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