCFS2_JOURNAL_ACCESS_WRITE);
1178			if (status < 0) {
1179				status = -EIO;
1180				mlog_errno(status);
1181				goto bail;
1182			}
1183			if (pde)
1184				le16_add_cpu(&pde->rec_len,
1185						le16_to_cpu(de->rec_len));
1186			de->inode = 0;
1187			dir->i_version++;
1188			ocfs2_journal_dirty(handle, bh);
1189			goto bail;
1190		}
1191		i += le16_to_cpu(de->rec_len);
1192		pde = de;
1193		de = (struct ocfs2_dir_entry *)((char *)de + le16_to_cpu(de->rec_len));
1194	}
1195bail:
1196	return status;
1197}
1198
1199static unsigned int ocfs2_figure_dirent_hole(struct ocfs2_dir_entry *de)
1200{
1201	unsigned int hole;
1202
1203	if (le64_to_cpu(de->inode) == 0)
1204		hole = le16_to_cpu(de->rec_len);
1205	else
1206		hole = le16_to_cpu(de->rec_len) -
1207			OCFS2_DIR_REC_LEN(de->name_len);
1208
1209	return hole;
1210}
1211
1212static int ocfs2_find_max_rec_len(struct super_block *sb,
1213				  struct buffer_head *dirblock_bh)
1214{
1215	int size, this_hole, largest_hole = 0;
1216	char *trailer, *de_buf, *limit, *start = dirblock_bh->b_data;
1217	struct ocfs2_dir_entry *de;
1218
1219	trailer = (char *)ocfs2_trailer_from_bh(dirblock_bh, sb);
1220	size = ocfs2_dir_trailer_blk_off(sb);
1221	limit = start + size;
1222	de_buf = start;
1223	de = (struct ocfs2_dir_entry *)de_buf;
1224	do {
1225		if (de_buf != trailer) {
1226			this_hole = ocfs2_figure_dirent_hole(de);
1227			if (this_hole > largest_hole)
1228				largest_hole = this_hole;
1229		}
1230
1231		de_buf += le16_to_cpu(de->rec_len);
1232		de = (struct ocfs2_dir_entry *)de_buf;
1233	} while (de_buf < limit);
1234
1235	if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
1236		return largest_hole;
1237	return 0;
1238}
1239
1240static void ocfs2_dx_list_remove_entry(struct ocfs2_dx_entry_list *entry_list,
1241				       int index)
1242{
1243	int num_used = le16_to_cpu(entry_list->de_num_used);
1244
1245	if (num_used == 1 || index == (num_used - 1))
1246		goto clear;
1247
1248	memmove(&entry_list->de_entries[index],
1249		&entry_list->de_entries[index + 1],
1250		(num_used - index - 1)*sizeof(struct ocfs2_dx_entry));
1251clear:
1252	num_used--;
1253	memset(&entry_list->de_entries[num_used], 0,
1254	       sizeof(struct ocfs2_dx_entry));
1255	entry_list->de_num_used = cpu_to_le16(num_used);
1256}
1257
1258static int ocfs2_delete_entry_dx(handle_t *handle, struct inode *dir,
1259				 struct ocfs2_dir_lookup_result *lookup)
1260{
1261	int ret, index, max_rec_len, add_to_free_list = 0;
1262	struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1263	struct buffer_head *leaf_bh = lookup->dl_leaf_bh;
1264	struct ocfs2_dx_leaf *dx_leaf;
1265	struct ocfs2_dx_entry *dx_entry = lookup->dl_dx_entry;
1266	struct ocfs2_dir_block_trailer *trailer;
1267	struct ocfs2_dx_root_block *dx_root;
1268	struct ocfs2_dx_entry_list *entry_list;
1269
1270	/*
1271	 * This function gets a bit messy because we might have to
1272	 * modify the root block, regardless of whether the indexed
1273	 * entries are stored inline.
1274	 */
1275
1276	/*
1277	 * *Only* set 'entry_list' here, based on where we're looking
1278	 * for the indexed entries. Later, we might still want to
1279	 * journal both blocks, based on free list state.
1280	 */
1281	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
1282	if (ocfs2_dx_root_inline(dx_root)) {
1283		entry_list = &dx_root->dr_entries;
1284	} else {
1285		dx_leaf = (struct ocfs2_dx_leaf *) lookup->dl_dx_leaf_bh->b_data;
1286		entry_list = &dx_leaf->dl_list;
1287	}
1288
1289	/* Neither of these are a disk corruption - that should have
1290	 * been caught by lookup, before we got here. */
1291	BUG_ON(le16_to_cpu(entry_list->de_count) <= 0);
1292	BUG_ON(le16_to_cpu(entry_list->de_num_used) <= 0);
1293
1294	index = (char *)dx_entry - (char *)entry_list->de_entries;
1295	index /= sizeof(*dx_entry);
1296
1297	if (index >= le16_to_cpu(entry_list->de_num_used)) {
1298		mlog(ML_ERROR, "Dir %llu: Bad dx_entry ptr idx %d, (%p, %p)\n",
1299		     (unsigned long long)OCFS2_I(dir)->ip_blkno, index,
1300		     entry_list, dx_entry);
1301		return -EIO;
1302	}
1303
1304	/*
1305	 * We know that removal of this dirent will leave enough room
1306	 * for a new one, so add this block to the free list if it
1307	 * isn't already there.
1308	 */
1309	trailer = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
1310	if (trailer->db_free_rec_len == 0)
1311		add_to_free_list = 1;
1312
1313	/*
1314	 * Add the block holding our index into the journal before
1315	 * removing the unindexed entry. If we get an error return
1316	 * from __ocfs2_delete_entry(), then it hasn't removed the
1317	 * entry yet. Likewise, successful return means we *must*
1318	 * remove the indexed entry.
1319	 *
1320	 * We're also careful to journal the root tree block here as
1321	 * the entry count needs to be updated. Also, we might be
1322	 * adding to the start of the free list.
1323	 */
1324	ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1325				      OCFS2_JOURNAL_ACCESS_WRITE);
1326	if (ret) {
1327		mlog_errno(ret);
1328		goto out;
1329	}
1330
1331	if (!ocfs2_dx_root_inline(dx_root)) {
1332		ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
1333					      lookup->dl_dx_leaf_bh,
1334					      OCFS2_JOURNAL_ACCESS_WRITE);
1335		if (ret) {
1336			mlog_errno(ret);
1337			goto out;
1338		}
1339	}
1340
1341	trace_ocfs2_delete_entry_dx((unsigned long long)OCFS2_I(dir)->ip_blkno,
1342				    index);
1343
1344	ret = __ocfs2_delete_entry(handle, dir, lookup->dl_entry,
1345				   leaf_bh, leaf_bh->b_data, leaf_bh->b_size);
1346	if (ret) {
1347		mlog_errno(ret);
1348		goto out;
1349	}
1350
1351	max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, leaf_bh);
1352	trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1353	if (add_to_free_list) {
1354		trailer->db_free_next = dx_root->dr_free_blk;
1355		dx_root->dr_free_blk = cpu_to_le64(leaf_bh->b_blocknr);
1356		ocfs2_journal_dirty(handle, dx_root_bh);
1357	}
1358
1359	/* leaf_bh was journal_accessed for us in __ocfs2_delete_entry */
1360	ocfs2_journal_dirty(handle, leaf_bh);
1361
1362	le32_add_cpu(&dx_root->dr_num_entries, -1);
1363	ocfs2_journal_dirty(handle, dx_root_bh);
1364
1365	ocfs2_dx_list_remove_entry(entry_list, index);
1366
1367	if (!ocfs2_dx_root_inline(dx_root))
1368		ocfs2_journal_dirty(handle, lookup->dl_dx_leaf_bh);
1369
1370out:
1371	return ret;
1372}
1373
1374static inline int ocfs2_delete_entry_id(handle_t *handle,
1375					struct inode *dir,
1376					struct ocfs2_dir_entry *de_del,
1377					struct buffer_head *bh)
1378{
1379	int ret;
1380	struct buffer_head *di_bh = NULL;
1381	struct ocfs2_dinode *di;
1382	struct ocfs2_inline_data *data;
1383
1384	ret = ocfs2_read_inode_block(dir, &di_bh);
1385	if (ret) {
1386		mlog_errno(ret);
1387		goto out;
1388	}
1389
1390	di = (struct ocfs2_dinode *)di_bh->b_data;
1391	data = &di->id2.i_data;
1392
1393	ret = __ocfs2_delete_entry(handle, dir, de_del, bh, data->id_data,
1394				   i_size_read(dir));
1395
1396	brelse(di_bh);
1397out:
1398	return ret;
1399}
1400
1401static inline int ocfs2_delete_entry_el(handle_t *handle,
1402					struct inode *dir,
1403					struct ocfs2_dir_entry *de_del,
1404					struct buffer_head *bh)
1405{
1406	return __ocfs2_delete_entry(handle, dir, de_del, bh, bh->b_data,
1407				    bh->b_size);
1408}
1409
1410/*
1411 * Delete a directory entry. Hide the details of directory
1412 * implementation from the caller.
1413 */
1414int ocfs2_delete_entry(handle_t *handle,
1415		       struct inode *dir,
1416		       struct ocfs2_dir_lookup_result *res)
1417{
1418	if (ocfs2_dir_indexed(dir))
1419		return ocfs2_delete_entry_dx(handle, dir, res);
1420
1421	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1422		return ocfs2_delete_entry_id(handle, dir, res->dl_entry,
1423					     res->dl_leaf_bh);
1424
1425	return ocfs2_delete_entry_el(handle, dir, res->dl_entry,
1426				     res->dl_leaf_bh);
1427}
1428
1429/*
1430 * Check whether 'de' has enough room to hold an entry of
1431 * 'new_rec_len' bytes.
1432 */
1433static inline int ocfs2_dirent_would_fit(struct ocfs2_dir_entry *de,
1434					 unsigned int new_rec_len)
1435{
1436	unsigned int de_really_used;
1437
1438	/* Check whether this is an empty record with enough space */
1439	if (le64_to_cpu(de->inode) == 0 &&
1440	    le16_to_cpu(de->rec_len) >= new_rec_len)
1441		return 1;
1442
1443	/*
1444	 * Record might have free space at the end which we can
1445	 * use.
1446	 */
1447	de_really_used = OCFS2_DIR_REC_LEN(de->name_len);
1448	if (le16_to_cpu(de->rec_len) >= (de_really_used + new_rec_len))
1449	    return 1;
1450
1451	return 0;
1452}
1453
1454static void ocfs2_dx_dir_leaf_insert_tail(struct ocfs2_dx_leaf *dx_leaf,
1455					  struct ocfs2_dx_entry *dx_new_entry)
1456{
1457	int i;
1458
1459	i = le16_to_cpu(dx_leaf->dl_list.de_num_used);
1460	dx_leaf->dl_list.de_entries[i] = *dx_new_entry;
1461
1462	le16_add_cpu(&dx_leaf->dl_list.de_num_used, 1);
1463}
1464
1465static void ocfs2_dx_entry_list_insert(struct ocfs2_dx_entry_list *entry_list,
1466				       struct ocfs2_dx_hinfo *hinfo,
1467				       u64 dirent_blk)
1468{
1469	int i;
1470	struct ocfs2_dx_entry *dx_entry;
1471
1472	i = le16_to_cpu(entry_list->de_num_used);
1473	dx_entry = &entry_list->de_entries[i];
1474
1475	memset(dx_entry, 0, sizeof(*dx_entry));
1476	dx_entry->dx_major_hash = cpu_to_le32(hinfo->major_hash);
1477	dx_entry->dx_minor_hash = cpu_to_le32(hinfo->minor_hash);
1478	dx_entry->dx_dirent_blk = cpu_to_le64(dirent_blk);
1479
1480	le16_add_cpu(&entry_list->de_num_used, 1);
1481}
1482
1483static int __ocfs2_dx_dir_leaf_insert(struct inode *dir, handle_t *handle,
1484				      struct ocfs2_dx_hinfo *hinfo,
1485				      u64 dirent_blk,
1486				      struct buffer_head *dx_leaf_bh)
1487{
1488	int ret;
1489	struct ocfs2_dx_leaf *dx_leaf;
1490
1491	ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
1492				      OCFS2_JOURNAL_ACCESS_WRITE);
1493	if (ret) {
1494		mlog_errno(ret);
1495		goto out;
1496	}
1497
1498	dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
1499	ocfs2_dx_entry_list_insert(&dx_leaf->dl_list, hinfo, dirent_blk);
1500	ocfs2_journal_dirty(handle, dx_leaf_bh);
1501
1502out:
1503	return ret;
1504}
1505
1506static void ocfs2_dx_inline_root_insert(struct inode *dir, handle_t *handle,
1507					struct ocfs2_dx_hinfo *hinfo,
1508					u64 dirent_blk,
1509					struct ocfs2_dx_root_block *dx_root)
1510{
1511	ocfs2_dx_entry_list_insert(&dx_root->dr_entries, hinfo, dirent_blk);
1512}
1513
1514static int ocfs2_dx_dir_insert(struct inode *dir, handle_t *handle,
1515			       struct ocfs2_dir_lookup_result *lookup)
1516{
1517	int ret = 0;
1518	struct ocfs2_dx_root_block *dx_root;
1519	struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1520
1521	ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1522				      OCFS2_JOURNAL_ACCESS_WRITE);
1523	if (ret) {
1524		mlog_errno(ret);
1525		goto out;
1526	}
1527
1528	dx_root = (struct ocfs2_dx_root_block *)lookup->dl_dx_root_bh->b_data;
1529	if (ocfs2_dx_root_inline(dx_root)) {
1530		ocfs2_dx_inline_root_insert(dir, handle,
1531					    &lookup->dl_hinfo,
1532					    lookup->dl_leaf_bh->b_blocknr,
1533					    dx_root);
1534	} else {
1535		ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &lookup->dl_hinfo,
1536						 lookup->dl_leaf_bh->b_blocknr,
1537						 lookup->dl_dx_leaf_bh);
1538		if (ret)
1539			goto out;
1540	}
1541
1542	le32_add_cpu(&dx_root->dr_num_entries, 1);
1543	ocfs2_journal_dirty(handle, dx_root_bh);
1544
1545out:
1546	return ret;
1547}
1548
1549static void ocfs2_remove_block_from_free_list(struct inode *dir,
1550				       handle_t *handle,
1551				       struct ocfs2_dir_lookup_result *lookup)
1552{
1553	struct ocfs2_dir_block_trailer *trailer, *prev;
1554	struct ocfs2_dx_root_block *dx_root;
1555	struct buffer_head *bh;
1556
1557	trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1558
1559	if (ocfs2_free_list_at_root(lookup)) {
1560		bh = lookup->dl_dx_root_bh;
1561		dx_root = (struct ocfs2_dx_root_block *)bh->b_data;
1562		dx_root->dr_free_blk = trailer->db_free_next;
1563	} else {
1564		bh = lookup->dl_prev_leaf_bh;
1565		prev = ocfs2_trailer_from_bh(bh, dir->i_sb);
1566		prev->db_free_next = trailer->db_free_next;
1567	}
1568
1569	trailer->db_free_rec_len = cpu_to_le16(0);
1570	trailer->db_free_next = cpu_to_le64(0);
1571
1572	ocfs2_journal_dirty(handle, bh);
1573	ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1574}
1575
1576/*
1577 * This expects that a journal write has been reserved on
1578 * lookup->dl_prev_leaf_bh or lookup->dl_dx_root_bh
1579 */
1580static void ocfs2_recalc_free_list(struct inode *dir, handle_t *handle,
1581				   struct ocfs2_dir_lookup_result *lookup)
1582{
1583	int max_rec_len;
1584	struct ocfs2_dir_block_trailer *trailer;
1585
1586	/* Walk dl_leaf_bh to figure out what the new free rec_len is. */
1587	max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, lookup->dl_leaf_bh);
1588	if (max_rec_len) {
1589		/*
1590		 * There's still room in this block, so no need to remove it
1591		 * from the free list. In this case, we just want to update
1592		 * the rec len accounting.
1593		 */
1594		trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1595		trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1596		ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1597	} else {
1598		ocfs2_remove_block_from_free_list(dir, handle, lookup);
1599	}
1600}
1601
1602/* we don't always have a dentry for what we want to add, so people
1603 * like orphan dir can call this instead.
1604 *
1605 * The lookup context must have been filled from
1606 * ocfs2_prepare_dir_for_insert.
1607 */
1608int __ocfs2_add_entry(handle_t *handle,
1609		      struct inode *dir,
1610		      const char *name, int namelen,
1611		      struct inode *inode, u64 blkno,
1612		      struct buffer_head *parent_fe_bh,
1613		      struct ocfs2_dir_lookup_result *lookup)
1614{
1615	unsigned long offset;
1616	unsigned short rec_len;
1617	struct ocfs2_dir_entry *de, *de1;
1618	struct ocfs2_dinode *di = (struct ocfs2_dinode *)parent_fe_bh->b_data;
1619	struct super_block *sb = dir->i_sb;
1620	int retval, status;
1621	unsigned int size = sb->s_blocksize;
1622	struct buffer_head *insert_bh = lookup->dl_leaf_bh;
1623	char *data_start = insert_bh->b_data;
1624
1625	if (!namelen)
1626		return -EINVAL;
1627
1628	if (ocfs2_dir_indexed(dir)) {
1629		struct buffer_head *bh;
1630
1631		/*
1632		 * An indexed dir may require that we update the free space
1633		 * list. Reserve a write to the previous node in the list so
1634		 * that we don't fail later.
1635		 *
1636		 * XXX: This can be either a dx_root_block, or an unindexed
1637		 * directory tree leaf block.
1638		 */
1639		if (ocfs2_free_list_at_root(lookup)) {
1640			bh = lookup->dl_dx_root_bh;
1641			retval = ocfs2_journal_access_dr(handle,
1642						 INODE_CACHE(dir), bh,
1643						 OCFS2_JOURNAL_ACCESS_WRITE);
1644		} else {
1645			bh = lookup->dl_prev_leaf_bh;
1646			retval = ocfs2_journal_access_db(handle,
1647						 INODE_CACHE(dir), bh,
1648						 OCFS2_JOURNAL_ACCESS_WRITE);
1649		}
1650		if (retval) {
1651			mlog_errno(retval);
1652			return retval;
1653		}
1654	} else if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1655		data_start = di->id2.i_data.id_data;
1656		size = i_size_read(dir);
1657
1658		BUG_ON(insert_bh != parent_fe_bh);
1659	}
1660
1661	rec_len = OCFS2_DIR_REC_LEN(namelen);
1662	offset = 0;
1663	de = (struct ocfs2_dir_entry *) data_start;
1664	while (1) {
1665		BUG_ON((char *)de >= (size + data_start));
1666
1667		/* These checks should've already been passed by the
1668		 * prepare function, but I guess we can leave them
1669		 * here anyway. */
1670		if (!ocfs2_check_dir_entry(dir, de, insert_bh, offset)) {
1671			retval = -ENOENT;
1672			goto bail;
1673		}
1674		if (ocfs2_match(namelen, name, de)) {
1675			retval = -EEXIST;
1676			goto bail;
1677		}
1678
1679		/* We're guaranteed that we should have space, so we
1680		 * can't possibly have hit the trailer...right? */
1681		mlog_bug_on_msg(ocfs2_skip_dir_trailer(dir, de, offset, size),
1682				"Hit dir trailer trying to insert %.*s "
1683			        "(namelen %d) into directory %llu.  "
1684				"offset is %lu, trailer offset is %d\n",
1685				namelen, name, namelen,
1686				(unsigned long long)parent_fe_bh->b_blocknr,
1687				offset, ocfs2_dir_trailer_blk_off(dir->i_sb));
1688
1689		if (ocfs2_dirent_would_fit(de, rec_len)) {
1690			dir->i_mtime = dir->i_ctime = CURRENT_TIME;
1691			retval = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
1692			if (retval < 0) {
1693				mlog_errno(retval);
1694				goto bail;
1695			}
1696
1697			if (insert_bh == parent_fe_bh)
1698				status = ocfs2_journal_access_di(handle,
1699								 INODE_CACHE(dir),
1700								 insert_bh,
1701								 OCFS2_JOURNAL_ACCESS_WRITE);
1702			else {
1703				status = ocfs2_journal_access_db(handle,
1704								 INODE_CACHE(dir),
1705								 insert_bh,
1706					      OCFS2_JOURNAL_ACCESS_WRITE);
1707
1708				if (ocfs2_dir_indexed(dir)) {
1709					status = ocfs2_dx_dir_insert(dir,
1710								handle,
1711								lookup);
1712					if (status) {
1713						mlog_errno(status);
1714						goto bail;
1715					}
1716				}
1717			}
1718
1719			/* By now the buffer is marked for journaling */
1720			offset += le16_to_cpu(de->rec_len);
1721			if (le64_to_cpu(de->inode)) {
1722				de1 = (struct ocfs2_dir_entry *)((char *) de +
1723					OCFS2_DIR_REC_LEN(de->name_len));
1724				de1->rec_len =
1725					cpu_to_le16(le16_to_cpu(de->rec_len) -
1726					OCFS2_DIR_REC_LEN(de->name_len));
1727				de->rec_len = cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
1728				de = de1;
1729			}
1730			de->file_type = OCFS2_FT_UNKNOWN;
1731			if (blkno) {
1732				de->inode = cpu_to_le64(blkno);
1733				ocfs2_set_de_type(de, inode->i_mode);
1734			} else
1735				de->inode = 0;
1736			de->name_len = namelen;
1737			memcpy(de->name, name, namelen);
1738
1739			if (ocfs2_dir_indexed(dir))
1740				ocfs2_recalc_free_list(dir, handle, lookup);
1741
1742			dir->i_version++;
1743			ocfs2_journal_dirty(handle, insert_bh);
1744			retval = 0;
1745			goto bail;
1746		}
1747
1748		offset += le16_to_cpu(de->rec_len);
1749		de = (struct ocfs2_dir_entry *) ((char *) de + le16_to_cpu(de->rec_len));
1750	}
1751
1752	/* when you think about it, the assert above should prevent us
1753	 * from ever getting here. */
1754	retval = -ENOSPC;
1755bail:
1756	if (retval)
1757		mlog_errno(retval);
1758
1759	return retval;
1760}
1761
1762static int ocfs2_dir_foreach_blk_id(struct inode *inode,
1763				    u64 *f_version,
1764				    struct dir_context *ctx)
1765{
1766	int ret, i;
1767	unsigned long offset = ctx->pos;
1768	struct buffer_head *di_bh = NULL;
1769	struct ocfs2_dinode *di;
1770	struct ocfs2_inline_data *data;
1771	struct ocfs2_dir_entry *de;
1772
1773	ret = ocfs2_read_inode_block(inode, &di_bh);
1774	if (ret) {
1775		mlog(ML_ERROR, "Unable to read inode block for dir %llu\n",
1776		     (unsigned long long)OCFS2_I(inode)->ip_blkno);
1777		goto out;
1778	}
1779
1780	di = (struct ocfs2_dinode *)di_bh->b_data;
1781	data = &di->id2.i_data;
1782
1783	while (ctx->pos < i_size_read(inode)) {
1784		/* If the dir block has changed since the last call to
1785		 * readdir(2), then we might be pointing to an invalid
1786		 * dirent right now.  Scan from the start of the block
1787		 * to make sure. */
1788		if (*f_version != inode->i_version) {
1789			for (i = 0; i < i_size_read(inode) && i < offset; ) {
1790				de = (struct ocfs2_dir_entry *)
1791					(data->id_data + i);
1792				/* It's too expensive to do a full
1793				 * dirent test each time round this
1794				 * loop, but we do have to test at
1795				 * least that it is non-zero.  A
1796				 * failure will be detected in the
1797				 * dirent test below. */
1798				if (le16_to_cpu(de->rec_len) <
1799				    OCFS2_DIR_REC_LEN(1))
1800					break;
1801				i += le16_to_cpu(de->rec_len);
1802			}
1803			ctx->pos = offset = i;
1804			*f_version = inode->i_version;
1805		}
1806
1807		de = (struct ocfs2_dir_entry *) (data->id_data + ctx->pos);
1808		if (!ocfs2_check_dir_entry(inode, de, di_bh, ctx->pos)) {
1809			/* On error, skip the f_pos to the end. */
1810			ctx->pos = i_size_read(inode);
1811			break;
1812		}
1813		offset += le16_to_cpu(de->rec_len);
1814		if (le64_to_cpu(de->inode)) {
1815			unsigned char d_type = DT_UNKNOWN;
1816
1817			if (de->file_type < OCFS2_FT_MAX)
1818				d_type = ocfs2_filetype_table[de->file_type];
1819
1820			if (!dir_emit(ctx, de->name, de->name_len,
1821				      le64_to_cpu(de->inode), d_type))
1822				goto out;
1823		}
1824		ctx->pos += le16_to_cpu(de->rec_len);
1825	}
1826out:
1827	brelse(di_bh);
1828	return 0;
1829}
1830
1831/*
1832 * NOTE: This function can be called against unindexed directories,
1833 * and indexed ones.
1834 */
1835static int ocfs2_dir_foreach_blk_el(struct inode *inode,
1836				    u64 *f_version,
1837				    struct dir_context *ctx,
1838				    bool persist)
1839{
1840	unsigned long offset, blk, last_ra_blk = 0;
1841	int i;
1842	struct buffer_head * bh, * tmp;
1843	struct ocfs2_dir_entry * de;
1844	struct super_block * sb = inode->i_sb;
1845	unsigned int ra_sectors = 16;
1846	int stored = 0;
1847
1848	bh = NULL;
1849
1850	offset = ctx->pos & (sb->s_blocksize - 1);
1851
1852	while (ctx->pos < i_size_read(inode)) {
1853		blk = ctx->pos >> sb->s_blocksize_bits;
1854		if (ocfs2_read_dir_block(inode, blk, &bh, 0)) {
1855			/* Skip the corrupt dirblock and keep trying */
1856			ctx->pos += sb->s_blocksize - offset;
1857			continue;
1858		}
1859
1860		/* The idea here is to begin with 8k read-ahead and to stay
1861		 * 4k ahead of our current position.
1862		 *
1863		 * TODO: Use the pagecache for this. We just need to
1864		 * make sure it's cluster-safe... */
1865		if (!last_ra_blk
1866		    || (((last_ra_blk - blk) << 9) <= (ra_sectors / 2))) {
1867			for (i = ra_sectors >> (sb->s_blocksize_bits - 9);
1868			     i > 0; i--) {
1869				tmp = NULL;
1870				if (!ocfs2_read_dir_block(inode, ++blk, &tmp,
1871							  OCFS2_BH_READAHEAD))
1872					brelse(tmp);
1873			}
1874			last_ra_blk = blk;
1875			ra_sectors = 8;
1876		}
1877
1878		/* If the dir block has changed since the last call to
1879		 * readdir(2), then we might be pointing to an invalid
1880		 * dirent right now.  Scan from the start of the block
1881		 * to make sure. */
1882		if (*f_version != inode->i_version) {
1883			for (i = 0; i < sb->s_blocksize && i < offset; ) {
1884				de = (struct ocfs2_dir_entry *) (bh->b_data + i);
1885				/* It's too expensive to do a full
1886				 * dirent test each time round this
1887				 * loop, but we do have to test at
1888				 * least that it is non-zero.  A
1889				 * failure will be detected in the
1890				 * dirent test below. */
1891				if (le16_to_cpu(de->rec_len) <
1892				    OCFS2_DIR_REC_LEN(1))
1893					break;
1894				i += le16_to_cpu(de->rec_len);
1895			}
1896			offset = i;
1897			ctx->pos = (ctx->pos & ~(sb->s_blocksize - 1))
1898				| offset;
1899			*f_version = inode->i_version;
1900		}
1901
1902		while (ctx->pos < i_size_read(inode)
1903		       && offset < sb->s_blocksize) {
1904			de = (struct ocfs2_dir_entry *) (bh->b_data + offset);
1905			if (!ocfs2_check_dir_entry(inode, de, bh, offset)) {
1906				/* On error, skip the f_pos to the
1907				   next block. */
1908				ctx->pos = (ctx->pos | (sb->s_blocksize - 1)) + 1;
1909				brelse(bh);
1910				continue;
1911			}
1912			if (le64_to_cpu(de->inode)) {
1913				unsigned char d_type = DT_UNKNOWN;
1914
1915				if (de->file_type < OCFS2_FT_MAX)
1916					d_type = ocfs2_filetype_table[de->file_type];
1917				if (!dir_emit(ctx, de->name,
1918						de->name_len,
1919						le64_to_cpu(de->inode),
1920						d_type)) {
1921					brelse(bh);
1922					return 0;
1923				}
1924				stored++;
1925			}
1926			offset += le16_to_cpu(de->rec_len);
1927			ctx->pos += le16_to_cpu(de->rec_len);
1928		}
1929		offset = 0;
1930		brelse(bh);
1931		bh = NULL;
1932		if (!persist && stored)
1933			break;
1934	}
1935	return 0;
1936}
1937
1938static int ocfs2_dir_foreach_blk(struct inode *inode, u64 *f_version,
1939				 struct dir_context *ctx,
1940				 bool persist)
1941{
1942	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1943		return ocfs2_dir_foreach_blk_id(inode, f_version, ctx);
1944	return ocfs2_dir_foreach_blk_el(inode, f_version, ctx, persist);
1945}
1946
1947/*
1948 * This is intended to be called from inside other kernel functions,
1949 * so we fake some arguments.
1950 */
1951int ocfs2_dir_foreach(struct inode *inode, struct dir_context *ctx)
1952{
1953	u64 version = inode->i_version;
1954	ocfs2_dir_foreach_blk(inode, &version, ctx, true);
1955	return 0;
1956}
1957
1958/*
1959 * ocfs2_readdir()
1960 *
1961 */
1962int ocfs2_readdir(struct file *file, struct dir_context *ctx)
1963{
1964	int error = 0;
1965	struct inode *inode = file_inode(file);
1966	int lock_level = 0;
1967
1968	trace_ocfs2_readdir((unsigned long long)OCFS2_I(inode)->ip_blkno);
1969
1970	error = ocfs2_inode_lock_atime(inode, file->f_path.mnt, &lock_level);
1971	if (lock_level && error >= 0) {
1972		/* We release EX lock which used to update atime
1973		 * and get PR lock again to reduce contention
1974		 * on commonly accessed directories. */
1975		ocfs2_inode_unlock(inode, 1);
1976		lock_level = 0;
1977		error = ocfs2_inode_lock(inode, NULL, 0);
1978	}
1979	if (error < 0) {
1980		if (error != -ENOENT)
1981			mlog_errno(error);
1982		/* we haven't got any yet, so propagate the error. */
1983		goto bail_nolock;
1984	}
1985
1986	error = ocfs2_dir_foreach_blk(inode, &file->f_version, ctx, false);
1987
1988	ocfs2_inode_unlock(inode, lock_level);
1989	if (error)
1990		mlog_errno(error);
1991
1992bail_nolock:
1993
1994	return error;
1995}
1996
1997/*
1998 * NOTE: this should always be called with parent dir i_mutex taken.
1999 */
2000int ocfs2_find_files_on_disk(const char *name,
2001			     int namelen,
2002			     u64 *blkno,
2003			     struct inode *inode,
2004			     struct ocfs2_dir_lookup_result *lookup)
2005{
2006	int status = -ENOENT;
2007
2008	trace_ocfs2_find_files_on_disk(namelen, name, blkno,
2009				(unsigned long long)OCFS2_I(inode)->ip_blkno);
2010
2011	status = ocfs2_find_entry(name, namelen, inode, lookup);
2012	if (status)
2013		goto leave;
2014
2015	*blkno = le64_to_cpu(lookup->dl_entry->inode);
2016
2017	status = 0;
2018leave:
2019
2020	return status;
2021}
2022
2023/*
2024 * Convenience function for callers which just want the block number
2025 * mapped to a name and don't require the full dirent info, etc.
2026 */
2027int ocfs2_lookup_ino_from_name(struct inode *dir, const char *name,
2028			       int namelen, u64 *blkno)
2029{
2030	int ret;
2031	struct ocfs2_dir_lookup_result lookup = { NULL, };
2032
2033	ret = ocfs2_find_files_on_disk(name, namelen, blkno, dir, &lookup);
2034	ocfs2_free_dir_lookup_result(&lookup);
2035
2036	return ret;
2037}
2038
2039/* Check for a name within a directory.
2040 *
2041 * Return 0 if the name does not exist
2042 * Return -EEXIST if the directory contains the name
2043 *
2044 * Callers should have i_mutex + a cluster lock on dir
2045 */
2046int ocfs2_check_dir_for_entry(struct inode *dir,
2047			      const char *name,
2048			      int namelen)
2049{
2050	int ret = 0;
2051	struct ocfs2_dir_lookup_result lookup = { NULL, };
2052
2053	trace_ocfs2_check_dir_for_entry(
2054		(unsigned long long)OCFS2_I(dir)->ip_blkno, namelen, name);
2055
2056	if (ocfs2_find_entry(name, namelen, dir, &lookup) == 0) {
2057		ret = -EEXIST;
2058		mlog_errno(ret);
2059	}
2060
2061	ocfs2_free_dir_lookup_result(&lookup);
2062
2063	return ret;
2064}
2065
2066struct ocfs2_empty_dir_priv {
2067	struct dir_context ctx;
2068	unsigned seen_dot;
2069	unsigned seen_dot_dot;
2070	unsigned seen_other;
2071	unsigned dx_dir;
2072};
2073static int ocfs2_empty_dir_filldir(struct dir_context *ctx, const char *name,
2074				   int name_len, loff_t pos, u64 ino,
2075				   unsigned type)
2076{
2077	struct ocfs2_empty_dir_priv *p =
2078		container_of(ctx, struct ocfs2_empty_dir_priv, ctx);
2079
2080	/*
2081	 * Check the positions of "." and ".." records to be sure
2082	 * they're in the correct place.
2083	 *
2084	 * Indexed directories don't need to proceed past the first
2085	 * two entries, so we end the scan after seeing '..'. Despite
2086	 * that, we allow the scan to proceed In the event that we
2087	 * have a corrupted indexed directory (no dot or dot dot
2088	 * entries). This allows us to double check for existing
2089	 * entries which might not have been found in the index.
2090	 */
2091	if (name_len == 1 && !strncmp(".", name, 1) && pos == 0) {
2092		p->seen_dot = 1;
2093		return 0;
2094	}
2095
2096	if (name_len == 2 && !strncmp("..", name, 2) &&
2097	    pos == OCFS2_DIR_REC_LEN(1)) {
2098		p->seen_dot_dot = 1;
2099
2100		if (p->dx_dir && p->seen_dot)
2101			return 1;
2102
2103		return 0;
2104	}
2105
2106	p->seen_other = 1;
2107	return 1;
2108}
2109
2110static int ocfs2_empty_dir_dx(struct inode *inode,
2111			      struct ocfs2_empty_dir_priv *priv)
2112{
2113	int ret;
2114	struct buffer_head *di_bh = NULL;
2115	struct buffer_head *dx_root_bh = NULL;
2116	struct ocfs2_dinode *di;
2117	struct ocfs2_dx_root_block *dx_root;
2118
2119	priv->dx_dir = 1;
2120
2121	ret = ocfs2_read_inode_block(inode, &di_bh);
2122	if (ret) {
2123		mlog_errno(ret);
2124		goto out;
2125	}
2126	di = (struct ocfs2_dinode *)di_bh->b_data;
2127
2128	ret = ocfs2_read_dx_root(inode, di, &dx_root_bh);
2129	if (ret) {
2130		mlog_errno(ret);
2131		goto out;
2132	}
2133	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2134
2135	if (le32_to_cpu(dx_root->dr_num_entries) != 2)
2136		priv->seen_other = 1;
2137
2138out:
2139	brelse(di_bh);
2140	brelse(dx_root_bh);
2141	return ret;
2142}
2143
2144/*
2145 * routine to check that the specified directory is empty (for rmdir)
2146 *
2147 * Returns 1 if dir is empty, zero otherwise.
2148 *
2149 * XXX: This is a performance problem for unindexed directories.
2150 */
2151int ocfs2_empty_dir(struct inode *inode)
2152{
2153	int ret;
2154	struct ocfs2_empty_dir_priv priv = {
2155		.ctx.actor = ocfs2_empty_dir_filldir,
2156	};
2157
2158	if (ocfs2_dir_indexed(inode)) {
2159		ret = ocfs2_empty_dir_dx(inode, &priv);
2160		if (ret)
2161			mlog_errno(ret);
2162		/*
2163		 * We still run ocfs2_dir_foreach to get the checks
2164		 * for "." and "..".
2165		 */
2166	}
2167
2168	ret = ocfs2_dir_foreach(inode, &priv.ctx);
2169	if (ret)
2170		mlog_errno(ret);
2171
2172	if (!priv.seen_dot || !priv.seen_dot_dot) {
2173		mlog(ML_ERROR, "bad directory (dir #%llu) - no `.' or `..'\n",
2174		     (unsigned long long)OCFS2_I(inode)->ip_blkno);
2175		/*
2176		 * XXX: Is it really safe to allow an unlink to continue?
2177		 */
2178		return 1;
2179	}
2180
2181	return !priv.seen_other;
2182}
2183
2184/*
2185 * Fills "." and ".." dirents in a new directory block. Returns dirent for
2186 * "..", which might be used during creation of a directory with a trailing
2187 * header. It is otherwise safe to ignore the return code.
2188 */
2189static struct ocfs2_dir_entry *ocfs2_fill_initial_dirents(struct inode *inode,
2190							  struct inode *parent,
2191							  char *start,
2192							  unsigned int size)
2193{
2194	struct ocfs2_dir_entry *de = (struct ocfs2_dir_entry *)start;
2195
2196	de->inode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
2197	de->name_len = 1;
2198	de->rec_len =
2199		cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
2200	strcpy(de->name, ".");
2201	ocfs2_set_de_type(de, S_IFDIR);
2202
2203	de = (struct ocfs2_dir_entry *) ((char *)de + le16_to_cpu(de->rec_len));
2204	de->inode = cpu_to_le64(OCFS2_I(parent)->ip_blkno);
2205	de->rec_len = cpu_to_le16(size - OCFS2_DIR_REC_LEN(1));
2206	de->name_len = 2;
2207	strcpy(de->name, "..");
2208	ocfs2_set_de_type(de, S_IFDIR);
2209
2210	return de;
2211}
2212
2213/*
2214 * This works together with code in ocfs2_mknod_locked() which sets
2215 * the inline-data flag and initializes the inline-data section.
2216 */
2217static int ocfs2_fill_new_dir_id(struct ocfs2_super *osb,
2218				 handle_t *handle,
2219				 struct inode *parent,
2220				 struct inode *inode,
2221				 struct buffer_head *di_bh)
2222{
2223	int ret;
2224	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2225	struct ocfs2_inline_data *data = &di->id2.i_data;
2226	unsigned int size = le16_to_cpu(data->id_count);
2227
2228	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2229				      OCFS2_JOURNAL_ACCESS_WRITE);
2230	if (ret) {
2231		mlog_errno(ret);
2232		goto out;
2233	}
2234
2235	ocfs2_fill_initial_dirents(inode, parent, data->id_data, size);
2236	ocfs2_journal_dirty(handle, di_bh);
2237
2238	i_size_write(inode, size);
2239	set_nlink(inode, 2);
2240	inode->i_blocks = ocfs2_inode_sector_count(inode);
2241
2242	ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
2243	if (ret < 0)
2244		mlog_errno(ret);
2245
2246out:
2247	return ret;
2248}
2249
2250static int ocfs2_fill_new_dir_el(struct ocfs2_super *osb,
2251				 handle_t *handle,
2252				 struct inode *parent,
2253				 struct inode *inode,
2254				 struct buffer_head *fe_bh,
2255				 struct ocfs2_alloc_context *data_ac,
2256				 struct buffer_head **ret_new_bh)
2257{
2258	int status;
2259	unsigned int size = osb->sb->s_blocksize;
2260	struct buffer_head *new_bh = NULL;
2261	struct ocfs2_dir_entry *de;
2262
2263	if (ocfs2_new_dir_wants_trailer(inode))
2264		size = ocfs2_dir_trailer_blk_off(parent->i_sb);
2265
2266	status = ocfs2_do_extend_dir(osb->sb, handle, inode, fe_bh,
2267				     data_ac, NULL, &new_bh);
2268	if (status < 0) {
2269		mlog_errno(status);
2270		goto bail;
2271	}
2272
2273	ocfs2_set_new_buffer_uptodate(INODE_CACHE(inode), new_bh);
2274
2275	status = ocfs2_journal_access_db(handle, INODE_CACHE(inode), new_bh,
2276					 OCFS2_JOURNAL_ACCESS_CREATE);
2277	if (status < 0) {
2278		mlog_errno(status);
2279		goto bail;
2280	}
2281	memset(new_bh->b_data, 0, osb->sb->s_blocksize);
2282
2283	de = ocfs2_fill_initial_dirents(inode, parent, new_bh->b_data, size);
2284	if (ocfs2_new_dir_wants_trailer(inode)) {
2285		int size = le16_to_cpu(de->rec_len);
2286
2287		/*
2288		 * Figure out the size of the hole left over after
2289		 * insertion of '.' and '..'. The trailer wants this
2290		 * information.
2291		 */
2292		size -= OCFS2_DIR_REC_LEN(2);
2293		size -= sizeof(struct ocfs2_dir_block_trailer);
2294
2295		ocfs2_init_dir_trailer(inode, new_bh, size);
2296	}
2297
2298	ocfs2_journal_dirty(handle, new_bh);
2299
2300	i_size_write(inode, inode->i_sb->s_blocksize);
2301	set_nlink(inode, 2);
2302	inode->i_blocks = ocfs2_inode_sector_count(inode);
2303	status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
2304	if (status < 0) {
2305		mlog_errno(status);
2306		goto bail;
2307	}
2308
2309	status = 0;
2310	if (ret_new_bh) {
2311		*ret_new_bh = new_bh;
2312		new_bh = NULL;
2313	}
2314bail:
2315	brelse(new_bh);
2316
2317	return status;
2318}
2319
2320static int ocfs2_dx_dir_attach_index(struct ocfs2_super *osb,
2321				     handle_t *handle, struct inode *dir,
2322				     struct buffer_head *di_bh,
2323				     struct buffer_head *dirdata_bh,
2324				     struct ocfs2_alloc_context *meta_ac,
2325				     int dx_inline, u32 num_entries,
2326				     struct buffer_head **ret_dx_root_bh)
2327{
2328	int ret;
2329	struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
2330	u16 dr_suballoc_bit;
2331	u64 suballoc_loc, dr_blkno;
2332	unsigned int num_bits;
2333	struct buffer_head *dx_root_bh = NULL;
2334	struct ocfs2_dx_root_block *dx_root;
2335	struct ocfs2_dir_block_trailer *trailer =
2336		ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
2337
2338	ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
2339				   &dr_suballoc_bit, &num_bits, &dr_blkno);
2340	if (ret) {
2341		mlog_errno(ret);
2342		goto out;
2343	}
2344
2345	trace_ocfs2_dx_dir_attach_index(
2346				(unsigned long long)OCFS2_I(dir)->ip_blkno,
2347				(unsigned long long)dr_blkno);
2348
2349	dx_root_bh = sb_getblk(osb->sb, dr_blkno);
2350	if (dx_root_bh == NULL) {
2351		ret = -ENOMEM;
2352		goto out;
2353	}
2354	ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dx_root_bh);
2355
2356	ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
2357				      OCFS2_JOURNAL_ACCESS_CREATE);
2358	if (ret < 0) {
2359		mlog_errno(ret);
2360		goto out;
2361	}
2362
2363	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2364	memset(dx_root, 0, osb->sb->s_blocksize);
2365	strcpy(dx_root->dr_signature, OCFS2_DX_ROOT_SIGNATURE);
2366	dx_root->dr_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
2367	dx_root->dr_suballoc_loc = cpu_to_le64(suballoc_loc);
2368	dx_root->dr_suballoc_bit = cpu_to_le16(dr_suballoc_bit);
2369	dx_root->dr_fs_generation = cpu_to_le32(osb->fs_generation);
2370	dx_root->dr_blkno = cpu_to_le64(dr_blkno);
2371	dx_root->dr_dir_blkno = cpu_to_le64(OCFS2_I(dir)->ip_blkno);
2372	dx_root->dr_num_entries = cpu_to_le32(num_entries);
2373	if (le16_to_cpu(trailer->db_free_rec_len))
2374		dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
2375	else
2376		dx_root->dr_free_blk = cpu_to_le64(0);
2377
2378	if (dx_inline) {
2379		dx_root->dr_flags |= OCFS2_DX_FLAG_INLINE;
2380		dx_root->dr_entries.de_count =
2381			cpu_to_le16(ocfs2_dx_entries_per_root(osb->sb));
2382	} else {
2383		dx_root->dr_list.l_count =
2384			cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
2385	}
2386	ocfs2_journal_dirty(handle, dx_root_bh);
2387
2388	ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2389				      OCFS2_JOURNAL_ACCESS_CREATE);
2390	if (ret) {
2391		mlog_errno(ret);
2392		goto out;
2393	}
2394
2395	di->i_dx_root = cpu_to_le64(dr_blkno);
2396
2397	spin_lock(&OCFS2_I(dir)->ip_lock);
2398	OCFS2_I(dir)->ip_dyn_features |= OCFS2_INDEXED_DIR_FL;
2399	di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
2400	spin_unlock(&OCFS2_I(dir)->ip_lock);
2401
2402	ocfs2_journal_dirty(handle, di_bh);
2403
2404	*ret_dx_root_bh = dx_root_bh;
2405	dx_root_bh = NULL;
2406
2407out:
2408	brelse(dx_root_bh);
2409	return ret;
2410}
2411
2412static int ocfs2_dx_dir_format_cluster(struct ocfs2_super *osb,
2413				       handle_t *handle, struct inode *dir,
2414				       struct buffer_head **dx_leaves,
2415				       int num_dx_leaves, u64 start_blk)
2416{
2417	int ret, i;
2418	struct ocfs2_dx_leaf *dx_leaf;
2419	struct buffer_head *bh;
2420
2421	for (i = 0; i < num_dx_leaves; i++) {
2422		bh = sb_getblk(osb->sb, start_blk + i);
2423		if (bh == NULL) {
2424			ret = -ENOMEM;
2425			goto out;
2426		}
2427		dx_leaves[i] = bh;
2428
2429		ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), bh);
2430
2431		ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), bh,
2432					      OCFS2_JOURNAL_ACCESS_CREATE);
2433		if (ret < 0) {
2434			mlog_errno(ret);
2435			goto out;
2436		}
2437
2438		dx_leaf = (struct ocfs2_dx_leaf *) bh->b_data;
2439
2440		memset(dx_leaf, 0, osb->sb->s_blocksize);
2441		strcpy(dx_leaf->dl_signature, OCFS2_DX_LEAF_SIGNATURE);
2442		dx_leaf->dl_fs_generation = cpu_to_le32(osb->fs_generation);
2443		dx_leaf->dl_blkno = cpu_to_le64(bh->b_blocknr);
2444		dx_leaf->dl_list.de_count =
2445			cpu_to_le16(ocfs2_dx_entries_per_leaf(osb->sb));
2446
2447		trace_ocfs2_dx_dir_format_cluster(
2448				(unsigned long long)OCFS2_I(dir)->ip_blkno,
2449				(unsigned long long)bh->b_blocknr,
2450				le16_to_cpu(dx_leaf->dl_list.de_count));
2451
2452		ocfs2_journal_dirty(handle, bh);
2453	}
2454
2455	ret = 0;
2456out:
2457	return ret;
2458}
2459
2460/*
2461 * Allocates and formats a new cluster for use in an indexed dir
2462 * leaf. This version will not do the extent insert, so that it can be
2463 * used by operations which need careful ordering.
2464 */
2465static int __ocfs2_dx_dir_new_cluster(struct inode *dir,
2466				      u32 cpos, handle_t *handle,
2467				      struct ocfs2_alloc_context *data_ac,
2468				      struct buffer_head **dx_leaves,
2469				      int num_dx_leaves, u64 *ret_phys_blkno)
2470{
2471	int ret;
2472	u32 phys, num;
2473	u64 phys_blkno;
2474	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2475
2476	/*
2477	 * XXX: For create, this should claim cluster for the index
2478	 * *before* the unindexed insert so that we have a better
2479	 * chance of contiguousness as the directory grows in number
2480	 * of entries.
2481	 */
2482	ret = __ocfs2_claim_clusters(handle, data_ac, 1, 1, &phys, &num);
2483	if (ret) {
2484		mlog_errno(ret);
2485		goto out;
2486	}
2487
2488	/*
2489	 * Format the new cluster first. That way, we're inserting
2490	 * valid data.
2491	 */
2492	phys_blkno = ocfs2_clusters_to_blocks(osb->sb, phys);
2493	ret = ocfs2_dx_dir_format_cluster(osb, handle, dir, dx_leaves,
2494					  num_dx_leaves, phys_blkno);
2495	if (ret) {
2496		mlog_errno(ret);
2497		goto out;
2498	}
2499
2500	*ret_phys_blkno = phys_blkno;
2501out:
2502	return ret;
2503}
2504
2505static int ocfs2_dx_dir_new_cluster(struct inode *dir,
2506				    struct ocfs2_extent_tree *et,
2507				    u32 cpos, handle_t *handle,
2508				    struct ocfs2_alloc_context *data_ac,
2509				    struct ocfs2_alloc_context *meta_ac,
2510				    struct buffer_head **dx_leaves,
2511				    int num_dx_leaves)
2512{
2513	int ret;
2514	u64 phys_blkno;
2515
2516	ret = __ocfs2_dx_dir_new_cluster(dir, cpos, handle, data_ac, dx_leaves,
2517					 num_dx_leaves, &phys_blkno);
2518	if (ret) {
2519		mlog_errno(ret);
2520		goto out;
2521	}
2522
2523	ret = ocfs2_insert_extent(handle, et, cpos, phys_blkno, 1, 0,
2524				  meta_ac);
2525	if (ret)
2526		mlog_errno(ret);
2527out:
2528	return ret;
2529}
2530
2531static struct buffer_head **ocfs2_dx_dir_kmalloc_leaves(struct super_block *sb,
2532							int *ret_num_leaves)
2533{
2534	int num_dx_leaves = ocfs2_clusters_to_blocks(sb, 1);
2535	struct buffer_head **dx_leaves;
2536
2537	dx_leaves = kcalloc(num_dx_leaves, sizeof(struct buffer_head *),
2538			    GFP_NOFS);
2539	if (dx_leaves && ret_num_leaves)
2540		*ret_num_leaves = num_dx_leaves;
2541
2542	return dx_leaves;
2543}
2544
2545static int ocfs2_fill_new_dir_dx(struct ocfs2_super *osb,
2546				 handle_t *handle,
2547				 struct inode *parent,
2548				 struct inode *inode,
2549				 struct buffer_head *di_bh,
2550				 struct ocfs2_alloc_context *data_ac,
2551				 struct ocfs2_alloc_context *meta_ac)
2552{
2553	int ret;
2554	struct buffer_head *leaf_bh = NULL;
2555	struct buffer_head *dx_root_bh = NULL;
2556	struct ocfs2_dx_hinfo hinfo;
2557	struct ocfs2_dx_root_block *dx_root;
2558	struct ocfs2_dx_entry_list *entry_list;
2559
2560	/*
2561	 * Our strategy is to create the directory as though it were
2562	 * unindexed, then add the index block. This works with very
2563	 * little complication since the state of a new directory is a
2564	 * very well known quantity.
2565	 *
2566	 * Essentially, we have two dirents ("." and ".."), in the 1st
2567	 * block which need indexing. These are easily inserted into
2568	 * the index block.
2569	 */
2570
2571	ret = ocfs2_fill_new_dir_el(osb, handle, parent, inode, di_bh,
2572				    data_ac, &leaf_bh);
2573	if (ret) {
2574		mlog_errno(ret);
2575		goto out;
2576	}
2577
2578	ret = ocfs2_dx_dir_attach_index(osb, handle, inode, di_bh, leaf_bh,
2579					meta_ac, 1, 2, &dx_root_bh);
2580	if (ret) {
2581		mlog_errno(ret);
2582		goto out;
2583	}
2584	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2585	entry_list = &dx_root->dr_entries;
2586
2587	/* Buffer has been journaled for us by ocfs2_dx_dir_attach_index */
2588	ocfs2_dx_dir_name_hash(inode, ".", 1, &hinfo);
2589	ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2590
2591	ocfs2_dx_dir_name_hash(inode, "..", 2, &hinfo);
2592	ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2593
2594out:
2595	brelse(dx_root_bh);
2596	brelse(leaf_bh);
2597	return ret;
2598}
2599
2600int ocfs2_fill_new_dir(struct ocfs2_super *osb,
2601		       handle_t *handle,
2602		       struct inode *parent,
2603		       struct inode *inode,
2604		       struct buffer_head *fe_bh,
2605		       struct ocfs2_alloc_context *data_ac,
2606		       struct ocfs2_alloc_context *meta_ac)
2607
2608{
2609	BUG_ON(!ocfs2_supports_inline_data(osb) && data_ac == NULL);
2610
2611	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2612		return ocfs2_fill_new_dir_id(osb, handle, parent, inode, fe_bh);
2613
2614	if (ocfs2_supports_indexed_dirs(osb))
2615		return ocfs2_fill_new_dir_dx(osb, handle, parent, inode, fe_bh,
2616					     data_ac, meta_ac);
2617
2618	return ocfs2_fill_new_dir_el(osb, handle, parent, inode, fe_bh,
2619				     data_ac, NULL);
2620}
2621
2622static int ocfs2_dx_dir_index_block(struct inode *dir,
2623				    handle_t *handle,
2624				    struct buffer_head **dx_leaves,
2625				    int num_dx_leaves,
2626				    u32 *num_dx_entries,
2627				    struct buffer_head *dirent_bh)
2628{
2629	int ret = 0, namelen, i;
2630	char *de_buf, *limit;
2631	struct ocfs2_dir_entry *de;
2632	struct buffer_head *dx_leaf_bh;
2633	struct ocfs2_dx_hinfo hinfo;
2634	u64 dirent_blk = dirent_bh->b_blocknr;
2635
2636	de_buf = dirent_bh->b_data;
2637	limit = de_buf + dir->i_sb->s_blocksize;
2638
2639	while (de_buf < limit) {
2640		de = (struct ocfs2_dir_entry *)de_buf;
2641
2642		namelen = de->name_len;
2643		if (!namelen || !de->inode)
2644			goto inc;
2645
2646		ocfs2_dx_dir_name_hash(dir, de->name, namelen, &hinfo);
2647
2648		i = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb), &hinfo);
2649		dx_leaf_bh = dx_leaves[i];
2650
2651		ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &hinfo,
2652						 dirent_blk, dx_leaf_bh);
2653		if (ret) {
2654			mlog_errno(ret);
2655			goto out;
2656		}
2657
2658		*num_dx_entries = *num_dx_entries + 1;
2659
2660inc:
2661		de_buf += le16_to_cpu(de->rec_len);
2662	}
2663
2664out:
2665	return ret;
2666}
2667
2668/*
2669 * XXX: This expects dx_root_bh to already be part of the transaction.
2670 */
2671static void ocfs2_dx_dir_index_root_block(struct inode *dir,
2672					 struct buffer_head *dx_root_bh,
2673					 struct buffer_head *dirent_bh)
2674{
2675	char *de_buf, *limit;
2676	struct ocfs2_dx_root_block *dx_root;
2677	struct ocfs2_dir_entry *de;
2678	struct ocfs2_dx_hinfo hinfo;
2679	u64 dirent_blk = dirent_bh->b_blocknr;
2680
2681	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2682
2683	de_buf = dirent_bh->b_data;
2684	limit = de_buf + dir->i_sb->s_blocksize;
2685
2686	while (de_buf < limit) {
2687		de = (struct ocfs2_dir_entry *)de_buf;
2688
2689		if (!de->name_len || !de->inode)
2690			goto inc;
2691
2692		ocfs2_dx_dir_name_hash(dir, de->name, de->name_len, &hinfo);
2693
2694		trace_ocfs2_dx_dir_index_root_block(
2695				(unsigned long long)dir->i_ino,
2696				hinfo.major_hash, hinfo.minor_hash,
2697				de->name_len, de->name,
2698				le16_to_cpu(dx_root->dr_entries.de_num_used));
2699
2700		ocfs2_dx_entry_list_insert(&dx_root->dr_entries, &hinfo,
2701					   dirent_blk);
2702
2703		le32_add_cpu(&dx_root->dr_num_entries, 1);
2704inc:
2705		de_buf += le16_to_cpu(de->rec_len);
2706	}
2707}
2708
2709/*
2710 * Count the number of inline directory entries in di_bh and compare
2711 * them against the number of entries we can hold in an inline dx root
2712 * block.
2713 */
2714static int ocfs2_new_dx_should_be_inline(struct inode *dir,
2715					 struct buffer_head *di_bh)
2716{
2717	int dirent_count = 0;
2718	char *de_buf, *limit;
2719	struct ocfs2_dir_entry *de;
2720	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2721
2722	de_buf = di->id2.i_data.id_data;
2723	limit = de_buf + i_size_read(dir);
2724
2725	while (de_buf < limit) {
2726		de = (struct ocfs2_dir_entry *)de_buf;
2727
2728		if (de->name_len && de->inode)
2729			dirent_count++;
2730
2731		de_buf += le16_to_cpu(de->rec_len);
2732	}
2733
2734	/* We are careful to leave room for one extra record. */
2735	return dirent_count < ocfs2_dx_entries_per_root(dir->i_sb);
2736}
2737
2738/*
2739 * Expand rec_len of the rightmost dirent in a directory block so that it
2740 * contains the end of our valid space for dirents. We do this during
2741 * expansion from an inline directory to one with extents. The first dir block
2742 * in that case is taken from the inline data portion of the inode block.
2743 *
2744 * This will also return the largest amount of contiguous space for a dirent
2745 * in the block. That value is *not* necessarily the last dirent, even after
2746 * expansion. The directory indexing code wants this value for free space
2747 * accounting. We do this here since we're already walking the entire dir
2748 * block.
2749 *
2750 * We add the dir trailer if this filesystem wants it.
2751 */
2752static unsigned int ocfs2_expand_last_dirent(char *start, unsigned int old_size,
2753					     struct inode *dir)
2754{
2755	struct super_block *sb = dir->i_sb;
2756	struct ocfs2_dir_entry *de;
2757	struct ocfs2_dir_entry *prev_de;
2758	char *de_buf, *limit;
2759	unsigned int new_size = sb->s_blocksize;
2760	unsigned int bytes, this_hole;
2761	unsigned int largest_hole = 0;
2762
2763	if (ocfs2_new_dir_wants_trailer(dir))
2764		new_size = ocfs2_dir_trailer_blk_off(sb);
2765
2766	bytes = new_size - old_size;
2767
2768	limit = start + old_size;
2769	de_buf = start;
2770	de = (struct ocfs2_dir_entry *)de_buf;
2771	do {
2772		this_hole = ocfs2_figure_dirent_hole(de);
2773		if (this_hole > largest_hole)
2774			largest_hole = this_hole;
2775
2776		prev_de = de;
2777		de_buf += le16_to_cpu(de->rec_len);
2778		de = (struct ocfs2_dir_entry *)de_buf;
2779	} while (de_buf < limit);
2780
2781	le16_add_cpu(&prev_de->rec_len, bytes);
2782
2783	/* We need to double check this after modification of the final
2784	 * dirent. */
2785	this_hole = ocfs2_figure_dirent_hole(prev_de);
2786	if (this_hole > largest_hole)
2787		largest_hole = this_hole;
2788
2789	if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
2790		return largest_hole;
2791	return 0;
2792}
2793
2794/*
2795 * We allocate enough clusters to fulfill "blocks_wanted", but set
2796 * i_size to exactly one block. Ocfs2_extend_dir() will handle the
2797 * rest automatically for us.
2798 *
2799 * *first_block_bh is a pointer to the 1st data block allocated to the
2800 *  directory.
2801 */
2802static int ocfs2_expand_inline_dir(struct inode *dir, struct buffer_head *di_bh,
2803				   unsigned int blocks_wanted,
2804				   struct ocfs2_dir_lookup_result *lookup,
2805				   struct buffer_head **first_block_bh)
2806{
2807	u32 alloc, dx_alloc, bit_off, len, num_dx_entries = 0;
2808	struct super_block *sb = dir->i_sb;
2809	int ret, i, num_dx_leaves = 0, dx_inline = 0,
2810		credits = ocfs2_inline_to_extents_credits(sb);
2811	u64 dx_insert_blkno, blkno,
2812		bytes = blocks_wanted << sb->s_blocksize_bits;
2813	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2814	struct ocfs2_inode_info *oi = OCFS2_I(dir);
2815	struct ocfs2_alloc_context *data_ac = NULL;
2816	struct ocfs2_alloc_context *meta_ac = NULL;
2817	struct buffer_head *dirdata_bh = NULL;
2818	struct buffer_head *dx_root_bh = NULL;
2819	struct buffer_head **dx_leaves = NULL;
2820	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2821	handle_t *handle;
2822	struct ocfs2_extent_tree et;
2823	struct ocfs2_extent_tree dx_et;
2824	int did_quota = 0, bytes_allocated = 0;
2825
2826	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir), di_bh);
2827
2828	alloc = ocfs2_clusters_for_bytes(sb, bytes);
2829	dx_alloc = 0;
2830
2831	down_write(&oi->ip_alloc_sem);
2832
2833	if (ocfs2_supports_indexed_dirs(osb)) {
2834		credits += ocfs2_add_dir_index_credits(sb);
2835
2836		dx_inline = ocfs2_new_dx_should_be_inline(dir, di_bh);
2837		if (!dx_inline) {
2838			/* Add one more cluster for an index leaf */
2839			dx_alloc++;
2840			dx_leaves = ocfs2_dx_dir_kmalloc_leaves(sb,
2841								&num_dx_leaves);
2842			if (!dx_leaves) {
2843				ret = -ENOMEM;
2844				mlog_errno(ret);
2845				goto out;
2846			}
2847		}
2848
2849		/* This gets us the dx_root */
2850		ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
2851		if (ret) {
2852			mlog_errno(ret);
2853			goto out;
2854		}
2855	}
2856
2857	/*
2858	 * We should never need more than 2 clusters for the unindexed
2859	 * tree - maximum dirent size is far less than one block. In
2860	 * fact, the only time we'd need more than one cluster is if
2861	 * blocksize == clustersize and the dirent won't fit in the
2862	 * extra space that the expansion to a single block gives. As
2863	 * of today, that only happens on 4k/4k file systems.
2864	 */
2865	BUG_ON(alloc > 2);
2866
2867	ret = ocfs2_reserve_clusters(osb, alloc + dx_alloc, &data_ac);
2868	if (ret) {
2869		mlog_errno(ret);
2870		goto out;
2871	}
2872
2873	/*
2874	 * Prepare for worst case allocation scenario of two separate
2875	 * extents in the unindexed tree.
2876	 */
2877	if (alloc == 2)
2878		credits += OCFS2_SUBALLOC_ALLOC;
2879
2880	handle = ocfs2_start_trans(osb, credits);
2881	if (IS_ERR(handle)) {
2882		ret = PTR_ERR(handle);
2883		mlog_errno(ret);
2884		goto out;
2885	}
2886
2887	ret = dquot_alloc_space_nodirty(dir,
2888		ocfs2_clusters_to_bytes(osb->sb, alloc + dx_alloc));
2889	if (ret)
2890		goto out_commit;
2891	did_quota = 1;
2892
2893	if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2894		/*
2895		 * Allocate our index cluster first, to maximize the
2896		 * possibility that unindexed leaves grow
2897		 * contiguously.
2898		 */
2899		ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac,
2900						 dx_leaves, num_dx_leaves,
2901						 &dx_insert_blkno);
2902		if (ret) {
2903			mlog_errno(ret);
2904			goto out_commit;
2905		}
2906		bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2907	}
2908
2909	/*
2910	 * Try to claim as many clusters as the bitmap can give though
2911	 * if we only get one now, that's enough to continue. The rest
2912	 * will be claimed after the conversion to extents.
2913	 */
2914	if (ocfs2_dir_resv_allowed(osb))
2915		data_ac->ac_resv = &oi->ip_la_data_resv;
2916	ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off, &len);
2917	if (ret) {
2918		mlog_errno(ret);
2919		goto out_commit;
2920	}
2921	bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2922
2923	/*
2924	 * Operations are carefully ordered so that we set up the new
2925	 * data block first. The conversion from inline data to
2926	 * extents follows.
2927	 */
2928	blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
2929	dirdata_bh = sb_getblk(sb, blkno);
2930	if (!dirdata_bh) {
2931		ret = -ENOMEM;
2932		mlog_errno(ret);
2933		goto out_commit;
2934	}
2935
2936	ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dirdata_bh);
2937
2938	ret = ocfs2_journal_access_db(handle, INODE_CACHE(dir), dirdata_bh,
2939				      OCFS2_JOURNAL_ACCESS_CREATE);
2940	if (ret) {
2941		mlog_errno(ret);
2942		goto out_commit;
2943	}
2944
2945	memcpy(dirdata_bh->b_data, di->id2.i_data.id_data, i_size_read(dir));
2946	memset(dirdata_bh->b_data + i_size_read(dir), 0,
2947	       sb->s_blocksize - i_size_read(dir));
2948	i = ocfs2_expand_last_dirent(dirdata_bh->b_data, i_size_read(dir), dir);
2949	if (ocfs2_new_dir_wants_trailer(dir)) {
2950		/*
2951		 * Prepare the dir trailer up front. It will otherwise look
2952		 * like a valid dirent. Even if inserting the index fails
2953		 * (unlikely), then all we'll have done is given first dir
2954		 * block a small amount of fragmentation.
2955		 */
2956		ocfs2_init_dir_trailer(dir, dirdata_bh, i);
2957	}
2958
2959	ocfs2_update_inode_fsync_trans(handle, dir, 1);
2960	ocfs2_journal_dirty(handle, dirdata_bh);
2961
2962	if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2963		/*
2964		 * Dx dirs with an external cluster need to do this up
2965		 * front. Inline dx root's get handled later, after
2966		 * we've allocated our root block. We get passed back
2967		 * a total number of items so that dr_num_entries can
2968		 * be correctly set once the dx_root has been
2969		 * allocated.
2970		 */
2971		ret = ocfs2_dx_dir_index_block(dir, handle, dx_leaves,
2972					       num_dx_leaves, &num_dx_entries,
2973					       dirdata_bh);
2974		if (ret) {
2975			mlog_errno(ret);
2976			goto out_commit;
2977		}
2978	}
2979
2980	/*
2981	 * Set extent, i_size, etc on the directory. After this, the
2982	 * inode should contain the same exact dirents as before and
2983	 * be fully accessible from system calls.
2984	 *
2985	 * We let the later dirent insert modify c/mtime - to the user
2986	 * the data hasn't changed.
2987	 */
2988	ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2989				      OCFS2_JOURNAL_ACCESS_CREATE);
2990	if (ret) {
2991		mlog_errno(ret);
2992		goto out_commit;
2993	}
2994
2995	spin_lock(&oi->ip_lock);
2996	oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
2997	di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
2998	spin_unlock(&oi->ip_lock);
2999
3000	ocfs2_dinode_new_extent_list(dir, di);
3001
3002	i_size_write(dir, sb->s_blocksize);
3003	dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3004
3005	di->i_size = cpu_to_le64(sb->s_blocksize);
3006	di->i_ctime = di->i_mtime = cpu_to_le64(dir->i_ctime.tv_sec);
3007	di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(dir->i_ctime.tv_nsec);
3008	ocfs2_update_inode_fsync_trans(handle, dir, 1);
3009
3010	/*
3011	 * This should never fail as our extent list is empty and all
3012	 * related blocks have been journaled already.
3013	 */
3014	ret = ocfs2_insert_extent(handle, &et, 0, blkno, len,
3015				  0, NULL);
3016	if (ret) {
3017		mlog_errno(ret);
3018		goto out_commit;
3019	}
3020
3021	/*
3022	 * Set i_blocks after the extent insert for the most up to
3023	 * date ip_clusters value.
3024	 */
3025	dir->i_blocks = ocfs2_inode_sector_count(dir);
3026
3027	ocfs2_journal_dirty(handle, di_bh);
3028
3029	if (ocfs2_supports_indexed_dirs(osb)) {
3030		ret = ocfs2_dx_dir_attach_index(osb, handle, dir, di_bh,
3031						dirdata_bh, meta_ac, dx_inline,
3032						num_dx_entries, &dx_root_bh);
3033		if (ret) {
3034			mlog_errno(ret);
3035			goto out_commit;
3036		}
3037
3038		if (dx_inline) {
3039			ocfs2_dx_dir_index_root_block(dir, dx_root_bh,
3040						      dirdata_bh);
3041		} else {
3042			ocfs2_init_dx_root_extent_tree(&dx_et,
3043						       INODE_CACHE(dir),
3044						       dx_root_bh);
3045			ret = ocfs2_insert_extent(handle, &dx_et, 0,
3046						  dx_insert_blkno, 1, 0, NULL);
3047			if (ret)
3048				mlog_errno(ret);
3049		}
3050	}
3051
3052	/*
3053	 * We asked for two clusters, but only got one in the 1st
3054	 * pass. Claim the 2nd cluster as a separate extent.
3055	 */
3056	if (alloc > len) {
3057		ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
3058					   &len);
3059		if (ret) {
3060			mlog_errno(ret);
3061			goto out_commit;
3062		}
3063		blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3064
3065		ret = ocfs2_insert_extent(handle, &et, 1,
3066					  blkno, len, 0, NULL);
3067		if (ret) {
3068			mlog_errno(ret);
3069			goto out_commit;
3070		}
3071		bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
3072	}
3073
3074	*first_block_bh = dirdata_bh;
3075	dirdata_bh = NULL;
3076	if (ocfs2_supports_indexed_dirs(osb)) {
3077		unsigned int off;
3078
3079		if (!dx_inline) {
3080			/*
3081			 * We need to return the correct block within the
3082			 * cluster which should hold our entry.
3083			 */
3084			off = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb),
3085						    &lookup->dl_hinfo);
3086			get_bh(dx_leaves[off]);
3087			lookup->dl_dx_leaf_bh = dx_leaves[off];
3088		}
3089		lookup->dl_dx_root_bh = dx_root_bh;
3090		dx_root_bh = NULL;
3091	}
3092
3093out_commit:
3094	if (ret < 0 && did_quota)
3095		dquot_free_space_nodirty(dir, bytes_allocated);
3096
3097	ocfs2_commit_trans(osb, handle);
3098
3099out:
3100	up_write(&oi->ip_alloc_sem);
3101	if (data_ac)
3102		ocfs2_free_alloc_context(data_ac);
3103	if (meta_ac)
3104		ocfs2_free_alloc_context(meta_ac);
3105
3106	if (dx_leaves) {
3107		for (i = 0; i < num_dx_leaves; i++)
3108			brelse(dx_leaves[i]);
3109		kfree(dx_leaves);
3110	}
3111
3112	brelse(dirdata_bh);
3113	brelse(dx_root_bh);
3114
3115	return ret;
3116}
3117
3118/* returns a bh of the 1st new block in the allocation. */
3119static int ocfs2_do_extend_dir(struct super_block *sb,
3120			       handle_t *handle,
3121			       struct inode *dir,
3122			       struct buffer_head *parent_fe_bh,
3123			       struct ocfs2_alloc_context *data_ac,
3124			       struct ocfs2_alloc_context *meta_ac,
3125			       struct buffer_head **new_bh)
3126{
3127	int status;
3128	int extend, did_quota = 0;
3129	u64 p_blkno, v_blkno;
3130
3131	spin_lock(&OCFS2_I(dir)->ip_lock);
3132	extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters));
3133	spin_unlock(&OCFS2_I(dir)->ip_lock);
3134
3135	if (extend) {
3136		u32 offset = OCFS2_I(dir)->ip_clusters;
3137
3138		status = dquot_alloc_space_nodirty(dir,
3139					ocfs2_clusters_to_bytes(sb, 1));
3140		if (status)
3141			goto bail;
3142		did_quota = 1;
3143
3144		status = ocfs2_add_inode_data(OCFS2_SB(sb), dir, &offset,
3145					      1, 0, parent_fe_bh, handle,
3146					      data_ac, meta_ac, NULL);
3147		BUG_ON(status == -EAGAIN);
3148		if (status < 0) {
3149			mlog_errno(status);
3150			goto bail;
3151		}
3152	}
3153
3154	v_blkno = ocfs2_blocks_for_bytes(sb, i_size_read(dir));
3155	status = ocfs2_extent_map_get_blocks(dir, v_blkno, &p_blkno, NULL, NULL);
3156	if (status < 0) {
3157		mlog_errno(status);
3158		goto bail;
3159	}
3160
3161	*new_bh = sb_getblk(sb, p_blkno);
3162	if (!*new_bh) {
3163		status = -ENOMEM;
3164		mlog_errno(status);
3165		goto bail;
3166	}
3167	status = 0;
3168bail:
3169	if (did_quota && status < 0)
3170		dquot_free_space_nodirty(dir, ocfs2_clusters_to_bytes(sb, 1));
3171	return status;
3172}
3173
3174/*
3175 * Assumes you already have a cluster lock on the directory.
3176 *
3177 * 'blocks_wanted' is only used if we have an inline directory which
3178 * is to be turned into an extent based one. The size of the dirent to
3179 * insert might be larger than the space gained by growing to just one
3180 * block, so we may have to grow the inode by two blocks in that case.
3181 *
3182 * If the directory is already indexed, dx_root_bh must be provided.
3183 */
3184static int ocfs2_extend_dir(struct ocfs2_super *osb,
3185			    struct inode *dir,
3186			    struct buffer_head *parent_fe_bh,
3187			    unsigned int blocks_wanted,
3188			    struct ocfs2_dir_lookup_result *lookup,
3189			    struct buffer_head **new_de_bh)
3190{
3191	int status = 0;
3192	int credits, num_free_extents, drop_alloc_sem = 0;
3193	loff_t dir_i_size;
3194	struct ocfs2_dinode *fe = (struct ocfs2_dinode *) parent_fe_bh->b_data;
3195	struct ocfs2_extent_list *el = &fe->id2.i_list;
3196	struct ocfs2_alloc_context *data_ac = NULL;
3197	struct ocfs2_alloc_context *meta_ac = NULL;
3198	handle_t *handle = NULL;
3199	struct buffer_head *new_bh = NULL;
3200	struct ocfs2_dir_entry * de;
3201	struct super_block *sb = osb->sb;
3202	struct ocfs2_extent_tree et;
3203	struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
3204
3205	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
3206		/*
3207		 * This would be a code error as an inline directory should
3208		 * never have an index root.
3209		 */
3210		BUG_ON(dx_root_bh);
3211
3212		status = ocfs2_expand_inline_dir(dir, parent_fe_bh,
3213						 blocks_wanted, lookup,
3214						 &new_bh);
3215		if (status) {
3216			mlog_errno(status);
3217			goto bail;
3218		}
3219
3220		/* Expansion from inline to an indexed directory will
3221		 * have given us this. */
3222		dx_root_bh = lookup->dl_dx_root_bh;
3223
3224		if (blocks_wanted == 1) {
3225			/*
3226			 * If the new dirent will fit inside the space
3227			 * created by pushing out to one block, then
3228			 * we can complete the operation
3229			 * here. Otherwise we have to expand i_size
3230			 * and format the 2nd block below.
3231			 */
3232			BUG_ON(new_bh == NULL);
3233			goto bail_bh;
3234		}
3235
3236		/*
3237		 * Get rid of 'new_bh' - we want to format the 2nd
3238		 * data block and return that instead.
3239		 */
3240		brelse(new_bh);
3241		new_bh = NULL;
3242
3243		down_write(&OCFS2_I(dir)->ip_alloc_sem);
3244		drop_alloc_sem = 1;
3245		dir_i_size = i_size_read(dir);
3246		credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3247		goto do_extend;
3248	}
3249
3250	down_write(&OCFS2_I(dir)->ip_alloc_sem);
3251	drop_alloc_sem = 1;
3252	dir_i_size = i_size_read(dir);
3253	trace_ocfs2_extend_dir((unsigned long long)OCFS2_I(dir)->ip_blkno,
3254			       dir_i_size);
3255
3256	/* dir->i_size is always block aligned. */
3257	spin_lock(&OCFS2_I(dir)->ip_lock);
3258	if (dir_i_size == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters)) {
3259		spin_unlock(&OCFS2_I(dir)->ip_lock);
3260		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir),
3261					      parent_fe_bh);
3262		num_free_extents = ocfs2_num_free_extents(osb, &et);
3263		if (num_free_extents < 0) {
3264			status = num_free_extents;
3265			mlog_errno(status);
3266			goto bail;
3267		}
3268
3269		if (!num_free_extents) {
3270			status = ocfs2_reserve_new_metadata(osb, el, &meta_ac);
3271			if (status < 0) {
3272				if (status != -ENOSPC)
3273					mlog_errno(status);
3274				goto bail;
3275			}
3276		}
3277
3278		status = ocfs2_reserve_clusters(osb, 1, &data_ac);
3279		if (status < 0) {
3280			if (status != -ENOSPC)
3281				mlog_errno(status);
3282			goto bail;
3283		}
3284
3285		if (ocfs2_dir_resv_allowed(osb))
3286			data_ac->ac_resv = &OCFS2_I(dir)->ip_la_data_resv;
3287
3288		credits = ocfs2_calc_extend_credits(sb, el);
3289	} else {
3290		spin_unlock(&OCFS2_I(dir)->ip_lock);
3291		credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3292	}
3293
3294do_extend:
3295	if (ocfs2_dir_indexed(dir))
3296		credits++; /* For attaching the new dirent block to the
3297			    * dx_root */
3298
3299	handle = ocfs2_start_trans(osb, credits);
3300	if (IS_ERR(handle)) {
3301		status = PTR_ERR(handle);
3302		handle = NULL;
3303		mlog_errno(status);
3304		goto bail;
3305	}
3306
3307	status = ocfs2_do_extend_dir(osb->sb, handle, dir, parent_fe_bh,
3308				     data_ac, meta_ac, &new_bh);
3309	if (status < 0) {
3310		mlog_errno(status);
3311		goto bail;
3312	}
3313
3314	ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), new_bh);
3315
3316	status = ocfs2_journal_access_db(handle, INODE_CACHE(dir), new_bh,
3317					 OCFS2_JOURNAL_ACCESS_CREATE);
3318	if (status < 0) {
3319		mlog_errno(status);
3320		goto bail;
3321	}
3322	memset(new_bh->b_data, 0, sb->s_blocksize);
3323
3324	de = (struct ocfs2_dir_entry *) new_bh->b_data;
3325	de->inode = 0;
3326	if (ocfs2_supports_dir_trailer(dir)) {
3327		de->rec_len = cpu_to_le16(ocfs2_dir_trailer_blk_off(sb));
3328
3329		ocfs2_init_dir_trailer(dir, new_bh, le16_to_cpu(de->rec_len));
3330
3331		if (ocfs2_dir_indexed(dir)) {
3332			status = ocfs2_dx_dir_link_trailer(dir, handle,
3333							   dx_root_bh, new_bh);
3334			if (status) {
3335				mlog_errno(status);
3336				goto bail;
3337			}
3338		}
3339	} else {
3340		de->rec_len = cpu_to_le16(sb->s_blocksize);
3341	}
3342	ocfs2_update_inode_fsync_trans(handle, dir, 1);
3343	ocfs2_journal_dirty(handle, new_bh);
3344
3345	dir_i_size += dir->i_sb->s_blocksize;
3346	i_size_write(dir, dir_i_size);
3347	dir->i_blocks = ocfs2_inode_sector_count(dir);
3348	status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
3349	if (status < 0) {
3350		mlog_errno(status);
3351		goto bail;
3352	}
3353
3354bail_bh:
3355	*new_de_bh = new_bh;
3356	get_bh(*new_de_bh);
3357bail:
3358	if (handle)
3359		ocfs2_commit_trans(osb, handle);
3360	if (drop_alloc_sem)
3361		up_write(&OCFS2_I(dir)->ip_alloc_sem);
3362
3363	if (data_ac)
3364		ocfs2_free_alloc_context(data_ac);
3365	if (meta_ac)
3366		ocfs2_free_alloc_context(meta_ac);
3367
3368	brelse(new_bh);
3369
3370	return status;
3371}
3372
3373static int ocfs2_find_dir_space_id(struct inode *dir, struct buffer_head *di_bh,
3374				   const char *name, int namelen,
3375				   struct buffer_head **ret_de_bh,
3376				   unsigned int *blocks_wanted)
3377{
3378	int ret;
3379	struct super_block *sb = dir->i_sb;
3380	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3381	struct ocfs2_dir_entry *de, *last_de = NULL;
3382	char *de_buf, *limit;
3383	unsigned long offset = 0;
3384	unsigned int rec_len, new_rec_len, free_space = dir->i_sb->s_blocksize;
3385
3386	/*
3387	 * This calculates how many free bytes we'd have in block zero, should
3388	 * this function force expansion to an extent tree.
3389	 */
3390	if (ocfs2_new_dir_wants_trailer(dir))
3391		free_space = ocfs2_dir_trailer_blk_off(sb) - i_size_read(dir);
3392	else
3393		free_space = dir->i_sb->s_blocksize - i_size_read(dir);
3394
3395	de_buf = di->id2.i_data.id_data;
3396	limit = de_buf + i_size_read(dir);
3397	rec_len = OCFS2_DIR_REC_LEN(namelen);
3398
3399	while (de_buf < limit) {
3400		de = (struct ocfs2_dir_entry *)de_buf;
3401
3402		if (!ocfs2_check_dir_entry(dir, de, di_bh, offset)) {
3403			ret = -ENOENT;
3404			goto out;
3405		}
3406		if (ocfs2_match(namelen, name, de)) {
3407			ret = -EEXIST;
3408			goto out;
3409		}
3410		/*
3411		 * No need to check for a trailing dirent record here as
3412		 * they're not used for inline dirs.
3413		 */
3414
3415		if (ocfs2_dirent_would_fit(de, rec_len)) {
3416			/* Ok, we found a spot. Return this bh and let
3417			 * the caller actually fill it in. */
3418			*ret_de_bh = di_bh;
3419			get_bh(*ret_de_bh);
3420			ret = 0;
3421			goto out;
3422		}
3423
3424		last_de = de;
3425		de_buf += le16_to_cpu(de->rec_len);
3426		offset += le16_to_cpu(de->rec_len);
3427	}
3428
3429	/*
3430	 * We're going to require expansion of the directory - figure
3431	 * out how many blocks we'll need so that a place for the
3432	 * dirent can be found.
3433	 */
3434	*blocks_wanted = 1;
3435	new_rec_len = le16_to_cpu(last_de->rec_len) + free_space;
3436	if (new_rec_len < (rec_len + OCFS2_DIR_REC_LEN(last_de->name_len)))
3437		*blocks_wanted = 2;
3438
3439	ret = -ENOSPC;
3440out:
3441	return ret;
3442}
3443
3444static int ocfs2_find_dir_space_el(struct inode *dir, const char *name,
3445				   int namelen, struct buffer_head **ret_de_bh)
3446{
3447	unsigned long offset;
3448	struct buffer_head *bh = NULL;
3449	unsigned short rec_len;
3450	struct ocfs2_dir_entry *de;
3451	struct super_block *sb = dir->i_sb;
3452	int status;
3453	int blocksize = dir->i_sb->s_blocksize;
3454
3455	status = ocfs2_read_dir_block(dir, 0, &bh, 0);
3456	if (status)
3457		goto bail;
3458
3459	rec_len = OCFS2_DIR_REC_LEN(namelen);
3460	offset = 0;
3461	de = (struct ocfs2_dir_entry *) bh->b_data;
3462	while (1) {
3463		if ((char *)de >= sb->s_blocksize + bh->b_data) {
3464			brelse(bh);
3465			bh = NULL;
3466
3467			if (i_size_read(dir) <= offset) {
3468				/*
3469				 * Caller will have to expand this
3470				 * directory.
3471				 */
3472				status = -ENOSPC;
3473				goto bail;
3474			}
3475			status = ocfs2_read_dir_block(dir,
3476					     offset >> sb->s_blocksize_bits,
3477					     &bh, 0);
3478			if (status)
3479				goto bail;
3480
3481			/* move to next block */
3482			de = (struct ocfs2_dir_entry *) bh->b_data;
3483		}
3484		if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
3485			status = -ENOENT;
3486			goto bail;
3487		}
3488		if (ocfs2_match(namelen, name, de)) {
3489			status = -EEXIST;
3490			goto bail;
3491		}
3492
3493		if (ocfs2_skip_dir_trailer(dir, de, offset % blocksize,
3494					   blocksize))
3495			goto next;
3496
3497		if (ocfs2_dirent_would_fit(de, rec_len)) {
3498			/* Ok, we found a spot. Return this bh and let
3499			 * the caller actually fill it in. */
3500			*ret_de_bh = bh;
3501			get_bh(*ret_de_bh);
3502			status = 0;
3503			goto bail;
3504		}
3505next:
3506		offset += le16_to_cpu(de->rec_len);
3507		de = (struct ocfs2_dir_entry *)((char *) de + le16_to_cpu(de->rec_len));
3508	}
3509
3510bail:
3511	brelse(bh);
3512	if (status)
3513		mlog_errno(status);
3514
3515	return status;
3516}
3517
3518static int dx_leaf_sort_cmp(const void *a, const void *b)
3519{
3520	const struct ocfs2_dx_entry *entry1 = a;
3521	const struct ocfs2_dx_entry *entry2 = b;
3522	u32 major_hash1 = le32_to_cpu(entry1->dx_major_hash);
3523	u32 major_hash2 = le32_to_cpu(entry2->dx_major_hash);
3524	u32 minor_hash1 = le32_to_cpu(entry1->dx_minor_hash);
3525	u32 minor_hash2 = le32_to_cpu(entry2->dx_minor_hash);
3526
3527	if (major_hash1 > major_hash2)
3528		return 1;
3529	if (major_hash1 < major_hash2)
3530		return -1;
3531
3532	/*
3533	 * It is not strictly necessary to sort by minor
3534	 */
3535	if (minor_hash1 > minor_hash2)
3536		return 1;
3537	if (minor_hash1 < minor_hash2)
3538		return -1;
3539	return 0;
3540}
3541
3542static void dx_leaf_sort_swap(void *a, void *b, int size)
3543{
3544	struct ocfs2_dx_entry *entry1 = a;
3545	struct ocfs2_dx_entry *entry2 = b;
3546	struct ocfs2_dx_entry tmp;
3547
3548	BUG_ON(size != sizeof(*entry1));
3549
3550	tmp = *entry1;
3551	*entry1 = *entry2;
3552	*entry2 = tmp;
3553}
3554
3555static int ocfs2_dx_leaf_same_major(struct ocfs2_dx_leaf *dx_leaf)
3556{
3557	struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3558	int i, num = le16_to_cpu(dl_list->de_num_used);
3559
3560	for (i = 0; i < (num - 1); i++) {
3561		if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) !=
3562		    le32_to_cpu(dl_list->de_entries[i + 1].dx_major_hash))
3563			return 0;
3564	}
3565
3566	return 1;
3567}
3568
3569/*
3570 * Find the optimal value to split this leaf on. This expects the leaf
3571 * entries to be in sorted order.
3572 *
3573 * leaf_cpos is the cpos of the leaf we're splitting. insert_hash is
3574 * the hash we want to insert.
3575 *
3576 * This function is only concerned with the major hash - that which
3577 * determines which cluster an item belongs to.
3578 */
3579static int ocfs2_dx_dir_find_leaf_split(struct ocfs2_dx_leaf *dx_leaf,
3580					u32 leaf_cpos, u32 insert_hash,
3581					u32 *split_hash)
3582{
3583	struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3584	int i, num_used = le16_to_cpu(dl_list->de_num_used);
3585	int allsame;
3586
3587	/*
3588	 * There's a couple rare, but nasty corner cases we have to
3589	 * check for here. All of them involve a leaf where all value
3590	 * have the same hash, which is what we look for first.
3591	 *
3592	 * Most of the time, all of the above is false, and we simply
3593	 * pick the median value for a split.
3594	 */
3595	allsame = ocfs2_dx_leaf_same_major(dx_leaf);
3596	if (allsame) {
3597		u32 val = le32_to_cpu(dl_list->de_entries[0].dx_major_hash);
3598
3599		if (val == insert_hash) {
3600			/*
3601			 * No matter where we would choose to split,
3602			 * the new entry would want to occupy the same
3603			 * block as these. Since there's no space left
3604			 * in their existing block, we know there
3605			 * won't be space after the split.
3606			 */
3607			return -ENOSPC;
3608		}
3609
3610		if (val == leaf_cpos) {
3611			/*
3612			 * Because val is the same as leaf_cpos (which
3613			 * is the smallest value this leaf can have),
3614			 * yet is not equal to insert_hash, then we
3615			 * know that insert_hash *must* be larger than
3616			 * val (and leaf_cpos). At least cpos+1 in value.
3617			 *
3618			 * We also know then, that there cannot be an
3619			 * adjacent extent (otherwise we'd be looking
3620			 * at it). Choosing this value gives us a
3621			 * chance to get some contiguousness.
3622			 */
3623			*split_hash = leaf_cpos + 1;
3624			return 0;
3625		}
3626
3627		if (val > insert_hash) {
3628			/*
3629			 * val can not be the same as insert hash, and
3630			 * also must be larger than leaf_cpos. Also,
3631			 * we know that there can't be a leaf between
3632			 * cpos and val, otherwise the entries with
3633			 * hash 'val' would be there.
3634			 */
3635			*split_hash = val;
3636			return 0;
3637		}
3638
3639		*split_hash = insert_hash;
3640		return 0;
3641	}
3642
3643	/*
3644	 * Since the records are sorted and the checks above
3645	 * guaranteed that not all records in this block are the same,
3646	 * we simple travel forward, from the median, and pick the 1st
3647	 * record whose value is larger than leaf_cpos.
3648	 */
3649	for (i = (num_used / 2); i < num_used; i++)
3650		if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) >
3651		    leaf_cpos)
3652			break;
3653
3654	BUG_ON(i == num_used); /* Should be impossible */
3655	*split_hash = le32_to_cpu(dl_list->de_entries[i].dx_major_hash);
3656	return 0;
3657}
3658
3659/*
3660 * Transfer all entries in orig_dx_leaves whose major hash is equal to or
3661 * larger than split_hash into new_dx_leaves. We use a temporary
3662 * buffer (tmp_dx_leaf) to make the changes to the original leaf blocks.
3663 *
3664 * Since the block offset inside a leaf (cluster) is a constant mask
3665 * of minor_hash, we can optimize - an item at block offset X within
3666 * the original cluster, will be at offset X within the new cluster.
3667 */
3668static void ocfs2_dx_dir_transfer_leaf(struct inode *dir, u32 split_hash,
3669				       handle_t *handle,
3670				       struct ocfs2_dx_leaf *tmp_dx_leaf,
3671				       struct buffer_head **orig_dx_leaves,
3672				       struct buffer_head **new_dx_leaves,
3673				       int num_dx_leaves)
3674{
3675	int i, j, num_used;
3676	u32 major_hash;
3677	struct ocfs2_dx_leaf *orig_dx_leaf, *new_dx_leaf;
3678	struct ocfs2_dx_entry_list *orig_list, *new_list, *tmp_list;
3679	struct ocfs2_dx_entry *dx_entry;
3680
3681	tmp_list = &tmp_dx_leaf->dl_list;
3682
3683	for (i = 0; i < num_dx_leaves; i++) {
3684		orig_dx_leaf = (struct ocfs2_dx_leaf *) orig_dx_leaves[i]->b_data;
3685		orig_list = &orig_dx_leaf->dl_list;
3686		new_dx_leaf = (struct ocfs2_dx_leaf *) new_dx_leaves[i]->b_data;
3687		new_list = &new_dx_leaf->dl_list;
3688
3689		num_used = le16_to_cpu(orig_list->de_num_used);
3690
3691		memcpy(tmp_dx_leaf, orig_dx_leaf, dir->i_sb->s_blocksize);
3692		tmp_list->de_num_used = cpu_to_le16(0);
3693		memset(&tmp_list->de_entries, 0, sizeof(*dx_entry)*num_used);
3694
3695		for (j = 0; j < num_used; j++) {
3696			dx_entry = &orig_list->de_entries[j];
3697			major_hash = le32_to_cpu(dx_entry->dx_major_hash);
3698			if (major_hash >= split_hash)
3699				ocfs2_dx_dir_leaf_insert_tail(new_dx_leaf,
3700							      dx_entry);
3701			else
3702				ocfs2_dx_dir_leaf_insert_tail(tmp_dx_leaf,
3703							      dx_entry);
3704		}
3705		memcpy(orig_dx_leaf, tmp_dx_leaf, dir->i_sb->s_blocksize);
3706
3707		ocfs2_journal_dirty(handle, orig_dx_leaves[i]);
3708		ocfs2_journal_dirty(handle, new_dx_leaves[i]);
3709	}
3710}
3711
3712static int ocfs2_dx_dir_rebalance_credits(struct ocfs2_super *osb,
3713					  struct ocfs2_dx_root_block *dx_root)
3714{
3715	int credits = ocfs2_clusters_to_blocks(osb->sb, 2);
3716
3717	credits += ocfs2_calc_extend_credits(osb->sb, &dx_root->dr_list);
3718	credits += ocfs2_quota_trans_credits(osb->sb);
3719	return credits;
3720}
3721
3722/*
3723 * Find the median value in dx_leaf_bh and allocate a new leaf to move
3724 * half our entries into.
3725 */
3726static int ocfs2_dx_dir_rebalance(struct ocfs2_super *osb, struct inode *dir,
3727				  struct buffer_head *dx_root_bh,
3728				  struct buffer_head *dx_leaf_bh,
3729				  struct ocfs2_dx_hinfo *hinfo, u32 leaf_cpos,
3730				  u64 leaf_blkno)
3731{
3732	struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3733	int credits, ret, i, num_used, did_quota = 0;
3734	u32 cpos, split_hash, insert_hash = hinfo->major_hash;
3735	u64 orig_leaves_start;
3736	int num_dx_leaves;
3737	struct buffer_head **orig_dx_leaves = NULL;
3738	struct buffer_head **new_dx_leaves = NULL;
3739	struct ocfs2_alloc_context *data_ac = NULL, *meta_ac = NULL;
3740	struct ocfs2_extent_tree et;
3741	handle_t *handle = NULL;
3742	struct ocfs2_dx_root_block *dx_root;
3743	struct ocfs2_dx_leaf *tmp_dx_leaf = NULL;
3744
3745	trace_ocfs2_dx_dir_rebalance((unsigned long long)OCFS2_I(dir)->ip_blkno,
3746				     (unsigned long long)leaf_blkno,
3747				     insert_hash);
3748
3749	ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
3750
3751	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3752	/*
3753	 * XXX: This is a rather large limit. We should use a more
3754	 * realistic value.
3755	 */
3756	if (le32_to_cpu(dx_root->dr_clusters) == UINT_MAX)
3757		return -ENOSPC;
3758
3759	num_used = le16_to_cpu(dx_leaf->dl_list.de_num_used);
3760	if (num_used < le16_to_cpu(dx_leaf->dl_list.de_count)) {
3761		mlog(ML_ERROR, "DX Dir: %llu, Asked to rebalance empty leaf: "
3762		     "%llu, %d\n", (unsigned long long)OCFS2_I(dir)->ip_blkno,
3763		     (unsigned long long)leaf_blkno, num_used);
3764		ret = -EIO;
3765		goto out;
3766	}
3767
3768	orig_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
3769	if (!orig_dx_leaves) {
3770		ret = -ENOMEM;
3771		mlog_errno(ret);
3772		goto out;
3773	}
3774
3775	new_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, NULL);
3776	if (!new_dx_leaves) {
3777		ret = -ENOMEM;
3778		mlog_errno(ret);
3779		goto out;
3780	}
3781
3782	ret = ocfs2_lock_allocators(dir, &et, 1, 0, &data_ac, &meta_ac);
3783	if (ret) {
3784		if (ret != -ENOSPC)
3785			mlog_errno(ret);
3786		goto out;
3787	}
3788
3789	credits = ocfs2_dx_dir_rebalance_credits(osb, dx_root);
3790	handle = ocfs2_start_trans(osb, credits);
3791	if (IS_ERR(handle)) {
3792		ret = PTR_ERR(handle);
3793		handle = NULL;
3794		mlog_errno(ret);
3795		goto out;
3796	}
3797
3798	ret = dquot_alloc_space_nodirty(dir,
3799				       ocfs2_clusters_to_bytes(dir->i_sb, 1));
3800	if (ret)
3801		goto out_commit;
3802	did_quota = 1;
3803
3804	ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
3805				      OCFS2_JOURNAL_ACCESS_WRITE);
3806	if (ret) {
3807		mlog_errno(ret);
3808		goto out_commit;
3809	}
3810
3811	/*
3812	 * This block is changing anyway, so we can sort it in place.
3813	 */
3814	sort(dx_leaf->dl_list.de_entries, num_used,
3815	     sizeof(struct ocfs2_dx_entry), dx_leaf_sort_cmp,
3816	     dx_leaf_sort_swap);
3817
3818	ocfs2_journal_dirty(handle, dx_leaf_bh);
3819
3820	ret = ocfs2_dx_dir_find_leaf_split(dx_leaf, leaf_cpos, insert_hash,
3821					   &split_hash);
3822	if (ret) {
3823		mlog_errno(ret);
3824		goto  out_commit;
3825	}
3826
3827	trace_ocfs2_dx_dir_rebalance_split(leaf_cpos, split_hash, insert_hash);
3828
3829	/*
3830	 * We have to carefully order operations here. There are items
3831	 * which want to be in the new cluster before insert, but in
3832	 * order to put those items in the new cluster, we alter the
3833	 * old cluster. A failure to insert gets nasty.
3834	 *
3835	 * So, start by reserving writes to the old
3836	 * cluster. ocfs2_dx_dir_new_cluster will reserve writes on
3837	 * the new cluster for us, before inserting it. The insert
3838	 * won't happen if there's an error before that. Once the
3839	 * insert is done then, we can transfer from one leaf into the
3840	 * other without fear of hitting any error.
3841	 */
3842
3843	/*
3844	 * The leaf transfer wants some scratch space so that we don't
3845	 * wind up doing a bunch of expensive memmove().
3846	 */
3847	tmp_dx_leaf = kmalloc(osb->sb->s_blocksize, GFP_NOFS);
3848	if (!tmp_dx_leaf) {
3849		ret = -ENOMEM;
3850		mlog_errno(ret);
3851		goto out_commit;
3852	}
3853
3854	orig_leaves_start = ocfs2_block_to_cluster_start(dir->i_sb, leaf_blkno);
3855	ret = ocfs2_read_dx_leaves(dir, orig_leaves_start, num_dx_leaves,
3856				   orig_dx_leaves);
3857	if (ret) {
3858		mlog_errno(ret);
3859		goto out_commit;
3860	}
3861
3862	cpos = split_hash;
3863	ret = ocfs2_dx_dir_new_cluster(dir, &et, cpos, handle,
3864				       data_ac, meta_ac, new_dx_leaves,
3865				       num_dx_leaves);
3866	if (ret) {
3867		mlog_errno(ret);
3868		goto out_commit;
3869	}
3870
3871	for (i = 0; i < num_dx_leaves; i++) {
3872		ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3873					      orig_dx_leaves[i],
3874					      OCFS2_JOURNAL_ACCESS_WRITE);
3875		if (ret) {
3876			mlog_errno(ret);
3877			goto out_commit;
3878		}
3879
3880		ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3881					      new_dx_leaves[i],
3882					      OCFS2_JOURNAL_ACCESS_WRITE);
3883		if (ret) {
3884			mlog_errno(ret);
3885			goto out_commit;
3886		}
3887	}
3888
3889	ocfs2_dx_dir_transfer_leaf(dir, split_hash, handle, tmp_dx_leaf,
3890				   orig_dx_leaves, new_dx_leaves, num_dx_leaves);
3891
3892out_commit:
3893	if (ret < 0 && did_quota)
3894		dquot_free_space_nodirty(dir,
3895				ocfs2_clusters_to_bytes(dir->i_sb, 1));
3896
3897	ocfs2_update_inode_fsync_trans(handle, dir, 1);
3898	ocfs2_commit_trans(osb, handle);
3899
3900out:
3901	if (orig_dx_leaves || new_dx_leaves) {
3902		for (i = 0; i < num_dx_leaves; i++) {
3903			if (orig_dx_leaves)
3904				brelse(orig_dx_leaves[i]);
3905			if (new_dx_leaves)
3906				brelse(new_dx_leaves[i]);
3907		}
3908		kfree(orig_dx_leaves);
3909		kfree(new_dx_leaves);
3910	}
3911
3912	if (meta_ac)
3913		ocfs2_free_alloc_context(meta_ac);
3914	if (data_ac)
3915		ocfs2_free_alloc_context(data_ac);
3916
3917	kfree(tmp_dx_leaf);
3918	return ret;
3919}
3920
3921static int ocfs2_find_dir_space_dx(struct ocfs2_super *osb, struct inode *dir,
3922				   struct buffer_head *di_bh,
3923				   struct buffer_head *dx_root_bh,
3924				   const char *name, int namelen,
3925				   struct ocfs2_dir_lookup_result *lookup)
3926{
3927	int ret, rebalanced = 0;
3928	struct ocfs2_dx_root_block *dx_root;
3929	struct buffer_head *dx_leaf_bh = NULL;
3930	struct ocfs2_dx_leaf *dx_leaf;
3931	u64 blkno;
3932	u32 leaf_cpos;
3933
3934	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3935
3936restart_search:
3937	ret = ocfs2_dx_dir_lookup(dir, &dx_root->dr_list, &lookup->dl_hinfo,
3938				  &leaf_cpos, &blkno);
3939	if (ret) {
3940		mlog_errno(ret);
3941		goto out;
3942	}
3943
3944	ret = ocfs2_read_dx_leaf(dir, blkno, &dx_leaf_bh);
3945	if (ret) {
3946		mlog_errno(ret);
3947		goto out;
3948	}
3949
3950	dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3951
3952	if (le16_to_cpu(dx_leaf->dl_list.de_num_used) >=
3953	    le16_to_cpu(dx_leaf->dl_list.de_count)) {
3954		if (rebalanced) {
3955			/*
3956			 * Rebalancing should have provided us with
3957			 * space in an appropriate leaf.
3958			 *
3959			 * XXX: Is this an abnormal condition then?
3960			 * Should we print a message here?
3961			 */
3962			ret = -ENOSPC;
3963			goto out;
3964		}
3965
3966		ret = ocfs2_dx_dir_rebalance(osb, dir, dx_root_bh, dx_leaf_bh,
3967					     &lookup->dl_hinfo, leaf_cpos,
3968					     blkno);
3969		if (ret) {
3970			if (ret != -ENOSPC)
3971				mlog_errno(ret);
3972			goto out;
3973		}
3974
3975		/*
3976		 * Restart the lookup. The rebalance might have
3977		 * changed which block our item fits into. Mark our
3978		 * progress, so we only execute this once.
3979		 */
3980		brelse(dx_leaf_bh);
3981		dx_leaf_bh = NULL;
3982		rebalanced = 1;
3983		goto restart_search;
3984	}
3985
3986	lookup->dl_dx_leaf_bh = dx_leaf_bh;
3987	dx_leaf_bh = NULL;
3988
3989out:
3990	brelse(dx_leaf_bh);
3991	return ret;
3992}
3993
3994static int ocfs2_search_dx_free_list(struct inode *dir,
3995				     struct buffer_head *dx_root_bh,
3996				     int namelen,
3997				     struct ocfs2_dir_lookup_result *lookup)
3998{
3999	int ret = -ENOSPC;
4000	struct buffer_head *leaf_bh = NULL, *prev_leaf_bh = NULL;
4001	struct ocfs2_dir_block_trailer *db;
4002	u64 next_block;
4003	int rec_len = OCFS2_DIR_REC_LEN(namelen);
4004	struct ocfs2_dx_root_block *dx_root;
4005
4006	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4007	next_block = le64_to_cpu(dx_root->dr_free_blk);
4008
4009	while (next_block) {
4010		brelse(prev_leaf_bh);
4011		prev_leaf_bh = leaf_bh;
4012		leaf_bh = NULL;
4013
4014		ret = ocfs2_read_dir_block_direct(dir, next_block, &leaf_bh);
4015		if (ret) {
4016			mlog_errno(ret);
4017			goto out;
4018		}
4019
4020		db = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
4021		if (rec_len <= le16_to_cpu(db->db_free_rec_len)) {
4022			lookup->dl_leaf_bh = leaf_bh;
4023			lookup->dl_prev_leaf_bh = prev_leaf_bh;
4024			leaf_bh = NULL;
4025			prev_leaf_bh = NULL;
4026			break;
4027		}
4028
4029		next_block = le64_to_cpu(db->db_free_next);
4030	}
4031
4032	if (!next_block)
4033		ret = -ENOSPC;
4034
4035out:
4036
4037	brelse(leaf_bh);
4038	brelse(prev_leaf_bh);
4039	return ret;
4040}
4041
4042static int ocfs2_expand_inline_dx_root(struct inode *dir,
4043				       struct buffer_head *dx_root_bh)
4044{
4045	int ret, num_dx_leaves, i, j, did_quota = 0;
4046	struct buffer_head **dx_leaves = NULL;
4047	struct ocfs2_extent_tree et;
4048	u64 insert_blkno;
4049	struct ocfs2_alloc_context *data_ac = NULL;
4050	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4051	handle_t *handle = NULL;
4052	struct ocfs2_dx_root_block *dx_root;
4053	struct ocfs2_dx_entry_list *entry_list;
4054	struct ocfs2_dx_entry *dx_entry;
4055	struct ocfs2_dx_leaf *target_leaf;
4056
4057	ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
4058	if (ret) {
4059		mlog_errno(ret);
4060		goto out;
4061	}
4062
4063	dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
4064	if (!dx_leaves) {
4065		ret = -ENOMEM;
4066		mlog_errno(ret);
4067		goto out;
4068	}
4069
4070	handle = ocfs2_start_trans(osb, ocfs2_calc_dxi_expand_credits(osb->sb));
4071	if (IS_ERR(handle)) {
4072		ret = PTR_ERR(handle);
4073		mlog_errno(ret);
4074		goto out;
4075	}
4076
4077	ret = dquot_alloc_space_nodirty(dir,
4078				       ocfs2_clusters_to_bytes(osb->sb, 1));
4079	if (ret)
4080		goto out_commit;
4081	did_quota = 1;
4082
4083	/*
4084	 * We do this up front, before the allocation, so that a
4085	 * failure to add the dx_root_bh to the journal won't result
4086	 * us losing clusters.
4087	 */
4088	ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
4089				      OCFS2_JOURNAL_ACCESS_WRITE);
4090	if (ret) {
4091		mlog_errno(ret);
4092		goto out_commit;
4093	}
4094
4095	ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac, dx_leaves,
4096					 num_dx_leaves, &insert_blkno);
4097	if (ret) {
4098		mlog_errno(ret);
4099		goto out_commit;
4100	}
4101
4102	/*
4103	 * Transfer the entries from our dx_root into the appropriate
4104	 * block
4105	 */
4106	dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4107	entry_list = &dx_root->dr_entries;
4108
4109	for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
4110		dx_entry = &entry_list->de_entries[i];
4111
4112		j = __ocfs2_dx_dir_hash_idx(osb,
4113					    le32_to_cpu(dx_entry->dx_minor_hash));
4114		target_leaf = (struct ocfs2_dx_leaf *)dx_leaves[j]->b_data;
4115
4116		ocfs2_dx_dir_leaf_insert_tail(target_leaf, dx_entry);
4117
4118		/* Each leaf has been passed to the journal already
4119		 * via __ocfs2_dx_dir_new_cluster() */
4120	}
4121
4122	dx_root->dr_flags &= ~OCFS2_DX_FLAG_INLINE;
4123	memset(&dx_root->dr_list, 0, osb->sb->s_blocksize -
4124	       offsetof(struct ocfs2_dx_root_block, dr_list));
4125	dx_root->dr_list.l_count =
4126		cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
4127
4128	/* This should never fail considering we start with an empty
4129	 * dx_root. */
4130	ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4131	ret = ocfs2_insert_extent(handle, &et, 0, insert_blkno, 1, 0, NULL);
4132	if (ret)
4133		mlog_errno(ret);
4134	did_quota = 0;
4135
4136	ocfs2_update_inode_fsync_trans(handle, dir, 1);
4137	ocfs2_journal_dirty(handle, dx_root_bh);
4138
4139out_commit:
4140	if (ret < 0 && did_quota)
4141		dquot_free_space_nodirty(dir,
4142					  ocfs2_clusters_to_bytes(dir->i_sb, 1));
4143
4144	ocfs2_commit_trans(osb, handle);
4145
4146out:
4147	if (data_ac)
4148		ocfs2_free_alloc_context(data_ac);
4149
4150	if (dx_leaves) {
4151		for (i = 0; i < num_dx_leaves; i++)
4152			brelse(dx_leaves[i]);
4153		kfree(dx_leaves);
4154	}
4155	return ret;
4156}
4157
4158static int ocfs2_inline_dx_has_space(struct buffer_head *dx_root_bh)
4159{
4160	struct ocfs2_dx_root_block *dx_root;
4161	struct ocfs2_dx_entry_list *entry_list;
4162
4163	dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4164	entry_list = &dx_root->dr_entries;
4165
4166	if (le16_to_cpu(entry_list->de_num_used) >=
4167	    le16_to_cpu(entry_list->de_count))
4168		return -ENOSPC;
4169
4170	return 0;
4171}
4172
4173static int ocfs2_prepare_dx_dir_for_insert(struct inode *dir,
4174					   struct buffer_head *di_bh,
4175					   const char *name,
4176					   int namelen,
4177					   struct ocfs2_dir_lookup_result *lookup)
4178{
4179	int ret, free_dx_root = 1;
4180	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4181	struct buffer_head *dx_root_bh = NULL;
4182	struct buffer_head *leaf_bh = NULL;
4183	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4184	struct ocfs2_dx_root_block *dx_root;
4185
4186	ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4187	if (ret) {
4188		mlog_errno(ret);
4189		goto out;
4190	}
4191
4192	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4193	if (le32_to_cpu(dx_root->dr_num_entries) == OCFS2_DX_ENTRIES_MAX) {
4194		ret = -ENOSPC;
4195		mlog_errno(ret);
4196		goto out;
4197	}
4198
4199	if (ocfs2_dx_root_inline(dx_root)) {
4200		ret = ocfs2_inline_dx_has_space(dx_root_bh);
4201
4202		if (ret == 0)
4203			goto search_el;
4204
4205		/*
4206		 * We ran out of room in the root block. Expand it to
4207		 * an extent, then allow ocfs2_find_dir_space_dx to do
4208		 * the rest.
4209		 */
4210		ret = ocfs2_expand_inline_dx_root(dir, dx_root_bh);
4211		if (ret) {
4212			mlog_errno(ret);
4213			goto out;
4214		}
4215	}
4216
4217	/*
4218	 * Insert preparation for an indexed directory is split into two
4219	 * steps. The call to find_dir_space_dx reserves room in the index for
4220	 * an additional item. If we run out of space there, it's a real error
4221	 * we can't continue on.
4222	 */
4223	ret = ocfs2_find_dir_space_dx(osb, dir, di_bh, dx_root_bh, name,
4224				      namelen, lookup);
4225	if (ret) {
4226		mlog_errno(ret);
4227		goto out;
4228	}
4229
4230search_el:
4231	/*
4232	 * Next, we need to find space in the unindexed tree. This call
4233	 * searches using the free space linked list. If the unindexed tree
4234	 * lacks sufficient space, we'll expand it below. The expansion code
4235	 * is smart enough to add any new blocks to the free space list.
4236	 */
4237	ret = ocfs2_search_dx_free_list(dir, dx_root_bh, namelen, lookup);
4238	if (ret && ret != -ENOSPC) {
4239		mlog_errno(ret);
4240		goto out;
4241	}
4242
4243	/* Do this up here - ocfs2_extend_dir might need the dx_root */
4244	lookup->dl_dx_root_bh = dx_root_bh;
4245	free_dx_root = 0;
4246
4247	if (ret == -ENOSPC) {
4248		ret = ocfs2_extend_dir(osb, dir, di_bh, 1, lookup, &leaf_bh);
4249
4250		if (ret) {
4251			mlog_errno(ret);
4252			goto out;
4253		}
4254
4255		/*
4256		 * We make the assumption here that new leaf blocks are added
4257		 * to the front of our free list.
4258		 */
4259		lookup->dl_prev_leaf_bh = NULL;
4260		lookup->dl_leaf_bh = leaf_bh;
4261	}
4262
4263out:
4264	if (free_dx_root)
4265		brelse(dx_root_bh);
4266	return ret;
4267}
4268
4269/*
4270 * Get a directory ready for insert. Any directory allocation required
4271 * happens here. Success returns zero, and enough context in the dir
4272 * lookup result that ocfs2_add_entry() will be able complete the task
4273 * with minimal performance impact.
4274 */
4275int ocfs2_prepare_dir_for_insert(struct ocfs2_super *osb,
4276				 struct inode *dir,
4277				 struct buffer_head *parent_fe_bh,
4278				 const char *name,
4279				 int namelen,
4280				 struct ocfs2_dir_lookup_result *lookup)
4281{
4282	int ret;
4283	unsigned int blocks_wanted = 1;
4284	struct buffer_head *bh = NULL;
4285
4286	trace_ocfs2_prepare_dir_for_insert(
4287		(unsigned long long)OCFS2_I(dir)->ip_blkno, namelen);
4288
4289	if (!namelen) {
4290		ret = -EINVAL;
4291		mlog_errno(ret);
4292		goto out;
4293	}
4294
4295	/*
4296	 * Do this up front to reduce confusion.
4297	 *
4298	 * The directory might start inline, then be turned into an
4299	 * indexed one, in which case we'd need to hash deep inside
4300	 * ocfs2_find_dir_space_id(). Since
4301	 * ocfs2_prepare_dx_dir_for_insert() also needs this hash
4302	 * done, there seems no point in spreading out the calls. We
4303	 * can optimize away the case where the file system doesn't
4304	 * support indexing.
4305	 */
4306	if (ocfs2_supports_indexed_dirs(osb))
4307		ocfs2_dx_dir_name_hash(dir, name, namelen, &lookup->dl_hinfo);
4308
4309	if (ocfs2_dir_indexed(dir)) {
4310		ret = ocfs2_prepare_dx_dir_for_insert(dir, parent_fe_bh,
4311						      name, namelen, lookup);
4312		if (ret)
4313			mlog_errno(ret);
4314		goto out;
4315	}
4316
4317	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4318		ret = ocfs2_find_dir_space_id(dir, parent_fe_bh, name,
4319					      namelen, &bh, &blocks_wanted);
4320	} else
4321		ret = ocfs2_find_dir_space_el(dir, name, namelen, &bh);
4322
4323	if (ret && ret != -ENOSPC) {
4324		mlog_errno(ret);
4325		goto out;
4326	}
4327
4328	if (ret == -ENOSPC) {
4329		/*
4330		 * We have to expand the directory to add this name.
4331		 */
4332		BUG_ON(bh);
4333
4334		ret = ocfs2_extend_dir(osb, dir, parent_fe_bh, blocks_wanted,
4335				       lookup, &bh);
4336		if (ret) {
4337			if (ret != -ENOSPC)
4338				mlog_errno(ret);
4339			goto out;
4340		}
4341
4342		BUG_ON(!bh);
4343	}
4344
4345	lookup->dl_leaf_bh = bh;
4346	bh = NULL;
4347out:
4348	brelse(bh);
4349	return ret;
4350}
4351
4352static int ocfs2_dx_dir_remove_index(struct inode *dir,
4353				     struct buffer_head *di_bh,
4354				     struct buffer_head *dx_root_bh)
4355{
4356	int ret;
4357	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4358	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4359	struct ocfs2_dx_root_block *dx_root;
4360	struct inode *dx_alloc_inode = NULL;
4361	struct buffer_head *dx_alloc_bh = NULL;
4362	handle_t *handle;
4363	u64 blk;
4364	u16 bit;
4365	u64 bg_blkno;
4366
4367	dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4368
4369	dx_alloc_inode = ocfs2_get_system_file_inode(osb,
4370					EXTENT_ALLOC_SYSTEM_INODE,
4371					le16_to_cpu(dx_root->dr_suballoc_slot));
4372	if (!dx_alloc_inode) {
4373		ret = -ENOMEM;
4374		mlog_errno(ret);
4375		goto out;
4376	}
4377	mutex_lock(&dx_alloc_inode->i_mutex);
4378
4379	ret = ocfs2_inode_lock(dx_alloc_inode, &dx_alloc_bh, 1);
4380	if (ret) {
4381		mlog_errno(ret);
4382		goto out_mutex;
4383	}
4384
4385	handle = ocfs2_start_trans(osb, OCFS2_DX_ROOT_REMOVE_CREDITS);
4386	if (IS_ERR(handle)) {
4387		ret = PTR_ERR(handle);
4388		mlog_errno(ret);
4389		goto out_unlock;
4390	}
4391
4392	ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
4393				      OCFS2_JOURNAL_ACCESS_WRITE);
4394	if (ret) {
4395		mlog_errno(ret);
4396		goto out_commit;
4397	}
4398
4399	spin_lock(&OCFS2_I(dir)->ip_lock);
4400	OCFS2_I(dir)->ip_dyn_features &= ~OCFS2_INDEXED_DIR_FL;
4401	di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
4402	spin_unlock(&OCFS2_I(dir)->ip_lock);
4403	di->i_dx_root = cpu_to_le64(0ULL);
4404	ocfs2_update_inode_fsync_trans(handle, dir, 1);
4405
4406	ocfs2_journal_dirty(handle, di_bh);
4407
4408	blk = le64_to_cpu(dx_root->dr_blkno);
4409	bit = le16_to_cpu(dx_root->dr_suballoc_bit);
4410	if (dx_root->dr_suballoc_loc)
4411		bg_blkno = le64_to_cpu(dx_root->dr_suballoc_loc);
4412	else
4413		bg_blkno = ocfs2_which_suballoc_group(blk, bit);
4414	ret = ocfs2_free_suballoc_bits(handle, dx_alloc_inode, dx_alloc_bh,
4415				       bit, bg_blkno, 1);
4416	if (ret)
4417		mlog_errno(ret);
4418
4419out_commit:
4420	ocfs2_commit_trans(osb, handle);
4421
4422out_unlock:
4423	ocfs2_inode_unlock(dx_alloc_inode, 1);
4424
4425out_mutex:
4426	mutex_unlock(&dx_alloc_inode->i_mutex);
4427	brelse(dx_alloc_bh);
4428out:
4429	iput(dx_alloc_inode);
4430	return ret;
4431}
4432
4433int ocfs2_dx_dir_truncate(struct inode *dir, struct buffer_head *di_bh)
4434{
4435	int ret;
4436	unsigned int uninitialized_var(clen);
4437	u32 major_hash = UINT_MAX, p_cpos, uninitialized_var(cpos);
4438	u64 uninitialized_var(blkno);
4439	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4440	struct buffer_head *dx_root_bh = NULL;
4441	struct ocfs2_dx_root_block *dx_root;
4442	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4443	struct ocfs2_cached_dealloc_ctxt dealloc;
4444	struct ocfs2_extent_tree et;
4445
4446	ocfs2_init_dealloc_ctxt(&dealloc);
4447
4448	if (!ocfs2_dir_indexed(dir))
4449		return 0;
4450
4451	ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4452	if (ret) {
4453		mlog_errno(ret);
4454		goto out;
4455	}
4456	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4457
4458	if (ocfs2_dx_root_inline(dx_root))
4459		goto remove_index;
4460
4461	ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4462
4463	/* XXX: What if dr_clusters is too large? */
4464	while (le32_to_cpu(dx_root->dr_clusters)) {
4465		ret = ocfs2_dx_dir_lookup_rec(dir, &dx_root->dr_list,
4466					      major_hash, &cpos, &blkno, &clen);
4467		if (ret) {
4468			mlog_errno(ret);
4469			goto out;
4470		}
4471
4472		p_cpos = ocfs2_blocks_to_clusters(dir->i_sb, blkno);
4473
4474		ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen, 0,
4475					       &dealloc, 0, false);
4476		if (ret) {
4477			mlog_errno(ret);
4478			goto out;
4479		}
4480
4481		if (cpos == 0)
4482			break;
4483
4484		major_hash = cpos - 1;
4485	}
4486
4487remove_index:
4488	ret = ocfs2_dx_dir_remove_index(dir, di_bh, dx_root_bh);
4489	if (ret) {
4490		mlog_errno(ret);
4491		goto out;
4492	}
4493
4494	ocfs2_remove_from_cache(INODE_CACHE(dir), dx_root_bh);
4495out:
4496	ocfs2_schedule_truncate_log_flush(osb, 1);
4497	ocfs2_run_deallocs(osb, &dealloc);
4498
4499	brelse(dx_root_bh);
4500	return ret;
4501}
4502