xref: /openbmc/linux/fs/ocfs2/dir.c (revision b6dcefde)
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 Linux 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 #define MLOG_MASK_PREFIX ML_NAMEI
47 #include <cluster/masklog.h>
48 
49 #include "ocfs2.h"
50 
51 #include "alloc.h"
52 #include "blockcheck.h"
53 #include "dir.h"
54 #include "dlmglue.h"
55 #include "extent_map.h"
56 #include "file.h"
57 #include "inode.h"
58 #include "journal.h"
59 #include "namei.h"
60 #include "suballoc.h"
61 #include "super.h"
62 #include "sysfile.h"
63 #include "uptodate.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 #define NAMEI_RA_INDEX(c,b)  (((c) * NAMEI_RA_BLOCKS) + (b))
71 
72 static unsigned char ocfs2_filetype_table[] = {
73 	DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
74 };
75 
76 static int ocfs2_do_extend_dir(struct super_block *sb,
77 			       handle_t *handle,
78 			       struct inode *dir,
79 			       struct buffer_head *parent_fe_bh,
80 			       struct ocfs2_alloc_context *data_ac,
81 			       struct ocfs2_alloc_context *meta_ac,
82 			       struct buffer_head **new_bh);
83 static int ocfs2_dir_indexed(struct inode *inode);
84 
85 /*
86  * These are distinct checks because future versions of the file system will
87  * want to have a trailing dirent structure independent of indexing.
88  */
89 static int ocfs2_supports_dir_trailer(struct inode *dir)
90 {
91 	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
92 
93 	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
94 		return 0;
95 
96 	return ocfs2_meta_ecc(osb) || ocfs2_dir_indexed(dir);
97 }
98 
99 /*
100  * "new' here refers to the point at which we're creating a new
101  * directory via "mkdir()", but also when we're expanding an inline
102  * directory. In either case, we don't yet have the indexing bit set
103  * on the directory, so the standard checks will fail in when metaecc
104  * is turned off. Only directory-initialization type functions should
105  * use this then. Everything else wants ocfs2_supports_dir_trailer()
106  */
107 static int ocfs2_new_dir_wants_trailer(struct inode *dir)
108 {
109 	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
110 
111 	return ocfs2_meta_ecc(osb) ||
112 		ocfs2_supports_indexed_dirs(osb);
113 }
114 
115 static inline unsigned int ocfs2_dir_trailer_blk_off(struct super_block *sb)
116 {
117 	return sb->s_blocksize - sizeof(struct ocfs2_dir_block_trailer);
118 }
119 
120 #define ocfs2_trailer_from_bh(_bh, _sb) ((struct ocfs2_dir_block_trailer *) ((_bh)->b_data + ocfs2_dir_trailer_blk_off((_sb))))
121 
122 /* XXX ocfs2_block_dqtrailer() is similar but not quite - can we make
123  * them more consistent? */
124 struct ocfs2_dir_block_trailer *ocfs2_dir_trailer_from_size(int blocksize,
125 							    void *data)
126 {
127 	char *p = data;
128 
129 	p += blocksize - sizeof(struct ocfs2_dir_block_trailer);
130 	return (struct ocfs2_dir_block_trailer *)p;
131 }
132 
133 /*
134  * XXX: This is executed once on every dirent. We should consider optimizing
135  * it.
136  */
137 static int ocfs2_skip_dir_trailer(struct inode *dir,
138 				  struct ocfs2_dir_entry *de,
139 				  unsigned long offset,
140 				  unsigned long blklen)
141 {
142 	unsigned long toff = blklen - sizeof(struct ocfs2_dir_block_trailer);
143 
144 	if (!ocfs2_supports_dir_trailer(dir))
145 		return 0;
146 
147 	if (offset != toff)
148 		return 0;
149 
150 	return 1;
151 }
152 
153 static void ocfs2_init_dir_trailer(struct inode *inode,
154 				   struct buffer_head *bh, u16 rec_len)
155 {
156 	struct ocfs2_dir_block_trailer *trailer;
157 
158 	trailer = ocfs2_trailer_from_bh(bh, inode->i_sb);
159 	strcpy(trailer->db_signature, OCFS2_DIR_TRAILER_SIGNATURE);
160 	trailer->db_compat_rec_len =
161 			cpu_to_le16(sizeof(struct ocfs2_dir_block_trailer));
162 	trailer->db_parent_dinode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
163 	trailer->db_blkno = cpu_to_le64(bh->b_blocknr);
164 	trailer->db_free_rec_len = cpu_to_le16(rec_len);
165 }
166 /*
167  * Link an unindexed block with a dir trailer structure into the index free
168  * list. This function will modify dirdata_bh, but assumes you've already
169  * passed it to the journal.
170  */
171 static int ocfs2_dx_dir_link_trailer(struct inode *dir, handle_t *handle,
172 				     struct buffer_head *dx_root_bh,
173 				     struct buffer_head *dirdata_bh)
174 {
175 	int ret;
176 	struct ocfs2_dx_root_block *dx_root;
177 	struct ocfs2_dir_block_trailer *trailer;
178 
179 	ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
180 				      OCFS2_JOURNAL_ACCESS_WRITE);
181 	if (ret) {
182 		mlog_errno(ret);
183 		goto out;
184 	}
185 	trailer = ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
186 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
187 
188 	trailer->db_free_next = dx_root->dr_free_blk;
189 	dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
190 
191 	ocfs2_journal_dirty(handle, dx_root_bh);
192 
193 out:
194 	return ret;
195 }
196 
197 static int ocfs2_free_list_at_root(struct ocfs2_dir_lookup_result *res)
198 {
199 	return res->dl_prev_leaf_bh == NULL;
200 }
201 
202 void ocfs2_free_dir_lookup_result(struct ocfs2_dir_lookup_result *res)
203 {
204 	brelse(res->dl_dx_root_bh);
205 	brelse(res->dl_leaf_bh);
206 	brelse(res->dl_dx_leaf_bh);
207 	brelse(res->dl_prev_leaf_bh);
208 }
209 
210 static int ocfs2_dir_indexed(struct inode *inode)
211 {
212 	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INDEXED_DIR_FL)
213 		return 1;
214 	return 0;
215 }
216 
217 static inline int ocfs2_dx_root_inline(struct ocfs2_dx_root_block *dx_root)
218 {
219 	return dx_root->dr_flags & OCFS2_DX_FLAG_INLINE;
220 }
221 
222 /*
223  * Hashing code adapted from ext3
224  */
225 #define DELTA 0x9E3779B9
226 
227 static void TEA_transform(__u32 buf[4], __u32 const in[])
228 {
229 	__u32	sum = 0;
230 	__u32	b0 = buf[0], b1 = buf[1];
231 	__u32	a = in[0], b = in[1], c = in[2], d = in[3];
232 	int	n = 16;
233 
234 	do {
235 		sum += DELTA;
236 		b0 += ((b1 << 4)+a) ^ (b1+sum) ^ ((b1 >> 5)+b);
237 		b1 += ((b0 << 4)+c) ^ (b0+sum) ^ ((b0 >> 5)+d);
238 	} while (--n);
239 
240 	buf[0] += b0;
241 	buf[1] += b1;
242 }
243 
244 static void str2hashbuf(const char *msg, int len, __u32 *buf, int num)
245 {
246 	__u32	pad, val;
247 	int	i;
248 
249 	pad = (__u32)len | ((__u32)len << 8);
250 	pad |= pad << 16;
251 
252 	val = pad;
253 	if (len > num*4)
254 		len = num * 4;
255 	for (i = 0; i < len; i++) {
256 		if ((i % 4) == 0)
257 			val = pad;
258 		val = msg[i] + (val << 8);
259 		if ((i % 4) == 3) {
260 			*buf++ = val;
261 			val = pad;
262 			num--;
263 		}
264 	}
265 	if (--num >= 0)
266 		*buf++ = val;
267 	while (--num >= 0)
268 		*buf++ = pad;
269 }
270 
271 static void ocfs2_dx_dir_name_hash(struct inode *dir, const char *name, int len,
272 				   struct ocfs2_dx_hinfo *hinfo)
273 {
274 	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
275 	const char	*p;
276 	__u32		in[8], buf[4];
277 
278 	/*
279 	 * XXX: Is this really necessary, if the index is never looked
280 	 * at by readdir? Is a hash value of '0' a bad idea?
281 	 */
282 	if ((len == 1 && !strncmp(".", name, 1)) ||
283 	    (len == 2 && !strncmp("..", name, 2))) {
284 		buf[0] = buf[1] = 0;
285 		goto out;
286 	}
287 
288 #ifdef OCFS2_DEBUG_DX_DIRS
289 	/*
290 	 * This makes it very easy to debug indexing problems. We
291 	 * should never allow this to be selected without hand editing
292 	 * this file though.
293 	 */
294 	buf[0] = buf[1] = len;
295 	goto out;
296 #endif
297 
298 	memcpy(buf, osb->osb_dx_seed, sizeof(buf));
299 
300 	p = name;
301 	while (len > 0) {
302 		str2hashbuf(p, len, in, 4);
303 		TEA_transform(buf, in);
304 		len -= 16;
305 		p += 16;
306 	}
307 
308 out:
309 	hinfo->major_hash = buf[0];
310 	hinfo->minor_hash = buf[1];
311 }
312 
313 /*
314  * bh passed here can be an inode block or a dir data block, depending
315  * on the inode inline data flag.
316  */
317 static int ocfs2_check_dir_entry(struct inode * dir,
318 				 struct ocfs2_dir_entry * de,
319 				 struct buffer_head * bh,
320 				 unsigned long offset)
321 {
322 	const char *error_msg = NULL;
323 	const int rlen = le16_to_cpu(de->rec_len);
324 
325 	if (rlen < OCFS2_DIR_REC_LEN(1))
326 		error_msg = "rec_len is smaller than minimal";
327 	else if (rlen % 4 != 0)
328 		error_msg = "rec_len % 4 != 0";
329 	else if (rlen < OCFS2_DIR_REC_LEN(de->name_len))
330 		error_msg = "rec_len is too small for name_len";
331 	else if (((char *) de - bh->b_data) + rlen > dir->i_sb->s_blocksize)
332 		error_msg = "directory entry across blocks";
333 
334 	if (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 	return error_msg == NULL ? 1 : 0;
341 }
342 
343 static inline int ocfs2_match(int len,
344 			      const char * const name,
345 			      struct ocfs2_dir_entry *de)
346 {
347 	if (len != de->name_len)
348 		return 0;
349 	if (!de->inode)
350 		return 0;
351 	return !memcmp(name, de->name, len);
352 }
353 
354 /*
355  * Returns 0 if not found, -1 on failure, and 1 on success
356  */
357 static int inline ocfs2_search_dirblock(struct buffer_head *bh,
358 					struct inode *dir,
359 					const char *name, int namelen,
360 					unsigned long offset,
361 					char *first_de,
362 					unsigned int bytes,
363 					struct ocfs2_dir_entry **res_dir)
364 {
365 	struct ocfs2_dir_entry *de;
366 	char *dlimit, *de_buf;
367 	int de_len;
368 	int ret = 0;
369 
370 	mlog_entry_void();
371 
372 	de_buf = first_de;
373 	dlimit = de_buf + bytes;
374 
375 	while (de_buf < dlimit) {
376 		/* this code is executed quadratically often */
377 		/* do minimal checking `by hand' */
378 
379 		de = (struct ocfs2_dir_entry *) de_buf;
380 
381 		if (de_buf + namelen <= dlimit &&
382 		    ocfs2_match(namelen, name, de)) {
383 			/* found a match - just to be sure, do a full check */
384 			if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
385 				ret = -1;
386 				goto bail;
387 			}
388 			*res_dir = de;
389 			ret = 1;
390 			goto bail;
391 		}
392 
393 		/* prevent looping on a bad block */
394 		de_len = le16_to_cpu(de->rec_len);
395 		if (de_len <= 0) {
396 			ret = -1;
397 			goto bail;
398 		}
399 
400 		de_buf += de_len;
401 		offset += de_len;
402 	}
403 
404 bail:
405 	mlog_exit(ret);
406 	return ret;
407 }
408 
409 static struct buffer_head *ocfs2_find_entry_id(const char *name,
410 					       int namelen,
411 					       struct inode *dir,
412 					       struct ocfs2_dir_entry **res_dir)
413 {
414 	int ret, found;
415 	struct buffer_head *di_bh = NULL;
416 	struct ocfs2_dinode *di;
417 	struct ocfs2_inline_data *data;
418 
419 	ret = ocfs2_read_inode_block(dir, &di_bh);
420 	if (ret) {
421 		mlog_errno(ret);
422 		goto out;
423 	}
424 
425 	di = (struct ocfs2_dinode *)di_bh->b_data;
426 	data = &di->id2.i_data;
427 
428 	found = ocfs2_search_dirblock(di_bh, dir, name, namelen, 0,
429 				      data->id_data, i_size_read(dir), res_dir);
430 	if (found == 1)
431 		return di_bh;
432 
433 	brelse(di_bh);
434 out:
435 	return NULL;
436 }
437 
438 static int ocfs2_validate_dir_block(struct super_block *sb,
439 				    struct buffer_head *bh)
440 {
441 	int rc;
442 	struct ocfs2_dir_block_trailer *trailer =
443 		ocfs2_trailer_from_bh(bh, sb);
444 
445 
446 	/*
447 	 * We don't validate dirents here, that's handled
448 	 * in-place when the code walks them.
449 	 */
450 	mlog(0, "Validating dirblock %llu\n",
451 	     (unsigned long long)bh->b_blocknr);
452 
453 	BUG_ON(!buffer_uptodate(bh));
454 
455 	/*
456 	 * If the ecc fails, we return the error but otherwise
457 	 * leave the filesystem running.  We know any error is
458 	 * local to this block.
459 	 *
460 	 * Note that we are safe to call this even if the directory
461 	 * doesn't have a trailer.  Filesystems without metaecc will do
462 	 * nothing, and filesystems with it will have one.
463 	 */
464 	rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &trailer->db_check);
465 	if (rc)
466 		mlog(ML_ERROR, "Checksum failed for dinode %llu\n",
467 		     (unsigned long long)bh->b_blocknr);
468 
469 	return rc;
470 }
471 
472 /*
473  * Validate a directory trailer.
474  *
475  * We check the trailer here rather than in ocfs2_validate_dir_block()
476  * because that function doesn't have the inode to test.
477  */
478 static int ocfs2_check_dir_trailer(struct inode *dir, struct buffer_head *bh)
479 {
480 	int rc = 0;
481 	struct ocfs2_dir_block_trailer *trailer;
482 
483 	trailer = ocfs2_trailer_from_bh(bh, dir->i_sb);
484 	if (!OCFS2_IS_VALID_DIR_TRAILER(trailer)) {
485 		rc = -EINVAL;
486 		ocfs2_error(dir->i_sb,
487 			    "Invalid dirblock #%llu: "
488 			    "signature = %.*s\n",
489 			    (unsigned long long)bh->b_blocknr, 7,
490 			    trailer->db_signature);
491 		goto out;
492 	}
493 	if (le64_to_cpu(trailer->db_blkno) != bh->b_blocknr) {
494 		rc = -EINVAL;
495 		ocfs2_error(dir->i_sb,
496 			    "Directory block #%llu has an invalid "
497 			    "db_blkno of %llu",
498 			    (unsigned long long)bh->b_blocknr,
499 			    (unsigned long long)le64_to_cpu(trailer->db_blkno));
500 		goto out;
501 	}
502 	if (le64_to_cpu(trailer->db_parent_dinode) !=
503 	    OCFS2_I(dir)->ip_blkno) {
504 		rc = -EINVAL;
505 		ocfs2_error(dir->i_sb,
506 			    "Directory block #%llu on dinode "
507 			    "#%llu has an invalid parent_dinode "
508 			    "of %llu",
509 			    (unsigned long long)bh->b_blocknr,
510 			    (unsigned long long)OCFS2_I(dir)->ip_blkno,
511 			    (unsigned long long)le64_to_cpu(trailer->db_blkno));
512 		goto out;
513 	}
514 out:
515 	return rc;
516 }
517 
518 /*
519  * This function forces all errors to -EIO for consistency with its
520  * predecessor, ocfs2_bread().  We haven't audited what returning the
521  * real error codes would do to callers.  We log the real codes with
522  * mlog_errno() before we squash them.
523  */
524 static int ocfs2_read_dir_block(struct inode *inode, u64 v_block,
525 				struct buffer_head **bh, int flags)
526 {
527 	int rc = 0;
528 	struct buffer_head *tmp = *bh;
529 
530 	rc = ocfs2_read_virt_blocks(inode, v_block, 1, &tmp, flags,
531 				    ocfs2_validate_dir_block);
532 	if (rc) {
533 		mlog_errno(rc);
534 		goto out;
535 	}
536 
537 	if (!(flags & OCFS2_BH_READAHEAD) &&
538 	    ocfs2_supports_dir_trailer(inode)) {
539 		rc = ocfs2_check_dir_trailer(inode, tmp);
540 		if (rc) {
541 			if (!*bh)
542 				brelse(tmp);
543 			mlog_errno(rc);
544 			goto out;
545 		}
546 	}
547 
548 	/* If ocfs2_read_virt_blocks() got us a new bh, pass it up. */
549 	if (!*bh)
550 		*bh = tmp;
551 
552 out:
553 	return rc ? -EIO : 0;
554 }
555 
556 /*
557  * Read the block at 'phys' which belongs to this directory
558  * inode. This function does no virtual->physical block translation -
559  * what's passed in is assumed to be a valid directory block.
560  */
561 static int ocfs2_read_dir_block_direct(struct inode *dir, u64 phys,
562 				       struct buffer_head **bh)
563 {
564 	int ret;
565 	struct buffer_head *tmp = *bh;
566 
567 	ret = ocfs2_read_block(INODE_CACHE(dir), phys, &tmp,
568 			       ocfs2_validate_dir_block);
569 	if (ret) {
570 		mlog_errno(ret);
571 		goto out;
572 	}
573 
574 	if (ocfs2_supports_dir_trailer(dir)) {
575 		ret = ocfs2_check_dir_trailer(dir, tmp);
576 		if (ret) {
577 			if (!*bh)
578 				brelse(tmp);
579 			mlog_errno(ret);
580 			goto out;
581 		}
582 	}
583 
584 	if (!ret && !*bh)
585 		*bh = tmp;
586 out:
587 	return ret;
588 }
589 
590 static int ocfs2_validate_dx_root(struct super_block *sb,
591 				  struct buffer_head *bh)
592 {
593 	int ret;
594 	struct ocfs2_dx_root_block *dx_root;
595 
596 	BUG_ON(!buffer_uptodate(bh));
597 
598 	dx_root = (struct ocfs2_dx_root_block *) bh->b_data;
599 
600 	ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_root->dr_check);
601 	if (ret) {
602 		mlog(ML_ERROR,
603 		     "Checksum failed for dir index root block %llu\n",
604 		     (unsigned long long)bh->b_blocknr);
605 		return ret;
606 	}
607 
608 	if (!OCFS2_IS_VALID_DX_ROOT(dx_root)) {
609 		ocfs2_error(sb,
610 			    "Dir Index Root # %llu has bad signature %.*s",
611 			    (unsigned long long)le64_to_cpu(dx_root->dr_blkno),
612 			    7, dx_root->dr_signature);
613 		return -EINVAL;
614 	}
615 
616 	return 0;
617 }
618 
619 static int ocfs2_read_dx_root(struct inode *dir, struct ocfs2_dinode *di,
620 			      struct buffer_head **dx_root_bh)
621 {
622 	int ret;
623 	u64 blkno = le64_to_cpu(di->i_dx_root);
624 	struct buffer_head *tmp = *dx_root_bh;
625 
626 	ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
627 			       ocfs2_validate_dx_root);
628 
629 	/* If ocfs2_read_block() got us a new bh, pass it up. */
630 	if (!ret && !*dx_root_bh)
631 		*dx_root_bh = tmp;
632 
633 	return ret;
634 }
635 
636 static int ocfs2_validate_dx_leaf(struct super_block *sb,
637 				  struct buffer_head *bh)
638 {
639 	int ret;
640 	struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)bh->b_data;
641 
642 	BUG_ON(!buffer_uptodate(bh));
643 
644 	ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_leaf->dl_check);
645 	if (ret) {
646 		mlog(ML_ERROR,
647 		     "Checksum failed for dir index leaf block %llu\n",
648 		     (unsigned long long)bh->b_blocknr);
649 		return ret;
650 	}
651 
652 	if (!OCFS2_IS_VALID_DX_LEAF(dx_leaf)) {
653 		ocfs2_error(sb, "Dir Index Leaf has bad signature %.*s",
654 			    7, dx_leaf->dl_signature);
655 		return -EROFS;
656 	}
657 
658 	return 0;
659 }
660 
661 static int ocfs2_read_dx_leaf(struct inode *dir, u64 blkno,
662 			      struct buffer_head **dx_leaf_bh)
663 {
664 	int ret;
665 	struct buffer_head *tmp = *dx_leaf_bh;
666 
667 	ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
668 			       ocfs2_validate_dx_leaf);
669 
670 	/* If ocfs2_read_block() got us a new bh, pass it up. */
671 	if (!ret && !*dx_leaf_bh)
672 		*dx_leaf_bh = tmp;
673 
674 	return ret;
675 }
676 
677 /*
678  * Read a series of dx_leaf blocks. This expects all buffer_head
679  * pointers to be NULL on function entry.
680  */
681 static int ocfs2_read_dx_leaves(struct inode *dir, u64 start, int num,
682 				struct buffer_head **dx_leaf_bhs)
683 {
684 	int ret;
685 
686 	ret = ocfs2_read_blocks(INODE_CACHE(dir), start, num, dx_leaf_bhs, 0,
687 				ocfs2_validate_dx_leaf);
688 	if (ret)
689 		mlog_errno(ret);
690 
691 	return ret;
692 }
693 
694 static struct buffer_head *ocfs2_find_entry_el(const char *name, int namelen,
695 					       struct inode *dir,
696 					       struct ocfs2_dir_entry **res_dir)
697 {
698 	struct super_block *sb;
699 	struct buffer_head *bh_use[NAMEI_RA_SIZE];
700 	struct buffer_head *bh, *ret = NULL;
701 	unsigned long start, block, b;
702 	int ra_max = 0;		/* Number of bh's in the readahead
703 				   buffer, bh_use[] */
704 	int ra_ptr = 0;		/* Current index into readahead
705 				   buffer */
706 	int num = 0;
707 	int nblocks, i, err;
708 
709 	mlog_entry_void();
710 
711 	sb = dir->i_sb;
712 
713 	nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
714 	start = OCFS2_I(dir)->ip_dir_start_lookup;
715 	if (start >= nblocks)
716 		start = 0;
717 	block = start;
718 
719 restart:
720 	do {
721 		/*
722 		 * We deal with the read-ahead logic here.
723 		 */
724 		if (ra_ptr >= ra_max) {
725 			/* Refill the readahead buffer */
726 			ra_ptr = 0;
727 			b = block;
728 			for (ra_max = 0; ra_max < NAMEI_RA_SIZE; ra_max++) {
729 				/*
730 				 * Terminate if we reach the end of the
731 				 * directory and must wrap, or if our
732 				 * search has finished at this block.
733 				 */
734 				if (b >= nblocks || (num && block == start)) {
735 					bh_use[ra_max] = NULL;
736 					break;
737 				}
738 				num++;
739 
740 				bh = NULL;
741 				err = ocfs2_read_dir_block(dir, b++, &bh,
742 							   OCFS2_BH_READAHEAD);
743 				bh_use[ra_max] = bh;
744 			}
745 		}
746 		if ((bh = bh_use[ra_ptr++]) == NULL)
747 			goto next;
748 		if (ocfs2_read_dir_block(dir, block, &bh, 0)) {
749 			/* read error, skip block & hope for the best.
750 			 * ocfs2_read_dir_block() has released the bh. */
751 			ocfs2_error(dir->i_sb, "reading directory %llu, "
752 				    "offset %lu\n",
753 				    (unsigned long long)OCFS2_I(dir)->ip_blkno,
754 				    block);
755 			goto next;
756 		}
757 		i = ocfs2_search_dirblock(bh, dir, name, namelen,
758 					  block << sb->s_blocksize_bits,
759 					  bh->b_data, sb->s_blocksize,
760 					  res_dir);
761 		if (i == 1) {
762 			OCFS2_I(dir)->ip_dir_start_lookup = block;
763 			ret = bh;
764 			goto cleanup_and_exit;
765 		} else {
766 			brelse(bh);
767 			if (i < 0)
768 				goto cleanup_and_exit;
769 		}
770 	next:
771 		if (++block >= nblocks)
772 			block = 0;
773 	} while (block != start);
774 
775 	/*
776 	 * If the directory has grown while we were searching, then
777 	 * search the last part of the directory before giving up.
778 	 */
779 	block = nblocks;
780 	nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
781 	if (block < nblocks) {
782 		start = 0;
783 		goto restart;
784 	}
785 
786 cleanup_and_exit:
787 	/* Clean up the read-ahead blocks */
788 	for (; ra_ptr < ra_max; ra_ptr++)
789 		brelse(bh_use[ra_ptr]);
790 
791 	mlog_exit_ptr(ret);
792 	return ret;
793 }
794 
795 static int ocfs2_dx_dir_lookup_rec(struct inode *inode,
796 				   struct ocfs2_extent_list *el,
797 				   u32 major_hash,
798 				   u32 *ret_cpos,
799 				   u64 *ret_phys_blkno,
800 				   unsigned int *ret_clen)
801 {
802 	int ret = 0, i, found;
803 	struct buffer_head *eb_bh = NULL;
804 	struct ocfs2_extent_block *eb;
805 	struct ocfs2_extent_rec *rec = NULL;
806 
807 	if (el->l_tree_depth) {
808 		ret = ocfs2_find_leaf(INODE_CACHE(inode), el, major_hash,
809 				      &eb_bh);
810 		if (ret) {
811 			mlog_errno(ret);
812 			goto out;
813 		}
814 
815 		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
816 		el = &eb->h_list;
817 
818 		if (el->l_tree_depth) {
819 			ocfs2_error(inode->i_sb,
820 				    "Inode %lu has non zero tree depth in "
821 				    "btree tree block %llu\n", inode->i_ino,
822 				    (unsigned long long)eb_bh->b_blocknr);
823 			ret = -EROFS;
824 			goto out;
825 		}
826 	}
827 
828 	found = 0;
829 	for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
830 		rec = &el->l_recs[i];
831 
832 		if (le32_to_cpu(rec->e_cpos) <= major_hash) {
833 			found = 1;
834 			break;
835 		}
836 	}
837 
838 	if (!found) {
839 		ocfs2_error(inode->i_sb, "Inode %lu has bad extent "
840 			    "record (%u, %u, 0) in btree", inode->i_ino,
841 			    le32_to_cpu(rec->e_cpos),
842 			    ocfs2_rec_clusters(el, rec));
843 		ret = -EROFS;
844 		goto out;
845 	}
846 
847 	if (ret_phys_blkno)
848 		*ret_phys_blkno = le64_to_cpu(rec->e_blkno);
849 	if (ret_cpos)
850 		*ret_cpos = le32_to_cpu(rec->e_cpos);
851 	if (ret_clen)
852 		*ret_clen = le16_to_cpu(rec->e_leaf_clusters);
853 
854 out:
855 	brelse(eb_bh);
856 	return ret;
857 }
858 
859 /*
860  * Returns the block index, from the start of the cluster which this
861  * hash belongs too.
862  */
863 static inline unsigned int __ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
864 						   u32 minor_hash)
865 {
866 	return minor_hash & osb->osb_dx_mask;
867 }
868 
869 static inline unsigned int ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
870 					  struct ocfs2_dx_hinfo *hinfo)
871 {
872 	return __ocfs2_dx_dir_hash_idx(osb, hinfo->minor_hash);
873 }
874 
875 static int ocfs2_dx_dir_lookup(struct inode *inode,
876 			       struct ocfs2_extent_list *el,
877 			       struct ocfs2_dx_hinfo *hinfo,
878 			       u32 *ret_cpos,
879 			       u64 *ret_phys_blkno)
880 {
881 	int ret = 0;
882 	unsigned int cend, uninitialized_var(clen);
883 	u32 uninitialized_var(cpos);
884 	u64 uninitialized_var(blkno);
885 	u32 name_hash = hinfo->major_hash;
886 
887 	ret = ocfs2_dx_dir_lookup_rec(inode, el, name_hash, &cpos, &blkno,
888 				      &clen);
889 	if (ret) {
890 		mlog_errno(ret);
891 		goto out;
892 	}
893 
894 	cend = cpos + clen;
895 	if (name_hash >= cend) {
896 		/* We want the last cluster */
897 		blkno += ocfs2_clusters_to_blocks(inode->i_sb, clen - 1);
898 		cpos += clen - 1;
899 	} else {
900 		blkno += ocfs2_clusters_to_blocks(inode->i_sb,
901 						  name_hash - cpos);
902 		cpos = name_hash;
903 	}
904 
905 	/*
906 	 * We now have the cluster which should hold our entry. To
907 	 * find the exact block from the start of the cluster to
908 	 * search, we take the lower bits of the hash.
909 	 */
910 	blkno += ocfs2_dx_dir_hash_idx(OCFS2_SB(inode->i_sb), hinfo);
911 
912 	if (ret_phys_blkno)
913 		*ret_phys_blkno = blkno;
914 	if (ret_cpos)
915 		*ret_cpos = cpos;
916 
917 out:
918 
919 	return ret;
920 }
921 
922 static int ocfs2_dx_dir_search(const char *name, int namelen,
923 			       struct inode *dir,
924 			       struct ocfs2_dx_root_block *dx_root,
925 			       struct ocfs2_dir_lookup_result *res)
926 {
927 	int ret, i, found;
928 	u64 uninitialized_var(phys);
929 	struct buffer_head *dx_leaf_bh = NULL;
930 	struct ocfs2_dx_leaf *dx_leaf;
931 	struct ocfs2_dx_entry *dx_entry = NULL;
932 	struct buffer_head *dir_ent_bh = NULL;
933 	struct ocfs2_dir_entry *dir_ent = NULL;
934 	struct ocfs2_dx_hinfo *hinfo = &res->dl_hinfo;
935 	struct ocfs2_extent_list *dr_el;
936 	struct ocfs2_dx_entry_list *entry_list;
937 
938 	ocfs2_dx_dir_name_hash(dir, name, namelen, &res->dl_hinfo);
939 
940 	if (ocfs2_dx_root_inline(dx_root)) {
941 		entry_list = &dx_root->dr_entries;
942 		goto search;
943 	}
944 
945 	dr_el = &dx_root->dr_list;
946 
947 	ret = ocfs2_dx_dir_lookup(dir, dr_el, hinfo, NULL, &phys);
948 	if (ret) {
949 		mlog_errno(ret);
950 		goto out;
951 	}
952 
953 	mlog(0, "Dir %llu: name: \"%.*s\", lookup of hash: %u.0x%x "
954 	     "returns: %llu\n",
955 	     (unsigned long long)OCFS2_I(dir)->ip_blkno,
956 	     namelen, name, hinfo->major_hash, hinfo->minor_hash,
957 	     (unsigned long long)phys);
958 
959 	ret = ocfs2_read_dx_leaf(dir, phys, &dx_leaf_bh);
960 	if (ret) {
961 		mlog_errno(ret);
962 		goto out;
963 	}
964 
965 	dx_leaf = (struct ocfs2_dx_leaf *) dx_leaf_bh->b_data;
966 
967 	mlog(0, "leaf info: num_used: %d, count: %d\n",
968 	     le16_to_cpu(dx_leaf->dl_list.de_num_used),
969 	     le16_to_cpu(dx_leaf->dl_list.de_count));
970 
971 	entry_list = &dx_leaf->dl_list;
972 
973 search:
974 	/*
975 	 * Empty leaf is legal, so no need to check for that.
976 	 */
977 	found = 0;
978 	for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
979 		dx_entry = &entry_list->de_entries[i];
980 
981 		if (hinfo->major_hash != le32_to_cpu(dx_entry->dx_major_hash)
982 		    || hinfo->minor_hash != le32_to_cpu(dx_entry->dx_minor_hash))
983 			continue;
984 
985 		/*
986 		 * Search unindexed leaf block now. We're not
987 		 * guaranteed to find anything.
988 		 */
989 		ret = ocfs2_read_dir_block_direct(dir,
990 					  le64_to_cpu(dx_entry->dx_dirent_blk),
991 					  &dir_ent_bh);
992 		if (ret) {
993 			mlog_errno(ret);
994 			goto out;
995 		}
996 
997 		/*
998 		 * XXX: We should check the unindexed block here,
999 		 * before using it.
1000 		 */
1001 
1002 		found = ocfs2_search_dirblock(dir_ent_bh, dir, name, namelen,
1003 					      0, dir_ent_bh->b_data,
1004 					      dir->i_sb->s_blocksize, &dir_ent);
1005 		if (found == 1)
1006 			break;
1007 
1008 		if (found == -1) {
1009 			/* This means we found a bad directory entry. */
1010 			ret = -EIO;
1011 			mlog_errno(ret);
1012 			goto out;
1013 		}
1014 
1015 		brelse(dir_ent_bh);
1016 		dir_ent_bh = NULL;
1017 	}
1018 
1019 	if (found <= 0) {
1020 		ret = -ENOENT;
1021 		goto out;
1022 	}
1023 
1024 	res->dl_leaf_bh = dir_ent_bh;
1025 	res->dl_entry = dir_ent;
1026 	res->dl_dx_leaf_bh = dx_leaf_bh;
1027 	res->dl_dx_entry = dx_entry;
1028 
1029 	ret = 0;
1030 out:
1031 	if (ret) {
1032 		brelse(dx_leaf_bh);
1033 		brelse(dir_ent_bh);
1034 	}
1035 	return ret;
1036 }
1037 
1038 static int ocfs2_find_entry_dx(const char *name, int namelen,
1039 			       struct inode *dir,
1040 			       struct ocfs2_dir_lookup_result *lookup)
1041 {
1042 	int ret;
1043 	struct buffer_head *di_bh = NULL;
1044 	struct ocfs2_dinode *di;
1045 	struct buffer_head *dx_root_bh = NULL;
1046 	struct ocfs2_dx_root_block *dx_root;
1047 
1048 	ret = ocfs2_read_inode_block(dir, &di_bh);
1049 	if (ret) {
1050 		mlog_errno(ret);
1051 		goto out;
1052 	}
1053 
1054 	di = (struct ocfs2_dinode *)di_bh->b_data;
1055 
1056 	ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
1057 	if (ret) {
1058 		mlog_errno(ret);
1059 		goto out;
1060 	}
1061 	dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
1062 
1063 	ret = ocfs2_dx_dir_search(name, namelen, dir, dx_root, lookup);
1064 	if (ret) {
1065 		if (ret != -ENOENT)
1066 			mlog_errno(ret);
1067 		goto out;
1068 	}
1069 
1070 	lookup->dl_dx_root_bh = dx_root_bh;
1071 	dx_root_bh = NULL;
1072 out:
1073 	brelse(di_bh);
1074 	brelse(dx_root_bh);
1075 	return ret;
1076 }
1077 
1078 /*
1079  * Try to find an entry of the provided name within 'dir'.
1080  *
1081  * If nothing was found, -ENOENT is returned. Otherwise, zero is
1082  * returned and the struct 'res' will contain information useful to
1083  * other directory manipulation functions.
1084  *
1085  * Caller can NOT assume anything about the contents of the
1086  * buffer_heads - they are passed back only so that it can be passed
1087  * into any one of the manipulation functions (add entry, delete
1088  * entry, etc). As an example, bh in the extent directory case is a
1089  * data block, in the inline-data case it actually points to an inode,
1090  * in the indexed directory case, multiple buffers are involved.
1091  */
1092 int ocfs2_find_entry(const char *name, int namelen,
1093 		     struct inode *dir, struct ocfs2_dir_lookup_result *lookup)
1094 {
1095 	struct buffer_head *bh;
1096 	struct ocfs2_dir_entry *res_dir = NULL;
1097 
1098 	if (ocfs2_dir_indexed(dir))
1099 		return ocfs2_find_entry_dx(name, namelen, dir, lookup);
1100 
1101 	/*
1102 	 * The unindexed dir code only uses part of the lookup
1103 	 * structure, so there's no reason to push it down further
1104 	 * than this.
1105 	 */
1106 	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1107 		bh = ocfs2_find_entry_id(name, namelen, dir, &res_dir);
1108 	else
1109 		bh = ocfs2_find_entry_el(name, namelen, dir, &res_dir);
1110 
1111 	if (bh == NULL)
1112 		return -ENOENT;
1113 
1114 	lookup->dl_leaf_bh = bh;
1115 	lookup->dl_entry = res_dir;
1116 	return 0;
1117 }
1118 
1119 /*
1120  * Update inode number and type of a previously found directory entry.
1121  */
1122 int ocfs2_update_entry(struct inode *dir, handle_t *handle,
1123 		       struct ocfs2_dir_lookup_result *res,
1124 		       struct inode *new_entry_inode)
1125 {
1126 	int ret;
1127 	ocfs2_journal_access_func access = ocfs2_journal_access_db;
1128 	struct ocfs2_dir_entry *de = res->dl_entry;
1129 	struct buffer_head *de_bh = res->dl_leaf_bh;
1130 
1131 	/*
1132 	 * The same code works fine for both inline-data and extent
1133 	 * based directories, so no need to split this up.  The only
1134 	 * difference is the journal_access function.
1135 	 */
1136 
1137 	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1138 		access = ocfs2_journal_access_di;
1139 
1140 	ret = access(handle, INODE_CACHE(dir), de_bh,
1141 		     OCFS2_JOURNAL_ACCESS_WRITE);
1142 	if (ret) {
1143 		mlog_errno(ret);
1144 		goto out;
1145 	}
1146 
1147 	de->inode = cpu_to_le64(OCFS2_I(new_entry_inode)->ip_blkno);
1148 	ocfs2_set_de_type(de, new_entry_inode->i_mode);
1149 
1150 	ocfs2_journal_dirty(handle, de_bh);
1151 
1152 out:
1153 	return ret;
1154 }
1155 
1156 /*
1157  * __ocfs2_delete_entry deletes a directory entry by merging it with the
1158  * previous entry
1159  */
1160 static int __ocfs2_delete_entry(handle_t *handle, struct inode *dir,
1161 				struct ocfs2_dir_entry *de_del,
1162 				struct buffer_head *bh, char *first_de,
1163 				unsigned int bytes)
1164 {
1165 	struct ocfs2_dir_entry *de, *pde;
1166 	int i, status = -ENOENT;
1167 	ocfs2_journal_access_func access = ocfs2_journal_access_db;
1168 
1169 	mlog_entry("(0x%p, 0x%p, 0x%p, 0x%p)\n", handle, dir, de_del, bh);
1170 
1171 	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1172 		access = ocfs2_journal_access_di;
1173 
1174 	i = 0;
1175 	pde = NULL;
1176 	de = (struct ocfs2_dir_entry *) first_de;
1177 	while (i < bytes) {
1178 		if (!ocfs2_check_dir_entry(dir, de, bh, i)) {
1179 			status = -EIO;
1180 			mlog_errno(status);
1181 			goto bail;
1182 		}
1183 		if (de == de_del)  {
1184 			status = access(handle, INODE_CACHE(dir), bh,
1185 					OCFS2_JOURNAL_ACCESS_WRITE);
1186 			if (status < 0) {
1187 				status = -EIO;
1188 				mlog_errno(status);
1189 				goto bail;
1190 			}
1191 			if (pde)
1192 				le16_add_cpu(&pde->rec_len,
1193 						le16_to_cpu(de->rec_len));
1194 			else
1195 				de->inode = 0;
1196 			dir->i_version++;
1197 			status = ocfs2_journal_dirty(handle, bh);
1198 			goto bail;
1199 		}
1200 		i += le16_to_cpu(de->rec_len);
1201 		pde = de;
1202 		de = (struct ocfs2_dir_entry *)((char *)de + le16_to_cpu(de->rec_len));
1203 	}
1204 bail:
1205 	mlog_exit(status);
1206 	return status;
1207 }
1208 
1209 static unsigned int ocfs2_figure_dirent_hole(struct ocfs2_dir_entry *de)
1210 {
1211 	unsigned int hole;
1212 
1213 	if (le64_to_cpu(de->inode) == 0)
1214 		hole = le16_to_cpu(de->rec_len);
1215 	else
1216 		hole = le16_to_cpu(de->rec_len) -
1217 			OCFS2_DIR_REC_LEN(de->name_len);
1218 
1219 	return hole;
1220 }
1221 
1222 static int ocfs2_find_max_rec_len(struct super_block *sb,
1223 				  struct buffer_head *dirblock_bh)
1224 {
1225 	int size, this_hole, largest_hole = 0;
1226 	char *trailer, *de_buf, *limit, *start = dirblock_bh->b_data;
1227 	struct ocfs2_dir_entry *de;
1228 
1229 	trailer = (char *)ocfs2_trailer_from_bh(dirblock_bh, sb);
1230 	size = ocfs2_dir_trailer_blk_off(sb);
1231 	limit = start + size;
1232 	de_buf = start;
1233 	de = (struct ocfs2_dir_entry *)de_buf;
1234 	do {
1235 		if (de_buf != trailer) {
1236 			this_hole = ocfs2_figure_dirent_hole(de);
1237 			if (this_hole > largest_hole)
1238 				largest_hole = this_hole;
1239 		}
1240 
1241 		de_buf += le16_to_cpu(de->rec_len);
1242 		de = (struct ocfs2_dir_entry *)de_buf;
1243 	} while (de_buf < limit);
1244 
1245 	if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
1246 		return largest_hole;
1247 	return 0;
1248 }
1249 
1250 static void ocfs2_dx_list_remove_entry(struct ocfs2_dx_entry_list *entry_list,
1251 				       int index)
1252 {
1253 	int num_used = le16_to_cpu(entry_list->de_num_used);
1254 
1255 	if (num_used == 1 || index == (num_used - 1))
1256 		goto clear;
1257 
1258 	memmove(&entry_list->de_entries[index],
1259 		&entry_list->de_entries[index + 1],
1260 		(num_used - index - 1)*sizeof(struct ocfs2_dx_entry));
1261 clear:
1262 	num_used--;
1263 	memset(&entry_list->de_entries[num_used], 0,
1264 	       sizeof(struct ocfs2_dx_entry));
1265 	entry_list->de_num_used = cpu_to_le16(num_used);
1266 }
1267 
1268 static int ocfs2_delete_entry_dx(handle_t *handle, struct inode *dir,
1269 				 struct ocfs2_dir_lookup_result *lookup)
1270 {
1271 	int ret, index, max_rec_len, add_to_free_list = 0;
1272 	struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1273 	struct buffer_head *leaf_bh = lookup->dl_leaf_bh;
1274 	struct ocfs2_dx_leaf *dx_leaf;
1275 	struct ocfs2_dx_entry *dx_entry = lookup->dl_dx_entry;
1276 	struct ocfs2_dir_block_trailer *trailer;
1277 	struct ocfs2_dx_root_block *dx_root;
1278 	struct ocfs2_dx_entry_list *entry_list;
1279 
1280 	/*
1281 	 * This function gets a bit messy because we might have to
1282 	 * modify the root block, regardless of whether the indexed
1283 	 * entries are stored inline.
1284 	 */
1285 
1286 	/*
1287 	 * *Only* set 'entry_list' here, based on where we're looking
1288 	 * for the indexed entries. Later, we might still want to
1289 	 * journal both blocks, based on free list state.
1290 	 */
1291 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
1292 	if (ocfs2_dx_root_inline(dx_root)) {
1293 		entry_list = &dx_root->dr_entries;
1294 	} else {
1295 		dx_leaf = (struct ocfs2_dx_leaf *) lookup->dl_dx_leaf_bh->b_data;
1296 		entry_list = &dx_leaf->dl_list;
1297 	}
1298 
1299 	/* Neither of these are a disk corruption - that should have
1300 	 * been caught by lookup, before we got here. */
1301 	BUG_ON(le16_to_cpu(entry_list->de_count) <= 0);
1302 	BUG_ON(le16_to_cpu(entry_list->de_num_used) <= 0);
1303 
1304 	index = (char *)dx_entry - (char *)entry_list->de_entries;
1305 	index /= sizeof(*dx_entry);
1306 
1307 	if (index >= le16_to_cpu(entry_list->de_num_used)) {
1308 		mlog(ML_ERROR, "Dir %llu: Bad dx_entry ptr idx %d, (%p, %p)\n",
1309 		     (unsigned long long)OCFS2_I(dir)->ip_blkno, index,
1310 		     entry_list, dx_entry);
1311 		return -EIO;
1312 	}
1313 
1314 	/*
1315 	 * We know that removal of this dirent will leave enough room
1316 	 * for a new one, so add this block to the free list if it
1317 	 * isn't already there.
1318 	 */
1319 	trailer = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
1320 	if (trailer->db_free_rec_len == 0)
1321 		add_to_free_list = 1;
1322 
1323 	/*
1324 	 * Add the block holding our index into the journal before
1325 	 * removing the unindexed entry. If we get an error return
1326 	 * from __ocfs2_delete_entry(), then it hasn't removed the
1327 	 * entry yet. Likewise, successful return means we *must*
1328 	 * remove the indexed entry.
1329 	 *
1330 	 * We're also careful to journal the root tree block here as
1331 	 * the entry count needs to be updated. Also, we might be
1332 	 * adding to the start of the free list.
1333 	 */
1334 	ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1335 				      OCFS2_JOURNAL_ACCESS_WRITE);
1336 	if (ret) {
1337 		mlog_errno(ret);
1338 		goto out;
1339 	}
1340 
1341 	if (!ocfs2_dx_root_inline(dx_root)) {
1342 		ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
1343 					      lookup->dl_dx_leaf_bh,
1344 					      OCFS2_JOURNAL_ACCESS_WRITE);
1345 		if (ret) {
1346 			mlog_errno(ret);
1347 			goto out;
1348 		}
1349 	}
1350 
1351 	mlog(0, "Dir %llu: delete entry at index: %d\n",
1352 	     (unsigned long long)OCFS2_I(dir)->ip_blkno, index);
1353 
1354 	ret = __ocfs2_delete_entry(handle, dir, lookup->dl_entry,
1355 				   leaf_bh, leaf_bh->b_data, leaf_bh->b_size);
1356 	if (ret) {
1357 		mlog_errno(ret);
1358 		goto out;
1359 	}
1360 
1361 	max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, leaf_bh);
1362 	trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1363 	if (add_to_free_list) {
1364 		trailer->db_free_next = dx_root->dr_free_blk;
1365 		dx_root->dr_free_blk = cpu_to_le64(leaf_bh->b_blocknr);
1366 		ocfs2_journal_dirty(handle, dx_root_bh);
1367 	}
1368 
1369 	/* leaf_bh was journal_accessed for us in __ocfs2_delete_entry */
1370 	ocfs2_journal_dirty(handle, leaf_bh);
1371 
1372 	le32_add_cpu(&dx_root->dr_num_entries, -1);
1373 	ocfs2_journal_dirty(handle, dx_root_bh);
1374 
1375 	ocfs2_dx_list_remove_entry(entry_list, index);
1376 
1377 	if (!ocfs2_dx_root_inline(dx_root))
1378 		ocfs2_journal_dirty(handle, lookup->dl_dx_leaf_bh);
1379 
1380 out:
1381 	return ret;
1382 }
1383 
1384 static inline int ocfs2_delete_entry_id(handle_t *handle,
1385 					struct inode *dir,
1386 					struct ocfs2_dir_entry *de_del,
1387 					struct buffer_head *bh)
1388 {
1389 	int ret;
1390 	struct buffer_head *di_bh = NULL;
1391 	struct ocfs2_dinode *di;
1392 	struct ocfs2_inline_data *data;
1393 
1394 	ret = ocfs2_read_inode_block(dir, &di_bh);
1395 	if (ret) {
1396 		mlog_errno(ret);
1397 		goto out;
1398 	}
1399 
1400 	di = (struct ocfs2_dinode *)di_bh->b_data;
1401 	data = &di->id2.i_data;
1402 
1403 	ret = __ocfs2_delete_entry(handle, dir, de_del, bh, data->id_data,
1404 				   i_size_read(dir));
1405 
1406 	brelse(di_bh);
1407 out:
1408 	return ret;
1409 }
1410 
1411 static inline int ocfs2_delete_entry_el(handle_t *handle,
1412 					struct inode *dir,
1413 					struct ocfs2_dir_entry *de_del,
1414 					struct buffer_head *bh)
1415 {
1416 	return __ocfs2_delete_entry(handle, dir, de_del, bh, bh->b_data,
1417 				    bh->b_size);
1418 }
1419 
1420 /*
1421  * Delete a directory entry. Hide the details of directory
1422  * implementation from the caller.
1423  */
1424 int ocfs2_delete_entry(handle_t *handle,
1425 		       struct inode *dir,
1426 		       struct ocfs2_dir_lookup_result *res)
1427 {
1428 	if (ocfs2_dir_indexed(dir))
1429 		return ocfs2_delete_entry_dx(handle, dir, res);
1430 
1431 	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1432 		return ocfs2_delete_entry_id(handle, dir, res->dl_entry,
1433 					     res->dl_leaf_bh);
1434 
1435 	return ocfs2_delete_entry_el(handle, dir, res->dl_entry,
1436 				     res->dl_leaf_bh);
1437 }
1438 
1439 /*
1440  * Check whether 'de' has enough room to hold an entry of
1441  * 'new_rec_len' bytes.
1442  */
1443 static inline int ocfs2_dirent_would_fit(struct ocfs2_dir_entry *de,
1444 					 unsigned int new_rec_len)
1445 {
1446 	unsigned int de_really_used;
1447 
1448 	/* Check whether this is an empty record with enough space */
1449 	if (le64_to_cpu(de->inode) == 0 &&
1450 	    le16_to_cpu(de->rec_len) >= new_rec_len)
1451 		return 1;
1452 
1453 	/*
1454 	 * Record might have free space at the end which we can
1455 	 * use.
1456 	 */
1457 	de_really_used = OCFS2_DIR_REC_LEN(de->name_len);
1458 	if (le16_to_cpu(de->rec_len) >= (de_really_used + new_rec_len))
1459 	    return 1;
1460 
1461 	return 0;
1462 }
1463 
1464 static void ocfs2_dx_dir_leaf_insert_tail(struct ocfs2_dx_leaf *dx_leaf,
1465 					  struct ocfs2_dx_entry *dx_new_entry)
1466 {
1467 	int i;
1468 
1469 	i = le16_to_cpu(dx_leaf->dl_list.de_num_used);
1470 	dx_leaf->dl_list.de_entries[i] = *dx_new_entry;
1471 
1472 	le16_add_cpu(&dx_leaf->dl_list.de_num_used, 1);
1473 }
1474 
1475 static void ocfs2_dx_entry_list_insert(struct ocfs2_dx_entry_list *entry_list,
1476 				       struct ocfs2_dx_hinfo *hinfo,
1477 				       u64 dirent_blk)
1478 {
1479 	int i;
1480 	struct ocfs2_dx_entry *dx_entry;
1481 
1482 	i = le16_to_cpu(entry_list->de_num_used);
1483 	dx_entry = &entry_list->de_entries[i];
1484 
1485 	memset(dx_entry, 0, sizeof(*dx_entry));
1486 	dx_entry->dx_major_hash = cpu_to_le32(hinfo->major_hash);
1487 	dx_entry->dx_minor_hash = cpu_to_le32(hinfo->minor_hash);
1488 	dx_entry->dx_dirent_blk = cpu_to_le64(dirent_blk);
1489 
1490 	le16_add_cpu(&entry_list->de_num_used, 1);
1491 }
1492 
1493 static int __ocfs2_dx_dir_leaf_insert(struct inode *dir, handle_t *handle,
1494 				      struct ocfs2_dx_hinfo *hinfo,
1495 				      u64 dirent_blk,
1496 				      struct buffer_head *dx_leaf_bh)
1497 {
1498 	int ret;
1499 	struct ocfs2_dx_leaf *dx_leaf;
1500 
1501 	ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
1502 				      OCFS2_JOURNAL_ACCESS_WRITE);
1503 	if (ret) {
1504 		mlog_errno(ret);
1505 		goto out;
1506 	}
1507 
1508 	dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
1509 	ocfs2_dx_entry_list_insert(&dx_leaf->dl_list, hinfo, dirent_blk);
1510 	ocfs2_journal_dirty(handle, dx_leaf_bh);
1511 
1512 out:
1513 	return ret;
1514 }
1515 
1516 static void ocfs2_dx_inline_root_insert(struct inode *dir, handle_t *handle,
1517 					struct ocfs2_dx_hinfo *hinfo,
1518 					u64 dirent_blk,
1519 					struct ocfs2_dx_root_block *dx_root)
1520 {
1521 	ocfs2_dx_entry_list_insert(&dx_root->dr_entries, hinfo, dirent_blk);
1522 }
1523 
1524 static int ocfs2_dx_dir_insert(struct inode *dir, handle_t *handle,
1525 			       struct ocfs2_dir_lookup_result *lookup)
1526 {
1527 	int ret = 0;
1528 	struct ocfs2_dx_root_block *dx_root;
1529 	struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1530 
1531 	ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1532 				      OCFS2_JOURNAL_ACCESS_WRITE);
1533 	if (ret) {
1534 		mlog_errno(ret);
1535 		goto out;
1536 	}
1537 
1538 	dx_root = (struct ocfs2_dx_root_block *)lookup->dl_dx_root_bh->b_data;
1539 	if (ocfs2_dx_root_inline(dx_root)) {
1540 		ocfs2_dx_inline_root_insert(dir, handle,
1541 					    &lookup->dl_hinfo,
1542 					    lookup->dl_leaf_bh->b_blocknr,
1543 					    dx_root);
1544 	} else {
1545 		ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &lookup->dl_hinfo,
1546 						 lookup->dl_leaf_bh->b_blocknr,
1547 						 lookup->dl_dx_leaf_bh);
1548 		if (ret)
1549 			goto out;
1550 	}
1551 
1552 	le32_add_cpu(&dx_root->dr_num_entries, 1);
1553 	ocfs2_journal_dirty(handle, dx_root_bh);
1554 
1555 out:
1556 	return ret;
1557 }
1558 
1559 static void ocfs2_remove_block_from_free_list(struct inode *dir,
1560 				       handle_t *handle,
1561 				       struct ocfs2_dir_lookup_result *lookup)
1562 {
1563 	struct ocfs2_dir_block_trailer *trailer, *prev;
1564 	struct ocfs2_dx_root_block *dx_root;
1565 	struct buffer_head *bh;
1566 
1567 	trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1568 
1569 	if (ocfs2_free_list_at_root(lookup)) {
1570 		bh = lookup->dl_dx_root_bh;
1571 		dx_root = (struct ocfs2_dx_root_block *)bh->b_data;
1572 		dx_root->dr_free_blk = trailer->db_free_next;
1573 	} else {
1574 		bh = lookup->dl_prev_leaf_bh;
1575 		prev = ocfs2_trailer_from_bh(bh, dir->i_sb);
1576 		prev->db_free_next = trailer->db_free_next;
1577 	}
1578 
1579 	trailer->db_free_rec_len = cpu_to_le16(0);
1580 	trailer->db_free_next = cpu_to_le64(0);
1581 
1582 	ocfs2_journal_dirty(handle, bh);
1583 	ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1584 }
1585 
1586 /*
1587  * This expects that a journal write has been reserved on
1588  * lookup->dl_prev_leaf_bh or lookup->dl_dx_root_bh
1589  */
1590 static void ocfs2_recalc_free_list(struct inode *dir, handle_t *handle,
1591 				   struct ocfs2_dir_lookup_result *lookup)
1592 {
1593 	int max_rec_len;
1594 	struct ocfs2_dir_block_trailer *trailer;
1595 
1596 	/* Walk dl_leaf_bh to figure out what the new free rec_len is. */
1597 	max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, lookup->dl_leaf_bh);
1598 	if (max_rec_len) {
1599 		/*
1600 		 * There's still room in this block, so no need to remove it
1601 		 * from the free list. In this case, we just want to update
1602 		 * the rec len accounting.
1603 		 */
1604 		trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1605 		trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1606 		ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1607 	} else {
1608 		ocfs2_remove_block_from_free_list(dir, handle, lookup);
1609 	}
1610 }
1611 
1612 /* we don't always have a dentry for what we want to add, so people
1613  * like orphan dir can call this instead.
1614  *
1615  * The lookup context must have been filled from
1616  * ocfs2_prepare_dir_for_insert.
1617  */
1618 int __ocfs2_add_entry(handle_t *handle,
1619 		      struct inode *dir,
1620 		      const char *name, int namelen,
1621 		      struct inode *inode, u64 blkno,
1622 		      struct buffer_head *parent_fe_bh,
1623 		      struct ocfs2_dir_lookup_result *lookup)
1624 {
1625 	unsigned long offset;
1626 	unsigned short rec_len;
1627 	struct ocfs2_dir_entry *de, *de1;
1628 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)parent_fe_bh->b_data;
1629 	struct super_block *sb = dir->i_sb;
1630 	int retval, status;
1631 	unsigned int size = sb->s_blocksize;
1632 	struct buffer_head *insert_bh = lookup->dl_leaf_bh;
1633 	char *data_start = insert_bh->b_data;
1634 
1635 	mlog_entry_void();
1636 
1637 	if (!namelen)
1638 		return -EINVAL;
1639 
1640 	if (ocfs2_dir_indexed(dir)) {
1641 		struct buffer_head *bh;
1642 
1643 		/*
1644 		 * An indexed dir may require that we update the free space
1645 		 * list. Reserve a write to the previous node in the list so
1646 		 * that we don't fail later.
1647 		 *
1648 		 * XXX: This can be either a dx_root_block, or an unindexed
1649 		 * directory tree leaf block.
1650 		 */
1651 		if (ocfs2_free_list_at_root(lookup)) {
1652 			bh = lookup->dl_dx_root_bh;
1653 			retval = ocfs2_journal_access_dr(handle,
1654 						 INODE_CACHE(dir), bh,
1655 						 OCFS2_JOURNAL_ACCESS_WRITE);
1656 		} else {
1657 			bh = lookup->dl_prev_leaf_bh;
1658 			retval = ocfs2_journal_access_db(handle,
1659 						 INODE_CACHE(dir), bh,
1660 						 OCFS2_JOURNAL_ACCESS_WRITE);
1661 		}
1662 		if (retval) {
1663 			mlog_errno(retval);
1664 			return retval;
1665 		}
1666 	} else if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1667 		data_start = di->id2.i_data.id_data;
1668 		size = i_size_read(dir);
1669 
1670 		BUG_ON(insert_bh != parent_fe_bh);
1671 	}
1672 
1673 	rec_len = OCFS2_DIR_REC_LEN(namelen);
1674 	offset = 0;
1675 	de = (struct ocfs2_dir_entry *) data_start;
1676 	while (1) {
1677 		BUG_ON((char *)de >= (size + data_start));
1678 
1679 		/* These checks should've already been passed by the
1680 		 * prepare function, but I guess we can leave them
1681 		 * here anyway. */
1682 		if (!ocfs2_check_dir_entry(dir, de, insert_bh, offset)) {
1683 			retval = -ENOENT;
1684 			goto bail;
1685 		}
1686 		if (ocfs2_match(namelen, name, de)) {
1687 			retval = -EEXIST;
1688 			goto bail;
1689 		}
1690 
1691 		/* We're guaranteed that we should have space, so we
1692 		 * can't possibly have hit the trailer...right? */
1693 		mlog_bug_on_msg(ocfs2_skip_dir_trailer(dir, de, offset, size),
1694 				"Hit dir trailer trying to insert %.*s "
1695 			        "(namelen %d) into directory %llu.  "
1696 				"offset is %lu, trailer offset is %d\n",
1697 				namelen, name, namelen,
1698 				(unsigned long long)parent_fe_bh->b_blocknr,
1699 				offset, ocfs2_dir_trailer_blk_off(dir->i_sb));
1700 
1701 		if (ocfs2_dirent_would_fit(de, rec_len)) {
1702 			dir->i_mtime = dir->i_ctime = CURRENT_TIME;
1703 			retval = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
1704 			if (retval < 0) {
1705 				mlog_errno(retval);
1706 				goto bail;
1707 			}
1708 
1709 			if (insert_bh == parent_fe_bh)
1710 				status = ocfs2_journal_access_di(handle,
1711 								 INODE_CACHE(dir),
1712 								 insert_bh,
1713 								 OCFS2_JOURNAL_ACCESS_WRITE);
1714 			else {
1715 				status = ocfs2_journal_access_db(handle,
1716 								 INODE_CACHE(dir),
1717 								 insert_bh,
1718 					      OCFS2_JOURNAL_ACCESS_WRITE);
1719 
1720 				if (ocfs2_dir_indexed(dir)) {
1721 					status = ocfs2_dx_dir_insert(dir,
1722 								handle,
1723 								lookup);
1724 					if (status) {
1725 						mlog_errno(status);
1726 						goto bail;
1727 					}
1728 				}
1729 			}
1730 
1731 			/* By now the buffer is marked for journaling */
1732 			offset += le16_to_cpu(de->rec_len);
1733 			if (le64_to_cpu(de->inode)) {
1734 				de1 = (struct ocfs2_dir_entry *)((char *) de +
1735 					OCFS2_DIR_REC_LEN(de->name_len));
1736 				de1->rec_len =
1737 					cpu_to_le16(le16_to_cpu(de->rec_len) -
1738 					OCFS2_DIR_REC_LEN(de->name_len));
1739 				de->rec_len = cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
1740 				de = de1;
1741 			}
1742 			de->file_type = OCFS2_FT_UNKNOWN;
1743 			if (blkno) {
1744 				de->inode = cpu_to_le64(blkno);
1745 				ocfs2_set_de_type(de, inode->i_mode);
1746 			} else
1747 				de->inode = 0;
1748 			de->name_len = namelen;
1749 			memcpy(de->name, name, namelen);
1750 
1751 			if (ocfs2_dir_indexed(dir))
1752 				ocfs2_recalc_free_list(dir, handle, lookup);
1753 
1754 			dir->i_version++;
1755 			status = ocfs2_journal_dirty(handle, insert_bh);
1756 			retval = 0;
1757 			goto bail;
1758 		}
1759 
1760 		offset += le16_to_cpu(de->rec_len);
1761 		de = (struct ocfs2_dir_entry *) ((char *) de + le16_to_cpu(de->rec_len));
1762 	}
1763 
1764 	/* when you think about it, the assert above should prevent us
1765 	 * from ever getting here. */
1766 	retval = -ENOSPC;
1767 bail:
1768 
1769 	mlog_exit(retval);
1770 	return retval;
1771 }
1772 
1773 static int ocfs2_dir_foreach_blk_id(struct inode *inode,
1774 				    u64 *f_version,
1775 				    loff_t *f_pos, void *priv,
1776 				    filldir_t filldir, int *filldir_err)
1777 {
1778 	int ret, i, filldir_ret;
1779 	unsigned long offset = *f_pos;
1780 	struct buffer_head *di_bh = NULL;
1781 	struct ocfs2_dinode *di;
1782 	struct ocfs2_inline_data *data;
1783 	struct ocfs2_dir_entry *de;
1784 
1785 	ret = ocfs2_read_inode_block(inode, &di_bh);
1786 	if (ret) {
1787 		mlog(ML_ERROR, "Unable to read inode block for dir %llu\n",
1788 		     (unsigned long long)OCFS2_I(inode)->ip_blkno);
1789 		goto out;
1790 	}
1791 
1792 	di = (struct ocfs2_dinode *)di_bh->b_data;
1793 	data = &di->id2.i_data;
1794 
1795 	while (*f_pos < i_size_read(inode)) {
1796 revalidate:
1797 		/* If the dir block has changed since the last call to
1798 		 * readdir(2), then we might be pointing to an invalid
1799 		 * dirent right now.  Scan from the start of the block
1800 		 * to make sure. */
1801 		if (*f_version != inode->i_version) {
1802 			for (i = 0; i < i_size_read(inode) && i < offset; ) {
1803 				de = (struct ocfs2_dir_entry *)
1804 					(data->id_data + i);
1805 				/* It's too expensive to do a full
1806 				 * dirent test each time round this
1807 				 * loop, but we do have to test at
1808 				 * least that it is non-zero.  A
1809 				 * failure will be detected in the
1810 				 * dirent test below. */
1811 				if (le16_to_cpu(de->rec_len) <
1812 				    OCFS2_DIR_REC_LEN(1))
1813 					break;
1814 				i += le16_to_cpu(de->rec_len);
1815 			}
1816 			*f_pos = offset = i;
1817 			*f_version = inode->i_version;
1818 		}
1819 
1820 		de = (struct ocfs2_dir_entry *) (data->id_data + *f_pos);
1821 		if (!ocfs2_check_dir_entry(inode, de, di_bh, *f_pos)) {
1822 			/* On error, skip the f_pos to the end. */
1823 			*f_pos = i_size_read(inode);
1824 			goto out;
1825 		}
1826 		offset += le16_to_cpu(de->rec_len);
1827 		if (le64_to_cpu(de->inode)) {
1828 			/* We might block in the next section
1829 			 * if the data destination is
1830 			 * currently swapped out.  So, use a
1831 			 * version stamp to detect whether or
1832 			 * not the directory has been modified
1833 			 * during the copy operation.
1834 			 */
1835 			u64 version = *f_version;
1836 			unsigned char d_type = DT_UNKNOWN;
1837 
1838 			if (de->file_type < OCFS2_FT_MAX)
1839 				d_type = ocfs2_filetype_table[de->file_type];
1840 
1841 			filldir_ret = filldir(priv, de->name,
1842 					      de->name_len,
1843 					      *f_pos,
1844 					      le64_to_cpu(de->inode),
1845 					      d_type);
1846 			if (filldir_ret) {
1847 				if (filldir_err)
1848 					*filldir_err = filldir_ret;
1849 				break;
1850 			}
1851 			if (version != *f_version)
1852 				goto revalidate;
1853 		}
1854 		*f_pos += le16_to_cpu(de->rec_len);
1855 	}
1856 
1857 out:
1858 	brelse(di_bh);
1859 
1860 	return 0;
1861 }
1862 
1863 /*
1864  * NOTE: This function can be called against unindexed directories,
1865  * and indexed ones.
1866  */
1867 static int ocfs2_dir_foreach_blk_el(struct inode *inode,
1868 				    u64 *f_version,
1869 				    loff_t *f_pos, void *priv,
1870 				    filldir_t filldir, int *filldir_err)
1871 {
1872 	int error = 0;
1873 	unsigned long offset, blk, last_ra_blk = 0;
1874 	int i, stored;
1875 	struct buffer_head * bh, * tmp;
1876 	struct ocfs2_dir_entry * de;
1877 	struct super_block * sb = inode->i_sb;
1878 	unsigned int ra_sectors = 16;
1879 
1880 	stored = 0;
1881 	bh = NULL;
1882 
1883 	offset = (*f_pos) & (sb->s_blocksize - 1);
1884 
1885 	while (!error && !stored && *f_pos < i_size_read(inode)) {
1886 		blk = (*f_pos) >> sb->s_blocksize_bits;
1887 		if (ocfs2_read_dir_block(inode, blk, &bh, 0)) {
1888 			/* Skip the corrupt dirblock and keep trying */
1889 			*f_pos += sb->s_blocksize - offset;
1890 			continue;
1891 		}
1892 
1893 		/* The idea here is to begin with 8k read-ahead and to stay
1894 		 * 4k ahead of our current position.
1895 		 *
1896 		 * TODO: Use the pagecache for this. We just need to
1897 		 * make sure it's cluster-safe... */
1898 		if (!last_ra_blk
1899 		    || (((last_ra_blk - blk) << 9) <= (ra_sectors / 2))) {
1900 			for (i = ra_sectors >> (sb->s_blocksize_bits - 9);
1901 			     i > 0; i--) {
1902 				tmp = NULL;
1903 				if (!ocfs2_read_dir_block(inode, ++blk, &tmp,
1904 							  OCFS2_BH_READAHEAD))
1905 					brelse(tmp);
1906 			}
1907 			last_ra_blk = blk;
1908 			ra_sectors = 8;
1909 		}
1910 
1911 revalidate:
1912 		/* If the dir block has changed since the last call to
1913 		 * readdir(2), then we might be pointing to an invalid
1914 		 * dirent right now.  Scan from the start of the block
1915 		 * to make sure. */
1916 		if (*f_version != inode->i_version) {
1917 			for (i = 0; i < sb->s_blocksize && i < offset; ) {
1918 				de = (struct ocfs2_dir_entry *) (bh->b_data + i);
1919 				/* It's too expensive to do a full
1920 				 * dirent test each time round this
1921 				 * loop, but we do have to test at
1922 				 * least that it is non-zero.  A
1923 				 * failure will be detected in the
1924 				 * dirent test below. */
1925 				if (le16_to_cpu(de->rec_len) <
1926 				    OCFS2_DIR_REC_LEN(1))
1927 					break;
1928 				i += le16_to_cpu(de->rec_len);
1929 			}
1930 			offset = i;
1931 			*f_pos = ((*f_pos) & ~(sb->s_blocksize - 1))
1932 				| offset;
1933 			*f_version = inode->i_version;
1934 		}
1935 
1936 		while (!error && *f_pos < i_size_read(inode)
1937 		       && offset < sb->s_blocksize) {
1938 			de = (struct ocfs2_dir_entry *) (bh->b_data + offset);
1939 			if (!ocfs2_check_dir_entry(inode, de, bh, offset)) {
1940 				/* On error, skip the f_pos to the
1941 				   next block. */
1942 				*f_pos = ((*f_pos) | (sb->s_blocksize - 1)) + 1;
1943 				brelse(bh);
1944 				goto out;
1945 			}
1946 			offset += le16_to_cpu(de->rec_len);
1947 			if (le64_to_cpu(de->inode)) {
1948 				/* We might block in the next section
1949 				 * if the data destination is
1950 				 * currently swapped out.  So, use a
1951 				 * version stamp to detect whether or
1952 				 * not the directory has been modified
1953 				 * during the copy operation.
1954 				 */
1955 				unsigned long version = *f_version;
1956 				unsigned char d_type = DT_UNKNOWN;
1957 
1958 				if (de->file_type < OCFS2_FT_MAX)
1959 					d_type = ocfs2_filetype_table[de->file_type];
1960 				error = filldir(priv, de->name,
1961 						de->name_len,
1962 						*f_pos,
1963 						le64_to_cpu(de->inode),
1964 						d_type);
1965 				if (error) {
1966 					if (filldir_err)
1967 						*filldir_err = error;
1968 					break;
1969 				}
1970 				if (version != *f_version)
1971 					goto revalidate;
1972 				stored ++;
1973 			}
1974 			*f_pos += le16_to_cpu(de->rec_len);
1975 		}
1976 		offset = 0;
1977 		brelse(bh);
1978 		bh = NULL;
1979 	}
1980 
1981 	stored = 0;
1982 out:
1983 	return stored;
1984 }
1985 
1986 static int ocfs2_dir_foreach_blk(struct inode *inode, u64 *f_version,
1987 				 loff_t *f_pos, void *priv, filldir_t filldir,
1988 				 int *filldir_err)
1989 {
1990 	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1991 		return ocfs2_dir_foreach_blk_id(inode, f_version, f_pos, priv,
1992 						filldir, filldir_err);
1993 
1994 	return ocfs2_dir_foreach_blk_el(inode, f_version, f_pos, priv, filldir,
1995 					filldir_err);
1996 }
1997 
1998 /*
1999  * This is intended to be called from inside other kernel functions,
2000  * so we fake some arguments.
2001  */
2002 int ocfs2_dir_foreach(struct inode *inode, loff_t *f_pos, void *priv,
2003 		      filldir_t filldir)
2004 {
2005 	int ret = 0, filldir_err = 0;
2006 	u64 version = inode->i_version;
2007 
2008 	while (*f_pos < i_size_read(inode)) {
2009 		ret = ocfs2_dir_foreach_blk(inode, &version, f_pos, priv,
2010 					    filldir, &filldir_err);
2011 		if (ret || filldir_err)
2012 			break;
2013 	}
2014 
2015 	if (ret > 0)
2016 		ret = -EIO;
2017 
2018 	return 0;
2019 }
2020 
2021 /*
2022  * ocfs2_readdir()
2023  *
2024  */
2025 int ocfs2_readdir(struct file * filp, void * dirent, filldir_t filldir)
2026 {
2027 	int error = 0;
2028 	struct inode *inode = filp->f_path.dentry->d_inode;
2029 	int lock_level = 0;
2030 
2031 	mlog_entry("dirino=%llu\n",
2032 		   (unsigned long long)OCFS2_I(inode)->ip_blkno);
2033 
2034 	error = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2035 	if (lock_level && error >= 0) {
2036 		/* We release EX lock which used to update atime
2037 		 * and get PR lock again to reduce contention
2038 		 * on commonly accessed directories. */
2039 		ocfs2_inode_unlock(inode, 1);
2040 		lock_level = 0;
2041 		error = ocfs2_inode_lock(inode, NULL, 0);
2042 	}
2043 	if (error < 0) {
2044 		if (error != -ENOENT)
2045 			mlog_errno(error);
2046 		/* we haven't got any yet, so propagate the error. */
2047 		goto bail_nolock;
2048 	}
2049 
2050 	error = ocfs2_dir_foreach_blk(inode, &filp->f_version, &filp->f_pos,
2051 				      dirent, filldir, NULL);
2052 
2053 	ocfs2_inode_unlock(inode, lock_level);
2054 
2055 bail_nolock:
2056 	mlog_exit(error);
2057 
2058 	return error;
2059 }
2060 
2061 /*
2062  * NOTE: this should always be called with parent dir i_mutex taken.
2063  */
2064 int ocfs2_find_files_on_disk(const char *name,
2065 			     int namelen,
2066 			     u64 *blkno,
2067 			     struct inode *inode,
2068 			     struct ocfs2_dir_lookup_result *lookup)
2069 {
2070 	int status = -ENOENT;
2071 
2072 	mlog(0, "name=%.*s, blkno=%p, inode=%llu\n", namelen, name, blkno,
2073 	     (unsigned long long)OCFS2_I(inode)->ip_blkno);
2074 
2075 	status = ocfs2_find_entry(name, namelen, inode, lookup);
2076 	if (status)
2077 		goto leave;
2078 
2079 	*blkno = le64_to_cpu(lookup->dl_entry->inode);
2080 
2081 	status = 0;
2082 leave:
2083 
2084 	return status;
2085 }
2086 
2087 /*
2088  * Convenience function for callers which just want the block number
2089  * mapped to a name and don't require the full dirent info, etc.
2090  */
2091 int ocfs2_lookup_ino_from_name(struct inode *dir, const char *name,
2092 			       int namelen, u64 *blkno)
2093 {
2094 	int ret;
2095 	struct ocfs2_dir_lookup_result lookup = { NULL, };
2096 
2097 	ret = ocfs2_find_files_on_disk(name, namelen, blkno, dir, &lookup);
2098 	ocfs2_free_dir_lookup_result(&lookup);
2099 
2100 	return ret;
2101 }
2102 
2103 /* Check for a name within a directory.
2104  *
2105  * Return 0 if the name does not exist
2106  * Return -EEXIST if the directory contains the name
2107  *
2108  * Callers should have i_mutex + a cluster lock on dir
2109  */
2110 int ocfs2_check_dir_for_entry(struct inode *dir,
2111 			      const char *name,
2112 			      int namelen)
2113 {
2114 	int ret;
2115 	struct ocfs2_dir_lookup_result lookup = { NULL, };
2116 
2117 	mlog_entry("dir %llu, name '%.*s'\n",
2118 		   (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen, name);
2119 
2120 	ret = -EEXIST;
2121 	if (ocfs2_find_entry(name, namelen, dir, &lookup) == 0)
2122 		goto bail;
2123 
2124 	ret = 0;
2125 bail:
2126 	ocfs2_free_dir_lookup_result(&lookup);
2127 
2128 	mlog_exit(ret);
2129 	return ret;
2130 }
2131 
2132 struct ocfs2_empty_dir_priv {
2133 	unsigned seen_dot;
2134 	unsigned seen_dot_dot;
2135 	unsigned seen_other;
2136 	unsigned dx_dir;
2137 };
2138 static int ocfs2_empty_dir_filldir(void *priv, const char *name, int name_len,
2139 				   loff_t pos, u64 ino, unsigned type)
2140 {
2141 	struct ocfs2_empty_dir_priv *p = priv;
2142 
2143 	/*
2144 	 * Check the positions of "." and ".." records to be sure
2145 	 * they're in the correct place.
2146 	 *
2147 	 * Indexed directories don't need to proceed past the first
2148 	 * two entries, so we end the scan after seeing '..'. Despite
2149 	 * that, we allow the scan to proceed In the event that we
2150 	 * have a corrupted indexed directory (no dot or dot dot
2151 	 * entries). This allows us to double check for existing
2152 	 * entries which might not have been found in the index.
2153 	 */
2154 	if (name_len == 1 && !strncmp(".", name, 1) && pos == 0) {
2155 		p->seen_dot = 1;
2156 		return 0;
2157 	}
2158 
2159 	if (name_len == 2 && !strncmp("..", name, 2) &&
2160 	    pos == OCFS2_DIR_REC_LEN(1)) {
2161 		p->seen_dot_dot = 1;
2162 
2163 		if (p->dx_dir && p->seen_dot)
2164 			return 1;
2165 
2166 		return 0;
2167 	}
2168 
2169 	p->seen_other = 1;
2170 	return 1;
2171 }
2172 
2173 static int ocfs2_empty_dir_dx(struct inode *inode,
2174 			      struct ocfs2_empty_dir_priv *priv)
2175 {
2176 	int ret;
2177 	struct buffer_head *di_bh = NULL;
2178 	struct buffer_head *dx_root_bh = NULL;
2179 	struct ocfs2_dinode *di;
2180 	struct ocfs2_dx_root_block *dx_root;
2181 
2182 	priv->dx_dir = 1;
2183 
2184 	ret = ocfs2_read_inode_block(inode, &di_bh);
2185 	if (ret) {
2186 		mlog_errno(ret);
2187 		goto out;
2188 	}
2189 	di = (struct ocfs2_dinode *)di_bh->b_data;
2190 
2191 	ret = ocfs2_read_dx_root(inode, di, &dx_root_bh);
2192 	if (ret) {
2193 		mlog_errno(ret);
2194 		goto out;
2195 	}
2196 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2197 
2198 	if (le32_to_cpu(dx_root->dr_num_entries) != 2)
2199 		priv->seen_other = 1;
2200 
2201 out:
2202 	brelse(di_bh);
2203 	brelse(dx_root_bh);
2204 	return ret;
2205 }
2206 
2207 /*
2208  * routine to check that the specified directory is empty (for rmdir)
2209  *
2210  * Returns 1 if dir is empty, zero otherwise.
2211  *
2212  * XXX: This is a performance problem for unindexed directories.
2213  */
2214 int ocfs2_empty_dir(struct inode *inode)
2215 {
2216 	int ret;
2217 	loff_t start = 0;
2218 	struct ocfs2_empty_dir_priv priv;
2219 
2220 	memset(&priv, 0, sizeof(priv));
2221 
2222 	if (ocfs2_dir_indexed(inode)) {
2223 		ret = ocfs2_empty_dir_dx(inode, &priv);
2224 		if (ret)
2225 			mlog_errno(ret);
2226 		/*
2227 		 * We still run ocfs2_dir_foreach to get the checks
2228 		 * for "." and "..".
2229 		 */
2230 	}
2231 
2232 	ret = ocfs2_dir_foreach(inode, &start, &priv, ocfs2_empty_dir_filldir);
2233 	if (ret)
2234 		mlog_errno(ret);
2235 
2236 	if (!priv.seen_dot || !priv.seen_dot_dot) {
2237 		mlog(ML_ERROR, "bad directory (dir #%llu) - no `.' or `..'\n",
2238 		     (unsigned long long)OCFS2_I(inode)->ip_blkno);
2239 		/*
2240 		 * XXX: Is it really safe to allow an unlink to continue?
2241 		 */
2242 		return 1;
2243 	}
2244 
2245 	return !priv.seen_other;
2246 }
2247 
2248 /*
2249  * Fills "." and ".." dirents in a new directory block. Returns dirent for
2250  * "..", which might be used during creation of a directory with a trailing
2251  * header. It is otherwise safe to ignore the return code.
2252  */
2253 static struct ocfs2_dir_entry *ocfs2_fill_initial_dirents(struct inode *inode,
2254 							  struct inode *parent,
2255 							  char *start,
2256 							  unsigned int size)
2257 {
2258 	struct ocfs2_dir_entry *de = (struct ocfs2_dir_entry *)start;
2259 
2260 	de->inode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
2261 	de->name_len = 1;
2262 	de->rec_len =
2263 		cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
2264 	strcpy(de->name, ".");
2265 	ocfs2_set_de_type(de, S_IFDIR);
2266 
2267 	de = (struct ocfs2_dir_entry *) ((char *)de + le16_to_cpu(de->rec_len));
2268 	de->inode = cpu_to_le64(OCFS2_I(parent)->ip_blkno);
2269 	de->rec_len = cpu_to_le16(size - OCFS2_DIR_REC_LEN(1));
2270 	de->name_len = 2;
2271 	strcpy(de->name, "..");
2272 	ocfs2_set_de_type(de, S_IFDIR);
2273 
2274 	return de;
2275 }
2276 
2277 /*
2278  * This works together with code in ocfs2_mknod_locked() which sets
2279  * the inline-data flag and initializes the inline-data section.
2280  */
2281 static int ocfs2_fill_new_dir_id(struct ocfs2_super *osb,
2282 				 handle_t *handle,
2283 				 struct inode *parent,
2284 				 struct inode *inode,
2285 				 struct buffer_head *di_bh)
2286 {
2287 	int ret;
2288 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2289 	struct ocfs2_inline_data *data = &di->id2.i_data;
2290 	unsigned int size = le16_to_cpu(data->id_count);
2291 
2292 	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2293 				      OCFS2_JOURNAL_ACCESS_WRITE);
2294 	if (ret) {
2295 		mlog_errno(ret);
2296 		goto out;
2297 	}
2298 
2299 	ocfs2_fill_initial_dirents(inode, parent, data->id_data, size);
2300 
2301 	ocfs2_journal_dirty(handle, di_bh);
2302 	if (ret) {
2303 		mlog_errno(ret);
2304 		goto out;
2305 	}
2306 
2307 	i_size_write(inode, size);
2308 	inode->i_nlink = 2;
2309 	inode->i_blocks = ocfs2_inode_sector_count(inode);
2310 
2311 	ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
2312 	if (ret < 0)
2313 		mlog_errno(ret);
2314 
2315 out:
2316 	return ret;
2317 }
2318 
2319 static int ocfs2_fill_new_dir_el(struct ocfs2_super *osb,
2320 				 handle_t *handle,
2321 				 struct inode *parent,
2322 				 struct inode *inode,
2323 				 struct buffer_head *fe_bh,
2324 				 struct ocfs2_alloc_context *data_ac,
2325 				 struct buffer_head **ret_new_bh)
2326 {
2327 	int status;
2328 	unsigned int size = osb->sb->s_blocksize;
2329 	struct buffer_head *new_bh = NULL;
2330 	struct ocfs2_dir_entry *de;
2331 
2332 	mlog_entry_void();
2333 
2334 	if (ocfs2_new_dir_wants_trailer(inode))
2335 		size = ocfs2_dir_trailer_blk_off(parent->i_sb);
2336 
2337 	status = ocfs2_do_extend_dir(osb->sb, handle, inode, fe_bh,
2338 				     data_ac, NULL, &new_bh);
2339 	if (status < 0) {
2340 		mlog_errno(status);
2341 		goto bail;
2342 	}
2343 
2344 	ocfs2_set_new_buffer_uptodate(INODE_CACHE(inode), new_bh);
2345 
2346 	status = ocfs2_journal_access_db(handle, INODE_CACHE(inode), new_bh,
2347 					 OCFS2_JOURNAL_ACCESS_CREATE);
2348 	if (status < 0) {
2349 		mlog_errno(status);
2350 		goto bail;
2351 	}
2352 	memset(new_bh->b_data, 0, osb->sb->s_blocksize);
2353 
2354 	de = ocfs2_fill_initial_dirents(inode, parent, new_bh->b_data, size);
2355 	if (ocfs2_new_dir_wants_trailer(inode)) {
2356 		int size = le16_to_cpu(de->rec_len);
2357 
2358 		/*
2359 		 * Figure out the size of the hole left over after
2360 		 * insertion of '.' and '..'. The trailer wants this
2361 		 * information.
2362 		 */
2363 		size -= OCFS2_DIR_REC_LEN(2);
2364 		size -= sizeof(struct ocfs2_dir_block_trailer);
2365 
2366 		ocfs2_init_dir_trailer(inode, new_bh, size);
2367 	}
2368 
2369 	status = ocfs2_journal_dirty(handle, new_bh);
2370 	if (status < 0) {
2371 		mlog_errno(status);
2372 		goto bail;
2373 	}
2374 
2375 	i_size_write(inode, inode->i_sb->s_blocksize);
2376 	inode->i_nlink = 2;
2377 	inode->i_blocks = ocfs2_inode_sector_count(inode);
2378 	status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
2379 	if (status < 0) {
2380 		mlog_errno(status);
2381 		goto bail;
2382 	}
2383 
2384 	status = 0;
2385 	if (ret_new_bh) {
2386 		*ret_new_bh = new_bh;
2387 		new_bh = NULL;
2388 	}
2389 bail:
2390 	brelse(new_bh);
2391 
2392 	mlog_exit(status);
2393 	return status;
2394 }
2395 
2396 static int ocfs2_dx_dir_attach_index(struct ocfs2_super *osb,
2397 				     handle_t *handle, struct inode *dir,
2398 				     struct buffer_head *di_bh,
2399 				     struct buffer_head *dirdata_bh,
2400 				     struct ocfs2_alloc_context *meta_ac,
2401 				     int dx_inline, u32 num_entries,
2402 				     struct buffer_head **ret_dx_root_bh)
2403 {
2404 	int ret;
2405 	struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
2406 	u16 dr_suballoc_bit;
2407 	u64 dr_blkno;
2408 	unsigned int num_bits;
2409 	struct buffer_head *dx_root_bh = NULL;
2410 	struct ocfs2_dx_root_block *dx_root;
2411 	struct ocfs2_dir_block_trailer *trailer =
2412 		ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
2413 
2414 	ret = ocfs2_claim_metadata(osb, handle, meta_ac, 1, &dr_suballoc_bit,
2415 				   &num_bits, &dr_blkno);
2416 	if (ret) {
2417 		mlog_errno(ret);
2418 		goto out;
2419 	}
2420 
2421 	mlog(0, "Dir %llu, attach new index block: %llu\n",
2422 	     (unsigned long long)OCFS2_I(dir)->ip_blkno,
2423 	     (unsigned long long)dr_blkno);
2424 
2425 	dx_root_bh = sb_getblk(osb->sb, dr_blkno);
2426 	if (dx_root_bh == NULL) {
2427 		ret = -EIO;
2428 		goto out;
2429 	}
2430 	ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dx_root_bh);
2431 
2432 	ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
2433 				      OCFS2_JOURNAL_ACCESS_CREATE);
2434 	if (ret < 0) {
2435 		mlog_errno(ret);
2436 		goto out;
2437 	}
2438 
2439 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2440 	memset(dx_root, 0, osb->sb->s_blocksize);
2441 	strcpy(dx_root->dr_signature, OCFS2_DX_ROOT_SIGNATURE);
2442 	dx_root->dr_suballoc_slot = cpu_to_le16(osb->slot_num);
2443 	dx_root->dr_suballoc_bit = cpu_to_le16(dr_suballoc_bit);
2444 	dx_root->dr_fs_generation = cpu_to_le32(osb->fs_generation);
2445 	dx_root->dr_blkno = cpu_to_le64(dr_blkno);
2446 	dx_root->dr_dir_blkno = cpu_to_le64(OCFS2_I(dir)->ip_blkno);
2447 	dx_root->dr_num_entries = cpu_to_le32(num_entries);
2448 	if (le16_to_cpu(trailer->db_free_rec_len))
2449 		dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
2450 	else
2451 		dx_root->dr_free_blk = cpu_to_le64(0);
2452 
2453 	if (dx_inline) {
2454 		dx_root->dr_flags |= OCFS2_DX_FLAG_INLINE;
2455 		dx_root->dr_entries.de_count =
2456 			cpu_to_le16(ocfs2_dx_entries_per_root(osb->sb));
2457 	} else {
2458 		dx_root->dr_list.l_count =
2459 			cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
2460 	}
2461 
2462 	ret = ocfs2_journal_dirty(handle, dx_root_bh);
2463 	if (ret)
2464 		mlog_errno(ret);
2465 
2466 	ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2467 				      OCFS2_JOURNAL_ACCESS_CREATE);
2468 	if (ret) {
2469 		mlog_errno(ret);
2470 		goto out;
2471 	}
2472 
2473 	di->i_dx_root = cpu_to_le64(dr_blkno);
2474 
2475 	OCFS2_I(dir)->ip_dyn_features |= OCFS2_INDEXED_DIR_FL;
2476 	di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
2477 
2478 	ret = ocfs2_journal_dirty(handle, di_bh);
2479 	if (ret)
2480 		mlog_errno(ret);
2481 
2482 	*ret_dx_root_bh = dx_root_bh;
2483 	dx_root_bh = NULL;
2484 
2485 out:
2486 	brelse(dx_root_bh);
2487 	return ret;
2488 }
2489 
2490 static int ocfs2_dx_dir_format_cluster(struct ocfs2_super *osb,
2491 				       handle_t *handle, struct inode *dir,
2492 				       struct buffer_head **dx_leaves,
2493 				       int num_dx_leaves, u64 start_blk)
2494 {
2495 	int ret, i;
2496 	struct ocfs2_dx_leaf *dx_leaf;
2497 	struct buffer_head *bh;
2498 
2499 	for (i = 0; i < num_dx_leaves; i++) {
2500 		bh = sb_getblk(osb->sb, start_blk + i);
2501 		if (bh == NULL) {
2502 			ret = -EIO;
2503 			goto out;
2504 		}
2505 		dx_leaves[i] = bh;
2506 
2507 		ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), bh);
2508 
2509 		ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), bh,
2510 					      OCFS2_JOURNAL_ACCESS_CREATE);
2511 		if (ret < 0) {
2512 			mlog_errno(ret);
2513 			goto out;
2514 		}
2515 
2516 		dx_leaf = (struct ocfs2_dx_leaf *) bh->b_data;
2517 
2518 		memset(dx_leaf, 0, osb->sb->s_blocksize);
2519 		strcpy(dx_leaf->dl_signature, OCFS2_DX_LEAF_SIGNATURE);
2520 		dx_leaf->dl_fs_generation = cpu_to_le32(osb->fs_generation);
2521 		dx_leaf->dl_blkno = cpu_to_le64(bh->b_blocknr);
2522 		dx_leaf->dl_list.de_count =
2523 			cpu_to_le16(ocfs2_dx_entries_per_leaf(osb->sb));
2524 
2525 		mlog(0,
2526 		     "Dir %llu, format dx_leaf: %llu, entry count: %u\n",
2527 		     (unsigned long long)OCFS2_I(dir)->ip_blkno,
2528 		     (unsigned long long)bh->b_blocknr,
2529 		     le16_to_cpu(dx_leaf->dl_list.de_count));
2530 
2531 		ocfs2_journal_dirty(handle, bh);
2532 	}
2533 
2534 	ret = 0;
2535 out:
2536 	return ret;
2537 }
2538 
2539 /*
2540  * Allocates and formats a new cluster for use in an indexed dir
2541  * leaf. This version will not do the extent insert, so that it can be
2542  * used by operations which need careful ordering.
2543  */
2544 static int __ocfs2_dx_dir_new_cluster(struct inode *dir,
2545 				      u32 cpos, handle_t *handle,
2546 				      struct ocfs2_alloc_context *data_ac,
2547 				      struct buffer_head **dx_leaves,
2548 				      int num_dx_leaves, u64 *ret_phys_blkno)
2549 {
2550 	int ret;
2551 	u32 phys, num;
2552 	u64 phys_blkno;
2553 	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2554 
2555 	/*
2556 	 * XXX: For create, this should claim cluster for the index
2557 	 * *before* the unindexed insert so that we have a better
2558 	 * chance of contiguousness as the directory grows in number
2559 	 * of entries.
2560 	 */
2561 	ret = __ocfs2_claim_clusters(osb, handle, data_ac, 1, 1, &phys, &num);
2562 	if (ret) {
2563 		mlog_errno(ret);
2564 		goto out;
2565 	}
2566 
2567 	/*
2568 	 * Format the new cluster first. That way, we're inserting
2569 	 * valid data.
2570 	 */
2571 	phys_blkno = ocfs2_clusters_to_blocks(osb->sb, phys);
2572 	ret = ocfs2_dx_dir_format_cluster(osb, handle, dir, dx_leaves,
2573 					  num_dx_leaves, phys_blkno);
2574 	if (ret) {
2575 		mlog_errno(ret);
2576 		goto out;
2577 	}
2578 
2579 	*ret_phys_blkno = phys_blkno;
2580 out:
2581 	return ret;
2582 }
2583 
2584 static int ocfs2_dx_dir_new_cluster(struct inode *dir,
2585 				    struct ocfs2_extent_tree *et,
2586 				    u32 cpos, handle_t *handle,
2587 				    struct ocfs2_alloc_context *data_ac,
2588 				    struct ocfs2_alloc_context *meta_ac,
2589 				    struct buffer_head **dx_leaves,
2590 				    int num_dx_leaves)
2591 {
2592 	int ret;
2593 	u64 phys_blkno;
2594 
2595 	ret = __ocfs2_dx_dir_new_cluster(dir, cpos, handle, data_ac, dx_leaves,
2596 					 num_dx_leaves, &phys_blkno);
2597 	if (ret) {
2598 		mlog_errno(ret);
2599 		goto out;
2600 	}
2601 
2602 	ret = ocfs2_insert_extent(handle, et, cpos, phys_blkno, 1, 0,
2603 				  meta_ac);
2604 	if (ret)
2605 		mlog_errno(ret);
2606 out:
2607 	return ret;
2608 }
2609 
2610 static struct buffer_head **ocfs2_dx_dir_kmalloc_leaves(struct super_block *sb,
2611 							int *ret_num_leaves)
2612 {
2613 	int num_dx_leaves = ocfs2_clusters_to_blocks(sb, 1);
2614 	struct buffer_head **dx_leaves;
2615 
2616 	dx_leaves = kcalloc(num_dx_leaves, sizeof(struct buffer_head *),
2617 			    GFP_NOFS);
2618 	if (dx_leaves && ret_num_leaves)
2619 		*ret_num_leaves = num_dx_leaves;
2620 
2621 	return dx_leaves;
2622 }
2623 
2624 static int ocfs2_fill_new_dir_dx(struct ocfs2_super *osb,
2625 				 handle_t *handle,
2626 				 struct inode *parent,
2627 				 struct inode *inode,
2628 				 struct buffer_head *di_bh,
2629 				 struct ocfs2_alloc_context *data_ac,
2630 				 struct ocfs2_alloc_context *meta_ac)
2631 {
2632 	int ret;
2633 	struct buffer_head *leaf_bh = NULL;
2634 	struct buffer_head *dx_root_bh = NULL;
2635 	struct ocfs2_dx_hinfo hinfo;
2636 	struct ocfs2_dx_root_block *dx_root;
2637 	struct ocfs2_dx_entry_list *entry_list;
2638 
2639 	/*
2640 	 * Our strategy is to create the directory as though it were
2641 	 * unindexed, then add the index block. This works with very
2642 	 * little complication since the state of a new directory is a
2643 	 * very well known quantity.
2644 	 *
2645 	 * Essentially, we have two dirents ("." and ".."), in the 1st
2646 	 * block which need indexing. These are easily inserted into
2647 	 * the index block.
2648 	 */
2649 
2650 	ret = ocfs2_fill_new_dir_el(osb, handle, parent, inode, di_bh,
2651 				    data_ac, &leaf_bh);
2652 	if (ret) {
2653 		mlog_errno(ret);
2654 		goto out;
2655 	}
2656 
2657 	ret = ocfs2_dx_dir_attach_index(osb, handle, inode, di_bh, leaf_bh,
2658 					meta_ac, 1, 2, &dx_root_bh);
2659 	if (ret) {
2660 		mlog_errno(ret);
2661 		goto out;
2662 	}
2663 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2664 	entry_list = &dx_root->dr_entries;
2665 
2666 	/* Buffer has been journaled for us by ocfs2_dx_dir_attach_index */
2667 	ocfs2_dx_dir_name_hash(inode, ".", 1, &hinfo);
2668 	ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2669 
2670 	ocfs2_dx_dir_name_hash(inode, "..", 2, &hinfo);
2671 	ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2672 
2673 out:
2674 	brelse(dx_root_bh);
2675 	brelse(leaf_bh);
2676 	return ret;
2677 }
2678 
2679 int ocfs2_fill_new_dir(struct ocfs2_super *osb,
2680 		       handle_t *handle,
2681 		       struct inode *parent,
2682 		       struct inode *inode,
2683 		       struct buffer_head *fe_bh,
2684 		       struct ocfs2_alloc_context *data_ac,
2685 		       struct ocfs2_alloc_context *meta_ac)
2686 
2687 {
2688 	BUG_ON(!ocfs2_supports_inline_data(osb) && data_ac == NULL);
2689 
2690 	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2691 		return ocfs2_fill_new_dir_id(osb, handle, parent, inode, fe_bh);
2692 
2693 	if (ocfs2_supports_indexed_dirs(osb))
2694 		return ocfs2_fill_new_dir_dx(osb, handle, parent, inode, fe_bh,
2695 					     data_ac, meta_ac);
2696 
2697 	return ocfs2_fill_new_dir_el(osb, handle, parent, inode, fe_bh,
2698 				     data_ac, NULL);
2699 }
2700 
2701 static int ocfs2_dx_dir_index_block(struct inode *dir,
2702 				    handle_t *handle,
2703 				    struct buffer_head **dx_leaves,
2704 				    int num_dx_leaves,
2705 				    u32 *num_dx_entries,
2706 				    struct buffer_head *dirent_bh)
2707 {
2708 	int ret = 0, namelen, i;
2709 	char *de_buf, *limit;
2710 	struct ocfs2_dir_entry *de;
2711 	struct buffer_head *dx_leaf_bh;
2712 	struct ocfs2_dx_hinfo hinfo;
2713 	u64 dirent_blk = dirent_bh->b_blocknr;
2714 
2715 	de_buf = dirent_bh->b_data;
2716 	limit = de_buf + dir->i_sb->s_blocksize;
2717 
2718 	while (de_buf < limit) {
2719 		de = (struct ocfs2_dir_entry *)de_buf;
2720 
2721 		namelen = de->name_len;
2722 		if (!namelen || !de->inode)
2723 			goto inc;
2724 
2725 		ocfs2_dx_dir_name_hash(dir, de->name, namelen, &hinfo);
2726 
2727 		i = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb), &hinfo);
2728 		dx_leaf_bh = dx_leaves[i];
2729 
2730 		ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &hinfo,
2731 						 dirent_blk, dx_leaf_bh);
2732 		if (ret) {
2733 			mlog_errno(ret);
2734 			goto out;
2735 		}
2736 
2737 		*num_dx_entries = *num_dx_entries + 1;
2738 
2739 inc:
2740 		de_buf += le16_to_cpu(de->rec_len);
2741 	}
2742 
2743 out:
2744 	return ret;
2745 }
2746 
2747 /*
2748  * XXX: This expects dx_root_bh to already be part of the transaction.
2749  */
2750 static void ocfs2_dx_dir_index_root_block(struct inode *dir,
2751 					 struct buffer_head *dx_root_bh,
2752 					 struct buffer_head *dirent_bh)
2753 {
2754 	char *de_buf, *limit;
2755 	struct ocfs2_dx_root_block *dx_root;
2756 	struct ocfs2_dir_entry *de;
2757 	struct ocfs2_dx_hinfo hinfo;
2758 	u64 dirent_blk = dirent_bh->b_blocknr;
2759 
2760 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2761 
2762 	de_buf = dirent_bh->b_data;
2763 	limit = de_buf + dir->i_sb->s_blocksize;
2764 
2765 	while (de_buf < limit) {
2766 		de = (struct ocfs2_dir_entry *)de_buf;
2767 
2768 		if (!de->name_len || !de->inode)
2769 			goto inc;
2770 
2771 		ocfs2_dx_dir_name_hash(dir, de->name, de->name_len, &hinfo);
2772 
2773 		mlog(0,
2774 		     "dir: %llu, major: 0x%x minor: 0x%x, index: %u, name: %.*s\n",
2775 		     (unsigned long long)dir->i_ino, hinfo.major_hash,
2776 		     hinfo.minor_hash,
2777 		     le16_to_cpu(dx_root->dr_entries.de_num_used),
2778 		     de->name_len, de->name);
2779 
2780 		ocfs2_dx_entry_list_insert(&dx_root->dr_entries, &hinfo,
2781 					   dirent_blk);
2782 
2783 		le32_add_cpu(&dx_root->dr_num_entries, 1);
2784 inc:
2785 		de_buf += le16_to_cpu(de->rec_len);
2786 	}
2787 }
2788 
2789 /*
2790  * Count the number of inline directory entries in di_bh and compare
2791  * them against the number of entries we can hold in an inline dx root
2792  * block.
2793  */
2794 static int ocfs2_new_dx_should_be_inline(struct inode *dir,
2795 					 struct buffer_head *di_bh)
2796 {
2797 	int dirent_count = 0;
2798 	char *de_buf, *limit;
2799 	struct ocfs2_dir_entry *de;
2800 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2801 
2802 	de_buf = di->id2.i_data.id_data;
2803 	limit = de_buf + i_size_read(dir);
2804 
2805 	while (de_buf < limit) {
2806 		de = (struct ocfs2_dir_entry *)de_buf;
2807 
2808 		if (de->name_len && de->inode)
2809 			dirent_count++;
2810 
2811 		de_buf += le16_to_cpu(de->rec_len);
2812 	}
2813 
2814 	/* We are careful to leave room for one extra record. */
2815 	return dirent_count < ocfs2_dx_entries_per_root(dir->i_sb);
2816 }
2817 
2818 /*
2819  * Expand rec_len of the rightmost dirent in a directory block so that it
2820  * contains the end of our valid space for dirents. We do this during
2821  * expansion from an inline directory to one with extents. The first dir block
2822  * in that case is taken from the inline data portion of the inode block.
2823  *
2824  * This will also return the largest amount of contiguous space for a dirent
2825  * in the block. That value is *not* necessarily the last dirent, even after
2826  * expansion. The directory indexing code wants this value for free space
2827  * accounting. We do this here since we're already walking the entire dir
2828  * block.
2829  *
2830  * We add the dir trailer if this filesystem wants it.
2831  */
2832 static unsigned int ocfs2_expand_last_dirent(char *start, unsigned int old_size,
2833 					     struct inode *dir)
2834 {
2835 	struct super_block *sb = dir->i_sb;
2836 	struct ocfs2_dir_entry *de;
2837 	struct ocfs2_dir_entry *prev_de;
2838 	char *de_buf, *limit;
2839 	unsigned int new_size = sb->s_blocksize;
2840 	unsigned int bytes, this_hole;
2841 	unsigned int largest_hole = 0;
2842 
2843 	if (ocfs2_new_dir_wants_trailer(dir))
2844 		new_size = ocfs2_dir_trailer_blk_off(sb);
2845 
2846 	bytes = new_size - old_size;
2847 
2848 	limit = start + old_size;
2849 	de_buf = start;
2850 	de = (struct ocfs2_dir_entry *)de_buf;
2851 	do {
2852 		this_hole = ocfs2_figure_dirent_hole(de);
2853 		if (this_hole > largest_hole)
2854 			largest_hole = this_hole;
2855 
2856 		prev_de = de;
2857 		de_buf += le16_to_cpu(de->rec_len);
2858 		de = (struct ocfs2_dir_entry *)de_buf;
2859 	} while (de_buf < limit);
2860 
2861 	le16_add_cpu(&prev_de->rec_len, bytes);
2862 
2863 	/* We need to double check this after modification of the final
2864 	 * dirent. */
2865 	this_hole = ocfs2_figure_dirent_hole(prev_de);
2866 	if (this_hole > largest_hole)
2867 		largest_hole = this_hole;
2868 
2869 	if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
2870 		return largest_hole;
2871 	return 0;
2872 }
2873 
2874 /*
2875  * We allocate enough clusters to fulfill "blocks_wanted", but set
2876  * i_size to exactly one block. Ocfs2_extend_dir() will handle the
2877  * rest automatically for us.
2878  *
2879  * *first_block_bh is a pointer to the 1st data block allocated to the
2880  *  directory.
2881  */
2882 static int ocfs2_expand_inline_dir(struct inode *dir, struct buffer_head *di_bh,
2883 				   unsigned int blocks_wanted,
2884 				   struct ocfs2_dir_lookup_result *lookup,
2885 				   struct buffer_head **first_block_bh)
2886 {
2887 	u32 alloc, dx_alloc, bit_off, len, num_dx_entries = 0;
2888 	struct super_block *sb = dir->i_sb;
2889 	int ret, i, num_dx_leaves = 0, dx_inline = 0,
2890 		credits = ocfs2_inline_to_extents_credits(sb);
2891 	u64 dx_insert_blkno, blkno,
2892 		bytes = blocks_wanted << sb->s_blocksize_bits;
2893 	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2894 	struct ocfs2_inode_info *oi = OCFS2_I(dir);
2895 	struct ocfs2_alloc_context *data_ac;
2896 	struct ocfs2_alloc_context *meta_ac = NULL;
2897 	struct buffer_head *dirdata_bh = NULL;
2898 	struct buffer_head *dx_root_bh = NULL;
2899 	struct buffer_head **dx_leaves = NULL;
2900 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2901 	handle_t *handle;
2902 	struct ocfs2_extent_tree et;
2903 	struct ocfs2_extent_tree dx_et;
2904 	int did_quota = 0, bytes_allocated = 0;
2905 
2906 	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir), di_bh);
2907 
2908 	alloc = ocfs2_clusters_for_bytes(sb, bytes);
2909 	dx_alloc = 0;
2910 
2911 	down_write(&oi->ip_alloc_sem);
2912 
2913 	if (ocfs2_supports_indexed_dirs(osb)) {
2914 		credits += ocfs2_add_dir_index_credits(sb);
2915 
2916 		dx_inline = ocfs2_new_dx_should_be_inline(dir, di_bh);
2917 		if (!dx_inline) {
2918 			/* Add one more cluster for an index leaf */
2919 			dx_alloc++;
2920 			dx_leaves = ocfs2_dx_dir_kmalloc_leaves(sb,
2921 								&num_dx_leaves);
2922 			if (!dx_leaves) {
2923 				ret = -ENOMEM;
2924 				mlog_errno(ret);
2925 				goto out;
2926 			}
2927 		}
2928 
2929 		/* This gets us the dx_root */
2930 		ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
2931 		if (ret) {
2932 			mlog_errno(ret);
2933 			goto out;
2934 		}
2935 	}
2936 
2937 	/*
2938 	 * We should never need more than 2 clusters for the unindexed
2939 	 * tree - maximum dirent size is far less than one block. In
2940 	 * fact, the only time we'd need more than one cluster is if
2941 	 * blocksize == clustersize and the dirent won't fit in the
2942 	 * extra space that the expansion to a single block gives. As
2943 	 * of today, that only happens on 4k/4k file systems.
2944 	 */
2945 	BUG_ON(alloc > 2);
2946 
2947 	ret = ocfs2_reserve_clusters(osb, alloc + dx_alloc, &data_ac);
2948 	if (ret) {
2949 		mlog_errno(ret);
2950 		goto out;
2951 	}
2952 
2953 	/*
2954 	 * Prepare for worst case allocation scenario of two separate
2955 	 * extents in the unindexed tree.
2956 	 */
2957 	if (alloc == 2)
2958 		credits += OCFS2_SUBALLOC_ALLOC;
2959 
2960 	handle = ocfs2_start_trans(osb, credits);
2961 	if (IS_ERR(handle)) {
2962 		ret = PTR_ERR(handle);
2963 		mlog_errno(ret);
2964 		goto out;
2965 	}
2966 
2967 	if (vfs_dq_alloc_space_nodirty(dir,
2968 				ocfs2_clusters_to_bytes(osb->sb,
2969 							alloc + dx_alloc))) {
2970 		ret = -EDQUOT;
2971 		goto out_commit;
2972 	}
2973 	did_quota = 1;
2974 
2975 	if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2976 		/*
2977 		 * Allocate our index cluster first, to maximize the
2978 		 * possibility that unindexed leaves grow
2979 		 * contiguously.
2980 		 */
2981 		ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac,
2982 						 dx_leaves, num_dx_leaves,
2983 						 &dx_insert_blkno);
2984 		if (ret) {
2985 			mlog_errno(ret);
2986 			goto out_commit;
2987 		}
2988 		bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2989 	}
2990 
2991 	/*
2992 	 * Try to claim as many clusters as the bitmap can give though
2993 	 * if we only get one now, that's enough to continue. The rest
2994 	 * will be claimed after the conversion to extents.
2995 	 */
2996 	ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off, &len);
2997 	if (ret) {
2998 		mlog_errno(ret);
2999 		goto out_commit;
3000 	}
3001 	bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
3002 
3003 	/*
3004 	 * Operations are carefully ordered so that we set up the new
3005 	 * data block first. The conversion from inline data to
3006 	 * extents follows.
3007 	 */
3008 	blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3009 	dirdata_bh = sb_getblk(sb, blkno);
3010 	if (!dirdata_bh) {
3011 		ret = -EIO;
3012 		mlog_errno(ret);
3013 		goto out_commit;
3014 	}
3015 
3016 	ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dirdata_bh);
3017 
3018 	ret = ocfs2_journal_access_db(handle, INODE_CACHE(dir), dirdata_bh,
3019 				      OCFS2_JOURNAL_ACCESS_CREATE);
3020 	if (ret) {
3021 		mlog_errno(ret);
3022 		goto out_commit;
3023 	}
3024 
3025 	memcpy(dirdata_bh->b_data, di->id2.i_data.id_data, i_size_read(dir));
3026 	memset(dirdata_bh->b_data + i_size_read(dir), 0,
3027 	       sb->s_blocksize - i_size_read(dir));
3028 	i = ocfs2_expand_last_dirent(dirdata_bh->b_data, i_size_read(dir), dir);
3029 	if (ocfs2_new_dir_wants_trailer(dir)) {
3030 		/*
3031 		 * Prepare the dir trailer up front. It will otherwise look
3032 		 * like a valid dirent. Even if inserting the index fails
3033 		 * (unlikely), then all we'll have done is given first dir
3034 		 * block a small amount of fragmentation.
3035 		 */
3036 		ocfs2_init_dir_trailer(dir, dirdata_bh, i);
3037 	}
3038 
3039 	ret = ocfs2_journal_dirty(handle, dirdata_bh);
3040 	if (ret) {
3041 		mlog_errno(ret);
3042 		goto out_commit;
3043 	}
3044 
3045 	if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
3046 		/*
3047 		 * Dx dirs with an external cluster need to do this up
3048 		 * front. Inline dx root's get handled later, after
3049 		 * we've allocated our root block. We get passed back
3050 		 * a total number of items so that dr_num_entries can
3051 		 * be correctly set once the dx_root has been
3052 		 * allocated.
3053 		 */
3054 		ret = ocfs2_dx_dir_index_block(dir, handle, dx_leaves,
3055 					       num_dx_leaves, &num_dx_entries,
3056 					       dirdata_bh);
3057 		if (ret) {
3058 			mlog_errno(ret);
3059 			goto out_commit;
3060 		}
3061 	}
3062 
3063 	/*
3064 	 * Set extent, i_size, etc on the directory. After this, the
3065 	 * inode should contain the same exact dirents as before and
3066 	 * be fully accessible from system calls.
3067 	 *
3068 	 * We let the later dirent insert modify c/mtime - to the user
3069 	 * the data hasn't changed.
3070 	 */
3071 	ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
3072 				      OCFS2_JOURNAL_ACCESS_CREATE);
3073 	if (ret) {
3074 		mlog_errno(ret);
3075 		goto out_commit;
3076 	}
3077 
3078 	spin_lock(&oi->ip_lock);
3079 	oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
3080 	di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
3081 	spin_unlock(&oi->ip_lock);
3082 
3083 	ocfs2_dinode_new_extent_list(dir, di);
3084 
3085 	i_size_write(dir, sb->s_blocksize);
3086 	dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3087 
3088 	di->i_size = cpu_to_le64(sb->s_blocksize);
3089 	di->i_ctime = di->i_mtime = cpu_to_le64(dir->i_ctime.tv_sec);
3090 	di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(dir->i_ctime.tv_nsec);
3091 
3092 	/*
3093 	 * This should never fail as our extent list is empty and all
3094 	 * related blocks have been journaled already.
3095 	 */
3096 	ret = ocfs2_insert_extent(handle, &et, 0, blkno, len,
3097 				  0, NULL);
3098 	if (ret) {
3099 		mlog_errno(ret);
3100 		goto out_commit;
3101 	}
3102 
3103 	/*
3104 	 * Set i_blocks after the extent insert for the most up to
3105 	 * date ip_clusters value.
3106 	 */
3107 	dir->i_blocks = ocfs2_inode_sector_count(dir);
3108 
3109 	ret = ocfs2_journal_dirty(handle, di_bh);
3110 	if (ret) {
3111 		mlog_errno(ret);
3112 		goto out_commit;
3113 	}
3114 
3115 	if (ocfs2_supports_indexed_dirs(osb)) {
3116 		ret = ocfs2_dx_dir_attach_index(osb, handle, dir, di_bh,
3117 						dirdata_bh, meta_ac, dx_inline,
3118 						num_dx_entries, &dx_root_bh);
3119 		if (ret) {
3120 			mlog_errno(ret);
3121 			goto out_commit;
3122 		}
3123 
3124 		if (dx_inline) {
3125 			ocfs2_dx_dir_index_root_block(dir, dx_root_bh,
3126 						      dirdata_bh);
3127 		} else {
3128 			ocfs2_init_dx_root_extent_tree(&dx_et,
3129 						       INODE_CACHE(dir),
3130 						       dx_root_bh);
3131 			ret = ocfs2_insert_extent(handle, &dx_et, 0,
3132 						  dx_insert_blkno, 1, 0, NULL);
3133 			if (ret)
3134 				mlog_errno(ret);
3135 		}
3136 	}
3137 
3138 	/*
3139 	 * We asked for two clusters, but only got one in the 1st
3140 	 * pass. Claim the 2nd cluster as a separate extent.
3141 	 */
3142 	if (alloc > len) {
3143 		ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off,
3144 					   &len);
3145 		if (ret) {
3146 			mlog_errno(ret);
3147 			goto out_commit;
3148 		}
3149 		blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3150 
3151 		ret = ocfs2_insert_extent(handle, &et, 1,
3152 					  blkno, len, 0, NULL);
3153 		if (ret) {
3154 			mlog_errno(ret);
3155 			goto out_commit;
3156 		}
3157 		bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
3158 	}
3159 
3160 	*first_block_bh = dirdata_bh;
3161 	dirdata_bh = NULL;
3162 	if (ocfs2_supports_indexed_dirs(osb)) {
3163 		unsigned int off;
3164 
3165 		if (!dx_inline) {
3166 			/*
3167 			 * We need to return the correct block within the
3168 			 * cluster which should hold our entry.
3169 			 */
3170 			off = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb),
3171 						    &lookup->dl_hinfo);
3172 			get_bh(dx_leaves[off]);
3173 			lookup->dl_dx_leaf_bh = dx_leaves[off];
3174 		}
3175 		lookup->dl_dx_root_bh = dx_root_bh;
3176 		dx_root_bh = NULL;
3177 	}
3178 
3179 out_commit:
3180 	if (ret < 0 && did_quota)
3181 		vfs_dq_free_space_nodirty(dir, bytes_allocated);
3182 
3183 	ocfs2_commit_trans(osb, handle);
3184 
3185 out:
3186 	up_write(&oi->ip_alloc_sem);
3187 	if (data_ac)
3188 		ocfs2_free_alloc_context(data_ac);
3189 	if (meta_ac)
3190 		ocfs2_free_alloc_context(meta_ac);
3191 
3192 	if (dx_leaves) {
3193 		for (i = 0; i < num_dx_leaves; i++)
3194 			brelse(dx_leaves[i]);
3195 		kfree(dx_leaves);
3196 	}
3197 
3198 	brelse(dirdata_bh);
3199 	brelse(dx_root_bh);
3200 
3201 	return ret;
3202 }
3203 
3204 /* returns a bh of the 1st new block in the allocation. */
3205 static int ocfs2_do_extend_dir(struct super_block *sb,
3206 			       handle_t *handle,
3207 			       struct inode *dir,
3208 			       struct buffer_head *parent_fe_bh,
3209 			       struct ocfs2_alloc_context *data_ac,
3210 			       struct ocfs2_alloc_context *meta_ac,
3211 			       struct buffer_head **new_bh)
3212 {
3213 	int status;
3214 	int extend, did_quota = 0;
3215 	u64 p_blkno, v_blkno;
3216 
3217 	spin_lock(&OCFS2_I(dir)->ip_lock);
3218 	extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters));
3219 	spin_unlock(&OCFS2_I(dir)->ip_lock);
3220 
3221 	if (extend) {
3222 		u32 offset = OCFS2_I(dir)->ip_clusters;
3223 
3224 		if (vfs_dq_alloc_space_nodirty(dir,
3225 					ocfs2_clusters_to_bytes(sb, 1))) {
3226 			status = -EDQUOT;
3227 			goto bail;
3228 		}
3229 		did_quota = 1;
3230 
3231 		status = ocfs2_add_inode_data(OCFS2_SB(sb), dir, &offset,
3232 					      1, 0, parent_fe_bh, handle,
3233 					      data_ac, meta_ac, NULL);
3234 		BUG_ON(status == -EAGAIN);
3235 		if (status < 0) {
3236 			mlog_errno(status);
3237 			goto bail;
3238 		}
3239 	}
3240 
3241 	v_blkno = ocfs2_blocks_for_bytes(sb, i_size_read(dir));
3242 	status = ocfs2_extent_map_get_blocks(dir, v_blkno, &p_blkno, NULL, NULL);
3243 	if (status < 0) {
3244 		mlog_errno(status);
3245 		goto bail;
3246 	}
3247 
3248 	*new_bh = sb_getblk(sb, p_blkno);
3249 	if (!*new_bh) {
3250 		status = -EIO;
3251 		mlog_errno(status);
3252 		goto bail;
3253 	}
3254 	status = 0;
3255 bail:
3256 	if (did_quota && status < 0)
3257 		vfs_dq_free_space_nodirty(dir, ocfs2_clusters_to_bytes(sb, 1));
3258 	mlog_exit(status);
3259 	return status;
3260 }
3261 
3262 /*
3263  * Assumes you already have a cluster lock on the directory.
3264  *
3265  * 'blocks_wanted' is only used if we have an inline directory which
3266  * is to be turned into an extent based one. The size of the dirent to
3267  * insert might be larger than the space gained by growing to just one
3268  * block, so we may have to grow the inode by two blocks in that case.
3269  *
3270  * If the directory is already indexed, dx_root_bh must be provided.
3271  */
3272 static int ocfs2_extend_dir(struct ocfs2_super *osb,
3273 			    struct inode *dir,
3274 			    struct buffer_head *parent_fe_bh,
3275 			    unsigned int blocks_wanted,
3276 			    struct ocfs2_dir_lookup_result *lookup,
3277 			    struct buffer_head **new_de_bh)
3278 {
3279 	int status = 0;
3280 	int credits, num_free_extents, drop_alloc_sem = 0;
3281 	loff_t dir_i_size;
3282 	struct ocfs2_dinode *fe = (struct ocfs2_dinode *) parent_fe_bh->b_data;
3283 	struct ocfs2_extent_list *el = &fe->id2.i_list;
3284 	struct ocfs2_alloc_context *data_ac = NULL;
3285 	struct ocfs2_alloc_context *meta_ac = NULL;
3286 	handle_t *handle = NULL;
3287 	struct buffer_head *new_bh = NULL;
3288 	struct ocfs2_dir_entry * de;
3289 	struct super_block *sb = osb->sb;
3290 	struct ocfs2_extent_tree et;
3291 	struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
3292 
3293 	mlog_entry_void();
3294 
3295 	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
3296 		/*
3297 		 * This would be a code error as an inline directory should
3298 		 * never have an index root.
3299 		 */
3300 		BUG_ON(dx_root_bh);
3301 
3302 		status = ocfs2_expand_inline_dir(dir, parent_fe_bh,
3303 						 blocks_wanted, lookup,
3304 						 &new_bh);
3305 		if (status) {
3306 			mlog_errno(status);
3307 			goto bail;
3308 		}
3309 
3310 		/* Expansion from inline to an indexed directory will
3311 		 * have given us this. */
3312 		dx_root_bh = lookup->dl_dx_root_bh;
3313 
3314 		if (blocks_wanted == 1) {
3315 			/*
3316 			 * If the new dirent will fit inside the space
3317 			 * created by pushing out to one block, then
3318 			 * we can complete the operation
3319 			 * here. Otherwise we have to expand i_size
3320 			 * and format the 2nd block below.
3321 			 */
3322 			BUG_ON(new_bh == NULL);
3323 			goto bail_bh;
3324 		}
3325 
3326 		/*
3327 		 * Get rid of 'new_bh' - we want to format the 2nd
3328 		 * data block and return that instead.
3329 		 */
3330 		brelse(new_bh);
3331 		new_bh = NULL;
3332 
3333 		down_write(&OCFS2_I(dir)->ip_alloc_sem);
3334 		drop_alloc_sem = 1;
3335 		dir_i_size = i_size_read(dir);
3336 		credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3337 		goto do_extend;
3338 	}
3339 
3340 	down_write(&OCFS2_I(dir)->ip_alloc_sem);
3341 	drop_alloc_sem = 1;
3342 	dir_i_size = i_size_read(dir);
3343 	mlog(0, "extending dir %llu (i_size = %lld)\n",
3344 	     (unsigned long long)OCFS2_I(dir)->ip_blkno, dir_i_size);
3345 
3346 	/* dir->i_size is always block aligned. */
3347 	spin_lock(&OCFS2_I(dir)->ip_lock);
3348 	if (dir_i_size == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters)) {
3349 		spin_unlock(&OCFS2_I(dir)->ip_lock);
3350 		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir),
3351 					      parent_fe_bh);
3352 		num_free_extents = ocfs2_num_free_extents(osb, &et);
3353 		if (num_free_extents < 0) {
3354 			status = num_free_extents;
3355 			mlog_errno(status);
3356 			goto bail;
3357 		}
3358 
3359 		if (!num_free_extents) {
3360 			status = ocfs2_reserve_new_metadata(osb, el, &meta_ac);
3361 			if (status < 0) {
3362 				if (status != -ENOSPC)
3363 					mlog_errno(status);
3364 				goto bail;
3365 			}
3366 		}
3367 
3368 		status = ocfs2_reserve_clusters(osb, 1, &data_ac);
3369 		if (status < 0) {
3370 			if (status != -ENOSPC)
3371 				mlog_errno(status);
3372 			goto bail;
3373 		}
3374 
3375 		credits = ocfs2_calc_extend_credits(sb, el, 1);
3376 	} else {
3377 		spin_unlock(&OCFS2_I(dir)->ip_lock);
3378 		credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3379 	}
3380 
3381 do_extend:
3382 	if (ocfs2_dir_indexed(dir))
3383 		credits++; /* For attaching the new dirent block to the
3384 			    * dx_root */
3385 
3386 	handle = ocfs2_start_trans(osb, credits);
3387 	if (IS_ERR(handle)) {
3388 		status = PTR_ERR(handle);
3389 		handle = NULL;
3390 		mlog_errno(status);
3391 		goto bail;
3392 	}
3393 
3394 	status = ocfs2_do_extend_dir(osb->sb, handle, dir, parent_fe_bh,
3395 				     data_ac, meta_ac, &new_bh);
3396 	if (status < 0) {
3397 		mlog_errno(status);
3398 		goto bail;
3399 	}
3400 
3401 	ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), new_bh);
3402 
3403 	status = ocfs2_journal_access_db(handle, INODE_CACHE(dir), new_bh,
3404 					 OCFS2_JOURNAL_ACCESS_CREATE);
3405 	if (status < 0) {
3406 		mlog_errno(status);
3407 		goto bail;
3408 	}
3409 	memset(new_bh->b_data, 0, sb->s_blocksize);
3410 
3411 	de = (struct ocfs2_dir_entry *) new_bh->b_data;
3412 	de->inode = 0;
3413 	if (ocfs2_supports_dir_trailer(dir)) {
3414 		de->rec_len = cpu_to_le16(ocfs2_dir_trailer_blk_off(sb));
3415 
3416 		ocfs2_init_dir_trailer(dir, new_bh, le16_to_cpu(de->rec_len));
3417 
3418 		if (ocfs2_dir_indexed(dir)) {
3419 			status = ocfs2_dx_dir_link_trailer(dir, handle,
3420 							   dx_root_bh, new_bh);
3421 			if (status) {
3422 				mlog_errno(status);
3423 				goto bail;
3424 			}
3425 		}
3426 	} else {
3427 		de->rec_len = cpu_to_le16(sb->s_blocksize);
3428 	}
3429 	status = ocfs2_journal_dirty(handle, new_bh);
3430 	if (status < 0) {
3431 		mlog_errno(status);
3432 		goto bail;
3433 	}
3434 
3435 	dir_i_size += dir->i_sb->s_blocksize;
3436 	i_size_write(dir, dir_i_size);
3437 	dir->i_blocks = ocfs2_inode_sector_count(dir);
3438 	status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
3439 	if (status < 0) {
3440 		mlog_errno(status);
3441 		goto bail;
3442 	}
3443 
3444 bail_bh:
3445 	*new_de_bh = new_bh;
3446 	get_bh(*new_de_bh);
3447 bail:
3448 	if (handle)
3449 		ocfs2_commit_trans(osb, handle);
3450 	if (drop_alloc_sem)
3451 		up_write(&OCFS2_I(dir)->ip_alloc_sem);
3452 
3453 	if (data_ac)
3454 		ocfs2_free_alloc_context(data_ac);
3455 	if (meta_ac)
3456 		ocfs2_free_alloc_context(meta_ac);
3457 
3458 	brelse(new_bh);
3459 
3460 	mlog_exit(status);
3461 	return status;
3462 }
3463 
3464 static int ocfs2_find_dir_space_id(struct inode *dir, struct buffer_head *di_bh,
3465 				   const char *name, int namelen,
3466 				   struct buffer_head **ret_de_bh,
3467 				   unsigned int *blocks_wanted)
3468 {
3469 	int ret;
3470 	struct super_block *sb = dir->i_sb;
3471 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3472 	struct ocfs2_dir_entry *de, *last_de = NULL;
3473 	char *de_buf, *limit;
3474 	unsigned long offset = 0;
3475 	unsigned int rec_len, new_rec_len, free_space = dir->i_sb->s_blocksize;
3476 
3477 	/*
3478 	 * This calculates how many free bytes we'd have in block zero, should
3479 	 * this function force expansion to an extent tree.
3480 	 */
3481 	if (ocfs2_new_dir_wants_trailer(dir))
3482 		free_space = ocfs2_dir_trailer_blk_off(sb) - i_size_read(dir);
3483 	else
3484 		free_space = dir->i_sb->s_blocksize - i_size_read(dir);
3485 
3486 	de_buf = di->id2.i_data.id_data;
3487 	limit = de_buf + i_size_read(dir);
3488 	rec_len = OCFS2_DIR_REC_LEN(namelen);
3489 
3490 	while (de_buf < limit) {
3491 		de = (struct ocfs2_dir_entry *)de_buf;
3492 
3493 		if (!ocfs2_check_dir_entry(dir, de, di_bh, offset)) {
3494 			ret = -ENOENT;
3495 			goto out;
3496 		}
3497 		if (ocfs2_match(namelen, name, de)) {
3498 			ret = -EEXIST;
3499 			goto out;
3500 		}
3501 		/*
3502 		 * No need to check for a trailing dirent record here as
3503 		 * they're not used for inline dirs.
3504 		 */
3505 
3506 		if (ocfs2_dirent_would_fit(de, rec_len)) {
3507 			/* Ok, we found a spot. Return this bh and let
3508 			 * the caller actually fill it in. */
3509 			*ret_de_bh = di_bh;
3510 			get_bh(*ret_de_bh);
3511 			ret = 0;
3512 			goto out;
3513 		}
3514 
3515 		last_de = de;
3516 		de_buf += le16_to_cpu(de->rec_len);
3517 		offset += le16_to_cpu(de->rec_len);
3518 	}
3519 
3520 	/*
3521 	 * We're going to require expansion of the directory - figure
3522 	 * out how many blocks we'll need so that a place for the
3523 	 * dirent can be found.
3524 	 */
3525 	*blocks_wanted = 1;
3526 	new_rec_len = le16_to_cpu(last_de->rec_len) + free_space;
3527 	if (new_rec_len < (rec_len + OCFS2_DIR_REC_LEN(last_de->name_len)))
3528 		*blocks_wanted = 2;
3529 
3530 	ret = -ENOSPC;
3531 out:
3532 	return ret;
3533 }
3534 
3535 static int ocfs2_find_dir_space_el(struct inode *dir, const char *name,
3536 				   int namelen, struct buffer_head **ret_de_bh)
3537 {
3538 	unsigned long offset;
3539 	struct buffer_head *bh = NULL;
3540 	unsigned short rec_len;
3541 	struct ocfs2_dir_entry *de;
3542 	struct super_block *sb = dir->i_sb;
3543 	int status;
3544 	int blocksize = dir->i_sb->s_blocksize;
3545 
3546 	status = ocfs2_read_dir_block(dir, 0, &bh, 0);
3547 	if (status) {
3548 		mlog_errno(status);
3549 		goto bail;
3550 	}
3551 
3552 	rec_len = OCFS2_DIR_REC_LEN(namelen);
3553 	offset = 0;
3554 	de = (struct ocfs2_dir_entry *) bh->b_data;
3555 	while (1) {
3556 		if ((char *)de >= sb->s_blocksize + bh->b_data) {
3557 			brelse(bh);
3558 			bh = NULL;
3559 
3560 			if (i_size_read(dir) <= offset) {
3561 				/*
3562 				 * Caller will have to expand this
3563 				 * directory.
3564 				 */
3565 				status = -ENOSPC;
3566 				goto bail;
3567 			}
3568 			status = ocfs2_read_dir_block(dir,
3569 					     offset >> sb->s_blocksize_bits,
3570 					     &bh, 0);
3571 			if (status) {
3572 				mlog_errno(status);
3573 				goto bail;
3574 			}
3575 			/* move to next block */
3576 			de = (struct ocfs2_dir_entry *) bh->b_data;
3577 		}
3578 		if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
3579 			status = -ENOENT;
3580 			goto bail;
3581 		}
3582 		if (ocfs2_match(namelen, name, de)) {
3583 			status = -EEXIST;
3584 			goto bail;
3585 		}
3586 
3587 		if (ocfs2_skip_dir_trailer(dir, de, offset % blocksize,
3588 					   blocksize))
3589 			goto next;
3590 
3591 		if (ocfs2_dirent_would_fit(de, rec_len)) {
3592 			/* Ok, we found a spot. Return this bh and let
3593 			 * the caller actually fill it in. */
3594 			*ret_de_bh = bh;
3595 			get_bh(*ret_de_bh);
3596 			status = 0;
3597 			goto bail;
3598 		}
3599 next:
3600 		offset += le16_to_cpu(de->rec_len);
3601 		de = (struct ocfs2_dir_entry *)((char *) de + le16_to_cpu(de->rec_len));
3602 	}
3603 
3604 	status = 0;
3605 bail:
3606 	brelse(bh);
3607 
3608 	mlog_exit(status);
3609 	return status;
3610 }
3611 
3612 static int dx_leaf_sort_cmp(const void *a, const void *b)
3613 {
3614 	const struct ocfs2_dx_entry *entry1 = a;
3615 	const struct ocfs2_dx_entry *entry2 = b;
3616 	u32 major_hash1 = le32_to_cpu(entry1->dx_major_hash);
3617 	u32 major_hash2 = le32_to_cpu(entry2->dx_major_hash);
3618 	u32 minor_hash1 = le32_to_cpu(entry1->dx_minor_hash);
3619 	u32 minor_hash2 = le32_to_cpu(entry2->dx_minor_hash);
3620 
3621 	if (major_hash1 > major_hash2)
3622 		return 1;
3623 	if (major_hash1 < major_hash2)
3624 		return -1;
3625 
3626 	/*
3627 	 * It is not strictly necessary to sort by minor
3628 	 */
3629 	if (minor_hash1 > minor_hash2)
3630 		return 1;
3631 	if (minor_hash1 < minor_hash2)
3632 		return -1;
3633 	return 0;
3634 }
3635 
3636 static void dx_leaf_sort_swap(void *a, void *b, int size)
3637 {
3638 	struct ocfs2_dx_entry *entry1 = a;
3639 	struct ocfs2_dx_entry *entry2 = b;
3640 	struct ocfs2_dx_entry tmp;
3641 
3642 	BUG_ON(size != sizeof(*entry1));
3643 
3644 	tmp = *entry1;
3645 	*entry1 = *entry2;
3646 	*entry2 = tmp;
3647 }
3648 
3649 static int ocfs2_dx_leaf_same_major(struct ocfs2_dx_leaf *dx_leaf)
3650 {
3651 	struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3652 	int i, num = le16_to_cpu(dl_list->de_num_used);
3653 
3654 	for (i = 0; i < (num - 1); i++) {
3655 		if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) !=
3656 		    le32_to_cpu(dl_list->de_entries[i + 1].dx_major_hash))
3657 			return 0;
3658 	}
3659 
3660 	return 1;
3661 }
3662 
3663 /*
3664  * Find the optimal value to split this leaf on. This expects the leaf
3665  * entries to be in sorted order.
3666  *
3667  * leaf_cpos is the cpos of the leaf we're splitting. insert_hash is
3668  * the hash we want to insert.
3669  *
3670  * This function is only concerned with the major hash - that which
3671  * determines which cluster an item belongs to.
3672  */
3673 static int ocfs2_dx_dir_find_leaf_split(struct ocfs2_dx_leaf *dx_leaf,
3674 					u32 leaf_cpos, u32 insert_hash,
3675 					u32 *split_hash)
3676 {
3677 	struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3678 	int i, num_used = le16_to_cpu(dl_list->de_num_used);
3679 	int allsame;
3680 
3681 	/*
3682 	 * There's a couple rare, but nasty corner cases we have to
3683 	 * check for here. All of them involve a leaf where all value
3684 	 * have the same hash, which is what we look for first.
3685 	 *
3686 	 * Most of the time, all of the above is false, and we simply
3687 	 * pick the median value for a split.
3688 	 */
3689 	allsame = ocfs2_dx_leaf_same_major(dx_leaf);
3690 	if (allsame) {
3691 		u32 val = le32_to_cpu(dl_list->de_entries[0].dx_major_hash);
3692 
3693 		if (val == insert_hash) {
3694 			/*
3695 			 * No matter where we would choose to split,
3696 			 * the new entry would want to occupy the same
3697 			 * block as these. Since there's no space left
3698 			 * in their existing block, we know there
3699 			 * won't be space after the split.
3700 			 */
3701 			return -ENOSPC;
3702 		}
3703 
3704 		if (val == leaf_cpos) {
3705 			/*
3706 			 * Because val is the same as leaf_cpos (which
3707 			 * is the smallest value this leaf can have),
3708 			 * yet is not equal to insert_hash, then we
3709 			 * know that insert_hash *must* be larger than
3710 			 * val (and leaf_cpos). At least cpos+1 in value.
3711 			 *
3712 			 * We also know then, that there cannot be an
3713 			 * adjacent extent (otherwise we'd be looking
3714 			 * at it). Choosing this value gives us a
3715 			 * chance to get some contiguousness.
3716 			 */
3717 			*split_hash = leaf_cpos + 1;
3718 			return 0;
3719 		}
3720 
3721 		if (val > insert_hash) {
3722 			/*
3723 			 * val can not be the same as insert hash, and
3724 			 * also must be larger than leaf_cpos. Also,
3725 			 * we know that there can't be a leaf between
3726 			 * cpos and val, otherwise the entries with
3727 			 * hash 'val' would be there.
3728 			 */
3729 			*split_hash = val;
3730 			return 0;
3731 		}
3732 
3733 		*split_hash = insert_hash;
3734 		return 0;
3735 	}
3736 
3737 	/*
3738 	 * Since the records are sorted and the checks above
3739 	 * guaranteed that not all records in this block are the same,
3740 	 * we simple travel forward, from the median, and pick the 1st
3741 	 * record whose value is larger than leaf_cpos.
3742 	 */
3743 	for (i = (num_used / 2); i < num_used; i++)
3744 		if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) >
3745 		    leaf_cpos)
3746 			break;
3747 
3748 	BUG_ON(i == num_used); /* Should be impossible */
3749 	*split_hash = le32_to_cpu(dl_list->de_entries[i].dx_major_hash);
3750 	return 0;
3751 }
3752 
3753 /*
3754  * Transfer all entries in orig_dx_leaves whose major hash is equal to or
3755  * larger than split_hash into new_dx_leaves. We use a temporary
3756  * buffer (tmp_dx_leaf) to make the changes to the original leaf blocks.
3757  *
3758  * Since the block offset inside a leaf (cluster) is a constant mask
3759  * of minor_hash, we can optimize - an item at block offset X within
3760  * the original cluster, will be at offset X within the new cluster.
3761  */
3762 static void ocfs2_dx_dir_transfer_leaf(struct inode *dir, u32 split_hash,
3763 				       handle_t *handle,
3764 				       struct ocfs2_dx_leaf *tmp_dx_leaf,
3765 				       struct buffer_head **orig_dx_leaves,
3766 				       struct buffer_head **new_dx_leaves,
3767 				       int num_dx_leaves)
3768 {
3769 	int i, j, num_used;
3770 	u32 major_hash;
3771 	struct ocfs2_dx_leaf *orig_dx_leaf, *new_dx_leaf;
3772 	struct ocfs2_dx_entry_list *orig_list, *new_list, *tmp_list;
3773 	struct ocfs2_dx_entry *dx_entry;
3774 
3775 	tmp_list = &tmp_dx_leaf->dl_list;
3776 
3777 	for (i = 0; i < num_dx_leaves; i++) {
3778 		orig_dx_leaf = (struct ocfs2_dx_leaf *) orig_dx_leaves[i]->b_data;
3779 		orig_list = &orig_dx_leaf->dl_list;
3780 		new_dx_leaf = (struct ocfs2_dx_leaf *) new_dx_leaves[i]->b_data;
3781 		new_list = &new_dx_leaf->dl_list;
3782 
3783 		num_used = le16_to_cpu(orig_list->de_num_used);
3784 
3785 		memcpy(tmp_dx_leaf, orig_dx_leaf, dir->i_sb->s_blocksize);
3786 		tmp_list->de_num_used = cpu_to_le16(0);
3787 		memset(&tmp_list->de_entries, 0, sizeof(*dx_entry)*num_used);
3788 
3789 		for (j = 0; j < num_used; j++) {
3790 			dx_entry = &orig_list->de_entries[j];
3791 			major_hash = le32_to_cpu(dx_entry->dx_major_hash);
3792 			if (major_hash >= split_hash)
3793 				ocfs2_dx_dir_leaf_insert_tail(new_dx_leaf,
3794 							      dx_entry);
3795 			else
3796 				ocfs2_dx_dir_leaf_insert_tail(tmp_dx_leaf,
3797 							      dx_entry);
3798 		}
3799 		memcpy(orig_dx_leaf, tmp_dx_leaf, dir->i_sb->s_blocksize);
3800 
3801 		ocfs2_journal_dirty(handle, orig_dx_leaves[i]);
3802 		ocfs2_journal_dirty(handle, new_dx_leaves[i]);
3803 	}
3804 }
3805 
3806 static int ocfs2_dx_dir_rebalance_credits(struct ocfs2_super *osb,
3807 					  struct ocfs2_dx_root_block *dx_root)
3808 {
3809 	int credits = ocfs2_clusters_to_blocks(osb->sb, 2);
3810 
3811 	credits += ocfs2_calc_extend_credits(osb->sb, &dx_root->dr_list, 1);
3812 	credits += ocfs2_quota_trans_credits(osb->sb);
3813 	return credits;
3814 }
3815 
3816 /*
3817  * Find the median value in dx_leaf_bh and allocate a new leaf to move
3818  * half our entries into.
3819  */
3820 static int ocfs2_dx_dir_rebalance(struct ocfs2_super *osb, struct inode *dir,
3821 				  struct buffer_head *dx_root_bh,
3822 				  struct buffer_head *dx_leaf_bh,
3823 				  struct ocfs2_dx_hinfo *hinfo, u32 leaf_cpos,
3824 				  u64 leaf_blkno)
3825 {
3826 	struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3827 	int credits, ret, i, num_used, did_quota = 0;
3828 	u32 cpos, split_hash, insert_hash = hinfo->major_hash;
3829 	u64 orig_leaves_start;
3830 	int num_dx_leaves;
3831 	struct buffer_head **orig_dx_leaves = NULL;
3832 	struct buffer_head **new_dx_leaves = NULL;
3833 	struct ocfs2_alloc_context *data_ac = NULL, *meta_ac = NULL;
3834 	struct ocfs2_extent_tree et;
3835 	handle_t *handle = NULL;
3836 	struct ocfs2_dx_root_block *dx_root;
3837 	struct ocfs2_dx_leaf *tmp_dx_leaf = NULL;
3838 
3839 	mlog(0, "DX Dir: %llu, rebalance leaf leaf_blkno: %llu insert: %u\n",
3840 	     (unsigned long long)OCFS2_I(dir)->ip_blkno,
3841 	     (unsigned long long)leaf_blkno, insert_hash);
3842 
3843 	ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
3844 
3845 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3846 	/*
3847 	 * XXX: This is a rather large limit. We should use a more
3848 	 * realistic value.
3849 	 */
3850 	if (le32_to_cpu(dx_root->dr_clusters) == UINT_MAX)
3851 		return -ENOSPC;
3852 
3853 	num_used = le16_to_cpu(dx_leaf->dl_list.de_num_used);
3854 	if (num_used < le16_to_cpu(dx_leaf->dl_list.de_count)) {
3855 		mlog(ML_ERROR, "DX Dir: %llu, Asked to rebalance empty leaf: "
3856 		     "%llu, %d\n", (unsigned long long)OCFS2_I(dir)->ip_blkno,
3857 		     (unsigned long long)leaf_blkno, num_used);
3858 		ret = -EIO;
3859 		goto out;
3860 	}
3861 
3862 	orig_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
3863 	if (!orig_dx_leaves) {
3864 		ret = -ENOMEM;
3865 		mlog_errno(ret);
3866 		goto out;
3867 	}
3868 
3869 	new_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, NULL);
3870 	if (!new_dx_leaves) {
3871 		ret = -ENOMEM;
3872 		mlog_errno(ret);
3873 		goto out;
3874 	}
3875 
3876 	ret = ocfs2_lock_allocators(dir, &et, 1, 0, &data_ac, &meta_ac);
3877 	if (ret) {
3878 		if (ret != -ENOSPC)
3879 			mlog_errno(ret);
3880 		goto out;
3881 	}
3882 
3883 	credits = ocfs2_dx_dir_rebalance_credits(osb, dx_root);
3884 	handle = ocfs2_start_trans(osb, credits);
3885 	if (IS_ERR(handle)) {
3886 		ret = PTR_ERR(handle);
3887 		handle = NULL;
3888 		mlog_errno(ret);
3889 		goto out;
3890 	}
3891 
3892 	if (vfs_dq_alloc_space_nodirty(dir,
3893 				       ocfs2_clusters_to_bytes(dir->i_sb, 1))) {
3894 		ret = -EDQUOT;
3895 		goto out_commit;
3896 	}
3897 	did_quota = 1;
3898 
3899 	ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
3900 				      OCFS2_JOURNAL_ACCESS_WRITE);
3901 	if (ret) {
3902 		mlog_errno(ret);
3903 		goto out_commit;
3904 	}
3905 
3906 	/*
3907 	 * This block is changing anyway, so we can sort it in place.
3908 	 */
3909 	sort(dx_leaf->dl_list.de_entries, num_used,
3910 	     sizeof(struct ocfs2_dx_entry), dx_leaf_sort_cmp,
3911 	     dx_leaf_sort_swap);
3912 
3913 	ret = ocfs2_journal_dirty(handle, dx_leaf_bh);
3914 	if (ret) {
3915 		mlog_errno(ret);
3916 		goto out_commit;
3917 	}
3918 
3919 	ret = ocfs2_dx_dir_find_leaf_split(dx_leaf, leaf_cpos, insert_hash,
3920 					   &split_hash);
3921 	if (ret) {
3922 		mlog_errno(ret);
3923 		goto  out_commit;
3924 	}
3925 
3926 	mlog(0, "Split leaf (%u) at %u, insert major hash is %u\n",
3927 	     leaf_cpos, split_hash, insert_hash);
3928 
3929 	/*
3930 	 * We have to carefully order operations here. There are items
3931 	 * which want to be in the new cluster before insert, but in
3932 	 * order to put those items in the new cluster, we alter the
3933 	 * old cluster. A failure to insert gets nasty.
3934 	 *
3935 	 * So, start by reserving writes to the old
3936 	 * cluster. ocfs2_dx_dir_new_cluster will reserve writes on
3937 	 * the new cluster for us, before inserting it. The insert
3938 	 * won't happen if there's an error before that. Once the
3939 	 * insert is done then, we can transfer from one leaf into the
3940 	 * other without fear of hitting any error.
3941 	 */
3942 
3943 	/*
3944 	 * The leaf transfer wants some scratch space so that we don't
3945 	 * wind up doing a bunch of expensive memmove().
3946 	 */
3947 	tmp_dx_leaf = kmalloc(osb->sb->s_blocksize, GFP_NOFS);
3948 	if (!tmp_dx_leaf) {
3949 		ret = -ENOMEM;
3950 		mlog_errno(ret);
3951 		goto out_commit;
3952 	}
3953 
3954 	orig_leaves_start = ocfs2_block_to_cluster_start(dir->i_sb, leaf_blkno);
3955 	ret = ocfs2_read_dx_leaves(dir, orig_leaves_start, num_dx_leaves,
3956 				   orig_dx_leaves);
3957 	if (ret) {
3958 		mlog_errno(ret);
3959 		goto out_commit;
3960 	}
3961 
3962 	for (i = 0; i < num_dx_leaves; i++) {
3963 		ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3964 					      orig_dx_leaves[i],
3965 					      OCFS2_JOURNAL_ACCESS_WRITE);
3966 		if (ret) {
3967 			mlog_errno(ret);
3968 			goto out_commit;
3969 		}
3970 	}
3971 
3972 	cpos = split_hash;
3973 	ret = ocfs2_dx_dir_new_cluster(dir, &et, cpos, handle,
3974 				       data_ac, meta_ac, new_dx_leaves,
3975 				       num_dx_leaves);
3976 	if (ret) {
3977 		mlog_errno(ret);
3978 		goto out_commit;
3979 	}
3980 
3981 	ocfs2_dx_dir_transfer_leaf(dir, split_hash, handle, tmp_dx_leaf,
3982 				   orig_dx_leaves, new_dx_leaves, num_dx_leaves);
3983 
3984 out_commit:
3985 	if (ret < 0 && did_quota)
3986 		vfs_dq_free_space_nodirty(dir,
3987 				ocfs2_clusters_to_bytes(dir->i_sb, 1));
3988 
3989 	ocfs2_commit_trans(osb, handle);
3990 
3991 out:
3992 	if (orig_dx_leaves || new_dx_leaves) {
3993 		for (i = 0; i < num_dx_leaves; i++) {
3994 			if (orig_dx_leaves)
3995 				brelse(orig_dx_leaves[i]);
3996 			if (new_dx_leaves)
3997 				brelse(new_dx_leaves[i]);
3998 		}
3999 		kfree(orig_dx_leaves);
4000 		kfree(new_dx_leaves);
4001 	}
4002 
4003 	if (meta_ac)
4004 		ocfs2_free_alloc_context(meta_ac);
4005 	if (data_ac)
4006 		ocfs2_free_alloc_context(data_ac);
4007 
4008 	kfree(tmp_dx_leaf);
4009 	return ret;
4010 }
4011 
4012 static int ocfs2_find_dir_space_dx(struct ocfs2_super *osb, struct inode *dir,
4013 				   struct buffer_head *di_bh,
4014 				   struct buffer_head *dx_root_bh,
4015 				   const char *name, int namelen,
4016 				   struct ocfs2_dir_lookup_result *lookup)
4017 {
4018 	int ret, rebalanced = 0;
4019 	struct ocfs2_dx_root_block *dx_root;
4020 	struct buffer_head *dx_leaf_bh = NULL;
4021 	struct ocfs2_dx_leaf *dx_leaf;
4022 	u64 blkno;
4023 	u32 leaf_cpos;
4024 
4025 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4026 
4027 restart_search:
4028 	ret = ocfs2_dx_dir_lookup(dir, &dx_root->dr_list, &lookup->dl_hinfo,
4029 				  &leaf_cpos, &blkno);
4030 	if (ret) {
4031 		mlog_errno(ret);
4032 		goto out;
4033 	}
4034 
4035 	ret = ocfs2_read_dx_leaf(dir, blkno, &dx_leaf_bh);
4036 	if (ret) {
4037 		mlog_errno(ret);
4038 		goto out;
4039 	}
4040 
4041 	dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
4042 
4043 	if (le16_to_cpu(dx_leaf->dl_list.de_num_used) >=
4044 	    le16_to_cpu(dx_leaf->dl_list.de_count)) {
4045 		if (rebalanced) {
4046 			/*
4047 			 * Rebalancing should have provided us with
4048 			 * space in an appropriate leaf.
4049 			 *
4050 			 * XXX: Is this an abnormal condition then?
4051 			 * Should we print a message here?
4052 			 */
4053 			ret = -ENOSPC;
4054 			goto out;
4055 		}
4056 
4057 		ret = ocfs2_dx_dir_rebalance(osb, dir, dx_root_bh, dx_leaf_bh,
4058 					     &lookup->dl_hinfo, leaf_cpos,
4059 					     blkno);
4060 		if (ret) {
4061 			if (ret != -ENOSPC)
4062 				mlog_errno(ret);
4063 			goto out;
4064 		}
4065 
4066 		/*
4067 		 * Restart the lookup. The rebalance might have
4068 		 * changed which block our item fits into. Mark our
4069 		 * progress, so we only execute this once.
4070 		 */
4071 		brelse(dx_leaf_bh);
4072 		dx_leaf_bh = NULL;
4073 		rebalanced = 1;
4074 		goto restart_search;
4075 	}
4076 
4077 	lookup->dl_dx_leaf_bh = dx_leaf_bh;
4078 	dx_leaf_bh = NULL;
4079 
4080 out:
4081 	brelse(dx_leaf_bh);
4082 	return ret;
4083 }
4084 
4085 static int ocfs2_search_dx_free_list(struct inode *dir,
4086 				     struct buffer_head *dx_root_bh,
4087 				     int namelen,
4088 				     struct ocfs2_dir_lookup_result *lookup)
4089 {
4090 	int ret = -ENOSPC;
4091 	struct buffer_head *leaf_bh = NULL, *prev_leaf_bh = NULL;
4092 	struct ocfs2_dir_block_trailer *db;
4093 	u64 next_block;
4094 	int rec_len = OCFS2_DIR_REC_LEN(namelen);
4095 	struct ocfs2_dx_root_block *dx_root;
4096 
4097 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4098 	next_block = le64_to_cpu(dx_root->dr_free_blk);
4099 
4100 	while (next_block) {
4101 		brelse(prev_leaf_bh);
4102 		prev_leaf_bh = leaf_bh;
4103 		leaf_bh = NULL;
4104 
4105 		ret = ocfs2_read_dir_block_direct(dir, next_block, &leaf_bh);
4106 		if (ret) {
4107 			mlog_errno(ret);
4108 			goto out;
4109 		}
4110 
4111 		db = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
4112 		if (rec_len <= le16_to_cpu(db->db_free_rec_len)) {
4113 			lookup->dl_leaf_bh = leaf_bh;
4114 			lookup->dl_prev_leaf_bh = prev_leaf_bh;
4115 			leaf_bh = NULL;
4116 			prev_leaf_bh = NULL;
4117 			break;
4118 		}
4119 
4120 		next_block = le64_to_cpu(db->db_free_next);
4121 	}
4122 
4123 	if (!next_block)
4124 		ret = -ENOSPC;
4125 
4126 out:
4127 
4128 	brelse(leaf_bh);
4129 	brelse(prev_leaf_bh);
4130 	return ret;
4131 }
4132 
4133 static int ocfs2_expand_inline_dx_root(struct inode *dir,
4134 				       struct buffer_head *dx_root_bh)
4135 {
4136 	int ret, num_dx_leaves, i, j, did_quota = 0;
4137 	struct buffer_head **dx_leaves = NULL;
4138 	struct ocfs2_extent_tree et;
4139 	u64 insert_blkno;
4140 	struct ocfs2_alloc_context *data_ac = NULL;
4141 	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4142 	handle_t *handle = NULL;
4143 	struct ocfs2_dx_root_block *dx_root;
4144 	struct ocfs2_dx_entry_list *entry_list;
4145 	struct ocfs2_dx_entry *dx_entry;
4146 	struct ocfs2_dx_leaf *target_leaf;
4147 
4148 	ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
4149 	if (ret) {
4150 		mlog_errno(ret);
4151 		goto out;
4152 	}
4153 
4154 	dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
4155 	if (!dx_leaves) {
4156 		ret = -ENOMEM;
4157 		mlog_errno(ret);
4158 		goto out;
4159 	}
4160 
4161 	handle = ocfs2_start_trans(osb, ocfs2_calc_dxi_expand_credits(osb->sb));
4162 	if (IS_ERR(handle)) {
4163 		ret = PTR_ERR(handle);
4164 		mlog_errno(ret);
4165 		goto out;
4166 	}
4167 
4168 	if (vfs_dq_alloc_space_nodirty(dir,
4169 				       ocfs2_clusters_to_bytes(osb->sb, 1))) {
4170 		ret = -EDQUOT;
4171 		goto out_commit;
4172 	}
4173 	did_quota = 1;
4174 
4175 	/*
4176 	 * We do this up front, before the allocation, so that a
4177 	 * failure to add the dx_root_bh to the journal won't result
4178 	 * us losing clusters.
4179 	 */
4180 	ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
4181 				      OCFS2_JOURNAL_ACCESS_WRITE);
4182 	if (ret) {
4183 		mlog_errno(ret);
4184 		goto out_commit;
4185 	}
4186 
4187 	ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac, dx_leaves,
4188 					 num_dx_leaves, &insert_blkno);
4189 	if (ret) {
4190 		mlog_errno(ret);
4191 		goto out_commit;
4192 	}
4193 
4194 	/*
4195 	 * Transfer the entries from our dx_root into the appropriate
4196 	 * block
4197 	 */
4198 	dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4199 	entry_list = &dx_root->dr_entries;
4200 
4201 	for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
4202 		dx_entry = &entry_list->de_entries[i];
4203 
4204 		j = __ocfs2_dx_dir_hash_idx(osb,
4205 					    le32_to_cpu(dx_entry->dx_minor_hash));
4206 		target_leaf = (struct ocfs2_dx_leaf *)dx_leaves[j]->b_data;
4207 
4208 		ocfs2_dx_dir_leaf_insert_tail(target_leaf, dx_entry);
4209 
4210 		/* Each leaf has been passed to the journal already
4211 		 * via __ocfs2_dx_dir_new_cluster() */
4212 	}
4213 
4214 	dx_root->dr_flags &= ~OCFS2_DX_FLAG_INLINE;
4215 	memset(&dx_root->dr_list, 0, osb->sb->s_blocksize -
4216 	       offsetof(struct ocfs2_dx_root_block, dr_list));
4217 	dx_root->dr_list.l_count =
4218 		cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
4219 
4220 	/* This should never fail considering we start with an empty
4221 	 * dx_root. */
4222 	ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4223 	ret = ocfs2_insert_extent(handle, &et, 0, insert_blkno, 1, 0, NULL);
4224 	if (ret)
4225 		mlog_errno(ret);
4226 	did_quota = 0;
4227 
4228 	ocfs2_journal_dirty(handle, dx_root_bh);
4229 
4230 out_commit:
4231 	if (ret < 0 && did_quota)
4232 		vfs_dq_free_space_nodirty(dir,
4233 					  ocfs2_clusters_to_bytes(dir->i_sb, 1));
4234 
4235 	ocfs2_commit_trans(osb, handle);
4236 
4237 out:
4238 	if (data_ac)
4239 		ocfs2_free_alloc_context(data_ac);
4240 
4241 	if (dx_leaves) {
4242 		for (i = 0; i < num_dx_leaves; i++)
4243 			brelse(dx_leaves[i]);
4244 		kfree(dx_leaves);
4245 	}
4246 	return ret;
4247 }
4248 
4249 static int ocfs2_inline_dx_has_space(struct buffer_head *dx_root_bh)
4250 {
4251 	struct ocfs2_dx_root_block *dx_root;
4252 	struct ocfs2_dx_entry_list *entry_list;
4253 
4254 	dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4255 	entry_list = &dx_root->dr_entries;
4256 
4257 	if (le16_to_cpu(entry_list->de_num_used) >=
4258 	    le16_to_cpu(entry_list->de_count))
4259 		return -ENOSPC;
4260 
4261 	return 0;
4262 }
4263 
4264 static int ocfs2_prepare_dx_dir_for_insert(struct inode *dir,
4265 					   struct buffer_head *di_bh,
4266 					   const char *name,
4267 					   int namelen,
4268 					   struct ocfs2_dir_lookup_result *lookup)
4269 {
4270 	int ret, free_dx_root = 1;
4271 	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4272 	struct buffer_head *dx_root_bh = NULL;
4273 	struct buffer_head *leaf_bh = NULL;
4274 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4275 	struct ocfs2_dx_root_block *dx_root;
4276 
4277 	ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4278 	if (ret) {
4279 		mlog_errno(ret);
4280 		goto out;
4281 	}
4282 
4283 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4284 	if (le32_to_cpu(dx_root->dr_num_entries) == OCFS2_DX_ENTRIES_MAX) {
4285 		ret = -ENOSPC;
4286 		mlog_errno(ret);
4287 		goto out;
4288 	}
4289 
4290 	if (ocfs2_dx_root_inline(dx_root)) {
4291 		ret = ocfs2_inline_dx_has_space(dx_root_bh);
4292 
4293 		if (ret == 0)
4294 			goto search_el;
4295 
4296 		/*
4297 		 * We ran out of room in the root block. Expand it to
4298 		 * an extent, then allow ocfs2_find_dir_space_dx to do
4299 		 * the rest.
4300 		 */
4301 		ret = ocfs2_expand_inline_dx_root(dir, dx_root_bh);
4302 		if (ret) {
4303 			mlog_errno(ret);
4304 			goto out;
4305 		}
4306 	}
4307 
4308 	/*
4309 	 * Insert preparation for an indexed directory is split into two
4310 	 * steps. The call to find_dir_space_dx reserves room in the index for
4311 	 * an additional item. If we run out of space there, it's a real error
4312 	 * we can't continue on.
4313 	 */
4314 	ret = ocfs2_find_dir_space_dx(osb, dir, di_bh, dx_root_bh, name,
4315 				      namelen, lookup);
4316 	if (ret) {
4317 		mlog_errno(ret);
4318 		goto out;
4319 	}
4320 
4321 search_el:
4322 	/*
4323 	 * Next, we need to find space in the unindexed tree. This call
4324 	 * searches using the free space linked list. If the unindexed tree
4325 	 * lacks sufficient space, we'll expand it below. The expansion code
4326 	 * is smart enough to add any new blocks to the free space list.
4327 	 */
4328 	ret = ocfs2_search_dx_free_list(dir, dx_root_bh, namelen, lookup);
4329 	if (ret && ret != -ENOSPC) {
4330 		mlog_errno(ret);
4331 		goto out;
4332 	}
4333 
4334 	/* Do this up here - ocfs2_extend_dir might need the dx_root */
4335 	lookup->dl_dx_root_bh = dx_root_bh;
4336 	free_dx_root = 0;
4337 
4338 	if (ret == -ENOSPC) {
4339 		ret = ocfs2_extend_dir(osb, dir, di_bh, 1, lookup, &leaf_bh);
4340 
4341 		if (ret) {
4342 			mlog_errno(ret);
4343 			goto out;
4344 		}
4345 
4346 		/*
4347 		 * We make the assumption here that new leaf blocks are added
4348 		 * to the front of our free list.
4349 		 */
4350 		lookup->dl_prev_leaf_bh = NULL;
4351 		lookup->dl_leaf_bh = leaf_bh;
4352 	}
4353 
4354 out:
4355 	if (free_dx_root)
4356 		brelse(dx_root_bh);
4357 	return ret;
4358 }
4359 
4360 /*
4361  * Get a directory ready for insert. Any directory allocation required
4362  * happens here. Success returns zero, and enough context in the dir
4363  * lookup result that ocfs2_add_entry() will be able complete the task
4364  * with minimal performance impact.
4365  */
4366 int ocfs2_prepare_dir_for_insert(struct ocfs2_super *osb,
4367 				 struct inode *dir,
4368 				 struct buffer_head *parent_fe_bh,
4369 				 const char *name,
4370 				 int namelen,
4371 				 struct ocfs2_dir_lookup_result *lookup)
4372 {
4373 	int ret;
4374 	unsigned int blocks_wanted = 1;
4375 	struct buffer_head *bh = NULL;
4376 
4377 	mlog(0, "getting ready to insert namelen %d into dir %llu\n",
4378 	     namelen, (unsigned long long)OCFS2_I(dir)->ip_blkno);
4379 
4380 	if (!namelen) {
4381 		ret = -EINVAL;
4382 		mlog_errno(ret);
4383 		goto out;
4384 	}
4385 
4386 	/*
4387 	 * Do this up front to reduce confusion.
4388 	 *
4389 	 * The directory might start inline, then be turned into an
4390 	 * indexed one, in which case we'd need to hash deep inside
4391 	 * ocfs2_find_dir_space_id(). Since
4392 	 * ocfs2_prepare_dx_dir_for_insert() also needs this hash
4393 	 * done, there seems no point in spreading out the calls. We
4394 	 * can optimize away the case where the file system doesn't
4395 	 * support indexing.
4396 	 */
4397 	if (ocfs2_supports_indexed_dirs(osb))
4398 		ocfs2_dx_dir_name_hash(dir, name, namelen, &lookup->dl_hinfo);
4399 
4400 	if (ocfs2_dir_indexed(dir)) {
4401 		ret = ocfs2_prepare_dx_dir_for_insert(dir, parent_fe_bh,
4402 						      name, namelen, lookup);
4403 		if (ret)
4404 			mlog_errno(ret);
4405 		goto out;
4406 	}
4407 
4408 	if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4409 		ret = ocfs2_find_dir_space_id(dir, parent_fe_bh, name,
4410 					      namelen, &bh, &blocks_wanted);
4411 	} else
4412 		ret = ocfs2_find_dir_space_el(dir, name, namelen, &bh);
4413 
4414 	if (ret && ret != -ENOSPC) {
4415 		mlog_errno(ret);
4416 		goto out;
4417 	}
4418 
4419 	if (ret == -ENOSPC) {
4420 		/*
4421 		 * We have to expand the directory to add this name.
4422 		 */
4423 		BUG_ON(bh);
4424 
4425 		ret = ocfs2_extend_dir(osb, dir, parent_fe_bh, blocks_wanted,
4426 				       lookup, &bh);
4427 		if (ret) {
4428 			if (ret != -ENOSPC)
4429 				mlog_errno(ret);
4430 			goto out;
4431 		}
4432 
4433 		BUG_ON(!bh);
4434 	}
4435 
4436 	lookup->dl_leaf_bh = bh;
4437 	bh = NULL;
4438 out:
4439 	brelse(bh);
4440 	return ret;
4441 }
4442 
4443 static int ocfs2_dx_dir_remove_index(struct inode *dir,
4444 				     struct buffer_head *di_bh,
4445 				     struct buffer_head *dx_root_bh)
4446 {
4447 	int ret;
4448 	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4449 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4450 	struct ocfs2_dx_root_block *dx_root;
4451 	struct inode *dx_alloc_inode = NULL;
4452 	struct buffer_head *dx_alloc_bh = NULL;
4453 	handle_t *handle;
4454 	u64 blk;
4455 	u16 bit;
4456 	u64 bg_blkno;
4457 
4458 	dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4459 
4460 	dx_alloc_inode = ocfs2_get_system_file_inode(osb,
4461 					EXTENT_ALLOC_SYSTEM_INODE,
4462 					le16_to_cpu(dx_root->dr_suballoc_slot));
4463 	if (!dx_alloc_inode) {
4464 		ret = -ENOMEM;
4465 		mlog_errno(ret);
4466 		goto out;
4467 	}
4468 	mutex_lock(&dx_alloc_inode->i_mutex);
4469 
4470 	ret = ocfs2_inode_lock(dx_alloc_inode, &dx_alloc_bh, 1);
4471 	if (ret) {
4472 		mlog_errno(ret);
4473 		goto out_mutex;
4474 	}
4475 
4476 	handle = ocfs2_start_trans(osb, OCFS2_DX_ROOT_REMOVE_CREDITS);
4477 	if (IS_ERR(handle)) {
4478 		ret = PTR_ERR(handle);
4479 		mlog_errno(ret);
4480 		goto out_unlock;
4481 	}
4482 
4483 	ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
4484 				      OCFS2_JOURNAL_ACCESS_WRITE);
4485 	if (ret) {
4486 		mlog_errno(ret);
4487 		goto out_commit;
4488 	}
4489 
4490 	OCFS2_I(dir)->ip_dyn_features &= ~OCFS2_INDEXED_DIR_FL;
4491 	di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
4492 	di->i_dx_root = cpu_to_le64(0ULL);
4493 
4494 	ocfs2_journal_dirty(handle, di_bh);
4495 
4496 	blk = le64_to_cpu(dx_root->dr_blkno);
4497 	bit = le16_to_cpu(dx_root->dr_suballoc_bit);
4498 	bg_blkno = ocfs2_which_suballoc_group(blk, bit);
4499 	ret = ocfs2_free_suballoc_bits(handle, dx_alloc_inode, dx_alloc_bh,
4500 				       bit, bg_blkno, 1);
4501 	if (ret)
4502 		mlog_errno(ret);
4503 
4504 out_commit:
4505 	ocfs2_commit_trans(osb, handle);
4506 
4507 out_unlock:
4508 	ocfs2_inode_unlock(dx_alloc_inode, 1);
4509 
4510 out_mutex:
4511 	mutex_unlock(&dx_alloc_inode->i_mutex);
4512 	brelse(dx_alloc_bh);
4513 out:
4514 	iput(dx_alloc_inode);
4515 	return ret;
4516 }
4517 
4518 int ocfs2_dx_dir_truncate(struct inode *dir, struct buffer_head *di_bh)
4519 {
4520 	int ret;
4521 	unsigned int uninitialized_var(clen);
4522 	u32 major_hash = UINT_MAX, p_cpos, uninitialized_var(cpos);
4523 	u64 uninitialized_var(blkno);
4524 	struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4525 	struct buffer_head *dx_root_bh = NULL;
4526 	struct ocfs2_dx_root_block *dx_root;
4527 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4528 	struct ocfs2_cached_dealloc_ctxt dealloc;
4529 	struct ocfs2_extent_tree et;
4530 
4531 	ocfs2_init_dealloc_ctxt(&dealloc);
4532 
4533 	if (!ocfs2_dir_indexed(dir))
4534 		return 0;
4535 
4536 	ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4537 	if (ret) {
4538 		mlog_errno(ret);
4539 		goto out;
4540 	}
4541 	dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4542 
4543 	if (ocfs2_dx_root_inline(dx_root))
4544 		goto remove_index;
4545 
4546 	ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4547 
4548 	/* XXX: What if dr_clusters is too large? */
4549 	while (le32_to_cpu(dx_root->dr_clusters)) {
4550 		ret = ocfs2_dx_dir_lookup_rec(dir, &dx_root->dr_list,
4551 					      major_hash, &cpos, &blkno, &clen);
4552 		if (ret) {
4553 			mlog_errno(ret);
4554 			goto out;
4555 		}
4556 
4557 		p_cpos = ocfs2_blocks_to_clusters(dir->i_sb, blkno);
4558 
4559 		ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen,
4560 					       &dealloc);
4561 		if (ret) {
4562 			mlog_errno(ret);
4563 			goto out;
4564 		}
4565 
4566 		if (cpos == 0)
4567 			break;
4568 
4569 		major_hash = cpos - 1;
4570 	}
4571 
4572 remove_index:
4573 	ret = ocfs2_dx_dir_remove_index(dir, di_bh, dx_root_bh);
4574 	if (ret) {
4575 		mlog_errno(ret);
4576 		goto out;
4577 	}
4578 
4579 	ocfs2_remove_from_cache(INODE_CACHE(dir), dx_root_bh);
4580 out:
4581 	ocfs2_schedule_truncate_log_flush(osb, 1);
4582 	ocfs2_run_deallocs(osb, &dealloc);
4583 
4584 	brelse(dx_root_bh);
4585 	return ret;
4586 }
4587