nfsd4: Fix filp leak
[linux-2.6.git] / fs / ocfs2 / dir.c
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 #include <cluster/masklog.h>
47
48 #include "ocfs2.h"
49
50 #include "alloc.h"
51 #include "blockcheck.h"
52 #include "dir.h"
53 #include "dlmglue.h"
54 #include "extent_map.h"
55 #include "file.h"
56 #include "inode.h"
57 #include "journal.h"
58 #include "namei.h"
59 #include "suballoc.h"
60 #include "super.h"
61 #include "sysfile.h"
62 #include "uptodate.h"
63 #include "ocfs2_trace.h"
64
65 #include "buffer_head_io.h"
66
67 #define NAMEI_RA_CHUNKS  2
68 #define NAMEI_RA_BLOCKS  4
69 #define NAMEI_RA_SIZE        (NAMEI_RA_CHUNKS * NAMEI_RA_BLOCKS)
70 #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 (unlikely(rlen < OCFS2_DIR_REC_LEN(1)))
326                 error_msg = "rec_len is smaller than minimal";
327         else if (unlikely(rlen % 4 != 0))
328                 error_msg = "rec_len % 4 != 0";
329         else if (unlikely(rlen < OCFS2_DIR_REC_LEN(de->name_len)))
330                 error_msg = "rec_len is too small for name_len";
331         else if (unlikely(
332                  ((char *) de - bh->b_data) + rlen > dir->i_sb->s_blocksize))
333                 error_msg = "directory entry across blocks";
334
335         if (unlikely(error_msg != NULL))
336                 mlog(ML_ERROR, "bad entry in directory #%llu: %s - "
337                      "offset=%lu, inode=%llu, rec_len=%d, name_len=%d\n",
338                      (unsigned long long)OCFS2_I(dir)->ip_blkno, error_msg,
339                      offset, (unsigned long long)le64_to_cpu(de->inode), rlen,
340                      de->name_len);
341
342         return error_msg == NULL ? 1 : 0;
343 }
344
345 static inline int ocfs2_match(int len,
346                               const char * const name,
347                               struct ocfs2_dir_entry *de)
348 {
349         if (len != de->name_len)
350                 return 0;
351         if (!de->inode)
352                 return 0;
353         return !memcmp(name, de->name, len);
354 }
355
356 /*
357  * Returns 0 if not found, -1 on failure, and 1 on success
358  */
359 static inline int ocfs2_search_dirblock(struct buffer_head *bh,
360                                         struct inode *dir,
361                                         const char *name, int namelen,
362                                         unsigned long offset,
363                                         char *first_de,
364                                         unsigned int bytes,
365                                         struct ocfs2_dir_entry **res_dir)
366 {
367         struct ocfs2_dir_entry *de;
368         char *dlimit, *de_buf;
369         int de_len;
370         int ret = 0;
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         trace_ocfs2_search_dirblock(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         trace_ocfs2_validate_dir_block((unsigned long long)bh->b_blocknr);
451
452         BUG_ON(!buffer_uptodate(bh));
453
454         /*
455          * If the ecc fails, we return the error but otherwise
456          * leave the filesystem running.  We know any error is
457          * local to this block.
458          *
459          * Note that we are safe to call this even if the directory
460          * doesn't have a trailer.  Filesystems without metaecc will do
461          * nothing, and filesystems with it will have one.
462          */
463         rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &trailer->db_check);
464         if (rc)
465                 mlog(ML_ERROR, "Checksum failed for dinode %llu\n",
466                      (unsigned long long)bh->b_blocknr);
467
468         return rc;
469 }
470
471 /*
472  * Validate a directory trailer.
473  *
474  * We check the trailer here rather than in ocfs2_validate_dir_block()
475  * because that function doesn't have the inode to test.
476  */
477 static int ocfs2_check_dir_trailer(struct inode *dir, struct buffer_head *bh)
478 {
479         int rc = 0;
480         struct ocfs2_dir_block_trailer *trailer;
481
482         trailer = ocfs2_trailer_from_bh(bh, dir->i_sb);
483         if (!OCFS2_IS_VALID_DIR_TRAILER(trailer)) {
484                 rc = -EINVAL;
485                 ocfs2_error(dir->i_sb,
486                             "Invalid dirblock #%llu: "
487                             "signature = %.*s\n",
488                             (unsigned long long)bh->b_blocknr, 7,
489                             trailer->db_signature);
490                 goto out;
491         }
492         if (le64_to_cpu(trailer->db_blkno) != bh->b_blocknr) {
493                 rc = -EINVAL;
494                 ocfs2_error(dir->i_sb,
495                             "Directory block #%llu has an invalid "
496                             "db_blkno of %llu",
497                             (unsigned long long)bh->b_blocknr,
498                             (unsigned long long)le64_to_cpu(trailer->db_blkno));
499                 goto out;
500         }
501         if (le64_to_cpu(trailer->db_parent_dinode) !=
502             OCFS2_I(dir)->ip_blkno) {
503                 rc = -EINVAL;
504                 ocfs2_error(dir->i_sb,
505                             "Directory block #%llu on dinode "
506                             "#%llu has an invalid parent_dinode "
507                             "of %llu",
508                             (unsigned long long)bh->b_blocknr,
509                             (unsigned long long)OCFS2_I(dir)->ip_blkno,
510                             (unsigned long long)le64_to_cpu(trailer->db_blkno));
511                 goto out;
512         }
513 out:
514         return rc;
515 }
516
517 /*
518  * This function forces all errors to -EIO for consistency with its
519  * predecessor, ocfs2_bread().  We haven't audited what returning the
520  * real error codes would do to callers.  We log the real codes with
521  * mlog_errno() before we squash them.
522  */
523 static int ocfs2_read_dir_block(struct inode *inode, u64 v_block,
524                                 struct buffer_head **bh, int flags)
525 {
526         int rc = 0;
527         struct buffer_head *tmp = *bh;
528
529         rc = ocfs2_read_virt_blocks(inode, v_block, 1, &tmp, flags,
530                                     ocfs2_validate_dir_block);
531         if (rc) {
532                 mlog_errno(rc);
533                 goto out;
534         }
535
536         if (!(flags & OCFS2_BH_READAHEAD) &&
537             ocfs2_supports_dir_trailer(inode)) {
538                 rc = ocfs2_check_dir_trailer(inode, tmp);
539                 if (rc) {
540                         if (!*bh)
541                                 brelse(tmp);
542                         mlog_errno(rc);
543                         goto out;
544                 }
545         }
546
547         /* If ocfs2_read_virt_blocks() got us a new bh, pass it up. */
548         if (!*bh)
549                 *bh = tmp;
550
551 out:
552         return rc ? -EIO : 0;
553 }
554
555 /*
556  * Read the block at 'phys' which belongs to this directory
557  * inode. This function does no virtual->physical block translation -
558  * what's passed in is assumed to be a valid directory block.
559  */
560 static int ocfs2_read_dir_block_direct(struct inode *dir, u64 phys,
561                                        struct buffer_head **bh)
562 {
563         int ret;
564         struct buffer_head *tmp = *bh;
565
566         ret = ocfs2_read_block(INODE_CACHE(dir), phys, &tmp,
567                                ocfs2_validate_dir_block);
568         if (ret) {
569                 mlog_errno(ret);
570                 goto out;
571         }
572
573         if (ocfs2_supports_dir_trailer(dir)) {
574                 ret = ocfs2_check_dir_trailer(dir, tmp);
575                 if (ret) {
576                         if (!*bh)
577                                 brelse(tmp);
578                         mlog_errno(ret);
579                         goto out;
580                 }
581         }
582
583         if (!ret && !*bh)
584                 *bh = tmp;
585 out:
586         return ret;
587 }
588
589 static int ocfs2_validate_dx_root(struct super_block *sb,
590                                   struct buffer_head *bh)
591 {
592         int ret;
593         struct ocfs2_dx_root_block *dx_root;
594
595         BUG_ON(!buffer_uptodate(bh));
596
597         dx_root = (struct ocfs2_dx_root_block *) bh->b_data;
598
599         ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_root->dr_check);
600         if (ret) {
601                 mlog(ML_ERROR,
602                      "Checksum failed for dir index root block %llu\n",
603                      (unsigned long long)bh->b_blocknr);
604                 return ret;
605         }
606
607         if (!OCFS2_IS_VALID_DX_ROOT(dx_root)) {
608                 ocfs2_error(sb,
609                             "Dir Index Root # %llu has bad signature %.*s",
610                             (unsigned long long)le64_to_cpu(dx_root->dr_blkno),
611                             7, dx_root->dr_signature);
612                 return -EINVAL;
613         }
614
615         return 0;
616 }
617
618 static int ocfs2_read_dx_root(struct inode *dir, struct ocfs2_dinode *di,
619                               struct buffer_head **dx_root_bh)
620 {
621         int ret;
622         u64 blkno = le64_to_cpu(di->i_dx_root);
623         struct buffer_head *tmp = *dx_root_bh;
624
625         ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
626                                ocfs2_validate_dx_root);
627
628         /* If ocfs2_read_block() got us a new bh, pass it up. */
629         if (!ret && !*dx_root_bh)
630                 *dx_root_bh = tmp;
631
632         return ret;
633 }
634
635 static int ocfs2_validate_dx_leaf(struct super_block *sb,
636                                   struct buffer_head *bh)
637 {
638         int ret;
639         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)bh->b_data;
640
641         BUG_ON(!buffer_uptodate(bh));
642
643         ret = ocfs2_validate_meta_ecc(sb, bh->b_data, &dx_leaf->dl_check);
644         if (ret) {
645                 mlog(ML_ERROR,
646                      "Checksum failed for dir index leaf block %llu\n",
647                      (unsigned long long)bh->b_blocknr);
648                 return ret;
649         }
650
651         if (!OCFS2_IS_VALID_DX_LEAF(dx_leaf)) {
652                 ocfs2_error(sb, "Dir Index Leaf has bad signature %.*s",
653                             7, dx_leaf->dl_signature);
654                 return -EROFS;
655         }
656
657         return 0;
658 }
659
660 static int ocfs2_read_dx_leaf(struct inode *dir, u64 blkno,
661                               struct buffer_head **dx_leaf_bh)
662 {
663         int ret;
664         struct buffer_head *tmp = *dx_leaf_bh;
665
666         ret = ocfs2_read_block(INODE_CACHE(dir), blkno, &tmp,
667                                ocfs2_validate_dx_leaf);
668
669         /* If ocfs2_read_block() got us a new bh, pass it up. */
670         if (!ret && !*dx_leaf_bh)
671                 *dx_leaf_bh = tmp;
672
673         return ret;
674 }
675
676 /*
677  * Read a series of dx_leaf blocks. This expects all buffer_head
678  * pointers to be NULL on function entry.
679  */
680 static int ocfs2_read_dx_leaves(struct inode *dir, u64 start, int num,
681                                 struct buffer_head **dx_leaf_bhs)
682 {
683         int ret;
684
685         ret = ocfs2_read_blocks(INODE_CACHE(dir), start, num, dx_leaf_bhs, 0,
686                                 ocfs2_validate_dx_leaf);
687         if (ret)
688                 mlog_errno(ret);
689
690         return ret;
691 }
692
693 static struct buffer_head *ocfs2_find_entry_el(const char *name, int namelen,
694                                                struct inode *dir,
695                                                struct ocfs2_dir_entry **res_dir)
696 {
697         struct super_block *sb;
698         struct buffer_head *bh_use[NAMEI_RA_SIZE];
699         struct buffer_head *bh, *ret = NULL;
700         unsigned long start, block, b;
701         int ra_max = 0;         /* Number of bh's in the readahead
702                                    buffer, bh_use[] */
703         int ra_ptr = 0;         /* Current index into readahead
704                                    buffer */
705         int num = 0;
706         int nblocks, i, err;
707
708         sb = dir->i_sb;
709
710         nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
711         start = OCFS2_I(dir)->ip_dir_start_lookup;
712         if (start >= nblocks)
713                 start = 0;
714         block = start;
715
716 restart:
717         do {
718                 /*
719                  * We deal with the read-ahead logic here.
720                  */
721                 if (ra_ptr >= ra_max) {
722                         /* Refill the readahead buffer */
723                         ra_ptr = 0;
724                         b = block;
725                         for (ra_max = 0; ra_max < NAMEI_RA_SIZE; ra_max++) {
726                                 /*
727                                  * Terminate if we reach the end of the
728                                  * directory and must wrap, or if our
729                                  * search has finished at this block.
730                                  */
731                                 if (b >= nblocks || (num && block == start)) {
732                                         bh_use[ra_max] = NULL;
733                                         break;
734                                 }
735                                 num++;
736
737                                 bh = NULL;
738                                 err = ocfs2_read_dir_block(dir, b++, &bh,
739                                                            OCFS2_BH_READAHEAD);
740                                 bh_use[ra_max] = bh;
741                         }
742                 }
743                 if ((bh = bh_use[ra_ptr++]) == NULL)
744                         goto next;
745                 if (ocfs2_read_dir_block(dir, block, &bh, 0)) {
746                         /* read error, skip block & hope for the best.
747                          * ocfs2_read_dir_block() has released the bh. */
748                         ocfs2_error(dir->i_sb, "reading directory %llu, "
749                                     "offset %lu\n",
750                                     (unsigned long long)OCFS2_I(dir)->ip_blkno,
751                                     block);
752                         goto next;
753                 }
754                 i = ocfs2_search_dirblock(bh, dir, name, namelen,
755                                           block << sb->s_blocksize_bits,
756                                           bh->b_data, sb->s_blocksize,
757                                           res_dir);
758                 if (i == 1) {
759                         OCFS2_I(dir)->ip_dir_start_lookup = block;
760                         ret = bh;
761                         goto cleanup_and_exit;
762                 } else {
763                         brelse(bh);
764                         if (i < 0)
765                                 goto cleanup_and_exit;
766                 }
767         next:
768                 if (++block >= nblocks)
769                         block = 0;
770         } while (block != start);
771
772         /*
773          * If the directory has grown while we were searching, then
774          * search the last part of the directory before giving up.
775          */
776         block = nblocks;
777         nblocks = i_size_read(dir) >> sb->s_blocksize_bits;
778         if (block < nblocks) {
779                 start = 0;
780                 goto restart;
781         }
782
783 cleanup_and_exit:
784         /* Clean up the read-ahead blocks */
785         for (; ra_ptr < ra_max; ra_ptr++)
786                 brelse(bh_use[ra_ptr]);
787
788         trace_ocfs2_find_entry_el(ret);
789         return ret;
790 }
791
792 static int ocfs2_dx_dir_lookup_rec(struct inode *inode,
793                                    struct ocfs2_extent_list *el,
794                                    u32 major_hash,
795                                    u32 *ret_cpos,
796                                    u64 *ret_phys_blkno,
797                                    unsigned int *ret_clen)
798 {
799         int ret = 0, i, found;
800         struct buffer_head *eb_bh = NULL;
801         struct ocfs2_extent_block *eb;
802         struct ocfs2_extent_rec *rec = NULL;
803
804         if (el->l_tree_depth) {
805                 ret = ocfs2_find_leaf(INODE_CACHE(inode), el, major_hash,
806                                       &eb_bh);
807                 if (ret) {
808                         mlog_errno(ret);
809                         goto out;
810                 }
811
812                 eb = (struct ocfs2_extent_block *) eb_bh->b_data;
813                 el = &eb->h_list;
814
815                 if (el->l_tree_depth) {
816                         ocfs2_error(inode->i_sb,
817                                     "Inode %lu has non zero tree depth in "
818                                     "btree tree block %llu\n", inode->i_ino,
819                                     (unsigned long long)eb_bh->b_blocknr);
820                         ret = -EROFS;
821                         goto out;
822                 }
823         }
824
825         found = 0;
826         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
827                 rec = &el->l_recs[i];
828
829                 if (le32_to_cpu(rec->e_cpos) <= major_hash) {
830                         found = 1;
831                         break;
832                 }
833         }
834
835         if (!found) {
836                 ocfs2_error(inode->i_sb, "Inode %lu has bad extent "
837                             "record (%u, %u, 0) in btree", inode->i_ino,
838                             le32_to_cpu(rec->e_cpos),
839                             ocfs2_rec_clusters(el, rec));
840                 ret = -EROFS;
841                 goto out;
842         }
843
844         if (ret_phys_blkno)
845                 *ret_phys_blkno = le64_to_cpu(rec->e_blkno);
846         if (ret_cpos)
847                 *ret_cpos = le32_to_cpu(rec->e_cpos);
848         if (ret_clen)
849                 *ret_clen = le16_to_cpu(rec->e_leaf_clusters);
850
851 out:
852         brelse(eb_bh);
853         return ret;
854 }
855
856 /*
857  * Returns the block index, from the start of the cluster which this
858  * hash belongs too.
859  */
860 static inline unsigned int __ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
861                                                    u32 minor_hash)
862 {
863         return minor_hash & osb->osb_dx_mask;
864 }
865
866 static inline unsigned int ocfs2_dx_dir_hash_idx(struct ocfs2_super *osb,
867                                           struct ocfs2_dx_hinfo *hinfo)
868 {
869         return __ocfs2_dx_dir_hash_idx(osb, hinfo->minor_hash);
870 }
871
872 static int ocfs2_dx_dir_lookup(struct inode *inode,
873                                struct ocfs2_extent_list *el,
874                                struct ocfs2_dx_hinfo *hinfo,
875                                u32 *ret_cpos,
876                                u64 *ret_phys_blkno)
877 {
878         int ret = 0;
879         unsigned int cend, uninitialized_var(clen);
880         u32 uninitialized_var(cpos);
881         u64 uninitialized_var(blkno);
882         u32 name_hash = hinfo->major_hash;
883
884         ret = ocfs2_dx_dir_lookup_rec(inode, el, name_hash, &cpos, &blkno,
885                                       &clen);
886         if (ret) {
887                 mlog_errno(ret);
888                 goto out;
889         }
890
891         cend = cpos + clen;
892         if (name_hash >= cend) {
893                 /* We want the last cluster */
894                 blkno += ocfs2_clusters_to_blocks(inode->i_sb, clen - 1);
895                 cpos += clen - 1;
896         } else {
897                 blkno += ocfs2_clusters_to_blocks(inode->i_sb,
898                                                   name_hash - cpos);
899                 cpos = name_hash;
900         }
901
902         /*
903          * We now have the cluster which should hold our entry. To
904          * find the exact block from the start of the cluster to
905          * search, we take the lower bits of the hash.
906          */
907         blkno += ocfs2_dx_dir_hash_idx(OCFS2_SB(inode->i_sb), hinfo);
908
909         if (ret_phys_blkno)
910                 *ret_phys_blkno = blkno;
911         if (ret_cpos)
912                 *ret_cpos = cpos;
913
914 out:
915
916         return ret;
917 }
918
919 static int ocfs2_dx_dir_search(const char *name, int namelen,
920                                struct inode *dir,
921                                struct ocfs2_dx_root_block *dx_root,
922                                struct ocfs2_dir_lookup_result *res)
923 {
924         int ret, i, found;
925         u64 uninitialized_var(phys);
926         struct buffer_head *dx_leaf_bh = NULL;
927         struct ocfs2_dx_leaf *dx_leaf;
928         struct ocfs2_dx_entry *dx_entry = NULL;
929         struct buffer_head *dir_ent_bh = NULL;
930         struct ocfs2_dir_entry *dir_ent = NULL;
931         struct ocfs2_dx_hinfo *hinfo = &res->dl_hinfo;
932         struct ocfs2_extent_list *dr_el;
933         struct ocfs2_dx_entry_list *entry_list;
934
935         ocfs2_dx_dir_name_hash(dir, name, namelen, &res->dl_hinfo);
936
937         if (ocfs2_dx_root_inline(dx_root)) {
938                 entry_list = &dx_root->dr_entries;
939                 goto search;
940         }
941
942         dr_el = &dx_root->dr_list;
943
944         ret = ocfs2_dx_dir_lookup(dir, dr_el, hinfo, NULL, &phys);
945         if (ret) {
946                 mlog_errno(ret);
947                 goto out;
948         }
949
950         trace_ocfs2_dx_dir_search((unsigned long long)OCFS2_I(dir)->ip_blkno,
951                                   namelen, name, hinfo->major_hash,
952                                   hinfo->minor_hash, (unsigned long long)phys);
953
954         ret = ocfs2_read_dx_leaf(dir, phys, &dx_leaf_bh);
955         if (ret) {
956                 mlog_errno(ret);
957                 goto out;
958         }
959
960         dx_leaf = (struct ocfs2_dx_leaf *) dx_leaf_bh->b_data;
961
962         trace_ocfs2_dx_dir_search_leaf_info(
963                         le16_to_cpu(dx_leaf->dl_list.de_num_used),
964                         le16_to_cpu(dx_leaf->dl_list.de_count));
965
966         entry_list = &dx_leaf->dl_list;
967
968 search:
969         /*
970          * Empty leaf is legal, so no need to check for that.
971          */
972         found = 0;
973         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
974                 dx_entry = &entry_list->de_entries[i];
975
976                 if (hinfo->major_hash != le32_to_cpu(dx_entry->dx_major_hash)
977                     || hinfo->minor_hash != le32_to_cpu(dx_entry->dx_minor_hash))
978                         continue;
979
980                 /*
981                  * Search unindexed leaf block now. We're not
982                  * guaranteed to find anything.
983                  */
984                 ret = ocfs2_read_dir_block_direct(dir,
985                                           le64_to_cpu(dx_entry->dx_dirent_blk),
986                                           &dir_ent_bh);
987                 if (ret) {
988                         mlog_errno(ret);
989                         goto out;
990                 }
991
992                 /*
993                  * XXX: We should check the unindexed block here,
994                  * before using it.
995                  */
996
997                 found = ocfs2_search_dirblock(dir_ent_bh, dir, name, namelen,
998                                               0, dir_ent_bh->b_data,
999                                               dir->i_sb->s_blocksize, &dir_ent);
1000                 if (found == 1)
1001                         break;
1002
1003                 if (found == -1) {
1004                         /* This means we found a bad directory entry. */
1005                         ret = -EIO;
1006                         mlog_errno(ret);
1007                         goto out;
1008                 }
1009
1010                 brelse(dir_ent_bh);
1011                 dir_ent_bh = NULL;
1012         }
1013
1014         if (found <= 0) {
1015                 ret = -ENOENT;
1016                 goto out;
1017         }
1018
1019         res->dl_leaf_bh = dir_ent_bh;
1020         res->dl_entry = dir_ent;
1021         res->dl_dx_leaf_bh = dx_leaf_bh;
1022         res->dl_dx_entry = dx_entry;
1023
1024         ret = 0;
1025 out:
1026         if (ret) {
1027                 brelse(dx_leaf_bh);
1028                 brelse(dir_ent_bh);
1029         }
1030         return ret;
1031 }
1032
1033 static int ocfs2_find_entry_dx(const char *name, int namelen,
1034                                struct inode *dir,
1035                                struct ocfs2_dir_lookup_result *lookup)
1036 {
1037         int ret;
1038         struct buffer_head *di_bh = NULL;
1039         struct ocfs2_dinode *di;
1040         struct buffer_head *dx_root_bh = NULL;
1041         struct ocfs2_dx_root_block *dx_root;
1042
1043         ret = ocfs2_read_inode_block(dir, &di_bh);
1044         if (ret) {
1045                 mlog_errno(ret);
1046                 goto out;
1047         }
1048
1049         di = (struct ocfs2_dinode *)di_bh->b_data;
1050
1051         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
1052         if (ret) {
1053                 mlog_errno(ret);
1054                 goto out;
1055         }
1056         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
1057
1058         ret = ocfs2_dx_dir_search(name, namelen, dir, dx_root, lookup);
1059         if (ret) {
1060                 if (ret != -ENOENT)
1061                         mlog_errno(ret);
1062                 goto out;
1063         }
1064
1065         lookup->dl_dx_root_bh = dx_root_bh;
1066         dx_root_bh = NULL;
1067 out:
1068         brelse(di_bh);
1069         brelse(dx_root_bh);
1070         return ret;
1071 }
1072
1073 /*
1074  * Try to find an entry of the provided name within 'dir'.
1075  *
1076  * If nothing was found, -ENOENT is returned. Otherwise, zero is
1077  * returned and the struct 'res' will contain information useful to
1078  * other directory manipulation functions.
1079  *
1080  * Caller can NOT assume anything about the contents of the
1081  * buffer_heads - they are passed back only so that it can be passed
1082  * into any one of the manipulation functions (add entry, delete
1083  * entry, etc). As an example, bh in the extent directory case is a
1084  * data block, in the inline-data case it actually points to an inode,
1085  * in the indexed directory case, multiple buffers are involved.
1086  */
1087 int ocfs2_find_entry(const char *name, int namelen,
1088                      struct inode *dir, struct ocfs2_dir_lookup_result *lookup)
1089 {
1090         struct buffer_head *bh;
1091         struct ocfs2_dir_entry *res_dir = NULL;
1092
1093         if (ocfs2_dir_indexed(dir))
1094                 return ocfs2_find_entry_dx(name, namelen, dir, lookup);
1095
1096         /*
1097          * The unindexed dir code only uses part of the lookup
1098          * structure, so there's no reason to push it down further
1099          * than this.
1100          */
1101         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1102                 bh = ocfs2_find_entry_id(name, namelen, dir, &res_dir);
1103         else
1104                 bh = ocfs2_find_entry_el(name, namelen, dir, &res_dir);
1105
1106         if (bh == NULL)
1107                 return -ENOENT;
1108
1109         lookup->dl_leaf_bh = bh;
1110         lookup->dl_entry = res_dir;
1111         return 0;
1112 }
1113
1114 /*
1115  * Update inode number and type of a previously found directory entry.
1116  */
1117 int ocfs2_update_entry(struct inode *dir, handle_t *handle,
1118                        struct ocfs2_dir_lookup_result *res,
1119                        struct inode *new_entry_inode)
1120 {
1121         int ret;
1122         ocfs2_journal_access_func access = ocfs2_journal_access_db;
1123         struct ocfs2_dir_entry *de = res->dl_entry;
1124         struct buffer_head *de_bh = res->dl_leaf_bh;
1125
1126         /*
1127          * The same code works fine for both inline-data and extent
1128          * based directories, so no need to split this up.  The only
1129          * difference is the journal_access function.
1130          */
1131
1132         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1133                 access = ocfs2_journal_access_di;
1134
1135         ret = access(handle, INODE_CACHE(dir), de_bh,
1136                      OCFS2_JOURNAL_ACCESS_WRITE);
1137         if (ret) {
1138                 mlog_errno(ret);
1139                 goto out;
1140         }
1141
1142         de->inode = cpu_to_le64(OCFS2_I(new_entry_inode)->ip_blkno);
1143         ocfs2_set_de_type(de, new_entry_inode->i_mode);
1144
1145         ocfs2_journal_dirty(handle, de_bh);
1146
1147 out:
1148         return ret;
1149 }
1150
1151 /*
1152  * __ocfs2_delete_entry deletes a directory entry by merging it with the
1153  * previous entry
1154  */
1155 static int __ocfs2_delete_entry(handle_t *handle, struct inode *dir,
1156                                 struct ocfs2_dir_entry *de_del,
1157                                 struct buffer_head *bh, char *first_de,
1158                                 unsigned int bytes)
1159 {
1160         struct ocfs2_dir_entry *de, *pde;
1161         int i, status = -ENOENT;
1162         ocfs2_journal_access_func access = ocfs2_journal_access_db;
1163
1164         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1165                 access = ocfs2_journal_access_di;
1166
1167         i = 0;
1168         pde = NULL;
1169         de = (struct ocfs2_dir_entry *) first_de;
1170         while (i < bytes) {
1171                 if (!ocfs2_check_dir_entry(dir, de, bh, i)) {
1172                         status = -EIO;
1173                         mlog_errno(status);
1174                         goto bail;
1175                 }
1176                 if (de == de_del)  {
1177                         status = access(handle, INODE_CACHE(dir), bh,
1178                                         OCFS2_JOURNAL_ACCESS_WRITE);
1179                         if (status < 0) {
1180                                 status = -EIO;
1181                                 mlog_errno(status);
1182                                 goto bail;
1183                         }
1184                         if (pde)
1185                                 le16_add_cpu(&pde->rec_len,
1186                                                 le16_to_cpu(de->rec_len));
1187                         else
1188                                 de->inode = 0;
1189                         dir->i_version++;
1190                         ocfs2_journal_dirty(handle, bh);
1191                         goto bail;
1192                 }
1193                 i += le16_to_cpu(de->rec_len);
1194                 pde = de;
1195                 de = (struct ocfs2_dir_entry *)((char *)de + le16_to_cpu(de->rec_len));
1196         }
1197 bail:
1198         return status;
1199 }
1200
1201 static unsigned int ocfs2_figure_dirent_hole(struct ocfs2_dir_entry *de)
1202 {
1203         unsigned int hole;
1204
1205         if (le64_to_cpu(de->inode) == 0)
1206                 hole = le16_to_cpu(de->rec_len);
1207         else
1208                 hole = le16_to_cpu(de->rec_len) -
1209                         OCFS2_DIR_REC_LEN(de->name_len);
1210
1211         return hole;
1212 }
1213
1214 static int ocfs2_find_max_rec_len(struct super_block *sb,
1215                                   struct buffer_head *dirblock_bh)
1216 {
1217         int size, this_hole, largest_hole = 0;
1218         char *trailer, *de_buf, *limit, *start = dirblock_bh->b_data;
1219         struct ocfs2_dir_entry *de;
1220
1221         trailer = (char *)ocfs2_trailer_from_bh(dirblock_bh, sb);
1222         size = ocfs2_dir_trailer_blk_off(sb);
1223         limit = start + size;
1224         de_buf = start;
1225         de = (struct ocfs2_dir_entry *)de_buf;
1226         do {
1227                 if (de_buf != trailer) {
1228                         this_hole = ocfs2_figure_dirent_hole(de);
1229                         if (this_hole > largest_hole)
1230                                 largest_hole = this_hole;
1231                 }
1232
1233                 de_buf += le16_to_cpu(de->rec_len);
1234                 de = (struct ocfs2_dir_entry *)de_buf;
1235         } while (de_buf < limit);
1236
1237         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
1238                 return largest_hole;
1239         return 0;
1240 }
1241
1242 static void ocfs2_dx_list_remove_entry(struct ocfs2_dx_entry_list *entry_list,
1243                                        int index)
1244 {
1245         int num_used = le16_to_cpu(entry_list->de_num_used);
1246
1247         if (num_used == 1 || index == (num_used - 1))
1248                 goto clear;
1249
1250         memmove(&entry_list->de_entries[index],
1251                 &entry_list->de_entries[index + 1],
1252                 (num_used - index - 1)*sizeof(struct ocfs2_dx_entry));
1253 clear:
1254         num_used--;
1255         memset(&entry_list->de_entries[num_used], 0,
1256                sizeof(struct ocfs2_dx_entry));
1257         entry_list->de_num_used = cpu_to_le16(num_used);
1258 }
1259
1260 static int ocfs2_delete_entry_dx(handle_t *handle, struct inode *dir,
1261                                  struct ocfs2_dir_lookup_result *lookup)
1262 {
1263         int ret, index, max_rec_len, add_to_free_list = 0;
1264         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1265         struct buffer_head *leaf_bh = lookup->dl_leaf_bh;
1266         struct ocfs2_dx_leaf *dx_leaf;
1267         struct ocfs2_dx_entry *dx_entry = lookup->dl_dx_entry;
1268         struct ocfs2_dir_block_trailer *trailer;
1269         struct ocfs2_dx_root_block *dx_root;
1270         struct ocfs2_dx_entry_list *entry_list;
1271
1272         /*
1273          * This function gets a bit messy because we might have to
1274          * modify the root block, regardless of whether the indexed
1275          * entries are stored inline.
1276          */
1277
1278         /*
1279          * *Only* set 'entry_list' here, based on where we're looking
1280          * for the indexed entries. Later, we might still want to
1281          * journal both blocks, based on free list state.
1282          */
1283         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
1284         if (ocfs2_dx_root_inline(dx_root)) {
1285                 entry_list = &dx_root->dr_entries;
1286         } else {
1287                 dx_leaf = (struct ocfs2_dx_leaf *) lookup->dl_dx_leaf_bh->b_data;
1288                 entry_list = &dx_leaf->dl_list;
1289         }
1290
1291         /* Neither of these are a disk corruption - that should have
1292          * been caught by lookup, before we got here. */
1293         BUG_ON(le16_to_cpu(entry_list->de_count) <= 0);
1294         BUG_ON(le16_to_cpu(entry_list->de_num_used) <= 0);
1295
1296         index = (char *)dx_entry - (char *)entry_list->de_entries;
1297         index /= sizeof(*dx_entry);
1298
1299         if (index >= le16_to_cpu(entry_list->de_num_used)) {
1300                 mlog(ML_ERROR, "Dir %llu: Bad dx_entry ptr idx %d, (%p, %p)\n",
1301                      (unsigned long long)OCFS2_I(dir)->ip_blkno, index,
1302                      entry_list, dx_entry);
1303                 return -EIO;
1304         }
1305
1306         /*
1307          * We know that removal of this dirent will leave enough room
1308          * for a new one, so add this block to the free list if it
1309          * isn't already there.
1310          */
1311         trailer = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
1312         if (trailer->db_free_rec_len == 0)
1313                 add_to_free_list = 1;
1314
1315         /*
1316          * Add the block holding our index into the journal before
1317          * removing the unindexed entry. If we get an error return
1318          * from __ocfs2_delete_entry(), then it hasn't removed the
1319          * entry yet. Likewise, successful return means we *must*
1320          * remove the indexed entry.
1321          *
1322          * We're also careful to journal the root tree block here as
1323          * the entry count needs to be updated. Also, we might be
1324          * adding to the start of the free list.
1325          */
1326         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1327                                       OCFS2_JOURNAL_ACCESS_WRITE);
1328         if (ret) {
1329                 mlog_errno(ret);
1330                 goto out;
1331         }
1332
1333         if (!ocfs2_dx_root_inline(dx_root)) {
1334                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
1335                                               lookup->dl_dx_leaf_bh,
1336                                               OCFS2_JOURNAL_ACCESS_WRITE);
1337                 if (ret) {
1338                         mlog_errno(ret);
1339                         goto out;
1340                 }
1341         }
1342
1343         trace_ocfs2_delete_entry_dx((unsigned long long)OCFS2_I(dir)->ip_blkno,
1344                                     index);
1345
1346         ret = __ocfs2_delete_entry(handle, dir, lookup->dl_entry,
1347                                    leaf_bh, leaf_bh->b_data, leaf_bh->b_size);
1348         if (ret) {
1349                 mlog_errno(ret);
1350                 goto out;
1351         }
1352
1353         max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, leaf_bh);
1354         trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1355         if (add_to_free_list) {
1356                 trailer->db_free_next = dx_root->dr_free_blk;
1357                 dx_root->dr_free_blk = cpu_to_le64(leaf_bh->b_blocknr);
1358                 ocfs2_journal_dirty(handle, dx_root_bh);
1359         }
1360
1361         /* leaf_bh was journal_accessed for us in __ocfs2_delete_entry */
1362         ocfs2_journal_dirty(handle, leaf_bh);
1363
1364         le32_add_cpu(&dx_root->dr_num_entries, -1);
1365         ocfs2_journal_dirty(handle, dx_root_bh);
1366
1367         ocfs2_dx_list_remove_entry(entry_list, index);
1368
1369         if (!ocfs2_dx_root_inline(dx_root))
1370                 ocfs2_journal_dirty(handle, lookup->dl_dx_leaf_bh);
1371
1372 out:
1373         return ret;
1374 }
1375
1376 static inline int ocfs2_delete_entry_id(handle_t *handle,
1377                                         struct inode *dir,
1378                                         struct ocfs2_dir_entry *de_del,
1379                                         struct buffer_head *bh)
1380 {
1381         int ret;
1382         struct buffer_head *di_bh = NULL;
1383         struct ocfs2_dinode *di;
1384         struct ocfs2_inline_data *data;
1385
1386         ret = ocfs2_read_inode_block(dir, &di_bh);
1387         if (ret) {
1388                 mlog_errno(ret);
1389                 goto out;
1390         }
1391
1392         di = (struct ocfs2_dinode *)di_bh->b_data;
1393         data = &di->id2.i_data;
1394
1395         ret = __ocfs2_delete_entry(handle, dir, de_del, bh, data->id_data,
1396                                    i_size_read(dir));
1397
1398         brelse(di_bh);
1399 out:
1400         return ret;
1401 }
1402
1403 static inline int ocfs2_delete_entry_el(handle_t *handle,
1404                                         struct inode *dir,
1405                                         struct ocfs2_dir_entry *de_del,
1406                                         struct buffer_head *bh)
1407 {
1408         return __ocfs2_delete_entry(handle, dir, de_del, bh, bh->b_data,
1409                                     bh->b_size);
1410 }
1411
1412 /*
1413  * Delete a directory entry. Hide the details of directory
1414  * implementation from the caller.
1415  */
1416 int ocfs2_delete_entry(handle_t *handle,
1417                        struct inode *dir,
1418                        struct ocfs2_dir_lookup_result *res)
1419 {
1420         if (ocfs2_dir_indexed(dir))
1421                 return ocfs2_delete_entry_dx(handle, dir, res);
1422
1423         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1424                 return ocfs2_delete_entry_id(handle, dir, res->dl_entry,
1425                                              res->dl_leaf_bh);
1426
1427         return ocfs2_delete_entry_el(handle, dir, res->dl_entry,
1428                                      res->dl_leaf_bh);
1429 }
1430
1431 /*
1432  * Check whether 'de' has enough room to hold an entry of
1433  * 'new_rec_len' bytes.
1434  */
1435 static inline int ocfs2_dirent_would_fit(struct ocfs2_dir_entry *de,
1436                                          unsigned int new_rec_len)
1437 {
1438         unsigned int de_really_used;
1439
1440         /* Check whether this is an empty record with enough space */
1441         if (le64_to_cpu(de->inode) == 0 &&
1442             le16_to_cpu(de->rec_len) >= new_rec_len)
1443                 return 1;
1444
1445         /*
1446          * Record might have free space at the end which we can
1447          * use.
1448          */
1449         de_really_used = OCFS2_DIR_REC_LEN(de->name_len);
1450         if (le16_to_cpu(de->rec_len) >= (de_really_used + new_rec_len))
1451             return 1;
1452
1453         return 0;
1454 }
1455
1456 static void ocfs2_dx_dir_leaf_insert_tail(struct ocfs2_dx_leaf *dx_leaf,
1457                                           struct ocfs2_dx_entry *dx_new_entry)
1458 {
1459         int i;
1460
1461         i = le16_to_cpu(dx_leaf->dl_list.de_num_used);
1462         dx_leaf->dl_list.de_entries[i] = *dx_new_entry;
1463
1464         le16_add_cpu(&dx_leaf->dl_list.de_num_used, 1);
1465 }
1466
1467 static void ocfs2_dx_entry_list_insert(struct ocfs2_dx_entry_list *entry_list,
1468                                        struct ocfs2_dx_hinfo *hinfo,
1469                                        u64 dirent_blk)
1470 {
1471         int i;
1472         struct ocfs2_dx_entry *dx_entry;
1473
1474         i = le16_to_cpu(entry_list->de_num_used);
1475         dx_entry = &entry_list->de_entries[i];
1476
1477         memset(dx_entry, 0, sizeof(*dx_entry));
1478         dx_entry->dx_major_hash = cpu_to_le32(hinfo->major_hash);
1479         dx_entry->dx_minor_hash = cpu_to_le32(hinfo->minor_hash);
1480         dx_entry->dx_dirent_blk = cpu_to_le64(dirent_blk);
1481
1482         le16_add_cpu(&entry_list->de_num_used, 1);
1483 }
1484
1485 static int __ocfs2_dx_dir_leaf_insert(struct inode *dir, handle_t *handle,
1486                                       struct ocfs2_dx_hinfo *hinfo,
1487                                       u64 dirent_blk,
1488                                       struct buffer_head *dx_leaf_bh)
1489 {
1490         int ret;
1491         struct ocfs2_dx_leaf *dx_leaf;
1492
1493         ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
1494                                       OCFS2_JOURNAL_ACCESS_WRITE);
1495         if (ret) {
1496                 mlog_errno(ret);
1497                 goto out;
1498         }
1499
1500         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
1501         ocfs2_dx_entry_list_insert(&dx_leaf->dl_list, hinfo, dirent_blk);
1502         ocfs2_journal_dirty(handle, dx_leaf_bh);
1503
1504 out:
1505         return ret;
1506 }
1507
1508 static void ocfs2_dx_inline_root_insert(struct inode *dir, handle_t *handle,
1509                                         struct ocfs2_dx_hinfo *hinfo,
1510                                         u64 dirent_blk,
1511                                         struct ocfs2_dx_root_block *dx_root)
1512 {
1513         ocfs2_dx_entry_list_insert(&dx_root->dr_entries, hinfo, dirent_blk);
1514 }
1515
1516 static int ocfs2_dx_dir_insert(struct inode *dir, handle_t *handle,
1517                                struct ocfs2_dir_lookup_result *lookup)
1518 {
1519         int ret = 0;
1520         struct ocfs2_dx_root_block *dx_root;
1521         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
1522
1523         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
1524                                       OCFS2_JOURNAL_ACCESS_WRITE);
1525         if (ret) {
1526                 mlog_errno(ret);
1527                 goto out;
1528         }
1529
1530         dx_root = (struct ocfs2_dx_root_block *)lookup->dl_dx_root_bh->b_data;
1531         if (ocfs2_dx_root_inline(dx_root)) {
1532                 ocfs2_dx_inline_root_insert(dir, handle,
1533                                             &lookup->dl_hinfo,
1534                                             lookup->dl_leaf_bh->b_blocknr,
1535                                             dx_root);
1536         } else {
1537                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &lookup->dl_hinfo,
1538                                                  lookup->dl_leaf_bh->b_blocknr,
1539                                                  lookup->dl_dx_leaf_bh);
1540                 if (ret)
1541                         goto out;
1542         }
1543
1544         le32_add_cpu(&dx_root->dr_num_entries, 1);
1545         ocfs2_journal_dirty(handle, dx_root_bh);
1546
1547 out:
1548         return ret;
1549 }
1550
1551 static void ocfs2_remove_block_from_free_list(struct inode *dir,
1552                                        handle_t *handle,
1553                                        struct ocfs2_dir_lookup_result *lookup)
1554 {
1555         struct ocfs2_dir_block_trailer *trailer, *prev;
1556         struct ocfs2_dx_root_block *dx_root;
1557         struct buffer_head *bh;
1558
1559         trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1560
1561         if (ocfs2_free_list_at_root(lookup)) {
1562                 bh = lookup->dl_dx_root_bh;
1563                 dx_root = (struct ocfs2_dx_root_block *)bh->b_data;
1564                 dx_root->dr_free_blk = trailer->db_free_next;
1565         } else {
1566                 bh = lookup->dl_prev_leaf_bh;
1567                 prev = ocfs2_trailer_from_bh(bh, dir->i_sb);
1568                 prev->db_free_next = trailer->db_free_next;
1569         }
1570
1571         trailer->db_free_rec_len = cpu_to_le16(0);
1572         trailer->db_free_next = cpu_to_le64(0);
1573
1574         ocfs2_journal_dirty(handle, bh);
1575         ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1576 }
1577
1578 /*
1579  * This expects that a journal write has been reserved on
1580  * lookup->dl_prev_leaf_bh or lookup->dl_dx_root_bh
1581  */
1582 static void ocfs2_recalc_free_list(struct inode *dir, handle_t *handle,
1583                                    struct ocfs2_dir_lookup_result *lookup)
1584 {
1585         int max_rec_len;
1586         struct ocfs2_dir_block_trailer *trailer;
1587
1588         /* Walk dl_leaf_bh to figure out what the new free rec_len is. */
1589         max_rec_len = ocfs2_find_max_rec_len(dir->i_sb, lookup->dl_leaf_bh);
1590         if (max_rec_len) {
1591                 /*
1592                  * There's still room in this block, so no need to remove it
1593                  * from the free list. In this case, we just want to update
1594                  * the rec len accounting.
1595                  */
1596                 trailer = ocfs2_trailer_from_bh(lookup->dl_leaf_bh, dir->i_sb);
1597                 trailer->db_free_rec_len = cpu_to_le16(max_rec_len);
1598                 ocfs2_journal_dirty(handle, lookup->dl_leaf_bh);
1599         } else {
1600                 ocfs2_remove_block_from_free_list(dir, handle, lookup);
1601         }
1602 }
1603
1604 /* we don't always have a dentry for what we want to add, so people
1605  * like orphan dir can call this instead.
1606  *
1607  * The lookup context must have been filled from
1608  * ocfs2_prepare_dir_for_insert.
1609  */
1610 int __ocfs2_add_entry(handle_t *handle,
1611                       struct inode *dir,
1612                       const char *name, int namelen,
1613                       struct inode *inode, u64 blkno,
1614                       struct buffer_head *parent_fe_bh,
1615                       struct ocfs2_dir_lookup_result *lookup)
1616 {
1617         unsigned long offset;
1618         unsigned short rec_len;
1619         struct ocfs2_dir_entry *de, *de1;
1620         struct ocfs2_dinode *di = (struct ocfs2_dinode *)parent_fe_bh->b_data;
1621         struct super_block *sb = dir->i_sb;
1622         int retval, status;
1623         unsigned int size = sb->s_blocksize;
1624         struct buffer_head *insert_bh = lookup->dl_leaf_bh;
1625         char *data_start = insert_bh->b_data;
1626
1627         if (!namelen)
1628                 return -EINVAL;
1629
1630         if (ocfs2_dir_indexed(dir)) {
1631                 struct buffer_head *bh;
1632
1633                 /*
1634                  * An indexed dir may require that we update the free space
1635                  * list. Reserve a write to the previous node in the list so
1636                  * that we don't fail later.
1637                  *
1638                  * XXX: This can be either a dx_root_block, or an unindexed
1639                  * directory tree leaf block.
1640                  */
1641                 if (ocfs2_free_list_at_root(lookup)) {
1642                         bh = lookup->dl_dx_root_bh;
1643                         retval = ocfs2_journal_access_dr(handle,
1644                                                  INODE_CACHE(dir), bh,
1645                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1646                 } else {
1647                         bh = lookup->dl_prev_leaf_bh;
1648                         retval = ocfs2_journal_access_db(handle,
1649                                                  INODE_CACHE(dir), bh,
1650                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1651                 }
1652                 if (retval) {
1653                         mlog_errno(retval);
1654                         return retval;
1655                 }
1656         } else if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1657                 data_start = di->id2.i_data.id_data;
1658                 size = i_size_read(dir);
1659
1660                 BUG_ON(insert_bh != parent_fe_bh);
1661         }
1662
1663         rec_len = OCFS2_DIR_REC_LEN(namelen);
1664         offset = 0;
1665         de = (struct ocfs2_dir_entry *) data_start;
1666         while (1) {
1667                 BUG_ON((char *)de >= (size + data_start));
1668
1669                 /* These checks should've already been passed by the
1670                  * prepare function, but I guess we can leave them
1671                  * here anyway. */
1672                 if (!ocfs2_check_dir_entry(dir, de, insert_bh, offset)) {
1673                         retval = -ENOENT;
1674                         goto bail;
1675                 }
1676                 if (ocfs2_match(namelen, name, de)) {
1677                         retval = -EEXIST;
1678                         goto bail;
1679                 }
1680
1681                 /* We're guaranteed that we should have space, so we
1682                  * can't possibly have hit the trailer...right? */
1683                 mlog_bug_on_msg(ocfs2_skip_dir_trailer(dir, de, offset, size),
1684                                 "Hit dir trailer trying to insert %.*s "
1685                                 "(namelen %d) into directory %llu.  "
1686                                 "offset is %lu, trailer offset is %d\n",
1687                                 namelen, name, namelen,
1688                                 (unsigned long long)parent_fe_bh->b_blocknr,
1689                                 offset, ocfs2_dir_trailer_blk_off(dir->i_sb));
1690
1691                 if (ocfs2_dirent_would_fit(de, rec_len)) {
1692                         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
1693                         retval = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
1694                         if (retval < 0) {
1695                                 mlog_errno(retval);
1696                                 goto bail;
1697                         }
1698
1699                         if (insert_bh == parent_fe_bh)
1700                                 status = ocfs2_journal_access_di(handle,
1701                                                                  INODE_CACHE(dir),
1702                                                                  insert_bh,
1703                                                                  OCFS2_JOURNAL_ACCESS_WRITE);
1704                         else {
1705                                 status = ocfs2_journal_access_db(handle,
1706                                                                  INODE_CACHE(dir),
1707                                                                  insert_bh,
1708                                               OCFS2_JOURNAL_ACCESS_WRITE);
1709
1710                                 if (ocfs2_dir_indexed(dir)) {
1711                                         status = ocfs2_dx_dir_insert(dir,
1712                                                                 handle,
1713                                                                 lookup);
1714                                         if (status) {
1715                                                 mlog_errno(status);
1716                                                 goto bail;
1717                                         }
1718                                 }
1719                         }
1720
1721                         /* By now the buffer is marked for journaling */
1722                         offset += le16_to_cpu(de->rec_len);
1723                         if (le64_to_cpu(de->inode)) {
1724                                 de1 = (struct ocfs2_dir_entry *)((char *) de +
1725                                         OCFS2_DIR_REC_LEN(de->name_len));
1726                                 de1->rec_len =
1727                                         cpu_to_le16(le16_to_cpu(de->rec_len) -
1728                                         OCFS2_DIR_REC_LEN(de->name_len));
1729                                 de->rec_len = cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
1730                                 de = de1;
1731                         }
1732                         de->file_type = OCFS2_FT_UNKNOWN;
1733                         if (blkno) {
1734                                 de->inode = cpu_to_le64(blkno);
1735                                 ocfs2_set_de_type(de, inode->i_mode);
1736                         } else
1737                                 de->inode = 0;
1738                         de->name_len = namelen;
1739                         memcpy(de->name, name, namelen);
1740
1741                         if (ocfs2_dir_indexed(dir))
1742                                 ocfs2_recalc_free_list(dir, handle, lookup);
1743
1744                         dir->i_version++;
1745                         ocfs2_journal_dirty(handle, insert_bh);
1746                         retval = 0;
1747                         goto bail;
1748                 }
1749
1750                 offset += le16_to_cpu(de->rec_len);
1751                 de = (struct ocfs2_dir_entry *) ((char *) de + le16_to_cpu(de->rec_len));
1752         }
1753
1754         /* when you think about it, the assert above should prevent us
1755          * from ever getting here. */
1756         retval = -ENOSPC;
1757 bail:
1758         if (retval)
1759                 mlog_errno(retval);
1760
1761         return retval;
1762 }
1763
1764 static int ocfs2_dir_foreach_blk_id(struct inode *inode,
1765                                     u64 *f_version,
1766                                     loff_t *f_pos, void *priv,
1767                                     filldir_t filldir, int *filldir_err)
1768 {
1769         int ret, i, filldir_ret;
1770         unsigned long offset = *f_pos;
1771         struct buffer_head *di_bh = NULL;
1772         struct ocfs2_dinode *di;
1773         struct ocfs2_inline_data *data;
1774         struct ocfs2_dir_entry *de;
1775
1776         ret = ocfs2_read_inode_block(inode, &di_bh);
1777         if (ret) {
1778                 mlog(ML_ERROR, "Unable to read inode block for dir %llu\n",
1779                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
1780                 goto out;
1781         }
1782
1783         di = (struct ocfs2_dinode *)di_bh->b_data;
1784         data = &di->id2.i_data;
1785
1786         while (*f_pos < i_size_read(inode)) {
1787 revalidate:
1788                 /* If the dir block has changed since the last call to
1789                  * readdir(2), then we might be pointing to an invalid
1790                  * dirent right now.  Scan from the start of the block
1791                  * to make sure. */
1792                 if (*f_version != inode->i_version) {
1793                         for (i = 0; i < i_size_read(inode) && i < offset; ) {
1794                                 de = (struct ocfs2_dir_entry *)
1795                                         (data->id_data + i);
1796                                 /* It's too expensive to do a full
1797                                  * dirent test each time round this
1798                                  * loop, but we do have to test at
1799                                  * least that it is non-zero.  A
1800                                  * failure will be detected in the
1801                                  * dirent test below. */
1802                                 if (le16_to_cpu(de->rec_len) <
1803                                     OCFS2_DIR_REC_LEN(1))
1804                                         break;
1805                                 i += le16_to_cpu(de->rec_len);
1806                         }
1807                         *f_pos = offset = i;
1808                         *f_version = inode->i_version;
1809                 }
1810
1811                 de = (struct ocfs2_dir_entry *) (data->id_data + *f_pos);
1812                 if (!ocfs2_check_dir_entry(inode, de, di_bh, *f_pos)) {
1813                         /* On error, skip the f_pos to the end. */
1814                         *f_pos = i_size_read(inode);
1815                         goto out;
1816                 }
1817                 offset += le16_to_cpu(de->rec_len);
1818                 if (le64_to_cpu(de->inode)) {
1819                         /* We might block in the next section
1820                          * if the data destination is
1821                          * currently swapped out.  So, use a
1822                          * version stamp to detect whether or
1823                          * not the directory has been modified
1824                          * during the copy operation.
1825                          */
1826                         u64 version = *f_version;
1827                         unsigned char d_type = DT_UNKNOWN;
1828
1829                         if (de->file_type < OCFS2_FT_MAX)
1830                                 d_type = ocfs2_filetype_table[de->file_type];
1831
1832                         filldir_ret = filldir(priv, de->name,
1833                                               de->name_len,
1834                                               *f_pos,
1835                                               le64_to_cpu(de->inode),
1836                                               d_type);
1837                         if (filldir_ret) {
1838                                 if (filldir_err)
1839                                         *filldir_err = filldir_ret;
1840                                 break;
1841                         }
1842                         if (version != *f_version)
1843                                 goto revalidate;
1844                 }
1845                 *f_pos += le16_to_cpu(de->rec_len);
1846         }
1847
1848 out:
1849         brelse(di_bh);
1850
1851         return 0;
1852 }
1853
1854 /*
1855  * NOTE: This function can be called against unindexed directories,
1856  * and indexed ones.
1857  */
1858 static int ocfs2_dir_foreach_blk_el(struct inode *inode,
1859                                     u64 *f_version,
1860                                     loff_t *f_pos, void *priv,
1861                                     filldir_t filldir, int *filldir_err)
1862 {
1863         int error = 0;
1864         unsigned long offset, blk, last_ra_blk = 0;
1865         int i, stored;
1866         struct buffer_head * bh, * tmp;
1867         struct ocfs2_dir_entry * de;
1868         struct super_block * sb = inode->i_sb;
1869         unsigned int ra_sectors = 16;
1870
1871         stored = 0;
1872         bh = NULL;
1873
1874         offset = (*f_pos) & (sb->s_blocksize - 1);
1875
1876         while (!error && !stored && *f_pos < i_size_read(inode)) {
1877                 blk = (*f_pos) >> sb->s_blocksize_bits;
1878                 if (ocfs2_read_dir_block(inode, blk, &bh, 0)) {
1879                         /* Skip the corrupt dirblock and keep trying */
1880                         *f_pos += sb->s_blocksize - offset;
1881                         continue;
1882                 }
1883
1884                 /* The idea here is to begin with 8k read-ahead and to stay
1885                  * 4k ahead of our current position.
1886                  *
1887                  * TODO: Use the pagecache for this. We just need to
1888                  * make sure it's cluster-safe... */
1889                 if (!last_ra_blk
1890                     || (((last_ra_blk - blk) << 9) <= (ra_sectors / 2))) {
1891                         for (i = ra_sectors >> (sb->s_blocksize_bits - 9);
1892                              i > 0; i--) {
1893                                 tmp = NULL;
1894                                 if (!ocfs2_read_dir_block(inode, ++blk, &tmp,
1895                                                           OCFS2_BH_READAHEAD))
1896                                         brelse(tmp);
1897                         }
1898                         last_ra_blk = blk;
1899                         ra_sectors = 8;
1900                 }
1901
1902 revalidate:
1903                 /* If the dir block has changed since the last call to
1904                  * readdir(2), then we might be pointing to an invalid
1905                  * dirent right now.  Scan from the start of the block
1906                  * to make sure. */
1907                 if (*f_version != inode->i_version) {
1908                         for (i = 0; i < sb->s_blocksize && i < offset; ) {
1909                                 de = (struct ocfs2_dir_entry *) (bh->b_data + i);
1910                                 /* It's too expensive to do a full
1911                                  * dirent test each time round this
1912                                  * loop, but we do have to test at
1913                                  * least that it is non-zero.  A
1914                                  * failure will be detected in the
1915                                  * dirent test below. */
1916                                 if (le16_to_cpu(de->rec_len) <
1917                                     OCFS2_DIR_REC_LEN(1))
1918                                         break;
1919                                 i += le16_to_cpu(de->rec_len);
1920                         }
1921                         offset = i;
1922                         *f_pos = ((*f_pos) & ~(sb->s_blocksize - 1))
1923                                 | offset;
1924                         *f_version = inode->i_version;
1925                 }
1926
1927                 while (!error && *f_pos < i_size_read(inode)
1928                        && offset < sb->s_blocksize) {
1929                         de = (struct ocfs2_dir_entry *) (bh->b_data + offset);
1930                         if (!ocfs2_check_dir_entry(inode, de, bh, offset)) {
1931                                 /* On error, skip the f_pos to the
1932                                    next block. */
1933                                 *f_pos = ((*f_pos) | (sb->s_blocksize - 1)) + 1;
1934                                 brelse(bh);
1935                                 goto out;
1936                         }
1937                         offset += le16_to_cpu(de->rec_len);
1938                         if (le64_to_cpu(de->inode)) {
1939                                 /* We might block in the next section
1940                                  * if the data destination is
1941                                  * currently swapped out.  So, use a
1942                                  * version stamp to detect whether or
1943                                  * not the directory has been modified
1944                                  * during the copy operation.
1945                                  */
1946                                 unsigned long version = *f_version;
1947                                 unsigned char d_type = DT_UNKNOWN;
1948
1949                                 if (de->file_type < OCFS2_FT_MAX)
1950                                         d_type = ocfs2_filetype_table[de->file_type];
1951                                 error = filldir(priv, de->name,
1952                                                 de->name_len,
1953                                                 *f_pos,
1954                                                 le64_to_cpu(de->inode),
1955                                                 d_type);
1956                                 if (error) {
1957                                         if (filldir_err)
1958                                                 *filldir_err = error;
1959                                         break;
1960                                 }
1961                                 if (version != *f_version)
1962                                         goto revalidate;
1963                                 stored ++;
1964                         }
1965                         *f_pos += le16_to_cpu(de->rec_len);
1966                 }
1967                 offset = 0;
1968                 brelse(bh);
1969                 bh = NULL;
1970         }
1971
1972         stored = 0;
1973 out:
1974         return stored;
1975 }
1976
1977 static int ocfs2_dir_foreach_blk(struct inode *inode, u64 *f_version,
1978                                  loff_t *f_pos, void *priv, filldir_t filldir,
1979                                  int *filldir_err)
1980 {
1981         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1982                 return ocfs2_dir_foreach_blk_id(inode, f_version, f_pos, priv,
1983                                                 filldir, filldir_err);
1984
1985         return ocfs2_dir_foreach_blk_el(inode, f_version, f_pos, priv, filldir,
1986                                         filldir_err);
1987 }
1988
1989 /*
1990  * This is intended to be called from inside other kernel functions,
1991  * so we fake some arguments.
1992  */
1993 int ocfs2_dir_foreach(struct inode *inode, loff_t *f_pos, void *priv,
1994                       filldir_t filldir)
1995 {
1996         int ret = 0, filldir_err = 0;
1997         u64 version = inode->i_version;
1998
1999         while (*f_pos < i_size_read(inode)) {
2000                 ret = ocfs2_dir_foreach_blk(inode, &version, f_pos, priv,
2001                                             filldir, &filldir_err);
2002                 if (ret || filldir_err)
2003                         break;
2004         }
2005
2006         if (ret > 0)
2007                 ret = -EIO;
2008
2009         return 0;
2010 }
2011
2012 /*
2013  * ocfs2_readdir()
2014  *
2015  */
2016 int ocfs2_readdir(struct file * filp, void * dirent, filldir_t filldir)
2017 {
2018         int error = 0;
2019         struct inode *inode = filp->f_path.dentry->d_inode;
2020         int lock_level = 0;
2021
2022         trace_ocfs2_readdir((unsigned long long)OCFS2_I(inode)->ip_blkno);
2023
2024         error = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2025         if (lock_level && error >= 0) {
2026                 /* We release EX lock which used to update atime
2027                  * and get PR lock again to reduce contention
2028                  * on commonly accessed directories. */
2029                 ocfs2_inode_unlock(inode, 1);
2030                 lock_level = 0;
2031                 error = ocfs2_inode_lock(inode, NULL, 0);
2032         }
2033         if (error < 0) {
2034                 if (error != -ENOENT)
2035                         mlog_errno(error);
2036                 /* we haven't got any yet, so propagate the error. */
2037                 goto bail_nolock;
2038         }
2039
2040         error = ocfs2_dir_foreach_blk(inode, &filp->f_version, &filp->f_pos,
2041                                       dirent, filldir, NULL);
2042
2043         ocfs2_inode_unlock(inode, lock_level);
2044         if (error)
2045                 mlog_errno(error);
2046
2047 bail_nolock:
2048
2049         return error;
2050 }
2051
2052 /*
2053  * NOTE: this should always be called with parent dir i_mutex taken.
2054  */
2055 int ocfs2_find_files_on_disk(const char *name,
2056                              int namelen,
2057                              u64 *blkno,
2058                              struct inode *inode,
2059                              struct ocfs2_dir_lookup_result *lookup)
2060 {
2061         int status = -ENOENT;
2062
2063         trace_ocfs2_find_files_on_disk(namelen, name, blkno,
2064                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
2065
2066         status = ocfs2_find_entry(name, namelen, inode, lookup);
2067         if (status)
2068                 goto leave;
2069
2070         *blkno = le64_to_cpu(lookup->dl_entry->inode);
2071
2072         status = 0;
2073 leave:
2074
2075         return status;
2076 }
2077
2078 /*
2079  * Convenience function for callers which just want the block number
2080  * mapped to a name and don't require the full dirent info, etc.
2081  */
2082 int ocfs2_lookup_ino_from_name(struct inode *dir, const char *name,
2083                                int namelen, u64 *blkno)
2084 {
2085         int ret;
2086         struct ocfs2_dir_lookup_result lookup = { NULL, };
2087
2088         ret = ocfs2_find_files_on_disk(name, namelen, blkno, dir, &lookup);
2089         ocfs2_free_dir_lookup_result(&lookup);
2090
2091         return ret;
2092 }
2093
2094 /* Check for a name within a directory.
2095  *
2096  * Return 0 if the name does not exist
2097  * Return -EEXIST if the directory contains the name
2098  *
2099  * Callers should have i_mutex + a cluster lock on dir
2100  */
2101 int ocfs2_check_dir_for_entry(struct inode *dir,
2102                               const char *name,
2103                               int namelen)
2104 {
2105         int ret;
2106         struct ocfs2_dir_lookup_result lookup = { NULL, };
2107
2108         trace_ocfs2_check_dir_for_entry(
2109                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen, name);
2110
2111         ret = -EEXIST;
2112         if (ocfs2_find_entry(name, namelen, dir, &lookup) == 0)
2113                 goto bail;
2114
2115         ret = 0;
2116 bail:
2117         ocfs2_free_dir_lookup_result(&lookup);
2118
2119         if (ret)
2120                 mlog_errno(ret);
2121         return ret;
2122 }
2123
2124 struct ocfs2_empty_dir_priv {
2125         unsigned seen_dot;
2126         unsigned seen_dot_dot;
2127         unsigned seen_other;
2128         unsigned dx_dir;
2129 };
2130 static int ocfs2_empty_dir_filldir(void *priv, const char *name, int name_len,
2131                                    loff_t pos, u64 ino, unsigned type)
2132 {
2133         struct ocfs2_empty_dir_priv *p = priv;
2134
2135         /*
2136          * Check the positions of "." and ".." records to be sure
2137          * they're in the correct place.
2138          *
2139          * Indexed directories don't need to proceed past the first
2140          * two entries, so we end the scan after seeing '..'. Despite
2141          * that, we allow the scan to proceed In the event that we
2142          * have a corrupted indexed directory (no dot or dot dot
2143          * entries). This allows us to double check for existing
2144          * entries which might not have been found in the index.
2145          */
2146         if (name_len == 1 && !strncmp(".", name, 1) && pos == 0) {
2147                 p->seen_dot = 1;
2148                 return 0;
2149         }
2150
2151         if (name_len == 2 && !strncmp("..", name, 2) &&
2152             pos == OCFS2_DIR_REC_LEN(1)) {
2153                 p->seen_dot_dot = 1;
2154
2155                 if (p->dx_dir && p->seen_dot)
2156                         return 1;
2157
2158                 return 0;
2159         }
2160
2161         p->seen_other = 1;
2162         return 1;
2163 }
2164
2165 static int ocfs2_empty_dir_dx(struct inode *inode,
2166                               struct ocfs2_empty_dir_priv *priv)
2167 {
2168         int ret;
2169         struct buffer_head *di_bh = NULL;
2170         struct buffer_head *dx_root_bh = NULL;
2171         struct ocfs2_dinode *di;
2172         struct ocfs2_dx_root_block *dx_root;
2173
2174         priv->dx_dir = 1;
2175
2176         ret = ocfs2_read_inode_block(inode, &di_bh);
2177         if (ret) {
2178                 mlog_errno(ret);
2179                 goto out;
2180         }
2181         di = (struct ocfs2_dinode *)di_bh->b_data;
2182
2183         ret = ocfs2_read_dx_root(inode, di, &dx_root_bh);
2184         if (ret) {
2185                 mlog_errno(ret);
2186                 goto out;
2187         }
2188         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2189
2190         if (le32_to_cpu(dx_root->dr_num_entries) != 2)
2191                 priv->seen_other = 1;
2192
2193 out:
2194         brelse(di_bh);
2195         brelse(dx_root_bh);
2196         return ret;
2197 }
2198
2199 /*
2200  * routine to check that the specified directory is empty (for rmdir)
2201  *
2202  * Returns 1 if dir is empty, zero otherwise.
2203  *
2204  * XXX: This is a performance problem for unindexed directories.
2205  */
2206 int ocfs2_empty_dir(struct inode *inode)
2207 {
2208         int ret;
2209         loff_t start = 0;
2210         struct ocfs2_empty_dir_priv priv;
2211
2212         memset(&priv, 0, sizeof(priv));
2213
2214         if (ocfs2_dir_indexed(inode)) {
2215                 ret = ocfs2_empty_dir_dx(inode, &priv);
2216                 if (ret)
2217                         mlog_errno(ret);
2218                 /*
2219                  * We still run ocfs2_dir_foreach to get the checks
2220                  * for "." and "..".
2221                  */
2222         }
2223
2224         ret = ocfs2_dir_foreach(inode, &start, &priv, ocfs2_empty_dir_filldir);
2225         if (ret)
2226                 mlog_errno(ret);
2227
2228         if (!priv.seen_dot || !priv.seen_dot_dot) {
2229                 mlog(ML_ERROR, "bad directory (dir #%llu) - no `.' or `..'\n",
2230                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
2231                 /*
2232                  * XXX: Is it really safe to allow an unlink to continue?
2233                  */
2234                 return 1;
2235         }
2236
2237         return !priv.seen_other;
2238 }
2239
2240 /*
2241  * Fills "." and ".." dirents in a new directory block. Returns dirent for
2242  * "..", which might be used during creation of a directory with a trailing
2243  * header. It is otherwise safe to ignore the return code.
2244  */
2245 static struct ocfs2_dir_entry *ocfs2_fill_initial_dirents(struct inode *inode,
2246                                                           struct inode *parent,
2247                                                           char *start,
2248                                                           unsigned int size)
2249 {
2250         struct ocfs2_dir_entry *de = (struct ocfs2_dir_entry *)start;
2251
2252         de->inode = cpu_to_le64(OCFS2_I(inode)->ip_blkno);
2253         de->name_len = 1;
2254         de->rec_len =
2255                 cpu_to_le16(OCFS2_DIR_REC_LEN(de->name_len));
2256         strcpy(de->name, ".");
2257         ocfs2_set_de_type(de, S_IFDIR);
2258
2259         de = (struct ocfs2_dir_entry *) ((char *)de + le16_to_cpu(de->rec_len));
2260         de->inode = cpu_to_le64(OCFS2_I(parent)->ip_blkno);
2261         de->rec_len = cpu_to_le16(size - OCFS2_DIR_REC_LEN(1));
2262         de->name_len = 2;
2263         strcpy(de->name, "..");
2264         ocfs2_set_de_type(de, S_IFDIR);
2265
2266         return de;
2267 }
2268
2269 /*
2270  * This works together with code in ocfs2_mknod_locked() which sets
2271  * the inline-data flag and initializes the inline-data section.
2272  */
2273 static int ocfs2_fill_new_dir_id(struct ocfs2_super *osb,
2274                                  handle_t *handle,
2275                                  struct inode *parent,
2276                                  struct inode *inode,
2277                                  struct buffer_head *di_bh)
2278 {
2279         int ret;
2280         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2281         struct ocfs2_inline_data *data = &di->id2.i_data;
2282         unsigned int size = le16_to_cpu(data->id_count);
2283
2284         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2285                                       OCFS2_JOURNAL_ACCESS_WRITE);
2286         if (ret) {
2287                 mlog_errno(ret);
2288                 goto out;
2289         }
2290
2291         ocfs2_fill_initial_dirents(inode, parent, data->id_data, size);
2292         ocfs2_journal_dirty(handle, di_bh);
2293
2294         i_size_write(inode, size);
2295         inode->i_nlink = 2;
2296         inode->i_blocks = ocfs2_inode_sector_count(inode);
2297
2298         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
2299         if (ret < 0)
2300                 mlog_errno(ret);
2301
2302 out:
2303         return ret;
2304 }
2305
2306 static int ocfs2_fill_new_dir_el(struct ocfs2_super *osb,
2307                                  handle_t *handle,
2308                                  struct inode *parent,
2309                                  struct inode *inode,
2310                                  struct buffer_head *fe_bh,
2311                                  struct ocfs2_alloc_context *data_ac,
2312                                  struct buffer_head **ret_new_bh)
2313 {
2314         int status;
2315         unsigned int size = osb->sb->s_blocksize;
2316         struct buffer_head *new_bh = NULL;
2317         struct ocfs2_dir_entry *de;
2318
2319         if (ocfs2_new_dir_wants_trailer(inode))
2320                 size = ocfs2_dir_trailer_blk_off(parent->i_sb);
2321
2322         status = ocfs2_do_extend_dir(osb->sb, handle, inode, fe_bh,
2323                                      data_ac, NULL, &new_bh);
2324         if (status < 0) {
2325                 mlog_errno(status);
2326                 goto bail;
2327         }
2328
2329         ocfs2_set_new_buffer_uptodate(INODE_CACHE(inode), new_bh);
2330
2331         status = ocfs2_journal_access_db(handle, INODE_CACHE(inode), new_bh,
2332                                          OCFS2_JOURNAL_ACCESS_CREATE);
2333         if (status < 0) {
2334                 mlog_errno(status);
2335                 goto bail;
2336         }
2337         memset(new_bh->b_data, 0, osb->sb->s_blocksize);
2338
2339         de = ocfs2_fill_initial_dirents(inode, parent, new_bh->b_data, size);
2340         if (ocfs2_new_dir_wants_trailer(inode)) {
2341                 int size = le16_to_cpu(de->rec_len);
2342
2343                 /*
2344                  * Figure out the size of the hole left over after
2345                  * insertion of '.' and '..'. The trailer wants this
2346                  * information.
2347                  */
2348                 size -= OCFS2_DIR_REC_LEN(2);
2349                 size -= sizeof(struct ocfs2_dir_block_trailer);
2350
2351                 ocfs2_init_dir_trailer(inode, new_bh, size);
2352         }
2353
2354         ocfs2_journal_dirty(handle, new_bh);
2355
2356         i_size_write(inode, inode->i_sb->s_blocksize);
2357         inode->i_nlink = 2;
2358         inode->i_blocks = ocfs2_inode_sector_count(inode);
2359         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
2360         if (status < 0) {
2361                 mlog_errno(status);
2362                 goto bail;
2363         }
2364
2365         status = 0;
2366         if (ret_new_bh) {
2367                 *ret_new_bh = new_bh;
2368                 new_bh = NULL;
2369         }
2370 bail:
2371         brelse(new_bh);
2372
2373         return status;
2374 }
2375
2376 static int ocfs2_dx_dir_attach_index(struct ocfs2_super *osb,
2377                                      handle_t *handle, struct inode *dir,
2378                                      struct buffer_head *di_bh,
2379                                      struct buffer_head *dirdata_bh,
2380                                      struct ocfs2_alloc_context *meta_ac,
2381                                      int dx_inline, u32 num_entries,
2382                                      struct buffer_head **ret_dx_root_bh)
2383 {
2384         int ret;
2385         struct ocfs2_dinode *di = (struct ocfs2_dinode *) di_bh->b_data;
2386         u16 dr_suballoc_bit;
2387         u64 suballoc_loc, dr_blkno;
2388         unsigned int num_bits;
2389         struct buffer_head *dx_root_bh = NULL;
2390         struct ocfs2_dx_root_block *dx_root;
2391         struct ocfs2_dir_block_trailer *trailer =
2392                 ocfs2_trailer_from_bh(dirdata_bh, dir->i_sb);
2393
2394         ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc,
2395                                    &dr_suballoc_bit, &num_bits, &dr_blkno);
2396         if (ret) {
2397                 mlog_errno(ret);
2398                 goto out;
2399         }
2400
2401         trace_ocfs2_dx_dir_attach_index(
2402                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2403                                 (unsigned long long)dr_blkno);
2404
2405         dx_root_bh = sb_getblk(osb->sb, dr_blkno);
2406         if (dx_root_bh == NULL) {
2407                 ret = -EIO;
2408                 goto out;
2409         }
2410         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dx_root_bh);
2411
2412         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
2413                                       OCFS2_JOURNAL_ACCESS_CREATE);
2414         if (ret < 0) {
2415                 mlog_errno(ret);
2416                 goto out;
2417         }
2418
2419         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2420         memset(dx_root, 0, osb->sb->s_blocksize);
2421         strcpy(dx_root->dr_signature, OCFS2_DX_ROOT_SIGNATURE);
2422         dx_root->dr_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot);
2423         dx_root->dr_suballoc_loc = cpu_to_le64(suballoc_loc);
2424         dx_root->dr_suballoc_bit = cpu_to_le16(dr_suballoc_bit);
2425         dx_root->dr_fs_generation = cpu_to_le32(osb->fs_generation);
2426         dx_root->dr_blkno = cpu_to_le64(dr_blkno);
2427         dx_root->dr_dir_blkno = cpu_to_le64(OCFS2_I(dir)->ip_blkno);
2428         dx_root->dr_num_entries = cpu_to_le32(num_entries);
2429         if (le16_to_cpu(trailer->db_free_rec_len))
2430                 dx_root->dr_free_blk = cpu_to_le64(dirdata_bh->b_blocknr);
2431         else
2432                 dx_root->dr_free_blk = cpu_to_le64(0);
2433
2434         if (dx_inline) {
2435                 dx_root->dr_flags |= OCFS2_DX_FLAG_INLINE;
2436                 dx_root->dr_entries.de_count =
2437                         cpu_to_le16(ocfs2_dx_entries_per_root(osb->sb));
2438         } else {
2439                 dx_root->dr_list.l_count =
2440                         cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
2441         }
2442         ocfs2_journal_dirty(handle, dx_root_bh);
2443
2444         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
2445                                       OCFS2_JOURNAL_ACCESS_CREATE);
2446         if (ret) {
2447                 mlog_errno(ret);
2448                 goto out;
2449         }
2450
2451         di->i_dx_root = cpu_to_le64(dr_blkno);
2452
2453         spin_lock(&OCFS2_I(dir)->ip_lock);
2454         OCFS2_I(dir)->ip_dyn_features |= OCFS2_INDEXED_DIR_FL;
2455         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
2456         spin_unlock(&OCFS2_I(dir)->ip_lock);
2457
2458         ocfs2_journal_dirty(handle, di_bh);
2459
2460         *ret_dx_root_bh = dx_root_bh;
2461         dx_root_bh = NULL;
2462
2463 out:
2464         brelse(dx_root_bh);
2465         return ret;
2466 }
2467
2468 static int ocfs2_dx_dir_format_cluster(struct ocfs2_super *osb,
2469                                        handle_t *handle, struct inode *dir,
2470                                        struct buffer_head **dx_leaves,
2471                                        int num_dx_leaves, u64 start_blk)
2472 {
2473         int ret, i;
2474         struct ocfs2_dx_leaf *dx_leaf;
2475         struct buffer_head *bh;
2476
2477         for (i = 0; i < num_dx_leaves; i++) {
2478                 bh = sb_getblk(osb->sb, start_blk + i);
2479                 if (bh == NULL) {
2480                         ret = -EIO;
2481                         goto out;
2482                 }
2483                 dx_leaves[i] = bh;
2484
2485                 ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), bh);
2486
2487                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), bh,
2488                                               OCFS2_JOURNAL_ACCESS_CREATE);
2489                 if (ret < 0) {
2490                         mlog_errno(ret);
2491                         goto out;
2492                 }
2493
2494                 dx_leaf = (struct ocfs2_dx_leaf *) bh->b_data;
2495
2496                 memset(dx_leaf, 0, osb->sb->s_blocksize);
2497                 strcpy(dx_leaf->dl_signature, OCFS2_DX_LEAF_SIGNATURE);
2498                 dx_leaf->dl_fs_generation = cpu_to_le32(osb->fs_generation);
2499                 dx_leaf->dl_blkno = cpu_to_le64(bh->b_blocknr);
2500                 dx_leaf->dl_list.de_count =
2501                         cpu_to_le16(ocfs2_dx_entries_per_leaf(osb->sb));
2502
2503                 trace_ocfs2_dx_dir_format_cluster(
2504                                 (unsigned long long)OCFS2_I(dir)->ip_blkno,
2505                                 (unsigned long long)bh->b_blocknr,
2506                                 le16_to_cpu(dx_leaf->dl_list.de_count));
2507
2508                 ocfs2_journal_dirty(handle, bh);
2509         }
2510
2511         ret = 0;
2512 out:
2513         return ret;
2514 }
2515
2516 /*
2517  * Allocates and formats a new cluster for use in an indexed dir
2518  * leaf. This version will not do the extent insert, so that it can be
2519  * used by operations which need careful ordering.
2520  */
2521 static int __ocfs2_dx_dir_new_cluster(struct inode *dir,
2522                                       u32 cpos, handle_t *handle,
2523                                       struct ocfs2_alloc_context *data_ac,
2524                                       struct buffer_head **dx_leaves,
2525                                       int num_dx_leaves, u64 *ret_phys_blkno)
2526 {
2527         int ret;
2528         u32 phys, num;
2529         u64 phys_blkno;
2530         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2531
2532         /*
2533          * XXX: For create, this should claim cluster for the index
2534          * *before* the unindexed insert so that we have a better
2535          * chance of contiguousness as the directory grows in number
2536          * of entries.
2537          */
2538         ret = __ocfs2_claim_clusters(handle, data_ac, 1, 1, &phys, &num);
2539         if (ret) {
2540                 mlog_errno(ret);
2541                 goto out;
2542         }
2543
2544         /*
2545          * Format the new cluster first. That way, we're inserting
2546          * valid data.
2547          */
2548         phys_blkno = ocfs2_clusters_to_blocks(osb->sb, phys);
2549         ret = ocfs2_dx_dir_format_cluster(osb, handle, dir, dx_leaves,
2550                                           num_dx_leaves, phys_blkno);
2551         if (ret) {
2552                 mlog_errno(ret);
2553                 goto out;
2554         }
2555
2556         *ret_phys_blkno = phys_blkno;
2557 out:
2558         return ret;
2559 }
2560
2561 static int ocfs2_dx_dir_new_cluster(struct inode *dir,
2562                                     struct ocfs2_extent_tree *et,
2563                                     u32 cpos, handle_t *handle,
2564                                     struct ocfs2_alloc_context *data_ac,
2565                                     struct ocfs2_alloc_context *meta_ac,
2566                                     struct buffer_head **dx_leaves,
2567                                     int num_dx_leaves)
2568 {
2569         int ret;
2570         u64 phys_blkno;
2571
2572         ret = __ocfs2_dx_dir_new_cluster(dir, cpos, handle, data_ac, dx_leaves,
2573                                          num_dx_leaves, &phys_blkno);
2574         if (ret) {
2575                 mlog_errno(ret);
2576                 goto out;
2577         }
2578
2579         ret = ocfs2_insert_extent(handle, et, cpos, phys_blkno, 1, 0,
2580                                   meta_ac);
2581         if (ret)
2582                 mlog_errno(ret);
2583 out:
2584         return ret;
2585 }
2586
2587 static struct buffer_head **ocfs2_dx_dir_kmalloc_leaves(struct super_block *sb,
2588                                                         int *ret_num_leaves)
2589 {
2590         int num_dx_leaves = ocfs2_clusters_to_blocks(sb, 1);
2591         struct buffer_head **dx_leaves;
2592
2593         dx_leaves = kcalloc(num_dx_leaves, sizeof(struct buffer_head *),
2594                             GFP_NOFS);
2595         if (dx_leaves && ret_num_leaves)
2596                 *ret_num_leaves = num_dx_leaves;
2597
2598         return dx_leaves;
2599 }
2600
2601 static int ocfs2_fill_new_dir_dx(struct ocfs2_super *osb,
2602                                  handle_t *handle,
2603                                  struct inode *parent,
2604                                  struct inode *inode,
2605                                  struct buffer_head *di_bh,
2606                                  struct ocfs2_alloc_context *data_ac,
2607                                  struct ocfs2_alloc_context *meta_ac)
2608 {
2609         int ret;
2610         struct buffer_head *leaf_bh = NULL;
2611         struct buffer_head *dx_root_bh = NULL;
2612         struct ocfs2_dx_hinfo hinfo;
2613         struct ocfs2_dx_root_block *dx_root;
2614         struct ocfs2_dx_entry_list *entry_list;
2615
2616         /*
2617          * Our strategy is to create the directory as though it were
2618          * unindexed, then add the index block. This works with very
2619          * little complication since the state of a new directory is a
2620          * very well known quantity.
2621          *
2622          * Essentially, we have two dirents ("." and ".."), in the 1st
2623          * block which need indexing. These are easily inserted into
2624          * the index block.
2625          */
2626
2627         ret = ocfs2_fill_new_dir_el(osb, handle, parent, inode, di_bh,
2628                                     data_ac, &leaf_bh);
2629         if (ret) {
2630                 mlog_errno(ret);
2631                 goto out;
2632         }
2633
2634         ret = ocfs2_dx_dir_attach_index(osb, handle, inode, di_bh, leaf_bh,
2635                                         meta_ac, 1, 2, &dx_root_bh);
2636         if (ret) {
2637                 mlog_errno(ret);
2638                 goto out;
2639         }
2640         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2641         entry_list = &dx_root->dr_entries;
2642
2643         /* Buffer has been journaled for us by ocfs2_dx_dir_attach_index */
2644         ocfs2_dx_dir_name_hash(inode, ".", 1, &hinfo);
2645         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2646
2647         ocfs2_dx_dir_name_hash(inode, "..", 2, &hinfo);
2648         ocfs2_dx_entry_list_insert(entry_list, &hinfo, leaf_bh->b_blocknr);
2649
2650 out:
2651         brelse(dx_root_bh);
2652         brelse(leaf_bh);
2653         return ret;
2654 }
2655
2656 int ocfs2_fill_new_dir(struct ocfs2_super *osb,
2657                        handle_t *handle,
2658                        struct inode *parent,
2659                        struct inode *inode,
2660                        struct buffer_head *fe_bh,
2661                        struct ocfs2_alloc_context *data_ac,
2662                        struct ocfs2_alloc_context *meta_ac)
2663
2664 {
2665         BUG_ON(!ocfs2_supports_inline_data(osb) && data_ac == NULL);
2666
2667         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2668                 return ocfs2_fill_new_dir_id(osb, handle, parent, inode, fe_bh);
2669
2670         if (ocfs2_supports_indexed_dirs(osb))
2671                 return ocfs2_fill_new_dir_dx(osb, handle, parent, inode, fe_bh,
2672                                              data_ac, meta_ac);
2673
2674         return ocfs2_fill_new_dir_el(osb, handle, parent, inode, fe_bh,
2675                                      data_ac, NULL);
2676 }
2677
2678 static int ocfs2_dx_dir_index_block(struct inode *dir,
2679                                     handle_t *handle,
2680                                     struct buffer_head **dx_leaves,
2681                                     int num_dx_leaves,
2682                                     u32 *num_dx_entries,
2683                                     struct buffer_head *dirent_bh)
2684 {
2685         int ret = 0, namelen, i;
2686         char *de_buf, *limit;
2687         struct ocfs2_dir_entry *de;
2688         struct buffer_head *dx_leaf_bh;
2689         struct ocfs2_dx_hinfo hinfo;
2690         u64 dirent_blk = dirent_bh->b_blocknr;
2691
2692         de_buf = dirent_bh->b_data;
2693         limit = de_buf + dir->i_sb->s_blocksize;
2694
2695         while (de_buf < limit) {
2696                 de = (struct ocfs2_dir_entry *)de_buf;
2697
2698                 namelen = de->name_len;
2699                 if (!namelen || !de->inode)
2700                         goto inc;
2701
2702                 ocfs2_dx_dir_name_hash(dir, de->name, namelen, &hinfo);
2703
2704                 i = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb), &hinfo);
2705                 dx_leaf_bh = dx_leaves[i];
2706
2707                 ret = __ocfs2_dx_dir_leaf_insert(dir, handle, &hinfo,
2708                                                  dirent_blk, dx_leaf_bh);
2709                 if (ret) {
2710                         mlog_errno(ret);
2711                         goto out;
2712                 }
2713
2714                 *num_dx_entries = *num_dx_entries + 1;
2715
2716 inc:
2717                 de_buf += le16_to_cpu(de->rec_len);
2718         }
2719
2720 out:
2721         return ret;
2722 }
2723
2724 /*
2725  * XXX: This expects dx_root_bh to already be part of the transaction.
2726  */
2727 static void ocfs2_dx_dir_index_root_block(struct inode *dir,
2728                                          struct buffer_head *dx_root_bh,
2729                                          struct buffer_head *dirent_bh)
2730 {
2731         char *de_buf, *limit;
2732         struct ocfs2_dx_root_block *dx_root;
2733         struct ocfs2_dir_entry *de;
2734         struct ocfs2_dx_hinfo hinfo;
2735         u64 dirent_blk = dirent_bh->b_blocknr;
2736
2737         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
2738
2739         de_buf = dirent_bh->b_data;
2740         limit = de_buf + dir->i_sb->s_blocksize;
2741
2742         while (de_buf < limit) {
2743                 de = (struct ocfs2_dir_entry *)de_buf;
2744
2745                 if (!de->name_len || !de->inode)
2746                         goto inc;
2747
2748                 ocfs2_dx_dir_name_hash(dir, de->name, de->name_len, &hinfo);
2749
2750                 trace_ocfs2_dx_dir_index_root_block(
2751                                 (unsigned long long)dir->i_ino,
2752                                 hinfo.major_hash, hinfo.minor_hash,
2753                                 de->name_len, de->name,
2754                                 le16_to_cpu(dx_root->dr_entries.de_num_used));
2755
2756                 ocfs2_dx_entry_list_insert(&dx_root->dr_entries, &hinfo,
2757                                            dirent_blk);
2758
2759                 le32_add_cpu(&dx_root->dr_num_entries, 1);
2760 inc:
2761                 de_buf += le16_to_cpu(de->rec_len);
2762         }
2763 }
2764
2765 /*
2766  * Count the number of inline directory entries in di_bh and compare
2767  * them against the number of entries we can hold in an inline dx root
2768  * block.
2769  */
2770 static int ocfs2_new_dx_should_be_inline(struct inode *dir,
2771                                          struct buffer_head *di_bh)
2772 {
2773         int dirent_count = 0;
2774         char *de_buf, *limit;
2775         struct ocfs2_dir_entry *de;
2776         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2777
2778         de_buf = di->id2.i_data.id_data;
2779         limit = de_buf + i_size_read(dir);
2780
2781         while (de_buf < limit) {
2782                 de = (struct ocfs2_dir_entry *)de_buf;
2783
2784                 if (de->name_len && de->inode)
2785                         dirent_count++;
2786
2787                 de_buf += le16_to_cpu(de->rec_len);
2788         }
2789
2790         /* We are careful to leave room for one extra record. */
2791         return dirent_count < ocfs2_dx_entries_per_root(dir->i_sb);
2792 }
2793
2794 /*
2795  * Expand rec_len of the rightmost dirent in a directory block so that it
2796  * contains the end of our valid space for dirents. We do this during
2797  * expansion from an inline directory to one with extents. The first dir block
2798  * in that case is taken from the inline data portion of the inode block.
2799  *
2800  * This will also return the largest amount of contiguous space for a dirent
2801  * in the block. That value is *not* necessarily the last dirent, even after
2802  * expansion. The directory indexing code wants this value for free space
2803  * accounting. We do this here since we're already walking the entire dir
2804  * block.
2805  *
2806  * We add the dir trailer if this filesystem wants it.
2807  */
2808 static unsigned int ocfs2_expand_last_dirent(char *start, unsigned int old_size,
2809                                              struct inode *dir)
2810 {
2811         struct super_block *sb = dir->i_sb;
2812         struct ocfs2_dir_entry *de;
2813         struct ocfs2_dir_entry *prev_de;
2814         char *de_buf, *limit;
2815         unsigned int new_size = sb->s_blocksize;
2816         unsigned int bytes, this_hole;
2817         unsigned int largest_hole = 0;
2818
2819         if (ocfs2_new_dir_wants_trailer(dir))
2820                 new_size = ocfs2_dir_trailer_blk_off(sb);
2821
2822         bytes = new_size - old_size;
2823
2824         limit = start + old_size;
2825         de_buf = start;
2826         de = (struct ocfs2_dir_entry *)de_buf;
2827         do {
2828                 this_hole = ocfs2_figure_dirent_hole(de);
2829                 if (this_hole > largest_hole)
2830                         largest_hole = this_hole;
2831
2832                 prev_de = de;
2833                 de_buf += le16_to_cpu(de->rec_len);
2834                 de = (struct ocfs2_dir_entry *)de_buf;
2835         } while (de_buf < limit);
2836
2837         le16_add_cpu(&prev_de->rec_len, bytes);
2838
2839         /* We need to double check this after modification of the final
2840          * dirent. */
2841         this_hole = ocfs2_figure_dirent_hole(prev_de);
2842         if (this_hole > largest_hole)
2843                 largest_hole = this_hole;
2844
2845         if (largest_hole >= OCFS2_DIR_MIN_REC_LEN)
2846                 return largest_hole;
2847         return 0;
2848 }
2849
2850 /*
2851  * We allocate enough clusters to fulfill "blocks_wanted", but set
2852  * i_size to exactly one block. Ocfs2_extend_dir() will handle the
2853  * rest automatically for us.
2854  *
2855  * *first_block_bh is a pointer to the 1st data block allocated to the
2856  *  directory.
2857  */
2858 static int ocfs2_expand_inline_dir(struct inode *dir, struct buffer_head *di_bh,
2859                                    unsigned int blocks_wanted,
2860                                    struct ocfs2_dir_lookup_result *lookup,
2861                                    struct buffer_head **first_block_bh)
2862 {
2863         u32 alloc, dx_alloc, bit_off, len, num_dx_entries = 0;
2864         struct super_block *sb = dir->i_sb;
2865         int ret, i, num_dx_leaves = 0, dx_inline = 0,
2866                 credits = ocfs2_inline_to_extents_credits(sb);
2867         u64 dx_insert_blkno, blkno,
2868                 bytes = blocks_wanted << sb->s_blocksize_bits;
2869         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
2870         struct ocfs2_inode_info *oi = OCFS2_I(dir);
2871         struct ocfs2_alloc_context *data_ac;
2872         struct ocfs2_alloc_context *meta_ac = NULL;
2873         struct buffer_head *dirdata_bh = NULL;
2874         struct buffer_head *dx_root_bh = NULL;
2875         struct buffer_head **dx_leaves = NULL;
2876         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
2877         handle_t *handle;
2878         struct ocfs2_extent_tree et;
2879         struct ocfs2_extent_tree dx_et;
2880         int did_quota = 0, bytes_allocated = 0;
2881
2882         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir), di_bh);
2883
2884         alloc = ocfs2_clusters_for_bytes(sb, bytes);
2885         dx_alloc = 0;
2886
2887         down_write(&oi->ip_alloc_sem);
2888
2889         if (ocfs2_supports_indexed_dirs(osb)) {
2890                 credits += ocfs2_add_dir_index_credits(sb);
2891
2892                 dx_inline = ocfs2_new_dx_should_be_inline(dir, di_bh);
2893                 if (!dx_inline) {
2894                         /* Add one more cluster for an index leaf */
2895                         dx_alloc++;
2896                         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(sb,
2897                                                                 &num_dx_leaves);
2898                         if (!dx_leaves) {
2899                                 ret = -ENOMEM;
2900                                 mlog_errno(ret);
2901                                 goto out;
2902                         }
2903                 }
2904
2905                 /* This gets us the dx_root */
2906                 ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac);
2907                 if (ret) {
2908                         mlog_errno(ret);
2909                         goto out;
2910                 }
2911         }
2912
2913         /*
2914          * We should never need more than 2 clusters for the unindexed
2915          * tree - maximum dirent size is far less than one block. In
2916          * fact, the only time we'd need more than one cluster is if
2917          * blocksize == clustersize and the dirent won't fit in the
2918          * extra space that the expansion to a single block gives. As
2919          * of today, that only happens on 4k/4k file systems.
2920          */
2921         BUG_ON(alloc > 2);
2922
2923         ret = ocfs2_reserve_clusters(osb, alloc + dx_alloc, &data_ac);
2924         if (ret) {
2925                 mlog_errno(ret);
2926                 goto out;
2927         }
2928
2929         /*
2930          * Prepare for worst case allocation scenario of two separate
2931          * extents in the unindexed tree.
2932          */
2933         if (alloc == 2)
2934                 credits += OCFS2_SUBALLOC_ALLOC;
2935
2936         handle = ocfs2_start_trans(osb, credits);
2937         if (IS_ERR(handle)) {
2938                 ret = PTR_ERR(handle);
2939                 mlog_errno(ret);
2940                 goto out;
2941         }
2942
2943         ret = dquot_alloc_space_nodirty(dir,
2944                 ocfs2_clusters_to_bytes(osb->sb, alloc + dx_alloc));
2945         if (ret)
2946                 goto out_commit;
2947         did_quota = 1;
2948
2949         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
2950                 /*
2951                  * Allocate our index cluster first, to maximize the
2952                  * possibility that unindexed leaves grow
2953                  * contiguously.
2954                  */
2955                 ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac,
2956                                                  dx_leaves, num_dx_leaves,
2957                                                  &dx_insert_blkno);
2958                 if (ret) {
2959                         mlog_errno(ret);
2960                         goto out_commit;
2961                 }
2962                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2963         }
2964
2965         /*
2966          * Try to claim as many clusters as the bitmap can give though
2967          * if we only get one now, that's enough to continue. The rest
2968          * will be claimed after the conversion to extents.
2969          */
2970         if (ocfs2_dir_resv_allowed(osb))
2971                 data_ac->ac_resv = &oi->ip_la_data_resv;
2972         ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off, &len);
2973         if (ret) {
2974                 mlog_errno(ret);
2975                 goto out_commit;
2976         }
2977         bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
2978
2979         /*
2980          * Operations are carefully ordered so that we set up the new
2981          * data block first. The conversion from inline data to
2982          * extents follows.
2983          */
2984         blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
2985         dirdata_bh = sb_getblk(sb, blkno);
2986         if (!dirdata_bh) {
2987                 ret = -EIO;
2988                 mlog_errno(ret);
2989                 goto out_commit;
2990         }
2991
2992         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), dirdata_bh);
2993
2994         ret = ocfs2_journal_access_db(handle, INODE_CACHE(dir), dirdata_bh,
2995                                       OCFS2_JOURNAL_ACCESS_CREATE);
2996         if (ret) {
2997                 mlog_errno(ret);
2998                 goto out_commit;
2999         }
3000
3001         memcpy(dirdata_bh->b_data, di->id2.i_data.id_data, i_size_read(dir));
3002         memset(dirdata_bh->b_data + i_size_read(dir), 0,
3003                sb->s_blocksize - i_size_read(dir));
3004         i = ocfs2_expand_last_dirent(dirdata_bh->b_data, i_size_read(dir), dir);
3005         if (ocfs2_new_dir_wants_trailer(dir)) {
3006                 /*
3007                  * Prepare the dir trailer up front. It will otherwise look
3008                  * like a valid dirent. Even if inserting the index fails
3009                  * (unlikely), then all we'll have done is given first dir
3010                  * block a small amount of fragmentation.
3011                  */
3012                 ocfs2_init_dir_trailer(dir, dirdata_bh, i);
3013         }
3014
3015         ocfs2_journal_dirty(handle, dirdata_bh);
3016
3017         if (ocfs2_supports_indexed_dirs(osb) && !dx_inline) {
3018                 /*
3019                  * Dx dirs with an external cluster need to do this up
3020                  * front. Inline dx root's get handled later, after
3021                  * we've allocated our root block. We get passed back
3022                  * a total number of items so that dr_num_entries can
3023                  * be correctly set once the dx_root has been
3024                  * allocated.
3025                  */
3026                 ret = ocfs2_dx_dir_index_block(dir, handle, dx_leaves,
3027                                                num_dx_leaves, &num_dx_entries,
3028                                                dirdata_bh);
3029                 if (ret) {
3030                         mlog_errno(ret);
3031                         goto out_commit;
3032                 }
3033         }
3034
3035         /*
3036          * Set extent, i_size, etc on the directory. After this, the
3037          * inode should contain the same exact dirents as before and
3038          * be fully accessible from system calls.
3039          *
3040          * We let the later dirent insert modify c/mtime - to the user
3041          * the data hasn't changed.
3042          */
3043         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
3044                                       OCFS2_JOURNAL_ACCESS_CREATE);
3045         if (ret) {
3046                 mlog_errno(ret);
3047                 goto out_commit;
3048         }
3049
3050         spin_lock(&oi->ip_lock);
3051         oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL;
3052         di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features);
3053         spin_unlock(&oi->ip_lock);
3054
3055         ocfs2_dinode_new_extent_list(dir, di);
3056
3057         i_size_write(dir, sb->s_blocksize);
3058         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3059
3060         di->i_size = cpu_to_le64(sb->s_blocksize);
3061         di->i_ctime = di->i_mtime = cpu_to_le64(dir->i_ctime.tv_sec);
3062         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(dir->i_ctime.tv_nsec);
3063
3064         /*
3065          * This should never fail as our extent list is empty and all
3066          * related blocks have been journaled already.
3067          */
3068         ret = ocfs2_insert_extent(handle, &et, 0, blkno, len,
3069                                   0, NULL);
3070         if (ret) {
3071                 mlog_errno(ret);
3072                 goto out_commit;
3073         }
3074
3075         /*
3076          * Set i_blocks after the extent insert for the most up to
3077          * date ip_clusters value.
3078          */
3079         dir->i_blocks = ocfs2_inode_sector_count(dir);
3080
3081         ocfs2_journal_dirty(handle, di_bh);
3082
3083         if (ocfs2_supports_indexed_dirs(osb)) {
3084                 ret = ocfs2_dx_dir_attach_index(osb, handle, dir, di_bh,
3085                                                 dirdata_bh, meta_ac, dx_inline,
3086                                                 num_dx_entries, &dx_root_bh);
3087                 if (ret) {
3088                         mlog_errno(ret);
3089                         goto out_commit;
3090                 }
3091
3092                 if (dx_inline) {
3093                         ocfs2_dx_dir_index_root_block(dir, dx_root_bh,
3094                                                       dirdata_bh);
3095                 } else {
3096                         ocfs2_init_dx_root_extent_tree(&dx_et,
3097                                                        INODE_CACHE(dir),
3098                                                        dx_root_bh);
3099                         ret = ocfs2_insert_extent(handle, &dx_et, 0,
3100                                                   dx_insert_blkno, 1, 0, NULL);
3101                         if (ret)
3102                                 mlog_errno(ret);
3103                 }
3104         }
3105
3106         /*
3107          * We asked for two clusters, but only got one in the 1st
3108          * pass. Claim the 2nd cluster as a separate extent.
3109          */
3110         if (alloc > len) {
3111                 ret = ocfs2_claim_clusters(handle, data_ac, 1, &bit_off,
3112                                            &len);
3113                 if (ret) {
3114                         mlog_errno(ret);
3115                         goto out_commit;
3116                 }
3117                 blkno = ocfs2_clusters_to_blocks(dir->i_sb, bit_off);
3118
3119                 ret = ocfs2_insert_extent(handle, &et, 1,
3120                                           blkno, len, 0, NULL);
3121                 if (ret) {
3122                         mlog_errno(ret);
3123                         goto out_commit;
3124                 }
3125                 bytes_allocated += ocfs2_clusters_to_bytes(dir->i_sb, 1);
3126         }
3127
3128         *first_block_bh = dirdata_bh;
3129         dirdata_bh = NULL;
3130         if (ocfs2_supports_indexed_dirs(osb)) {
3131                 unsigned int off;
3132
3133                 if (!dx_inline) {
3134                         /*
3135                          * We need to return the correct block within the
3136                          * cluster which should hold our entry.
3137                          */
3138                         off = ocfs2_dx_dir_hash_idx(OCFS2_SB(dir->i_sb),
3139                                                     &lookup->dl_hinfo);
3140                         get_bh(dx_leaves[off]);
3141                         lookup->dl_dx_leaf_bh = dx_leaves[off];
3142                 }
3143                 lookup->dl_dx_root_bh = dx_root_bh;
3144                 dx_root_bh = NULL;
3145         }
3146
3147 out_commit:
3148         if (ret < 0 && did_quota)
3149                 dquot_free_space_nodirty(dir, bytes_allocated);
3150
3151         ocfs2_commit_trans(osb, handle);
3152
3153 out:
3154         up_write(&oi->ip_alloc_sem);
3155         if (data_ac)
3156                 ocfs2_free_alloc_context(data_ac);
3157         if (meta_ac)
3158                 ocfs2_free_alloc_context(meta_ac);
3159
3160         if (dx_leaves) {
3161                 for (i = 0; i < num_dx_leaves; i++)
3162                         brelse(dx_leaves[i]);
3163                 kfree(dx_leaves);
3164         }
3165
3166         brelse(dirdata_bh);
3167         brelse(dx_root_bh);
3168
3169         return ret;
3170 }
3171
3172 /* returns a bh of the 1st new block in the allocation. */
3173 static int ocfs2_do_extend_dir(struct super_block *sb,
3174                                handle_t *handle,
3175                                struct inode *dir,
3176                                struct buffer_head *parent_fe_bh,
3177                                struct ocfs2_alloc_context *data_ac,
3178                                struct ocfs2_alloc_context *meta_ac,
3179                                struct buffer_head **new_bh)
3180 {
3181         int status;
3182         int extend, did_quota = 0;
3183         u64 p_blkno, v_blkno;
3184
3185         spin_lock(&OCFS2_I(dir)->ip_lock);
3186         extend = (i_size_read(dir) == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters));
3187         spin_unlock(&OCFS2_I(dir)->ip_lock);
3188
3189         if (extend) {
3190                 u32 offset = OCFS2_I(dir)->ip_clusters;
3191
3192                 status = dquot_alloc_space_nodirty(dir,
3193                                         ocfs2_clusters_to_bytes(sb, 1));
3194                 if (status)
3195                         goto bail;
3196                 did_quota = 1;
3197
3198                 status = ocfs2_add_inode_data(OCFS2_SB(sb), dir, &offset,
3199                                               1, 0, parent_fe_bh, handle,
3200                                               data_ac, meta_ac, NULL);
3201                 BUG_ON(status == -EAGAIN);
3202                 if (status < 0) {
3203                         mlog_errno(status);
3204                         goto bail;
3205                 }
3206         }
3207
3208         v_blkno = ocfs2_blocks_for_bytes(sb, i_size_read(dir));
3209         status = ocfs2_extent_map_get_blocks(dir, v_blkno, &p_blkno, NULL, NULL);
3210         if (status < 0) {
3211                 mlog_errno(status);
3212                 goto bail;
3213         }
3214
3215         *new_bh = sb_getblk(sb, p_blkno);
3216         if (!*new_bh) {
3217                 status = -EIO;
3218                 mlog_errno(status);
3219                 goto bail;
3220         }
3221         status = 0;
3222 bail:
3223         if (did_quota && status < 0)
3224                 dquot_free_space_nodirty(dir, ocfs2_clusters_to_bytes(sb, 1));
3225         return status;
3226 }
3227
3228 /*
3229  * Assumes you already have a cluster lock on the directory.
3230  *
3231  * 'blocks_wanted' is only used if we have an inline directory which
3232  * is to be turned into an extent based one. The size of the dirent to
3233  * insert might be larger than the space gained by growing to just one
3234  * block, so we may have to grow the inode by two blocks in that case.
3235  *
3236  * If the directory is already indexed, dx_root_bh must be provided.
3237  */
3238 static int ocfs2_extend_dir(struct ocfs2_super *osb,
3239                             struct inode *dir,
3240                             struct buffer_head *parent_fe_bh,
3241                             unsigned int blocks_wanted,
3242                             struct ocfs2_dir_lookup_result *lookup,
3243                             struct buffer_head **new_de_bh)
3244 {
3245         int status = 0;
3246         int credits, num_free_extents, drop_alloc_sem = 0;
3247         loff_t dir_i_size;
3248         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) parent_fe_bh->b_data;
3249         struct ocfs2_extent_list *el = &fe->id2.i_list;
3250         struct ocfs2_alloc_context *data_ac = NULL;
3251         struct ocfs2_alloc_context *meta_ac = NULL;
3252         handle_t *handle = NULL;
3253         struct buffer_head *new_bh = NULL;
3254         struct ocfs2_dir_entry * de;
3255         struct super_block *sb = osb->sb;
3256         struct ocfs2_extent_tree et;
3257         struct buffer_head *dx_root_bh = lookup->dl_dx_root_bh;
3258
3259         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
3260                 /*
3261                  * This would be a code error as an inline directory should
3262                  * never have an index root.
3263                  */
3264                 BUG_ON(dx_root_bh);
3265
3266                 status = ocfs2_expand_inline_dir(dir, parent_fe_bh,
3267                                                  blocks_wanted, lookup,
3268                                                  &new_bh);
3269                 if (status) {
3270                         mlog_errno(status);
3271                         goto bail;
3272                 }
3273
3274                 /* Expansion from inline to an indexed directory will
3275                  * have given us this. */
3276                 dx_root_bh = lookup->dl_dx_root_bh;
3277
3278                 if (blocks_wanted == 1) {
3279                         /*
3280                          * If the new dirent will fit inside the space
3281                          * created by pushing out to one block, then
3282                          * we can complete the operation
3283                          * here. Otherwise we have to expand i_size
3284                          * and format the 2nd block below.
3285                          */
3286                         BUG_ON(new_bh == NULL);
3287                         goto bail_bh;
3288                 }
3289
3290                 /*
3291                  * Get rid of 'new_bh' - we want to format the 2nd
3292                  * data block and return that instead.
3293                  */
3294                 brelse(new_bh);
3295                 new_bh = NULL;
3296
3297                 down_write(&OCFS2_I(dir)->ip_alloc_sem);
3298                 drop_alloc_sem = 1;
3299                 dir_i_size = i_size_read(dir);
3300                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3301                 goto do_extend;
3302         }
3303
3304         down_write(&OCFS2_I(dir)->ip_alloc_sem);
3305         drop_alloc_sem = 1;
3306         dir_i_size = i_size_read(dir);
3307         trace_ocfs2_extend_dir((unsigned long long)OCFS2_I(dir)->ip_blkno,
3308                                dir_i_size);
3309
3310         /* dir->i_size is always block aligned. */
3311         spin_lock(&OCFS2_I(dir)->ip_lock);
3312         if (dir_i_size == ocfs2_clusters_to_bytes(sb, OCFS2_I(dir)->ip_clusters)) {
3313                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3314                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(dir),
3315                                               parent_fe_bh);
3316                 num_free_extents = ocfs2_num_free_extents(osb, &et);
3317                 if (num_free_extents < 0) {
3318                         status = num_free_extents;
3319                         mlog_errno(status);
3320                         goto bail;
3321                 }
3322
3323                 if (!num_free_extents) {
3324                         status = ocfs2_reserve_new_metadata(osb, el, &meta_ac);
3325                         if (status < 0) {
3326                                 if (status != -ENOSPC)
3327                                         mlog_errno(status);
3328                                 goto bail;
3329                         }
3330                 }
3331
3332                 status = ocfs2_reserve_clusters(osb, 1, &data_ac);
3333                 if (status < 0) {
3334                         if (status != -ENOSPC)
3335                                 mlog_errno(status);
3336                         goto bail;
3337                 }
3338
3339                 if (ocfs2_dir_resv_allowed(osb))
3340                         data_ac->ac_resv = &OCFS2_I(dir)->ip_la_data_resv;
3341
3342                 credits = ocfs2_calc_extend_credits(sb, el, 1);
3343         } else {
3344                 spin_unlock(&OCFS2_I(dir)->ip_lock);
3345                 credits = OCFS2_SIMPLE_DIR_EXTEND_CREDITS;
3346         }
3347
3348 do_extend:
3349         if (ocfs2_dir_indexed(dir))
3350                 credits++; /* For attaching the new dirent block to the
3351                             * dx_root */
3352
3353         handle = ocfs2_start_trans(osb, credits);
3354         if (IS_ERR(handle)) {
3355                 status = PTR_ERR(handle);
3356                 handle = NULL;
3357                 mlog_errno(status);
3358                 goto bail;
3359         }
3360
3361         status = ocfs2_do_extend_dir(osb->sb, handle, dir, parent_fe_bh,
3362                                      data_ac, meta_ac, &new_bh);
3363         if (status < 0) {
3364                 mlog_errno(status);
3365                 goto bail;
3366         }
3367
3368         ocfs2_set_new_buffer_uptodate(INODE_CACHE(dir), new_bh);
3369
3370         status = ocfs2_journal_access_db(handle, INODE_CACHE(dir), new_bh,
3371                                          OCFS2_JOURNAL_ACCESS_CREATE);
3372         if (status < 0) {
3373                 mlog_errno(status);
3374                 goto bail;
3375         }
3376         memset(new_bh->b_data, 0, sb->s_blocksize);
3377
3378         de = (struct ocfs2_dir_entry *) new_bh->b_data;
3379         de->inode = 0;
3380         if (ocfs2_supports_dir_trailer(dir)) {
3381                 de->rec_len = cpu_to_le16(ocfs2_dir_trailer_blk_off(sb));
3382
3383                 ocfs2_init_dir_trailer(dir, new_bh, le16_to_cpu(de->rec_len));
3384
3385                 if (ocfs2_dir_indexed(dir)) {
3386                         status = ocfs2_dx_dir_link_trailer(dir, handle,
3387                                                            dx_root_bh, new_bh);
3388                         if (status) {
3389                                 mlog_errno(status);
3390                                 goto bail;
3391                         }
3392                 }
3393         } else {
3394                 de->rec_len = cpu_to_le16(sb->s_blocksize);
3395         }
3396         ocfs2_journal_dirty(handle, new_bh);
3397
3398         dir_i_size += dir->i_sb->s_blocksize;
3399         i_size_write(dir, dir_i_size);
3400         dir->i_blocks = ocfs2_inode_sector_count(dir);
3401         status = ocfs2_mark_inode_dirty(handle, dir, parent_fe_bh);
3402         if (status < 0) {
3403                 mlog_errno(status);
3404                 goto bail;
3405         }
3406
3407 bail_bh:
3408         *new_de_bh = new_bh;
3409         get_bh(*new_de_bh);
3410 bail:
3411         if (handle)
3412                 ocfs2_commit_trans(osb, handle);
3413         if (drop_alloc_sem)
3414                 up_write(&OCFS2_I(dir)->ip_alloc_sem);
3415
3416         if (data_ac)
3417                 ocfs2_free_alloc_context(data_ac);
3418         if (meta_ac)
3419                 ocfs2_free_alloc_context(meta_ac);
3420
3421         brelse(new_bh);
3422
3423         return status;
3424 }
3425
3426 static int ocfs2_find_dir_space_id(struct inode *dir, struct buffer_head *di_bh,
3427                                    const char *name, int namelen,
3428                                    struct buffer_head **ret_de_bh,
3429                                    unsigned int *blocks_wanted)
3430 {
3431         int ret;
3432         struct super_block *sb = dir->i_sb;
3433         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
3434         struct ocfs2_dir_entry *de, *last_de = NULL;
3435         char *de_buf, *limit;
3436         unsigned long offset = 0;
3437         unsigned int rec_len, new_rec_len, free_space = dir->i_sb->s_blocksize;
3438
3439         /*
3440          * This calculates how many free bytes we'd have in block zero, should
3441          * this function force expansion to an extent tree.
3442          */
3443         if (ocfs2_new_dir_wants_trailer(dir))
3444                 free_space = ocfs2_dir_trailer_blk_off(sb) - i_size_read(dir);
3445         else
3446                 free_space = dir->i_sb->s_blocksize - i_size_read(dir);
3447
3448         de_buf = di->id2.i_data.id_data;
3449         limit = de_buf + i_size_read(dir);
3450         rec_len = OCFS2_DIR_REC_LEN(namelen);
3451
3452         while (de_buf < limit) {
3453                 de = (struct ocfs2_dir_entry *)de_buf;
3454
3455                 if (!ocfs2_check_dir_entry(dir, de, di_bh, offset)) {
3456                         ret = -ENOENT;
3457                         goto out;
3458                 }
3459                 if (ocfs2_match(namelen, name, de)) {
3460                         ret = -EEXIST;
3461                         goto out;
3462                 }
3463                 /*
3464                  * No need to check for a trailing dirent record here as
3465                  * they're not used for inline dirs.
3466                  */
3467
3468                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3469                         /* Ok, we found a spot. Return this bh and let
3470                          * the caller actually fill it in. */
3471                         *ret_de_bh = di_bh;
3472                         get_bh(*ret_de_bh);
3473                         ret = 0;
3474                         goto out;
3475                 }
3476
3477                 last_de = de;
3478                 de_buf += le16_to_cpu(de->rec_len);
3479                 offset += le16_to_cpu(de->rec_len);
3480         }
3481
3482         /*
3483          * We're going to require expansion of the directory - figure
3484          * out how many blocks we'll need so that a place for the
3485          * dirent can be found.
3486          */
3487         *blocks_wanted = 1;
3488         new_rec_len = le16_to_cpu(last_de->rec_len) + free_space;
3489         if (new_rec_len < (rec_len + OCFS2_DIR_REC_LEN(last_de->name_len)))
3490                 *blocks_wanted = 2;
3491
3492         ret = -ENOSPC;
3493 out:
3494         return ret;
3495 }
3496
3497 static int ocfs2_find_dir_space_el(struct inode *dir, const char *name,
3498                                    int namelen, struct buffer_head **ret_de_bh)
3499 {
3500         unsigned long offset;
3501         struct buffer_head *bh = NULL;
3502         unsigned short rec_len;
3503         struct ocfs2_dir_entry *de;
3504         struct super_block *sb = dir->i_sb;
3505         int status;
3506         int blocksize = dir->i_sb->s_blocksize;
3507
3508         status = ocfs2_read_dir_block(dir, 0, &bh, 0);
3509         if (status) {
3510                 mlog_errno(status);
3511                 goto bail;
3512         }
3513
3514         rec_len = OCFS2_DIR_REC_LEN(namelen);
3515         offset = 0;
3516         de = (struct ocfs2_dir_entry *) bh->b_data;
3517         while (1) {
3518                 if ((char *)de >= sb->s_blocksize + bh->b_data) {
3519                         brelse(bh);
3520                         bh = NULL;
3521
3522                         if (i_size_read(dir) <= offset) {
3523                                 /*
3524                                  * Caller will have to expand this
3525                                  * directory.
3526                                  */
3527                                 status = -ENOSPC;
3528                                 goto bail;
3529                         }
3530                         status = ocfs2_read_dir_block(dir,
3531                                              offset >> sb->s_blocksize_bits,
3532                                              &bh, 0);
3533                         if (status) {
3534                                 mlog_errno(status);
3535                                 goto bail;
3536                         }
3537                         /* move to next block */
3538                         de = (struct ocfs2_dir_entry *) bh->b_data;
3539                 }
3540                 if (!ocfs2_check_dir_entry(dir, de, bh, offset)) {
3541                         status = -ENOENT;
3542                         goto bail;
3543                 }
3544                 if (ocfs2_match(namelen, name, de)) {
3545                         status = -EEXIST;
3546                         goto bail;
3547                 }
3548
3549                 if (ocfs2_skip_dir_trailer(dir, de, offset % blocksize,
3550                                            blocksize))
3551                         goto next;
3552
3553                 if (ocfs2_dirent_would_fit(de, rec_len)) {
3554                         /* Ok, we found a spot. Return this bh and let
3555                          * the caller actually fill it in. */
3556                         *ret_de_bh = bh;
3557                         get_bh(*ret_de_bh);
3558                         status = 0;
3559                         goto bail;
3560                 }
3561 next:
3562                 offset += le16_to_cpu(de->rec_len);
3563                 de = (struct ocfs2_dir_entry *)((char *) de + le16_to_cpu(de->rec_len));
3564         }
3565
3566         status = 0;
3567 bail:
3568         brelse(bh);
3569         if (status)
3570                 mlog_errno(status);
3571
3572         return status;
3573 }
3574
3575 static int dx_leaf_sort_cmp(const void *a, const void *b)
3576 {
3577         const struct ocfs2_dx_entry *entry1 = a;
3578         const struct ocfs2_dx_entry *entry2 = b;
3579         u32 major_hash1 = le32_to_cpu(entry1->dx_major_hash);
3580         u32 major_hash2 = le32_to_cpu(entry2->dx_major_hash);
3581         u32 minor_hash1 = le32_to_cpu(entry1->dx_minor_hash);
3582         u32 minor_hash2 = le32_to_cpu(entry2->dx_minor_hash);
3583
3584         if (major_hash1 > major_hash2)
3585                 return 1;
3586         if (major_hash1 < major_hash2)
3587                 return -1;
3588
3589         /*
3590          * It is not strictly necessary to sort by minor
3591          */
3592         if (minor_hash1 > minor_hash2)
3593                 return 1;
3594         if (minor_hash1 < minor_hash2)
3595                 return -1;
3596         return 0;
3597 }
3598
3599 static void dx_leaf_sort_swap(void *a, void *b, int size)
3600 {
3601         struct ocfs2_dx_entry *entry1 = a;
3602         struct ocfs2_dx_entry *entry2 = b;
3603         struct ocfs2_dx_entry tmp;
3604
3605         BUG_ON(size != sizeof(*entry1));
3606
3607         tmp = *entry1;
3608         *entry1 = *entry2;
3609         *entry2 = tmp;
3610 }
3611
3612 static int ocfs2_dx_leaf_same_major(struct ocfs2_dx_leaf *dx_leaf)
3613 {
3614         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3615         int i, num = le16_to_cpu(dl_list->de_num_used);
3616
3617         for (i = 0; i < (num - 1); i++) {
3618                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) !=
3619                     le32_to_cpu(dl_list->de_entries[i + 1].dx_major_hash))
3620                         return 0;
3621         }
3622
3623         return 1;
3624 }
3625
3626 /*
3627  * Find the optimal value to split this leaf on. This expects the leaf
3628  * entries to be in sorted order.
3629  *
3630  * leaf_cpos is the cpos of the leaf we're splitting. insert_hash is
3631  * the hash we want to insert.
3632  *
3633  * This function is only concerned with the major hash - that which
3634  * determines which cluster an item belongs to.
3635  */
3636 static int ocfs2_dx_dir_find_leaf_split(struct ocfs2_dx_leaf *dx_leaf,
3637                                         u32 leaf_cpos, u32 insert_hash,
3638                                         u32 *split_hash)
3639 {
3640         struct ocfs2_dx_entry_list *dl_list = &dx_leaf->dl_list;
3641         int i, num_used = le16_to_cpu(dl_list->de_num_used);
3642         int allsame;
3643
3644         /*
3645          * There's a couple rare, but nasty corner cases we have to
3646          * check for here. All of them involve a leaf where all value
3647          * have the same hash, which is what we look for first.
3648          *
3649          * Most of the time, all of the above is false, and we simply
3650          * pick the median value for a split.
3651          */
3652         allsame = ocfs2_dx_leaf_same_major(dx_leaf);
3653         if (allsame) {
3654                 u32 val = le32_to_cpu(dl_list->de_entries[0].dx_major_hash);
3655
3656                 if (val == insert_hash) {
3657                         /*
3658                          * No matter where we would choose to split,
3659                          * the new entry would want to occupy the same
3660                          * block as these. Since there's no space left
3661                          * in their existing block, we know there
3662                          * won't be space after the split.
3663                          */
3664                         return -ENOSPC;
3665                 }
3666
3667                 if (val == leaf_cpos) {
3668                         /*
3669                          * Because val is the same as leaf_cpos (which
3670                          * is the smallest value this leaf can have),
3671                          * yet is not equal to insert_hash, then we
3672                          * know that insert_hash *must* be larger than
3673                          * val (and leaf_cpos). At least cpos+1 in value.
3674                          *
3675                          * We also know then, that there cannot be an
3676                          * adjacent extent (otherwise we'd be looking
3677                          * at it). Choosing this value gives us a
3678                          * chance to get some contiguousness.
3679                          */
3680                         *split_hash = leaf_cpos + 1;
3681                         return 0;
3682                 }
3683
3684                 if (val > insert_hash) {
3685                         /*
3686                          * val can not be the same as insert hash, and
3687                          * also must be larger than leaf_cpos. Also,
3688                          * we know that there can't be a leaf between
3689                          * cpos and val, otherwise the entries with
3690                          * hash 'val' would be there.
3691                          */
3692                         *split_hash = val;
3693                         return 0;
3694                 }
3695
3696                 *split_hash = insert_hash;
3697                 return 0;
3698         }
3699
3700         /*
3701          * Since the records are sorted and the checks above
3702          * guaranteed that not all records in this block are the same,
3703          * we simple travel forward, from the median, and pick the 1st
3704          * record whose value is larger than leaf_cpos.
3705          */
3706         for (i = (num_used / 2); i < num_used; i++)
3707                 if (le32_to_cpu(dl_list->de_entries[i].dx_major_hash) >
3708                     leaf_cpos)
3709                         break;
3710
3711         BUG_ON(i == num_used); /* Should be impossible */
3712         *split_hash = le32_to_cpu(dl_list->de_entries[i].dx_major_hash);
3713         return 0;
3714 }
3715
3716 /*
3717  * Transfer all entries in orig_dx_leaves whose major hash is equal to or
3718  * larger than split_hash into new_dx_leaves. We use a temporary
3719  * buffer (tmp_dx_leaf) to make the changes to the original leaf blocks.
3720  *
3721  * Since the block offset inside a leaf (cluster) is a constant mask
3722  * of minor_hash, we can optimize - an item at block offset X within
3723  * the original cluster, will be at offset X within the new cluster.
3724  */
3725 static void ocfs2_dx_dir_transfer_leaf(struct inode *dir, u32 split_hash,
3726                                        handle_t *handle,
3727                                        struct ocfs2_dx_leaf *tmp_dx_leaf,
3728                                        struct buffer_head **orig_dx_leaves,
3729                                        struct buffer_head **new_dx_leaves,
3730                                        int num_dx_leaves)
3731 {
3732         int i, j, num_used;
3733         u32 major_hash;
3734         struct ocfs2_dx_leaf *orig_dx_leaf, *new_dx_leaf;
3735         struct ocfs2_dx_entry_list *orig_list, *new_list, *tmp_list;
3736         struct ocfs2_dx_entry *dx_entry;
3737
3738         tmp_list = &tmp_dx_leaf->dl_list;
3739
3740         for (i = 0; i < num_dx_leaves; i++) {
3741                 orig_dx_leaf = (struct ocfs2_dx_leaf *) orig_dx_leaves[i]->b_data;
3742                 orig_list = &orig_dx_leaf->dl_list;
3743                 new_dx_leaf = (struct ocfs2_dx_leaf *) new_dx_leaves[i]->b_data;
3744                 new_list = &new_dx_leaf->dl_list;
3745
3746                 num_used = le16_to_cpu(orig_list->de_num_used);
3747
3748                 memcpy(tmp_dx_leaf, orig_dx_leaf, dir->i_sb->s_blocksize);
3749                 tmp_list->de_num_used = cpu_to_le16(0);
3750                 memset(&tmp_list->de_entries, 0, sizeof(*dx_entry)*num_used);
3751
3752                 for (j = 0; j < num_used; j++) {
3753                         dx_entry = &orig_list->de_entries[j];
3754                         major_hash = le32_to_cpu(dx_entry->dx_major_hash);
3755                         if (major_hash >= split_hash)
3756                                 ocfs2_dx_dir_leaf_insert_tail(new_dx_leaf,
3757                                                               dx_entry);
3758                         else
3759                                 ocfs2_dx_dir_leaf_insert_tail(tmp_dx_leaf,
3760                                                               dx_entry);
3761                 }
3762                 memcpy(orig_dx_leaf, tmp_dx_leaf, dir->i_sb->s_blocksize);
3763
3764                 ocfs2_journal_dirty(handle, orig_dx_leaves[i]);
3765                 ocfs2_journal_dirty(handle, new_dx_leaves[i]);
3766         }
3767 }
3768
3769 static int ocfs2_dx_dir_rebalance_credits(struct ocfs2_super *osb,
3770                                           struct ocfs2_dx_root_block *dx_root)
3771 {
3772         int credits = ocfs2_clusters_to_blocks(osb->sb, 2);
3773
3774         credits += ocfs2_calc_extend_credits(osb->sb, &dx_root->dr_list, 1);
3775         credits += ocfs2_quota_trans_credits(osb->sb);
3776         return credits;
3777 }
3778
3779 /*
3780  * Find the median value in dx_leaf_bh and allocate a new leaf to move
3781  * half our entries into.
3782  */
3783 static int ocfs2_dx_dir_rebalance(struct ocfs2_super *osb, struct inode *dir,
3784                                   struct buffer_head *dx_root_bh,
3785                                   struct buffer_head *dx_leaf_bh,
3786                                   struct ocfs2_dx_hinfo *hinfo, u32 leaf_cpos,
3787                                   u64 leaf_blkno)
3788 {
3789         struct ocfs2_dx_leaf *dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
3790         int credits, ret, i, num_used, did_quota = 0;
3791         u32 cpos, split_hash, insert_hash = hinfo->major_hash;
3792         u64 orig_leaves_start;
3793         int num_dx_leaves;
3794         struct buffer_head **orig_dx_leaves = NULL;
3795         struct buffer_head **new_dx_leaves = NULL;
3796         struct ocfs2_alloc_context *data_ac = NULL, *meta_ac = NULL;
3797         struct ocfs2_extent_tree et;
3798         handle_t *handle = NULL;
3799         struct ocfs2_dx_root_block *dx_root;
3800         struct ocfs2_dx_leaf *tmp_dx_leaf = NULL;
3801
3802         trace_ocfs2_dx_dir_rebalance((unsigned long long)OCFS2_I(dir)->ip_blkno,
3803                                      (unsigned long long)leaf_blkno,
3804                                      insert_hash);
3805
3806         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
3807
3808         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3809         /*
3810          * XXX: This is a rather large limit. We should use a more
3811          * realistic value.
3812          */
3813         if (le32_to_cpu(dx_root->dr_clusters) == UINT_MAX)
3814                 return -ENOSPC;
3815
3816         num_used = le16_to_cpu(dx_leaf->dl_list.de_num_used);
3817         if (num_used < le16_to_cpu(dx_leaf->dl_list.de_count)) {
3818                 mlog(ML_ERROR, "DX Dir: %llu, Asked to rebalance empty leaf: "
3819                      "%llu, %d\n", (unsigned long long)OCFS2_I(dir)->ip_blkno,
3820                      (unsigned long long)leaf_blkno, num_used);
3821                 ret = -EIO;
3822                 goto out;
3823         }
3824
3825         orig_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
3826         if (!orig_dx_leaves) {
3827                 ret = -ENOMEM;
3828                 mlog_errno(ret);
3829                 goto out;
3830         }
3831
3832         new_dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, NULL);
3833         if (!new_dx_leaves) {
3834                 ret = -ENOMEM;
3835                 mlog_errno(ret);
3836                 goto out;
3837         }
3838
3839         ret = ocfs2_lock_allocators(dir, &et, 1, 0, &data_ac, &meta_ac);
3840         if (ret) {
3841                 if (ret != -ENOSPC)
3842                         mlog_errno(ret);
3843                 goto out;
3844         }
3845
3846         credits = ocfs2_dx_dir_rebalance_credits(osb, dx_root);
3847         handle = ocfs2_start_trans(osb, credits);
3848         if (IS_ERR(handle)) {
3849                 ret = PTR_ERR(handle);
3850                 handle = NULL;
3851                 mlog_errno(ret);
3852                 goto out;
3853         }
3854
3855         ret = dquot_alloc_space_nodirty(dir,
3856                                        ocfs2_clusters_to_bytes(dir->i_sb, 1));
3857         if (ret)
3858                 goto out_commit;
3859         did_quota = 1;
3860
3861         ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir), dx_leaf_bh,
3862                                       OCFS2_JOURNAL_ACCESS_WRITE);
3863         if (ret) {
3864                 mlog_errno(ret);
3865                 goto out_commit;
3866         }
3867
3868         /*
3869          * This block is changing anyway, so we can sort it in place.
3870          */
3871         sort(dx_leaf->dl_list.de_entries, num_used,
3872              sizeof(struct ocfs2_dx_entry), dx_leaf_sort_cmp,
3873              dx_leaf_sort_swap);
3874
3875         ocfs2_journal_dirty(handle, dx_leaf_bh);
3876
3877         ret = ocfs2_dx_dir_find_leaf_split(dx_leaf, leaf_cpos, insert_hash,
3878                                            &split_hash);
3879         if (ret) {
3880                 mlog_errno(ret);
3881                 goto  out_commit;
3882         }
3883
3884         trace_ocfs2_dx_dir_rebalance_split(leaf_cpos, split_hash, insert_hash);
3885
3886         /*
3887          * We have to carefully order operations here. There are items
3888          * which want to be in the new cluster before insert, but in
3889          * order to put those items in the new cluster, we alter the
3890          * old cluster. A failure to insert gets nasty.
3891          *
3892          * So, start by reserving writes to the old
3893          * cluster. ocfs2_dx_dir_new_cluster will reserve writes on
3894          * the new cluster for us, before inserting it. The insert
3895          * won't happen if there's an error before that. Once the
3896          * insert is done then, we can transfer from one leaf into the
3897          * other without fear of hitting any error.
3898          */
3899
3900         /*
3901          * The leaf transfer wants some scratch space so that we don't
3902          * wind up doing a bunch of expensive memmove().
3903          */
3904         tmp_dx_leaf = kmalloc(osb->sb->s_blocksize, GFP_NOFS);
3905         if (!tmp_dx_leaf) {
3906                 ret = -ENOMEM;
3907                 mlog_errno(ret);
3908                 goto out_commit;
3909         }
3910
3911         orig_leaves_start = ocfs2_block_to_cluster_start(dir->i_sb, leaf_blkno);
3912         ret = ocfs2_read_dx_leaves(dir, orig_leaves_start, num_dx_leaves,
3913                                    orig_dx_leaves);
3914         if (ret) {
3915                 mlog_errno(ret);
3916                 goto out_commit;
3917         }
3918
3919         cpos = split_hash;
3920         ret = ocfs2_dx_dir_new_cluster(dir, &et, cpos, handle,
3921                                        data_ac, meta_ac, new_dx_leaves,
3922                                        num_dx_leaves);
3923         if (ret) {
3924                 mlog_errno(ret);
3925                 goto out_commit;
3926         }
3927
3928         for (i = 0; i < num_dx_leaves; i++) {
3929                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3930                                               orig_dx_leaves[i],
3931                                               OCFS2_JOURNAL_ACCESS_WRITE);
3932                 if (ret) {
3933                         mlog_errno(ret);
3934                         goto out_commit;
3935                 }
3936
3937                 ret = ocfs2_journal_access_dl(handle, INODE_CACHE(dir),
3938                                               new_dx_leaves[i],
3939                                               OCFS2_JOURNAL_ACCESS_WRITE);
3940                 if (ret) {
3941                         mlog_errno(ret);
3942                         goto out_commit;
3943                 }
3944         }
3945
3946         ocfs2_dx_dir_transfer_leaf(dir, split_hash, handle, tmp_dx_leaf,
3947                                    orig_dx_leaves, new_dx_leaves, num_dx_leaves);
3948
3949 out_commit:
3950         if (ret < 0 && did_quota)
3951                 dquot_free_space_nodirty(dir,
3952                                 ocfs2_clusters_to_bytes(dir->i_sb, 1));
3953
3954         ocfs2_commit_trans(osb, handle);
3955
3956 out:
3957         if (orig_dx_leaves || new_dx_leaves) {
3958                 for (i = 0; i < num_dx_leaves; i++) {
3959                         if (orig_dx_leaves)
3960                                 brelse(orig_dx_leaves[i]);
3961                         if (new_dx_leaves)
3962                                 brelse(new_dx_leaves[i]);
3963                 }
3964                 kfree(orig_dx_leaves);
3965                 kfree(new_dx_leaves);
3966         }
3967
3968         if (meta_ac)
3969                 ocfs2_free_alloc_context(meta_ac);
3970         if (data_ac)
3971                 ocfs2_free_alloc_context(data_ac);
3972
3973         kfree(tmp_dx_leaf);
3974         return ret;
3975 }
3976
3977 static int ocfs2_find_dir_space_dx(struct ocfs2_super *osb, struct inode *dir,
3978                                    struct buffer_head *di_bh,
3979                                    struct buffer_head *dx_root_bh,
3980                                    const char *name, int namelen,
3981                                    struct ocfs2_dir_lookup_result *lookup)
3982 {
3983         int ret, rebalanced = 0;
3984         struct ocfs2_dx_root_block *dx_root;
3985         struct buffer_head *dx_leaf_bh = NULL;
3986         struct ocfs2_dx_leaf *dx_leaf;
3987         u64 blkno;
3988         u32 leaf_cpos;
3989
3990         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
3991
3992 restart_search:
3993         ret = ocfs2_dx_dir_lookup(dir, &dx_root->dr_list, &lookup->dl_hinfo,
3994                                   &leaf_cpos, &blkno);
3995         if (ret) {
3996                 mlog_errno(ret);
3997                 goto out;
3998         }
3999
4000         ret = ocfs2_read_dx_leaf(dir, blkno, &dx_leaf_bh);
4001         if (ret) {
4002                 mlog_errno(ret);
4003                 goto out;
4004         }
4005
4006         dx_leaf = (struct ocfs2_dx_leaf *)dx_leaf_bh->b_data;
4007
4008         if (le16_to_cpu(dx_leaf->dl_list.de_num_used) >=
4009             le16_to_cpu(dx_leaf->dl_list.de_count)) {
4010                 if (rebalanced) {
4011                         /*
4012                          * Rebalancing should have provided us with
4013                          * space in an appropriate leaf.
4014                          *
4015                          * XXX: Is this an abnormal condition then?
4016                          * Should we print a message here?
4017                          */
4018                         ret = -ENOSPC;
4019                         goto out;
4020                 }
4021
4022                 ret = ocfs2_dx_dir_rebalance(osb, dir, dx_root_bh, dx_leaf_bh,
4023                                              &lookup->dl_hinfo, leaf_cpos,
4024                                              blkno);
4025                 if (ret) {
4026                         if (ret != -ENOSPC)
4027                                 mlog_errno(ret);
4028                         goto out;
4029                 }
4030
4031                 /*
4032                  * Restart the lookup. The rebalance might have
4033                  * changed which block our item fits into. Mark our
4034                  * progress, so we only execute this once.
4035                  */
4036                 brelse(dx_leaf_bh);
4037                 dx_leaf_bh = NULL;
4038                 rebalanced = 1;
4039                 goto restart_search;
4040         }
4041
4042         lookup->dl_dx_leaf_bh = dx_leaf_bh;
4043         dx_leaf_bh = NULL;
4044
4045 out:
4046         brelse(dx_leaf_bh);
4047         return ret;
4048 }
4049
4050 static int ocfs2_search_dx_free_list(struct inode *dir,
4051                                      struct buffer_head *dx_root_bh,
4052                                      int namelen,
4053                                      struct ocfs2_dir_lookup_result *lookup)
4054 {
4055         int ret = -ENOSPC;
4056         struct buffer_head *leaf_bh = NULL, *prev_leaf_bh = NULL;
4057         struct ocfs2_dir_block_trailer *db;
4058         u64 next_block;
4059         int rec_len = OCFS2_DIR_REC_LEN(namelen);
4060         struct ocfs2_dx_root_block *dx_root;
4061
4062         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4063         next_block = le64_to_cpu(dx_root->dr_free_blk);
4064
4065         while (next_block) {
4066                 brelse(prev_leaf_bh);
4067                 prev_leaf_bh = leaf_bh;
4068                 leaf_bh = NULL;
4069
4070                 ret = ocfs2_read_dir_block_direct(dir, next_block, &leaf_bh);
4071                 if (ret) {
4072                         mlog_errno(ret);
4073                         goto out;
4074                 }
4075
4076                 db = ocfs2_trailer_from_bh(leaf_bh, dir->i_sb);
4077                 if (rec_len <= le16_to_cpu(db->db_free_rec_len)) {
4078                         lookup->dl_leaf_bh = leaf_bh;
4079                         lookup->dl_prev_leaf_bh = prev_leaf_bh;
4080                         leaf_bh = NULL;
4081                         prev_leaf_bh = NULL;
4082                         break;
4083                 }
4084
4085                 next_block = le64_to_cpu(db->db_free_next);
4086         }
4087
4088         if (!next_block)
4089                 ret = -ENOSPC;
4090
4091 out:
4092
4093         brelse(leaf_bh);
4094         brelse(prev_leaf_bh);
4095         return ret;
4096 }
4097
4098 static int ocfs2_expand_inline_dx_root(struct inode *dir,
4099                                        struct buffer_head *dx_root_bh)
4100 {
4101         int ret, num_dx_leaves, i, j, did_quota = 0;
4102         struct buffer_head **dx_leaves = NULL;
4103         struct ocfs2_extent_tree et;
4104         u64 insert_blkno;
4105         struct ocfs2_alloc_context *data_ac = NULL;
4106         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4107         handle_t *handle = NULL;
4108         struct ocfs2_dx_root_block *dx_root;
4109         struct ocfs2_dx_entry_list *entry_list;
4110         struct ocfs2_dx_entry *dx_entry;
4111         struct ocfs2_dx_leaf *target_leaf;
4112
4113         ret = ocfs2_reserve_clusters(osb, 1, &data_ac);
4114         if (ret) {
4115                 mlog_errno(ret);
4116                 goto out;
4117         }
4118
4119         dx_leaves = ocfs2_dx_dir_kmalloc_leaves(osb->sb, &num_dx_leaves);
4120         if (!dx_leaves) {
4121                 ret = -ENOMEM;
4122                 mlog_errno(ret);
4123                 goto out;
4124         }
4125
4126         handle = ocfs2_start_trans(osb, ocfs2_calc_dxi_expand_credits(osb->sb));
4127         if (IS_ERR(handle)) {
4128                 ret = PTR_ERR(handle);
4129                 mlog_errno(ret);
4130                 goto out;
4131         }
4132
4133         ret = dquot_alloc_space_nodirty(dir,
4134                                        ocfs2_clusters_to_bytes(osb->sb, 1));
4135         if (ret)
4136                 goto out_commit;
4137         did_quota = 1;
4138
4139         /*
4140          * We do this up front, before the allocation, so that a
4141          * failure to add the dx_root_bh to the journal won't result
4142          * us losing clusters.
4143          */
4144         ret = ocfs2_journal_access_dr(handle, INODE_CACHE(dir), dx_root_bh,
4145                                       OCFS2_JOURNAL_ACCESS_WRITE);
4146         if (ret) {
4147                 mlog_errno(ret);
4148                 goto out_commit;
4149         }
4150
4151         ret = __ocfs2_dx_dir_new_cluster(dir, 0, handle, data_ac, dx_leaves,
4152                                          num_dx_leaves, &insert_blkno);
4153         if (ret) {
4154                 mlog_errno(ret);
4155                 goto out_commit;
4156         }
4157
4158         /*
4159          * Transfer the entries from our dx_root into the appropriate
4160          * block
4161          */
4162         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4163         entry_list = &dx_root->dr_entries;
4164
4165         for (i = 0; i < le16_to_cpu(entry_list->de_num_used); i++) {
4166                 dx_entry = &entry_list->de_entries[i];
4167
4168                 j = __ocfs2_dx_dir_hash_idx(osb,
4169                                             le32_to_cpu(dx_entry->dx_minor_hash));
4170                 target_leaf = (struct ocfs2_dx_leaf *)dx_leaves[j]->b_data;
4171
4172                 ocfs2_dx_dir_leaf_insert_tail(target_leaf, dx_entry);
4173
4174                 /* Each leaf has been passed to the journal already
4175                  * via __ocfs2_dx_dir_new_cluster() */
4176         }
4177
4178         dx_root->dr_flags &= ~OCFS2_DX_FLAG_INLINE;
4179         memset(&dx_root->dr_list, 0, osb->sb->s_blocksize -
4180                offsetof(struct ocfs2_dx_root_block, dr_list));
4181         dx_root->dr_list.l_count =
4182                 cpu_to_le16(ocfs2_extent_recs_per_dx_root(osb->sb));
4183
4184         /* This should never fail considering we start with an empty
4185          * dx_root. */
4186         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4187         ret = ocfs2_insert_extent(handle, &et, 0, insert_blkno, 1, 0, NULL);
4188         if (ret)
4189                 mlog_errno(ret);
4190         did_quota = 0;
4191
4192         ocfs2_journal_dirty(handle, dx_root_bh);
4193
4194 out_commit:
4195         if (ret < 0 && did_quota)
4196                 dquot_free_space_nodirty(dir,
4197                                           ocfs2_clusters_to_bytes(dir->i_sb, 1));
4198
4199         ocfs2_commit_trans(osb, handle);
4200
4201 out:
4202         if (data_ac)
4203                 ocfs2_free_alloc_context(data_ac);
4204
4205         if (dx_leaves) {
4206                 for (i = 0; i < num_dx_leaves; i++)
4207                         brelse(dx_leaves[i]);
4208                 kfree(dx_leaves);
4209         }
4210         return ret;
4211 }
4212
4213 static int ocfs2_inline_dx_has_space(struct buffer_head *dx_root_bh)
4214 {
4215         struct ocfs2_dx_root_block *dx_root;
4216         struct ocfs2_dx_entry_list *entry_list;
4217
4218         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4219         entry_list = &dx_root->dr_entries;
4220
4221         if (le16_to_cpu(entry_list->de_num_used) >=
4222             le16_to_cpu(entry_list->de_count))
4223                 return -ENOSPC;
4224
4225         return 0;
4226 }
4227
4228 static int ocfs2_prepare_dx_dir_for_insert(struct inode *dir,
4229                                            struct buffer_head *di_bh,
4230                                            const char *name,
4231                                            int namelen,
4232                                            struct ocfs2_dir_lookup_result *lookup)
4233 {
4234         int ret, free_dx_root = 1;
4235         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4236         struct buffer_head *dx_root_bh = NULL;
4237         struct buffer_head *leaf_bh = NULL;
4238         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4239         struct ocfs2_dx_root_block *dx_root;
4240
4241         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4242         if (ret) {
4243                 mlog_errno(ret);
4244                 goto out;
4245         }
4246
4247         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4248         if (le32_to_cpu(dx_root->dr_num_entries) == OCFS2_DX_ENTRIES_MAX) {
4249                 ret = -ENOSPC;
4250                 mlog_errno(ret);
4251                 goto out;
4252         }
4253
4254         if (ocfs2_dx_root_inline(dx_root)) {
4255                 ret = ocfs2_inline_dx_has_space(dx_root_bh);
4256
4257                 if (ret == 0)
4258                         goto search_el;
4259
4260                 /*
4261                  * We ran out of room in the root block. Expand it to
4262                  * an extent, then allow ocfs2_find_dir_space_dx to do
4263                  * the rest.
4264                  */
4265                 ret = ocfs2_expand_inline_dx_root(dir, dx_root_bh);
4266                 if (ret) {
4267                         mlog_errno(ret);
4268                         goto out;
4269                 }
4270         }
4271
4272         /*
4273          * Insert preparation for an indexed directory is split into two
4274          * steps. The call to find_dir_space_dx reserves room in the index for
4275          * an additional item. If we run out of space there, it's a real error
4276          * we can't continue on.
4277          */
4278         ret = ocfs2_find_dir_space_dx(osb, dir, di_bh, dx_root_bh, name,
4279                                       namelen, lookup);
4280         if (ret) {
4281                 mlog_errno(ret);
4282                 goto out;
4283         }
4284
4285 search_el:
4286         /*
4287          * Next, we need to find space in the unindexed tree. This call
4288          * searches using the free space linked list. If the unindexed tree
4289          * lacks sufficient space, we'll expand it below. The expansion code
4290          * is smart enough to add any new blocks to the free space list.
4291          */
4292         ret = ocfs2_search_dx_free_list(dir, dx_root_bh, namelen, lookup);
4293         if (ret && ret != -ENOSPC) {
4294                 mlog_errno(ret);
4295                 goto out;
4296         }
4297
4298         /* Do this up here - ocfs2_extend_dir might need the dx_root */
4299         lookup->dl_dx_root_bh = dx_root_bh;
4300         free_dx_root = 0;
4301
4302         if (ret == -ENOSPC) {
4303                 ret = ocfs2_extend_dir(osb, dir, di_bh, 1, lookup, &leaf_bh);
4304
4305                 if (ret) {
4306                         mlog_errno(ret);
4307                         goto out;
4308                 }
4309
4310                 /*
4311                  * We make the assumption here that new leaf blocks are added
4312                  * to the front of our free list.
4313                  */
4314                 lookup->dl_prev_leaf_bh = NULL;
4315                 lookup->dl_leaf_bh = leaf_bh;
4316         }
4317
4318 out:
4319         if (free_dx_root)
4320                 brelse(dx_root_bh);
4321         return ret;
4322 }
4323
4324 /*
4325  * Get a directory ready for insert. Any directory allocation required
4326  * happens here. Success returns zero, and enough context in the dir
4327  * lookup result that ocfs2_add_entry() will be able complete the task
4328  * with minimal performance impact.
4329  */
4330 int ocfs2_prepare_dir_for_insert(struct ocfs2_super *osb,
4331                                  struct inode *dir,
4332                                  struct buffer_head *parent_fe_bh,
4333                                  const char *name,
4334                                  int namelen,
4335                                  struct ocfs2_dir_lookup_result *lookup)
4336 {
4337         int ret;
4338         unsigned int blocks_wanted = 1;
4339         struct buffer_head *bh = NULL;
4340
4341         trace_ocfs2_prepare_dir_for_insert(
4342                 (unsigned long long)OCFS2_I(dir)->ip_blkno, namelen);
4343
4344         if (!namelen) {
4345                 ret = -EINVAL;
4346                 mlog_errno(ret);
4347                 goto out;
4348         }
4349
4350         /*
4351          * Do this up front to reduce confusion.
4352          *
4353          * The directory might start inline, then be turned into an
4354          * indexed one, in which case we'd need to hash deep inside
4355          * ocfs2_find_dir_space_id(). Since
4356          * ocfs2_prepare_dx_dir_for_insert() also needs this hash
4357          * done, there seems no point in spreading out the calls. We
4358          * can optimize away the case where the file system doesn't
4359          * support indexing.
4360          */
4361         if (ocfs2_supports_indexed_dirs(osb))
4362                 ocfs2_dx_dir_name_hash(dir, name, namelen, &lookup->dl_hinfo);
4363
4364         if (ocfs2_dir_indexed(dir)) {
4365                 ret = ocfs2_prepare_dx_dir_for_insert(dir, parent_fe_bh,
4366                                                       name, namelen, lookup);
4367                 if (ret)
4368                         mlog_errno(ret);
4369                 goto out;
4370         }
4371
4372         if (OCFS2_I(dir)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
4373                 ret = ocfs2_find_dir_space_id(dir, parent_fe_bh, name,
4374                                               namelen, &bh, &blocks_wanted);
4375         } else
4376                 ret = ocfs2_find_dir_space_el(dir, name, namelen, &bh);
4377
4378         if (ret && ret != -ENOSPC) {
4379                 mlog_errno(ret);
4380                 goto out;
4381         }
4382
4383         if (ret == -ENOSPC) {
4384                 /*
4385                  * We have to expand the directory to add this name.
4386                  */
4387                 BUG_ON(bh);
4388
4389                 ret = ocfs2_extend_dir(osb, dir, parent_fe_bh, blocks_wanted,
4390                                        lookup, &bh);
4391                 if (ret) {
4392                         if (ret != -ENOSPC)
4393                                 mlog_errno(ret);
4394                         goto out;
4395                 }
4396
4397                 BUG_ON(!bh);
4398         }
4399
4400         lookup->dl_leaf_bh = bh;
4401         bh = NULL;
4402 out:
4403         brelse(bh);
4404         return ret;
4405 }
4406
4407 static int ocfs2_dx_dir_remove_index(struct inode *dir,
4408                                      struct buffer_head *di_bh,
4409                                      struct buffer_head *dx_root_bh)
4410 {
4411         int ret;
4412         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4413         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4414         struct ocfs2_dx_root_block *dx_root;
4415         struct inode *dx_alloc_inode = NULL;
4416         struct buffer_head *dx_alloc_bh = NULL;
4417         handle_t *handle;
4418         u64 blk;
4419         u16 bit;
4420         u64 bg_blkno;
4421
4422         dx_root = (struct ocfs2_dx_root_block *) dx_root_bh->b_data;
4423
4424         dx_alloc_inode = ocfs2_get_system_file_inode(osb,
4425                                         EXTENT_ALLOC_SYSTEM_INODE,
4426                                         le16_to_cpu(dx_root->dr_suballoc_slot));
4427         if (!dx_alloc_inode) {
4428                 ret = -ENOMEM;
4429                 mlog_errno(ret);
4430                 goto out;
4431         }
4432         mutex_lock(&dx_alloc_inode->i_mutex);
4433
4434         ret = ocfs2_inode_lock(dx_alloc_inode, &dx_alloc_bh, 1);
4435         if (ret) {
4436                 mlog_errno(ret);
4437                 goto out_mutex;
4438         }
4439
4440         handle = ocfs2_start_trans(osb, OCFS2_DX_ROOT_REMOVE_CREDITS);
4441         if (IS_ERR(handle)) {
4442                 ret = PTR_ERR(handle);
4443                 mlog_errno(ret);
4444                 goto out_unlock;
4445         }
4446
4447         ret = ocfs2_journal_access_di(handle, INODE_CACHE(dir), di_bh,
4448                                       OCFS2_JOURNAL_ACCESS_WRITE);
4449         if (ret) {
4450                 mlog_errno(ret);
4451                 goto out_commit;
4452         }
4453
4454         spin_lock(&OCFS2_I(dir)->ip_lock);
4455         OCFS2_I(dir)->ip_dyn_features &= ~OCFS2_INDEXED_DIR_FL;
4456         di->i_dyn_features = cpu_to_le16(OCFS2_I(dir)->ip_dyn_features);
4457         spin_unlock(&OCFS2_I(dir)->ip_lock);
4458         di->i_dx_root = cpu_to_le64(0ULL);
4459
4460         ocfs2_journal_dirty(handle, di_bh);
4461
4462         blk = le64_to_cpu(dx_root->dr_blkno);
4463         bit = le16_to_cpu(dx_root->dr_suballoc_bit);
4464         if (dx_root->dr_suballoc_loc)
4465                 bg_blkno = le64_to_cpu(dx_root->dr_suballoc_loc);
4466         else
4467                 bg_blkno = ocfs2_which_suballoc_group(blk, bit);
4468         ret = ocfs2_free_suballoc_bits(handle, dx_alloc_inode, dx_alloc_bh,
4469                                        bit, bg_blkno, 1);
4470         if (ret)
4471                 mlog_errno(ret);
4472
4473 out_commit:
4474         ocfs2_commit_trans(osb, handle);
4475
4476 out_unlock:
4477         ocfs2_inode_unlock(dx_alloc_inode, 1);
4478
4479 out_mutex:
4480         mutex_unlock(&dx_alloc_inode->i_mutex);
4481         brelse(dx_alloc_bh);
4482 out:
4483         iput(dx_alloc_inode);
4484         return ret;
4485 }
4486
4487 int ocfs2_dx_dir_truncate(struct inode *dir, struct buffer_head *di_bh)
4488 {
4489         int ret;
4490         unsigned int uninitialized_var(clen);
4491         u32 major_hash = UINT_MAX, p_cpos, uninitialized_var(cpos);
4492         u64 uninitialized_var(blkno);
4493         struct ocfs2_super *osb = OCFS2_SB(dir->i_sb);
4494         struct buffer_head *dx_root_bh = NULL;
4495         struct ocfs2_dx_root_block *dx_root;
4496         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
4497         struct ocfs2_cached_dealloc_ctxt dealloc;
4498         struct ocfs2_extent_tree et;
4499
4500         ocfs2_init_dealloc_ctxt(&dealloc);
4501
4502         if (!ocfs2_dir_indexed(dir))
4503                 return 0;
4504
4505         ret = ocfs2_read_dx_root(dir, di, &dx_root_bh);
4506         if (ret) {
4507                 mlog_errno(ret);
4508                 goto out;
4509         }
4510         dx_root = (struct ocfs2_dx_root_block *)dx_root_bh->b_data;
4511
4512         if (ocfs2_dx_root_inline(dx_root))
4513                 goto remove_index;
4514
4515         ocfs2_init_dx_root_extent_tree(&et, INODE_CACHE(dir), dx_root_bh);
4516
4517         /* XXX: What if dr_clusters is too large? */
4518         while (le32_to_cpu(dx_root->dr_clusters)) {
4519                 ret = ocfs2_dx_dir_lookup_rec(dir, &dx_root->dr_list,
4520                                               major_hash, &cpos, &blkno, &clen);
4521                 if (ret) {
4522                         mlog_errno(ret);
4523                         goto out;
4524                 }
4525
4526                 p_cpos = ocfs2_blocks_to_clusters(dir->i_sb, blkno);
4527
4528                 ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen, 0,
4529                                                &dealloc, 0);
4530                 if (ret) {
4531                         mlog_errno(ret);
4532                         goto out;
4533                 }
4534
4535                 if (cpos == 0)
4536                         break;
4537
4538                 major_hash = cpos - 1;
4539         }
4540
4541 remove_index:
4542         ret = ocfs2_dx_dir_remove_index(dir, di_bh, dx_root_bh);
4543         if (ret) {
4544                 mlog_errno(ret);
4545                 goto out;
4546         }
4547
4548         ocfs2_remove_from_cache(INODE_CACHE(dir), dx_root_bh);
4549 out:
4550         ocfs2_schedule_truncate_log_flush(osb, 1);
4551         ocfs2_run_deallocs(osb, &dealloc);
4552
4553         brelse(dx_root_bh);
4554         return ret;
4555 }